Introduction
We study pointwise gradient estimates for k k k -Hessian equations with measure data. For a smooth function u u u , the k k k -Hessian operator σ k ( D 2 u ) \sigma_k(D^2u) σ k ( D 2 u ) is the k k k -th elementary symmetric function of the eigenvalues of D 2 u D^2u D 2 u . Trudinger–Wang [14 , 15 ] extended this operator to a nonnegative Radon measure F k [ u ] F_k[u] F k [ u ] for proper weak k k k -convex functions and proved its weak continuity under local L 1 L^1 L 1 convergence. This formulation includes Hessian measures with singular parts.
Estimates for function values include Hölder bounds under density or ball-mass assumptions [13 , 15 ] and Wolff-potential estimates due to Labutin [9 ] and Phuc–Verbitsky [12 ] . Interior gradient estimates for smooth admissible solutions were obtained by Trudinger [13 ] and Chou–Wang [3 ] . The gradient potential theory for quasilinear equations [5 , 7 , 8 , 11 ] suggests a corresponding question for Hessian equations: how does the full Hessian measure control the gradient of a weak k k k -convex function?
For 2 ≤ k < n 2 \le k < n 2 ≤ k < n , we bound the gradient of a proper weak k k k -convex function u u u at Lebesgue points of D u Du D u by a local average of u u u and a Wolff potential of its full Hessian measure. In a comparison argument over successive scales, ordinary interior gradient estimates may introduce a fixed multiplier at each step. We prove a Lipschitz stability estimate for k k k -harmonic functions in which the preceding Lipschitz constant enters with coefficient one. The comparison errors can then be summed and bounded by the potential.
The mapping properties of potentials [4 ] then yield local Lipschitz regularity for F k [ u ] = f F_k[u] = f F k [ u ] = f with f ∈ L l o c n / k , 1 / k f \in L^{n/k,1/k}_{\mathrm{loc}} f ∈ L loc n / k , 1/ k . We also prove local Lipschitz regularity for F k [ u ] = ∣ D u ∣ p F_k[u] = |Du|^p F k [ u ] = ∣ D u ∣ p when 0 < p < n k / ( n − k ) 0 < p < nk/(n-k) 0 < p < nk / ( n − k ) . The second Lorentz index 1 / k 1/k 1/ k and the upper bound on p p p are sharp in the full weak class.
Main results
For a symmetric matrix A A A , let σ j ( A ) \sigma_j(A) σ j ( A ) denote the j j j -th elementary symmetric function of its eigenvalues, with σ 0 = 1 \sigma_0 = 1 σ 0 = 1 , and set
Γ k = { A : σ j ( A ) > 0 , 1 ≤ j ≤ k } . \Gamma_k = \{A : \sigma_j(A) > 0,\ 1 \le j \le k\}. Γ k = { A : σ j ( A ) > 0 , 1 ≤ j ≤ k } .
Let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be open.
2020 Mathematics Subject Classification . 35J60, 35B65, 31C45.
Key words and phrases. k k k -Hessian equations, Wolff potentials, Hessian measures, Lorentz spaces, gradient growth, critical regularity.
Definition 1.1. ([15 ] ) For 1 ≤ k ≤ n 1 \le k \le n 1 ≤ k ≤ n , a function u : Ω → [ − ∞ , ∞ ) u:\Omega\to[-\infty,\infty) u : Ω → [ − ∞ , ∞ ) is proper weak k k k -convex if
u u u is upper semicontinuous;
u u u is not identically − ∞ -\infty − ∞ on any connected component of Ω \Omega Ω ;
every C 2 C^2 C 2 function φ \varphi φ touching u u u from above at x ∈ Ω x \in\Omega x ∈ Ω satisfies D 2 φ ( x ) ∈ Γ ‾ k D^2\varphi(x) \in\overline{\Gamma}_k D 2 φ ( x ) ∈ Γ k .
We denote this class by Φ k ( Ω ) \Phi_k(\Omega) Φ k ( Ω ) .
For u ∈ Φ k ( Ω ) u \in\Phi_k(\Omega) u ∈ Φ k ( Ω ) , write μ = F k [ u ] \mu= F_k[u] μ = F k [ u ] for its full Hessian measure, following the notation in [12 Section 7, p. 903] . For smooth u u u ,
F k [ u ] = σ k ( D 2 u ) d x . F_k[u] = \sigma_k(D^2u)\,dx. F k [ u ] = σ k ( D 2 u ) d x .
Throughout the paper, F k [ u ] = f F_k[u] = f F k [ u ] = f means that the full Hessian measure equals f d x f\,dx f d x . For 2 ≤ k < n 2 \le k < n 2 ≤ k < n , weak gradients are locally integrable by (2.1), and their pointwise values are understood as Lebesgue values. We write \fint_E = |E|^{-1}\int_E .
For a nonnegative Radon measure μ \mu μ , define
W k R [ μ ] ( x ) = ∫ 0 R ( μ ( B t ( x ) ) t n − k ) 1 / k d t t , 0 < R < ∞ . \mathcal{W}_k^R[\mu](x)=\int_0^R\left(\frac{\mu(B_t(x))}{t^{n-k}}\right)^{1/k}\frac{dt}{t}, \qquad0<R<\infty. W k R [ μ ] ( x ) = ∫ 0 R ( t n − k μ ( B t ( x )) ) 1/ k t d t , 0 < R < ∞.
In the usual notation, this is W k / ( k + 1 ) , k + 1 R [ μ ] \mathbf{W}^{R}_{k/(k+1),k+1}[\mu] W k / ( k + 1 ) , k + 1 R [ μ ] .
Theorem 1.2. Let 2 ≤ k < n 2 \le k < n 2 ≤ k < n , let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be open, and let u ∈ Φ k ( Ω ) u \in\Phi_k(\Omega) u ∈ Φ k ( Ω ) . Set μ = F k [ u ] \mu= F_k[u] μ = F k [ u ] . If B R ( x 0 ) ⋐ Ω B_R(x_0) \Subset\Omega B R ( x 0 ) ⋐ Ω and x 0 x_0 x 0 is a Lebesgue point of D u Du D u , then
|Du(x_0)| \le\frac{C}{R}\fint_{B_R(x_0)} |u| + C\mathcal{W}_k^R[\mu](x_0), \tag*{(1)}
where C = C ( n , k ) C=C(n,k) C = C ( n , k ) .
Remark 1.3. The measure F k [ u ] F_k[u] F k [ u ] may contain a singular part, and u u u may be unbounded. Subtracting an affine function preserves both k k k -convexity and the Hessian measure. Thus, for ℓ ( x ) = a ⋅ ( x − x 0 ) + b \ell(x)=a\cdot(x-x_0)+b ℓ ( x ) = a ⋅ ( x − x 0 ) + b , applying (1.1) to u − ℓ u-\ell u − ℓ gives
|Du(x_0)-a| \le\frac{C}{R}\fint_{B_R(x_0)} |u-\ell| + C\mathcal{W}_k^R[\mu](x_0). \tag*{(2)}
Choosing b=\fint_{B_R(x_0)}u and using Poincaré’s inequality, we obtain
|Du(x_0)-a| \le C\fint_{B_R(x_0)} |Du-a| + C\mathcal{W}_k^R[\mu](x_0). \tag*{(3)}
For k = 1 k=1 k = 1 , the gradient estimate follows from linear potential theory, and W 1 R \mathcal{W}_1^R W 1 R is a truncated first-order Riesz potential, up to normalization. For k = n k=n k = n , convexity gives interior gradient bounds in terms of local averages.
Applying Hölder’s inequality to the potential in Theorem 1.2 gives the following local Lipschitz estimate.
Corollary 1.4. Let 2 ≤ k < n 2 \le k < n 2 ≤ k < n , u ∈ Φ k ( Ω ) u \in\Phi_k(\Omega) u ∈ Φ k ( Ω ) , and B R ( x 0 ) ⋐ Ω B_R(x_0) \Subset\Omega B R ( x 0 ) ⋐ Ω . Suppose that F k [ u ] = f F_k[u]=f F k [ u ] = f in B R ( x 0 ) B_R(x_0) B R ( x 0 ) , with f ∈ L q ( B R ( x 0 ) ) f \in L^q(B_R(x_0)) f ∈ L q ( B R ( x 0 )) for some n / k < q ≤ ∞ n/k<q\le\infty n / k < q ≤ ∞ . Set
α q = 1 − n k q > 0 and α ∞ = 1. \alpha_q=1-\frac{n}{kq}>0 \quad\text{and}\quad\alpha_\infty=1. α q = 1 − k q n > 0 and α ∞ = 1. Then, for every affine function ℓ \ell ℓ with D ℓ = a D\ell=a D ℓ = a ,
\|Du-a\|_{L^\infty(B_{R/2}(x_0))} \le\frac{C}{R}\fint_{B_R(x_0)} |u-\ell|\,dx+\frac{C}{\alpha_q}R^{\alpha_q}\|f\|_{L^q(B_R(x_0))}^{1/q}, \tag*{(4)}
where C = C ( n , k ) C=C(n,k) C = C ( n , k ) . For q = ∞ q=\infty q = ∞ , the second term on the right-hand side reduces to C R ∥ f ∥ L ∞ ( B R ( x 0 ) ) 1 / k CR\|f\|_{L^\infty(B_R(x_0))}^{1/k} C R ∥ f ∥ L ∞ ( B R ( x 0 )) 1/ k . In particular, if F k [ u ] = f F_k[u]=f F k [ u ] = f throughout Ω \Omega Ω , then
f ∈ L l o c q ( Ω ) , n / k < q ≤ ∞ ⟹ u ∈ W l o c 1 , ∞ ( Ω ) . (5) f\in L^q_{\mathrm{loc}}(\Omega),\quad n/k<q\le\infty\quad\Longrightarrow\quad u\in W^{1,\infty}_{\mathrm{loc}}(\Omega). \tag*{(5)} f ∈ L loc q ( Ω ) , n / k < q ≤ ∞ ⟹ u ∈ W loc 1 , ∞ ( Ω ) . (5)
If F k [ u ] = f F_k[u]=f F k [ u ] = f in Ω \Omega Ω , the critical Lorentz condition f ∈ L l o c n / k , 1 / k ( Ω ) f \in L_{\mathrm{loc}}^{n/k,1/k}(\Omega) f ∈ L loc n / k , 1/ k ( Ω ) gives local Lipschitz regularity by Proposition 4.1(iii). The example in Proposition 1 shows that the second Lorentz exponent 1 / k 1/k 1/ k is sharp. For max { 0 , 2 − n / k } < β < 1 \max\{0,2-n/k\}<\beta<1 max { 0 , 2 − n / k } < β < 1 , the function u ( x ) = ∣ x ∣ β u(x)=|x|^\beta u ( x ) = ∣ x ∣ β is bounded and k k k -convex in B 1 B_1 B 1 , while ∣ D u ( x ) ∣ = β ∣ x ∣ β − 1 → ∞ |Du(x)|=\beta|x|^{\beta-1}\to\infty ∣ D u ( x ) ∣ = β ∣ x ∣ β − 1 → ∞ as x → 0 x\to0 x → 0 .
We next use Theorem 1.2 to obtain local Lipschitz regularity for equations with gradient growth. Related equations for smooth solutions were studied by Mohammed–Porru [10 ] .
Theorem 1.5. Let 2 ≤ k < n 2\le k<n 2 ≤ k < n , 0 ≤ Λ < ∞ 0\le\Lambda<\infty 0 ≤ Λ < ∞ , and 0 ≤ h ∈ L l o c q ( Ω ) 0\le h\in L_{\mathrm{loc}}^q(\Omega) 0 ≤ h ∈ L loc q ( Ω ) for some n / k < q ≤ ∞ n/k<q\le\infty n / k < q ≤ ∞ . Suppose that u ∈ Φ k ( Ω ) u\in\Phi_k(\Omega) u ∈ Φ k ( Ω ) satisfies
F k [ u ] = f in Ω , 0 ≤ f ≤ Λ ∣ D u ∣ p + h almost everywhere in Ω , (6) F_k[u]=f\quad\text{in }\Omega,\qquad0\le f\le\Lambda|Du|^p+h\quad\text{almost everywhere in }\Omega, \tag*{(6)} F k [ u ] = f in Ω , 0 ≤ f ≤ Λ∣ D u ∣ p + h almost everywhere in Ω , (6) with
0 < p < p ∗ : = n k n − k . (7) 0<p<p_*:=\frac{nk}{n-k}. \tag*{(7)} 0 < p < p ∗ := n − k nk . (7) Then u ∈ W l o c 1 , ∞ ( Ω ) u\in W_{\mathrm{loc}}^{1,\infty}(\Omega) u ∈ W loc 1 , ∞ ( Ω ) .
In particular, every weak k k k -convex solution of F k [ u ] = ∣ D u ∣ p F_k[u]=|Du|^p F k [ u ] = ∣ D u ∣ p in Ω \Omega Ω is locally Lipschitz whenever p p p satisfies (1.7).
The proof of Theorem 1.5 is given in Section 3. The restriction p < p ∗ p<p_* p < p ∗ is sharp among weak k k k -convex solutions: Proposition 3.1 constructs a solution of F k [ u ] = ∣ D u ∣ p ∗ F_k[u]=|Du|^{p_*} F k [ u ] = ∣ D u ∣ p ∗ in B 1 B_1 B 1 whose gradient belongs to L p ∗ ( B 1 ) L^{p_*}(B_1) L p ∗ ( B 1 ) but is unbounded. The example is continuous when n < 2 k n<2k n < 2 k and unbounded when n ≥ 2 k n\ge2k n ≥ 2 k . In the latter range, the endpoint case for locally bounded solutions falls outside this construction.
Section 4 also gives estimates for the oscillation of u u u , including a log-Lipschitz modulus under a uniform ball-mass bound μ ( B r ( x ) ) ≤ M r n − k \mu(B_r(x))\le Mr^{n-k} μ ( B r ( x )) ≤ M r n − k . The gradient can be discontinuous: for k ≥ 2 k\ge2 k ≥ 2 , the function u ( x ) = ∣ x 1 ∣ u(x)=|x_1| u ( x ) = ∣ x 1 ∣ satisfies F k [ u ] = 0 F_k[u]=0 F k [ u ] = 0 . Gradient continuity is available under stronger hypotheses. Caffarelli [1 ] proved interior C 1 , α C^{1,\alpha} C 1 , α regularity for strictly convex Alexandrov solutions of the Monge–Ampère equation with density bounded between positive constants. Chen–Wang–Wu [2 Theorem 1.4] obtained C l o c 1 , α C_{\mathrm{loc}}^{1,\alpha} C loc 1 , α regularity for convex weak k k k -Hessian solutions in a bounded convex domain, with zero or affine boundary data and 0 < c 0 ≤ f ≤ c 1 0<c_0\le f\le c_1 0 < c 0 ≤ f ≤ c 1 .
The appendix gives examples with vanishing critical Morrey density and gradients outside B M O l o c \mathrm{BMO}_{\mathrm{loc}} BMO loc . For k ≥ 3 k\ge3 k ≥ 3 , gradient BMO can also fail for densities in L n / k L^{n/k} L n / k .
Note added. The pointwise gradient potential estimates in this paper and in He [6 ] were obtained independently. In the common range 2 ≤ k < n 2\le k<n 2 ≤ k < n , the Wolff-potential estimate in [6 Theorem 1.1] corresponds to (1.3) with a = 0 a=0 a = 0 , up to fixed changes of the radii. Both proofs use a Lipschitz stability estimate with coefficient one for k k k -harmonic functions.
Proof strategy and organization
The proof of Theorem 1.2 begins with a two-point maximum argument based on the negative-direction Newton tensor inequality [13 (3.5)] . For v ∈ Φ k ( B r ) ∩ C ( B r ‾ ) v\in\Phi_k(B_r)\cap C(\overline{B_r}) v ∈ Φ k ( B r ) ∩ C ( B r ) with F k [ v ] = 0 F_k[v]=0 F k [ v ] = 0 and L , δ ≥ 0 L,\delta\ge0 L , δ ≥ 0 , we prove
∣ v ( x ) − v ( y ) ∣ ≤ L ∣ x − y ∣ + δ ( x , y ∈ B r ) ⟹ Lip ( v , B r / 2 ) ≤ L + C ( n ) δ r . |v(x)-v(y)|\le L|x-y|+\delta\quad(x,y\in B_r)
\quad\Longrightarrow\quad
\operatorname{Lip}(v,B_{r/2})\le L+C(n)\frac{\delta}{r}. ∣ v ( x ) − v ( y ) ∣ ≤ L ∣ x − y ∣ + δ ( x , y ∈ B r ) ⟹ Lip ( v , B r /2 ) ≤ L + C ( n ) r δ .
An integral estimate gives the initial Lipschitz bound. We then restore the Hessian measure on successive annuli and apply this stability estimate to the comparison functions. The coefficient of L L L remains one, so the errors add over scales. The zero-boundary estimate of Phuc–Verbitsky [12 ] controls these errors when n > 2 k n>2k n > 2 k ; dimensional extension handles n ≤ 2 k n\le2k n ≤ 2 k . An averaging argument gives the estimate for weak solutions. Remark 2.10 explains the role of the stability estimate in this summation.
Section 2 contains the proof of the pointwise estimate and Corollary 1.4. Section 3 treats equations with gradient growth, and Section 4 develops the regularity consequences. The appendix contains sharpness constructions and examples of gradient BMO failure.
The pointwise gradient estimate
In this section we prove Theorem 1.2 and Corollary 1.4. We begin with the basic properties of weak k k k -convex functions and the potential estimates needed below.
Preliminaries
We write B r = B r ( 0 ) B_r = B_r(0) B r = B r ( 0 ) and ω m = ∣ B 1 m ∣ \omega_m = |B_1^m| ω m = ∣ B 1 m ∣ . Unless explicitly fixed, constants may vary from line to line, with their dependence indicated in each statement. Norms of weak gradients are understood in the essential sense. Inequalities between measures are understood on all Borel sets. Measures appearing in potentials are extended by zero outside their domains when necessary.
The approximation and comparison arguments below use the following properties of weak k k k -convex functions.
Lemma 2.1. [14 , 15 , 16 ] Let 2 ≤ k < n 2 \le k < n 2 ≤ k < n and let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be a bounded domain.
(i) We have
Φ k ( Ω ) ⊂ W l o c 1 , q ( Ω ) , 1 ≤ q < n k n − k . (8) \Phi_k(\Omega) \subset W_{\mathrm{loc}}^{1,q}(\Omega), \qquad1 \le q < \frac{nk}{n-k}. \tag*{(8)} Φ k ( Ω ) ⊂ W loc 1 , q ( Ω ) , 1 ≤ q < n − k nk . (8) (ii) For u ∈ Φ k ( Ω ) u \in\Phi_k(\Omega) u ∈ Φ k ( Ω ) , its standard mollifications u ε u_\varepsilon u ε are smooth and k k k -convex on each subdomain Ω ′ ⋐ Ω \Omega' \Subset\Omega Ω ′ ⋐ Ω for sufficiently small ε > 0 \varepsilon> 0 ε > 0 , and u ε → u u_\varepsilon\to u u ε → u in L l o c 1 ( Ω ) L_{\mathrm{loc}}^1(\Omega) L loc 1 ( Ω ) . Moreover, if u m ∈ Φ k ( Ω ) u_m \in\Phi_k(\Omega) u m ∈ Φ k ( Ω ) and u m → u u_m \to u u m → u in L l o c 1 ( Ω ) L_{\mathrm{loc}}^1(\Omega) L loc 1 ( Ω ) , then u u u admits a unique representative in Φ k ( Ω ) \Phi_k(\Omega) Φ k ( Ω ) . For this representative,
F k [ u m ] → F k [ u ] locally weakly as Radon measures . (9) F_k[u_m] \to F_k[u] \quad\text{locally weakly as Radon measures}. \tag*{(9)} F k [ u m ] → F k [ u ] locally weakly as Radon measures . (9) (iii) If a , b ∈ Φ k ( Ω ) a,b \in\Phi_k(\Omega) a , b ∈ Φ k ( Ω ) , then a + b ∈ Φ k ( Ω ) a+b \in\Phi_k(\Omega) a + b ∈ Φ k ( Ω ) and
F k [ a + b ] ≥ F k [ a ] + F k [ b ] . F_k[a+b] \ge F_k[a] + F_k[b]. F k [ a + b ] ≥ F k [ a ] + F k [ b ] . (iv) For every ball B ⊂ R n B \subset\mathbb{R}^n B ⊂ R n and a , b ∈ Φ k ( B ) ∩ C ( B ‾ ) a,b \in\Phi_k(B) \cap C(\overline{B}) a , b ∈ Φ k ( B ) ∩ C ( B ) ,
F k [ a ] ≥ F k [ b ] in B , a ≤ b on ∂ B ⟹ a ≤ b in B . \begin{aligned}
F_k[a] &\ge F_k[b] &&\text{in } B,\\
a &\le b &&\text{on } \partial B
\end{aligned}
\quad\Longrightarrow\quad a \le b \quad\text{in } B. F k [ a ] a ≥ F k [ b ] ≤ b in B , on ∂ B ⟹ a ≤ b in B .
The homogeneous comparison uses k k k -harmonic functions, which we define in the admissible viscosity sense.
Definition 2.2. [15 ] For 1 ≤ k ≤ n 1 \le k \le n 1 ≤ k ≤ n , a function u ∈ C ( Ω ) ∩ Φ k ( Ω ) u \in C(\Omega) \cap\Phi_k(\Omega) u ∈ C ( Ω ) ∩ Φ k ( Ω ) is called k k k -harmonic in Ω \Omega Ω if the following conditions hold at every x ∈ Ω x \in\Omega x ∈ Ω :
(1) Every C 2 C^2 C 2 , k k k -convex test function φ \varphi φ touching u u u from below at x x x satisfies
σ k ( D 2 φ ( x ) ) ≤ 0. \sigma_k(D^2\varphi(x)) \le0. σ k ( D 2 φ ( x )) ≤ 0. (2) Every C 2 C^2 C 2 test function φ \varphi φ touching u u u from above at x x x satisfies
σ k ( D 2 φ ( x ) ) ≥ 0. \sigma_k(D^2\varphi(x)) \ge0. σ k ( D 2 φ ( x )) ≥ 0.
For u ∈ C ( Ω ) ∩ Φ k ( Ω ) u \in C(\Omega) \cap\Phi_k(\Omega) u ∈ C ( Ω ) ∩ Φ k ( Ω ) , these conditions are equivalent to F k [ u ] = 0 F_k[u] = 0 F k [ u ] = 0 in Ω \Omega Ω ; see [15 ] . We construct the comparison functions by solving Dirichlet problems on balls.
Lemma 2.3. [13 , 15 ] Let 2 ≤ k < n 2 \le k < n 2 ≤ k < n and let B ⊂ R n B \subset\mathbb{R}^n B ⊂ R n be a ball. Suppose that φ ∈ Φ k ( B ) ∩ C ( B ‾ ) \varphi\in\Phi_k(B) \cap C(\overline{B}) φ ∈ Φ k ( B ) ∩ C ( B ) and 0 ≤ f ∈ L ∞ ( B ) 0 \le f \in L^\infty(B) 0 ≤ f ∈ L ∞ ( B ) . Then there is a unique v ∈ Φ k ( B ) ∩ C ( B ‾ ) v \in\Phi_k(B) \cap C(\overline{B}) v ∈ Φ k ( B ) ∩ C ( B ) satisfying
{ F k [ v ] = f in B , v = φ on ∂ B . \begin{cases}
F_k[v] = f & \text{in } B,\\
v = \varphi& \text{on } \partial B.
\end{cases} { F k [ v ] = f v = φ in B , on ∂ B . If f = 0 f = 0 f = 0 , then v v v is locally Lipschitz in B B B . If f ∈ C ∞ ( B ‾ ) f \in C^\infty(\overline{B}) f ∈ C ∞ ( B ) , φ ∣ ∂ B ∈ C ∞ ( ∂ B ) \varphi|_{\partial B} \in C^\infty(\partial B) φ ∣ ∂ B ∈ C ∞ ( ∂ B ) , and inf B f > 0 \inf_B f > 0 inf B f > 0 , then v v v is a classical solution.
To control the difference between successive comparison functions, we use a potential estimate for solutions with zero boundary data.
Lemma 2.4. ([9, 12]) Let 2 ≤ k < n / 2 2 \le k < n/2 2 ≤ k < n /2 and s > 0 s > 0 s > 0 . Let η \eta η be a nonnegative measure on B 2 s B_{2s} B 2 s with bounded density and supp η ⊂ B s \operatorname{supp}\eta\subset B_s supp η ⊂ B s . Suppose that z ∈ Φ k ( B 2 s ) ∩ C ( B ‾ 2 s ) z \in\Phi_k(B_{2s}) \cap C(\overline{B}_{2s}) z ∈ Φ k ( B 2 s ) ∩ C ( B 2 s ) solves
{ F k [ z ] = η in B 2 s , z = 0 on ∂ B 2 s . \begin{cases}
F_k[z] = \eta& \text{in } B_{2s},\\
z = 0 & \text{on } \partial B_{2s}.
\end{cases} { F k [ z ] = η z = 0 in B 2 s , on ∂ B 2 s . Then, for every x ∈ B 2 s x \in B_{2s} x ∈ B 2 s ,
0 ≤ − z ( x ) ≤ C ( n , k ) ∫ 0 8 s ( η ( B t ( x ) ) t n − 2 k ) 1 / k d t t . (10) 0 \le-z(x) \le C(n,k)\int_0^{8s}\left(\frac{\eta(B_t(x))}{t^{n-2k}}\right)^{1/k}\frac{dt}{t}. \tag*{(10)} 0 ≤ − z ( x ) ≤ C ( n , k ) ∫ 0 8 s ( t n − 2 k η ( B t ( x )) ) 1/ k t d t . (10) Here η \eta η is extended by zero outside B 2 s B_{2s} B 2 s .
In the passage to weak solutions, the weak convergence in (9) will be combined with the following L 1 L^1 L 1 convergence of truncated potentials.
