A sharp lower bound for the log canonical threshold

A sharp lower bound for the log canonical threshold

Jean-Pierre Demailly and Pha.m Ho`ang Hiˆe.p Institut Fourier, Universit´e de Grenoble I,

Hanoi National University of Education

Abstract. In this note, we prove a sharp lower bound for the log canonical threshold of a plurisubharmonic function ϕ with an isolated singularity at 0 in an open subset of Cⁿ. This threshold is defined as the supremum of constants c > 0 such that e^−2cϕ is integrable on a neighborhood of 0. We relate c(ϕ) to the intermediate multiplicity numbers e_j(ϕ), defined as the Lelong numbers of (dd^cϕ)^j at 0 (so that in particular e₀(ϕ) = 1). Our main result is that c(ϕ) ≥ P

ej(ϕ)/ej+1(ϕ), 0 ≤ j ≤ n− 1. This inequality is shown to be sharp; it simultaneously improves the classical result c(ϕ) ≥ 1/e1(ϕ) due to Skoda, as well as the lower estimate c(ϕ)≥n/en(ϕ)^1/n which has received crucial applications to birational geometry in recent years. The proof consists in a reduction to the toric case, i.e. singularities arising from monomial ideals.

2000 Mathematics Subject Classification: 14B05, 32S05, 32S10, 32U25

Keywords and Phrases: Lelong number, Monge-Amp`ere operator, log canonical threshold.

1. Notation and main results

Here we put d^c = _2πⁱ (∂−∂), so that dd^c = _πⁱ∂∂. The normalization of thed^c operator is cho- sen so that we have precisely (dd^clog|z|)ⁿ=δ0 for the Monge-Amp`ere operator in C<sup>n</sup>. The Monge-Amp`ere operator is defined on locally bounded plurisubharmonic functions according to the definition of Bedford-Taylor [BT76, BT82]; it can also be extended to plurisubhar- monic functions with isolated or compactly supported poles by [Dem93]. If Ω is an open subset of Cⁿ, we let PSH(Ω) (resp. PSH⁻(Ω)) be the set of plurisubharmonic (resp. psh≤0) functions on Ω.

Definition 1.1. LetΩbe a bounded hyperconvex domain(i.e. a domain possessing a negative psh exhaustion). Following Cegrell[Ce04], we introduce certain classes of psh functions onΩ, in relation with the definition of the Monge-Amp`ere operator :

(a) E0(Ω) ={ϕ∈PSH⁻(Ω) : lim

z→∂Ωϕ(z) = 0, Z

Ω

(dd^cϕ)ⁿ<+∞},

(b) F(Ω) ={ϕ ∈PSH⁻(Ω) : ∃ E₀(Ω)∋ϕ_p ցϕ, sup

p≥1

Z

Ω

(dd^cϕ_p)ⁿ<+∞},

(c) E(Ω) ={ϕ ∈PSH⁻(Ω) : ∃ ϕK ∈ F(Ω) such that ϕK =ϕ on K, ∀K ⊂⊂Ω}.

It is proved in [Ce04] that the class E(Ω) is the biggest subset of PSH⁻(Ω) on which the Monge-Amp`ere operator is well-defined. For a general complex manifold X, after removing

1

the negativity assumption of the functions involved, one can in fact extend the Monge- Amp`ere operator to the class

(1.2) Ee(X)⊂PSH(X)

of psh functions which, on a neighborhood Ω ∋x0 of an arbitrary pointx0 ∈X, are equal to a sum u+v with u ∈ E(Ω) andv ∈ C^∞(Ω); again, this is the biggest subclass of functions of PSH(X) on which the Monge-Amp`ere operator is locally well defined. It is easy to see that Ee(X) contains the class of psh functions which are locally bounded outside isolated singularities.

