betaFeedback

WorksMathematics

Chow Stability of Polarized K3 Surfaces with ADE singularities

Abstract

The asymptotic Chow stability of smooth polarized K3 surfaces is known through the existence of Ricci-flat metrics and Donaldson's theorem. We give a purely algebraic proof and extend the result to polarized K3 surfaces with at worst ADE singularities. The proof gives a uniform effective bound for the exponent, depending only on the degree of the polarization. Following the approach of Mumford and Morrison, we establish quantitative estimates for linear systems and their contributions to Chow weights, without restrictions on the Picard number.

Introduction

The study of GIT stability provides a natural approach to constructing moduli spaces and their compactifications. For curves, the classical results of Mumford and Gieseker establish the stability of sufficiently high pluricanonical embeddings and yield a GIT construction of the moduli space of stable curves [14, 4]. For surfaces, the picture is more varied. Gieseker established asymptotic stability for canonically polarized surfaces [3], while Chow stability for anticanonically polarized surfaces exhibits more subtle behavior (cf. [9]).

The Ricci-flat case provides another natural setting for these questions. For abelian surfaces, Chow stability of embeddings defined by complete very ample linear systems admits a classical algebraic proof using theta groups and Kempf’s results [8]. For projective K3 surfaces, the situation is more involved:

  • In [17 Vie95], Viehweg constructed quasi-projective moduli spaces in arbitrary degree, using weak positivity and suitable GIT linearizations.

  • For polarized K3 surfaces of small degree, the following results are known:

    • Shah gave GIT constructions in degrees 2 and 4 [15, 16].

    • In [10], Li and Tian proved GIT stability for complete-intersection K3 surfaces with at worst ADE singularities in degrees 6 and 8.

  • In [12], Morrison proved asymptotic Hilbert stability for K3 surfaces of Picard number one embedded by primitive polarizations of degree at least 12.

  • In [2], Donaldson used Ricci-flat Kähler metrics and the asymptotic expansion of Bergman kernels to construct balanced embeddings of smooth polarized K3 surfaces for sufficiently high powers of the polarization.

As mentioned in [7], Donaldson’s approach does not identify the ample line bundles on the moduli space arising from the construction, nor does it clarify their relation to the ample line bundles obtained from other constructions. A purely algebraic proof of asymptotic Chow stability remains a natural problem. More generally, R. Pandharipande asked whether asymptotic Chow stability also holds for polarized K3 surfaces with ADE singularities.

In this paper, we solve this problem via a purely algebraic method. More precisely, let XX be a complex projective K3 surface with at worst ADE singularities, and let LL be an ample line bundle with L2=2dL^2=2d. For r≥3r\geq3, the complete linear system ∣rL∣|rL| defines an embedding

X↪PN−1,X\hookrightarrow\mathbb{P}^{N-1},

where N=h0(X,rL)=dr2+2N=h^0(X,rL)=dr^2+2.

Theorem 1.1. For every positive integer dd, there exists an integer r0=r0(d)r_0=r_0(d) such that, for every complex polarized K3 surface (X,L)(X,L) of degree 2d2d with at worst ADE singularities, the embedding defined by ∣L⊗r∣|L^{\otimes r}| is Chow stable for all r≥r0r\geq r_0.

Our proof gives the explicit choice

r0(d)=231(2d+1)3.r_0(d)=2^{31}(2d+1)^3.

For smooth K3 surfaces, one may instead take r0(d)=224(2d+1)3r_0(d)=2^{24}(2d+1)^3. These bounds are not expected to be optimal.

Remark 1.2. Chow stability also implies asymptotic Hilbert stability [13]. For each fixed (X,L)(X,L) and r≥r0(d)r\geq r_0(d), the qq-th Hilbert point of the embedding defined by L⊗rL^{\otimes r} is stable for all sufficiently large qq.

Let F2d\mathcal{F}_{2d} be the moduli stack of primitively polarized K3 surfaces of degree 2d2d with at worst ADE singularities, and let F2dF_{2d} be its coarse moduli space. By the global Torelli theorem, F2dF_{2d} is an arithmetic quotient of a type IV Hermitian symmetric domain, hence is quasi-projective by the Baily–Borel theorem. By the standard construction of moduli spaces from embedded polarized K3 surfaces, applied to the ADE locus (cf. [7 Chapter 5, §1.4 and §2]), Theorem 1.1 gives the following consequence.

Corollary 1.3. Fix an integer r≥r0(d)r\geq r_0(d), and let Zd,rZ_{d,r} be the reduced closure in the Chow variety of the locus of primitively polarized K3 surfaces (X,L)(X,L) of degree 2d2d with at worst ADE singularities, embedded by ∣L⊗r∣|L^{\otimes r}|. Then the Chow GIT quotient

Zd,r//SLdr2+2Z_{d,r}//\mathrm{SL}_{dr^2+2}

is a projective compactification of F2dF_{2d}.

Strategy of the Proof

Our proof follows the algebraic approach of Morrison [12], using Mumford’s intersection-theoretic estimates for Chow weights [14]. The main difficulty in passing from Picard number one to arbitrary Picard number is to control linear systems composed with a pencil and those for which Clifford’s inequality is nearly an equality. We bound the total contribution of the former and show that the latter, when their moving parts have positive self-intersection, occur only O(r)O(r) times in any strict flag of subspaces of H0(X,rL)H^{0}(X,rL). These estimates allow us to prove Chow stability without imposing a condition on the Picard number.

Organization of the paper

In § 2, we use Mumford’s numerical criterion to reduce Chow stability to an inequality comparing the sum of the weights with an integral associated with the flag. § 3 establishes the required weight estimates, treating moving parts of square zero and positive square separately. In § 4, we combine these estimates to prove Theorem 1.1 and obtain an explicit bound depending only on the degree.

Acknowledgements

The author is supported by NSFC grant 12425105 and Shanghai Pilot Program for Basic Research (No. 21TQ001). He is also a member of LMNS.

AI disclosure

The author used an LLM to assist in developing the proof. It provided many useful suggestions concerning the function used in the integral estimate in § 2, several estimates in § 3, and, in particular, the subspace counting argument in Subsection 3.4. The author takes full responsibility for the mathematical content of the paper.

Mumford’s Numerical Criterion

In this section, we recall Mumford’s numerical criterion and reduce Chow stability to an inequality comparing a sum of weights with an integral.

Weighted Flags and the Degree of Contact

Throughout, let (X,L)(X,L) be a complex polarized K3 surface with at worst ADE singularities and L2=2dL^{2}=2d. Fix r≥3r \ge3, and write N=h0(X,rL)=dr2+2N=h^{0}(X,rL)=dr^{2}+2. The complete linear system ∣rL∣|rL| embeds XX in PN−1\mathbb{P}^{N-1}, as recalled in [7] and [11 §5–6.1]. By the Hilbert–Mumford criterion, Chow stability is tested by one-parameter subgroups. We express their weights in terms of Mumford’s degree of contact.

Definition 2.1 (Weighted flags and degree of contact). Let λ\lambda be a nontrivial one-parameter subgroup of SL⁡(H0(X,rL))\operatorname{SL}(H^{0}(X,rL)), with eigenbasis z0,…,zN−1z_{0},\ldots,z_{N-1}. After translating and rescaling its weights, write

0=u0≤⋯≤uN−1=1,Vj=⟨z0,…,zj⟩,W=∑j=0N−1uj.0=u_{0}\le\cdots\le u_{N-1}=1,\qquad V_{j}=\langle z_{0},\ldots,z_{j}\rangle,\qquad W=\sum_{j=0}^{N-1}u_{j}.

These data define a weighted flag F\mathcal{F}. By its weight filtration, we mean the subspaces VjV_j at the ends of the maximal blocks of equal weights. Let

Jj=im⁡(Vj⊗OX(−rL)⟶OX)\mathcal{J}_j=\operatorname{im}(V_j\otimes\mathcal{O}_X(-rL)\longrightarrow\mathcal{O}_X)

be the associated base ideals. Choose a positive integer TT with Tuj∈ZTu_j\in\mathbb{Z}, and let p:X×A1⟶Xp:X\times\mathbb{A}^1\longrightarrow X be the projection. Put

J~=∑j=0N−1tTujp∗Jj,L=p∗OX(rL).\widetilde{\mathcal{J}}=\sum_{j=0}^{N-1}t^{Tu_j}p^*\mathcal{J}_j,\qquad\mathcal{L}=p^*\mathcal{O}_X(rL).