Lemma 2.5. Let 2 ≤ k < n 2 \le k < n 2 ≤ k < n and let G ⊂ R n G \subset\mathbb{R}^n G ⊂ R n be open. Suppose that ν m \nu_m ν m and ν \nu ν are nonnegative Radon measures on G G G and that ν m → ν \nu_m \to\nu ν m → ν locally weakly in G G G . Let E ⊂ G E \subset G E ⊂ G be bounded and measurable, and let a > 0 a > 0 a > 0 . Assume that there is a compact set K ⊂ G K \subset G K ⊂ G such that B ‾ a ( y ) ⊂ K \overline{B}_a(y) \subset K B a ( y ) ⊂ K for every y ∈ E y \in E y ∈ E . Then
W k a [ ν m ] → W k a [ ν ] in L 1 ( E ) . \mathcal{W}_k^a[\nu_m] \to\mathcal{W}_k^a[\nu] \quad\text{in } L^1(E). W k a [ ν m ] → W k a [ ν ] in L 1 ( E ) .
Proof. Choose a compact neighbourhood of K K K contained in G G G . Local weak convergence gives a common bound M M M for the masses of ν m \nu_m ν m and ν \nu ν on this neighbourhood. Hölder’s inequality and Fubini’s theorem give, for 0 < ϵ < a 0 < \epsilon< a 0 < ϵ < a ,
sup m ∫ E W k ϵ [ ν m ] ( x ) d x ≤ C ∣ E ∣ 1 − 1 / k M 1 / k ϵ , \sup_m \int_E \mathcal{W}_k^\epsilon[\nu_m](x)\,dx \le C|E|^{1-1/k}M^{1/k}\epsilon, m sup ∫ E W k ϵ [ ν m ] ( x ) d x ≤ C ∣ E ∣ 1 − 1/ k M 1/ k ϵ ,
and the same estimate holds for ν \nu ν .
For each x ∈ E x \in E x ∈ E , weak convergence gives ν m ( B t ( x ) ) → ν ( B t ( x ) ) \nu_m(B_t(x)) \to\nu(B_t(x)) ν m ( B t ( x )) → ν ( B t ( x )) for all but countably many t ∈ [ ϵ , a ] t \in[\epsilon,a] t ∈ [ ϵ , a ] , since only countably many spheres ∂ B t ( x ) \partial B_t(x) ∂ B t ( x ) have positive ν \nu ν -measure. Dominated convergence on E × [ ϵ , a ] E \times[\epsilon,a] E × [ ϵ , a ] , followed by ϵ → 0 \epsilon\to0 ϵ → 0 , proves the claim. □
We also need to control an average of the potential by its value at the center, with a larger truncation radius.
Lemma 2.6. Let 2 ≤ k < n 2 \le k < n 2 ≤ k < n , and let ν \nu ν be a nonnegative Radon measure defined in a neighbourhood of B R ( x 0 ) B_R(x_0) B R ( x 0 ) . For every 0 < ρ < R / 4 0 < \rho< R/4 0 < ρ < R /4 ,
∫ B ρ ( x 0 ) W k R / 2 [ ν ] ( y ) d y ≤ C ( n , k ) W k R [ ν ] ( x 0 ) . \int_{B_\rho(x_0)} \mathcal{W}_k^{R/2}[\nu](y)\,dy \le C(n,k)\mathcal{W}_k^R[\nu](x_0). ∫ B ρ ( x 0 ) W k R /2 [ ν ] ( y ) d y ≤ C ( n , k ) W k R [ ν ] ( x 0 ) .
Proof. Split the potential integral at t = ρ t = \rho t = ρ . Concavity and Fubini’s theorem give
∫ B ρ ( x 0 ) W k ρ [ ν ] ( y ) d y ≤ C ( ν ( B 2 ρ ( x 0 ) ) ρ n − k ) 1 / k . \int_{B_\rho(x_0)} \mathcal{W}_k^\rho[\nu](y)\,dy \le C\left(\frac{\nu(B_{2\rho}(x_0))}{\rho^{n-k}}\right)^{1/k}. ∫ B ρ ( x 0 ) W k ρ [ ν ] ( y ) d y ≤ C ( ρ n − k ν ( B 2 ρ ( x 0 )) ) 1/ k .
The right-hand side is bounded by a constant times the contribution of [ 2 ρ , 4 ρ ] [2\rho,4\rho] [ 2 ρ , 4 ρ ] to W k R [ ν ] ( x 0 ) \mathcal{W}_k^R[\nu](x_0) W k R [ ν ] ( x 0 ) .
For ρ ≤ t ≤ R / 2 \rho\le t \le R/2 ρ ≤ t ≤ R /2 and y ∈ B ρ ( x 0 ) y \in B_\rho(x_0) y ∈ B ρ ( x 0 ) , we have B t ( y ) ⊂ B 2 t ( x 0 ) B_t(y) \subset B_{2t}(x_0) B t ( y ) ⊂ B 2 t ( x 0 ) . The change of variables t ↦ 2 t t \mapsto2t t ↦ 2 t therefore bounds the remaining part by C W k R [ ν ] ( x 0 ) C\mathcal{W}_k^R[\nu](x_0) C W k R [ ν ] ( x 0 ) . □
Lipschitz stability for k k k -harmonic functions
The homogeneous comparison rests on an algebraic perturbation in a direction where the Hessian is nonpositive. We first record this ingredient and then apply it in a two-point maximum argument. The cone Γ k \Gamma_k Γ k is convex and invariant under orthogonal conjugation and addition of nonnegative matrices. The function σ k 1 / k \sigma_k^{1/k} σ k 1/ k is concave on this cone. If A ∈ Γ k A \in\Gamma_k A ∈ Γ k , its restriction to a hyperplane belongs to Γ k − 1 \Gamma_{k-1} Γ k − 1 . The Newton tensors
T j − 1 ( A ) = D A σ j ( A ) , 1 ≤ j ≤ k , T_{j-1}(A) = D_A\sigma_j(A), \qquad1 \le j \le k, T j − 1 ( A ) = D A σ j ( A ) , 1 ≤ j ≤ k ,
are positive definite and satisfy
tr T j − 1 ( A ) = ( n − j + 1 ) σ j − 1 ( A ) . (11) \operatorname{tr} T_{j-1}(A) = (n-j+1)\sigma_{j-1}(A). \tag*{(11)} tr T j − 1 ( A ) = ( n − j + 1 ) σ j − 1 ( A ) . (11)
The corresponding non-strict statements hold on the closed cones. These algebraic properties are recalled in [17 Section 2] .
The negative-direction estimate (12) follows the Newton tensor argument in [13 Section 3] , applied also to lower-order operators. We prove the form needed here and derive the perturbation conclusion (13) .
Lemma 2.7. Let 2 ≤ k < n 2 \le k < n 2 ≤ k < n , A ∈ Γ k A \in\Gamma_k A ∈ Γ k and ∣ e ∣ = 1 |e| = 1 ∣ e ∣ = 1 . If A [ e , e ] ≤ 0 A[e,e] \le0 A [ e , e ] ≤ 0 , then
T j − 1 ( A ) [ e , e ] ≥ σ j − 1 ( A ) , 1 ≤ j ≤ k . (12) T_{j-1}(A)[e,e] \ge\sigma_{j-1}(A), \qquad1 \le j \le k. \tag*{(12)} T j − 1 ( A ) [ e , e ] ≥ σ j − 1 ( A ) , 1 ≤ j ≤ k . (12) Consequently, if b > 0 b > 0 b > 0 and A [ e , e ] + n b < 0 A[e,e] + nb < 0 A [ e , e ] + nb < 0 , then, with
V = b ( ( n + 1 ) e ⊗ e − I ) , V = b\left((n+1)e \otimes e - I\right), V = b ( ( n + 1 ) e ⊗ e − I ) , we have
A + V ∈ Γ k , σ k ( A + V ) > σ k ( A ) . (13) A + V \in\Gamma_k,\qquad\sigma_k(A+V) > \sigma_k(A). \tag*{(13)} A + V ∈ Γ k , σ k ( A + V ) > σ k ( A ) . (13)
Proof. Choose the first coordinate in the direction e e e and write
A = ( a ξ T ξ B ) , a ≤ 0. A = \begin{pmatrix} a & \xi^{T} \\ \xi& B \end{pmatrix}, \qquad a \le0. A = ( a ξ ξ T B ) , a ≤ 0.
Then B ∈ Γ k − 1 B \in\Gamma_{k-1} B ∈ Γ k − 1 and T j − 1 ( A ) [ e , e ] = σ j − 1 ( B ) T_{j-1}(A)[e,e] = \sigma_{j-1}(B) T j − 1 ( A ) [ e , e ] = σ j − 1 ( B ) . For j ≥ 3 j \ge3 j ≥ 3 , expansion in principal minors gives
σ j − 1 ( A ) = σ j − 1 ( B ) + a σ j − 2 ( B ) − ξ T T j − 3 ( B ) ξ ≤ σ j − 1 ( B ) . \sigma_{j-1}(A) = \sigma_{j-1}(B) + a\sigma_{j-2}(B) - \xi^{T}T_{j-3}(B)\xi\le\sigma_{j-1}(B). σ j − 1 ( A ) = σ j − 1 ( B ) + a σ j − 2 ( B ) − ξ T T j − 3 ( B ) ξ ≤ σ j − 1 ( B ) .
The cases j = 1 , 2 j = 1,2 j = 1 , 2 follow from T 0 = I T_0 = I T 0 = I and tr A = a + tr B \operatorname{tr} A = a + \operatorname{tr} B tr A = a + tr B , respectively.
Set A t = A + t V A_t = A + tV A t = A + t V , 0 ≤ t ≤ 1 0 \le t \le1 0 ≤ t ≤ 1 . Since A t [ e , e ] = A [ e , e ] + n b t < 0 A_t[e,e] = A[e,e] + nbt < 0 A t [ e , e ] = A [ e , e ] + nb t < 0 , (11) and (12) imply, as long as A t ∈ Γ k A_t \in\Gamma_k A t ∈ Γ k ,
d d t σ j ( A t ) = b ( ( n + 1 ) T j − 1 ( A t ) [ e , e ] − tr T j − 1 ( A t ) ) ≥ j b σ j − 1 ( A t ) > 0 , 1 ≤ j ≤ k . (14) \frac{d}{dt}\sigma_j(A_t) = b\left((n+1)T_{j-1}(A_t)[e,e] - \operatorname{tr}T_{j-1}(A_t)\right) \ge jb\sigma_{j-1}(A_t) > 0,\qquad1 \le j \le k. \tag*{(14)} d t d σ j ( A t ) = b ( ( n + 1 ) T j − 1 ( A t ) [ e , e ] − tr T j − 1 ( A t ) ) ≥ j b σ j − 1 ( A t ) > 0 , 1 ≤ j ≤ k . (14)
At a first exit from Γ k \Gamma_k Γ k , all the σ j \sigma_j σ j would still be at least their strictly positive initial values, a contradiction. Thus the entire segment lies in Γ k \Gamma_k Γ k , and integrating (14) with j = k j = k j = k proves (13) . □
The interior Lipschitz estimate for k k k -harmonic functions follows from [13 Theorem 3.1] by approximation. For the iteration below, we need the more precise estimate (16) , in which the coefficient of L L L is one. This allows us to sum the errors from successive comparisons, as in (2.29).
Proposition 2.8. Let 2 ≤ k < n 2 \le k < n 2 ≤ k < n and let v ∈ Φ k ( B r ) ∩ C ( B r ‾ ) v \in\Phi_k(B_r) \cap C(\overline{B_r}) v ∈ Φ k ( B r ) ∩ C ( B r ) be k k k -harmonic. There is a constant C 0 = C 0 ( n ) C_0 = C_0(n) C 0 = C 0 ( n ) such that v v v is Lipschitz in B r / 2 B_{r/2} B r /2 ; more precisely,
Lip ( v ; B r / 2 ) ≤ C 0 r osc B r v . \operatorname{Lip}(v;B_{r/2}) \le\frac{C_0}{r}\operatorname{osc}_{B_r} v. Lip ( v ; B r /2 ) ≤ r C 0 osc B r v . Moreover, whenever L , δ ≥ 0 L,\delta\ge0 L , δ ≥ 0 satisfy
∣ v ( x ) − v ( y ) ∣ ≤ L ∣ x − y ∣ + δ ( x , y ∈ B r ) , (15) |v(x)-v(y)| \le L|x-y|+\delta\qquad(x,y \in B_r), \tag*{(15)} ∣ v ( x ) − v ( y ) ∣ ≤ L ∣ x − y ∣ + δ ( x , y ∈ B r ) , (15) one has the sharper estimate
Lip ( v ; B r / 2 ) ≤ L + C 0 δ r . (16) \operatorname{Lip}(v;B_{r/2}) \le L + C_0\frac{\delta}{r}. \tag*{(16)} Lip ( v ; B r /2 ) ≤ L + C 0 r δ . (16) In particular, if w w w is Lipschitz in B r B_r B r and 0 ≤ w − v ≤ δ 0 \le w-v \le\delta 0 ≤ w − v ≤ δ , then
Lip ( v ; B r / 2 ) ≤ Lip ( w ; B r ) + C 0 δ r . (17) \operatorname{Lip}(v;B_{r/2}) \le\operatorname{Lip}(w;B_r) + C_0\frac{\delta}{r}. \tag*{(17)} Lip ( v ; B r /2 ) ≤ Lip ( w ; B r ) + C 0 r δ . (17) The interior Lipschitz estimate follows from (16) by taking L = 0 L = 0 L = 0 and δ = osc B r v \delta= \operatorname{osc}_{B_r} v δ = osc B r v .
Proof. Suppose first that v v v is smooth and admissible, with σ k ( D 2 v ) = c > 0 \sigma_k(D^2v) = c > 0 σ k ( D 2 v ) = c > 0 in B r B_r B r , so that D 2 v ∈ Γ k D^2v \in\Gamma_k D 2 v ∈ Γ k . The case δ = 0 \delta= 0 δ = 0 is immediate from (15) . For δ > 0 \delta> 0 δ > 0 , fix x 0 ∈ B r / 2 x_0 \in B_{r/2} x 0 ∈ B r /2 and set
K = 32 ( n + 1 ) , ϕ ( s ) = L s + K δ ( 2 s r − s 2 r 2 ) . K = 32(n+1), \qquad\phi(s) = Ls + K\delta\left(\frac{2s}{r} - \frac{s^2}{r^2}\right). K = 32 ( n + 1 ) , ϕ ( s ) = L s + K δ ( r 2 s − r 2 s 2 ) .
for 0 ≤ s ≤ r 0 \le s \le r 0 ≤ s ≤ r . Then
ϕ ′ ( s ) = L + 2 K δ ( 1 r − s r 2 ) , ϕ ′ ′ ( s ) = − 2 K δ r 2 \phi'(s)=L+2K\delta\left(\frac{1}{r}-\frac{s}{r^2}\right),\qquad\phi''(s)=-\frac{2K\delta}{r^2} ϕ ′ ( s ) = L + 2 K δ ( r 1 − r 2 s ) , ϕ ′′ ( s ) = − r 2 2 K δ
and
ϕ ( s ) ≥ L s (18) \phi(s)\ge Ls \tag*{(18)} ϕ ( s ) ≥ L s (18)
On B ‾ r / 2 ( x 0 ) × B ‾ r / 2 ( x 0 ) \overline{B}_{r/2}(x_0)\times\overline{B}_{r/2}(x_0) B r /2 ( x 0 ) × B r /2 ( x 0 ) consider
Ψ ( x , y ) = v ( x ) − v ( y ) − ϕ ( ∣ x − y ∣ ) − 8 δ r 2 ( ∣ x − x 0 ∣ 2 + ∣ y − x 0 ∣ 2 ) . \Psi(x,y)=v(x)-v(y)-\phi(|x-y|)-\frac{8\delta}{r^2}\left(|x-x_0|^2+|y-x_0|^2\right). Ψ ( x , y ) = v ( x ) − v ( y ) − ϕ ( ∣ x − y ∣ ) − r 2 8 δ ( ∣ x − x 0 ∣ 2 + ∣ y − x 0 ∣ 2 ) .
By (15) and (18) ,
Ψ ( x , y ) ≤ − δ < 0 if x or y ∈ ∂ B r / 2 ( x 0 ) , (19) \Psi(x,y)\le-\delta<0\quad\text{if }x\text{ or }y\in\partial B_{r/2}(x_0), \tag*{(19)} Ψ ( x , y ) ≤ − δ < 0 if x or y ∈ ∂ B r /2 ( x 0 ) , (19)
whereas
Ψ ( x , x ) = − 16 δ r 2 ∣ x − x 0 ∣ 2 ≤ 0. (20) \Psi(x,x)=-\frac{16\delta}{r^2}|x-x_0|^2\le0. \tag*{(20)} Ψ ( x , x ) = − r 2 16 δ ∣ x − x 0 ∣ 2 ≤ 0. (20)
We claim that
∣ D v ( x 0 ) ∣ ≤ ϕ ′ ( 0 ) = L + 2 K δ r . (21) |Dv(x_0)|\le\phi'(0)=L+\frac{2K\delta}{r}. \tag*{(21)} ∣ D v ( x 0 ) ∣ ≤ ϕ ′ ( 0 ) = L + r 2 K δ . (21)
Otherwise, with e 0 = D v ( x 0 ) / ∣ D v ( x 0 ) ∣ e_0=Dv(x_0)/|Dv(x_0)| e 0 = D v ( x 0 ) /∣ D v ( x 0 ) ∣ , Taylor’s formula gives
Ψ ( x 0 + t e 0 , x 0 ) = t ( ∣ D v ( x 0 ) ∣ − ϕ ′ ( 0 ) ) + o ( t ) > 0 \Psi(x_0+te_0,x_0)=t\left(|Dv(x_0)|-\phi'(0)\right)+o(t)>0 Ψ ( x 0 + t e 0 , x 0 ) = t ( ∣ D v ( x 0 ) ∣ − ϕ ′ ( 0 ) ) + o ( t ) > 0
for sufficiently small t > 0 t>0 t > 0 . Thus (19) and (20) place a positive maximum at an interior pair ( x ˉ , y ˉ ) (\bar{x},\bar{y}) ( x ˉ , y ˉ ) with x ˉ ≠ y ˉ \bar{x}\ne\bar{y} x ˉ = y ˉ .
At this pair set
e = x ˉ − y ˉ ∣ x ˉ − y ˉ ∣ , A = D 2 v ( x ˉ ) , B = D 2 v ( y ˉ ) , e=\frac{\bar{x}-\bar{y}}{|\bar{x}-\bar{y}|},\qquad A=D^2v(\bar{x}),\qquad B=D^2v(\bar{y}), e = ∣ x ˉ − y ˉ ∣ x ˉ − y ˉ , A = D 2 v ( x ˉ ) , B = D 2 v ( y ˉ ) ,
and let
H = ϕ ′ ′ e ⊗ e + ϕ ′ ∣ x ˉ − y ˉ ∣ ( I − e ⊗ e ) , H=\phi''e\otimes e+\frac{\phi'}{|\bar{x}-\bar{y}|}(I-e\otimes e), H = ϕ ′′ e ⊗ e + ∣ x ˉ − y ˉ ∣ ϕ ′ ( I − e ⊗ e ) ,
with the derivatives of ϕ \phi ϕ evaluated at ∣ x ˉ − y ˉ ∣ |\bar{x}-\bar{y}| ∣ x ˉ − y ˉ ∣ . The second-derivative test yields
\begin{pmatrix} A & 0 \\ 0 & -B \end{pmatrix}
\le
\begin{pmatrix} H & -H \\ -H & H \end{pmatrix}
+\frac{16\delta}{r^2}I_{2n}. \tag*{(22)}\end{pmatrix}
Testing on ( e , 0 ) (e,0) ( e , 0 ) gives the negative direction
A [ e , e ] ≤ ϕ ′ ′ + 16 δ r 2 = ( − 2 K + 16 ) δ r 2 . (23) A[e,e]\le\phi''+\frac{16\delta}{r^2}=(-2K+16)\frac{\delta}{r^2}. \tag*{(23)} A [ e , e ] ≤ ϕ ′′ + r 2 16 δ = ( − 2 K + 16 ) r 2 δ . (23)
Next let Q = I − 2 e ⊗ e Q=I-2e\otimes e Q = I − 2 e ⊗ e be the orthogonal reflection across e ⊥ e^\perp e ⊥ . Since
z − Q z = 2 ( z ⋅ e ) e , ∀ z ∈ R n , z-Qz=2(z\cdot e)e,\quad\forall z\in\mathbb{R}^n, z − Q z = 2 ( z ⋅ e ) e , ∀ z ∈ R n ,
testing (22) on ( z , Q z ) (z,Qz) ( z , Q z ) removes the tangential part of H H H and gives
z T ( A − Q B Q ) z ≤ 4 ϕ ′ ′ ( z ⋅ e ) 2 + 32 δ r 2 ∣ z ∣ 2 . z^{T}(A-QBQ)z\le4\phi''(z\cdot e)^2+\frac{32\delta}{r^2}|z|^2. z T ( A − QB Q ) z ≤ 4 ϕ ′′ ( z ⋅ e ) 2 + r 2 32 δ ∣ z ∣ 2 .
Equivalently,
Q B Q ≥ A + 8 K δ r 2 e ⊗ e − 32 δ r 2 I . (24) QBQ\ge A+\frac{8K\delta}{r^2}e\otimes e-\frac{32\delta}{r^2}I. \tag*{(24)} QB Q ≥ A + r 2 8 K δ e ⊗ e − r 2 32 δ I . (24)
Set b = 32 δ / r 2 b=32\delta/r^2 b = 32 δ / r 2 and V = b ( ( n + 1 ) e ⊗ e − I ) V=b((n+1)e\otimes e-I) V = b (( n + 1 ) e ⊗ e − I ) . By (23) and the choice of K K K ,
A [ e , e ] + n b ≤ − ( 32 n + 48 ) δ r 2 < 0. A[e,e]+nb\le-(32n+48)\frac{\delta}{r^2}<0. A [ e , e ] + nb ≤ − ( 32 n + 48 ) r 2 δ < 0.
Lemma 2.7 therefore implies
A + V ∈ Γ k , σ k ( A + V ) > σ k ( A ) . (25) A+V\in\Gamma_k,\qquad\sigma_k(A+V)>\sigma_k(A). \tag*{(25)} A + V ∈ Γ k , σ k ( A + V ) > σ k ( A ) . (25)
Moreover, 2 K ≥ 8 ( n + 1 ) 2K\ge8(n+1) 2 K ≥ 8 ( n + 1 ) , so (24) gives
P : = Q B Q − ( A + V ) ≥ 0. P:=QBQ-(A+V)\ge0. P := QB Q − ( A + V ) ≥ 0.
The segment C t = A + V + t P C_t=A+V+tP C t = A + V + tP , 0 ≤ t ≤ 1 0\le t\le1 0 ≤ t ≤ 1 , remains in Γ k \Gamma_k Γ k , and d d t σ k ( C t ) = tr ( T k − 1 ( C t ) P ) ≥ 0 \frac{d}{dt}\sigma_k(C_t)=\operatorname{tr}(T_{k-1}(C_t)P)\ge0 d t d σ k ( C t ) = tr ( T k − 1 ( C t ) P ) ≥ 0 . Consequently,
σ k ( Q B Q ) ≥ σ k ( A + V ) . \sigma_k(QBQ)\ge\sigma_k(A+V). σ k ( QB Q ) ≥ σ k ( A + V ) .
Orthogonal invariance and (25) now give
c = σ k ( B ) = σ k ( Q B Q ) ≥ σ k ( A + V ) > σ k ( A ) = c , c=\sigma_k(B)=\sigma_k(QBQ)\ge\sigma_k(A+V)>\sigma_k(A)=c, c = σ k ( B ) = σ k ( QB Q ) ≥ σ k ( A + V ) > σ k ( A ) = c ,
a contradiction. Thus (21) holds for every x 0 ∈ B r / 2 x_0\in B_{r/2} x 0 ∈ B r /2 . Integrating along line segments gives
Lip ( v ; B r / 2 ) ≤ L + C 0 δ r . (26) \operatorname{Lip}(v;B_{r/2})\le L+C_0\frac{\delta}{r}. \tag*{(26)} Lip ( v ; B r /2 ) ≤ L + C 0 r δ . (26)
with the fixed choice C 0 = 2 K C_0=2K C 0 = 2 K , independent of c c c .
Next, a sequence of smooth solutions is used to approximate the homogeneous weak solution. Choose φ m ∈ C ∞ ( ∂ B r ) \varphi_m\in C^\infty(\partial B_r) φ m ∈ C ∞ ( ∂ B r ) with ϵ m : = ∥ φ m − v ∥ L ∞ ( ∂ B r ) → 0 \epsilon_m:=\|\varphi_m-v\|_{L^\infty(\partial B_r)}\to0 ϵ m := ∥ φ m − v ∥ L ∞ ( ∂ B r ) → 0 , and constants c m > 0 c_m>0 c m > 0 tending to zero. The classical Dirichlet theorem [15 Theorem 3.2] on the uniformly ( k − 1 ) (k-1) ( k − 1 ) -convex ball gives smooth admissible solutions
{ F k [ v m ] = c m in B r , v m = φ m on ∂ B r . \begin{cases}
F_k[v_m]=c_m & \text{in } B_r,\\
v_m=\varphi_m & \text{on } \partial B_r.
\end{cases} { F k [ v m ] = c m v m = φ m in B r , on ∂ B r .
Define
α m = ( c m ( k n ) ) 1 / k , q m ( x ) = α m 2 ( ∣ x ∣ 2 − r 2 ) . \alpha_m=\left(\frac{c_m}{\binom{k}{n}}\right)^{1/k},\qquad q_m(x)=\frac{\alpha_m}{2}\left(|x|^2-r^2\right). α m = ( ( n k ) c m ) 1/ k , q m ( x ) = 2 α m ( ∣ x ∣ 2 − r 2 ) .