For ϕ∈PSH(Ω) and 0∈Ω, we introduce the log canonical threshold at 0

(1.3) c(ϕ) = sup

c >0 : e^−2cϕ is L¹ on a neighborhood of 0 , and for ϕ ∈Ee(Ω) we introduce the intersection numbers

(1.4) ej(ϕ) =

Z

{0}

(dd^cϕ)^j ∧(dd^clogkzk)^n−j

which can be seen also as the Lelong numbers of (dd^cϕ)^j at 0. Our main result is the following sharp estimate. It is a generalization and a sharpening of similar inequalities discussed in [Cor95], [Cor00], [dFEM03], [dFEM04]; such inequalities have fundamental applications to birational geometry (see [IM72], [Puk87], [Puk02], [Isk01], [Che05]).

Theorem 1.5. Let ϕ ∈Ee(Ω) and 0∈Ω. Then c(ϕ) = +∞ if e1(ϕ) = 0, and otherwise c(ϕ)≥

Xn−1

j=0

ej(ϕ) ej+1(ϕ).

Remark 1.6. By Lemma 2.1 below, we have (e1(ϕ), . . . , en(ϕ))∈D where

D=

t = (t1, . . . , tn)∈[0,+∞)ⁿ: t²₁ ≤t2, t²_j ≤tj−1tj+1, ∀j = 2, . . . , n−1 ,

i.e. logej(ϕ) is a convex sequence. In particular, we have ej(ϕ) ≥e1(ϕ)^j, and the denomi- nators do not vanish in 1.5 if e1(ϕ)>0. On the other hand, a well known inequality due to Skoda [Sko72] tells us that

1

e1(ϕ) ≤c(ϕ)≤ n e1(ϕ),

hence c(ϕ)<+∞ iff e₁(ϕ)>0. To see that Theorem 1.5 is optimal, let us choose ϕ(z) = max a1ln|z1|, . . . , anln|zn|

with 0 < a₁ ≤a₂ ≤ . . .≤ a_n. Then e_j(ϕ) = a₁a₂. . . a_j, and a change of variable z_j = ζ_j^1/aj on C r R₋ easily shows that

c(ϕ) = Xn

j=1

1 aj

.

Assume that we have a function f :D → [0,+∞) such that c(ϕ)≥ f(e1(ϕ), . . . , en(ϕ)) for all ϕ∈Ee(Ω). Then, by the above example, we must have

f(a1, a1a2, . . . , a1. . . an)≤ Xn

j=1

1 aj

for all aj as above. By taking aj =tj/tj−1, t0 = 1, this implies that f(t1, . . . , tn)≤ 1

t1

+ t1

t2

+. . .+ tn−1

tn

, ∀t ∈D,

whence the optimality of our inequality.

Remark 1.7. Theorem 1.5 is of course stronger than Skoda’s lower boundc(ϕ)≥1/e1(ϕ).

By the inequality between the arithmetic and geometric means, we infer the main inequality of [dFEM03], [dFEM04] and [Dem09]

(1.8) c(ϕ)≥ n

en(ϕ)^1/n.

By applying the arithmetic-geometric inequality for the indices 1 ≤ j ≤ n − 1 in our summation Pn−1

j=0 e_j(ϕ)/e_j+1(ϕ), we also infer the stronger inequality

(1.9) c(ϕ)≥ 1

e₁(ϕ) + (n−1)

e1(ϕ) e_n(ϕ)

_n−1¹ .

2. Log convexity of the multiplicity sequence

The log convexity of the multiplicity sequence can be derived from very elementary integra- tion by parts and the Cauchy-Schwarz inequality, using an argument from [Ce04].

Lemma 2.1. Letϕ∈Ee(Ω)and0∈Ω. We haveej(ϕ)² ≤ej−1(ϕ)ej+1(ϕ), ∀j = 1, . . . , n−1.