The normalized degree of contact e(F)e(\mathcal{F}) is defined, for k≫0k\gg0, by

χ(L⊗k/J~ kL⊗k)=Te(F)6k3+O(k2).(1)\chi\left(\mathcal{L}^{\otimes k}/\widetilde{\mathcal{J}}^{\,k}\mathcal{L}^{\otimes k}\right)=\frac{Te(\mathcal{F})}{6}k^3+O(k^2). \tag*{(1)}

Since tT∈J~t^T\in\widetilde{\mathcal{J}}, the quotient has proper support. The polynomial expansion follows from [14 Proposition 2.1 and Definition 2.2]. Replacing TT by qTqT amounts to the finite flat substitution t↦tqt\mapsto t^q, which multiplies the Euler characteristic by qq. Thus e(F)e(\mathcal{F}) is independent of TT.

Theorem 2.2. (Mumford [14 Theorem 2.9]). The complete embedding by ∣rL∣|rL| is Chow stable if and only if every normalized weighted flag satisfies

e(F)<3(rL)2NW.e(\mathcal{F})<\frac{3(rL)^2}{N}W.

We next bound the degree of contact using the moving parts of the flag.

The Degree of Contact and Moving Parts

Let π:X~⟶X\pi:\widetilde{X}\longrightarrow X be the minimal resolution. For a nonzero subspace V⊆H0(X,rL)V\subseteq H^0(X,rL), choose a sequence of point blowups σ:Y⟶X~\sigma:Y\longrightarrow\widetilde{X} resolving its base ideal. On YY, write A=B+DA=B+D, where AA represents the pullback of rLrL, and DD is the fixed part of the pulled-back linear system defined by VV. We call BB the moving part associated with VV. The induced evaluation map

V⊗OY⟶OY(B)V\otimes\mathcal{O}_Y\longrightarrow\mathcal{O}_Y(B)

is surjective, so BB is globally generated and nef. For a flag (Vj)(V_j), we choose a common resolution and write A=Bj+DjA=B_j+D_j for the corresponding decompositions.

Proposition 2.3. Let F\mathcal{F} be a normalized weighted flag as in Definition 2.1. There exist a sequence of point blowups σ:Y⟶X~\sigma:Y\longrightarrow\widetilde{X} and divisors A,B0,…,BN−1A,B_0,\ldots,B_{N-1} on YY such that

OY(A)=(π∘σ)∗OX(rL),\mathcal{O}_Y(A)=(\pi\circ\sigma)^*\mathcal{O}_X(rL),

where BjB_j is the moving part associated with VjV_j. In particular, the BjB_j are nef and satisfy

0=B0≤B1≤⋯≤BN−1=A,h0(Y,Bj)≥j+1.0=B_0\leq B_1\leq\cdots\leq B_{N-1}=A,\qquad h^0(Y,B_j)\geq j+1.

Moreover, for every subsequence 0=i0<⋯<iℓ=N−10=i_0<\cdots<i_\ell=N-1, one has

e(F)≤3A2−∑ν=0ℓ−1(uiν+1−uiν)(Biν2+Biν⋅Biν+1+Biν+12).(2)e(\mathcal{F}) \le3A^2-\sum_{\nu=0}^{\ell-1}(u_{i_{\nu+1}}-u_{i_\nu})(B_{i_\nu}^2+B_{i_\nu}\cdot B_{i_{\nu+1}}+B_{i_{\nu+1}}^2). \tag*{(2)}

Proof. Resolve the base ideals Jj\mathcal{J}_j simultaneously by point blowups σ:Y→X~\sigma:Y\to\widetilde{X}, and write

JjOY=OY(−Dj).\mathcal{J}_j\mathcal{O}_Y=\mathcal{O}_Y(-D_j).

Let AA be the divisor of the pullback of z0z_0, and set Bj=A−DjB_j=A-D_j. The evaluation maps give surjections

Vj⊗OY⟶OY(Bj).V_j\otimes\mathcal{O}_Y\longrightarrow\mathcal{O}_Y(B_j).

Thus the BjB_j are globally generated and nef, with h0(Y,Bj)≥j+1h^0(Y,B_j)\ge j+1. Since Jj⊆Jj+1\mathcal{J}_j\subseteq\mathcal{J}_{j+1}, we have Dj≥Dj+1D_j\ge D_{j+1}. Moreover, D0=AD_0=A and DN−1=0D_{N-1}=0, because V0=⟨z0⟩V_0=\langle z_0\rangle and ∣rL∣|rL| is base-point-free. This proves the assertions about the BjB_j.

Fix a subsequence 0=i0<⋯<iℓ=N−10=i_0<\cdots<i_\ell=N-1, and suppose first that its weights are distinct. Omitting subspaces from the flag can only increase the degree of contact. Applying Mumford’s estimates to this subsequence [14 Propositions 4.9–4.10 and Corollary 4.11], we obtain

e(F)≤∑ν=0ℓ−1(uiν+1−uiν)(eA(Jiν)+eA(Jiν,Jiν+1)+eA(Jiν+1)).e(\mathcal{F})\le\sum_{\nu=0}^{\ell-1}(u_{i_{\nu+1}}-u_{i_\nu})\left(e_A(\mathcal{J}_{i_\nu})+e_A(\mathcal{J}_{i_\nu},\mathcal{J}_{i_{\nu+1}})+e_A(\mathcal{J}_{i_{\nu+1}})\right).

where eAe_A denotes the global multiplicities and mixed multiplicities with respect to AA. By [14 Proposition 4.8 and Lemma 5.6],

eA(Ji)=A2−Bi2,eA(Ji,Jj)=A2−Bi⋅Bj,e_A(\mathcal{J}_i)=A^2-B_i^2,\qquad e_A(\mathcal{J}_i,\mathcal{J}_j)=A^2-B_i\cdot B_j,

with the mixed formula obtained by polarization. Substituting these identities and using ∑ν=0ℓ−1(uiν+1−uiν)=1\sum_{\nu=0}^{\ell-1}(u_{i_{\nu+1}}-u_{i_\nu})=1 gives (2).

For repeated weights, keep only the last subspace in each block of equal weights in the chosen subsequence. Indeed, if i<ji<j and ui=uju_i=u_j, then Vi⊆VjV_i\subseteq V_j, so ViV_i contributes no additional sections at that weight. Removing it therefore leaves the degree of contact of the subsequence unchanged. The remaining weights are distinct and still run from 0 to 1, so the preceding calculation applies.

To compare the resulting bound with the original sum, note that for i≤i′i\le i' and j≤j′j\le j',

Bi′⋅Bj′−Bi⋅Bj=(Bi′−Bi)⋅Bj′+Bi⋅(Bj′−Bj)≥0,B_{i'}\cdot B_{j'}-B_i\cdot B_j=(B_{i'}-B_i)\cdot B_{j'}+B_i\cdot(B_{j'}-B_j)\ge0,

since the differences are effective and the BjB_j are nef. Replacing each block by its last subspace therefore increases the intersection terms being subtracted, while terms between equal weights vanish. This proves (2) in general. ∎

A Convex Function Associated with the Flag

Fix YY and BjB_j as in Proposition 2.3, and put

a=A22=dr2=N−2,bj=Bj2.a = \frac{A^2}{2} = dr^2 = N - 2, \qquad b_j = B_j^2.

We associate with the weighted flag a function that records the largest self-intersection attainable by convex combinations of the moving parts, subject to a bound on their average weight. Convex combinations are needed to control the moving parts of square zero. Although such a divisor has zero self-intersection, its mixed intersection with a nef divisor of positive square can contribute to the square of their convex combination. We use this observation to estimate the contribution of pencils in Subsection 3.1.

Definition 2.4. For 0≤ρ≤10 \le\rho\le1, set

ϕ(ρ)=max⁡{Q22a:Q=∑jλjBj, λj≥0, ∑jλj=1, ∑jλjuj≤ρ},\phi(\rho) = \max\left\{ \sqrt{\frac{Q^2}{2a}} : Q = \sum_j \lambda_j B_j,\ \lambda_j \ge0,\ \sum_j \lambda_j = 1,\ \sum_j \lambda_j u_j \le\rho\right\},

and define

h(y)=min⁡{ρ∈[0,1]:ϕ(ρ)≥y},0≤y≤1.h(y) = \min\{\rho\in[0,1] : \phi(\rho) \ge y\}, \qquad0 \le y \le1.