Then
{ F k [ q m ] = c m in B r , q m = 0 on ∂ B r , \begin{cases}
F_k[q_m]=c_m & \text{in } B_r,\\
q_m=0 & \text{on } \partial B_r,
\end{cases} { F k [ q m ] = c m q m = 0 in B r , on ∂ B r ,
and q m ≤ 0 q_m\le0 q m ≤ 0 in B r B_r B r . By Lemma 2.1(iii), F k [ v + q m ] ≥ c m d x = F k [ v m ] F_k[v+q_m]\ge c_m\,dx=F_k[v_m] F k [ v + q m ] ≥ c m d x = F k [ v m ] , while v + q m − ϵ m ≤ φ m v+q_m-\epsilon_m\le\varphi_m v + q m − ϵ m ≤ φ m on ∂ B r \partial B_r ∂ B r . Comparison gives
v + q m − ϵ m ≤ v m in B r . v+q_m-\epsilon_m\le v_m\quad\text{in }B_r. v + q m − ϵ m ≤ v m in B r .
Conversely, F k [ v m ] ≥ F k [ v + ϵ m ] = 0 F_k[v_m]\ge F_k[v+\epsilon_m]=0 F k [ v m ] ≥ F k [ v + ϵ m ] = 0 and φ m ≤ v + ϵ m \varphi_m\le v+\epsilon_m φ m ≤ v + ϵ m on the boundary, so
v m ≤ v + ϵ m in B r . v_m\le v+\epsilon_m\quad\text{in }B_r. v m ≤ v + ϵ m in B r .
Thus
v + q m − ϵ m ≤ v m ≤ v + ϵ m , v+q_m-\epsilon_m\le v_m\le v+\epsilon_m, v + q m − ϵ m ≤ v m ≤ v + ϵ m ,
and, since ∥ q m ∥ ∞ = α m r 2 / 2 → 0 \|q_m\|_\infty=\alpha_mr^2/2\to0 ∥ q m ∥ ∞ = α m r 2 /2 → 0 ,
∥ v m − v ∥ L ∞ ( B r ) → 0. \|v_m-v\|_{L^\infty(B_r)}\to0. ∥ v m − v ∥ L ∞ ( B r ) → 0.
The approximants satisfy
∣ v m ( x ) − v m ( y ) ∣ ≤ L ∣ x − y ∣ + δ + 2 ∥ v m − v ∥ L ∞ ( B r ) . |v_m(x)-v_m(y)|\le L|x-y|+\delta+2\|v_m-v\|_{L^\infty(B_r)}. ∣ v m ( x ) − v m ( y ) ∣ ≤ L ∣ x − y ∣ + δ + 2∥ v m − v ∥ L ∞ ( B r ) .
Applying (26) and passing to the uniform limit proves (16) . Taking L = 0 L=0 L = 0 and δ = osc B r v \delta=\operatorname{osc}_{B_r}v δ = osc B r v gives the ordinary interior estimate. Finally, if 0 ≤ g : = w − v ≤ δ 0\le g:=w-v\le\delta 0 ≤ g := w − v ≤ δ , then ∣ g ( x ) − g ( y ) ∣ ≤ δ |g(x)-g(y)|\le\delta ∣ g ( x ) − g ( y ) ∣ ≤ δ , so (15) holds with L = Lip ( w ; B r ) L=\operatorname{Lip}(w;B_r) L = Lip ( w ; B r ) . This proves (17) . □
The following interpolation consequence of Proposition 2.8 supplies the initial term in the proof of Theorem 1.2; see also [15 Corollary 3.4] and [17 Lemma 4.1] .
Corollary 2.9. Let 2 ≤ k < n 2\le k<n 2 ≤ k < n . If h ∈ Φ k ( B r ) ∩ C ( B r ‾ ) h\in\Phi_k(B_r)\cap C(\overline{B_r}) h ∈ Φ k ( B r ) ∩ C ( B r ) is k k k -harmonic, then, for every a ∈ R a\in\mathbb{R} a ∈ R ,
\operatorname{Lip}(h;B_{r/4})\le\frac{C(n)}{r}\fint_{B_r}|h-a|. \tag*{(27)}
Proof. Replacing h ( x ) h(x) h ( x ) by h ( r x ) − a h(rx)-a h ( r x ) − a , it suffices to prove
Lip ( h ; B 1 / 4 ) ≤ C ( n ) ∫ B 1 ∣ h ∣ . (28) \operatorname{Lip}(h;B_{1/4}) \le C(n)\int_{B_1}|h|. \tag*{(28)} Lip ( h ; B 1/4 ) ≤ C ( n ) ∫ B 1 ∣ h ∣. (28)
With C 1 = 2 C 0 C_1=2C_0 C 1 = 2 C 0 , Proposition 2.8 gives
Lip ( h ; B t / 2 ( y ) ) ≤ C 1 t sup B t ( y ) ∣ h ∣ for all B t ( y ) ⊂ B 1 . (29) \operatorname{Lip}(h;B_{t/2}(y)) \le\frac{C_1}{t}\sup_{B_t(y)}|h| \quad\text{for all } B_t(y)\subset B_1. \tag*{(29)} Lip ( h ; B t /2 ( y )) ≤ t C 1 B t ( y ) sup ∣ h ∣ for all B t ( y ) ⊂ B 1 . (29)
Set M = max B 1 ‾ ( 1 − ∣ x ∣ ) n ∣ h ( x ) ∣ M=\max_{\overline{B_1}}(1-|x|)^n|h(x)| M = max B 1 ( 1 − ∣ x ∣ ) n ∣ h ( x ) ∣ . If M = 0 M=0 M = 0 , the conclusion is immediate. Otherwise the maximum occurs at some y ∈ B 1 y\in B_1 y ∈ B 1 . Maximality and 1 − ∣ x ∣ ≥ ( 1 − ∣ y ∣ ) / 2 1-|x|\ge(1-|y|)/2 1 − ∣ x ∣ ≥ ( 1 − ∣ y ∣ ) /2 on B ( y , ( 1 − ∣ y ∣ ) / 2 ) B(y,(1-|y|)/2) B ( y , ( 1 − ∣ y ∣ ) /2 ) imply
∣ h ( x ) ∣ ≤ 2 n ∣ h ( y ) ∣ , ∀ x ∈ B ( y , 1 − ∣ y ∣ 2 ) . |h(x)|\le2^n|h(y)|,\ \forall x\in B\left(y,\frac{1-|y|}{2}\right). ∣ h ( x ) ∣ ≤ 2 n ∣ h ( y ) ∣ , ∀ x ∈ B ( y , 2 1 − ∣ y ∣ ) .
Applying (29) with t = ( 1 − ∣ y ∣ ) / 2 t=(1-|y|)/2 t = ( 1 − ∣ y ∣ ) /2 and setting C 2 = 2 n + 1 C 1 C_2=2^{n+1}C_1 C 2 = 2 n + 1 C 1 yields
Lip ( h ; B ( y , 1 − ∣ y ∣ 4 ) ) ≤ C 2 ∣ h ( y ) ∣ 1 − ∣ y ∣ . \operatorname{Lip}\left(h;B\left(y,\frac{1-|y|}{4}\right)\right)\le\frac{C_2|h(y)|}{1-|y|}. Lip ( h ; B ( y , 4 1 − ∣ y ∣ ) ) ≤ 1 − ∣ y ∣ C 2 ∣ h ( y ) ∣ .
Choose τ = min { 1 / 4 , ( 2 C 2 ) − 1 } \tau=\min\{1/4,(2C_2)^{-1}\} τ = min { 1/4 , ( 2 C 2 ) − 1 } , so that C 2 τ ≤ 1 / 2 C_2\tau\le1/2 C 2 τ ≤ 1/2 . Then
∣ h ( x ) ∣ ≥ ∣ h ( y ) ∣ − ∣ h ( x ) − h ( y ) ∣ ≥ ∣ h ( y ) ∣ 2 , ∀ x ∈ B ( y , τ ( 1 − ∣ y ∣ ) ) , |h(x)|\ge|h(y)|-|h(x)-h(y)|\ge\frac{|h(y)|}{2},\ \forall x\in B(y,\tau(1-|y|)), ∣ h ( x ) ∣ ≥ ∣ h ( y ) ∣ − ∣ h ( x ) − h ( y ) ∣ ≥ 2 ∣ h ( y ) ∣ , ∀ x ∈ B ( y , τ ( 1 − ∣ y ∣ )) ,
and integration gives
M ≤ 2 ω n τ n ∫ B 1 ∣ h ∣ . M\le\frac{2}{\omega_n\tau^n}\int_{B_1}|h|. M ≤ ω n τ n 2 ∫ B 1 ∣ h ∣.
Since 1 − ∣ x ∣ ≥ 1 / 2 1-|x|\ge1/2 1 − ∣ x ∣ ≥ 1/2 on B 1 / 2 B_{1/2} B 1/2 ,
sup B 1 / 2 ∣ h ∣ ≤ 2 n M ≤ 2 n + 1 ω n τ n ∫ B 1 ∣ h ∣ . \sup_{B_{1/2}}|h|\le2^nM\le\frac{2^{n+1}}{\omega_n\tau^n}\int_{B_1}|h|. B 1/2 sup ∣ h ∣ ≤ 2 n M ≤ ω n τ n 2 n + 1 ∫ B 1 ∣ h ∣.
A final application of (29) with t = 1 / 2 t=1/2 t = 1/2 and y = 0 y=0 y = 0 proves (28) . Rescaling proves (27) . □
Proof of the main estimate and the L q L^q L q criterion
We now apply the preceding estimates to a sequence of comparisons on a fixed ball. Corollary 2.9 controls its first Lipschitz constant; Lemma 2.4 bounds the value error caused by one annulus; and Proposition 2.8 converts this error into a Lipschitz increment without multiplying the preceding bound. The proof first treats smooth functions in dimensions n > 2 k n>2k n > 2 k . A dimensional extension and an averaging argument then remove these restrictions.
Proof of Theorem 1.2. We prove (1.1); the affine-invariant form will then follow by subtraction of an affine function.
Step 1. Assume first that n > 2 k n>2k n > 2 k and that u ∈ C ∞ ∩ Φ k u\in C^\infty\cap\Phi_k u ∈ C ∞ ∩ Φ k in a neighborhood of B R ‾ \overline{B_R} B R ; strict admissibility is not required. By translation take x 0 = 0 x_0=0 x 0 = 0 , and set
f = σ k ( D 2 u ) , μ = f d x . f=\sigma_k(D^2u),\qquad\mu=f\,dx. f = σ k ( D 2 u ) , μ = f d x .
Lemma 2.3 provides continuous weak solutions for all auxiliary Dirichlet problems below, whose densities are bounded and whose boundary data are u u u or zero. Comparison applies, and absolute continuity ensures that spherical boundaries have zero mass.
By integration in polar coordinates, choose s ∈ [ R / 4 , R / 2 ] s\in[R/4,R/2] s ∈ [ R /4 , R /2 ] such that
∫ ∂ B s ∣ u ∣ ≤ C R ∫ B R ∣ u ∣ . (30) \int_{\partial B_s}|u|\le\frac{C}{R}\int_{B_R}|u|. \tag*{(30)} ∫ ∂ B s ∣ u ∣ ≤ R C ∫ B R ∣ u ∣. (30)
By the homogeneous Dirichlet theorem [15 Theorem 3.5] , let U 0 ∈ Φ k ( B s ) ∩ C ( B s ‾ ) U_0\in\Phi_k(B_s)\cap C(\overline{B_s}) U 0 ∈ Φ k ( B s ) ∩ C ( B s ) be the unique k k k -harmonic solution with boundary value u u u , and let H H H be the ordinary harmonic function with the same boundary value. Thus
{ F k [ U 0 ] = 0 in B s , U 0 = u on ∂ B s , { Δ H = 0 in B s , H = u on ∂ B s . \begin{cases}
F_k[U_0]=0 & \text{in } B_s,\\
U_0=u & \text{on } \partial B_s,
\end{cases}
\qquad
\begin{cases}
\Delta H=0 & \text{in } B_s,\\
H=u & \text{on } \partial B_s.
\end{cases} { F k [ U 0 ] = 0 U 0 = u in B s , on ∂ B s , { Δ H = 0 H = u in B s , on ∂ B s .
Comparison and the subharmonicity of U 0 U_0 U 0 give
u ≤ U 0 ≤ H in B s . u \le U_0 \le H \quad\text{in } B_s. u ≤ U 0 ≤ H in B s .
Since ∣ U 0 ∣ ≤ ∣ u ∣ + ∣ H ∣ |U_0| \le|u| + |H| ∣ U 0 ∣ ≤ ∣ u ∣ + ∣ H ∣ , and the Poisson formula gives
\fint_{B_s} |H| \le\frac{s}{n}\fint_{\partial B_s} |u|,
(2.23) yields
\fint_{B_s} |U_0| \le C\fint_{B_R} |u|.
By Corollary 2.9,
L_0 := \operatorname{Lip}(U_0;B_{s/4}) \le\frac{C}{R}\fint_{B_R} |u|. \tag*{(31)}
Step 2. Let r j = 4 − j s r_j=4^{-j}s r j = 4 − j s for j ≥ 0 j\ge0 j ≥ 0 . On the fixed ball B s B_s B s , let U j ∈ Φ k ( B s ) ∩ C ( B s ‾ ) U_j\in\Phi_k(B_s)\cap C(\overline{B_s}) U j ∈ Φ k ( B s ) ∩ C ( B s ) be the bounded-density weak Dirichlet solution
\begin{cases}
F_k[U_j]=\mu\mathbin{\vrule height 1.4ex depth -0.4ex width .07ex\setminus} (B_s\setminus B_{r_j}) & \text{in } B_s,\\
U_j=u & \text{on } \partial B_s. \tag*{(32)}
\end{cases}
The case j = 0 j=0 j = 0 agrees with the preceding definition. Since the measures increase with j j j and the boundary values agree, comparison gives
U j + 1 ≤ U j . U_{j+1}\le U_j. U j + 1 ≤ U j .
Comparison with u u u and the support of the measure in (2.25) also give
u ≤ U j + 1 ≤ U j , F k [ U j ] = 0 in B r j . (33) u\le U_{j+1}\le U_j,\qquad F_k[U_j]=0\quad\text{in }B_{r_j}. \tag*{(33)} u ≤ U j + 1 ≤ U j , F k [ U j ] = 0 in B r j . (33)
Take a sequence z j ∈ Φ k ( B 2 s ) ∩ C ( B 2 s ‾ ) z_j\in\Phi_k(B_{2s})\cap C(\overline{B_{2s}}) z j ∈ Φ k ( B 2 s ) ∩ C ( B 2 s ) satisfying the following equation:
\begin{cases}
F_k[z_j]=\mu\mathbin{\vrule height 1.4ex depth -0.4ex width .07ex\setminus}(B_{r_j}\setminus B_{r_j/4}) & \text{in } B_{2s},\\
z_j=0 & \text{on } \partial B_{2s}.
\end{cases}
Define
a j = ( F k [ z j ] ( B s ) r j n − k ) 1 / k . a_j=\left(\frac{F_k[z_j](B_s)}{r_j^{n-k}}\right)^{1/k}. a j = ( r j n − k F k [ z j ] ( B s ) ) 1/ k .
Since z j ≤ 0 z_j\le0 z j ≤ 0 by Lemma 2.4, we have
U j + z j ≤ U j = u = U j + 1 on ∂ B s . U_j+z_j\le U_j=u=U_{j+1}\quad\text{on }\partial B_s. U j + z j ≤ U j = u = U j + 1 on ∂ B s .
Moreover, Lemma 2.1(iii) gives
F k [ U j + z j ] ≥ F k [ U j ] + F k [ z j ] = F k [ U j + 1 ] . F_k[U_j+z_j]\ge F_k[U_j]+F_k[z_j]=F_k[U_{j+1}]. F k [ U j + z j ] ≥ F k [ U j ] + F k [ z j ] = F k [ U j + 1 ] .
The comparison principle therefore yields
U j + z j ≤ U j + 1 in B s . U_j+z_j\le U_{j+1}\quad\text{in }B_s. U j + z j ≤ U j + 1 in B s .
Together with (2.26), this yields
0 ≤ U j − U j + 1 ≤ − z j . (34) 0\le U_j-U_{j+1}\le-z_j. \tag*{(34)} 0 ≤ U j − U j + 1 ≤ − z j . (34)
For x ∈ B r j / 8 x\in B_{r_j/8} x ∈ B r j /8 , the support of F k [ z j ] F_k[z_j] F k [ z j ] lies at distance at least r j / 8 r_j/8 r j /8 from x x x . Lemma 2.4 therefore gives
− z j ( x ) ≤ C ( F k [ z j ] ( B s ) ) 1 / k ∫ r j / 8 8 s t 2 − n / k t d t ≤ C ( F k [ z j ] ( B s ) ) 1 / k r j 2 − n / k ≤ C a n n r j a j . (35) -z_j(x)\le C\bigl(F_k[z_j](B_s)\bigr)^{1/k}\int_{r_j/8}^{8s}\frac{t^{2-n/k}}{t}\,dt
\le C\bigl(F_k[z_j](B_s)\bigr)^{1/k}r_j^{2-n/k}
\le C_{\mathrm{ann}}r_ja_j. \tag*{(35)} − z j ( x ) ≤ C ( F k [ z j ] ( B s ) ) 1/ k ∫ r j /8 8 s t t 2 − n / k d t ≤ C ( F k [ z j ] ( B s ) ) 1/ k r j 2 − n / k ≤ C ann r j a j . (35)
The assumption n > 2 k n>2k n > 2 k makes the integral controlled by its lower limit, uniformly in s / r j s/r_j s / r j ; fix C a n n = C a n n ( n , k ) C_{\mathrm{ann}}=C_{\mathrm{ann}}(n,k) C ann = C ann ( n , k ) so that the last bound holds for every j j j .
Since U j U_j U j is continuous on B r j ‾ \overline{B_{r_j}} B r j and k k k -harmonic in B r j B_{r_j} B r j , Proposition 2.8 gives
Lip ( U j ; B r j / 2 ) ≤ C r j osc B r j U j < ∞ . \operatorname{Lip}(U_j;B_{r_j/2})\le\frac{C}{r_j}\operatorname{osc}_{B_{r_j}}U_j<\infty. Lip ( U j ; B r j /2 ) ≤ r j C osc B r j U j < ∞.
In particular,
L j : = Lip ( U j ; B r j / 4 ) < ∞ . L_j := \operatorname{Lip}(U_j; B_{r_j/4}) < \infty. L j := Lip ( U j ; B r j /4 ) < ∞.
The function U j + 1 U_{j+1} U j + 1 is k k k -harmonic in B r j / 4 B_{r_j/4} B r j /4 . On B r j / 8 B_{r_j/8} B r j /8 , U j U_j U j has Lipschitz constant at most L j L_j L j , and (2.27)–(2.28) give
0 ≤ U j − U j + 1 ≤ C a n n r j a j . 0 \le U_j-U_{j+1} \le C_{\mathrm{ann}}r_j a_j. 0 ≤ U j − U j + 1 ≤ C ann r j a j .
Applying (2.10) on B r j / 8 B_{r_j/8} B r j /8 with v = U j + 1 v=U_{j+1} v = U j + 1 , w = U j w=U_j w = U j and δ = C a n n r j a j \delta=C_{\mathrm{ann}}r_j a_j δ = C ann r j a j , we obtain
Lip ( U j + 1 ; B r j / 16 ) ≤ Lip ( U j ; B r j / 8 ) + C 0 C a n n r j a j r j / 8 ≤ L j + 8 C 0 C a n n a j . \operatorname{Lip}(U_{j+1};B_{r_j/16}) \le\operatorname{Lip}(U_j;B_{r_j/8})+C_0\frac{C_{\mathrm{ann}}r_j a_j}{r_j/8}\le L_j+8C_0C_{\mathrm{ann}}a_j. Lip ( U j + 1 ; B r j /16 ) ≤ Lip ( U j ; B r j /8 ) + C 0 r j /8 C ann r j a j ≤ L j + 8 C 0 C ann a j .
Since r j + 1 / 4 = r j / 16 r_{j+1}/4=r_j/16 r j + 1 /4 = r j /16 , the left-hand side is exactly L j + 1 L_{j+1} L j + 1 . Hence
L j + 1 ≤ L j + 8 C 0 C a n n a j . L_{j+1}\le L_j+8C_0C_{\mathrm{ann}}a_j. L j + 1 ≤ L j + 8 C 0 C ann a j .
Iteration gives
{#eq:2.29}
L_m\le L_0+8C_0C_{\mathrm{ann}}\sum_{j=0}^{m-1}a_j.
The coefficient one in (2.10) prevents geometric weights in this sum; see Remark 2.10.
For r j ≤ t ≤ 2 r j r_j\le t\le2r_j r j ≤ t ≤ 2 r j , we have F k [ z j ] ( B s ) ≤ μ ( B t ) F_k[z_j](B_s)\le\mu(B_t) F k [ z j ] ( B s ) ≤ μ ( B t ) and t n − k ≤ 2 n − k r j n − k t^{n-k}\le2^{n-k}r_j^{n-k} t n − k ≤ 2 n − k r j n − k . Therefore
( μ ( B t ) t n − k ) 1 / k ≥ 2 − ( n − k ) / k a j . \left(\frac{\mu(B_t)}{t^{n-k}}\right)^{1/k}\ge2^{-(n-k)/k}a_j. ( t n − k μ ( B t ) ) 1/ k ≥ 2 − ( n − k ) / k a j .
The intervals [ r j , 2 r j ] [r_j,2r_j] [ r j , 2 r j ] are pairwise disjoint because r j + 1 = r j / 4 r_{j+1}=r_j/4 r j + 1 = r j /4 , and they lie in ( 0 , R ] (0,R] ( 0 , R ] since 2 s ≤ R 2s\le R 2 s ≤ R . Thus
{#eq:2.30}
\sum_{j=0}^{\infty}a_j\le C W_k^R[\mu](0).
Combining (2.24), (2.29), and (2.30) gives the uniform bound
{#eq:2.31}
L_m\le\frac{C}{R}\int_{B_R}|u|+C W_k^R[\mu](0).
Let F = ∥ f ∥ L ∞ ( B s ) F=\|f\|_{L^\infty(B_s)} F = ∥ f ∥ L ∞ ( B s ) . Let z ( j ) ∈ Φ k ( B 2 s ) ∩ C ( B 2 s ) z^{(j)}\in\Phi_k(B_{2s})\cap C(B_{2s}) z ( j ) ∈ Φ k ( B 2 s ) ∩ C ( B 2 s ) be the following solution
{ F k [ z ( j ) ] = μ ⌞ B r j in B 2 s , z ( j ) = 0 on ∂ B 2 s . \begin{cases}
F_k[z^{(j)}]=\mu\mathbin{\llcorner} B_{r_j} & \text{in }B_{2s},\\
z^{(j)}=0 & \text{on }\partial B_{2s}.
\end{cases} { F k [ z ( j ) ] = μ └ B r j z ( j ) = 0 in B 2 s , on ∂ B 2 s .
By Lemma 2.1(iii),
F k [ U j + z ( j ) ] ≥ F k [ u ] . F_k[U_j+z^{(j)}]\ge F_k[u]. F k [ U j + z ( j ) ] ≥ F k [ u ] .
Since U j + z ( j ) ≤ u U_j+z^{(j)}\le u U j + z ( j ) ≤ u on ∂ B s \partial B_s ∂ B s , comparison and (2.26) yield
0 ≤ U j − u ≤ − z ( j ) in B s . 0\le U_j-u\le-z^{(j)}\quad\text{in }B_s. 0 ≤ U j − u ≤ − z ( j ) in B s .
For every x x x and t > 0 t>0 t > 0 ,
μ ( B t ( x ) ∩ B r j ) ≤ C F min { t n , r j n } . \mu(B_t(x)\cap B_{r_j})\le CF\min\{t^n,r_j^n\}. μ ( B t ( x ) ∩ B r j ) ≤ C F min { t n , r j n } .
Using Lemma 2.4 and splitting its integral at r j r_j r j , we get
{#eq:2.32}
0\le U_j-u\le CF^{1/k}\left(\int_0^{r_j}t^2\frac{dt}{t}+r_j^{n/k}\int_{r_j}^{8s}t^{-n/k}\frac{dt}{t}\right)\le CF^{1/k}r_j^2.
Define
V j ( y ) = U j ( r j y / 4 ) − u ( 0 ) r j / 4 , y ∈ B 1 . V_j(y)=\frac{U_j(r_jy/4)-u(0)}{r_j/4},\qquad y\in B_1. V j ( y ) = r j /4 U j ( r j y /4 ) − u ( 0 ) , y ∈ B 1 .
By (2.32),
sup B 1 ∣ V j ( y ) − u ( r j y / 4 ) − u ( 0 ) r j / 4 ∣ ≤ C F 1 / k 4 r j 2 r j ≤ C F 1 / k r j → 0. \sup_{B_1}\left|V_j(y)-\frac{u(r_jy/4)-u(0)}{r_j/4}\right|\le CF^{1/k}\frac{4r_j^2}{r_j}\le CF^{1/k}r_j\to0. B 1 sup V j ( y ) − r j /4 u ( r j y /4 ) − u ( 0 ) ≤ C F 1/ k r j 4 r j 2 ≤ C F 1/ k r j → 0.
Taylor’s formula gives
u ( r j y / 4 ) − u ( 0 ) r j / 4 → D u ( 0 ) ⋅ y \frac{u(r_jy/4)-u(0)}{r_j/4}\to Du(0)\cdot y r j /4 u ( r j y /4 ) − u ( 0 ) → D u ( 0 ) ⋅ y
uniformly on B ‾ 1 \overline{B}_1 B 1 , and consequently
V j → V ( y ) : = D u ( 0 ) ⋅ y V_j\to V(y):=Du(0)\cdot y V j → V ( y ) := D u ( 0 ) ⋅ y
uniformly on B ‾ 1 \overline{B}_1 B 1 . Also,
Lip ( V j ; B 1 ) = Lip ( U j ; B r j / 4 ) = L j . \operatorname{Lip}(V_j;B_1)=\operatorname{Lip}(U_j;B_{r_j/4})=L_j. Lip ( V j ; B 1 ) = Lip ( U j ; B r j /4 ) = L j .
