betaFeedback

WorksMathematics

A second-order estimate linear in the gradient for complex Hessian equations

Abstract

We use the maximum principle to prove a second-order estimate with linear dependence on the gradient for the nondegenerate complex Hessian equation on a compact Kähler manifold. The proof combines C0C^0 compactness and a refined concavity inequality with the use of nearby smooth admissible functions. A standard interpolation argument then yields the known uniform gradient and complex Hessian bounds.

Introduction

Fully nonlinear partial differential equations (PDEs) that incorporate elementary symmetric polynomials arise frequently in complex differential geometry. Among these, complex kk-Hessian equations are an important class of PDEs with applications in complex geometry, mathematical physics, and complex dynamical systems. Since Yau’s pioneering work that resolved the renowned Calabi conjecture in [25], extensive research has been dedicated to investigating the existence, uniqueness, and fine properties of solutions to these equations, as well as developing innovative solution techniques. A notable example is the Fu–Yau equation [11, 12] arising from the Strominger system in complex geometry. In higher dimensions, the Fu–Yau equation corresponds precisely to an equation of complex 2-Hessian type and has been extensively studied by Chu–Huang–Zhu [3, 4] and by Phong–Picard–Zhang [20, 21, 22, 23]. Another representative example is the quaternionic Calabi conjecture on hyperKähler with torsion (HKT) manifolds, proposed by Alesker–Verbitsky [1]. In their work, the quaternionic Monge–Ampère equation is reformulated as a special case of the complex Hessian equation on complex manifolds.

For the complex kk-Hessian equations on compact manifolds, the central technical challenge lies in establishing a priori gradient estimates and second-order estimates for solutions. Under the non-negativity assumption on the holomorphic bisectional curvature, Hou [15], Jbilou [18], and Kokarev [19] successfully derived the global a priori gradient estimates for solutions to these equations. Hou–Ma–Wu [16] obtained a second-order estimate with quadratic dependence on the gradient. Then, Dinew–Kołodziej [9] combined this estimate with a blow-up argument and a Liouville-type theorem to establish the gradient estimate. On compact almost Hermitian manifolds, Chu–Huang–Zhang [5] established a real Hessian estimate with quadratic gradient dependence for a general class of fully nonlinear elliptic equations admitting a CC-subsolution, including the complex Hessian equation. For equations that also contain linear gradient terms, Huang–Zhang [17] obtained a full real C2C^2 estimate with quadratic gradient dependence under a CC-subsolution assumption.

Very recently, Chu–Liu [6] have established a new gradient estimate that integrates a logarithmic modulus of continuity, the Hou–Ma–Wu estimate, and local W2,pW^{2,p} estimates. Moreover, they explicitly clarified the precise dependence on the density. Their proof is built upon the stability property and on the auxiliary Monge–Ampère equations, instead of relying on blow-up and Liouville-type theorems. However, their proof still relies heavily on the Hou–Ma–Wu estimate. Guo–Song [14] have obtained gradient estimates for a broader class of concave elliptic equations on compact Hermitian manifolds under a fixed smooth admissible subsolution with a uniform strict sup-slope gap. Their proof uses stability and

Key words and phrases. Complex Hessian equation, second-order estimate, maximum principle, stability. comparison with nearby smooth admissible functions, without the Hou–Ma–Wu estimate or a Liouville-type theorem. Under additional structural assumptions they also proved a separate Hessian estimate using Dirichlet Green functions. Their gradient theorem requires weaker assumptions on the density than those in [6].

In Yau’s proof of the Calabi conjecture [25], the second-order estimate for the complex Monge–Ampère equation on a compact Kähler manifold is obtained from a C0C^{0} bound and the prescribed data, independently of any a priori gradient bound. The gradient estimate then follows from standard interpolation. This suggests a similar strategy for complex Hessian equations: first establish a suitable second-order estimate, and then recover the gradient bound by interpolation. It is therefore natural to ask:

Can one obtain a second-order estimate for complex Hessian equations that yields a gradient bound by interpolation? {#eq:1.1}

An estimate with linear dependence on the gradient is sufficient for this purpose:

sup⁡MΔωu≤C(1+sup⁡M∣∇u∣ω).\sup_{M}\Delta_{\omega}u \leq C\left(1+\sup_{M}|\nabla u|_{\omega}\right).

Together with a C0C^{0} bound, this inequality allows the gradient term to be absorbed by interpolation. Zhang–Zhang [26] formulated the question in this linear form and proved the estimate under additional assumptions. In particular, they treated a regime with 2≤k≤n−12 \leq k \leq n-1 in which the wedge density is sufficiently close to 1 in L1(M)L^{1}(M).

In the present work, we prove such a linear second-order estimate by the maximum principle in the nondegenerate Kähler setting. Interpolation then recovers the known gradient and complex Hessian bounds. This provides an alternative proof through the second-order estimate, without using an a priori gradient estimate, the Hou–Ma–Wu estimate, or Green functions. We now state our results more precisely.

Let (M,ω)(M,\omega) be a compact Kähler manifold of complex dimension n≥2n \geq2. For a smooth real-valued function uu, we write

ωu=ω+−1∂∂ˉu,Au=ω−1ωu.\omega_{u}=\omega+\sqrt{-1}\partial\bar{\partial}u,\qquad A_{u}=\omega^{-1}\omega_{u}.

Let λ(A)\lambda(A) denote the eigenvalues of the Hermitian endomorphism AA. For every fixed integer 2≤k≤n2 \leq k \leq n, the kk-Hessian operator is given by

f(λ)=σk(λ)1/kf(\lambda)=\sigma_{k}(\lambda)^{1/k}

in the convex cone

Γk={λ∈Rn:σj(λ)>0 for 1≤j≤k},\Gamma_{k}=\{\lambda\in\mathbb{R}^{n}:\sigma_{j}(\lambda)>0\text{ for }1\leq j\leq k\},

where the jj-th elementary symmetric polynomial σj\sigma_{j} is defined by

σj(λ1,λ2,⋯ ,λn)=∑1≤i1<i2<⋯<ij≤nλi1λi2⋯λij.\sigma_{j}(\lambda_{1},\lambda_{2},\cdots,\lambda_{n})= \sum_{1\leq i_{1}<i_{2}<\cdots<i_{j}\leq n} \lambda_{i_{1}}\lambda_{i_{2}}\cdots\lambda_{i_{j}}.

A function uu is called admissible if λ(Au)∈Γk\lambda(A_{u})\in\Gamma_{k} on MM. Denote

F(A)=f(λ(A)).F(A)=f(\lambda(A)).

In the present work, we consider the nondegenerate complex kk-Hessian equation

F(Au)=h>0,sup⁡Mu=0,λ(Au)∈Γk.(1)F(A_{u})=h>0,\qquad\sup_{M}u=0,\qquad\lambda(A_{u})\in\Gamma_{k}. \tag*{(1)}

Equivalently,

ωuk∧ωn−k=ψωn,ψ=hk(nk).(2)\omega_{u}^{k}\wedge\omega^{n-k}=\psi\omega^{n},\qquad\psi=\frac{h^{k}}{\binom{n}{k}}. \tag*{(2)}

The compatibility condition ∫Mψωn=∫Mωn\int_{M}\psi\omega^{n}=\int_{M}\omega^{n} is automatic for an existing solution.

In local holomorphic coordinates, we write ω=−1gijˉ dzi∧dzˉj\omega=\sqrt{-1}g_{i\bar{j}}\,dz^{i}\wedge d\bar{z}^{j}. The inverse metric components gijˉg^{i\bar{j}} are characterized by

∑j=1ngijˉgkjˉ=δki.\sum_{j=1}^{n}g^{i\bar{j}}g_{k\bar{j}}=\delta^{i}_{k}.

Throughout the paper we use the conventions

Δωu=gijˉuijˉ,∣∇u∣ω2=gijˉuiujˉ.\Delta_{\omega}u=g^{i\bar{j}}u_{i\bar{j}},\qquad\lvert\nabla u\rvert_{\omega}^{2}=g^{i\bar{j}}u_{i}u_{\bar{j}}.

