Introduction
Let D ⊂ R n D \subset\mathbb{R}^{n} D ⊂ R n be an open domain, and consider a continuous nonnegative function u : D → [ 0 , ∞ ) u : D \to[0,\infty) u : D → [ 0 , ∞ ) with positivity set (positive phase) Ω = { x ∈ D : u ( x ) > 0 } \Omega= \{x \in D : u(x) > 0\} Ω = { x ∈ D : u ( x ) > 0 } and free boundary FB ( u ) = ∂ Ω ∩ D \operatorname{FB}(u) = \partial\Omega\cap D FB ( u ) = ∂ Ω ∩ D . We say that u u u is a classical solution to the one-phase Bernoulli free boundary problem in D D D if FB ( u ) \operatorname{FB}(u) FB ( u ) is a smooth hypersurface around which Ω \Omega Ω is locally a smooth one-sided domain, and u u u satisfies the overdetermined boundary value problem
Δ u = 0 in Ω , u = 0 , ∣ ∇ u ∣ = 1 on FB ( u ) , (1) \Delta u = 0 \quad\text{in } \Omega,\qquad u = 0,\qquad|\nabla u| = 1 \quad\text{on } \operatorname{FB}(u), \tag*{(1)} Δ u = 0 in Ω , u = 0 , ∣∇ u ∣ = 1 on FB ( u ) , (1)
where ∇ u \nabla u ∇ u along FB ( u ) \operatorname{FB}(u) FB ( u ) denotes the non-tangential trace from Ω \Omega Ω . Introduced in its modern variational formulation by Alt and Caffarelli [1 ] , (1) describes the Euler–Lagrange equilibrium of the functional
J ( u ; D ) = ∫ D ( ∣ ∇ u ∣ 2 + 1 Ω ) d x J(u;D) = \int_D \left(|\nabla u|^2 + \mathbf{1}_{\Omega}\right)\,dx J ( u ; D ) = ∫ D ( ∣∇ u ∣ 2 + 1 Ω ) d x
under compactly supported inner domain variations. Denoting by ν = − ∇ u \nu= -\nabla u ν = − ∇ u the outward unit normal to Ω \Omega Ω along FB ( u ) \operatorname{FB}(u) FB ( u ) and by H = − div FB ( u ) ν H = -\operatorname{div}_{\operatorname{FB}(u)} \nu H = − div FB ( u ) ν the corresponding scalar mean curvature (signed so that H = n − 1 > 0 H = n - 1 > 0 H = n − 1 > 0 on the unit sphere bounding the exterior-ball domain R n ∖ B ‾ 1 \mathbb{R}^{n} \setminus\overline{B}_{1} R n ∖ B 1 ), we use the scalar second-variation form of [8 Lemma 1] in the convention of [15 (1.2)]
Q u [ ϕ ; D ] = ∫ Ω ∣ ∇ ϕ ∣ 2 d x − ∫ FB ( u ) H ϕ 2 d H n − 1 . Q_u[\phi;D] = \int_{\Omega} |\nabla\phi|^2\,dx - \int_{\operatorname{FB}(u)} H\phi^2\,d\mathcal{H}^{n-1}. Q u [ ϕ ; D ] = ∫ Ω ∣∇ ϕ ∣ 2 d x − ∫ FB ( u ) H ϕ 2 d H n − 1 .
Here admissible test functions ϕ \phi ϕ are restrictions to Ω ∩ D ‾ \overline{\Omega\cap D} Ω ∩ D of smooth functions in C ∞ ( D ) C^\infty(D) C ∞ ( D ) , which need not vanish on FB ( u ) \operatorname{FB}(u) FB ( u ) . The Morse index ind ( u ; D ) \operatorname{ind}(u;D) ind ( u ; D ) is defined as the supremum of the dimensions of subspaces of admissible test functions on which Q u [ ⋅ ; D ] Q_u[\cdot;D] Q u [ ⋅ ; D ] is negative definite; when D = R n D = \mathbb{R}^{n} D = R n , we simply write ind ( u ) \operatorname{ind}(u) ind ( u ) . In particular, u u u is stable in D D D if ind ( u ; D ) = 0 \operatorname{ind}(u;D) = 0 ind ( u ; D ) = 0 , that is, Q u [ ϕ ; D ] ≥ 0 Q_u[\phi;D] \ge0 Q u [ ϕ ; D ] ≥ 0 for all admissible ϕ \phi ϕ .