Thus for every y , z ∈ B 1 y,z\in B_1 y , z ∈ B 1 ,
∣ V ( y ) − V ( z ) ∣ = lim j → ∞ ∣ V j ( y ) − V j ( z ) ∣ ≤ ( lim inf j → ∞ L j ) ∣ y − z ∣ . |V(y)-V(z)|=\lim_{j\to\infty}|V_j(y)-V_j(z)|\leq\left(\liminf_{j\to\infty}L_j\right)|y-z|. ∣ V ( y ) − V ( z ) ∣ = j → ∞ lim ∣ V j ( y ) − V j ( z ) ∣ ≤ ( j → ∞ lim inf L j ) ∣ y − z ∣.
Since Lip ( V ; B 1 ) = ∣ D u ( 0 ) ∣ \operatorname{Lip}(V;B_1)=|Du(0)| Lip ( V ; B 1 ) = ∣ D u ( 0 ) ∣ ,
∣ D u ( 0 ) ∣ ≤ lim inf j → ∞ L j . |Du(0)|\leq\liminf_{j\to\infty}L_j. ∣ D u ( 0 ) ∣ ≤ j → ∞ lim inf L j .
Together with (2.31), this proves (1) for smooth functions when n > 2 k n>2k n > 2 k . The bound F F F enters only the identification of the limit, not (2.31).
Step 4. If n ≤ 2 k n\leq2k n ≤ 2 k , still assume that u ∈ C ∞ ∩ Φ k u\in C^\infty\cap\Phi_k u ∈ C ∞ ∩ Φ k in a neighborhood U U U of B ‾ R \overline{B}_R B R . Choose N = 2 k + 1 N=2k+1 N = 2 k + 1 and m = N − n m=N-n m = N − n , and set
u ~ ( x , z ) = u ( x ) , ( x , z ) ∈ U × R m . \widetilde{u}(x,z)=u(x),\qquad(x,z)\in U\times\mathbb{R}^m. u ( x , z ) = u ( x ) , ( x , z ) ∈ U × R m .
The Hessian has the block form
D x , z 2 u ~ = ( D x 2 u 0 0 0 m ) . D^2_{x,z}\widetilde{u}=
\begin{pmatrix}
D_x^2u & 0\\
0 & 0_m
\end{pmatrix}. D x , z 2 u = ( D x 2 u 0 0 0 m ) .
For 1 ≤ j ≤ k 1\leq j\leq k 1 ≤ j ≤ k ,
σ j ( N ) ( D x , z 2 u ~ ) = σ j ( n ) ( D x 2 u ) . \sigma_j^{(N)}(D^2_{x,z}\widetilde{u})=\sigma_j^{(n)}(D_x^2u). σ j ( N ) ( D x , z 2 u ) = σ j ( n ) ( D x 2 u ) .
In particular, D x 2 u ∈ Γ ‾ k n D_x^2u\in\overline{\Gamma}_k^n D x 2 u ∈ Γ k n implies D x , z 2 u ~ ∈ Γ ‾ k N D^2_{x,z}\widetilde{u}\in\overline{\Gamma}_k^N D x , z 2 u ∈ Γ k N , and
σ k ( N ) ( D 2 u ~ ) = σ k ( n ) ( D 2 u ) . \sigma_k^{(N)}(D^2\widetilde{u})=\sigma_k^{(n)}(D^2u). σ k ( N ) ( D 2 u ) = σ k ( n ) ( D 2 u ) .
Hence u ~ ∈ C ∞ ∩ Φ k \widetilde{u}\in C^\infty\cap\Phi_k u ∈ C ∞ ∩ Φ k on U × R m U\times\mathbb{R}^m U × R m , with N > 2 k N>2k N > 2 k . If μ = f d x \mu=f\,dx μ = f d x and μ ~ = f ( x ) d x d z \widetilde{\mu}=f(x)\,dx\,dz μ = f ( x ) d x d z , then
μ ~ ( B t N ( x 0 , 0 ) ) = ω m ∫ B t n ( x 0 ) ( t 2 − ∣ x − x 0 ∣ 2 ) m / 2 d μ ( x ) ≤ ω m t m μ ( B t n ( x 0 ) ) . \widetilde{\mu}(B_t^N(x_0,0))=\omega_m\int_{B_t^n(x_0)}(t^2-|x-x_0|^2)^{m/2}\,d\mu(x)\leq\omega_m t^m\mu(B_t^n(x_0)). μ ( B t N ( x 0 , 0 )) = ω m ∫ B t n ( x 0 ) ( t 2 − ∣ x − x 0 ∣ 2 ) m /2 d μ ( x ) ≤ ω m t m μ ( B t n ( x 0 )) .
The factor t m t^m t m cancels the dimension change in the potential, so the N N N -dimensional potential is bounded by C ( n , k ) W k R [ μ ] ( x 0 ) C(n,k)W_k^R[\mu](x_0) C ( n , k ) W k R [ μ ] ( x 0 ) . Integration in the extra variables also gives
\fint_{B_R^N(x_0,0)}|\widetilde{u}|\leq C(n,k)\fint_{B_R^n(x_0)}|u|,\qquad|D\widetilde{u}(x_0,0)|=|Du(x_0)|.
The N N N -dimensional estimate therefore proves (1) for n ≤ 2 k n\leq2k n ≤ 2 k .
Step 5. We first pass to the weak solution at a fixed averaging radius, then shrink the averaging ball. Let u ∈ Φ k ( Ω ) u\in\Phi_k(\Omega) u ∈ Φ k ( Ω ) and fix B R ( x 0 ) ⋐ Ω B_R(x_0)\Subset\Omega B R ( x 0 ) ⋐ Ω . If W k R [ F k [ u ] ] ( x 0 ) = ∞ W_k^R[F_k[u]](x_0)=\infty W k R [ F k [ u ]] ( x 0 ) = ∞ , the estimate is immediate. We may therefore assume that this potential is finite. Choose standard local convolutions u m = u ∗ ρ ϵ m u_m=u*\rho_{\epsilon_m} u m = u ∗ ρ ϵ m in a fixed neighborhood of B R ( x 0 ) B_R(x_0) B R ( x 0 ) . By (8) and the approximation facts in Subsection 2.1,
u m ∈ C ∞ ∩ Φ k , D u m → D u in L 1 ( B R ( x 0 ) ) , μ m : = F k [ u m ] ⇀ F k [ u ] = μ . u_m\in C^\infty\cap\Phi_k,\qquad Du_m\to Du\ \text{in }L^1(B_R(x_0)),\qquad\mu_m:=F_k[u_m]\rightharpoonup F_k[u]=\mu. u m ∈ C ∞ ∩ Φ k , D u m → D u in L 1 ( B R ( x 0 )) , μ m := F k [ u m ] ⇀ F k [ u ] = μ .
Here the Hessian measures converge locally weakly. For y ∈ B R / 4 ( x 0 ) y\in B_{R/4}(x_0) y ∈ B R /4 ( x 0 ) , apply the smooth estimate on B R / 2 ( y ) ⋐ B R ( x 0 ) B_{R/2}(y)\Subset B_R(x_0) B R /2 ( y ) ⋐ B R ( x 0 ) to obtain
|Du_m(y)|\leq\frac{C}{R}\fint_{B_R(x_0)}|u_m|+CW_k^{R/2}[\mu_m](y).
Fix 0 < ρ < R / 4 0<\rho<R/4 0 < ρ < R /4 , average over B ρ ( x 0 ) B_\rho(x_0) B ρ ( x 0 ) , and let m → ∞ m\to\infty m → ∞ . The W 1 , 1 W^{1,1} W 1 , 1 convergence and the L 1 L^1 L 1 convergence of truncated potentials from Lemma 2.5 give
\fint_{B_\rho(x_0)}|Du|\leq\frac{C}{R}\fint_{B_R(x_0)}|u|+C\fint_{B_\rho(x_0)}W_k^{R/2}[\mu](y)\,dy.
Lemma 2.6 bounds the last term by C W k R [ μ ] ( x 0 ) C W_k^R[\mu](x_0) C W k R [ μ ] ( x 0 ) . Letting ρ → 0 \rho\to0 ρ → 0 at a Lebesgue point of D u Du D u proves (1.1). Since subtracting an affine function preserves k k k -convexity and the Hessian measure, application to u − ℓ u-\ell u − ℓ also gives (1.2). □
Remark 2.10. The construction in Steps 1–3 restores the measure on successive annuli of the fixed Dirichlet problem (2.25). The zero-order potential estimate gives 0 ≤ U j − U j + 1 ≤ C r j a j 0 \le U_j-U_{j+1} \le C r_j a_j 0 ≤ U j − U j + 1 ≤ C r j a j on the inner ball. Proposition 2.8 converts this value error into
L j + 1 ≤ L j + C a j , L m ≤ L 0 + C ∑ j < m a j . L_{j+1} \le L_j + C a_j,\qquad L_m \le L_0 + C\sum_{j<m}a_j. L j + 1 ≤ L j + C a j , L m ≤ L 0 + C j < m ∑ a j .
Since the coefficient is one, (2.30) makes the increments summable.
In the p p p -Laplace setting, the standard gradient potential argument combines comparison with homogeneous solutions on shrinking balls and decay of a suitable gradient excess, that is, an averaged gradient oscillation. The recurrence is schematically E j + 1 ≤ θ E j + data error E_{j+1} \le\theta E_j + \text{data error} E j + 1 ≤ θ E j + data error , with 0 < θ < 1 0 < \theta< 1 0 < θ < 1 at suitably chosen scales, and summability of the excess controls successive gradient averages; see [5 , 8 ] . The linear-potential refinement in [7 Section 3] uses an induction that identifies and controls the nondegenerate regime. Here we propagate Lipschitz bounds without decay of gradient oscillations. For general k k k -harmonic functions with k ≥ 2 k \ge2 k ≥ 2 such decay is false: u ( x ) = ∣ x 1 ∣ u(x)=|x_1| u ( x ) = ∣ x 1 ∣ is an example. Thus the iteration controls the size of the gradient but gives no gradient continuity. The function average in (1.1) enters through the initialization (2.24), not through the iteration; affine subtraction and Poincaré’s inequality already give the gradient-average form (1.3).
A direct use of Trudinger’s estimates shows the same difficulty quantitatively. For the comparison functions and radii in Step 2, an ordinary homogeneous Lipschitz estimate gives a recurrence with a multiplier on L j L_j L j . Assume n > 2 k n>2k n > 2 k . Let C T C_T C T denote a constant in the centered homogeneous estimate
∣ D h ( y ) ∣ ≤ C T ρ osc B ρ ( y ) h |Dh(y)|\le\frac{C_T}{\rho}\operatorname{osc}_{B_\rho(y)}h ∣ D h ( y ) ∣ ≤ ρ C T osc B ρ ( y ) h
of [13 Theorem 3.1] for smooth homogeneous solutions, and use its Lipschitz consequence for continuous weak solutions by approximation. For y ∈ B r j / 16 y\in B_{r_j/16} y ∈ B r j /16 , the ball B r j / 16 ( y ) B_{r_j/16}(y) B r j /16 ( y ) lies in B r j / 8 B_{r_j/8} B r j /8 , so (2.27)–(2.28) give
osc B r j / 16 ( y ) U j + 1 ≤ r j 8 L j + K r j a j . \operatorname{osc}_{B_{r_j/16}(y)}U_{j+1}\le\frac{r_j}{8}L_j+K r_j a_j. osc B r j /16 ( y ) U j + 1 ≤ 8 r j L j + K r j a j .
with a fixed K = K ( n , k ) K=K(n,k) K = K ( n , k ) . Applying the centered estimate directly gives only
L j + 1 ≤ 2 C T L j + 16 C T K a j . L_{j+1}\le2C_TL_j+16C_TK a_j. L j + 1 ≤ 2 C T L j + 16 C T K a j .
For k ≥ 2 k\ge2 k ≥ 2 , smooth convex approximations of ( x 1 + a ) + (x_1+a)_+ ( x 1 + a ) + on B 1 B_1 B 1 , with 0 < a < 1 0<a<1 0 < a < 1 , have Hessians of rank at most one and are k k k -harmonic. Applying the centered estimate to these functions gives C T ≥ ( 1 + a ) − 1 C_T\ge(1+a)^{-1} C T ≥ ( 1 + a ) − 1 , hence C T ≥ 1 C_T\ge1 C T ≥ 1 as a → 0 a\to0 a → 0 . Iterating the preceding bound therefore produces geometric weights:
L m ≤ ( 2 C T ) m L 0 + 16 C T K ∑ j = 0 m − 1 ( 2 C T ) m − 1 − j a j . L_m\le(2C_T)^mL_0+16C_TK\sum_{j=0}^{m-1}(2C_T)^{m-1-j}a_j. L m ≤ ( 2 C T ) m L 0 + 16 C T K j = 0 ∑ m − 1 ( 2 C T ) m − 1 − j a j .
The summability in (2.30) does not give a bound uniform in m m m , because L j L_j L j is multiplied repeatedly. Trudinger’s Hölder estimate [13 Theorem 4.1] has α < 1 \alpha<1 α < 1 . Its proof uses the small factor ( σ / τ ) 1 − α (\sigma/\tau)^{1-\alpha} ( σ / τ ) 1 − α to absorb the homogeneous oscillation term; see [13 (4.5)–(4.9)] . This absorption is unavailable at α = 1 \alpha=1 α = 1 . The comparison (2.10) gives instead the coefficient-one estimate leading to the additive bound (2.29).
The first consequence of the pointwise estimate follows by bounding the potential through Hölder’s inequality. Its scale exponent is positive precisely when q > n / k q>n/k q > n / k .
Proof of Corollary 1.4. Write N = ∥ f ∥ L q ( B R ( x 0 ) ) N=\|f\|_{L^q(B_R(x_0))} N = ∥ f ∥ L q ( B R ( x 0 )) and ω n = ∣ B 1 ∣ \omega_n=|B_1| ω n = ∣ B 1 ∣ . For x ∈ B R / 2 ( x 0 ) x\in B_{R/2}(x_0) x ∈ B R /2 ( x 0 ) and 0 < t ≤ R / 2 0<t\le R/2 0 < t ≤ R /2 , Hölder’s inequality gives
F k [ u ] ( B t ( x ) ) = ∫ B t ( x ) f d y ≤ ω n 1 − 1 / q N t n ( 1 − 1 / q ) , F_k[u](B_t(x))=\int_{B_t(x)}f\,dy\le\omega_n^{1-1/q}Nt^{n(1-1/q)}, F k [ u ] ( B t ( x )) = ∫ B t ( x ) f d y ≤ ω n 1 − 1/ q N t n ( 1 − 1/ q ) ,
with the same interpretation when q = ∞ q=\infty q = ∞ . Consequently,
sup x ∈ B R / 2 ( x 0 ) W k R / 2 [ F k [ u ] ] ( x ) ≤ C ( n , k ) N 1 / k ∫ 0 R / 2 t − n / ( k q ) d t ≤ C ( n , k ) α q R α q N 1 / k . (36) \sup_{x\in B_{R/2}(x_0)} \mathcal{W}_k^{R/2}[F_k[u]](x) \le C(n,k)N^{1/k}\int_0^{R/2} t^{-n/(kq)}\,dt \le\frac{C(n,k)}{\alpha_q}R^{\alpha_q}N^{1/k}. \tag*{(36)} x ∈ B R /2 ( x 0 ) sup W k R /2 [ F k [ u ]] ( x ) ≤ C ( n , k ) N 1/ k ∫ 0 R /2 t − n / ( k q ) d t ≤ α q C ( n , k ) R α q N 1/ k . (36)
Apply (2) on B R / 2 ( x ) ⊂ B R ( x 0 ) B_{R/2}(x)\subset B_R(x_0) B R /2 ( x ) ⊂ B R ( x 0 ) at Lebesgue points x ∈ B R / 2 ( x 0 ) x\in B_{R/2}(x_0) x ∈ B R /2 ( x 0 ) of D u Du D u ; the average term is bounded by 2 n 2^n 2 n times the corresponding average on B R ( x 0 ) B_R(x_0) B R ( x 0 ) . Taking the essential supremum and using (36) proves (4) . Since u ∈ W l o c 1 , 1 ( Ω ) u\in W_{\mathrm{loc}}^{1,1}(\Omega) u ∈ W loc 1 , 1 ( Ω ) , the locally bounded weak gradient gives (5) and a locally Lipschitz representative. □
Equations with gradient growth
This section proves Theorem 1.5 and establishes the sharpness of its exponent in the full weak class. The basic Sobolev integrability of k k k -convex functions starts an iteration, which closes once the density exponent exceeds n / k n/k n / k .
Proof of Theorem 1.5. All estimates are applied on nested relatively compact balls. We first record the density mapping needed below:
f ∈ L l o c t ( Ω ) , 1 < t < n k ⟹ D u ∈ L l o c n k t n − k t ( Ω ) . (37) f\in L_{\mathrm{loc}}^t(\Omega),\qquad1<t<\frac{n}{k}\quad\Longrightarrow\quad Du\in L_{\mathrm{loc}}^{\frac{nkt}{n-kt}}(\Omega). \tag*{(37)} f ∈ L loc t ( Ω ) , 1 < t < k n ⟹ D u ∈ L loc n − k t nk t ( Ω ) . (37)
Indeed, restrict f f f to a fixed ball and apply (1) on balls of a fixed smaller radius centered at its interior points. The mean term is uniformly bounded, while the nonlinear potential mapping in [4 Proposition 2.8 and Theorem 3.1(i)] , with parameters k / ( k + 1 ) k/(k+1) k / ( k + 1 ) and k + 1 k+1 k + 1 , maps into L n k t / ( n − k t ) L^{nkt/(n-kt)} L nk t / ( n − k t ) . This gives (37) . By (8) , D u ∈ L l o c r Du\in L_{\mathrm{loc}}^r D u ∈ L loc r for 1 ≤ r < p ∗ 1\le r<p_* 1 ≤ r < p ∗ . Since p < p ∗ p<p_* p < p ∗ and n ( p ∗ − k ) / k = p ∗ n(p_*-k)/k=p_* n ( p ∗ − k ) / k = p ∗ , choose
max { 1 , p , n ( p − k ) k } < r 0 < p ∗ . (38) \max\left\{1,p,\frac{n(p-k)}{k}\right\}<r_0<p_*. \tag*{(38)} max { 1 , p , k n ( p − k ) } < r 0 < p ∗ . (38)
Our aim is to reach r > n p / k r>np/k r > n p / k , so that ∣ D u ∣ p |Du|^p ∣ D u ∣ p has density exponent r / p > n / k r/p>n/k r / p > n / k . If r 0 r_0 r 0 already exceeds this threshold, no iteration is needed. Otherwise we use the following improvement. Suppose D u ∈ L l o c r Du\in L_{\mathrm{loc}}^r D u ∈ L loc r with r ≥ r 0 r\ge r_0 r ≥ r 0 and r < n p / k r<np/k r < n p / k . Then 1 < r / p < n / k < q 1<r/p<n/k<q 1 < r / p < n / k < q , so (6) gives f ∈ L l o c r / p f\in L_{\mathrm{loc}}^{r/p} f ∈ L loc r / p . Applying (37) with t = r / p t=r/p t = r / p yields
D u ∈ L l o c r + , r + = n k r n p − k r , 1 r + = p k 1 r − 1 n . Du\in L_{\mathrm{loc}}^{r_+},\qquad r_+=\frac{nkr}{np-kr},\qquad\frac{1}{r_+}=\frac{p}{k}\frac{1}{r}-\frac{1}{n}. D u ∈ L loc r + , r + = n p − k r nk r , r + 1 = k p r 1 − n 1 .
Moreover,
r + > r ⟺ r > n ( p − k ) k , r_+>r\quad\Longleftrightarrow\quad r>\frac{n(p-k)}{k}, r + > r ⟺ r > k n ( p − k ) ,
which holds by (38) . The iteration reaches r ≥ n p / k r\ge np/k r ≥ n p / k in finitely many steps. Otherwise s j = 1 / r j > 0 s_j=1/r_j>0 s j = 1/ r j > 0 would satisfy s j + 1 = ( p / k ) s j − 1 / n s_{j+1}=(p/k)s_j-1/n s j + 1 = ( p / k ) s j − 1/ n indefinitely. If p ≤ k p\le k p ≤ k , each step decreases s j s_j s j by at least 1 / n 1/n 1/ n . If p > k p>k p > k , let s ∗ = k / [ n ( p − k ) ] s_* = k/[n(p-k)] s ∗ = k / [ n ( p − k )] ; then s j − s ∗ = ( p / k ) j ( s 0 − s ∗ ) s_j-s_*=(p/k)^j(s_0-s_*) s j − s ∗ = ( p / k ) j ( s 0 − s ∗ ) , with s 0 < s ∗ s_0<s_* s 0 < s ∗ by (38) . Both cases eventually give a negative s j s_j s j , a contradiction. Only finitely many improvements are therefore needed; the intermediate balls may all be chosen to contain a prescribed compact subset of Ω \Omega Ω .
If r = n p / k r=np/k r = n p / k , we have f ∈ L l o c n / k f\in L_{\mathrm{loc}}^{n/k} f ∈ L loc n / k . For any 1 < t < n / k 1<t<n/k 1 < t < n / k , (37) then gives D u ∈ L l o c n k t / ( n − k t ) Du\in L_{\mathrm{loc}}^{nkt/(n-kt)} D u ∈ L loc nk t / ( n − k t ) . Choosing t t t sufficiently close to n / k n/k n / k makes this exponent exceed n p / k np/k n p / k ; relabel the larger exponent as r r r . Thus in all cases
f ∈ L l o c min { r / p , q } ( Ω ) , min { r / p , q } > n / k . f\in L_{\mathrm{loc}}^{\min\{r/p,q\}}(\Omega),\qquad\min\{r/p,q\}>n/k. f ∈ L loc m i n { r / p , q } ( Ω ) , min { r / p , q } > n / k .
Corollary 1.4 completes the proof. □
The conclusion concerns gradient size, not gradient continuity. For instance, u ( x ) = ∣ x 1 ∣ u(x)=|x_1| u ( x ) = ∣ x 1 ∣ satisfies the growth assumption with f = h = 0 f=h=0 f = h = 0 for k ≥ 2 k\ge2 k ≥ 2 but is not C 1 C^1 C 1 ; see Remark 4.2. The strict inequality in (7) is essential in the full weak class: Proposition 3.1 constructs a solution of F k [ u ] = ∣ D u ∣ p ∗ F_k[u]=|Du|^{p_*} F k [ u ] = ∣ D u ∣ p ∗ with D u ∈ L p ∗ Du\in L^{p_*} D u ∈ L p ∗ and an unbounded gradient, as the following construction shows. The density must now equal the stated power of the gradient. The following radial example verifies this equality for the full Hessian measure.
Proposition 3.1. For every 2 ≤ k < n 2 \le k < n 2 ≤ k < n , let p ∗ = n k / ( n − k ) p_* = nk/(n-k) p ∗ = nk / ( n − k ) . There is a radial function u ∈ Φ k ( B 1 ) ∩ W 1 , p ∗ ( B 1 ) u \in\Phi_k(B_1) \cap W^{1,p_*}(B_1) u ∈ Φ k ( B 1 ) ∩ W 1 , p ∗ ( B 1 ) , F k [ u ] = ∣ D u ∣ p ∗ F_k[u] = |Du|^{p_*} F k [ u ] = ∣ D u ∣ p ∗ in B 1 B_1 B 1 , which is smooth and strictly k k k -convex on B 1 ∖ { 0 } B_1 \setminus\{0\} B 1 ∖ { 0 } and satisfies ∣ D u ( x ) ∣ → ∞ |Du(x)| \to\infty ∣ D u ( x ) ∣ → ∞ as x → 0 x \to0 x → 0 , x ≠ 0 x \ne0 x = 0 . Its full Hessian measure has no singular part. If n < 2 k n < 2k n < 2 k , the function is also continuous at zero.
Proof. Set
a = n − k k , b = n − k k 2 , c = [ ( n − 1 k − 1 ) b ] 1 / ( p ∗ − k ) , L ( r ) = log e r , a = \frac{n-k}{k}, \qquad b = \frac{n-k}{k^2}, \qquad c = \left[\binom{n-1}{k-1}b\right]^{1/(p_*-k)}, \qquad L(r) = \log\frac{e}{r}, a = k n − k , b = k 2 n − k , c = [ ( k − 1 n − 1 ) b ] 1/ ( p ∗ − k ) , L ( r ) = log r e ,
and define
w ( r ) = c r − a L ( r ) − b , U ( r ) = − ∫ r 1 w ( t ) d t , u ( x ) = U ( ∣ x ∣ ) . w(r) = cr^{-a}L(r)^{-b}, \qquad U(r) = -\int_r^1 w(t)\,dt, \qquad u(x) = U(|x|). w ( r ) = c r − a L ( r ) − b , U ( r ) = − ∫ r 1 w ( t ) d t , u ( x ) = U ( ∣ x ∣ ) .