Proposition 2.5. The function ϕ\phi is continuous, nondecreasing and concave, and satisfies

ρ≤ϕ(ρ)≤1,ϕ(1)=1.\rho\le\phi(\rho) \le1, \qquad\phi(1) = 1.

Its generalized inverse hh is continuous, nondecreasing and convex. Moreover,

h(0)=0,h(y)≤y,uj≥h(bj2a).(3)h(0) = 0, \qquad h(y) \le y, \qquad u_j \ge h\left(\sqrt{\frac{b_j}{2a}}\right). \tag*{(3)}

Proof. The admissible coefficient sets in the definition of ϕ\phi are compact and increase with ρ\rho. Thus ϕ\phi is nondecreasing. Concavity follows from

(P+Q)2≥P2+Q2\sqrt{(P+Q)^2} \ge\sqrt{P^2} + \sqrt{Q^2}

for nef divisor classes PP, QQ, by the Hodge index theorem.

Since B0=0B_0 = 0 and BN−1=AB_{N-1} = A, the divisor ρA\rho A is admissible for ϕ(ρ)\phi(\rho). On the other hand, every admissible QQ is nef and A−QA-Q is effective, so Q2≤A2=2aQ^2 \le A^2 = 2a. Hence

ρ≤ϕ(ρ)≤1.\rho\le\phi(\rho) \le1.

Compactness gives continuity at 00, while these bounds give continuity at 11. Concavity gives continuity in the interior. The asserted properties of hh now follow from the corresponding properties of the generalized inverse of a continuous, nondecreasing concave function.

Finally, BjB_j is admissible at ρ=uj\rho= u_j, so

ϕ(uj)≥bj2a.\phi(u_j) \ge\sqrt{\frac{b_j}{2a}}.

The definition of hh gives the last inequality in (3), while ϕ(ρ)≥ρ\phi(\rho) \ge\rho gives h(y)≤yh(y) \le y. ∎

Proposition 2.6. For every weighted flag as above,

e(F)≤6a∫01(1−ϕ(ρ)2) dρ=6a∫01h(x) dx.(4)e(\mathcal{F}) \le6a \int_0^1 (1-\phi(\rho)^2)\,d\rho= 6a\int_0^1 h(\sqrt{x})\,dx. \tag*{(4)}

Proof. Choose a rational partition 0=ρ0<⋯<ρℓ=10=\rho_0<\cdots<\rho_\ell=1. For each ii, choose rational coefficients λij\lambda_{ij} such that

Qi=∑jλijBj,λij≥0,∑jλij=1,∑jλijuj≤ρi,Q_i=\sum_j\lambda_{ij}B_j,\qquad\lambda_{ij}\ge0,\qquad\sum_j\lambda_{ij}=1,\qquad\sum_j\lambda_{ij}u_j\le\rho_i,

and Qi2/(2a)Q_i^2/(2a) approximates ϕ(ρi)2\phi(\rho_i)^2. Such approximations exist because the admissible coefficient sets are rational polytopes. At the endpoint, choose Qℓ=AQ_\ell=A. Take qq sufficiently divisible that all qλijq\lambda_{ij} and qρiq\rho_i are integral. The coefficient ideals of JF\mathfrak{J}^{\mathcal{F}} are

Kq,ν=∑∑jnj=qT∑jnjuj≤ν∏jJjnj.\mathcal{K}_{q,\nu}=\sum_{\substack{\sum_j n_j=q\\ T\sum_j n_ju_j\le\nu}}\prod_j\mathfrak{J}_j^{n_j}.

Let Wi⊆H0(X,qrL)W_i\subseteq H^0(X,qrL) be generated by products of qq basis sections of total weight at most qρiq\rho_i. Its base ideal is Kq,qTρi\mathcal{K}_{q,qT\rho_i}.

Resolve these ideals on a common resolution dominating YY, and denote the moving parts by CiC_i. We use the same notation for the pullbacks of AA, BjB_j, and QiQ_i. Since VjV_j generates OY(Bj)\mathcal{O}_Y(B_j), products with qλijq\lambda_{ij} factors from each VjV_j generate the line bundle associated with qQiqQ_i. They form a subsystem of WiW_i, so Ci−qQiC_i-qQ_i is effective. Since these divisors are nef,

Ci2≥q2Qi2,Ci⋅Ci+1≥q2Qi⋅Qi+1.C_i^2\ge q^2Q_i^2,\qquad C_i\cdot C_{i+1}\ge q^2Q_i\cdot Q_{i+1}.

At ρℓ=1\rho_\ell=1, all products are allowed and generate OX(qrL)\mathcal{O}_X(qrL), hence Cℓ=qAC_\ell=qA.

The sheaves Kq,ν⊗OX(qrL)\mathcal{K}_{q,\nu}\otimes\mathcal{O}_X(qrL) and JFL⊗q\mathfrak{J}^{\mathcal{F}}\mathcal{L}^{\otimes q} are generated by the corresponding products. Mumford’s surface estimate [14 Proposition 4.10 and Corollary 4.11] therefore applies. For fixed qq, equation (1) gives

χ(L⊗qk/JqkL⊗qk)=Tq3e(F)6k3+O(k2).\chi\left(\mathcal{L}^{\otimes qk}/\mathfrak{J}^{qk}\mathcal{L}^{\otimes qk}\right)=\frac{Tq^3e(\mathcal{F})}{6}k^3+O(k^2).

Thus the left-hand side of Mumford’s estimate is Tq3e(F)Tq^3e(\mathcal{F}), while the mixed multiplicities are

eqA(Kq,qTρi)=q2A2−Ci2,e_{qA}(\mathcal{K}_{q,qT\rho_i})=q^2A^2-C_i^2,
eqA(Kq,qTρi,Kq,qTρi+1)=q2A2−Ci⋅Ci+1.e_{qA}(\mathcal{K}_{q,qT\rho_i},\mathcal{K}_{q,qT\rho_{i+1}})=q^2A^2-C_i\cdot C_{i+1}.

Dividing the resulting estimate by Tq3Tq^3 gives

e(F)≤6a−∑i=0ℓ−1(ρi+1−ρi)(Qi2+Qi⋅Qi+1+Qi+12).e(\mathcal{F})\le6a-\sum_{i=0}^{\ell-1}(\rho_{i+1}-\rho_i)(Q_i^2+Q_i\cdot Q_{i+1}+Q_{i+1}^2).

By Hodge index, Qi⋅Qi+1≥Qi2Qi+12Q_i \cdot Q_{i+1} \ge\sqrt{Q_i^2Q_{i+1}^2}. We now let the approximation errors tend to zero and refine the partition. By continuity of ϕ\phi, we obtain

e(F)≤6a∫01(1−ϕ(ρ)2) dρ.e(\mathcal{F}) \le6a \int_0^1 (1-\phi(\rho)^2)\,d\rho.

Finally, h(x)≤ρh(\sqrt{x}) \le\rho if and only if x≤ϕ(ρ)2x \le\phi(\rho)^2. Fubini’s theorem gives the asserted equality of integrals. □

Combining Proposition 2.6 with Mumford’s criterion, we obtain the following sufficient condition.

Proposition 2.7. The embedding defined by ∣rL∣|rL| is Chow stable if every normalized weighted flag F\mathcal{F} satisfies

N∫01h(x) dx<W,(5)N\int_0^1 h(\sqrt{x})\,dx < W, \tag*{(5)}

where hh and WW are associated with F\mathcal{F}.

§ 3 gives geometric estimates for the weights, which are used in § 4 to verify (5).

Weight Estimates for Linear Subsystems

To verify (5), we compare values of hh with the weights uju_j. We treat moving parts of square zero and positive square separately, then collect the comparisons in Proposition 3.5. Throughout, BB denotes the moving part associated with a subspace VV, as defined in Subsection 2.2.

Linear Systems Composed with a Pencil

For the moving parts with B2=0B^2=0, we bound the total error in the comparison with the weights. Let PP be the largest j∈{1,…,a}j \in\{1,\ldots,a\} with Bj2=0B_j^2=0, setting P=0P=0 if there is none. We write (x)+=max⁡{x,0}(x)_+ = \max\{x,0\}.