For convenience, we write λ(Au)=(λ1(Au),…,λn(Au))\lambda(A_{u})=(\lambda_{1}(A_{u}),\ldots,\lambda_{n}(A_{u})), with the eigenvalues arranged in decreasing order; that is,

λ1(Au)≥λ2(Au)≥⋯≥λn(Au).(3)\lambda_{1}(A_{u})\geq\lambda_{2}(A_{u})\geq\cdots\geq\lambda_{n}(A_{u}). \tag*{(3)}

Therefore,

tr⁡Au≡∑iλi(Au)=n+Δωu.\operatorname{tr}A_{u}\equiv\sum_{i}\lambda_{i}(A_{u})=n+\Delta_{\omega}u.

Our main result is the following estimate.

Theorem 1.1. Fix positive constants H1H_{1}, H2H_{2}, UU. Let uu be a smooth admissible solution of (1) such that

h≥H1>0,∥h∥C2(M)≤H2,osc⁡Mu≤U.(4)h\geq H_{1}>0,\qquad\lVert h\rVert_{C^{2}(M)}\leq H_{2},\qquad\operatorname{osc}_{M}u\leq U. \tag*{(4)}

There exists a constant CC, depending only on

D=(M,ω,n,k,H1,H2,U),\mathcal{D}=(M,\omega,n,k,H_{1},H_{2},U),

such that

sup⁡Mλ1(Au)≤C(1+sup⁡M∣∇u∣ω2)1/2.(5)\sup_{M}\lambda_{1}(A_{u})\leq C\left(1+\sup_{M}\lvert\nabla u\rvert_{\omega}^{2}\right)^{1/2}. \tag*{(5)}

Consequently,

sup⁡MΔωu≤C(1+sup⁡M∣∇u∣ω).(6)\sup_{M}\Delta_{\omega}u\leq C\left(1+\sup_{M}\lvert\nabla u\rvert_{\omega}\right). \tag*{(6)}

The linear dependence in (6) allows us to close the gradient estimate by interpolation.

Corollary 1.2. Under the assumptions of Theorem 1.1, there is a constant C=C(D)C=C(\mathcal{D}) such that

∥∇u∥L∞(M)+∥∂∂ˉu∥L∞(M)≤C.(7)\lVert\nabla u\rVert_{L^{\infty}(M)}+\lVert\partial\bar{\partial}u\rVert_{L^{\infty}(M)}\leq C. \tag*{(7)}

We now briefly outline the core technical ingredients of the present work:

  1. In the prior work [26], the authors omitted the crucial third-order term

Fiiˉ∑p>1∣up1ˉi∣2+∣u1pˉi∣2λ1−λp(8)F^{i\bar{i}}\sum_{p>1}\frac{\lvert u_{p\bar{1}i}\rvert^{2}+\lvert u_{1\bar{p}i}\rvert^{2}}{\lambda_{1}-\lambda_{p}} \tag*{(8)}

arising from the computation of L(λ1)\mathcal{L}(\lambda_{1}). In the present work, we invoke an improved lower bound for L(log⁡λ1)\mathcal{L}(\log\lambda_{1}) established by Guo–Song [14], which makes essential use of the term (8) and concavity.

  1. In order to derive (5), we employ a comparison argument inspired by the work of Guo–Song in [14]. More precisely, we choose a smooth admissible function vv close to uu such that ∥u−v∥L∞≤ε\lVert u-v\rVert_{L^{\infty}}\leq\varepsilon for a small constant ε>0\varepsilon>0. Then, we consider a new auxiliary function of the form

Q=log⁡λ1−12log⁡(1−∣∇u∣ω22K)−Bu+N(u−v−ε)2,(9)Q=\log\lambda_{1}-\frac{1}{2}\log\left(1-\frac{\lvert\nabla u\rvert_{\omega}^{2}}{2K}\right)-Bu+N(u-v-\varepsilon)^{2}, \tag*{(9)}

where K=1+sup⁡M∣∇u∣ω2K=1+\sup_{M}\lvert\nabla u\rvert_{\omega}^{2}, N=16B2N=16B^{2}, and B>1B>1 is a sufficiently large constant with ε=164B\varepsilon=\frac{1}{64B}. With this construction of the auxiliary function QQ, we are able to establish a comparison estimate for second derivatives that exhibits linear dependence on both the gradient bound and ∥∇v∥L∞\lVert\nabla v\rVert_{L^{\infty}}. The remaining ingredient is the C0(M)C^{0}(M) precompactness of the solution class, which follows from [8, 6] independently of gradient and second-order estimates. Following the argument in the proof of Theorem 1.1 of [14], we use a finite covering of the solution class in C0(M)C^{0}(M) to obtain the uniform estimate (5) with a global constant independent of the individual solutions in SD\mathcal{S}_{D}.

At the end of the section, we give some further discussion on the auxiliary function QQ defined in (9). Define a function γ(t)=Bt−N(t−v−ε)2\gamma(t)=Bt-N(t-v-\varepsilon)^2 on a suitable interval, so that QQ can be written as

Q=log⁡λ1−12log⁡(1−∣∇u∣ω22K)−γ(u).Q=\log\lambda_{1}-\frac{1}{2}\log\left(1-\frac{|\nabla u|_{\omega}^{2}}{2K}\right)-\gamma(u).

A direct computation gives

γ′(u)=B−2N(u−v−ε),γ′′(u)=−2N.\gamma'(u)=B-2N(u-v-\varepsilon),\qquad\gamma''(u)=-2N.

With the choices of N=16B2N=16B^{2} and ε=164B\varepsilon=\frac{1}{64B}, it is not hard to verify the following important properties:

γ′(u)≥B>1,−[γ′′(u)+(γ′(u))2]≥28B2.\gamma'(u)\ge B>1,\qquad-\left[\gamma''(u)+(\gamma'(u))^{2}\right]\ge28B^{2}.

As noted by Błocki in [2], such a function γ\gamma cannot be constructed without introducing the nearby smooth admissible function vv.

The paper is organized as follows. In § 2, we first collect some elementary inequalities about the kk-Hessian operator, then derive a concavity inequality motivated by recent work of Dong–Zhang [10], and lastly prove compactness of the set of normalized smooth admissible solutions to (1). In § 3, we prove an improved lower bound for L(log⁡λ1)\mathcal{L}(\log\lambda_{1}), by adapting an argument similar to that in Lemma 6.1 of [14]. In § 4, we prove a comparison estimate for second derivatives. Finally, we complete the proofs of Theorem 1.1 and Corollary 1.2 in § 5.

Preliminaries

Elementary inequalities

For f=σk1/kf=\sigma_{k}^{1/k}, we write

fi=∂f∂λi,fij=∂2f∂λi∂λj,F=∑ifi.f_i=\frac{\partial f}{\partial\lambda_i},\qquad f_{ij}=\frac{\partial^2f}{\partial\lambda_i\partial\lambda_j},\qquad\mathcal{F}=\sum_i f_i.

By homogeneity,

∑ifiλi=f.(10)\sum_i f_i\lambda_i=f. \tag*{(10)}

The Newton–Maclaurin inequalities imply

F≥τn,k>0,σ1(λ)≤Cn,khFk−1,(11)\mathcal{F}\ge\tau_{n,k}>0,\qquad\sigma_1(\lambda)\le C_{n,k}h\mathcal{F}^{k-1}, \tag*{(11)}

for some positive constants τn,k\tau_{n,k}, Cn,kC_{n,k} depending only on nn and kk. This is because

F=n−k+1kσk1/k−1σk−1,σk−1k−1≥cn,kσ1σkk−2.\mathcal{F}=\frac{n-k+1}{k}\sigma_k^{1/k-1}\sigma_{k-1},\qquad\sigma_{k-1}^{k-1}\ge c_{n,k}\sigma_1\sigma_k^{k-2}.

In particular, one can take τn,k=(nk)1/k\tau_{n,k}=\binom{n}{k}^{1/k} in (11). Indeed, the concavity of ff at λ\lambda, applied to 1=(1,…,1)\mathbf{1}=(1,\ldots,1), and homogeneity give

f(1)≤f(λ)+∑ifi(1−λi)=F.f(\mathbf{1})\le f(\lambda)+\sum_i f_i(1-\lambda_i)=\mathcal{F}.

The ordering of the eigenvalues gives

0<f1≤⋯≤fn,fn≥F/n.(12)0<f_1\le\cdots\le f_n,\qquad f_n\ge\mathcal{F}/n. \tag*{(12)}

Since Γk⊂Γ2\Gamma_k\subset\Gamma_2,

∑iλi2<σ1(λ)2,0<λ1≤σ1(λ)≤nλ1.(13)\sum_i\lambda_i^2<\sigma_1(\lambda)^2,\qquad0<\lambda_1\le\sigma_1(\lambda)\le n\lambda_1. \tag*{(13)}

These inequalities will also control the curvature terms in the covariant derivative calculations.