Give u ( 0 ) u(0) u ( 0 ) its radial limit, finite if n < 2 k n < 2k n < 2 k and equal to − ∞ -\infty − ∞ otherwise. Since U ′ = w > 0 U' = w > 0 U ′ = w > 0 and r w ′ / w = − a + b / L rw'/w = -a + b/L r w ′ / w = − a + b / L , the radial eigenvalue formula gives, for 1 ≤ j ≤ k 1 \le j \le k 1 ≤ j ≤ k ,
σ j ( D 2 u ) = ( n − 1 j − 1 ) ( w ( r ) r ) j ( n j − n k + b L ( r ) ) > 0. (39) \sigma_j(D^2u) = \binom{n-1}{j-1}\left(\frac{w(r)}{r}\right)^j\left(\frac{n}{j}-\frac{n}{k}+\frac{b}{L(r)}\right)>0. \tag*{(39)} σ j ( D 2 u ) = ( j − 1 n − 1 ) ( r w ( r ) ) j ( j n − k n + L ( r ) b ) > 0. (39)
The identities a p ∗ = n ap_* = n a p ∗ = n and b p ∗ = n / k = k b + 1 bp_* = n/k = kb + 1 b p ∗ = n / k = k b + 1 , together with the choice of c c c , yield
σ k ( D 2 u ) = 1 k ( n − 1 k − 1 ) r 1 − n ( r n − k w ( r ) k ) ′ = w ( r ) p ∗ = ∣ D u ∣ p ∗ , r > 0. \sigma_k(D^2u) = \frac{1}{k}\binom{n-1}{k-1}r^{1-n}(r^{n-k}w(r)^k)' = w(r)^{p_*} = |Du|^{p_*}, \qquad r > 0. σ k ( D 2 u ) = k 1 ( k − 1 n − 1 ) r 1 − n ( r n − k w ( r ) k ) ′ = w ( r ) p ∗ = ∣ D u ∣ p ∗ , r > 0.
Although w ( r ) → ∞ w(r) \to\infty w ( r ) → ∞ ,
∫ B 1 ∣ D u ∣ p ∗ = n ω n c p ∗ ∫ 0 1 d r r L ( r ) n / k < ∞ . (40) \int_{B_1}|Du|^{p_*} = n\omega_n c^{p_*}\int_0^1 \frac{dr}{rL(r)^{n/k}} < \infty. \tag*{(40)} ∫ B 1 ∣ D u ∣ p ∗ = n ω n c p ∗ ∫ 0 1 r L ( r ) n / k d r < ∞. (40)
Since L ( t ) − b ≤ 1 L(t)^{-b} \le1 L ( t ) − b ≤ 1 , the defining integral gives
∣ U ( r ) ∣ ≤ C { 1 , a < 1 , log ( e / r ) , a = 1 , r 1 − a , a > 1. |U(r)| \le C
\begin{cases}
1, & a < 1,\\
\log(e/r), & a = 1,\\
r^{1-a}, & a > 1.
\end{cases} ∣ U ( r ) ∣ ≤ C ⎩ ⎨ ⎧ 1 , log ( e / r ) , r 1 − a , a < 1 , a = 1 , a > 1.
In the last case n + ( 1 − a ) p ∗ = p ∗ > 0 n+(1-a)p_* = p_* > 0 n + ( 1 − a ) p ∗ = p ∗ > 0 , so these bounds give u ∈ L p ∗ ( B 1 ) u \in L^{p_*}(B_1) u ∈ L p ∗ ( B 1 ) in every case. They also imply r n − 1 ∣ U ( r ) ∣ → 0 r^{n-1}|U(r)| \to0 r n − 1 ∣ U ( r ) ∣ → 0 . Integration by parts on punctured balls, together with (3.4), therefore identifies D u = w ( r ) x / r Du = w(r)x/r D u = w ( r ) x / r as the weak gradient and proves u ∈ W 1 , p ∗ ( B 1 ) u \in W^{1,p_*}(B_1) u ∈ W 1 , p ∗ ( B 1 ) .
For the full measure, 0 < ε < 1 0 < \varepsilon< 1 0 < ε < 1 , keep U ε = U U_\varepsilon= U U ε = U on [ ε , 1 ) [\varepsilon,1) [ ε , 1 ) and set
U ε ( r ) = U ( ε ) + w ( ε ) 2 ε ( r 2 − ε 2 ) , 0 ≤ r ≤ ε . U_\varepsilon(r) = U(\varepsilon) + \frac{w(\varepsilon)}{2\varepsilon}(r^2-\varepsilon^2), \qquad0 \le r \le\varepsilon. U ε ( r ) = U ( ε ) + 2 ε w ( ε ) ( r 2 − ε 2 ) , 0 ≤ r ≤ ε .
The matching first derivatives give a locally C 1 , 1 C^{1,1} C 1 , 1 radial function u ε u_\varepsilon u ε whose Hessian is in Γ k \Gamma_k Γ k almost everywhere: it is a positive multiple of I I I inside B ε B_\varepsilon B ε , and (3.3) applies outside. For fixed ε \varepsilon ε , convolution preserves admissibility and the mollified Hessians converge in L l o c k L^k_{\mathrm{loc}} L loc k . Their k k k th symmetric functions therefore converge in L l o c 1 L^1_{\mathrm{loc}} L loc 1 ; (2.2) identifies F k [ u ε ] = σ k ( D 2 u ε ) d x F_k[u_\varepsilon] = \sigma_k(D^2u_\varepsilon)\,dx F k [ u ε ] = σ k ( D 2 u ε ) d x , with no measure on the joining sphere. Its inner mass is
F k [ u ε ] ( B ε ) = ω n ( n k ) ε n − k w ( ε ) k = ω n ( n k ) c k L ( ε ) − k b → 0 as ε → 0. (41) F_k[u_\varepsilon](B_\varepsilon) = \omega_n\binom{n}{k}\varepsilon^{n-k}w(\varepsilon)^k = \omega_n\binom{n}{k}c^kL(\varepsilon)^{-kb} \to0 \quad\text{as } \varepsilon\to0. \tag*{(41)} F k [ u ε ] ( B ε ) = ω n ( k n ) ε n − k w ( ε ) k = ω n ( k n ) c k L ( ε ) − k b → 0 as ε → 0. (41)
Outside this ball the density is ∣ D u ∣ p ∗ |Du|^{p_*} ∣ D u ∣ p ∗ . Moreover,
∫ B ε ∣ u ε − u ∣ ≤ C ε n ( ∣ U ( ε ) ∣ + ε w ( ε ) ) + ∫ B ε ∣ u ∣ → 0 \int_{B_\varepsilon}|u_\varepsilon-u| \le C\varepsilon^n\bigl(|U(\varepsilon)|+\varepsilon w(\varepsilon)\bigr)+\int_{B_\varepsilon}|u| \to0 ∫ B ε ∣ u ε − u ∣ ≤ C ε n ( ∣ U ( ε ) ∣ + εw ( ε ) ) + ∫ B ε ∣ u ∣ → 0
by the preceding bounds. Since u ε = u u_\varepsilon= u u ε = u off B ε B_\varepsilon B ε , this proves local L 1 L^1 L 1 convergence. Closure of Φ k \Phi_k Φ k and weak continuity (2.2), together with (3.4)–(3.5), show that F k [ u ] = ∣ D u ∣ p ∗ d x F_k[u] = |Du|^{p_*}\,dx F k [ u ] = ∣ D u ∣ p ∗ d x with no mass at zero. Finally, n < 2 k n < 2k n < 2 k is exactly a < 1 a < 1 a < 1 , which makes the defining integral finite at zero and gives the stated continuity. □ \square □
Regularity consequences
Throughout this section, we assume 2 ≤ k < n 2 \le k < n 2 ≤ k < n and u ∈ Φ k ( Ω ) u \in\Phi_k(\Omega) u ∈ Φ k ( Ω ) , and write μ = F k [ u ] \mu= F_k[u] μ = F k [ u ] . We apply Theorem 1.2 to obtain integrability and boundedness estimates for D u Du D u and oscillation estimates for u u u . Sharpness examples are given in the appendix.
Gradient integrability and boundedness
Fix B = B R ( z ) ⋐ Ω B = B_R(z) \Subset\Omega B = B R ( z ) ⋐ Ω and a ∈ R n a \in\mathbb{R}^n a ∈ R n , and set E = B R / 2 ( z ) E = B_{R/2}(z) E = B R /2 ( z ) and A = \fint_B |Du-a| . Let λ \lambda λ denote the extension of μ ⌊ B \mu\lfloor B μ ⌊ B by zero to R n \mathbb{R}^n R n . For almost every x ∈ E x \in E x ∈ E , we apply (1.3) on B R / 2 ( x ) ⊂ B B_{R/2}(x) \subset B B R /2 ( x ) ⊂ B . The average is bounded by 2 n A 2^n A 2 n A and the truncated potential by W k ∞ [ λ ] ( x ) W_k^\infty[\lambda](x) W k ∞ [ λ ] ( x ) , giving
∣ D u ( x ) − a ∣ ≤ C A + C W k ∞ [ λ ] ( x ) (42) |Du(x)-a| \le CA + CW_k^\infty[\lambda](x) \tag*{(42)} ∣ D u ( x ) − a ∣ ≤ C A + C W k ∞ [ λ ] ( x ) (42)
for almost every x ∈ E x \in E x ∈ E . Using (1.2) instead, we may replace A A A throughout by R^{-1}\fint_B |u-\ell| for any affine function ℓ \ell ℓ with D ℓ = a D\ell=a D ℓ = a .
For 0 < q < ∞ 0 < q < \infty 0 < q < ∞ and 0 < s < ∞ 0 < s < \infty 0 < s < ∞ , we use the Lorentz quasi-norms
∥ F ∥ L q , s ( G ) = ( ∫ 0 ∣ G ∣ [ t 1 / q F ∗ ( t ) ] s d t t ) 1 / s , ∥ F ∥ L q , ∞ ( G ) = sup 0 < t < ∣ G ∣ t 1 / q F ∗ ( t ) , \|F\|_{L^{q,s}(G)} = \left(\int_0^{|G|} [t^{1/q}F^*(t)]^s \frac{dt}{t}\right)^{1/s}, \qquad\|F\|_{L^{q,\infty}(G)} = \sup_{0<t<|G|} t^{1/q}F^*(t), ∥ F ∥ L q , s ( G ) = ( ∫ 0 ∣ G ∣ [ t 1/ q F ∗ ( t ) ] s t d t ) 1/ s , ∥ F ∥ L q , ∞ ( G ) = 0 < t < ∣ G ∣ sup t 1/ q F ∗ ( t ) ,
where F ∗ F^* F ∗ is the decreasing rearrangement of ∣ F ∣ |F| ∣ F ∣ extended by zero outside G G G . The upper limits are infinite when ∣ G ∣ = ∞ |G|=\infty ∣ G ∣ = ∞ . These definitions apply also for s < 1 s<1 s < 1 , and L q , q ( G ) = L q ( G ) L^{q,q}(G)=L^q(G) L q , q ( G ) = L q ( G ) .
We use the mapping estimates in [4 Proposition 2.8 and Theorem 3.1] with γ = k / ( k + 1 ) \gamma= k/(k+1) γ = k / ( k + 1 ) and p = k + 1 p=k+1 p = k + 1 . These parameters satisfy γ p = k < n \gamma p=k<n γ p = k < n and p − 1 = k p-1=k p − 1 = k . In particular, [4 (4.4)] gives, for F ≥ 0 F \ge0 F ≥ 0 ,
W k ∞ [ F d x ] ≤ C I γ [ ( I γ F ) 1 / k ] , I γ F ( x ) = ∫ R n ∣ x − y ∣ γ − n F ( y ) d y . (43) W_k^\infty[F\,dx] \le C I_\gamma[(I_\gamma F)^{1/k}], \qquad I_\gamma F(x) = \int_{\mathbb{R}^n} |x-y|^{\gamma-n}F(y)\,dy. \tag*{(43)} W k ∞ [ F d x ] ≤ C I γ [( I γ F ) 1/ k ] , I γ F ( x ) = ∫ R n ∣ x − y ∣ γ − n F ( y ) d y . (43)
Proposition 4.1. Let p ∗ = n k / ( n − k ) p_* = nk/(n-k) p ∗ = nk / ( n − k ) . With the notation above, the following estimates hold.
(i) For arbitrary Hessian measure,
∥ D u − a ∥ L p ∗ , ∞ ( E ) ≤ C R n / p ∗ A + C F k [ u ] ( B ) 1 / k , (44) \|Du-a\|_{L^{p_*,\infty}(E)} \le CR^{n/p_*}A + C F_k[u](B)^{1/k}, \tag*{(44)} ∥ D u − a ∥ L p ∗ , ∞ ( E ) ≤ C R n / p ∗ A + C F k [ u ] ( B ) 1/ k , (44) where C = C ( n , k ) C=C(n,k) C = C ( n , k ) .
(ii) If F k [ u ] = f F_k[u]=f F k [ u ] = f on B B B , 1 < q < n / k 1<q<n/k 1 < q < n / k , 0 < s ≤ ∞ 0<s\le\infty 0 < s ≤ ∞ , and p = n k q / ( n − k q ) p=nkq/(n-kq) p = nk q / ( n − k q ) , then
∥ D u − a ∥ L p , k s ( E ) ≤ C R n / p A + C ∥ f ∥ L q , s ( B ) 1 / k , (45) \|Du-a\|_{L^{p,ks}(E)} \le CR^{n/p}A + C\|f\|_{L^{q,s}(B)}^{1/k}, \tag*{(45)} ∥ D u − a ∥ L p , k s ( E ) ≤ C R n / p A + C ∥ f ∥ L q , s ( B ) 1/ k , (45) where k ∞ = ∞ k\infty=\infty k ∞ = ∞ and C = C ( n , k , q , s ) C=C(n,k,q,s) C = C ( n , k , q , s ) .
(iii) If F k [ u ] = f F_k[u]=f F k [ u ] = f on B B B and f ∈ L n / k , 1 / k ( B ) f\in L^{n/k,1/k}(B) f ∈ L n / k , 1/ k ( B ) , then
∥ D u − a ∥ L ∞ ( E ) ≤ C A + C ∥ f ∥ L n / k , 1 / k ( B ) 1 / k , (46) \|Du-a\|_{L^\infty(E)} \le CA + C\|f\|_{L^{n/k,1/k}(B)}^{1/k}, \tag*{(46)} ∥ D u − a ∥ L ∞ ( E ) ≤ C A + C ∥ f ∥ L n / k , 1/ k ( B ) 1/ k , (46) where C = C ( n , k ) C=C(n,k) C = C ( n , k ) . If this density condition holds locally throughout Ω \Omega Ω , then u u u is locally Lipschitz and, for every compact K ⋐ Ω K\Subset\Omega K ⋐ Ω ,
lim ρ → 0 sup x ∈ K W k ρ [ f d x ] ( x ) = 0. (47) \lim_{\rho\to0}\sup_{x\in K} W_k^\rho[f\,dx](x)=0. \tag*{(47)} ρ → 0 lim x ∈ K sup W k ρ [ f d x ] ( x ) = 0. (47) (iv) If F k [ u ] = f F_k[u]=f F k [ u ] = f on B B B and f ∈ L q , ∞ ( B ) f\in L^{q,\infty}(B) f ∈ L q , ∞ ( B ) for n / k < q ≤ ∞ n/k<q\le\infty n / k < q ≤ ∞ , let η q = 1 − n / ( k q ) > 0 \eta_q=1-n/(kq)>0 η q = 1 − n / ( k q ) > 0 , with η ∞ = 1 \eta_\infty=1 η ∞ = 1 and L ∞ , ∞ = L ∞ L^{\infty,\infty}=L^\infty L ∞ , ∞ = L ∞ . Then
∥ D u − a ∥ L ∞ ( E ) ≤ C A + C η q R η q ∥ f ∥ L q , ∞ ( B ) 1 / k . \|Du-a\|_{L^\infty(E)} \le CA + \frac{C}{\eta_q}R^{\eta_q}\|f\|_{L^{q,\infty}(B)}^{1/k}. ∥ D u − a ∥ L ∞ ( E ) ≤ C A + η q C R η q ∥ f ∥ L q , ∞ ( B ) 1/ k . Moreover, for every ρ > 0 \rho>0 ρ > 0 ,
sup x ∈ R n W k ρ [ f 1 B d x ] ( x ) ≤ C η q ρ η q ∥ f ∥ L q , ∞ ( B ) 1 / k , (48) \sup_{x\in\mathbb{R}^n} W_k^\rho[f1_B\,dx](x) \le\frac{C}{\eta_q}\rho^{\eta_q}\|f\|_{L^{q,\infty}(B)}^{1/k}, \tag*{(48)} x ∈ R n sup W k ρ [ f 1 B d x ] ( x ) ≤ η q C ρ η q ∥ f ∥ L q , ∞ ( B ) 1/ k , (48) where C = C ( n , k ) C=C(n,k) C = C ( n , k ) . In particular, the small-scale potentials vanish uniformly on compact subsets of $B.
Proof. (i) Let m = λ ( R n ) m=\lambda(\mathbb{R}^n) m = λ ( R n ) and let M λ M\lambda M λ be the centered maximal function of the measure. Splitting the Wolff integral at δ > 0 \delta>0 δ > 0 gives
W k ∞ [ λ ] ( x ) ≤ C ( M λ ( x ) ) 1 / k δ + C m 1 / k δ 1 − n / k . \mathcal{W}^{\infty}_{k}[\lambda](x)\leq C(M\lambda(x))^{1/k}\delta+C m^{1/k}\delta^{1-n/k}. W k ∞ [ λ ] ( x ) ≤ C ( M λ ( x ) ) 1/ k δ + C m 1/ k δ 1 − n / k .
For 0 < M λ ( x ) < ∞ 0<M\lambda(x)<\infty 0 < M λ ( x ) < ∞ , take δ = ( m / M λ ( x ) ) 1 / n \delta=(m/M\lambda(x))^{1/n} δ = ( m / M λ ( x ) ) 1/ n . Then W k ∞ [ λ ] ≤ C m 1 / n ( M λ ) 1 / k − 1 / n \mathcal{W}^{\infty}_{k}[\lambda]\leq C m^{1/n}(M\lambda)^{1/k-1/n} W k ∞ [ λ ] ≤ C m 1/ n ( M λ ) 1/ k − 1/ n . Since ( 1 / k − 1 / n ) − 1 = p ∗ (1/k-1/n)^{-1}=p_* ( 1/ k − 1/ n ) − 1 = p ∗ , the weak ( 1 , 1 ) (1,1) ( 1 , 1 ) maximal estimate yields
∣ { x : W k ∞ [ λ ] ( x ) > τ } ∣ ≤ C ( m 1 / k τ ) p ∗ , τ > 0. \left|\left\{x:\mathcal{W}^{\infty}_{k}[\lambda](x)>\tau\right\}\right|\leq C\left(\frac{m^{1/k}}{\tau}\right)^{p_*},\qquad\tau>0. ∣ { x : W k ∞ [ λ ] ( x ) > τ } ∣ ≤ C ( τ m 1/ k ) p ∗ , τ > 0.
Thus ∥ W k ∞ [ λ ] ∥ L p ∗ , ∞ ≤ C m 1 / k \|\mathcal{W}^{\infty}_{k}[\lambda]\|_{L^{p_*,\infty}}\leq C m^{1/k} ∥ W k ∞ [ λ ] ∥ L p ∗ , ∞ ≤ C m 1/ k ; the case m = 0 m=0 m = 0 is immediate. Combining this with (42) proves (44) .
(ii) Apply [4 Theorem 3.1(i)] to (43) , with F = f 1 B F=f\mathbf{1}_B F = f 1 B . The target space is L n k q / ( n − k q ) , k s L^{nkq/(n-kq),ks} L nk q / ( n − k q ) , k s for all 0 < s ≤ ∞ 0<s\leq\infty 0 < s ≤ ∞ , including k s < 1 ks<1 k s < 1 . Equation (42) and ∥ A ∥ L p , k s ( E ) = C R n / p A \|A\|_{L^{p,ks}(E)}=CR^{n/p}A ∥ A ∥ L p , k s ( E ) = C R n / p A give (45) .
(iii) By [4 Theorem 3.1(iv)] and (43) ,
sup x ∈ R n W k ∞ [ F d x ] ( x ) ≤ C ∥ F ∥ L n / k , 1 / k ( R n ) 1 / k . \sup_{x\in\mathbb{R}^n}\mathcal{W}^{\infty}_{k}[F\,dx](x)\leq C\|F\|_{L^{n/k,1/k}(\mathbb{R}^n)}^{1/k}. x ∈ R n sup W k ∞ [ F d x ] ( x ) ≤ C ∥ F ∥ L n / k , 1/ k ( R n ) 1/ k .
The essential bound holds at every point because the potential is lower semicontinuous. Taking F = f 1 B F=f\mathbf{1}_B F = f 1 B and using (42) proves (46) .
For the uniform tails, choose K ⋐ U ⋐ Ω K\Subset U\Subset\Omega K ⋐ U ⋐ Ω and extend F = f 1 U F=f\mathbf{1}_U F = f 1 U by zero. For small ρ \rho ρ and x ∈ K x\in K x ∈ K ,
W k ρ [ f d x ] ( x ) ≤ W k ∞ [ F 1 B ρ ( x ) d x ] ( x ) ≤ C ∥ F 1 B ρ ( x ) ∥ L n / k , 1 / k 1 / k . \mathcal{W}^{\rho}_{k}[f\,dx](x)\leq\mathcal{W}^{\infty}_{k}[F\mathbf{1}_{B_\rho(x)}\,dx](x)\leq C\|F\mathbf{1}_{B_\rho(x)}\|_{L^{n/k,1/k}}^{1/k}. W k ρ [ f d x ] ( x ) ≤ W k ∞ [ F 1 B ρ ( x ) d x ] ( x ) ≤ C ∥ F 1 B ρ ( x ) ∥ L n / k , 1/ k 1/ k .
For a measurable set G G G , ( F 1 G ) ∗ ( t ) ≤ F ∗ ( t ) 1 ( 0 , ∣ G ∣ ) ( t ) (F\mathbf{1}_G)^*(t)\leq F^*(t)\mathbf{1}_{(0,|G|)}(t) ( F 1 G ) ∗ ( t ) ≤ F ∗ ( t ) 1 ( 0 , ∣ G ∣ ) ( t ) , and hence
∥ F 1 G ∥ L n / k , 1 / k 1 / k ≤ ∫ 0 ∣ G ∣ [ t k / n F ∗ ( t ) ] 1 / k d t t ⟶ 0 ( ∣ G ∣ → 0 ) . \|F\mathbf{1}_G\|_{L^{n/k,1/k}}^{1/k}\leq\int_0^{|G|}[t^{k/n}F^*(t)]^{1/k}\frac{dt}{t}\longrightarrow0\quad(|G|\to0). ∥ F 1 G ∥ L n / k , 1/ k 1/ k ≤ ∫ 0 ∣ G ∣ [ t k / n F ∗ ( t ) ] 1/ k t d t ⟶ 0 ( ∣ G ∣ → 0 ) .
The bound depends only on ∣ G ∣ |G| ∣ G ∣ . Taking G = B ρ ( x ) G=B_\rho(x) G = B ρ ( x ) proves (47) . The bounded weak gradient gives a locally Lipschitz representative, which agrees with the canonical subharmonic representative of u u u .
(iv) For finite q q q , the weak-L q L^q L q estimate gives
∫ B t ( x ) f 1 B d x ≤ q q − 1 ∣ B t ∣ 1 − 1 / q ∥ f ∥ L q , ∞ ( B ) . \int_{B_t(x)}f\mathbf{1}_B\,dx\leq\frac{q}{q-1}|B_t|^{1-1/q}\|f\|_{L^{q,\infty}(B)}. ∫ B t ( x ) f 1 B d x ≤ q − 1 q ∣ B t ∣ 1 − 1/ q ∥ f ∥ L q , ∞ ( B ) .
Since q > n / k > 1 q>n/k>1 q > n / k > 1 , its coefficient is bounded in terms of n , k n,k n , k . Substitution in the Wolff integral gives (48) , because ∫ 0 ρ t η q d t / t = ρ η q / η q \int_0^\rho t^{\eta q}\,dt/t=\rho^{\eta q}/\eta q ∫ 0 ρ t η q d t / t = ρ η q / η q . For q = ∞ q=\infty q = ∞ , use ∫ B t ( x ) f 1 B d x ≤ ∣ B t ∣ ∥ f ∥ ∞ \int_{B_t(x)}f\mathbf{1}_B\,dx\leq|B_t|\|f\|_\infty ∫ B t ( x ) f 1 B d x ≤ ∣ B t ∣∥ f ∥ ∞ instead. Applying (1.3) on B R / 2 ( x ) ⊂ B B_{R/2}(x)\subset B B R /2 ( x ) ⊂ B for almost every x ∈ E x\in E x ∈ E proves the gradient bound; its average term is at most 2 n A 2^nA 2 n A . □
Remark 4.2. The space L p ∗ , ∞ L^{p_*,\infty} L p ∗ , ∞ in (44) cannot in general be replaced by L p ∗ L^{p_*} L p ∗ , where p ∗ = n k / ( n − k ) p_*=nk/(n-k) p ∗ = nk / ( n − k ) . The radial k k k -Hessian fundamental solution is a suitably signed multiple of ∣ x ∣ 2 − n / k |x|^{2-n/k} ∣ x ∣ 2 − n / k when n ≠ 2 k n\neq2k n = 2 k , and a positive multiple of log ∣ x ∣ \log|x| log ∣ x ∣ when n = 2 k n=2k n = 2 k ; see [15 ] . Its Hessian measure is c δ 0 c\delta_0 c δ 0 with c > 0 c>0 c > 0 , and ∣ D u ( x ) ∣ ≍ ∣ x ∣ 1 − n / k |Du(x)|\asymp|x|^{1-n/k} ∣ D u ( x ) ∣ ≍ ∣ x ∣ 1 − n / k , so D u ∈ L p ∗ , ∞ Du\in L^{p_*,\infty} D u ∈ L p ∗ , ∞ but D u ∉ L p ∗ Du\notin L^{p_*} D u ∈ / L p ∗ near the origin. The exponent 1 / k 1/k 1/ k in Proposition 4.1(iii) is sharp as well: the example in Proposition 1 has unbounded gradient, while its Hessian density satisfies
f ( x ) ≍ ∣ x ∣ − k [ log ( 4 / ∣ x ∣ ) ] − k , f ∈ L n / k , s ( B 1 ) for every s > 1 / k . f(x)\asymp|x|^{-k}[\log(4/|x|)]^{-k},\qquad f\in L^{n/k,s}(B_1)\text{ for every }s>1/k. f ( x ) ≍ ∣ x ∣ − k [ log ( 4/∣ x ∣ ) ] − k , f ∈ L n / k , s ( B 1 ) for every s > 1/ k .