Proof. Without loss generality, by replacing ϕ with a sequence of local approximations ϕp(z) = max(ϕ(z)−C, plog|z|) of ϕ(z)−C, C≫ 1, we can assume that Ω is the unit ball and ϕ ∈ E0(Ω). Take also h, ψ ∈ E0(Ω). Then integration by parts and the Cauchy-Schwarz inequality yield

Z

Ω

−h(dd^cϕ)^j ∧(dd^cψ)^n−j 2

= Z

Ω

dϕ∧d^cψ∧(dd^cϕ)^j−1∧(dd^cψ)^n−j−1∧dd^ch 2

≤ Z

Ω

dψ∧d^cψ∧(dd^cϕ)^j−1∧(dd^cψ)^n−j−1∧dd^ch Z

Ω

dϕ∧d^cϕ∧(dd^cϕ)^j−1∧(dd^cψ)^n−j−1∧dd^ch

= Z

Ω

−h(dd^cϕ)^j−1∧(dd^cψ)^n−j+1 Z

Ω

−h(dd^cϕ)^j+1∧(dd^cψ)^n−j−1. Now, as p→+∞, take

h(z) =hp(z) = max

−1,1

plogkzk ր

0 if z ∈Ωr{0}

−1 if z = 0.

By the monotone convergence theorem we get in the limit Z

{0}

(dd^cϕ)^j ∧(dd^cψ)^n−j 2

≤ Z

{0}

(dd^cϕ)^j−1∧(dd^cψ)^n−j+1 Z

{0}

(dd^cϕ)^j+1∧(dd^cψ)^n−j−1.

For ψ(z) = lnkzk, this is the desired estimate.

Corollary 2.2. Let ϕ∈Ee(Ω) and 0∈Ω. We have the inequalities ej(ϕ) ≥ e1(ϕ)^j, ∀j = 0,1, . . .≤n

e_k(ϕ) ≤ e_j(ϕ)^ll−−kje_l(ϕ)^kl−−jj, ∀0≤j < k < l ≤n.

In particular e1(ϕ) = 0 implies ek(ϕ) = 0 for k = 2, . . . , n−1 if n ≥3.

Proof. Ifej(ϕ)>0 for allj, Lemma 2.1 implies thatj 7→ej(ϕ)/ej−1(ϕ) is increasing, at least equal toe1(ϕ)/e0(ϕ) =e1(ϕ), and the inequalities follow from the log convexity. The general case can be proved by considering ϕε(z) = ϕ(z) +εlogkzk, since 0 < ε^j ≤ ej(ϕε)→ ej(ϕ) when ε→0. The last statement is obtained by taking j = 1 and l=n.

3. Proof of the main theorem We start with a monotonicity statement.

Lemma 3.1. Let ϕ, ψ∈Ee(Ω) be such that ϕ≤ψ (i.e. ϕ is “more singular” than ψ). Then Xn−1

j=0

ej(ϕ) ej+1(ϕ) ≤

Xn−1

j=0

ej(ψ) ej+1(ψ). Proof. As in Remark 1.6, we set

D={t = (t1, . . . , tn)∈[0,+∞)ⁿ: t²₁ ≤t2, t²_j ≤tj−1tj+1, ∀j = 2, . . . , n−1}.

Then D is a convex set in Rⁿ, as can be checked by a straightforward application of the Cauchy-Schwarz inequality. We consider the function f : intD→[0,+∞)

(3.2) f(t1, . . . , tn) = 1 t1

+t1

t2

. . .+tn−1

tn

.

We have

∂f

∂tj

(t) =−tj−1

t²_j + 1 tj+1

≤0, ∀t∈D.

Fora, b∈intDsuch thataj ≥bj,∀j = 1, . . . , n, [0,1]∋λ→f(b+λ(a−b)) is thus a decreas- ing function. This implies that f(a)≤ f(b) for a, b∈ intD, aj ≥bj, ∀j = 1, . . . , n. On the other hand, the hypothesis ϕ ≤ ψ implies ej(ϕ) ≥ej(ψ), ∀j = 1, . . . , n, by the comparison principle (see e.g. [Dem87]). Therefore f(e1(ϕ), . . . , en(ϕ))≤f(e1(ψ), . . . , en(ψ)).