The one-phase problem has antecedents in free-streamline models in fluid mechanics. The periodic hollow vortices constructed by Baker, Saffman and Sheffield [2 ] provide a classical example of this connection. Alt and Caffarelli [1 ] introduced the modern variational framework and established existence and regularity for minimizers. Caffarelli’s Harnack-inequality approach subsequently established regularity of Lipschitz free boundaries and a flatness-to-Lipschitz theorem [6 , 7 ] . We refer to [9 ] for the geometric theory of these problems. Weiss [23 , 24 ] developed the monotonicity formula and used one-homogeneous minimizing blow-ups to relate the dimension of the singular set to the first dimension admitting a nonflat minimizing cone.
For one-homogeneous stable solutions with free boundary smooth away from the vertex, the rigidity results of Caffarelli–Jerison–Kenig [8 ] and Jerison–Savin [17 ] give flatness through dimensions three and four, respectively. The classification of these smooth stable cones in dimensions five and six follows from the analysis underlying the minimizing-cone theorem in [21 Sections 3–7] , as recorded in [14 Theorem 3 and Remark 4] . In dimension seven, De Silva and Jerison [12 ] constructed a nonflat minimizing cone. The foliation theorem of Edelen, Spolaor and Velichkov [13 Theorem 1.5] gives smooth entire minimizing leaves with that cone as a blow-down.
A central problem in free boundary regularity—parallel to the Bernstein problem for minimal hypersurfaces and De Giorgi’s conjecture for the Allen–Cahn equation—is the global stability conjecture for (1) (see [14 Conjecture 6] ):
Up to which ambient dimension must every entire classical stable solution u : R n → [ 0 , ∞ ) u:\mathbb{R}^{n}\to[0,\infty) u : R n → [ 0 , ∞ ) be one-dimensional?
In dimension n = 2 n=2 n = 2 , rigidity follows from a logarithmic cutoff argument [14 discussion following Conjecture 6] . Chan, Fernández-Real, Figalli and Serra [10 ] resolved the conjecture in dimension n = 3 n=3 n = 3 , and Fernández-Real and Serra [16 ] established the case n = 4 n=4 n = 4 using a multiscale signed decomposition adapted to non-minimizing two-sheeted configurations. The resulting gap between the smooth stable cone classification and the entire stable classification is in dimensions five and six.
Main results
Our first objective is to close this dimensional gap completely. To separate our analytic PDE reduction cleanly from the external cone-classification input, recall following [16 Definition 4.1] that the critical cone dimension n c o n e ∗ n_{\mathrm{cone}}^* n cone ∗ is the smallest integer n n n admitting a nonzero, non-half-space one-homogeneous global stable solution that is classical away from the origin with 0 0 0 on its free boundary; similarly, let n e n t i r e ∗ n_{\mathrm{entire}}^* n entire ∗ denote the smallest dimension admitting a non-one-dimensional entire classical stable solution. Our first main theorem shows that cone rigidity transfers unconditionally to entire rigidity in all dimensions below n c o n e ∗ n_{\mathrm{cone}}^* n cone ∗ .
Theorem 1.1. If 4 ≤ n < n c o n e ∗ 4\le n<n_{\mathrm{cone}}^* 4 ≤ n < n cone ∗ , then every entire classical stable solution u u u to (1) in R n \mathbb{R}^{n} R n satisfies D 2 u ≡ 0 D^2u\equiv0 D 2 u ≡ 0 in its positive phase Ω \Omega Ω and is one-dimensional.
The stable-cone classification through dimension six [14 Theorem 3 and Remark 4] and the seven-dimensional De Silva–Jerison cone [12 ] give n c o n e ∗ = 7 n_{\mathrm{cone}}^*=7 n cone ∗ = 7 . Combining Theorem 1.1 with the two- and three-dimensional results [10 , 14 ] yields the full classification of entire classical stable solutions through dimension six.