The potential can also vanish identically while the gradient is discontinuous. For u ( x ) = ∣ x 1 ∣ u(x)=|x_1| u ( x ) = ∣ x 1 ∣ , the smooth convex approximations u ε ( x ) = x 1 2 + ε 2 u_\varepsilon(x)=\sqrt{x_1^2+\varepsilon^2} u ε ( x ) = x 1 2 + ε 2 have rank-one Hessians, so weak continuity gives F k [ u ] = 0 F_k[u]=0 F k [ u ] = 0 for k ≥ 2 k\geq2 k ≥ 2 , whereas D u = sgn ( x 1 ) e 1 Du=\operatorname{sgn}(x_1)e_1 D u = sgn ( x 1 ) e 1 almost everywhere. On every ball B ρ B_\rho B ρ centered at the origin,
∮ B ρ ∣ D u − ( D u ) B ρ ∣ = 1 , \oint_{B_\rho}|Du-(Du)_{B_\rho}|=1, ∮ B ρ ∣ D u − ( D u ) B ρ ∣ = 1 ,
so D u Du D u is not in VMO. Gradient BMO can fail even under vanishing critical Morrey density; see Proposition 3 .
Averaged estimates and moduli of continuity
The averaged estimates below hold at every center. For B R ( x ) ⊂ Ω B_R(x) \subset\Omega B R ( x ) ⊂ Ω , β ≥ 0 \beta\ge0 β ≥ 0 , and 0 ≤ α ≤ 1 0 \le\alpha\le1 0 ≤ α ≤ 1 , set
M_{\beta,R}F(x)=\sup_{0<r\le R}r^\beta\fint_{B_r(x)}|F|,\qquad M_{\alpha,R}^{\#}v(x)=\sup_{0<r\le R}r^{-\alpha}\fint_{B_r(x)}|v-(v)_{B_r(x)}|.
For 0 ≤ α ≤ 1 0 \le\alpha\le1 0 ≤ α ≤ 1 and σ ≥ 0 \sigma\ge0 σ ≥ 0 , we also define
P α R [ μ ] ( x ) = ∫ 0 R ( μ ( B t ( x ) ) t n − 2 k + k α ) 1 / k d t t , Q σ , R μ ( x ) = sup 0 < t ≤ R μ ( B t ( x ) ) t n − σ . P_{\alpha}^{R}[\mu](x)=\int_0^R\left(\frac{\mu(B_t(x))}{t^{n-2k+k\alpha}}\right)^{1/k}\frac{dt}{t},\qquad Q_{\sigma,R}\mu(x)=\sup_{0<t\le R}\frac{\mu(B_t(x))}{t^{n-\sigma}}. P α R [ μ ] ( x ) = ∫ 0 R ( t n − 2 k + k α μ ( B t ( x )) ) 1/ k t d t , Q σ , R μ ( x ) = 0 < t ≤ R sup t n − σ μ ( B t ( x )) .
With this notation, P 1 R = W k R P_1^R=\mathcal{W}_k^R P 1 R = W k R . We use Q σ , R μ ( x ) Q_{\sigma,R}\mu(x) Q σ , R μ ( x ) through the bound μ ( B t ( x ) ) ≤ Q σ , R μ ( x ) t n − σ \mu(B_t(x))\le Q_{\sigma,R}\mu(x)t^{n-\sigma} μ ( B t ( x )) ≤ Q σ , R μ ( x ) t n − σ for 0 < t ≤ R 0<t\le R 0 < t ≤ R , which holds also for σ ≥ n \sigma\ge n σ ≥ n .
Proposition 4.3. Let B R ( x ) ⋐ Ω B_R(x)\Subset\Omega B R ( x ) ⋐ Ω , let ℓ \ell ℓ be affine with D ℓ = a D\ell=a D ℓ = a , and set A=\fint_{B_R(x)}|Du-a| . Then
M α , R # ( u − ℓ ) ( x ) + M 1 − α , R ( D u − a ) ( x ) ≤ C P α R [ μ ] ( x ) + C R 1 − α A , 0 ≤ α ≤ 1 , (49) M_{\alpha,R}^{\#}(u-\ell)(x)+M_{1-\alpha,R}(Du-a)(x)\le CP_{\alpha}^{R}[\mu](x)+CR^{1-\alpha}A,\qquad0\le\alpha\le1, \tag*{(49)} M α , R # ( u − ℓ ) ( x ) + M 1 − α , R ( D u − a ) ( x ) ≤ C P α R [ μ ] ( x ) + C R 1 − α A , 0 ≤ α ≤ 1 , (49) M α , R # ( u − ℓ ) ( x ) + M 1 − α , R ( D u − a ) ( x ) ≤ C 1 − α [ Q k ( 2 − α ) , R μ ( x ) ] 1 / k + C R 1 − α A , 0 ≤ α < 1 , M_{\alpha,R}^{\#}(u-\ell)(x)+M_{1-\alpha,R}(Du-a)(x)\le\frac{C}{1-\alpha}\left[Q_{k(2-\alpha),R}\mu(x)\right]^{1/k}+CR^{1-\alpha}A,\qquad0\le\alpha<1, M α , R # ( u − ℓ ) ( x ) + M 1 − α , R ( D u − a ) ( x ) ≤ 1 − α C [ Q k ( 2 − α ) , R μ ( x ) ] 1/ k + C R 1 − α A , 0 ≤ α < 1 , where C = C ( n , k ) C=C(n,k) C = C ( n , k ) is independent of α \alpha α . Infinite right-hand sides are allowed.
Proof. The sharp maximal term follows from the averaged gradient term by Poincaré’s inequality. Indeed, since u ∈ W l o c 1 , 1 u\in W_{\mathrm{loc}}^{1,1} u ∈ W loc 1 , 1 , for every 0 < r ≤ R 0<r\le R 0 < r ≤ R we have
r^{-\alpha}\fint_{B_r(x)}|(u-\ell)-(u-\ell)_{B_r(x)}|\le Cr^{1-\alpha}\fint_{B_r(x)}|Du-a|.
Taking the supremum in r r r gives
M α , R # ( u − ℓ ) ( x ) ≤ C M 1 − α , R ( D u − a ) ( x ) . M_{\alpha,R}^{\#}(u-\ell)(x)\le CM_{1-\alpha,R}(Du-a)(x). M α , R # ( u − ℓ ) ( x ) ≤ C M 1 − α , R ( D u − a ) ( x ) .
It therefore suffices to establish the two stated estimates for M 1 − α , R ( D u − a ) ( x ) M_{1-\alpha,R}(Du-a)(x) M 1 − α , R ( D u − a ) ( x ) .
Fix 0 < r ≤ R / 4 0<r\le R/4 0 < r ≤ R /4 . For almost every y ∈ B r ( x ) y\in B_r(x) y ∈ B r ( x ) , apply (1.3) on B R / 2 ( y ) ⊂ B R ( x ) B_{R/2}(y)\subset B_R(x) B R /2 ( y ) ⊂ B R ( x ) . Its average term is at most 2 n A 2^nA 2 n A . We average the potential term over B r ( x ) B_r(x) B r ( x ) and split its radii at r r r . For 0 < t < r 0<t<r 0 < t < r , Fubini’s theorem gives
\fint_{B_r(x)}\mu(B_t(y))\,dy=\frac{1}{|B_r|}\int_{B_{r+t}(x)}|B_r(x)\cap B_t(z)|\,d\mu(z)\le C(t/r)^n\mu(B_{2r}(x)).
By concavity of the k k k th root,
\fint_{B_r(x)}\mu(B_t(y))^{1/k}\,dy\le C(t/r)^{n/k}\mu(B_{2r}(x))^{1/k}.
Integrating against t 1 − n / k d t / t t^{1-n/k}\,dt/t t 1 − n / k d t / t over ( 0 , r ) (0,r) ( 0 , r ) gives C μ ( B 2 r ( x ) ) 1 / k r 1 − n / k C\mu(B_{2r}(x))^{1/k}r^{1-n/k} C μ ( B 2 r ( x ) ) 1/ k r 1 − n / k . For r ≤ t ≤ R / 2 r\le t\le R/2 r ≤ t ≤ R /2 , use B t ( y ) ⊂ B 2 t ( x ) B_t(y)\subset B_{2t}(x) B t ( y ) ⊂ B 2 t ( x ) and then change variables from t t t to 2 t 2t 2 t . Thus
\fint_{B_r(x)}|Du-a|\le CA+C\mu(B_{2r}(x))^{1/k}r^{1-n/k}+C\int_{2r}^{R}\mu(B_t(x))^{1/k}t^{1-n/k}\frac{dt}{t}. \tag*{(50)}
All integrands are nonnegative, so these steps remain valid when a potential is infinite. Multiply (4.9) by r 1 − α r^{1-\alpha} r 1 − α . For 0 ≤ α ≤ 1 0\le\alpha\le1 0 ≤ α ≤ 1 , monotonicity of the ball mass gives
μ ( B 2 r ( x ) ) 1 / k r 2 − α − n / k ≤ C ∫ 2 r 4 r ( μ ( B t ( x ) ) t n − 2 k + k α ) 1 / k d t t . \mu(B_{2r}(x))^{1/k}r^{2-\alpha-n/k}\le C\int_{2r}^{4r}\left(\frac{\mu(B_t(x))}{t^{n-2k+k\alpha}}\right)^{1/k}\frac{dt}{t}. μ ( B 2 r ( x ) ) 1/ k r 2 − α − n / k ≤ C ∫ 2 r 4 r ( t n − 2 k + k α μ ( B t ( x )) ) 1/ k t d t .
The constant is uniform in α \alpha α , because t / r ∈ [ 2 , 4 ] t/r\in[2,4] t / r ∈ [ 2 , 4 ] . For the final term in (4.9), use r 1 − α ≤ t 1 − α r^{1-\alpha}\le t^{1-\alpha} r 1 − α ≤ t 1 − α . Both measure terms are consequently bounded by C P α R [ μ ] ( x ) CP_{\alpha}^{R}[\mu](x) C P α R [ μ ] ( x ) , and the average term is at most C R 1 − α A CR^{1-\alpha}A C R 1 − α A . For the second estimate, let 0 ≤ α < 1 0 \le\alpha< 1 0 ≤ α < 1 and set Q = Q k ( 2 − α ) , R μ ( x ) Q = Q_{k(2-\alpha),R}\mu(x) Q = Q k ( 2 − α ) , R μ ( x ) . We may assume Q < ∞ Q < \infty Q < ∞ . Then μ ( B t ( x ) ) ≤ Q t n − 2 k + k α \mu(B_t(x)) \le Qt^{n-2k+k\alpha} μ ( B t ( x )) ≤ Q t n − 2 k + k α for t ≤ R t \le R t ≤ R . Substitution into (4.9) yields
r^{1-\alpha}\fint_{B_r(x)} |Du-a| \le CR^{1-\alpha}A + CQ^{1/k}\left(1+r^{1-\alpha}\int_{2r}^{R}t^{\alpha-1}\frac{dt}{t}\right) \le CR^{1-\alpha}A+\frac{C}{1-\alpha}Q^{1/k}.
Finally, for R / 4 < r ≤ R R/4 < r \le R R /4 < r ≤ R ,
r^{1-\alpha}\fint_{B_r(x)} |Du-a| \le4^nR^{1-\alpha}A.
Taking the supremum over all radii proves both gradient maximal estimates. The Poincaré reduction at the beginning then supplies the sharp maximal term in each inequality. □
Corollary 4.4. Let B 2 R ( z ) ⋐ Ω B_{2R}(z) \Subset\Omega B 2 R ( z ) ⋐ Ω , let ℓ \ell ℓ be affine, and set A=\fint_{B_{2R}(z)}|Du-D\ell| .
(i) Suppose that, for some 0 ≤ α < 1 0 \le\alpha< 1 0 ≤ α < 1 ,
μ ( B t ( x ) ) ≤ M t n − 2 k + k α ( x ∈ B R ( z ) , 0 < t ≤ R ) . (51) \mu(B_t(x)) \le Mt^{n-2k+k\alpha}\qquad(x\in B_R(z),\ 0<t\le R). \tag*{(51)} μ ( B t ( x )) ≤ M t n − 2 k + k α ( x ∈ B R ( z ) , 0 < t ≤ R ) . (51) Then
[ u − ℓ ] C 0 , α ( B R / 4 ( z ) ) ≤ C α ( M 1 / k 1 − α + R 1 − α A ) , 0 < α < 1 , (52) [u-\ell]_{C^{0,\alpha}(B_{R/4}(z))}\le\frac{C}{\alpha}\left(\frac{M^{1/k}}{1-\alpha}+R^{1-\alpha}A\right),\qquad0<\alpha<1, \tag*{(52)} [ u − ℓ ] C 0 , α ( B R /4 ( z )) ≤ α C ( 1 − α M 1/ k + R 1 − α A ) , 0 < α < 1 , (52) [ u − ℓ ] B M O ( B R / 4 ( z ) ) ≤ C ( M 1 / k + R A ) , α = 0. (53) [u-\ell]_{\mathrm{BMO}(B_{R/4}(z))}\le C(M^{1/k}+RA),\qquad\alpha=0. \tag*{(53)} [ u − ℓ ] BMO ( B R /4 ( z )) ≤ C ( M 1/ k + R A ) , α = 0. (53) Here BMO uses mean oscillations on subballs.
(ii) Suppose instead that
μ ( B t ( x ) ) ≤ M t n − k ( x ∈ B R ( z ) , 0 < t ≤ R ) . (54) \mu(B_t(x))\le Mt^{n-k}\qquad(x\in B_R(z),\ 0<t\le R). \tag*{(54)} μ ( B t ( x )) ≤ M t n − k ( x ∈ B R ( z ) , 0 < t ≤ R ) . (54) Then, for distinct x , y ∈ B R / 4 ( z ) x,y\in B_{R/4}(z) x , y ∈ B R /4 ( z ) ,
∣ ( u − ℓ ) ( x ) − ( u − ℓ ) ( y ) ∣ ≤ C ∣ x − y ∣ [ A + M 1 / k log e R ∣ x − y ∣ ] . (55) |(u-\ell)(x)-(u-\ell)(y)|\le C|x-y|\left[A+M^{1/k}\log\frac{eR}{|x-y|}\right]. \tag*{(55)} ∣ ( u − ℓ ) ( x ) − ( u − ℓ ) ( y ) ∣ ≤ C ∣ x − y ∣ [ A + M 1/ k log ∣ x − y ∣ e R ] . (55) In both parts C = C ( n , k ) C=C(n,k) C = C ( n , k ) .
Proof. (i) For every x ∈ B R ( z ) x\in B_R(z) x ∈ B R ( z ) , one has B R ( x ) ⊂ B 2 R ( z ) B_R(x)\subset B_{2R}(z) B R ( x ) ⊂ B 2 R ( z ) and \fint_{B_R(x)}|Du-D\ell|\le2^nA . By (4.8) and (4.10),
M α , R # ( u − ℓ ) ( x ) ≤ C M 1 / k 1 − α + C R 1 − α A . M^\#_{\alpha,R}(u-\ell)(x)\le\frac{CM^{1/k}}{1-\alpha}+CR^{1-\alpha}A. M α , R # ( u − ℓ ) ( x ) ≤ 1 − α C M 1/ k + C R 1 − α A .
For 0 < α < 1 0<\alpha<1 0 < α < 1 , summing differences of averages on B 2 − j r ( x ) B_{2^{-j}r}(x) B 2 − j r ( x ) gives the factor ( 1 − 2 − α ) − 1 (1-2^{-\alpha})^{-1} ( 1 − 2 − α ) − 1 . Comparing averages at two centers on the scale of their distance then proves (4.11), since ( 1 − 2 − α ) − 1 ≤ C / α (1-2^{-\alpha})^{-1}\le C/\alpha ( 1 − 2 − α ) − 1 ≤ C / α . The resulting continuous representative agrees with the canonical subharmonic representative. For α = 0 \alpha=0 α = 0 , the same maximal estimate directly bounds mean oscillations on every subball of B R / 4 ( z ) B_{R/4}(z) B R /4 ( z ) and proves (4.12).
(ii) For 0 < β ≤ 1 / 2 0<\beta\le1/2 0 < β ≤ 1/2 , (4.13) implies Q k ( 1 + β ) , R μ ( x ) ≤ M R k β Q_{k(1+\beta),R}\mu(x)\le MR^{k\beta} Q k ( 1 + β ) , R μ ( x ) ≤ M R k β uniformly for x ∈ B R ( z ) x\in B_R(z) x ∈ B R ( z ) . Apply (4.8) with α = 1 − β \alpha=1-\beta α = 1 − β and the Campanato argument in part (i). Its constant is bounded because α ≥ 1 / 2 \alpha\ge1/2 α ≥ 1/2 . With d = ∣ x − y ∣ d=|x-y| d = ∣ x − y ∣ ,
∣ ( u − ℓ ) ( x ) − ( u − ℓ ) ( y ) ∣ ≤ C d 1 − β R β ( A + M 1 / k / β ) . |(u-\ell)(x)-(u-\ell)(y)|\le Cd^{1-\beta}R^\beta(A+M^{1/k}/\beta). ∣ ( u − ℓ ) ( x ) − ( u − ℓ ) ( y ) ∣ ≤ C d 1 − β R β ( A + M 1/ k / β ) .
Choose β = min { 1 / 2 , 1 / log ( e R / d ) } \beta=\min\{1/2,1/\log(eR/d)\} β = min { 1/2 , 1/ log ( e R / d )} . Then d 1 − β R β = d ( R / d ) β ≤ e d d^{1-\beta}R^\beta=d(R/d)^\beta\le ed d 1 − β R β = d ( R / d ) β ≤ e d and 1 / β ≤ 2 log ( e R / d ) 1/\beta\le2\log(eR/d) 1/ β ≤ 2 log ( e R / d ) , which give (4.14). □
Suppose that μ = f d x \mu=f\,dx μ = f d x on B 2 R ( z ) B_{2R}(z) B 2 R ( z ) . For 1 < q < ∞ 1<q<\infty 1 < q < ∞ and every measurable set G ⊂ B 2 R ( z ) G\subset B_{2R}(z) G ⊂ B 2 R ( z ) , we have
∫ G f d x ≤ q q − 1 ∣ G ∣ 1 − 1 / q ∥ f ∥ L q , ∞ ( B 2 R ( z ) ) . \int_G f\,dx\le\frac{q}{q-1}|G|^{1-1/q}\|f\|_{L^{q,\infty}(B_{2R}(z))}. ∫ G f d x ≤ q − 1 q ∣ G ∣ 1 − 1/ q ∥ f ∥ L q , ∞ ( B 2 R ( z )) .
Taking G = B t ( x ) G = B_t(x) G = B t ( x ) gives μ ( B t ( x ) ) ≤ C ( n , q ) ∥ f ∥ L q , ∞ ( B 2 R ( z ) ) t n − n / q \mu(B_t(x)) \le C(n,q)\|f\|_{L^{q,\infty}(B_{2R}(z))}t^{n-n/q} μ ( B t ( x )) ≤ C ( n , q ) ∥ f ∥ L q , ∞ ( B 2 R ( z )) t n − n / q for x ∈ B R ( z ) x \in B_R(z) x ∈ B R ( z ) and 0 < t ≤ R 0 < t \le R 0 < t ≤ R . If max { 1 , n / ( 2 k ) } < q < n / k \max\{1,n/(2k)\} < q < n/k max { 1 , n / ( 2 k )} < q < n / k , then α = 2 − n / ( k q ) ∈ ( 0 , 1 ) \alpha= 2-n/(kq) \in(0,1) α = 2 − n / ( k q ) ∈ ( 0 , 1 ) and n − n / q = n − 2 k + k α n-n/q = n-2k+k\alpha n − n / q = n − 2 k + k α . Corollary 4.4 therefore yields
[ u − ℓ ] C 0 , α ( B R / 4 ( z ) ) ≤ C ( n , k , q ) [ ∥ f ∥ L q , ∞ ( B 2 R ( z ) ) 1 / k + R 1 − α A ] . (56) [u-\ell]_{C^{0,\alpha}(B_{R/4}(z))} \le C(n,k,q)\left[\|f\|_{L^{q,\infty}(B_{2R}(z))}^{1/k}+R^{1-\alpha}A\right]. \tag*{(56)} [ u − ℓ ] C 0 , α ( B R /4 ( z )) ≤ C ( n , k , q ) [ ∥ f ∥ L q , ∞ ( B 2 R ( z )) 1/ k + R 1 − α A ] . (56)
At q = n / k q=n/k q = n / k , the same calculation gives the growth exponent n − k n-k n − k . Applying (55) , we obtain
∣ ( u − ℓ ) ( x ) − ( u − ℓ ) ( y ) ∣ ≤ C ∣ x − y ∣ [ A + ∥ f ∥ L n / k , ∞ ( B 2 R ( z ) ) 1 / k log e R ∣ x − y ∣ ] |(u-\ell)(x)-(u-\ell)(y)| \le C|x-y|\left[A+\|f\|_{L^{n/k,\infty}(B_{2R}(z))}^{1/k}\log\frac{eR}{|x-y|}\right] ∣ ( u − ℓ ) ( x ) − ( u − ℓ ) ( y ) ∣ ≤ C ∣ x − y ∣ [ A + ∥ f ∥ L n / k , ∞ ( B 2 R ( z )) 1/ k log ∣ x − y ∣ e R ]
for distinct x , y ∈ B R / 4 ( z ) x,y \in B_{R/4}(z) x , y ∈ B R /4 ( z ) , where C = C ( n , k ) C=C(n,k) C = C ( n , k ) .
Remark 4.5. The Hölder exponent in (56) is optimal. Let max { 1 , n / ( 2 k ) } < q < n / k \max\{1,n/(2k)\}<q<n/k max { 1 , n / ( 2 k )} < q < n / k , α = 2 − n / ( k q ) \alpha=2-n/(kq) α = 2 − n / ( k q ) , and u ( x ) = ∣ x ∣ α u(x)=|x|^\alpha u ( x ) = ∣ x ∣ α . For x ≠ 0 x\ne0 x = 0 ,
σ k ( D 2 u ) = 1 k ( n − 1 k − 1 ) α k ( n − 2 k + k α ) ∣ x ∣ k α − 2 k ∈ L q , ∞ ( B 1 ) . \sigma_k(D^2u)=\frac{1}{k}\binom{n-1}{k-1}\alpha^k(n-2k+k\alpha)|x|^{k\alpha-2k}\in L^{q,\infty}(B_1). σ k ( D 2 u ) = k 1 ( k − 1 n − 1 ) α k ( n − 2 k + k α ) ∣ x ∣ k α − 2 k ∈ L q , ∞ ( B 1 ) .
The radial flux tends to zero at the origin, so the full Hessian measure is given by this density. Since u ( r e ) − u ( 0 ) = r α u(re)-u(0)=r^\alpha u ( r e ) − u ( 0 ) = r α for every unit vector e e e , the Hölder exponent at the origin is exactly α \alpha α . The logarithmic factor in (55) is also sharp, as shown by the example
u ( x ) = x 1 log 4 ∣ x ∣ + 6 n n − k ∣ x ∣ u(x)=x_1\log\frac{4}{|x|}+\frac{6n}{n-k}|x| u ( x ) = x 1 log ∣ x ∣ 4 + n − k 6 n ∣ x ∣
in Proposition 2.
The folding construction in Proposition 3 gives a function with vanishing critical Morrey density and D u ∉ BMO l o c Du\notin\operatorname{BMO}_{\mathrm{loc}} D u ∈ / BMO loc . For k ≥ 3 k\ge3 k ≥ 3 , smoothing the fold at a variable scale gives a density in L n / k L^{n/k} L n / k while preserving the failure of gradient BMO.
Appendix. Sharpness and counterexamples
We give examples showing the sharpness and limitations of the preceding estimates. For the first example, we verify k k k -convexity and identify the full Hessian measure in detail. These arguments will also be used in the subsequent constructions.
Critical densities and logarithmic moduli.