(3.3) Proof of the main theorem in the “toric case”.

It will be convenient here to introduce Kiselman’s refined Lelong numbers (cf. [Kis87], [Kis94a]):

Definition 3.4. Let ϕ ∈PSH(Ω). Then the function ν_ϕ(x) = lim

t→−∞

max{ϕ(z) :|z1|=e^x1t, . . . ,|zn|=e^xnt} t

is called the refined Lelong number of ϕ at 0. This function is increasing in each variablexj

and concave on Rⁿ

+.

By “toric case”, we mean that ϕ(z1, . . . , zn) =ϕ(|z1|, . . . ,|zn|) depends only on |zj|for all j;

then ϕ is psh if and only if (t1, . . . , tn)7→ϕ(e^t1, . . . , e^tn) is increasing in each tj and convex.

By replacing ϕ with ϕ(λz)−ϕ(λ, ..., λ), 0 < λ ≪ 1, we can assume that Ω = ∆ⁿ is the unit polydisk, ϕ(1, . . . ,1) = 0 (so that ϕ ≤ 0 on Ω), and we have e1(ϕ) = n νϕ(_n¹, . . . ,_n¹).

By convexity, the slope ^{max{ϕ(z) :}_t^{|zj|=exj t}} is increasing in t for t < 0. Therefore, by taking t =−1 we get

νϕ(−ln|z1|, . . . ,−ln|zn|)≤ −ϕ(z1, . . . , zn).

Notice also that νϕ(x) satisfies the 1-homogeneity property νϕ(λx) = λνϕ(x) for λ ∈ R₊. As a consequence, νϕ is entirely characterized by its restriction to the set

Σ =n

x= (x₁, . . . , x_n)∈Rⁿ

+: Xn

j=1

x_j = 1o .

We choose x⁰ = (x⁰₁, . . . , x⁰_n)∈Σ such that

νϕ(x⁰) = max{νϕ(x) : x∈Σ} ∈he1(ϕ)

n , e1(ϕ)i .

By Theorem 5.8 in [Kis94a] (see also [Ho01] for similar results in an algebraic context) we have the formula

c(ϕ) = 1 νϕ(x⁰). Set

ζ(x) =νϕ(x⁰) minx1

x⁰₁, . . . ,xn

x⁰_n

, ∀x∈Rⁿ

+. Thenζ is the smallest nonnegative concave 1-homogeneous function onRⁿ

+that is increasing in each variable xj and such thatζ(x⁰) = νϕ(x⁰). Therefore we have ζ ≤νϕ, hence

ϕ(z1, . . . , zn) ≤ −νϕ(−ln|z1|, . . . ,−ln|zn|)

≤ −ζ(−ln|z1|, . . . ,−ln|zn|)

≤ νϕ(x⁰) max

ln|z1|

x⁰₁ , . . . , ln|zn| x⁰_n

:=ψ(z1, . . . , zn).

By Lemma 3.1 and Remark 1.6 we get

f(e1(ϕ), . . . , en(ϕ))≤f(e1(ψ), . . . , en(ψ)) =c(ψ) = 1

νϕ(x⁰) =c(ϕ).

(3.5) Reduction to the case of psh functions with analytic singularities.

In the second step, we reduce the proof to the case ϕ = log(|f1|² + . . .+|fN|²), where f1, . . . , fN are germs of holomorphic functions at 0. Following the technique introduced in [Dem92], we let Hmϕ(Ω) be the Hilbert space of holomorphic functions f on Ω such that

Z

Ω

|f|²e^−2mϕdV <+∞, and let ψm = _2m¹ logP

|gm,k|² where{gm,k}k≥1 is an orthonormal basis of Hmϕ(Ω). Thanks to Theorem 4.2 in [DK00], mainly based on to the Ohsawa-Takegoshi L² extension theorem [OT87] (see also [Dem92]), there are constants C1, C2 >0 independent of m such that

ϕ(z)− C1

m ≤ψm(z)≤ sup

|ζ−z|<r

ϕ(ζ) + 1

mlog C2

rⁿ for every z ∈Ω and r < d(z, ∂Ω) and

ν(ϕ)− n

m ≤ν(ψm)≤ν(ϕ), 1

c(ϕ)− 1

m ≤ 1

c(ψm) ≤ 1 c(ϕ). By Lemma 3.1, we have

f(e1(ϕ), . . . , en(ϕ))≤f(e1(ψm), . . . , en(ψm)), ∀m≥1.