A refined concavity inequality

We now give an algebraic fact that we will use in the next section. This is just a restatement of [10 Lemma 3.2] adapted to our choice f=σk1/kf=\sigma_k^{1/k}.

Lemma 2.1. For each β∈(0,1/2)\beta\in(0,1/2) there exist a small constant 0<δβ≤β0 < \delta_\beta\le\beta and a large constant 1≤Rβ<∞1 \le R_\beta< \infty, depending only on n,k,H2,βn,k,H_2,\beta, such that if λ∈Γk\lambda\in\Gamma_k satisfies

0<f(λ)≤H2(14)0 < f(\lambda) \le H_2 \tag*{(14)}

and

λ1≥Rβ,λn≥−δβλ1,(15)\lambda_1 \ge R_\beta,\qquad\lambda_n \ge-\delta_\beta\lambda_1, \tag*{(15)}

then, for every η∈Cn\eta\in\mathbb{C}^n,

−∑p,qfpqηpηq‾+∑i>1fi∣ηi∣2(1+δβ)λ1≥(1−β)f1∣η1∣2λ1−k+1f∣∑ifiηi∣2.(16)-\sum_{p,q} f_{pq}\eta_p\overline{\eta_q}+\sum_{i>1}\frac{f_i|\eta_i|^2}{(1+\delta_\beta)\lambda_1}\ge(1-\beta)\frac{f_1|\eta_1|^2}{\lambda_1}-\frac{k+1}{f}\left|\sum_i f_i\eta_i\right|^2. \tag*{(16)}

Proof. The inequality in [10 Lemma 3.2] gives

−∑p,qσk−2;pqηpηq‾σk+2∣∑iσk−1;iηi∣2σk2+∑i>1σk−1;i∣ηi∣2(1+δβ)λ1σk≥(1−β)σk−1;1∣η1∣2λ1σk,(17)\frac{-\sum_{p,q}\sigma_{k-2;pq}\eta_p\overline{\eta_q}}{\sigma_k}+2\frac{\left|\sum_i\sigma_{k-1;i}\eta_i\right|^2}{\sigma_k^2}+\sum_{i>1}\frac{\sigma_{k-1;i}|\eta_i|^2}{(1+\delta_\beta)\lambda_1\sigma_k}\ge(1-\beta)\frac{\sigma_{k-1;1}|\eta_1|^2}{\lambda_1\sigma_k}, \tag*{(17)}

where

σk−1;i=∂σk∂λi,σk−2;ij=∂2σk∂λi∂λj.\sigma_{k-1;i}=\frac{\partial\sigma_k}{\partial\lambda_i},\qquad\sigma_{k-2;ij}=\frac{\partial^2\sigma_k}{\partial\lambda_i\partial\lambda_j}.

Multiplication by f/kf/k, together with

fp=fkσkσk−1;p,fpq=fkσkσk−2;pq−(k−1)fk2σk2σk−1;pσk−1;q,f_p=\frac{f}{k\sigma_k}\sigma_{k-1;p},\qquad f_{pq}=\frac{f}{k\sigma_k}\sigma_{k-2;pq}-\frac{(k-1)f}{k^2\sigma_k^2}\sigma_{k-1;p}\sigma_{k-1;q},

gives the desired inequality (16).

To make these thresholds {Rβ,δβ}\{R_\beta,\delta_\beta\} uniform, we first apply [10 Lemma 3.2] at the fixed level f(μ)=1f(\mu)=1, which yields the thresholds {δβ,rβ}\{\delta_\beta,r_\beta\} that depend only on nn, kk and β\beta. Then for a general λ\lambda, we define

μ=λf(λ).\mu=\frac{\lambda}{f(\lambda)}.

The ratio λnλ1\frac{\lambda_n}{\lambda_1} remains invariant, i.e., λnλ1=μnμ1\frac{\lambda_n}{\lambda_1}=\frac{\mu_n}{\mu_1}. Moreover, for any fixed η∈Cn\eta\in\mathbb{C}^n, each term in (16) is homogeneous of degree −1-1 in λ\lambda. As a consequence, the unit-level inequality formulated in terms of μ\mu implies the desired inequality in terms of λ\lambda, provided that

μ1≥rβ,μn≥−δβμ1.\mu_1\ge r_\beta,\qquad\mu_n\ge-\delta_\beta\mu_1.

These conditions reduce to λ1≥rβf(λ)\lambda_1\ge r_\beta f(\lambda) and λn≥−δβλ1\lambda_n\ge-\delta_\beta\lambda_1. It therefore suffices to set Rβ=max⁡{1,H2rβ}R_\beta=\max\{1,H_2r_\beta\}. This completes the proof. □

Uniform continuity of the solutions

In the lemma below, we prove a compactness result that does not depend on any norm of the derivatives of uu, which serves as the crucial ingredient for establishing the Hessian estimate (5).

Lemma 2.2. For a fixed DD, the family of normalized smooth admissible solutions satisfying (1) and (4) is precompact in C0(M)C^0(M).

Proof. From the assumptions (4), we know that both the wedge densities ψ\psi in (2) and the polynomial densities σk(λ(Au))=hk\sigma_k(\lambda(A_u))=h^k are uniformly bounded in Lp(M)L^p(M) for all exponents p>n/kp>n/k.

  1. For 1<k<n1<k<n, we apply the equicontinuity result of Dinew–Kołodziej [8 Corollary 3.2] to this family.

(ii) For k=nk=n, equicontinuity follows from the modulus of continuity in [6 Theorem 3.1]. The same theorem also applies for each 1≤k≤n1 \le k \le n.

The normalization sup⁡Mu=0\sup_M u=0 and the bound osc⁡Mu≤U\operatorname{osc}_M u \le U give −U≤u≤0-U \le u \le0. The Arzelà–Ascoli theorem therefore gives precompactness in C0(M)C^0(M). This compactness property follows from zeroth order stability alone, independently of any gradient or second-order estimates.

Remark 2.3. It should be noted that the derivation of the modulus of continuity in [6 Section 3] does not make use of the Hou–Ma–Wu estimate, which only appears in their subsequent gradient estimates. For our argument, only compactness is required; no quantitative modulus of continuity appears in the maximum principle argument that follows.

Derivatives of the largest eigenvalue

Let ∇\nabla be the Chern connection of ω\omega, which agrees with the Levi–Civita connection in the Kähler setting. Derivative indices are appended in the order in which the covariant derivatives are taken:

T;ijˉ=∇jˉ∇iT,upqˉijˉ=∇jˉ∇i(upqˉ).T_{;i\bar j}=\nabla_{\bar j}\nabla_iT,\qquad u_{p\bar q i\bar j}=\nabla_{\bar j}\nabla_i(u_{p\bar q}).

We omit semicolons in derivative subscripts and use the curvature convention

Γipr=grsˉ∂igpsˉ,Rijˉpqˉ=−grqˉ∂jˉΓipr.\Gamma^r_{ip}=g^{r\bar s}\partial_i g_{p\bar s},\qquad R_{i\bar j p\bar q}=-g_{r\bar q}\partial_{\bar j}\Gamma^r_{ip}.

In normal holomorphic coordinates, this gives Rijˉpqˉ=−∂i∂jˉgpqˉR_{i\bar j p\bar q}=-\partial_i\partial_{\bar j}g_{p\bar q} at the center. Metric compatibility gives ∇g=0\nabla g=0, so all covariant derivatives of gg vanish. Lowering the endomorphism index of AuA_u with gg, we write

Apqˉu=(ωu)pqˉ=gpqˉ+upqˉ.A^u_{p\bar q}=(\omega_u)_{p\bar q}=g_{p\bar q}+u_{p\bar q}.

In an ω\omega-unitary frame, this Hermitian matrix represents AuA_u. For the matrix derivatives of FF, we write

Fijˉ=∂F∂Aijˉu,Fijˉ,klˉ=∂2F∂Aijˉu∂Aklˉu,∣∂w∣F2=Fijˉwiwjˉ.F^{i\bar j}=\frac{\partial F}{\partial A^u_{i\bar j}},\qquad F^{i\bar j,k\bar l}=\frac{\partial^2F}{\partial A^u_{i\bar j}\partial A^u_{k\bar l}},\qquad|\partial w|_F^2=F^{i\bar j}w_iw_{\bar j}.