Proposition 1. For every 2 ≤ k < n 2\le k<n 2 ≤ k < n , there is a function
u ∈ Φ k ( B 1 ) ∩ C ( B 1 ) ∩ W l o c 2 , k ( B 1 ) , F k [ u ] = f d x , u\in\Phi_k(B_1)\cap C(B_1)\cap W_{\mathrm{loc}}^{2,k}(B_1),\qquad F_k[u]=f\,dx, u ∈ Φ k ( B 1 ) ∩ C ( B 1 ) ∩ W loc 2 , k ( B 1 ) , F k [ u ] = f d x , smooth and strictly k k k -convex away from the origin, such that
f ∈ L n / k ( B 1 ) , ∣ D u ( x ) ∣ → ∞ as x → 0. f\in L^{n/k}(B_1),\qquad|Du(x)|\to\infty\text{ as }x\to0. f ∈ L n / k ( B 1 ) , ∣ D u ( x ) ∣ → ∞ as x → 0. Proof. Step 1. Set L ( r ) = log ( 4 / r ) L(r)=\log(4/r) L ( r ) = log ( 4/ r ) , g ( r ) = log L ( r ) g(r)=\log L(r) g ( r ) = log L ( r ) , H ( r ) = ∫ 0 r L ( t ) − 1 d t H(r)=\int_0^r L(t)^{-1}\,dt H ( r ) = ∫ 0 r L ( t ) − 1 d t , and A 0 = 6 n / ( n − k ) A_0=6n/(n-k) A 0 = 6 n / ( n − k ) , and set
u ( x ) = x 1 g ( ∣ x ∣ ) + A 0 H ( ∣ x ∣ ) ( x ≠ 0 ) , u ( 0 ) = 0. (57) u(x)=x_1g(|x|)+A_0H(|x|)\quad(x\ne0),\qquad u(0)=0. \tag*{(57)} u ( x ) = x 1 g ( ∣ x ∣ ) + A 0 H ( ∣ x ∣ ) ( x = 0 ) , u ( 0 ) = 0. (57) Since r log L ( r ) → 0 r\log L(r)\to0 r log L ( r ) → 0 , u u u is continuous. For e = x / r e=x/r e = x / r and β = ( r L ( r ) ) − 1 \beta=(rL(r))^{-1} β = ( r L ( r ) ) − 1 ,
D 2 H = β [ I + ( L − 1 − 1 ) e ⊗ e ] , ∥ D 2 ( x 1 g ) ∥ o p ≤ 3 β . D^2H=\beta[I+(L^{-1}-1)e\otimes e],\qquad\|D^2(x_1g)\|_{\mathrm{op}}\le3\beta. D 2 H = β [ I + ( L − 1 − 1 ) e ⊗ e ] , ∥ D 2 ( x 1 g ) ∥ op ≤ 3 β . In a radial frame, D 2 u ≥ β M 0 D^2u\ge\beta M_0 D 2 u ≥ β M 0 , where M 0 = diag ( − 3 , A 0 − 3 , … , A 0 − 3 ) M_0=\operatorname{diag}(-3,A_0-3,\ldots,A_0-3) M 0 = diag ( − 3 , A 0 − 3 , … , A 0 − 3 ) . For 1 ≤ j ≤ k 1\le j\le k 1 ≤ j ≤ k ,
σ j ( M 0 ) = ( A 0 − 3 ) j − 1 [ ( n − 1 j ) ( A 0 − 3 ) − 3 ( n − 1 j − 1 ) ] > 0 , \sigma_j(M_0)=(A_0-3)^{j-1}\left[\binom{n-1}{j}(A_0-3)-3\binom{n-1}{j-1}\right]>0, σ j ( M 0 ) = ( A 0 − 3 ) j − 1 [ ( j n − 1 ) ( A 0 − 3 ) − 3 ( j − 1 n − 1 ) ] > 0 , because A 0 > 3 n / ( n − k ) ≥ 3 n / ( n − j ) A_0>3n/(n-k)\ge3n/(n-j) A 0 > 3 n / ( n − k ) ≥ 3 n / ( n − j ) . Cone monotonicity and ∣ D 2 u ∣ ≤ C β |D^2u|\le C\beta ∣ D 2 u ∣ ≤ C β therefore yield
D 2 u ∈ Γ k , c ( r L ( r ) ) − k ≤ f : = σ k ( D 2 u ) ≤ C ( r L ( r ) ) − k . (58) D^2u\in\Gamma_k,\qquad c(rL(r))^{-k}\le f:=\sigma_k(D^2u)\le C(rL(r))^{-k}. \tag*{(58)} D 2 u ∈ Γ k , c ( r L ( r ) ) − k ≤ f := σ k ( D 2 u ) ≤ C ( r L ( r ) ) − k . (58) Step 2. Here ∣ D u ∣ ≤ g + A 0 + 1 |Du|\le g+A_0+1 ∣ D u ∣ ≤ g + A 0 + 1 and ∫ 0 1 r n − k − 1 L ( r ) − k d r < ∞ \int_0^1 r^{n-k-1}L(r)^{-k}\,dr<\infty ∫ 0 1 r n − k − 1 L ( r ) − k d r < ∞ . The inner boundary terms in integration by parts for the first and second derivatives are O ( r n g ( r ) ) O(r^ng(r)) O ( r n g ( r )) and O ( r n − 1 ( g ( r ) + 1 ) ) O(r^{n-1}(g(r)+1)) O ( r n − 1 ( g ( r ) + 1 )) , respectively. They vanish, so u ∈ W l o c 2 , k u\in W_{\mathrm{loc}}^{2,k} u ∈ W loc 2 , k with its classical Hessian.
We record the measure argument used also below. If continuous v ∈ W l o c 2 , k v \in W_{\mathrm{loc}}^{2,k} v ∈ W loc 2 , k satisfies D 2 v ∈ Γ ‾ k D^2v \in\overline{\Gamma}_k D 2 v ∈ Γ k almost everywhere, local convolution preserves this closed convex cone and converges locally uniformly to v v v . Hence v ∈ Φ k v \in\Phi_k v ∈ Φ k . Moreover,
∣ σ k ( A ) − σ k ( B ) ∣ ≤ C ( ∣ A ∣ + ∣ B ∣ ) k − 1 ∣ A − B ∣ |\sigma_k(A)-\sigma_k(B)| \le C(|A|+|B|)^{k-1}|A-B| ∣ σ k ( A ) − σ k ( B ) ∣ ≤ C ( ∣ A ∣ + ∣ B ∣ ) k − 1 ∣ A − B ∣
and strong L k L^k L k convergence of the Hessians imply strong L 1 L^1 L 1 convergence of their densities. Weak continuity of Hessian measures then gives F k [ v ] = σ k ( D 2 v ) d x F_k[v]=\sigma_k(D^2v)\,dx F k [ v ] = σ k ( D 2 v ) d x . For (1), this excludes any singular mass.
Step 3. By (2),
∫ B 1 f n / k d x ≤ C ∫ 0 1 d r r L ( r ) n < ∞ and D u = e 1 log L ( r ) + ( A 0 − x 1 r ) x r L ( r ) . \int_{B_1} f^{n/k}\,dx \le C\int_0^1 \frac{dr}{rL(r)^n}<\infty
\quad\text{and}\quad
Du=e_1\log L(r)+\left(A_0-\frac{x_1}{r}\right)\frac{x}{rL(r)}. ∫ B 1 f n / k d x ≤ C ∫ 0 1 r L ( r ) n d r < ∞ and D u = e 1 log L ( r ) + ( A 0 − r x 1 ) r L ( r ) x .
The second gradient term tends to zero, while the first diverges. □
Proposition 2. For every 2 ≤ k < n 2 \le k < n 2 ≤ k < n , there is a continuous weak k k k -convex function u u u on B 1 B_1 B 1 , smooth and strictly k k k -convex away from zero, such that F k [ u ] = f d x F_k[u]=f\,dx F k [ u ] = f d x with
0 < f ( x ) ≍ ∣ x ∣ − k , sup z ∈ B 1 / 2 , 0 < ρ < 1 / 4 F k [ u ] ( B ρ ( z ) ) ρ n − k < ∞ . 0<f(x)\asymp|x|^{-k},\qquad
\sup_{z\in B_{1/2},\,0<\rho<1/4}\frac{F_k[u](B_\rho(z))}{\rho^{n-k}}<\infty. 0 < f ( x ) ≍ ∣ x ∣ − k , z ∈ B 1/2 , 0 < ρ < 1/4 sup ρ n − k F k [ u ] ( B ρ ( z )) < ∞. The function has a local modulus C ρ log ( 4 / ρ ) C\rho\log(4/\rho) C ρ log ( 4/ ρ ) , and this order cannot be replaced by o ( ρ log ( 1 / ρ ) ) o(\rho\log(1/\rho)) o ( ρ log ( 1/ ρ )) at zero.
Proof. Set u ( x ) = x 1 log ( 4 / ∣ x ∣ ) + A 0 ∣ x ∣ u(x)=x_1\log(4/|x|)+A_0|x| u ( x ) = x 1 log ( 4/∣ x ∣ ) + A 0 ∣ x ∣ for x ≠ 0 x\ne0 x = 0 , u ( 0 ) = 0 u(0)=0 u ( 0 ) = 0 , and A 0 = 6 n / ( n − k ) A_0=6n/(n-k) A 0 = 6 n / ( n − k ) . With e = x / r e=x/r e = x / r and ξ = x 1 / r \xi=x_1/r ξ = x 1 / r ,
D 2 ( x 1 log ( 4 / r ) ) = r − 1 [ − e 1 ⊗ e − e ⊗ e 1 + ξ ( 2 e ⊗ e − I ) ] , D 2 ( A 0 r ) = A 0 r − 1 ( I − e ⊗ e ) . D^2(x_1\log(4/r))=r^{-1}[-e_1\otimes e-e\otimes e_1+\xi(2e\otimes e-I)],\qquad
D^2(A_0r)=A_0r^{-1}(I-e\otimes e). D 2 ( x 1 log ( 4/ r )) = r − 1 [ − e 1 ⊗ e − e ⊗ e 1 + ξ ( 2 e ⊗ e − I )] , D 2 ( A 0 r ) = A 0 r − 1 ( I − e ⊗ e ) . The first matrix has norm at most 3 / r 3/r 3/ r . The comparison in Proposition 1, Step 1, gives D 2 u ∈ Γ k D^2u\in\Gamma_k D 2 u ∈ Γ k and f = σ k ( D 2 u ) ≍ r − k f=\sigma_k(D^2u)\asymp r^{-k} f = σ k ( D 2 u ) ≍ r − k .
Since ∣ D 2 u ∣ ≤ C / r |D^2u|\le C/r ∣ D 2 u ∣ ≤ C / r and ∣ D u ∣ ≤ C log ( 4 / r ) |Du|\le C\log(4/r) ∣ D u ∣ ≤ C log ( 4/ r ) , Step 2 of that proof applies: u ∈ W l o c 2 , k ∩ Φ k u\in W_{\mathrm{loc}}^{2,k}\cap\Phi_k u ∈ W loc 2 , k ∩ Φ k and F k [ u ] = f d x F_k[u]=f\,dx F k [ u ] = f d x , without an atom at zero. The bound
∫ B ρ ( z ) ∣ x ∣ − k d x ≤ C ρ n − k \int_{B_\rho(z)} |x|^{-k}\,dx\le C\rho^{n-k} ∫ B ρ ( z ) ∣ x ∣ − k d x ≤ C ρ n − k is uniform in z z z : compare with B 3 ρ ( 0 ) B_{3\rho}(0) B 3 ρ ( 0 ) for ∣ z ∣ ≤ 2 ρ |z|\le2\rho ∣ z ∣ ≤ 2 ρ and use ∣ x ∣ ≥ ∣ z ∣ / 2 |x|\ge|z|/2 ∣ x ∣ ≥ ∣ z ∣/2 otherwise.
The upper modulus follows from Corollary 4.4(ii); its order is attained because u ( t e 1 ) − u ( 0 ) = t log ( 4 / t ) + A 0 t u(te_1)-u(0)=t\log(4/t)+A_0t u ( t e 1 ) − u ( 0 ) = t log ( 4/ t ) + A 0 t for t > 0 t>0 t > 0 . □
Failure of gradient BMO bounds
For a locally integrable vector field F F F , set
E F ( x , r ) = ∫ B r ( x ) ∣ F − ( F ) B r ( x ) ∣ . E_F(x,r)=\int_{B_r(x)}|F-(F)_{B_r(x)}|. E F ( x , r ) = ∫ B r ( x ) ∣ F − ( F ) B r ( x ) ∣.
The two constructions below separate critical size conditions from locally bounded mean oscillation of the gradient.
Proposition 3. (i) For every 2 ≤ k < n 2\le k<n 2 ≤ k < n , there is v ∈ Φ k ( B 1 ) ∩ C ( B 1 ) v\in\Phi_k(B_1)\cap C(B_1) v ∈ Φ k ( B 1 ) ∩ C ( B 1 ) with D v ∉ BMO l o c ( B 1 ) Dv\notin\operatorname{BMO}_{\mathrm{loc}}(B_1) D v ∈ / BMO loc ( B 1 ) and
lim ρ → 0 sup z ∈ B 1 / 2 F k [ v ] ( B ρ ( z ) ) ρ n − k = 0. (59) \lim_{\rho\to0}\sup_{z\in B_{1/2}}\frac{F_k[v](B_\rho(z))}{\rho^{n-k}}=0. \tag*{(59)} ρ → 0 lim z ∈ B 1/2 sup ρ n − k F k [ v ] ( B ρ ( z )) = 0. (59) Its measure has a bulk density and a part on { x 1 = 0 } \{x_1=0\} { x 1 = 0 } , with no atom at the origin.
(ii) For every 3 ≤ k < n 3\le k<n 3 ≤ k < n , there is u ∈ Φ k ( B 1 ) ∩ C ( B 1 ) ∩ W l o c 2 , k ( B 1 ) u\in\Phi_k(B_1)\cap C(B_1)\cap W_{\mathrm{loc}}^{2,k}(B_1) u ∈ Φ k ( B 1 ) ∩ C ( B 1 ) ∩ W loc 2 , k ( B 1 ) with F k [ u ] = f d x F_k[u]=f\,dx F k [ u ] = f d x , f ∈ L n / k ( B 1 ) f\in L^{n/k}(B_1) f ∈ L n / k ( B 1 ) , and D u ∉ BMO l o c ( B 1 ) Du\notin\operatorname{BMO}_{\mathrm{loc}}(B_1) D u ∈ / BMO loc ( B 1 ) . The function is smooth and strictly k k k -convex away from the origin.
Proof. Part (i). Use the notation
L ( r ) = log ( 4 / r ) , a ( r ) = log L ( r ) , b ( r ) = L ( r ) − 1 , H ( r ) = ∫ 0 r b ( t ) d t , A 0 = 6 n n − k . (60) L(r)=\log(4/r),\quad a(r)=\log L(r),\quad b(r)=L(r)^{-1},\quad H(r)=\int_0^r b(t)\,\mathrm{d}t,\quad A_0=\frac{6n}{n-k}. \tag*{(60)} L ( r ) = log ( 4/ r ) , a ( r ) = log L ( r ) , b ( r ) = L ( r ) − 1 , H ( r ) = ∫ 0 r b ( t ) d t , A 0 = n − k 6 n . (60)
Let g = x 1 a ( ∣ x ∣ ) g=x_1a(|x|) g = x 1 a ( ∣ x ∣ ) , h = A 0 H ( ∣ x ∣ ) h=A_0H(|x|) h = A 0 H ( ∣ x ∣ ) , and v = h + ∣ g ∣ v=h+|g| v = h + ∣ g ∣ , with v ( 0 ) = 0 v(0)=0 v ( 0 ) = 0 . Proposition 1 gives h ± g ∈ Φ k h\pm g\in\Phi_k h ± g ∈ Φ k ; their maximum v v v is therefore weakly k k k -convex and continuous.
On the half-balls, the Sobolev chain rule gives
D v = sgn ( x 1 ) e 1 a ( r ) + ( A 0 − ∣ x 1 ∣ r ) b ( r ) e , e = x / r . Dv=\operatorname{sgn}(x_1)e_1a(r)+\left(A_0-\frac{|x_1|}{r}\right)b(r)e,\qquad e=x/r. D v = sgn ( x 1 ) e 1 a ( r ) + ( A 0 − r ∣ x 1 ∣ ) b ( r ) e , e = x / r .
This gradient is locally integrable. Since v v v is even, ( D v ) B r = 0 (Dv)_{B_r}=0 ( D v ) B r = 0 ; projection onto sgn ( x 1 ) e 1 \operatorname{sgn}(x_1)e_1 sgn ( x 1 ) e 1 yields
E_{Dv}(0,r)=\fint_{B_r}|Dv|\geq a(r)=\log L(r)\to\infty.
Writing x = ( x 1 , y ) x=(x_1,y) x = ( x 1 , y ) , we claim
F_k[v]=f\,\mathrm{d}x+w(|y|)\mathcal{H}^{n-1}\mathbin{\vrule height 1.4ex depth -0.4ex width 0.07ex\vrule height 0.07ex depth -0.02ex width 0.8ex}\{x_1=0\},\qquad f(x)=\sigma_k\bigl(D^2(h+\operatorname{sgn}(x_1)g)\bigr)(x)\quad(x_1\ne0),
w ( r ) = 2 A 0 k − 1 a ( r ) ( r L ( r ) ) k − 1 [ ( n − 2 k − 1 ) + 1 L ( r ) ( n − 2 k − 2 ) ] . (61) w(r)=\frac{2A_0^{k-1}a(r)}{(rL(r))^{k-1}}\left[\binom{n-2}{k-1}+\frac{1}{L(r)}\binom{n-2}{k-2}\right]. \tag*{(61)} w ( r ) = ( r L ( r ) ) k − 1 2 A 0 k − 1 a ( r ) [ ( k − 1 n − 2 ) + L ( r ) 1 ( k − 2 n − 2 ) ] . (61)
In particular,
0 ≤ f ( x ) ≤ C ( r L ( r ) ) − k , w ( r ) ≍ log L ( r ) ( r L ( r ) ) k − 1 . 0\leq f(x)\leq C(rL(r))^{-k},\qquad w(r)\asymp\frac{\log L(r)}{(rL(r))^{k-1}}. 0 ≤ f ( x ) ≤ C ( r L ( r ) ) − k , w ( r ) ≍ ( r L ( r ) ) k − 1 log L ( r ) .
To identify the interface term, set ϕ δ ( t ) = t 2 + δ 2 − δ \phi_\delta(t)=\sqrt{t^2+\delta^2}-\delta ϕ δ ( t ) = t 2 + δ 2 − δ and v δ = h + ϕ δ ( g ) v_\delta=h+\phi_\delta(g) v δ = h + ϕ δ ( g ) . Then
D 2 v δ = B δ + ϕ δ ′ ′ ( g ) D g ⊗ D g , B δ = D 2 h + ϕ δ ′ ( g ) D 2 g , D^2v_\delta=B_\delta+\phi_\delta''(g)Dg\otimes Dg,\qquad B_\delta=D^2h+\phi_\delta'(g)D^2g, D 2 v δ = B δ + ϕ δ ′′ ( g ) D g ⊗ D g , B δ = D 2 h + ϕ δ ′ ( g ) D 2 g ,
where B δ ∈ Γ k B_\delta\in\Gamma_k B δ ∈ Γ k by convexity of the cone, and
∣ B δ ∣ ≤ C b / r , ∣ D g ∣ ≤ C a . |B_\delta|\leq Cb/r,\qquad|Dg|\leq Ca. ∣ B δ ∣ ≤ C b / r , ∣ D g ∣ ≤ C a .
For fixed δ \delta δ , these bounds and ϕ δ ′ ′ ( g ) ∣ D g ∣ 2 ≤ δ − 1 ∣ D g ∣ 2 \phi_\delta''(g)|Dg|^2\leq\delta^{-1}|Dg|^2 ϕ δ ′′ ( g ) ∣ D g ∣ 2 ≤ δ − 1 ∣ D g ∣ 2 imply v δ ∈ W l o c 2 , k v_\delta\in W_{\mathrm{loc}}^{2,k} v δ ∈ W loc 2 , k ; the logarithmic growth of its gradient makes the inner integration-by-parts terms vanish. Proposition 1, Step 2, therefore identifies its full measure with its density. Also ∣ v δ − v ∣ ≤ δ |v_\delta-v|\leq\delta ∣ v δ − v ∣ ≤ δ , so these measures converge weakly to F k [ v ] F_k[v] F k [ v ] .
The Newton tensor rank-one identity is
σ k ( D 2 v δ ) = σ k ( B δ ) + ϕ δ ′ ′ ( g ) ⟨ T k − 1 ( B δ ) D g , D g ⟩ . (62) \sigma_k(D^2v_\delta)=\sigma_k(B_\delta)+\phi_\delta''(g)\langle T_{k-1}(B_\delta)Dg,Dg\rangle. \tag*{(62)} σ k ( D 2 v δ ) = σ k ( B δ ) + ϕ δ ′′ ( g ) ⟨ T k − 1 ( B δ ) D g , D g ⟩ . (62)
On compact portions of { x 1 = 0 } ∖ { 0 } \{x_1=0\}\setminus\{0\} { x 1 = 0 } ∖ { 0 } , D g = a e 1 Dg=ae_1 D g = a e 1 , D y 2 g = 0 D_y^2g=0 D y 2 g = 0 , and ( T k − 1 ( B δ ) ) 11 = σ k − 1 ( D y 2 h ) (T_{k-1}(B_\delta))_{11}=\sigma_{k-1}(D_y^2h) ( T k − 1 ( B δ ) ) 11 = σ k − 1 ( D y 2 h ) . Work in a level-set coordinate neighborhood of the interface, away from the origin, and let ψ \psi ψ be supported there. The coarea formula writes the second term as
∫ R ϕ δ ′ ′ ( t ) ∫ { g = t } ψ ⟨ T k − 1 ( B δ ) D g , D g ⟩ ∣ D g ∣ d H n − 1 d t . \int_{\mathbb{R}}\phi_\delta''(t)\int_{\{g=t\}}\psi\frac{\langle T_{k-1}(B_\delta)Dg,Dg\rangle}{|Dg|}\,\mathrm{d}\mathcal{H}^{n-1}\,\mathrm{d}t. ∫ R ϕ δ ′′ ( t ) ∫ { g = t } ψ ∣ D g ∣ ⟨ T k − 1 ( B δ ) D g , D g ⟩ d H n − 1 d t .
As the level parameter t → 0 t\to0 t → 0 , the inner integral converges, uniformly in δ \delta δ , to the integral of ψ a σ k − 1 ( D y 2 h ) \psi a\sigma_{k-1}(D_y^2h) ψ a σ k − 1 ( D y 2 h ) over the interface. Indeed, D g → a e 1 Dg\to ae_1 D g → a e 1 , the tangential block D y 2 g → 0 D_y^2g\to0 D y 2 g → 0 , and ∣ ϕ δ ′ ∣ ≤ 1 |\phi_\delta'|\leq1 ∣ ϕ δ ′ ∣ ≤ 1 ; only that tangential block enters ( T k − 1 ) 11 (T_{k-1})_{11} ( T k − 1 ) 11 .
Since ϕ δ ′ ′ ( t ) d t → 2 δ 0 \phi_\delta''(t)\,\mathrm{d}t\to2\delta_0 ϕ δ ′′ ( t ) d t → 2 δ 0 weakly as Radon measures, the limiting surface density is 2 a σ k − 1 ( D y 2 h ) 2a\sigma_{k-1}(D_y^2h) 2 a σ k − 1 ( D y 2 h ) . The tangential eigenvalues A 0 / ( r L 2 ) A_0/(rL^2) A 0 / ( r L 2 ) and A 0 / ( r L ) A_0/(rL) A 0 / ( r L ) (multiplicity n − 2 n-2 n − 2 ) give w w w above. Dominated convergence gives the bulk term on compact sets away from zero.
The bulk and interface measures have now been identified away from zero, so the only remaining possible singular part is an atom at the origin. To exclude it, use (6):
σ k ( D 2 v δ ) ≤ C ( b / r ) k + C ( b / r ) k − 1 a 2 ϕ δ ′ ′ ( x 1 a ) . \sigma_k(D^2v_\delta)\leq C(b/r)^k+C(b/r)^{k-1}a^2\phi_\delta''(x_1a). σ k ( D 2 v δ ) ≤ C ( b / r ) k + C ( b / r ) k − 1 a 2 ϕ δ ′′ ( x 1 a ) .
Since n ≥ 3 n\geq3 n ≥ 3 ,
∫ S n − 1 ϕ δ ′ ′ ( r a θ 1 ) d θ ≤ C ∫ − 1 1 ϕ δ ′ ′ ( r a t ) d t ≤ C r a . \int_{S^{n-1}}\phi_\delta''(ra\theta_1)\,\mathrm{d}\theta\leq C\int_{-1}^{1}\phi_\delta''(rat)\,\mathrm{d}t\leq\frac{C}{ra}. ∫ S n − 1 ϕ δ ′′ ( r a θ 1 ) d θ ≤ C ∫ − 1 1 ϕ δ ′′ ( r a t ) d t ≤ r a C .
Consequently, uniformly in δ \delta δ ,
F k [ v δ ] ( B ρ ) ≤ C ∫ 0 ρ s n − k − 1 [ b ( s ) k + a ( s ) b ( s ) k − 1 ] d s ≤ C ρ n − k log L ( ρ ) L ( ρ ) k − 1 . F_k[v_\delta](B_\rho) \le C \int_0^\rho s^{n-k-1}\left[b(s)^k+a(s)b(s)^{k-1}\right]\,ds \le C\rho^{n-k}\frac{\log L(\rho)}{L(\rho)^{k-1}}. F k [ v δ ] ( B ρ ) ≤ C ∫ 0 ρ s n − k − 1 [ b ( s ) k + a ( s ) b ( s ) k − 1 ] d s ≤ C ρ n − k L ( ρ ) k − 1 log L ( ρ ) .
The last weights are increasing for sufficiently small s s s and n > k n>k n > k . The Portmanteau inequality on B ρ B_\rho B ρ gives the same bound for F k [ v ] F_k[v] F k [ v ] ; it tends to zero, so (61) is the full measure. Its surface term on ρ / 2 < ∣ y ∣ < ρ \rho/2<|y|<\rho ρ /2 < ∣ y ∣ < ρ also gives
F k [ v ] ( B ρ ) ≍ ρ n − k log L ( ρ ) L ( ρ ) k − 1 . F_k[v](B_\rho)\asymp\rho^{n-k}\frac{\log L(\rho)}{L(\rho)^{k-1}}. F k [ v ] ( B ρ ) ≍ ρ n − k L ( ρ ) k − 1 log L ( ρ ) .
The estimates
∫ B ρ ( z ) ∣ x ∣ − k d x ≤ C ρ n − k , ∫ B ρ ( z ) ∩ { x 1 = 0 } ∣ y ∣ − ( k − 1 ) d H n − 1 ≤ C ρ n − k \int_{B_\rho(z)}|x|^{-k}\,dx\le C\rho^{n-k},\qquad\int_{B_\rho(z)\cap\{x_1=0\}}|y|^{-(k-1)}\,d\mathcal{H}^{n-1}\le C\rho^{n-k} ∫ B ρ ( z ) ∣ x ∣ − k d x ≤ C ρ n − k , ∫ B ρ ( z ) ∩ { x 1 = 0 } ∣ y ∣ − ( k − 1 ) d H n − 1 ≤ C ρ n − k
hold for all centers, by the near/far split used in Proposition 2, also in dimension n − 1 n-1 n − 1 .
Given η > 0 \eta>0 η > 0 , both weights b k b^k b k and a b k − 1 ab^{k-1} a b k − 1 are at most η \eta η in a sufficiently small fixed ball. Its contribution to F k [ v ] ( B ρ ( z ) ) / ρ n − k F_k[v](B_\rho(z))/\rho^{n-k} F k [ v ] ( B ρ ( z )) / ρ n − k is at most C η C\eta C η . Outside it the volume and surface densities are bounded, contributing C η ρ k C\eta\rho^k C η ρ k and C η ρ k − 1 C\eta\rho^{k-1} C η ρ k − 1 . Letting ρ → 0 \rho\to0 ρ → 0 and then η → 0 \eta\to0 η → 0 proves (3).