The above inequalities show that in order to prove the lower bound of c(ϕ) in Theorem 1.5, we only need to prove it for c(ψm) and let m tend to infinity. Also notice that since the Lelong numbers of a function ϕ ∈Ee(Ω) occur only on a discrete set, the same is true for the functions ψm.

(3.6) Reduction of the main theorem to the case of monomial ideals.

The final step consists of proving the theorem forϕ = log(|f1|²+. . . .+|fN|²), wheref1, . . . , fN

are germs of holomorphic functions at 0 [this is because the ideals (g_m,k)_k∈Nin the Noetherian ring OCⁿ,0 are always finitely generated]. SetJ = (f1, . . . , fN),c(J) = c(ϕ),ej(J) =ej(ϕ),

∀j = 0, . . . , n. By the final observation of 3.5, we can assume that J has an isolated zero at 0. Now, by fixing a multiplicative order on the monomials z^α = z₁^α1. . . z^α_nⁿ (see [Eis95]

Chap. 15 and [dFEM04]), it is well known that one can construct a flat family (Js)s∈C of ideals of OCⁿ,0 depending on a complex parameter s∈C, such that J0 is a monomial ideal, J1 =J and dim(OCⁿ,0/J_s^t) = dim(OCⁿ,0/J^t) for allsand t∈N; in factJ0 is just the initial ideal associated to J with respect to the monomial order. Moreover, we can arrange by a generic rotation of coordinates C^p ⊂Cⁿ that the family of idealsJs|C^p is also flat, and that the dimensions

dim(OC^p,0/(Js|C^p)^t) = dim(OC^p,0/(J|C^p)^t) compute the intermediate multiplicities

ep(Js) = lim

t→+∞

p!

t^p dim(OC^p,0/(Js|C^p)^t) =ep(J)

(notice, in the analytic setting, that the Lelong number of the (p, p)-current (dd^cϕ)^p at 0 is the Lelong number of its slice on a generic C^p ⊂Cⁿ); in particulare_p(J₀) =e_p(J) for allp.

The semicontinuity property of the log canonical threshold (see for example [DK00]) now implies that c(J0) ≤ c(Js) for s small. As c(Js) = c(J) for s 6= 0 (Js being a pull-back of J by a biholomorphism, in other words OCⁿ,0/Js ≃ OCⁿ,0/J as rings, see again [Eis95], chap. 15), the lower bound is valid for c(J) if it is valid for c(J0).

References

[BT76] E. Bedford and B. A. Taylor, The Dirichlet problem for a complex Monge-Amp`ere equation, Invent. Math.37 (1976) 1–44.

[BT82] E. Bedford and B. A. Taylor,A new capacity for plurisubharmonic functions, Acta Math. 149 (1982) 1–41.

[Ce04] U. Cegrell, The general definition of the complex Monge-Amp`ere operator, Ann. Inst. Fourier (Grenoble).,54(2004), 159–179.

[Che05] I. Chel’tsov, Birationally rigid Fano manifolds, Uspekhi Mat. Nauk 60:5 (2005), 71-160 and Russian Math. Surveys60:5 (2005), 875–965.

[Cor95] A. Corti, Factoring birational maps of threefolds after Sarkisov, J. Algebraic Geom., 4(1995), 223–254.

[Cor00] A. Corti,Singularities of linear systems and 3-fold birational geometry, London Math. Soc. Lec- ture Note Ser. 281(2000) 259–312.