Lemma 3.1. Put s=16(2d+1)s=16(2d+1). Assume

r≥32(s+1),a≥16s,s(1+log⁡(2dr))r≤1.(6)r \ge32(s+1), \qquad a \ge16s, \qquad\frac{s(1+\log(2dr))}{r} \le1. \tag*{(6)}

Then

∑j=1P(h(j+sa)−uj)+≤16[2d+s(1+log⁡(2dr))r]h(1/2)(7)\sum_{j=1}^{P}\left(h\left(\sqrt{\frac{j+s}{a}}\right)-u_j\right)_+ \le16\left[2d+\frac{s(1+\log(2dr))}{r}\right]h(1/2) \tag*{(7)}
≤sh(1/2).\le s h(1/2).

Proof. Suppose P≥1P \ge1, since otherwise there is nothing to prove. The inequalities $0 \le B_i \le B_jandnefnessimplythatand nefness imply thatB_j^2is nondecreasing. For 1≤i<j≤P\text{is nondecreasing. For }1 \le i < j \le P,

0≤Bi⋅Bj=Bj2−Bj⋅(Bj−Bi)≤0.(8)0 \le B_i \cdot B_j = B_j^2-B_j\cdot(B_j-B_i) \le0. \tag*{(8)}

Thus the nonzero isotropic divisors are proportional and define the same fibration f:Y→P1f:Y\to\mathbb{P}^1, since q(Y)=0q(Y)=0. For its general connected fiber FF, we have

Bj∼qjF,h0(Y,Bj)=qj+1,qj≥j.B_j\sim q_jF,\qquad h^0(Y,B_j)=q_j+1,\qquad q_j\ge j.

Put ℓ=A⋅F\ell=A\cdot F. Since AA is nef and big and is rr times an integral divisor, ℓ≥r\ell\ge r. The effectivity of A−qPFA-q_PF gives

P≤qP≤2aℓ≤2dr.(9)P\le q_P\le\frac{2a}{\ell}\le2dr. \tag*{(9)}

In particular, the assumptions imply

P+s≤2ar+s≤a8+a16<a4.(10)P+s\le\frac{2a}{r}+s\le\frac{a}{8}+\frac{a}{16}<\frac{a}{4}. \tag*{(10)}

Set θ=h(1/2)\theta=h(1/2). Choose an admissible convex combination QQ at weight θ\theta with Q2≥a/2Q^2\ge a/2. If necessary, replace QQ by λQ+(1−λ)B0\lambda Q+(1-\lambda)B_0 for a suitable 0<λ≤10<\lambda\le1. This preserves admissibility and gives Q2=a/2Q^2=a/2. Since QQ is a convex combination of the BjB_j, the divisor A−QA-Q is effective. For a nonzero nef divisor GG with G2=0G^2=0, Hodge index on the orthogonal complement of the plane spanned by A,GA,G gives, for every real divisor class B′B',

(B′)2≤2(A⋅B′)(B′⋅G)A⋅G−2a(B′⋅G)2(A⋅G)2.(11)(B')^2\le\frac{2(A\cdot B')(B'\cdot G)}{A\cdot G}-\frac{2a(B'\cdot G)^2}{(A\cdot G)^2}. \tag*{(11)}

Applying this to Q,FQ,F, and using A⋅Q≤2aA\cdot Q\le2a, yields

a/2=Q2≤4a(F⋅Q)ℓ,F⋅Q≥ℓ/8.a/2=Q^2\le\frac{4a(F\cdot Q)}{\ell},\qquad F\cdot Q\ge\ell/8.

For 1≤j≤P1\le j\le P, put

tj=16(j+s)qjℓ≤16(s+1)r≤12.t_j=\frac{16(j+s)}{q_j\ell}\le\frac{16(s+1)}{r}\le\frac{1}{2}.

The convex combination (1−tj)Bj+tjQ(1-t_j)B_j+t_jQ is admissible at weight (1−tj)uj+tjθ(1-t_j)u_j+t_j\theta, and

((1−tj)Bj+tjQ)2≥2tj(1−tj)qj(F⋅Q)≥tjqjℓ8=2(j+s).\bigl((1-t_j)B_j+t_jQ\bigr)^2\ge2t_j(1-t_j)q_j(F\cdot Q)\ge\frac{t_jq_j\ell}{8}=2(j+s).

Consequently,

(h(j+sa)−uj)+≤tj(θ−uj)+≤tjθ.\left(h\left(\sqrt{\frac{j+s}{a}}\right)-u_j\right)_+\le t_j(\theta-u_j)_+\le t_j\theta.

Finally, qj≥jq_j\ge j, (3.3), and the harmonic sum bound give

∑j=1Ptj≤16ℓ(P+s(1+log⁡P))\sum_{j=1}^{P}t_j\le\frac{16}{\ell}\bigl(P+s(1+\log P)\bigr)
≤16[2d+s(1+log⁡(2dr))r]≤s.\le16\left[2d+\frac{s(1+\log(2dr))}{r}\right]\le s.

Multiplying by θ\theta proves the assertion. □

Linear Systems with Two-Dimensional Image

When B2>0B^2 > 0, Clifford’s inequality controls dim⁡V\dim V in terms of B2B^2. Combined with (3), this gives a pointwise comparison between hh and the weights. For a globally generated divisor BB on YY and a subspace V⊆H0(Y,B)V \subseteq H^0(Y,B) generating BB, put

κ(B,V)=B2−2(dim⁡V−2),κ(B)=B2−2(h0(Y,B)−2).\kappa(B,V) = B^2 - 2(\dim V - 2), \qquad\kappa(B) = B^2 - 2(h^0(Y,B) - 2).

Proposition 3.2. Suppose B2>0B^2 > 0. Then

h0(Y,B)≤B22+2,κ(B,V)≥κ(B)≥0.(12)h^0(Y,B) \le\frac{B^2}{2} + 2, \qquad\kappa(B,V) \ge\kappa(B) \ge0. \tag*{(12)}

If VV belongs to the filtration associated with λ\lambda, BB is its moving part, and κ(B,V)≥2s+2\kappa(B,V) \ge2s + 2 for an integer s≥1s \ge1, then every index jj in the corresponding block of equal weights satisfies

uj≥h(min⁡(j+s,a)a).(13)u_j \ge h\left(\sqrt{\frac{\min(j+s,a)}{a}}\right). \tag*{(13)}

Proof. Let C∈∣B∣C \in|B| be a general member. Since B2>0B^2 > 0 and BB is globally generated, CC is smooth and irreducible. Moreover, q(Y)=0q(Y) = 0, and the restriction sequence gives

h0(C,OC(B))=h0(Y,B)−1.h^0(C,\mathcal{O}_C(B)) = h^0(Y,B) - 1.

By adjunction,

KC⊗OC(−B)=OC(KY).K_C \otimes\mathcal{O}_C(-B) = \mathcal{O}_C(K_Y).

Since KYK_Y is effective and CC is general, OC(B)\mathcal{O}_C(B) is special. Clifford’s inequality gives

h0(C,OC(B))≤B22+1.h^0(C,\mathcal{O}_C(B)) \le\frac{B^2}{2} + 1.

The remaining inequalities follow from

κ(B,V)=κ(B)+2(h0(Y,B)−dim⁡V).\kappa(B,V) = \kappa(B) + 2(h^0(Y,B) - \dim V).

For the weight comparison, put v=dim⁡Vv = \dim V. The hypothesis κ(B,V)≥2s+2\kappa(B,V) \ge2s + 2 gives

B2≥2(v+s−1).B^2 \ge2(v+s-1).

Since VV is the last subspace in its block of equal weights, uj=uv−1u_j = u_{v-1} for every index jj in that block, and j≤v−1j \le v - 1. Applying (3) and using the monotonicity of hh, we obtain

uj=uv−1≥h(B22a)≥h(min⁡(j+s,a)a).u_j = u_{v-1} \ge h\left(\sqrt{\frac{B^2}{2a}}\right) \ge h\left(\sqrt{\frac{\min(j+s,a)}{a}}\right).

Bounds for Base-Point Multiplicities

The shifted comparison above applies when κ(B,V)\kappa(B,V) is large. To count the remaining subspaces, we first bound the base-point multiplicities when κ(B)\kappa(B) is small.

Lemma 3.3. Let η≥0\eta\ge0 be an integer, and let BB be a globally generated divisor on YY such that A−BA-B is effective. If

κ(B)≤η,B2>max⁡{8dr,4η+8},\kappa(B) \le\eta,\qquad B^2 > \max\{8dr,4\eta+8\},

then KY⋅B≤3ηK_Y \cdot B \le3\eta.