The pointwise component calculations below use an ω\omega-unitary frame diagonalizing AuA_u. In this frame,

gijˉ=δij,Aijˉu=λiδij,Fijˉ=fiδij,g_{i\bar j}=\delta_{ij},\qquad A^u_{i\bar j}=\lambda_i\delta_{ij},\qquad F^{i\bar j}=f_i\delta_{ij},

with no summation over the free indices i,ji,j in these identities. The linearized operator associated with (1.2) is given by

L(w)=Fijˉ∇i∇jˉw.\mathcal{L}(w)=F^{i\bar j}\nabla_i\nabla_{\bar j}w.

Notice that the assumption (2.6) will be checked at the auxiliary maximum in one case of the proof. As a result, our use of [10 Lemma 3.2] imposes no additional global eigenvalue condition on the solution.

In this section, we will establish an improved lower bound for L\mathcal{L} applied to the logarithm of the largest eigenvalue.

Lemma 3.1. Let uu be a smooth admissible solution of (1.2). Then

(1) There exists a constant C0=C0(D)C_0=C_0(D) such that

L(log⁡λ1)≥−∣∂log⁡λ1∣F2−C0F.(18)\mathcal{L}(\log\lambda_1)\ge-|\partial\log\lambda_1|_F^2-C_0F. \tag*{(18)}

(2) Moreover, for every β∈(0,1/2)\beta\in(0,1/2) such that λ(Au)\lambda(A_u) fulfills (2.6), we have

L(log⁡λ1)≥−β∣∂log⁡λ1∣F2−C1F.(19)\mathcal{L}(\log\lambda_1)\ge-\beta|\partial\log\lambda_1|_F^2-C_1F. \tag*{(19)}

Here the constant C1=C1(D)C_1=C_1(D) is independent of β\beta.

At multiple eigenvalues, both inequalities hold for smooth functions touching λ1\lambda_1 from above, with λ1\lambda_1 replaced by the test function on the left and in its gradient.

Proof. The proof originates from [14 Lemma 6.1], and we include the calculation here for the sake of completeness.

Step 1. First suppose that λ1\lambda_1 is simple. In a normal unitary frame diagonalizing AuA_u, for each p=1,⋯ ,np=1,\cdots,n we write

ηp=∇1(ωu)ppˉ,zp=∇p(ωu)11ˉ.(20)\eta_p=\nabla_1(\omega_u)_{p\bar p},\qquad z_p=\nabla_p(\omega_u)_{1\bar1}. \tag*{(20)}

Closedness of ωu\omega_u gives the Codazzi identities, in particular

z1=η1,zp=∇1(ωu)p1ˉ.(21)z_1=\eta_1,\qquad z_p=\nabla_1(\omega_u)_{p\bar1}. \tag*{(21)}

Differentiating equation (1) once yields

∑pfpηp=h1.(22)\sum_p f_p\eta_p=h_1. \tag*{(22)}

At a diagonal matrix with simple largest eigenvalue λ1\lambda_1, the standard derivative formulas are as follows (see [13, 24]):

λ1pq=δ1pδ1q,\lambda_1^{pq}=\delta_{1p}\delta_{1q},
λ1pq,rs={1λ1−λp,q=r=1, p=s>1,1λ1−λr,p=s=1, r=q>1,0,otherwise.(23)\lambda_1^{pq,rs}= \begin{cases} \dfrac{1}{\lambda_1-\lambda_p}, & q=r=1,\ p=s>1,\\ \dfrac{1}{\lambda_1-\lambda_r}, & p=s=1,\ r=q>1,\\ 0, & \text{otherwise}. \end{cases} \tag*{(23)}

where λ1pq\lambda_1^{pq} represents the derivative with respect to the entry located at the pp-th row and qq-th column. It follows from (23) along with the Kählerity of ω\omega that

λ1,i=λ1pq(Apqu)i=u11iˉ,\lambda_{1,i}=\lambda_1^{pq}(A^u_{pq})_i=u_{11\bar i},
λ1,iiˉ=λ1pq,rs(Apqu)i(Arsu)iˉ+λ1pq(Apqu)iiˉ=u11iˉi+∑p>1∣up1ˉi∣2+∣u1pˉi∣2λ1−λp.(24)\begin{aligned} \lambda_{1,i\bar i} &=\lambda_1^{pq,rs}(A^u_{pq})_i(A^u_{rs})_{\bar i}+\lambda_1^{pq}(A^u_{pq})_{i\bar i}\\ &=u_{11\bar i i}+\sum_{p>1}\frac{|u_{p\bar1 i}|^2+|u_{1\bar p i}|^2}{\lambda_1-\lambda_p}. \tag*{(24)} \end{aligned}

By commuting the covariant derivatives, we have

u11ˉiiˉ=uiiˉ11ˉ+(u11ˉ−uiiˉ)R11ˉiiˉ=uiiˉ11ˉ+(λ1−λi)R11ˉiiˉ≥uiiˉ11ˉ+(λ1−λi)inf⁡l≠1R11ˉllˉ.(25)\begin{aligned} u_{1\bar1 i\bar i} &=u_{i\bar i1\bar1}+(u_{1\bar1}-u_{i\bar i})R_{1\bar1 i\bar i}\\ &=u_{i\bar i1\bar1}+(\lambda_1-\lambda_i)R_{1\bar1 i\bar i}\\ &\ge u_{i\bar i1\bar1}+(\lambda_1-\lambda_i)\inf_{l\ne1}R_{1\bar1l\bar l}. \tag*{(25)} \end{aligned}

It follows that

L(λ1)≥Fiiˉuiiˉ11ˉ+inf⁡l≠1R11ˉllˉ∑i=1nFiiˉ(λ1−λi)+Fiiˉ∑p>1∣up1ˉi∣2+∣u1pˉi∣2λ1−λp≥Fiiˉuiiˉ11ˉ−CR∑i=1nFiiˉ(λ1−λi)+∑i>1Fiiˉ∣u1iˉi∣2λ1−λi+∑p>1F11ˉ∣up1ˉ1∣2λ1−λp≥Fiiˉuiiˉ11ˉ−CR∑i=1nFiiˉ(λ1−λi)+∑i>1fi∣ηi∣2λ1−λi+∑i>1f1∣zi∣2λ1−λi.(26)\begin{aligned} \mathcal{L}(\lambda_1) &\ge F^{i\bar i}u_{i\bar i1\bar1}+\inf_{l\ne1}R_{1\bar1l\bar l}\sum_{i=1}^{n}F^{i\bar i}(\lambda_1-\lambda_i) +F^{i\bar i}\sum_{p>1}\frac{|u_{p\bar1 i}|^2+|u_{1\bar p i}|^2}{\lambda_1-\lambda_p}\\ &\ge F^{i\bar i}u_{i\bar i1\bar1}-C_R\sum_{i=1}^{n}F^{i\bar i}(\lambda_1-\lambda_i) +\sum_{i>1}\frac{F^{i\bar i}|u_{1\bar i i}|^2}{\lambda_1-\lambda_i} +\sum_{p>1}\frac{F^{1\bar1}|u_{p\bar1 1}|^2}{\lambda_1-\lambda_p}\\ &\ge F^{i\bar i}u_{i\bar i1\bar1}-C_R\sum_{i=1}^{n}F^{i\bar i}(\lambda_1-\lambda_i) +\sum_{i>1}\frac{f_i|\eta_i|^2}{\lambda_1-\lambda_i} +\sum_{i>1}\frac{f_1|z_i|^2}{\lambda_1-\lambda_i}. \tag*{(26)} \end{aligned}

where CRC_R is a constant depending on the holomorphic bisectional curvature. Differentiating equation (1) twice, we see that

Fiiˉuiiˉ11ˉ+Fij,kluij1ˉuklˉ1=h11ˉ.F^{i\bar i}u_{i\bar i1\bar1}+F^{ij,kl}u_{ij\bar1}u_{k\bar l1}=h_{1\bar1}.