Corollary 1.2. Let 2 ≤ n ≤ 6 2\le n\le6 2 ≤ n ≤ 6 . Every entire classical stable solution u u u to (1) in R n \mathbb{R}^{n} R n is either a nonnegative constant or, for some unit vector e ∈ S n − 1 e\in\mathbb{S}^{n-1} e ∈ S n − 1 , one of the one-dimensional half-space or two-plane profiles
( x ⋅ e − a ) + ( a ∈ R ) , or ( x ⋅ e − a ) + + ( b − x ⋅ e ) + ( b < a ) . (x\cdot e-a)_+ \qquad(a\in\mathbb{R}), \qquad\text{or} \qquad(x\cdot e-a)_+ +(b-x\cdot e)_+ \qquad(b<a). ( x ⋅ e − a ) + ( a ∈ R ) , or ( x ⋅ e − a ) + + ( b − x ⋅ e ) + ( b < a ) . In particular, D 2 u ≡ 0 D^2u\equiv0 D 2 u ≡ 0 in Ω \Omega Ω , and n e n t i r e ∗ = 7 n_{\mathrm{entire}}^*=7 n entire ∗ = 7 is the sharp first dimension admitting a nonflat entire classical stable solution (see Corollary 2.4 below).
By a point-picking blow-up argument, entire rigidity immediately implies universal, scale-invariant interior bounds on both the positive-phase Hessian and the second fundamental form II of the free boundary. The connection between entire rigidity and local curvature estimates is already present in the conditional estimate of Kamburov and Wang [18 ] .
Corollary 1.3. Let 2 ≤ n ≤ 6 2 \le n \le6 2 ≤ n ≤ 6 . There exists a dimensional constant C n > 0 C_n > 0 C n > 0 such that if u u u is a classical stable solution to (1) in B R ( z ) ⊂ R n B_R(z) \subset\mathbb{R}^n B R ( z ) ⊂ R n with z ∈ F B ( u ) z \in\mathrm{FB}(u) z ∈ FB ( u ) , then
∥ D 2 u ∥ L ∞ ( B R / 2 ( z ) ∩ { u > 0 } ) + ∥ I I ∥ L ∞ ( B R / 2 ( z ) ∩ F B ( u ) ) ≤ C n R − 1 , \lVert D^2u\rVert_{L^\infty(B_{R/2}(z)\cap\{u>0\})}+\lVert\mathrm{II}\rVert_{L^\infty(B_{R/2}(z)\cap\mathrm{FB}(u))}\le C_nR^{-1}, ∥ D 2 u ∥ L ∞ ( B R /2 ( z ) ∩ { u > 0 }) + ∥ II ∥ L ∞ ( B R /2 ( z ) ∩ FB ( u )) ≤ C n R − 1 , where D 2 u D^2u D 2 u extends continuously up to F B ( u ) \mathrm{FB}(u) FB ( u ) from the positive phase.
For planar solutions with connected positivity set, Basulto and Kamburov [3 Theorem 1.1] classified the finite-index profiles. In higher dimensions, Liu, Wang and Wei [20 ] constructed axially symmetric catenoid-type solutions. The cone classification and Corollary 1.2 give, respectively, n c o n e ∗ > 6 n_{\mathrm{cone}}^*>6 n cone ∗ > 6 and n e n t i r e ∗ > 6 n_{\mathrm{entire}}^*>6 n entire ∗ > 6 . Together these verify the threshold required by the recent work of Fernández-Real, Florit-Simon and Serra [15 Theorem 1.2 and (1.3)–(1.5)] . Consequently, every non-flat entire classical solution of finite Morse index in dimensions 3 ≤ n ≤ 6 3 \le n \le6 3 ≤ n ≤ 6 is axially symmetric and belongs, up to rigid motion and Bernoulli rescaling u z , λ ( x ) = λ − 1 u ( z + λ x ) u_{z,\lambda}(x)=\lambda^{-1}u(z+\lambda x) u z , λ ( x ) = λ − 1 u ( z + λ x ) , either to the radial exterior-ball family or to the two-ended catenoid-type family (recorded precisely in Corollary 3.9).
Once all entire finite-Morse-index blow-up models are classified, it is natural to ask how curvature concentrates along sequences of classical solutions with uniformly bounded Morse index. A geometric precedent is the total-curvature quantization theorem of Buzano and Sharp [5 ] for closed minimal hypersurfaces of bounded area and index. In the Bernoulli problem, the scale-invariant curvature measure we study is the L n L^n L n Hessian measure
μ u ( E ) = ∫ E ∩ { u > 0 } ∣ D 2 u ∣ n d x . \mu_u(E)=\int_{E\cap\{u>0\}}|D^2u|^n\,dx. μ u ( E ) = ∫ E ∩ { u > 0 } ∣ D 2 u ∣ n d x .