Proof. Put b=B2b=B^2, m=h0(Y,B)−2m=h^0(Y,B)-2, κ=κ(B)\kappa=\kappa(B), and δ=KY⋅B\delta=K_Y\cdot B. The assumptions imply m≥η+3m\ge\eta+3. The morphism defined by ∣B∣|B| has nondegenerate image T⊂Pm+1T\subset\mathbb{P}^{m+1}. If its degree is ee, then

em≤edeg⁡T=b=2m+κ<3m,em\le e\deg T=b=2m+\kappa<3m,

so e=1e=1 or 22.

If e=1e=1, the complete linear series on a general member C∈∣B∣C\in|B| has degree 2m+κ2m+\kappa and dimension mm, since q(Y)=0q(Y)=0. Castelnuovo’s bound [5 §1] gives

m+1+κ+δ2=g(C)≤m+2κ+1.(14)m+1+\frac{\kappa+\delta}{2}=g(C)\le m+2\kappa+1. \tag*{(14)}

Hence δ≤3κ≤3η\delta\le3\kappa\le3\eta.

Suppose e=2e=2. Let QQ be the minimal resolution of the normalization of TT, and let DD be the pullback of its hyperplane divisor. Resolving the rational map gives morphisms

p:Y′⟶Y,f:Y′⟶Q,p∗B=f∗D,p:Y'\longrightarrow Y,\qquad f:Y'\longrightarrow Q,\qquad p^*B=f^*D,

with pp birational and deg⁡f=2\deg f=2. Pullback of sections and of one-forms gives

h0(Q,D)=m+2,D2=m+κ/2,q(Q)=0.h^0(Q,D)=m+2,\qquad D^2=m+\kappa/2,\qquad q(Q)=0.

Thus ∣D∣|D| is complete and birational onto TT. Castelnuovo’s bound gives g≤κ/2g\le\kappa/2 for the genus of a general member of ∣D∣|D|. In particular, KQ⋅D=2g−2−D2<0K_Q\cdot D=2g-2-D^2<0, so QQ is rational by Castelnuovo’s rationality criterion.

Apply [1 Propositions 4.1 and 5.1]. If g>0g>0, then D2≤4g+4≤2η+4D^2\le4g+4\le2\eta+4, unless TT is the Veronese surface of degree 9. Both alternatives contradict b>max⁡{8dr,4η+8}b>\max\{8dr,4\eta+8\}, since 8dr≥248dr\ge24. If g=0g=0, then TT is a scroll, since D2>4D^2>4. Its ruling on QQ has general fiber FF with F2=0F^2=0 and D⋅F=1D\cdot F=1.

Set G=f∗FG=f^*F. Then GG is nonzero and nef, with G2=0G^2=0 and p∗B⋅G=2p^*B\cdot G=2. Continue to denote the pullback of AA by AA. Since AA is nef and big and divisible by rr, A⋅G≥rA\cdot G\ge r. Moreover, A⋅p∗B≤A2=2aA\cdot p^*B\le A^2=2a. Applying (11), we obtain

b≤2(A⋅p∗B)(p∗B⋅G)A⋅G≤8ar=8dr,b\le\frac{2(A\cdot p^*B)(p^*B\cdot G)}{A\cdot G}\le\frac{8a}{r}=8dr,

a contradiction. ∎

Counting Subspaces in a Flag

We now show that only O(r)O(r) subspaces in a strict flag can have positive square and bounded κ(B,V)\kappa(B,V).

Fix an integral ample divisor HH on X~\widetilde{X}, and put c=H⋅π∗Lc=H\cdot\pi^{*}L. For a strict flag

0≠V0⊊⋯⊊Vn⊆H0(X,rL),0\ne V_{0}\subsetneq\cdots\subsetneq V_{n}\subseteq H^{0}(X,rL),

let BiB_{i} be the moving parts on a common resolution σ:Y→X~\sigma:Y\to\widetilde{X} of the base ideals.

Proposition 3.4. For every integer η≥0\eta\ge0, there is a constant C>0C>0, depending only on dd, cc, η\eta, such that

#{i:Bi2>0, κ(Bi,Vi)≤η}≤Cr.\#\{i:B_{i}^{2}>0,\ \kappa(B_{i},V_{i})\le\eta\}\le Cr.

The bound is independent of the flag and its resolution.

Proof. Put

R=max⁡{16dr,4η+8}.R=\max\{16dr,4\eta+8\}.

For 0<Bi2≤R0<B_{i}^{2}\le R, Clifford’s inequality gives

dim⁡Vi≤h0(Y,Bi)≤R/2+2.\dim V_{i}\le h^{0}(Y,B_{i})\le R/2+2.

Since the dimensions in a strict flag are distinct, there are at most ⌊R/2⌋+2\lfloor R/2\rfloor+2 such indices.

Let FiF_{i} be the fixed divisor of ViV_{i} on X~\widetilde{X}, and set Mi=rπ∗L−FiM_{i}=r\pi^{*}L-F_{i}. The inclusions Vi⊆VjV_{i}\subseteq V_{j} imply Fi≥FjF_{i}\ge F_{j}, hence Mi≤MjM_{i}\le M_{j}, for i<ji<j. Moreover,

0≤H⋅Mi≤rc.0\le H\cdot M_{i}\le rc.

Every strict increase of MiM_{i} increases its integral HH-degree, so at most rc+1rc+1 distinct divisors MiM_{i} occur.

Now consider an index with Bi2>RB_{i}^{2}>R and κ(Bi,Vi)≤η\kappa(B_{i},V_{i})\le\eta. Writing EαE_{\alpha} for the total transforms of the exceptional curves, we have

Bi=σ∗Mi−∑αmi,αEα,mi,α≥0.B_{i}=\sigma^{*}M_{i}-\sum_{\alpha}m_{i,\alpha}E_{\alpha},\qquad m_{i,\alpha}\ge0.

The blowup formulas and Lemma 3.3 give

∑αmi,α=KY⋅Bi≤3η,Mi2−Bi2=∑αmi,α2≤9η2.\sum_{\alpha}m_{i,\alpha}=K_{Y}\cdot B_{i}\le3\eta,\qquad M_{i}^{2}-B_{i}^{2}=\sum_{\alpha}m_{i,\alpha}^{2}\le9\eta^{2}.

Since

dim⁡Vi=Bi2−κ(Bi,Vi)2+2,0≤κ(Bi,Vi)≤η,\dim V_{i}=\frac{B_{i}^{2}-\kappa(B_{i},V_{i})}{2}+2,\qquad0\le\kappa(B_{i},V_{i})\le\eta,

all such indices with a fixed Mi=MM_{i}=M satisfy

M2−9η2−η2+2≤dim⁡Vi≤M22+2.\frac{M^{2}-9\eta^{2}-\eta}{2}+2\le\dim V_{i}\le\frac{M^{2}}{2}+2.

There are therefore at most (9η2+η)/2+1(9\eta^{2}+\eta)/2+1 such indices for each MM. Together with the preceding counts, this gives

#{i:Bi2>0, κ(Bi,Vi)≤η}≤⌊R2⌋+2+(9η2+η2+1)(rc+1).(15)\#\{i:B_{i}^{2}>0,\ \kappa(B_{i},V_{i})\le\eta\}\le\left\lfloor\frac{R}{2}\right\rfloor+2+\left(\frac{9\eta^{2}+\eta}{2}+1\right)(rc+1). \tag*{(15)}

For fixed dd, cc, η\eta, the right-hand side is bounded by CrCr. All estimates are independent of the chosen resolution.

Comparison with the Weights

Put s=16(2d+1)s=16(2d+1), and write

xj=j/a(0≤j≤a),yj=xmin⁡(j+s,a)(1≤j≤a).x_j=\sqrt{j/a}\qquad(0\le j\le a),\qquad y_j=x_{\min(j+s,a)}\qquad(1\le j\le a).

As in Subsection 3.1, PP is the largest index in {1,…,a}\{1,\ldots,a\} for which BP2=0B_P^2=0, or 0 if there is no such index.

Proposition 3.5. Assume (6). For the indices with Bj2=0B_j^2=0, the total error satisfies

∑j=1P(h(yj)−uj)+≤sh(1/2).\sum_{j=1}^{P}\left(h(y_j)-u_j\right)_{+}\le s h(1/2).