Inserting it into (26), we infer that

L(λ1)≥−Fiˉjˉ,klˉuijˉ1uklˉ1−CR∑i=1nFiˉi(λ1−λi)+h11ˉ+∑i>1fi∣ηi∣2λ1−λi+∑i>1f1∣zi∣2λ1−λi≥−Fiˉjˉ,klˉuijˉ1uklˉ1−CR(1+λ1)F+h11ˉ+∑i>1fi∣ηi∣2λ1−λi+∑i>1f1∣zi∣2λ1−λi.\begin{aligned} \mathcal{L}(\lambda_1) \ge-F^{\bar{i}\bar{j},k\bar{l}}u_{i\bar{j}1}u_{k\bar{l}1}-C_R\sum_{i=1}^{n}F^{\bar{i}i}(\lambda_1-\lambda_i)+h_{1\bar{1}}+\sum_{i>1}\frac{f_i|\eta_i|^2}{\lambda_1-\lambda_i}+\sum_{i>1}\frac{f_1|z_i|^2}{\lambda_1-\lambda_i} \\ &\ge-F^{\bar{i}\bar{j},k\bar{l}}u_{i\bar{j}1}u_{k\bar{l}1}-C_R(1+\lambda_1)F+h_{1\bar{1}}+\sum_{i>1}\frac{f_i|\eta_i|^2}{\lambda_1-\lambda_i}+\sum_{i>1}\frac{f_1|z_i|^2}{\lambda_1-\lambda_i}. \end{aligned}

Here, we have used the homogeneity of FF in the last inequality. Notice that

Fiˉjˉ,klˉ={Fiˉi,kˉk,if i=j, k=l,Fiˉkˉ,kˉi,if i=l, k=j, i≠k,0,otherwise.F^{\bar{i}\bar{j},k\bar{l}}= \begin{cases} F^{\bar{i}i,\bar{k}k}, & \text{if } i=j,\ k=l,\\ F^{\bar{i}\bar{k},\bar{k}i}, & \text{if } i=l,\ k=j,\ i\ne k,\\ 0, & \text{otherwise.} \end{cases}

Therefore, by concavity of FF,

−Fiˉjˉ,klˉuijˉ1uklˉ1≥−Fiˉi,kˉkuiiˉ1ukkˉ1−∑i>1Fiˉ1,1ˉi(∣ui1ˉ1∣2+∣u1iˉ1∣2)≥−∑p,qfpqηpηq‾+∑i>1(fi−f1)∣zi∣2λ1−λi.(27)\begin{aligned} -F^{\bar{i}\bar{j},k\bar{l}}u_{i\bar{j}1}u_{k\bar{l}1}\ge-F^{\bar{i}i,\bar{k}k}u_{i\bar{i}1}u_{k\bar{k}1}-\sum_{i>1}F^{\bar{i}1,\bar{1}i}\left(|u_{i\bar{1}1}|^2+|u_{1\bar{i}1}|^2\right) \\ &\ge-\sum_{p,q}f_{pq}\eta_p\overline{\eta_q}+\sum_{i>1}\frac{(f_i-f_1)|z_i|^2}{\lambda_1-\lambda_i}. \tag*{(27)} \end{aligned}

Consequently, we have

L(λ1)≥h11ˉ−CR(1+λ1)F−∑p,qfpqηpηq‾+∑i>1fi∣ηi∣2λ1−λi+∑i>1fi∣zi∣2λ1−λi.(28)\mathcal{L}(\lambda_1)\ge h_{1\bar{1}}-C_R(1+\lambda_1)F-\sum_{p,q}f_{pq}\eta_p\overline{\eta_q}+\sum_{i>1}\frac{f_i|\eta_i|^2}{\lambda_1-\lambda_i}+\sum_{i>1}\frac{f_i|z_i|^2}{\lambda_1-\lambda_i}. \tag*{(28)}

Discarding the nonnegative quadratic terms by concavity, dividing by λ1\lambda_1, and using F≥τn,kF\ge\tau_{n,k} proves (18).

In the region of (19), it suffices to prove that there exists a constant CC, independent of β\beta, such that

L(λ1)≥(1−β)∣∂λ1∣F2λ1−Cλ1F=(1−β)f1∣η1∣2λ1+(1−β)∑i>1fi∣zi∣2λ1−Cλ1F(29)\begin{aligned} \mathcal{L}(\lambda_1)\ge(1-\beta)\frac{|\partial\lambda_1|_F^2}{\lambda_1}-C\lambda_1F \\ &= (1-\beta)\frac{f_1|\eta_1|^2}{\lambda_1}+(1-\beta)\sum_{i>1}\frac{f_i|z_i|^2}{\lambda_1}-C\lambda_1F \tag*{(29)} \end{aligned}

provided that

λ1−λi≤(1+δβ)λ1(i>1).(30)\lambda_1-\lambda_i\le(1+\delta_\beta)\lambda_1\qquad(i>1). \tag*{(30)}

On the one hand, using (16), (22) and (30), we have

−∑p,qfpqηpηq‾+∑i>1fi∣ηi∣2λ1−λi≥−∑p,qfpqηpηq‾+∑i>1fi∣ηi∣2(1+δβ)λ1≥(1−β)f1∣η1∣2λ1−(k+1)∣h1∣2h.(31)\begin{aligned} -\sum_{p,q}f_{pq}\eta_p\overline{\eta_q}+\sum_{i>1}\frac{f_i|\eta_i|^2}{\lambda_1-\lambda_i} \ge-\sum_{p,q}f_{pq}\eta_p\overline{\eta_q}+\sum_{i>1}\frac{f_i|\eta_i|^2}{(1+\delta_\beta)\lambda_1} \\ &\ge(1-\beta)\frac{f_1|\eta_1|^2}{\lambda_1}-(k+1)\frac{|h_1|^2}{h}. \tag*{(31)} \end{aligned}

On the other hand, since δβ≤β\delta_\beta\le\beta, we have 11+δβ≥1−β\frac{1}{1+\delta_\beta}\ge1-\beta. Hence, it follows from (30) that

∑i>1fi∣zi∣2λ1−λi≥(1−β)∑i>1fi∣zi∣2λ1.(32)\sum_{i>1}\frac{f_i|z_i|^2}{\lambda_1-\lambda_i}\ge(1-\beta)\sum_{i>1}\frac{f_i|z_i|^2}{\lambda_1}. \tag*{(32)}

Putting (31) and (32) into (28), we infer that

L(λ1)≥h11ˉ−(k+1)∣h1∣2h−CR(1+λ1)F+(1−β)f1∣η1∣2λ1+(1−β)∑i>1fi∣zi∣2λ1.(33)\mathcal{L}(\lambda_1) \ge h_{1\bar{1}}-(k+1)\frac{\lvert h_1\rvert^2}{h}-C_R(1+\lambda_1)\mathcal{F}+(1-\beta)\frac{f_1\lvert\eta_1\rvert^2}{\lambda_1}+(1-\beta)\sum_{i>1}\frac{f_i\lvert z_i\rvert^2}{\lambda_1}. \tag*{(33)}

The term involving the prescribed density is bounded below independently of β\beta:

h11ˉ−(k+1)∣h1∣2h≥−C2h_{1\bar{1}}-(k+1)\frac{\lvert h_1\rvert^2}{h}\ge-C_2

for some positive constant C2C_2 depending only on H1H_1, H2H_2 and kk. Since F≥τn,k\mathcal{F}\ge\tau_{n,k} and λ1≥1\lambda_1\ge1, this proves (29) with CC independent of β\beta.

Step 2. Now suppose that λ1\lambda_1 has multiplicity m>1m>1, and that a smooth function ℓ\ell touches λ1\lambda_1 from above. First order contact implies

∇i(ωu)pqˉ=ℓiδpq(p,q≤m).\nabla_i(\omega_u)_{p\bar q}=\ell_i\delta_{pq}\qquad(p,q\le m).

For i≤mi\le m, choosing p≤mp\le m with p≠ip\ne i and applying the Codazzi identity shows ℓi=∇p(ωu)ipˉ=0\ell_i=\nabla_p(\omega_u)_{i\bar p}=0. It follows that ηp=0\eta_p=0 for p≤mp\le m and zi=0z_i=0 for i≤mi\le m. The upper-contact second derivative formula retains the variation terms with positive gaps, i>mi>m. All the omitted numerators vanish, so the same calculation applies to ℓ\ell. □

The comparison estimate

We first prove a comparison estimate in which the permitted C0C^0 distance is independent of the comparison function.

Proposition 4.1. There exists a small constant ε=ε(D)>0\varepsilon=\varepsilon(\mathcal{D})>0 such that the following holds. Let uu satisfy the assumptions of Theorem 1.1, and let vv be any smooth admissible function with

∥u−v∥L∞(M)≤ε.\lVert u-v\rVert_{L^\infty(M)}\le\varepsilon.