where ∣ D 2 u ∣ |D^2u| ∣ D 2 u ∣ denotes the Hilbert–Schmidt norm of the classical Hessian in { u > 0 } \{u>0\} { u > 0 } (extended by zero outside { u > 0 } \{u>0\} { u > 0 } ), which satisfies μ u z , λ ( E ) = μ u ( z + λ E ) \mu_{u_{z,\lambda}}(E)=\mu_u(z+\lambda E) μ u z , λ ( E ) = μ u ( z + λ E ) . In the Bernoulli setting, classical free boundaries can collapse onto two-sided unit wedge limits ∣ ( x − p ) ⋅ e ∣ |(x-p)\cdot e| ∣ ( x − p ) ⋅ e ∣ whose free boundaries develop self-contact without carrying positive-phase Hessian mass, while shrinking exterior-ball solutions converge uniformly to zero even as their positivity sets fill the ambient ball in L 1 L^1 L 1 . Kriventsov and Weiss [19 Theorems 1.2 and 9.3] developed a compact variational formulation that tracks the phase indicator independently of the function in such limits. Here the second main theorem establishes an exact interior bubble-tree identity for the Hessian measure. The proof controls its mass across intermediate neck scales.
Theorem 1.4. Let 3 ≤ n ≤ 6 3 \le n \le6 3 ≤ n ≤ 6 , and let { u j } \{u_j\} { u j } be a sequence of classical solutions to (1) in B 2 ⊂ R n B_2 \subset\mathbb{R}^n B 2 ⊂ R n satisfying ind ( u j ; B 2 ) ≤ I \operatorname{ind}(u_j;B_2)\le I ind ( u j ; B 2 ) ≤ I and F B ( u j ) ∩ B 1 ≠ ∅ \mathrm{FB}(u_j)\cap B_1\ne\varnothing FB ( u j ) ∩ B 1 = ∅ . After passing to a subsequence, u j u_j u j converges locally uniformly in B 2 B_2 B 2 to a locally Lipschitz function U U U that is harmonic on { U > 0 } \{U>0\} { U > 0 } (whose free boundary may exhibit two-sided sheet contact), and there exist an integer 0 ≤ N ≤ I 0\le N\le I 0 ≤ N ≤ I , nonflat entire classical solutions V 1 , … , V N V^1,\ldots,V^N V 1 , … , V N of finite Morse index, centers z j i ∈ F B ( u j ) z_j^i\in\mathrm{FB}(u_j) z j i ∈ FB ( u j ) converging to points p i ∈ B 2 p^i\in B_2 p i ∈ B 2 , and blow-up scales λ j i → 0 \lambda_j^i\to0 λ j i → 0 with the following properties:
(i) For each 1 ≤ i ≤ N 1\le i\le N 1 ≤ i ≤ N , the rescalings u j ( z j i + λ j i x ) / λ j i u_j(z_j^i+\lambda_j^ix)/\lambda_j^i u j ( z j i + λ j i x ) / λ j i converge locally uniformly on R n \mathbb{R}^n R n to V i V^i V i , smoothly on compact subsets of the positive phase in local free-boundary graph charts, and with 1 { u j ( z j i + λ j i ⋅ ) > 0 } → 1 { V i > 0 } \mathbf{1}_{\{u_j(z_j^i+\lambda_j^i\cdot)>0\}}\to\mathbf{1}_{\{V^i>0\}} 1 { u j ( z j i + λ j i ⋅ ) > 0 } → 1 { V i > 0 } in L l o c 1 ( R n ) L^1_{\mathrm{loc}}(\mathbb{R}^n) L loc 1 ( R n ) . Each bubble V i V^i V i is, up to rigid motion and Bernoulli scaling, either the radial exterior-ball solution or an axially symmetric catenoid-type solution, and has finite total Hessian energy μ V i ( R n ) < ∞ \mu_{V^i}(\mathbb{R}^n)<\infty μ V i ( R n ) < ∞ .
(ii) The bubble scales and centers are pairwise separated in the bubble-tree sense, and the total Morse index is bounded by I I I :
∑ i = 1 N ind ( V i ) ≤ I , ∣ z j i − z j k ∣ max ( λ j i , λ j k ) ⟶ ∞ ( 1 ≤ i ≠ k ≤ N ) . \sum_{i=1}^{N}\operatorname{ind}(V^i)\le I,\qquad
\frac{|z_j^i-z_j^k|}{\max(\lambda_j^i,\lambda_j^k)}\longrightarrow\infty
\quad(1\le i\ne k\le N). i = 1 ∑ N ind ( V i ) ≤ I , max ( λ j i , λ j k ) ∣ z j i − z j k ∣ ⟶ ∞ ( 1 ≤ i = k ≤ N ) .