For the remaining indices, there are sets E1⊆Es⊆{P+1,…,a}E_1\subseteq E_s\subseteq\{P+1,\ldots,a\} such that

h(xj−1)≤uj(P<j≤a),h(xj)≤uj(P<j≤a, j∉E1),h(yj)≤uj(P<j≤a, j∉Es).(16)\begin{aligned} h(x_{j-1})&\le u_j &&(P<j\le a),\\ h(x_j)&\le u_j &&(P<j\le a,\ j\notin E_1),\\ h(y_j)&\le u_j &&(P<j\le a,\ j\notin E_s). \tag*{(16)} \end{aligned}

The exceptional sets satisfy

∣E1∣≤k,(17)|E_1|\le k, \tag*{(17)}
∣Es∣≤(s+1)k,(18)|E_s|\le(s+1)k, \tag*{(18)}

where kk is the number of subspaces VV in the weight filtration for which B2>0B^2>0 and κ(B,V)≤2s+1\kappa(B,V)\le2s+1. Moreover, k≤Crk\le Cr, with CC depending only on d,cd,c.

Proof. The bound on kk follows from Proposition 3.4. Lemma 3.1 gives the asserted estimate for j≤Pj\le P, since (10) implies P+s<aP+s<a.

Consider a maximal block II of equal weights, and let v−1v-1 be its last index. For j∈Ij\in I with j>Pj>P, the divisor Bv−1B_{v-1} has positive square. Put

b=Bv−12,κ=κ(Bv−1,Vv−1)=b−2(v−2).b=B_{v-1}^{2},\qquad\kappa=\kappa(B_{v-1},V_{v-1})=b-2(v-2).

Clifford’s inequality and (3) give

uj=uv−1≥h(b2a)≥h(xj−1),u_j=u_{v-1}\ge h\left(\sqrt{\frac{b}{2a}}\right)\ge h(x_{j-1}),

because b≥2(v−2)b\ge2(v-2) and j≤v−1j\le v-1.

If κ≥2s+2\kappa\ge2s+2, Proposition 3.2 gives uj≥h(yj)u_j\ge h(y_j) throughout the block. For the remaining blocks, the same bound b≥2(v−2)b\ge2(v-2) gives

uj≥h(xj)when j≤v−2,u_j\ge h(x_j)\qquad\text{when }j\le v-2,
uj≥h(yj)when j+s≤v−2.u_j\ge h(y_j)\qquad\text{when }j+s\le v-2.

Thus the first comparison can fail only at the last index of the block, and the second only at its last s+1s+1 indices.

Define E1E_1 and EsE_s to consist of these respective indices in {P+1,…,a}\{P+1,\ldots,a\}, over the blocks with b>0b>0 and κ≤2s+1\kappa\le2s+1. There are at most kk such blocks, so (17) and (18) follow. The preceding comparisons prove (16). □

Proof of the Main Theorem

We compare the integral in (5) with finite sums of values of hh, then use Proposition 3.53.5 to bound those sums by the weights. This proves the required inequality once the exceptional sets are sufficiently small.

Riemann Sum Estimates

We first estimate the difference between the discrete sum and the integral for the functions (t−c)+(t-c)_+. The piecewise linear interpolation reduces the estimate for hh to nonnegative linear combinations of these functions.

Lemma 4.1. Let a≥8a\ge8 be an integer, and put xj=j/ax_j=\sqrt{j/a}. For 0≤c<10\le c<1, set ψc(t)=(t−c)+=max⁡{t−c,0}\psi_c(t)=(t-c)_+=\max\{t-c,0\} and

Q0[ψc]=ψc(1)+∑j=1aψc(xj)−(a+2)∫01ψc(x) dx.Q_0[\psi_c]=\psi_c(1)+\sum_{j=1}^{a}\psi_c(x_j)-(a+2)\int_0^1\psi_c(\sqrt{x})\,dx.

Then

Q0[ψc]>112(0≤c≤4/5).(19)Q_0[\psi_c]>\frac{1}{12}\qquad(0\le c\le4/5). \tag*{(19)}
Q0[ψc]>1−c(4/5≤c<1).(20)Q_0[\psi_c]>1-c\qquad(4/5\le c<1). \tag*{(20)}

Proof. Direct integration gives

∫01ψc(x) dx=23−c+c33.\int_0^1\psi_c(\sqrt{x})\,dx=\frac{2}{3}-c+\frac{c^3}{3}.

Strict concavity of the square root gives the midpoint estimate

∑j=1aj/a>1a∫1/2a+1/2t dt=23a((a+1/2)3/2−(1/2)3/2)>2a3+12−132a.\sum_{j=1}^{a}\sqrt{j/a}>\frac{1}{\sqrt{a}}\int_{1/2}^{a+1/2}\sqrt{t}\,dt =\frac{2}{3\sqrt{a}}\left((a+1/2)^{3/2}-(1/2)^{3/2}\right)>\frac{2a}{3}+\frac{1}{2}-\frac{1}{3\sqrt{2a}}.

For the convex function x↦(c−x)+x\mapsto(c-\sqrt{x})_+, the trapezoidal estimate gives

∑j=1a(c−xj)+−a∫01(c−x)+ dx≥−c2.\sum_{j=1}^{a}(c-x_j)_+-a\int_0^1(c-\sqrt{x})_+\,dx\ge-\frac{c}{2}.

Using ψc(x)=x−c+(c−x)+\psi_c(\sqrt{x})=\sqrt{x}-c+(c-\sqrt{x})_+ and adding these estimates, we obtain

∑j=1aψc(xj)−a∫01ψc(x) dx>12−132a−c2≥512−c2.(21)\sum_{j=1}^{a}\psi_c(x_j)-a\int_0^1\psi_c(\sqrt{x})\,dx>\frac{1}{2}-\frac{1}{3\sqrt{2a}}-\frac{c}{2}\ge\frac{5}{12}-\frac{c}{2}. \tag*{(21)}

Consequently,

Q0[ψc]>112+c2−2c33≥112(0≤c≤4/5).Q_0[\psi_c] > \frac{1}{12}+\frac{c}{2}-\frac{2c^3}{3}\ge\frac{1}{12}\qquad(0\le c\le4/5).

For 4/5≤c<14/5\le c<1, integration on the grid intervals yields

∑j=1aψc(xj)−a∫01ψc(x) dx=∫c21{ax}2x dx≥1−c24.\sum_{j=1}^{a}\psi_c(x_j)-a\int_0^1\psi_c(\sqrt{x})\,\mathrm{d}x=\int_{c^2}^1\frac{\{ax\}}{2\sqrt{x}}\,\mathrm{d}x\ge\frac{1-c^2}{4}.

Here {ax}\{ax\} is the fractional part of axax. Its average is 1/21/2 on each full grid interval and at least 1/21/2 on the initial partial interval. Hence

Q0[ψc]−(1−c)≥1−c12(8c2+11c−13)>0.Q_0[\psi_c]-(1-c)\ge\frac{1-c}{12}(8c^2+11c-13)>0.

Summation of the Weight Estimates

We now combine the weight estimates with the Riemann sum estimates of the preceding subsection. The following proposition gives the inequality required by Proposition 2.7.

Proposition 4.2. Let F\mathcal{F} be a normalized weighted flag, and let hh be its associated function. Put s=16(2d+1)s=16(2d+1), and let kk count the subspaces in its weight filtration with B2>0B^2>0 and κ(B,V)≤2s+1\kappa(B,V)\le2s+1. If (6) holds and

ka≤s264(s+1)3,(22)\frac{k}{a}\le\frac{s^2}{64(s+1)^3}, \tag*{(22)}

then

W>N∫01h(x) dx.(23)W>N\int_0^1h(\sqrt{x})\,\mathrm{d}x. \tag*{(23)}

Proof. If h=0h=0, then W≥ua+1=1W\ge u_{a+1}=1 proves the assertion. Assume that h≠0h\ne0, and recall

xj=j/a,yj=xmin⁡(j+s,a).x_j=\sqrt{j/a},\qquad y_j=x_{\min(j+s,a)}.