Writing V=∥∇v∥L∞(M)V=\lVert\nabla v\rVert_{L^\infty(M)}, one has

sup⁡Mλ1(Au)≤C(D,V)(1+sup⁡M∣∇u∣ω2)1/2.(34)\sup_M\lambda_1(A_u)\le C(\mathcal{D},V)\left(1+\sup_M\lvert\nabla u\rvert_\omega^2\right)^{1/2}. \tag*{(34)}

Remark 4.2. We remark that during the proof of (34), no second derivative bound for vv is required. Instead, we use only the fact that λ(Av)∈Γk\lambda(A_v)\in\Gamma_k.

Proof. Set

P=∣∇u∣ω2,K=1+sup⁡MP,ϕ(s)=−12log⁡(1−s2K).P=\lvert\nabla u\rvert_\omega^2,\qquad K=1+\sup_M P,\qquad\phi(s)=-\frac{1}{2}\log\left(1-\frac{s}{2K}\right).

For 0≤s≤K0\le s\le K, we have

14K≤ϕ′≤12K,ϕ′′=2(ϕ′)2,0≤ϕ(P)≤12log⁡2.(35)\frac{1}{4K}\le\phi'\le\frac{1}{2K},\qquad\phi''=2(\phi')^2,\qquad0\le\phi(P)\le\frac{1}{2}\log2. \tag*{(35)}

Choose a sufficiently large constant B≥1B\ge1, depending only on D\mathcal{D}, and fix

N=16B2,ε=164B.(36)N=16B^2,\qquad\varepsilon=\frac{1}{64B}. \tag*{(36)}

The required lower bound for BB will be specified below and will not involve vv. Given vv as in the proposition, define

w=u−v−ε∈[−2ε,0].w=u-v-\varepsilon\in[-2\varepsilon,0].

Consider the auxiliary function

Q=log⁡λ1+ϕ(P)−Bu+Nw2.(37)Q=\log\lambda_1+\phi(P)-Bu+Nw^2. \tag*{(37)}

We first calculate the terms in L(Q)\mathcal{L}(Q). Concavity of FF and admissibility of vv imply

L(w)≤F(Au)−F(Av)≤h.\mathcal{L}(w)\le F(A_u)-F(A_v)\le h.

Since w≤0w \le0 and h≤H2h \le H_2, it follows that

L(Nw2)=2NwL(w)+2N∣∂w∣F2≥−4NεH2+2N∣∂w∣F2.(38)\mathcal{L}(Nw^2)=2Nw\mathcal{L}(w)+2N|\partial w|_F^2\ge-4N\varepsilon H_2+2N|\partial w|_F^2. \tag*{(38)}

This is the point at which admissibility of vv replaces a bound for its second derivatives. A direct calculation shows that

L(P)=Fiˉi(∣uiiˉ∣2+∑p=1n∣uip∣2)+∑p=1nFiˉi(upiiˉupˉ+upˉiiˉup).\mathcal{L}(P)=F^{\bar{i}i}\left(|u_{i\bar{i}}|^2+\sum_{p=1}^{n}|u_{ip}|^2\right)+\sum_{p=1}^{n}F^{\bar{i}i}\left(u_{p i\bar{i}}u_{\bar{p}}+u_{\bar{p}i\bar{i}}u_p\right).

By commuting the covariant derivatives,

upiiˉ=uiiˉp+∑q=1nuqRiiˉpqˉ,upˉiiˉ=uiiˉpˉ.u_{p i\bar{i}}=u_{i\bar{i}p}+\sum_{q=1}^{n}u_qR_{i\bar{i}p\bar{q}},\qquad u_{\bar{p}i\bar{i}}=u_{i\bar{i}\bar{p}}.

Differentiating (1) once, we obtain

Fiˉiuiiˉp=hp,Fiˉiuiiˉpˉ=hpˉ.F^{\bar{i}i}u_{i\bar{i}p}=h_p,\qquad F^{\bar{i}i}u_{i\bar{i}\bar{p}}=h_{\bar{p}}.

Thus, we infer that

∑p=1nFiˉi(upiiˉupˉ+upˉiiˉup)=2∑p=1nRe⁡(hpupˉ)+Fiˉi∑p,q=1nuqupˉRiiˉpqˉ≥−C3P−CRPF.(39)\begin{aligned} \sum_{p=1}^{n}F^{\bar{i}i}\left(u_{p i\bar{i}}u_{\bar{p}}+u_{\bar{p}i\bar{i}}u_p\right)=2\sum_{p=1}^{n}\operatorname{Re}(h_pu_{\bar{p}})+F^{\bar{i}i}\sum_{p,q=1}^{n}u_qu_{\bar{p}}R_{i\bar{i}p\bar{q}} \\ &\ge-C_3\sqrt{P}-C_RPF. \tag*{(39)} \end{aligned}

for some constant C3C_3 depending only on ∥∇h∥L∞\|\nabla h\|_{L^\infty}. It follows from (39) that

L(P)≥∑ifi(λi−1)2−C3P−CRPF.(40)\mathcal{L}(P)\ge\sum_i f_i(\lambda_i-1)^2-C_3\sqrt{P}-C_RPF. \tag*{(40)}

Using the elementary inequality (λi−1)2≥12λi2−1(\lambda_i-1)^2\ge\frac{1}{2}\lambda_i^2-1 and (35), we get

ϕ′L(P)≥18K∑ifiλi2−CRF−C3.(41)\phi'\mathcal{L}(P)\ge\frac{1}{8K}\sum_i f_i\lambda_i^2-C_RF-C_3. \tag*{(41)}

In addition, (10) implies

L(−Bu)=BF−Bh.(42)\mathcal{L}(-Bu)=BF-Bh. \tag*{(42)}

Let x0x_0 be a maximum point of QQ. We consider two cases.

(i) Assume λ1\lambda_1 is simple at x0x_0; that is, λ1\lambda_1 is smooth there. At x0x_0,

∂log⁡λ1=−ϕ′∂P+(B−2Nw)∂w+B∂v.(43)\partial\log\lambda_1=-\phi'\partial P+(B-2Nw)\partial w+B\partial v. \tag*{(43)}

Since B−2Nw≤B+4Nε=2BB-2Nw\le B+4N\varepsilon=2B, two applications of the elementary square inequality yield

∣∂log⁡λ1∣F2≤2(ϕ′)2∣∂P∣F2+16B2∣∂w∣F2+4B2V2F.(44)|\partial\log\lambda_1|_F^2\le2(\phi')^2|\partial P|_F^2+16B^2|\partial w|_F^2+4B^2V^2\mathcal{F}. \tag*{(44)}

(ii) Otherwise, the function

ℓ(x)=exp⁡(Q(x0)−ϕ(P(x))+Bu(x)−Nw(x)2)\ell(x)=\exp(Q(x_0)-\phi(P(x))+Bu(x)-Nw(x)^2)

touches λ1\lambda_1 from above at x0x_0. Lemma 3.1 applied to ℓ\ell justifies the same calculation. In particular, the corresponding smooth test for QQ is constant, which is sufficient for the maximum principle inequality.

Let β~∈(0,1]\widetilde{\beta}\in(0,1] denote the loss coefficient in Lemma 3.1; that is,

β~={β,if (2.6) holds,1,otherwise.(45)\widetilde{\beta}= \begin{cases} \beta, & \text{if (2.6) holds},\\ 1, & \text{otherwise}. \end{cases} \tag*{(45)}

The first two terms on the right of (44), after multiplication by β~\widetilde{\beta}, are absorbed because

ϕ′′−2β~(ϕ′)2≥0,2N−16β~B2≥16B2.\phi''-2\widetilde{\beta}(\phi')^2\ge0,\qquad2N-16\widetilde{\beta}B^2\ge16B^2.