[dFEM03] T. de Fernex, T, L. Ein and Mustat¸ˇa,Bounds for log canonical thresholds with applications to birational rigidity, Math. Res. Lett. 10(2003) 219–236.

[dFEM04] T. de Fernex, T, L. Ein and Mustat¸ˇa, Multiplicities and log canonical thresholds, J. Algebraic Geom.13(2004) 603–615.

[Dem87] J.-P. Demailly, Nombres de Lelong g´en´eralis´es, th´eor`emes d’int´egralit´e et d’analyticit´e, Acta Math.159(1987) 153–169.

[Dem92] J.-P. Demailly,Regularization of closed positive currents and Intersection Theory, J. Alg. Geom.

1(1992), 361–409.

[Dem93] J.-P. Demailly,Monge-Amp`ere operators, Lelong numbers and intersection theory, Complex Anal- ysis and Geometry, Univ. Series in Math., edited by V. Ancona and A. Silva, Plenum Press, New-York, 1993.

[DK00] J.-P. Demailly and J. Koll´ar, Semi-continuity of complex singularity exponents and K¨ahler- Einstein metrics on Fano orbifolds, Ann. Sci. Ecole Norm. Sup. (4)34 (2001), 525–556.

[Dem09] J.-P. Demailly,Estimates on Monge-Amp`ere operators derived from a local algebra inequality, in:

Complex Analysis and Digital geometry, Proceedings of the Kiselmanfest 2006, Acta Universitatis Upsaliensis, 2009.

[Eis95] D. Eisenbud,Commutative algebra with a view toward algebraic geometry, Grad. Texts in Math.

150, Springer, New York, 1995.

[Ho01] J. Howald,Multiplier ideals of monomial ideals, Trans. Amer. Math. Soc.353(2001), 2665–2671.

[IM72] V.A. Iskovskikh and I.Yu. Manin,Three-dimensional quartics and counterexamples to the L¨uroth problem, Mat. Sb.86 (1971), 140–166; English transl., Math. Sb.15(1972), 141–166.

[Isk01] V.A. Iskovskikh, Birational rigidity and Mori theory, Uspekhi Mat. Nauk 56:2 (2001) 3-86, English transl., Russian Math. Surveys56:2 (2001), 207–291.

[Kis87] C.O. Kiselman, Un nombre de Lelong raffin´e, S´eminaire d’Analyse Complexe et G´eom´etrie 198587, Facult´e des Sciences de Tunis & Facult´e des Sciences et Techniques de Monastir, (1987) 61–70.

[Kis94a] C.O. Kiselman,Attenuating the singularities of plurisubharmonic functions, Ann. Polon. Math.

60(1994) 173–197.

[OhT87] T. Ohsawa and K. Takegoshi, On the extension of L² holomorphic functions, Math. Zeit. 195 (1987) 197–204.

[Puk87] A.V. Pukhlikov, Birational automorphisms of a four-dimensional quintic, Invent. Math. 87 (1987), 303–329.

[Puk02] A.V. Pukhlikov, Birationally rigid Fano hypersurfaces, Izv. Ross. Akad. Nauk Ser. Mat. 66:6 (2002), 159-186; English translation, Izv. Math.66: (2002), 1243–1269.

[Sko72] H. Skoda, Sous-ensembles analytiques d’ordre fini ou infini dans Cⁿ, Bull. Soc. Math. France 100(1972) 353–408.

Jean-Pierre Demailly

Universit´e de Grenoble I, D´epartement de Math´ematiques Institut Fourier, 38402 Saint-Martin d’H`eres, France e-mail: jean-pierre.demailly@ujf-grenoble.fr Pha.m Ho`ang Hiˆe.p

Department of Mathematics, National University of Education 136-Xuan Thuy, Cau Giay, Hanoi, Vietnam

and Institut Fourier (on a Post-Doctoral grant from Univ. Grenoble I) e-mail: phhiep−vn@yahoo.com