We first reduce to a finite nonnegative linear combination of the functions ψc\psi_c from Lemma 4.1. Let h~\widetilde{h} be the piecewise linear interpolation of hh at x0,…,xax_0,\ldots,x_a and 1/21/2. It is convex and nondecreasing, and convexity of hh gives h≤h~h\le\widetilde{h}. Moreover,

h~(xj)=h(xj),h~(yj)=h(yj),h~(1/2)=h(1/2),\widetilde{h}(x_j)=h(x_j),\qquad\widetilde{h}(y_j)=h(y_j),\qquad\widetilde{h}(1/2)=h(1/2),

because each yjy_j is one of the points x0,…,xax_0,\ldots,x_a. Thus replacing hh by h~\widetilde{h} leaves both the pointwise weight bounds and the total error bound in Proposition 3.5 unchanged. It suffices to prove

W>N∫01h~(x) dx.W>N\int_0^1\widetilde{h}(\sqrt{x})\,\mathrm{d}x.

Since h~(0)=0\widetilde{h}(0)=0 and its slopes are nonnegative and nondecreasing, we can write

h~=∑ℓ=1mβℓψcℓ,βℓ≥0,0≤cℓ<1.\widetilde{h}=\sum_{\ell=1}^{m}\beta_\ell\psi_{c_\ell},\qquad\beta_\ell\ge0,\quad0\le c_\ell<1.

Split this sum as

f=∑cℓ<1/2βℓψcℓ,g=∑cℓ≥1/2βℓψcℓ.f=\sum_{c_\ell<1/2}\beta_\ell\psi_{c_\ell},\qquad g=\sum_{c_\ell\ge1/2}\beta_\ell\psi_{c_\ell}.

Then h~=f+g\widetilde{h}=f+g, both functions are nondecreasing, and g=0g=0 on [0,1/2][0,1/2]. We will use the shifted weight comparison for ff and the unshifted comparison for gg.

Let P,E1,EsP,E_1,E_s be as in Proposition 3.5. For j>Pj>P, its pointwise bounds give

uj≥{f(yj)+g(xj),j∉Es,f(xj−1)+g(xj),j∈Es∖E1,f(xj−1)+g(xj−1),j∈E1.u_j\ge \begin{cases} f(y_j)+g(x_j), & j\notin E_s,\\ f(x_{j-1})+g(x_j), & j\in E_s\setminus E_1,\\ f(x_{j-1})+g(x_{j-1}), & j\in E_1. \end{cases}

For j≤Pj\le P, we have xj≤yj<1/2x_j\le y_j<1/2, so gg vanishes at both points and the total error bound becomes

∑j=1P(f(yj)−uj)+≤sf(1/2).\sum_{j=1}^{P}(f(y_j)-u_j)_+\le s f(1/2).

For a function vv, put

Q0[v]=v(1)+∑j=1av(xj)−N∫01v(x) dx,Q_0[v]=v(1)+\sum_{j=1}^{a}v(x_j)-N\int_0^1v(\sqrt{x})\,\mathrm{d}x,
Qs[v]=v(1)+∑j=1av(yj)−N∫01v(x) dx.Q_s[v]=v(1)+\sum_{j=1}^{a}v(y_j)-N\int_0^1v(\sqrt{x})\,\mathrm{d}x.

The first expression agrees with the notation of Lemma 4.1, since N=a+2N=a+2. Summing the preceding bounds and using ua+1=1≥h~(1)u_{a+1}=1\ge\widetilde{h}(1), we obtain

W−N∫01h~(x) dx≥Qs[f]−sf(1/2)−∑j∈Es(f(yj)−f(xj−1))+Q0[g]−∑j∈E1(g(xj)−g(xj−1)).(24)\begin{aligned} W-N\int_0^1\widetilde{h}(\sqrt{x})\,\mathrm{d}x &\ge Q_s[f]-s f(1/2)\\ &\quad-\sum_{j\in E_s}\bigl(f(y_j)-f(x_{j-1})\bigr)\\ &\quad+Q_0[g]-\sum_{j\in E_1}\bigl(g(x_j)-g(x_{j-1})\bigr). \tag*{(24)} \end{aligned}

By linearity, it remains to show that the contribution of each summand ψc\psi_c is positive.

The increments xj−xj−1x_j-x_{j-1} decrease, so any MM distinct increments sum to at most M/a\sqrt{M/a}. Since ψc\psi_c is nondecreasing and 1-Lipschitz,

∑j∈E1(ψc(xj)−ψc(xj−1))≤min⁡{1−c,k/a}.(25)\sum_{j\in E_1}\bigl(\psi_c(x_j)-\psi_c(x_{j-1})\bigr)\le\min\left\{1-c,\sqrt{k/a}\right\}. \tag*{(25)}

The bound 1−c1-c follows by telescoping over all indices. Each difference yj−xj−1y_j-x_{j-1} contains at most s+1s+1 consecutive increments. Summing separately at each offset gives

∑j∈Es(ψc(yj)−ψc(xj−1))≤(s+1)∣Es∣/a≤(s+1)3/2k/a≤s/8.(26)\sum_{j\in E_s}\bigl(\psi_c(y_j)-\psi_c(x_{j-1})\bigr) \le(s+1)\sqrt{|E_s|/a}\le(s+1)^{3/2}\sqrt{k}/a\le s/8. \tag*{(26)}

For 0≤c<1/20 \le c < 1/2, shifting the sum gives

Qs[ψc]=Q0[ψc]+s(1−c)−∑j=1sψc(xj)≥Q0[ψc]+s(1−c−s/a).(27)\begin{aligned} Q_s[\psi_c] = Q_0[\psi_c] + s(1-c) - \sum_{j=1}^{s}\psi_c(x_j) \\ &\ge Q_0[\psi_c] + s\left(1-c-\sqrt{s/a}\right). \tag*{(27)} \end{aligned}

Using ψc(1/2)=1/2−c\psi_c(1/2)=1/2-c, a≥16sa\ge16s, and (26), we obtain

Qs[ψc]−sψc(1/2)−∑j∈Es(ψc(yj)−ψc(xj−1))≥Q0[ψc]+s2−ss/a−s8≥Q0[ψc]+s8>0.\begin{aligned} Q_s[\psi_c]-s\psi_c(1/2)-\sum_{j\in E_s}\left(\psi_c(y_j)-\psi_c(x_{j-1})\right) \\ &\ge Q_0[\psi_c]+\frac{s}{2}-s\sqrt{s/a}-\frac{s}{8} \\ &\ge Q_0[\psi_c]+\frac{s}{8}>0. \end{aligned}

where the last inequality follows from Lemma 4.1.

For 1/2≤c<11/2\le c<1, the same lemma and (25) give

Q0[ψc]−∑j∈E1(ψc(xj)−ψc(xj−1))>{1/12−k/a,1/2≤c≤4/5,0,4/5<c<1.Q_0[\psi_c]-\sum_{j\in E_1}\left(\psi_c(x_j)-\psi_c(x_{j-1})\right)> \begin{cases} 1/12-\sqrt{k/a}, & 1/2\le c\le4/5,\\ 0, & 4/5<c<1. \end{cases}

The first bound is positive because (22) and s≥48s\ge48 imply k/a<1/144k/a<1/144.

Multiplying these inequalities by βℓ\beta_\ell and summing makes the right-hand side of (24) strictly positive. Indeed, at least one coefficient is positive because h~≠0\widetilde h\ne0. Therefore

W>N∫01h~(x) dx≥N∫01h(x) dx,W>N\int_0^1\widetilde h(\sqrt{x})\,dx\ge N\int_0^1h(\sqrt{x})\,dx,

as required. □

Combining this proposition with Proposition 2.7, we obtain the following numerical criterion.

Corollary 4.3. Put s=16(2d+1)s=16(2d+1), and suppose that kk bounds the number of subspaces with B2>0B^2>0 and κ(B,V)≤2s+1\kappa(B,V)\le2s+1 in every weight filtration. If r≥3r\ge3 satisfies

r≥32(s+1),s(1+log⁡(2dr))r≤1,a=dr2≥16s,r\ge32(s+1),\qquad\frac{s(1+\log(2dr))}{r}\le1,\qquad a=dr^2\ge16s,
ka≤s264(s+1)3(28)\frac{k}{a}\le\frac{s^2}{64(s+1)^3} \tag*{(28)}

then (X,L⊗r)(X,L^{\otimes r}) is Chow stable.

An Ample Divisor on the Resolution

The subspace count depends on the degree of an auxiliary ample divisor on the minimal resolution. To obtain a bound depending only on dd, we choose such a divisor as follows.

Lemma 4.4. Let (X,L)(X,L) be a complex polarized K3 surface with at worst ADE singularities and L2=2dL^2=2d. Its minimal resolution π:X~→X\pi:\widetilde{X}\to X admits an integral ample divisor HH with

H⋅π∗L=292d.H\cdot\pi^*L=292d.