Combining (38), (41), (42) and Lemma 3.1, we obtain at x0x_0, whenever the relevant eigenvalue threshold is satisfied,

0≥18K∑ifiλi2+(B−C0−C1−CR−4β~B2V2)F−C4.(46)0 \ge\frac{1}{8K}\sum_i f_i\lambda_i^2+\left(B-C_0-C_1-C_R-4\widetilde{\beta}B^2V^2\right)\mathcal{F}-C_4. \tag*{(46)}

Here C0C_0 and C1C_1 are the coefficients of F\mathcal{F} in Lemma 3.1. We may take

C4≡C3+(B+4Nε)H2=C3+2BH2,C_4\equiv C_3+(B+4N\varepsilon)H_2=C_3+2BH_2,

where C3C_3 is the constant term in (41). The constant C3C_3 depends only on DD and is independent of BB, KK, VV, β~\widetilde{\beta}. We now specify the earlier choice by taking

B≡C0+C1+CR+2.(47)B\equiv C_0+C_1+C_R+2. \tag*{(47)}

Thus BB, NN, ε\varepsilon are determined solely by DD.

After vv has been fixed, choose

0<β≤min⁡{12,18B2(1+V2)}.(48)0<\beta\leq\min\left\{\frac{1}{2},\frac{1}{8B^2(1+V^2)}\right\}. \tag*{(48)}

and let δβ,Rβ\delta_\beta,R_\beta be as in Lemma 3.1. If λ1(x0)<Rβ\lambda_1(x_0)<R_\beta, the required bound at x0x_0 is immediate. Otherwise, there are two cases.

Case 1: If λn(x0)≥−δβλ1(x0)\lambda_n(x_0)\geq-\delta_\beta\lambda_1(x_0). In this case we treat (46) with β~=β\widetilde{\beta}=\beta. By (47) and (48), we have

B−C0−C1−CR−4βB2V2≥1B-C_0-C_1-C_R-4\beta B^2V^2\geq1

and hence F(x0)≤C4\mathcal{F}(x_0)\leq C_4. It follows from (11) that

λ1(x0)≤σ1(λ(x0))≤Cn,kh(x0)F(x0)k−1≤C5(49)\lambda_1(x_0)\leq\sigma_1(\lambda(x_0))\leq C_{n,k}h(x_0)\mathcal{F}(x_0)^{k-1}\leq C_5 \tag*{(49)}

for a uniform constant C5C_5 depending only on DD.

Case 2: If λn(x0)<−δβλ1(x0)\lambda_n(x_0)<-\delta_\beta\lambda_1(x_0). In this case we treat (46) with β~=1\widetilde{\beta}=1. By (12), we have

∑ifiλi2≥fnλn2≥δβ2nλ12F.\sum_i f_i\lambda_i^2\geq f_n\lambda_n^2\geq\frac{\delta_\beta^2}{n}\lambda_1^2\mathcal{F}.

Consequently,

δβ28nKλ1(x0)2F≤4B2V2F+C4.\frac{\delta_\beta^2}{8nK}\lambda_1(x_0)^2\mathcal{F}\leq4B^2V^2\mathcal{F}+C_4.

Since F≥τn,k\mathcal{F}\geq\tau_{n,k}, we obtain

λ1(x0)≤C(D,V)K.(50)\lambda_1(x_0)\leq C(D,V)\sqrt{K}. \tag*{(50)}

Here the dependence on VV includes that of β\beta, δβ\delta_\beta, RβR_\beta.

In every case, we have λ1(x0)≤C(D,V)K\lambda_1(x_0)\leq C(D,V)\sqrt{K}. The oscillation of the remaining part of QQ is bounded by

osc⁡M(ϕ(P)−Bu+Nw2)≤12log⁡2+BU+4Nε2.\operatorname{osc}_M\left(\phi(P)-Bu+Nw^2\right)\leq\frac{1}{2}\log2+BU+4N\varepsilon^2.

Comparing Q(x)Q(x) with Q(x0)Q(x_0) proves (34). □

Uniformity and the gradient bound

Proof of Theorem 1.1

Proof of Theorem 1.1. Let SDS_D denote the family of normalized smooth admissible solutions of (1) satisfying (4). The right-hand side may vary within the prescribed bounds. If this class is empty, the assertion is vacuous. Suppose that the desired estimate (5) does not hold. Then there exists a sequence {uj}j∈N⊂SD\{u_j\}_{j\in\mathbb{N}}\subset S_D such that

sup⁡Mλ1(Auj)(1+sup⁡M∣∇uj∣ω2)1/2⟶∞.\frac{\sup_M\lambda_1(A_{u_j})}{\left(1+\sup_M|\nabla u_j|_\omega^2\right)^{1/2}}\longrightarrow\infty.

Fix ε=ε(D)\varepsilon=\varepsilon(D) from Proposition 4.1. By Lemma 2.2, SD\mathcal{S}_D is totally bounded in C0(M)C^0(M). It therefore admits a finite ε\varepsilon-net with centers in SD\mathcal{S}_D, so we may choose finitely many smooth admissible functions v1,…,vN∈SDv_1,\ldots,v_N\in\mathcal{S}_D such that

SD⊂⋃ℓ=1N{w∈C0(M):∥w−vℓ∥L∞(M)<ε}.(51)\mathcal{S}_D\subset\bigcup_{\ell=1}^{N}\left\{w\in C^0(M):\|w-v_\ell\|_{L^\infty(M)}<\varepsilon\right\}. \tag*{(51)}

The centers are chosen in the smooth solution class itself. Fix this finite family once for the class determined by DD. Therefore, there exist an index ℓ0∈{1,2,…,N}\ell_0\in\{1,2,\ldots,N\} and a subsequence {ujk}k∈N⊂SD\{u_{j_k}\}_{k\in\mathbb{N}}\subset\mathcal{S}_D such that

∥ujk−vℓ0∥L∞≤ε(for all sufficiently large k).\|u_{j_k}-v_{\ell_0}\|_{L^\infty}\leq\varepsilon\qquad\text{(for all sufficiently large } k\text{)}.

Then Proposition 4.1 gives

sup⁡Mλ1(Aujk)(1+sup⁡M∣∇ujk∣ω2)1/2≤C(D,Vℓ0),(52)\frac{\sup_M\lambda_1(A_{u_{j_k}})}{\left(1+\sup_M|\nabla u_{j_k}|_{\omega}^{2}\right)^{1/2}}\leq C(D,V_{\ell_0}), \tag*{(52)}

where we denote Vℓ=∥∇vℓ∥L∞(M)V_\ell=\|\nabla v_\ell\|_{L^\infty(M)}. The right-hand side is finite and independent of kk, which yields a contradiction. This completes the proof of Theorem 1.1.

Proof of Corollary 1.2

Proof. Without loss of generality, we may assume G=∥∇u∥L∞(M)>1G=\|\nabla u\|_{L^\infty(M)}>1. By compactness of MM, there exists a point p∈Mp\in M such that

G=∣∇u∣(p).G=|\nabla u|(p).

Fix an exponent q>2nq>2n. Local elliptic W2,qW^{2,q}-estimates and Sobolev embedding give constants r0∈(0,1)r_0\in(0,1) and C‾\overline C, depending only on (M,ω)(M,\omega) and the fixed exponent, such that for 0<r<r00<r<r_0,

r∣∇u∣(p)≤C‾osc⁡Br(p)u+C‾r2∥Δωu∥L∞(Br(p)).r|\nabla u|(p)\leq\overline C\operatorname{osc}_{B_r(p)}u+\overline Cr^2\|\Delta_\omega u\|_{L^\infty(B_r(p))}.

Here Br(p)⋐MB_r(p)\Subset M denotes the coordinate ball centered at pp with radius rr. Notice that Theorem 1.1 yields

∥Δωu∥L∞(M)≤C(D)(1+G)≤2C(D)G.\|\Delta_\omega u\|_{L^\infty(M)}\leq C(D)(1+G)\leq2C(D)G.

Therefore,

rG≤C‾U+2C‾C(D)r2G.(53)rG\leq\overline CU+2\overline CC(D)r^2G. \tag*{(53)}

If G≤r0−2G\leq r_0^{-2}, the gradient is already bounded. Otherwise, choose r=G−1/2<r0r=G^{-1/2}<r_0. Then (53) gives

G≤(C‾U+2C‾C(D))2.G\leq\left(\overline CU+2\overline CC(D)\right)^2.

Thus G≤C(D)G\leq C(D) in all cases. Theorem 1.1 now bounds λ1(Au)\lambda_1(A_u), and (13) bounds all the eigenvalues of AuA_u. Since Au=Id⁡+ω−1−1∂∂‾uA_u=\operatorname{Id}+\omega^{-1}\sqrt{-1}\partial\overline\partial u, this also gives the asserted complex Hessian bound and completes the proof. □

Our argument does not require an explicit modulus of continuity in the range 1<k<n1<k<n, in contrast to the proof of Chu–Liu [6 Theorem 1.2]. In this range, the equicontinuity result of Dinew–Kołodziej [8] suffices for the precompactness used in Lemma 2.2.