Part (ii). Keep (4), g = x 1 a ( ∣ x ∣ ) g=x_1a(|x|) g = x 1 a ( ∣ x ∣ ) , and h = A 0 H ( ∣ x ∣ ) h=A_0H(|x|) h = A 0 H ( ∣ x ∣ ) . The variable smoothing width
ε ( r ) = r b ( r ) , F = ( g 2 + ε 2 ) 1 / 2 , u = h + F , u ( 0 ) = 0 \varepsilon(r)=rb(r),\qquad F=(g^2+\varepsilon^2)^{1/2},\qquad u=h+F,\qquad u(0)=0 ε ( r ) = r b ( r ) , F = ( g 2 + ε 2 ) 1/2 , u = h + F , u ( 0 ) = 0
removes the interface measure. Since F ≤ r ( a + b ) F\le r(a+b) F ≤ r ( a + b ) , u u u is continuous and smooth away from zero. Also ε ′ = b + b 2 \varepsilon'=b+b^2 ε ′ = b + b 2 and ε ′ ′ = ( b 2 + 2 b 3 ) / r \varepsilon''=(b^2+2b^3)/r ε ′′ = ( b 2 + 2 b 3 ) / r , so D 2 ε ≥ 0 D^2\varepsilon\ge0 D 2 ε ≥ 0 and ∣ D 2 ε ∣ ≤ C b / r |D^2\varepsilon|\le Cb/r ∣ D 2 ε ∣ ≤ C b / r . Set τ = g / F \tau=g/F τ = g / F and η = ε / F \eta=\varepsilon/F η = ε / F ; then ∣ τ ∣ ≤ 1 |\tau|\le1 ∣ τ ∣ ≤ 1 , 0 ≤ η ≤ 1 0\le\eta\le1 0 ≤ η ≤ 1 , and τ 2 + η 2 = 1 \tau^2+\eta^2=1 τ 2 + η 2 = 1 .
The Hessian splits into an admissible base matrix and a nonnegative rank-one term:
D 2 u = B + Q , B = D 2 h + τ D 2 g + η D 2 ε , Q = ( ε D g − g D ε ) ⊗ 2 F 3 . D^2u=B+Q,\qquad B=D^2h+\tau D^2g+\eta D^2\varepsilon,\qquad Q=\frac{(\varepsilon Dg-gD\varepsilon)^{\otimes2}}{F^3}. D 2 u = B + Q , B = D 2 h + τ D 2 g + η D 2 ε , Q = F 3 ( ε D g − g D ε ) ⊗ 2 .
Here Q ≥ 0 Q\ge0 Q ≥ 0 , η ≥ 0 \eta\ge0 η ≥ 0 , and the first two terms of B B B are a convex combination of D 2 ( h ± g ) D^2(h\pm g) D 2 ( h ± g ) . Hence
B , B + Q ∈ Γ k , ∣ B ∣ ≤ C b / r . B,B+Q\in\Gamma_k,\qquad|B|\le Cb/r. B , B + Q ∈ Γ k , ∣ B ∣ ≤ C b / r .
For e = x / r e=x/r e = x / r and ξ = x 1 / r \xi=x_1/r ξ = x 1 / r , D g = a e 1 − ξ b e Dg=ae_1-\xi be D g = a e 1 − ξ b e and D ε = ( b + b 2 ) e D\varepsilon=(b+b^2)e D ε = ( b + b 2 ) e , whence
ε D g − g D ε = r a b [ e 1 − ξ ( 1 + b + b a ) e ] . \varepsilon Dg-gD\varepsilon=r ab\left[e_1-\xi\left(1+b+\frac{b}{a}\right)e\right]. ε D g − g D ε = r ab [ e 1 − ξ ( 1 + b + a b ) e ] .
The bracket is bounded, and F = r b ( 1 + a 2 ξ 2 / b 2 ) 1 / 2 F=rb(1+a^2\xi^2/b^2)^{1/2} F = r b ( 1 + a 2 ξ 2 / b 2 ) 1/2 , so
∣ Q ∣ ≤ C a 2 r b ( 1 + a 2 ξ 2 b 2 ) − 3 / 2 . |Q|\le\frac{Ca^2}{rb}\left(1+\frac{a^2\xi^2}{b^2}\right)^{-3/2}. ∣ Q ∣ ≤ r b C a 2 ( 1 + b 2 a 2 ξ 2 ) − 3/2 .
The largest second derivatives lie in the angular strip ∣ x 1 ∣ / r ≲ b / a |x_1|/r\lesssim b/a ∣ x 1 ∣/ r ≲ b / a . Its width provides the gain in the next integral.
For p = n / k > 1 p=n/k>1 p = n / k > 1 , the spherical density in ξ \xi ξ is bounded (n ≥ 3 n\ge3 n ≥ 3 ). Integration over ξ \xi ξ , followed by s = a ξ / b s=a\xi/b s = a ξ / b , gives
∫ S n − 1 ∣ Q ∣ p d θ ≤ C ( a 2 r b ) p ∫ − 1 1 ( 1 + a 2 ξ 2 b 2 ) − 3 p / 2 d ξ ≤ C ( a 2 r b ) p b a . \int_{\mathbb{S}^{n-1}}|Q|^p\,d\theta\le C\left(\frac{a^2}{rb}\right)^p\int_{-1}^1\left(1+\frac{a^2\xi^2}{b^2}\right)^{-3p/2}\,d\xi\le C\left(\frac{a^2}{rb}\right)^p\frac{b}{a}. ∫ S n − 1 ∣ Q ∣ p d θ ≤ C ( r b a 2 ) p ∫ − 1 1 ( 1 + b 2 a 2 ξ 2 ) − 3 p /2 d ξ ≤ C ( r b a 2 ) p a b .
Since Q Q Q has rank one, it enters σ k ( B + Q ) \sigma_k(B+Q) σ k ( B + Q ) only linearly. The rank-one identity therefore bounds the classical density by
0 < f = σ k ( D 2 u ) ≤ C ( b / r ) k + C ( b / r ) k − 1 ∣ Q ∣ . 0<f=\sigma_k(D^2u)\le C(b/r)^k+C(b/r)^{k-1}|Q|. 0 < f = σ k ( D 2 u ) ≤ C ( b / r ) k + C ( b / r ) k − 1 ∣ Q ∣.
Since k p = n kp=n k p = n , it follows that
∫ B 1 f p d x ≤ C ∫ 0 1 [ b k p + a 2 p − 1 b ( k − 2 ) p + 1 ] d r r = C ∫ log 4 ∞ [ L − n + ( log L ) 2 p − 1 L − ( ( k − 2 ) p + 1 ) ] d L < ∞ . \int_{B_1}f^p\,dx\le C\int_0^1\left[b^{kp}+a^{2p-1}b^{(k-2)p+1}\right]\frac{dr}{r}
=C\int_{\log4}^{\infty}\left[L^{-n}+(\log L)^{2p-1}L^{-((k-2)p+1)}\right]\,dL<\infty. ∫ B 1 f p d x ≤ C ∫ 0 1 [ b k p + a 2 p − 1 b ( k − 2 ) p + 1 ] r d r = C ∫ l o g 4 ∞ [ L − n + ( log L ) 2 p − 1 L − (( k − 2 ) p + 1 ) ] d L < ∞.
The last inequality uses ( k − 2 ) p + 1 > 1 (k-2)p+1>1 ( k − 2 ) p + 1 > 1 , precisely where k ≥ 3 k\ge3 k ≥ 3 is needed. The coarse bound ∣ D 2 u ∣ ≤ C r − 1 L ( r ) ( log L ( r ) ) 2 |D^2u|\le Cr^{-1}L(r)(\log L(r))^2 ∣ D 2 u ∣ ≤ C r − 1 L ( r ) ( log L ( r ) ) 2 is in L k L^k L k because n > k n>k n > k . Also ∣ D u ∣ ≤ a + C b |Du|\le a+Cb ∣ D u ∣ ≤ a + C b and ∣ u ∣ ≤ C r ( a + b ) |u|\le Cr(a+b) ∣ u ∣ ≤ C r ( a + b ) .
The inner boundary terms for the first and second weak derivatives are respectively O ( r n ( a + b ) ) O(r^{n}(a+b)) O ( r n ( a + b )) and O ( r n − 1 ( a + b ) ) O(r^{n-1}(a+b)) O ( r n − 1 ( a + b )) , and vanish. Thus u ∈ W l o c 2 , k u \in W_{\mathrm{loc}}^{2,k} u ∈ W loc 2 , k . The measure argument in Proposition 1, Step 2, now gives u ∈ Φ k u \in\Phi_k u ∈ Φ k and F k [ u ] = f d x F_k[u] = f\,dx F k [ u ] = f d x , including at the origin.
Finally, u u u is even, ( D u ) B r = 0 (Du)_{B_r}=0 ( D u ) B r = 0 , and
D u = τ a e 1 + [ A 0 b − τ ξ b + η ( b + b 2 ) ] e . Du=\tau ae_1+[A_0b-\tau\xi b+\eta(b+b^2)]e. D u = τ a e 1 + [ A 0 b − τ ξ b + η ( b + b 2 )] e .
The bracket is at least ( A 0 − 1 ) b > 0 (A_0-1)b>0 ( A 0 − 1 ) b > 0 . On the fixed angular set ∣ ξ ∣ ≥ 1 / 2 |\xi|\ge1/2 ∣ ξ ∣ ≥ 1/2 , ∣ τ ∣ ≥ 1 / 2 |\tau|\ge1/2 ∣ τ ∣ ≥ 1/2 at small radii since a / b → ∞ a/b\to\infty a / b → ∞ . Projection onto sgn ( ξ ) e 1 \operatorname{sgn}(\xi)e_1 sgn ( ξ ) e 1 therefore gives ∣ D u ∣ ≥ a / 2 |Du|\ge a/2 ∣ D u ∣ ≥ a /2 . This set has a fixed positive relative area and a ( s ) ≥ a ( r ) a(s)\ge a(r) a ( s ) ≥ a ( r ) for s < r s<r s < r , so
E D u ( 0 , r ) ≥ c log L ( r ) → ∞ . E_{Du}(0,r)\ge c\log L(r)\to\infty. E D u ( 0 , r ) ≥ c log L ( r ) → ∞.
Limitations of the potential estimate
Proposition 4. For every 2 ≤ k < n 2\le k<n 2 ≤ k < n :
(i) There is a convex v ∈ C 1 ( B 1 ) v\in C^1(B_1) v ∈ C 1 ( B 1 ) with bounded gradient whose potential W k R [ F k [ v ] ] W_k^R[F_k[v]] W k R [ F k [ v ]] is unbounded near zero, for every fixed 0 < R < 1 / 2 0<R<1/2 0 < R < 1/2 .
(ii) The Wolff term in (1.1) cannot be replaced by ( I k R [ F k [ u ] ] ) 1 / k (I_k^R[F_k[u]])^{1/k} ( I k R [ F k [ u ]] ) 1/ k , where
I k R [ μ ] ( x ) = ∫ 0 R μ ( B t ( x ) ) t n − k d t t . I_k^R[\mu](x)=\int_0^R\frac{\mu(B_t(x))}{t^{n-k}}\frac{dt}{t}. I k R [ μ ] ( x ) = ∫ 0 R t n − k μ ( B t ( x )) t d t .
Proof. (i) Use L ( r ) = log ( 4 / r ) L(r)=\log(4/r) L ( r ) = log ( 4/ r ) and H ( r ) = ∫ 0 r L ( t ) − 1 d t H(r)=\int_0^r L(t)^{-1}\,dt H ( r ) = ∫ 0 r L ( t ) − 1 d t from Proposition 1, and set v ( x ) = H ( ∣ x ∣ ) v(x)=H(|x|) v ( x ) = H ( ∣ x ∣ ) . Then ∣ D v ∣ = 1 / L ( r ) → 0 |Dv|=1/L(r)\to0 ∣ D v ∣ = 1/ L ( r ) → 0 , and the radial and tangential Hessian eigenvalues are ( r L 2 ) − 1 (rL^2)^{-1} ( r L 2 ) − 1 and ( r L ) − 1 (rL)^{-1} ( r L ) − 1 . Also H ( r ) / r ≤ 1 / L ( r ) → 0 H(r)/r\le1/L(r)\to0 H ( r ) / r ≤ 1/ L ( r ) → 0 , so v v v is convex and C 1 C^1 C 1 , with D v ( 0 ) = 0 Dv(0)=0 D v ( 0 ) = 0 .
The full-measure identification in the proof of Proposition 1 applies, giving F k [ v ] = f v d x F_k[v]=f_v\,dx F k [ v ] = f v d x with f v ≍ ( r L ) − k f_v\asymp(rL)^{-k} f v ≍ ( r L ) − k and no mass at zero. Hence
F k [ v ] ( B t ) ≍ t n − k L ( t ) − k . F_k[v](B_t)\asymp t^{n-k}L(t)^{-k}. F k [ v ] ( B t ) ≍ t n − k L ( t ) − k .
In particular the potential is infinite at zero; for 0 < ∣ x ∣ < R / 4 0<|x|<R/4 0 < ∣ x ∣ < R /4 , the inclusion B t / 2 ( 0 ) ⊂ B t ( x ) B_{t/2}(0)\subset B_t(x) B t /2 ( 0 ) ⊂ B t ( x ) for t ≥ 2 ∣ x ∣ t\ge2|x| t ≥ 2∣ x ∣ gives
W k R [ F k [ v ] ] ( x ) ≥ c ∫ 2 ∣ x ∣ R d t t L ( t ) → ∞ ( x → 0 ) . W_k^R[F_k[v]](x)\ge c\int_{2|x|}^{R}\frac{dt}{tL(t)}\to\infty\qquad(x\to0). W k R [ F k [ v ]] ( x ) ≥ c ∫ 2∣ x ∣ R t L ( t ) d t → ∞ ( x → 0 ) .
(ii) Take u u u from Proposition 1. Its full measure is f d x f\,dx f d x with f ( y ) ≤ C ( ∣ y ∣ L ( ∣ y ∣ ) ) − k f(y)\le C(|y|L(|y|))^{-k} f ( y ) ≤ C ( ∣ y ∣ L ( ∣ y ∣ ) ) − k . Tonelli’s theorem gives
I k 1 / 2 [ f d x ] ( x ) ≤ C ∫ B 1 f ( y ) ∣ x − y ∣ k − n d y . I_k^{1/2}[f\,dx](x)\le C\int_{B_1}f(y)|x-y|^{k-n}\,dy. I k 1/2 [ f d x ] ( x ) ≤ C ∫ B 1 f ( y ) ∣ x − y ∣ k − n d y .
For 0 < d = ∣ x ∣ < 1 / 4 0<d=|x|<1/4 0 < d = ∣ x ∣ < 1/4 , split the last integral into ∣ y ∣ < d / 2 |y|<d/2 ∣ y ∣ < d /2 , d / 2 ≤ ∣ y ∣ ≤ 2 d d/2\le|y|\le2d d /2 ≤ ∣ y ∣ ≤ 2 d , and ∣ y ∣ > 2 d |y|>2d ∣ y ∣ > 2 d . The first two contributions are each at most C L ( d ) − k CL(d)^{-k} C L ( d ) − k : use respectively ∣ x − y ∣ ≥ d / 2 |x-y|\ge d/2 ∣ x − y ∣ ≥ d /2 and f ( y ) ≤ C d − k L ( d ) − k f(y)\le Cd^{-k}L(d)^{-k} f ( y ) ≤ C d − k L ( d ) − k together with
∫ B 3 d ( x ) ∣ x − y ∣ k − n d y ≤ C d k . \int_{B_{3d}(x)}|x-y|^{k-n}\,dy\le Cd^k. ∫ B 3 d ( x ) ∣ x − y ∣ k − n d y ≤ C d k .
The third contribution is bounded by
C ∫ 2 d 1 d t t L ( t ) k . C\int_{2d}^{1}\frac{dt}{tL(t)^k}. C ∫ 2 d 1 t L ( t ) k d t .
This integral remains finite with lower limit zero, because k > 1 k>1 k > 1 , also covering x = 0 x=0 x = 0 . Therefore
sup x ∈ B 1 / 4 I k 1 / 2 [ F k [ u ] ] ( x ) < ∞ . \sup_{x\in B_{1/4}}I_k^{1/2}[F_k[u]](x)<\infty. x ∈ B 1/4 sup I k 1/2 [ F k [ u ]] ( x ) < ∞.
For the fixed radius R = 1 / 2 R=1/2 R = 1/2 , the average term in (1.1) is uniformly bounded as the center tends to zero, whereas ∣ D u ( x ) ∣ → ∞ |Du(x)|\to\infty ∣ D u ( x ) ∣ → ∞ through nonzero gradient Lebesgue points. This contradicts the proposed replacement. □
The following sequence shows why an interior estimate cannot be extended uniformly to the boundary using only the same measure data and the height of the solution.
Proof. Write a = t + ε a=t+\varepsilon a = t + ε , suppress subscripts, and set F = L 1 / ( 4 k ) F=L^{1/(4k)} F = L 1/ ( 4 k ) . Since L > 3 L>3 L > 3 ,
F ′ = − F 4 k a L , F ′ ′ = ∣ F ′ ∣ a ( 1 − 1 − 1 / ( 4 k ) L ) , 2 ( F ′ ) 2 F < F ′ ′ ≤ ∣ F ′ ∣ a . F'=-\frac{F}{4kaL},\qquad F''=\frac{\lvert F'\rvert}{a}\left(1-\frac{1-1/(4k)}{L}\right),\qquad\frac{2(F')^2}{F}<F''\leq\frac{\lvert F'\rvert}{a}. F ′ = − 4 k a L F , F ′′ = a ∣ F ′ ∣ ( 1 − L 1 − 1/ ( 4 k ) ) , F 2 ( F ′ ) 2 < F ′′ ≤ a ∣ F ′ ∣ .
In the orthogonal coordinates ( y , t ) (y,t) ( y , t ) , the Hessian has blocks
D 2 w ε = ( 2 F I n − 1 2 F ′ y 2 F ′ y T c ) , c = 2 F + 4 ( t − 1 ) F ′ + ( ∣ y ∣ 2 − 2 t + t 2 ) F ′ ′ . D^2w_\varepsilon=\begin{pmatrix}2F I_{n-1}&2F'y\\2F'y^{T}&c\end{pmatrix},\qquad c=2F+4(t-1)F'+(\lvert y\rvert^2-2t+t^2)F''. D 2 w ε = ( 2 F I n − 1 2 F ′ y T 2 F ′ y c ) , c = 2 F + 4 ( t − 1 ) F ′ + (∣ y ∣ 2 − 2 t + t 2 ) F ′′ .
Its Schur complement satisfies
S = 2 F + 4 ( t − 1 ) F ′ − t ( 2 − t ) F ′ ′ + ∣ y ∣ 2 ( F ′ ′ − 2 ( F ′ ) 2 F ) S=2F+4(t-1)F'-t(2-t)F''+\lvert y\rvert^2\left(F''-\frac{2(F')^2}{F}\right) S = 2 F + 4 ( t − 1 ) F ′ − t ( 2 − t ) F ′′ + ∣ y ∣ 2 ( F ′′ − F 2 ( F ′ ) 2 )
≥ 2 F + ∣ F ′ ∣ a [ 4 ( 1 − t ) a − t ( 2 − t ) ] ≥ 2 F − 2 a ∣ F ′ ∣ = 2 F ( 1 − 1 4 k L ) > 0. \geq2F+\frac{\lvert F'\rvert}{a}\left[4(1-t)a-t(2-t)\right]\geq2F-2a\lvert F'\rvert=2F\left(1-\frac{1}{4kL}\right)>0. ≥ 2 F + a ∣ F ′ ∣ [ 4 ( 1 − t ) a − t ( 2 − t ) ] ≥ 2 F − 2 a ∣ F ′ ∣ = 2 F ( 1 − 4 k L 1 ) > 0.
Here
4 ( 1 − t ) a − t ( 2 − t ) + 2 a 2 = t ( 2 − t ) + 4 ε + 2 ε 2 ≥ 0. 4(1-t)a-t(2-t)+2a^2=t(2-t)+4\varepsilon+2\varepsilon^2\geq0. 4 ( 1 − t ) a − t ( 2 − t ) + 2 a 2 = t ( 2 − t ) + 4 ε + 2 ε 2 ≥ 0.
Hence w ε w_\varepsilon w ε is strictly convex.
Expanding the principal minors, the off-diagonal contribution is nonpositive. Since c ≤ 2 F + 4 ∣ F ′ ∣ c\leq2F+4\lvert F'\rvert c ≤ 2 F + 4 ∣ F ′ ∣ , F k = L 1 / 4 F^k=L^{1/4} F k = L 1/4 , and F k − 1 ∣ F ′ ∣ = 1 / ( 4 k a L 3 / 4 ) F^{k-1}\lvert F'\rvert=1/(4kaL^{3/4}) F k − 1 ∣ F ′ ∣ = 1/ ( 4 k a L 3/4 ) , we obtain
0 < f ε 0 : = σ k ( D 2 w ε ) ≤ C ( F k + F k − 1 ∣ F ′ ∣ ) ≤ C a L 3 / 4 ≤ C t L 0 ( t ) 3 / 4 . 0<f_\varepsilon^0:=\sigma_k(D^2w_\varepsilon)\leq C(F^k+F^{k-1}\lvert F'\rvert)\leq\frac{C}{aL^{3/4}}\leq\frac{C}{tL_0(t)^{3/4}}. 0 < f ε 0 := σ k ( D 2 w ε ) ≤ C ( F k + F k − 1 ∣ F ′ ∣) ≤ a L 3/4 C ≤ t L 0 ( t ) 3/4 C .
Here a L aL a L is bounded, a L 3 / 4 aL^{3/4} a L 3/4 is increasing on ( 0 , 9 / 4 ] (0,9/4] ( 0 , 9/4 ] , and L 0 ( t ) = log ( e 4 / t ) L_0(t)=\log(e^4/t) L 0 ( t ) = log ( e 4 / t ) . The horizontal cross-section of B 1 B_1 B 1 has measure ω n − 1 ( 2 t − t 2 ) ( n − 1 ) / 2 \omega_{n-1}(2t-t^2)^{(n-1)/2} ω n − 1 ( 2 t − t 2 ) ( n − 1 ) /2 . Consequently, writing q = q ∂ q=q_\partial q = q ∂ ,
sup ε ∫ B 1 ( f ε 0 ) q d x ≤ C ( 1 + ∫ 0 1 d t t L 0 ( t ) 3 q / 4 ) < ∞ , \sup_\varepsilon\int_{B_1}(f_\varepsilon^0)^q\,dx\leq C\left(1+\int_0^1\frac{dt}{tL_0(t)^{3q/4}}\right)<\infty, ε sup ∫ B 1 ( f ε 0 ) q d x ≤ C ( 1 + ∫ 0 1 t L 0 ( t ) 3 q /4 d t ) < ∞ ,
because q ≥ 2 q\geq2 q ≥ 2 .
Set c ε = F ( 0 ) − 1 / 2 c_\varepsilon=F(0)^{-1/2} c ε = F ( 0 ) − 1/2 , u ε = c ε w ε u_\varepsilon=c_\varepsilon w_\varepsilon u ε = c ε w ε , and f ε = c ε k f ε 0 f_\varepsilon=c_\varepsilon^k f_\varepsilon^0 f ε = c ε k f ε 0 . The bound ∣ w ε ∣ ≤ 2 t L 0 ( t ) 1 / ( 4 k ) \lvert w_\varepsilon\rvert\leq2tL_0(t)^{1/(4k)} ∣ w ε ∣ ≤ 2 t L 0 ( t ) 1/ ( 4 k ) is uniform, so c ε → 0 c_\varepsilon\to0 c ε → 0 gives vanishing height and density norms, while
∂ ν u ε ( e n ) = 2 F ( 0 ) 1 / 2 → ∞ . \partial_\nu u_\varepsilon(e_n)=2F(0)^{1/2}\to\infty. ∂ ν u ε ( e n ) = 2 F ( 0 ) 1/2 → ∞.
Since q > n / k q>n/k q > n / k , Hölder’s inequality gives
sup x ∈ R n W k r [ f ε 1 B 1 d x ] ( x ) ≤ C ∥ f ε ∥ L q ( B 1 ) 1 / k r 1 − n / ( k q ) , 0 < r ≤ 4. \sup_{x\in\mathbb{R}^n}\mathcal{W}_k^r[f_\varepsilon\mathbf{1}_{B_1}\,dx](x)\leq C\lVert f_\varepsilon\rVert_{L^q(B_1)}^{1/k}r^{1-n/(kq)},\qquad0<r\leq4. x ∈ R n sup W k r [ f ε 1 B 1 d x ] ( x ) ≤ C ∥ f ε ∥ L q ( B 1 ) 1/ k r 1 − n / ( k q ) , 0 < r ≤ 4.
Taking r = 4 r=4 r = 4 proves the vanishing potential norms. The positive power of r r r , together with the uniform density bound, also gives uniformly vanishing small-scale tails. Letting ε → 0 \varepsilon\to0 ε → 0 along any sequence proves the claim. □
Declaration of AI-assisted research and writing. We used ChatGPT and Codex (OpenAI) to assist with mathematical exploration and with drafting and revising this manuscript. In particular, these tools assisted with some counterexample constructions and with the design of the two-point test function used in the proof of Proposition 2.8. The authors take full responsibility for the mathematical content and the final manuscript.
References
(Huajie Liu) School of Mathematical Sciences, Peking University, Beijing, 100871, China Email address: huajieliu@math.pku.edu.cn
(Shiguang Ma) School of Mathematical Sciences, Nankai University, Tianjin, 300071, China Email address: msgdyx8741@nankai.edu.cn
(Shengyang Zang) School of Mathematical Sciences, Nankai University, Tianjin, 300071, China Email address: 1120230032@mail.nankai.edu.cn