Proof. Put ℓ=π∗L\ell=\pi^*L. The exceptional curves generate a direct sum of negative-definite ADE lattices of total rank at most 19. In each component, take the sum of the positive roots, viewed as an effective exceptional divisor, and let EE be the sum of these divisors. The Weyl vector identities give

ℓ⋅E=0,E⋅Ei=−2,M:=−E2=∑νnνhν(hν+1)3,\ell\cdot E=0,\qquad E\cdot E_i=-2,\qquad M:=-E^2=\sum_\nu\frac{n_\nu h_\nu(h_\nu+1)}{3},

where EiE_i runs through the exceptional curves, and nν,hνn_\nu,h_\nu are the rank and Coxeter number of each component. The identities are recalled in [6 Remark 6.6 and Proposition 6.12]. Since ∑νnν≤19\sum_\nu n_\nu\le19 and hν≤36h_\nu\le36, we have M≤444⋅19=8436M\le444\cdot19=8436.

Set H=146ℓ−EH=146\ell-E. Then

H⋅ℓ=292d,H2=2⋅1462d−M>0,H⋅Ei=2.H\cdot\ell=292d,\qquad H^2=2\cdot146^2d-M>0,\qquad H\cdot E_i=2.

For any other irreducible curve CC, put q=ℓ⋅C≥1q=\ell\cdot C\ge1. Adjunction gives C2≥−2C^2\ge-2, and the Hodge index theorem on ℓ⊥\ell^\perp yields

(E⋅C)2≤M(q22d−C2)≤8436(q22+2)≤21090q2<1462q2.(E\cdot C)^2\le M\left(\frac{q^2}{2d}-C^2\right)\le8436\left(\frac{q^2}{2}+2\right)\le21090q^2<146^2q^2.

Thus H⋅C>0H\cdot C>0, and the Nakai–Moishezon criterion shows that HH is ample. For smooth XX, the given polarization LL already has the required ampleness. □

An Effective Bound for the Exponent

We now substitute the explicit subspace count into Corollary 4.3, keeping track of the degree of the auxiliary ample divisor.

Proposition 4.5. Suppose that X~\widetilde{X} admits an integral ample divisor HH with H⋅π∗L≤2dtH\cdot\pi^*L\le2dt, where t≥1t\ge1. Put s=16(2d+1)s=16(2d+1). Then (X,L⊗r)(X,L^{\otimes r}) is Chow stable for every integer

r≥3072ts3=3⋅222t(2d+1)3.(29)r\ge3072ts^3=3\cdot2^{22}t(2d+1)^3. \tag*{(29)}

Proof. We verify the conditions of Corollary 4.3. Since s≥48s\ge48 and r≥3072ts3r\ge3072ts^3, we have r≥32(s+1)r\ge32(s+1) and a=dr2≥16sa=dr^2\ge16s. The function (1+log⁡(2dr))/r(1+\log(2dr))/r is decreasing for 2dr>12dr>1. Using r≥s3r\ge s^3 and 2d≤s2d\le s, we obtain

s(1+log⁡(2dr))r≤1+4log⁡ss2≤1.\frac{s(1+\log(2dr))}{r}\le\frac{1+4\log s}{s^2}\le1.

It remains to bound the number of exceptional subspaces. Apply (15) with η=2s+1\eta=2s+1. Since r≥s3r\ge s^3, the threshold is R=16drR=16dr, and

9η2+η2+1=18s2+19s+6≤19s2.\frac{9\eta^2+\eta}{2}+1=18s^2+19s+6\le19s^2.

Together with c=H⋅π∗L≤2dtc = H \cdot\pi^{*}L \le2dt, this gives

k≤8dr+2+19s2(2dtr+1)≤40dtrs2.k \le8dr + 2 + 19s^{2}(2dtr + 1) \le40dtrs^{2}.

Consequently,

64(s+1)3s2ka≤2560t(s+1)3r≤56(1+1s)3<1.\frac{64(s+1)^{3}}{s^{2}}\frac{k}{a} \le\frac{2560t(s+1)^{3}}{r} \le\frac{5}{6}\left(1+\frac{1}{s}\right)^{3} < 1.

This verifies the last condition and proves the assertion.

Proof of Theorem 1.1

For an ADE K3 surface, Lemma 4.4 supplies HH with H⋅π∗L=2d⋅146H \cdot\pi^{*}L = 2d \cdot146. Proposition 4.5, applied with t=146t = 146, therefore gives Chow stability whenever

r≥231(2d+1)3.r \ge2^{31}(2d+1)^{3}.

since 3⋅146⋅222<2313 \cdot146 \cdot2^{22} < 2^{31}. If XX is smooth, applying the same proposition with H=LH = L and t=1t = 1 gives Chow stability whenever

r≥224(2d+1)3.r \ge2^{24}(2d+1)^{3}.

This proves Theorem 1.1.

References

Shanghai Center for Mathematical Sciences, Fudan University, 2005 Songhu Road 200438, Shanghai, China

Email address: [email protected]

References

  1. [1][CDML26] Ciro Ciliberto, Thomas Dedieu, and Margarida Mendes Lopes. On surfaces of high degree with respect to the sectional genus. Accepted for publication in Kyoto Journal of Mathematics. arXiv:2503.22299v2, 2026.
  2. [2][Don01] Simon K. Donaldson. Scalar curvature and projective embeddings, I. Journal of Differential Geometry, 59(3):479–522, 2001.DOI
  3. [3][Gie77] David Gieseker. Global moduli for surfaces of general type. Inventiones Mathematicae, 43(3):233–282, 1977.DOI
  4. [4][Gie82] David Gieseker. Lectures on Moduli of Curves, volume 69 of Tata Institute of Fundamental Research Lectures on Mathematics and Physics. Springer-Verlag, Berlin, 1982. Notes by D. R. Gokhale. Published for the Tata Institute of Fundamental Research, Bombay.
  5. [5][Har81] Joe Harris. A bound on the geometric genus of projective varieties. Annali della Scuola Normale Superiore di Pisa, Classe di Scienze, 8(1):35–68, 1981.
  6. [6][Hec18] Gert Heckman. Coxeter groups. Lecture notes, Radboud University Nijmegen, 2018.
  7. [7][Huy16] Daniel Huybrechts. Lectures on K3 Surfaces. Number 158 in Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 2016.DOI
  8. [8][Kem78] George R. Kempf. Instability in invariant theory. Annals of Mathematics, 108(2):299–316, 1978.
  9. [9][LLSW19] King-Leung Lee, Zhiyuan Li, Jacob Sturm, and Xiaowei Wang. Asymptotic Chow stability of toric Del Pezzo surfaces. Mathematical Research Letters, 26(6):1759–1787, 2019.arxiv.org/abs/1711.10099
  10. [10][LT21] Zhiyuan Li and Zhiyu Tian. Moduli space of quasi-polarized K3 surfaces of degree 6 and 8. Chinese Annals of Mathematics, Series B, 42(3):427–450, 2021.
  11. [11][Mor88] David R. Morrison. The geometry of K3 surfaces. Lectures delivered at the Scuola Matematica Interuniversitaria, Cortona, Italy, July 31–August 27, 1988. Preliminary version, 2009, 1988.
  12. [12][Mor99] Ian Morrison. Stability of Hilbert points of generic K3 surfaces. Centre de Recerca Matemàtica, Publication no. 401, 1999.
  13. [13][Mor09] Ian Morrison. GIT constructions of moduli spaces of stable curves and maps. Surveys in Differential Geometry, 14(1):315–369, 2009.arxiv.org/abs/0810.2340
  14. [14]David Mumford. Stability of projective varieties. L’Enseignement Mathématique, 23(1-2):39–110, 1977.
  15. [15]Jayant Shah. A complete moduli space for K3 surfaces of degree 2. Annals of Mathematics, 112(3):485–510, 1980.DOI
  16. [16]Jayant Shah. Degenerations of K3 surfaces of degree 4. Transactions of the American Mathematical Society, 263(2):271–308, 1981.DOI
  17. [17]Eckart Viehweg. Weak positivity and the stability of certain Hilbert points. III. Inventiones Mathematicae, 101(3):521–543, 1990.DOI

Paper details

vlnc:W_j5pqs9cy
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

LLM

View statement
Looking for related papers…
Contents