Further questions

We conclude with two questions concerning possible extensions of this proof.

The degenerate setting

Theorem 1.1 assumes a positive lower bound for hh. Under what regularity assumptions on a nonnegative wedge density ψ\psi can we use the present argument to obtain (5) uniformly along positive approximations? The distinction between the regularity of ψ\psi and that of its root hh is important: smoothness of a nonnegative ψ\psi need not give uniform C2C^2 control of hh. On the other hand, uniform C2C^2 bounds for nonnegative root densities control ∣∇h∣2/h|\nabla h|^2/h on {h>0}\{h>0\} independently of a positive lower bound. A passage to degenerate solutions would also require tracking the constants and allowing admissibility in Γ‾k\overline{\Gamma}_k. For related degenerate estimates, see Chu–McCleerey [7].

Non-Kähler setting

Can we adapt the present maximum principle argument to compact Hermitian manifolds? The Kähler condition enters the Codazzi identities and the largest-eigenvalue calculation. On a Hermitian background, torsion introduces additional third-order terms, although the gradient Bochner terms can still be controlled; see [14 Lemma 3.4]. The main issue for the present method is to control the torsion with error bounds independent of the small parameter β\beta. This uniformity allows us to fix the comparison distance before choosing the nearby smooth admissible function. Further cancellations or a modified auxiliary function may be needed to preserve this order of choices; see [10 Section 4] for the torsion terms arising in Hermitian estimates.

Acknowledgements The first author was supported by the NSFC (No. 12401075) and Fundamental Research Funds for the Central Universities (Grant No. 20720260220). The second author was supported by the NSFC (No. 12401065). The third author was supported by the National Key R and D Program of China 2020YFA0713100 and by NSFC (No. 12141104, 11801535 and 11721101).

Use of AI. This work is a continuation of arXiv:2610.07852. Compared with the previous version, the key improvement in the present work relies on the techniques of Guo and Song [14] for handling the concavity of the Hessian operators. During the preparation of this work, we use AI tools, particularly ChatGPT 6.1 Astra, to explore proof strategies and assist with calculations. They verified and revised all AI-generated suggestions incorporated into the paper and take full responsibility for its contents.

References

SCHOOL OF MATHEMATICAL SCIENCES, XIAMEN UNIVERSITY, XIAMEN 361005, P. R. CHINA

Email address: [email protected]

SCHOOL OF MATHEMATICS, HEFEI UNIVERSITY OF TECHNOLOGY, HEFEI 230009, P. R. CHINA

Email address: [email protected]

SCHOOL OF MATHEMATICS AND STATISTICS, NANJING UNIVERSITY OF SCIENCE AND TECHNOLOGY, NANJING 210094, CHINA

Email address: [email protected]

References

  1. [1]S. Alesker and M. Verbitsky, Quaternionic Monge-Ampère equation and Calabi problem for HKT-manifolds. Israel J. Math. 176, (2010), 109–138.arxiv.org/abs/0802.4202
  2. [2]Z. Blocki, A gradient estimate in the Calabi-Yau theorem. Math. Ann. 344, no. 2, (2009), 317–327.DOI
  3. [3]J. Chu, L. Huang and X. Zhu, The Fu-Yau equation on compact astheno-Kähler manifolds. Adv. Math. 346, (2019), 908–945.arxiv.org/abs/1803.01475
  4. [4]J. Chu, L. Huang and X. Zhu, The Fu-Yau equation in higher dimensions. Peking Math. J. 2, no. 1, (2019), 71–97.arxiv.org/abs/1801.09351
  5. [5]J. Chu, L. Huang and J. Zhang, Fully non-linear elliptic equations on compact almost Hermitian manifolds, Calc. Var. Partial Differential Equations 62 (2023), no. 3, Paper No. 105, 34 pp.arxiv.org/abs/2109.12566
  6. [6]J. Chu and Y. Liu, Alternative proof of the gradient estimate for complex Hessian equations, preprint, arXiv:2609.33091v1 (2026).arxiv.org/abs/2609.33091
  7. [7]J. Chu and N. McCleerey, Fully non-linear degenerate elliptic equations in complex geometry, J. Funct. Anal. 281 (2021), no. 9, Paper No. 109176, 45 pp.arxiv.org/abs/2010.03431
  8. [8]S. Dinew and S. Kołodziej, A priori estimates for complex Hessian equations, Anal. PDE 7 (2014), no. 1, 227–244.DOI
  9. [9]S. Dinew and S. Kołodziej, Liouville and Calabi–Yau type theorems for complex Hessian equations, Amer. J. Math. 139 (2017), no. 2, 403–415.arxiv.org/abs/1203.3995
  10. [10]W. Dong and R. Zhang, Second order estimates for equations with sums of Hessian operators on Hermitian manifolds, preprint, arXiv:2603.19045v1 (2026).arxiv.org/abs/2603.19045
  11. [11]J. Fu and S.-T. Yau, A Monge-Ampère-type equation motivated by string theory. Comm. Anal. Geom. 15, no. 1, (2007), 29–76.DOI
  12. [12]J. Fu and S.-T. Yau, The theory of superstring with flux on non-Kähler manifolds and the complex Monge-Ampère equation. J. Differential Geom. 78, no. 3, (2008), 369–428.
  13. [13]C. Gerhardt, Closed Weingarten hypersurfaces in Riemannian manifolds. J. Differential Geom. 43, (1996), 612–641.DOI
  14. [14]B. Guo and J. Song, Stability and gradient estimates for fully nonlinear elliptic equations on hermitian manifolds, preprint, arXiv:2609.39726v1 (2026).arxiv.org/abs/2609.39726
  15. [15]Z. Hou, Complex Hessian equation on Kähler manifold. Int. Math. Res. Not. IMRN (2009), no. 16, 3098–3111.arxiv.org/abs/0812.4549
  16. [16]Z. Hou, X.-N. Ma and D. Wu, A second order estimate for complex Hessian equations on a compact Kähler manifold, Math. Res. Lett. 17 (2010), no. 3, 547–561.DOI
  17. [17]L. Huang and J. Zhang, Fully nonlinear elliptic equations with gradient terms on compact almost Hermitian manifolds, Math. Z. 303 (2023), article 36.arxiv.org/abs/2112.02919
  18. [18]A. Jbilou, Équations hessiennes complexes sur des variétés kählériennes compactes. C. R. Math. Acad. Sci. Paris. 348, no. 1-2, (2010), 41–46.DOI
  19. [19]V. N. Kokarev, Mixed volume forms and a complex equation of Monge-Ampère type on Kähler manifolds of positive curvature. Izv. Ross. Akad. Nauk Ser. Mat. 74, no. 3, (2010), 65–78.DOI
  20. [20]D.-H. Phong, S. Picard and X.-W. Zhang, A second order estimate for general complex Hessian equations. Anal. PDE 9, no. 7, (2016), 1693–1709.arxiv.org/abs/1508.03254
  21. [21]D.-H. Phong, S. Picard and X.-W. Zhang, The Fu-Yau equation with negative slope parameter. Invent. Math. 209, (2017), 541–576.DOI
  22. [22]D.-H. Phong, S. Picard and X.-W. Zhang, On estimates for the Fu-Yau generalization of a Strominger system. J. Reine Angew. Math. 751, (2019), 243–274.
  23. [23]D.-H. Phong, S. Picard and X.-W. Zhang, Fu-Yau Hessian equations. J. Differential Geom. 118, no. 1, (2021), 147–187.arxiv.org/abs/1801.09842
  24. [24]J. Spruck, Geometric aspects of the theory of fully nonlinear elliptic equations. in Global theory of minimal surfaces, vol. 2, Amer. Math. Soc., Providence, RI, (2005), 283–309.
  25. [25]S.-T. Yau, On the Ricci curvature of a compact Kähler manifold and the complex Monge-Ampère equation. I. Comm. Pure Appl. Math. 31, no. 3, (1978), 339–411.DOI
  26. [26]J. Zhang and X. Zhang, The interpolation method for the complex Hessian equation on compact Kähler manifolds, preprint, arXiv:2610.07852.

Paper details

vlnc:W_36yu3rbm
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

Not recorded for this paper

View statement
Looking for related papers…
Contents