(iii) As Radon measures locally in B 2 B_2 B 2 , the scale-invariant Hessian measures satisfy the exact energy identity
μ u j ⇀ ∣ D 2 U ∣ n 1 { U > 0 } d x + ∑ i = 1 N μ V i ( R n ) δ p i , (2) \mu_{u_j} \rightharpoonup|D^2U|^n\mathbf{1}_{\{U>0\}}\,\mathrm{d}x+\sum_{i=1}^{N}\mu_{V^i}(\mathbb{R}^n)\delta_{p^i}, \tag*{(2)} μ u j ⇀ ∣ D 2 U ∣ n 1 { U > 0 } d x + i = 1 ∑ N μ V i ( R n ) δ p i , (2)
where the concentration points p 1 , … , p N p^1,\ldots,p^N p 1 , … , p N need not be distinct.
As immediate consequences of Theorem 1.4, we obtain uniform interior L n L^n L n -Hessian bounds depending only on n n n , I I I , and the subdomain (see Corollary 4.6 below). When curvature blows up at I I I distinct points, there is one index-one bubble at each point (see Corollary 4.7 below). Proposition 5.1 computes the exterior-ball bubble energy
E n b a l l = ∣ S n − 1 ∣ [ n ( n − 1 ) ] n / 2 − 1 . E_n^{\mathrm{ball}}=|\mathbb{S}^{n-1}|[n(n-1)]^{n/2-1}. E n ball = ∣ S n − 1 ∣ [ n ( n − 1 ) ] n /2 − 1 .
For a two-ended bubble, the weight in (2) retains its dependence on the profile. Proposition 5.3 gives index-zero rescalings in dimension seven for which the interior Hessian energy grows at least logarithmically. Thus the dimensional threshold n ≤ 6 n\le6 n ≤ 6 is sharp for both parts of the paper.
Strategy of the proofs and organization of the paper
Let us outline the main analytical mechanisms behind Theorems 1.1 and 1.4.
1. Localized signed tree and fixed-center excess contraction. Suppose toward a contradiction that a nonflat entire classical stable solution exists in dimension 4 ≤ n < n c o n e ∗ 4\le n<n_{\mathrm{cone}}^* 4 ≤ n < n cone ∗ . After normalizing ∥ ∇ u ∥ L ∞ ≤ 1 \|\nabla u\|_{L^\infty}\le1 ∥∇ u ∥ L ∞ ≤ 1 and ∥ D 2 u ∥ L ∞ = ∣ D 2 u ( 0 ) ∣ = 1 \|D^2u\|_{L^\infty}=|D^2u(0)|=1 ∥ D 2 u ∥ L ∞ = ∣ D 2 u ( 0 ) ∣ = 1 , the ε \varepsilon ε -regularity theory and cone classification below n c o n e ∗ n_{\mathrm{cone}}^* n cone ∗ produce a nonempty set of intrinsic neck centers Z \mathcal{Z} Z with regularity scales r ∗ ( ζ ) ≥ c 0 > 0 r_*(\zeta)\ge c_0>0 r ∗ ( ζ ) ≥ c 0 > 0 carrying a definite quantum η 0 n \eta_0^n η 0 n of L n L^n L n -Hessian mass. This setup comes from [10 ] and is developed in [16 Section 5] . A major obstacle when passing from n = 4 n=4 n = 4 to n = 5 , 6 n=5,6 n = 5 , 6 is that the ( n − 2 ) (n-2) ( n − 2 ) -dimensional neck content Q ( r ) = ∑ N ℓ N n − 2 Q(r)=\sum_{\mathcal{N}} \ell_{\mathcal{N}}^{n-2} Q ( r ) = ∑ N ℓ N n − 2 has supercritical scaling r n − 2 r^{n-2} r n − 2 at infinity. Rather than performing a global blow-down of trees across R n \mathbb{R}^n R n , we fix a single neck center z ∈ Z z\in\mathcal{Z} z ∈ Z and build a finite Calderón–Zygmund-type signed stopping-time tree inside one large root ball B R ( z ) B_R(z) B R ( z ) , decomposing Ω ∩ B R / 2 ( z ) \Omega\cap B_{R/2}(z) Ω ∩ B R /2 ( z ) outside a union of neck terminal balls into two disjoint Caccioppoli sheets U ± U_\pm U ± closely modeled on a symmetric unit wedge ∣ ( x − z ) ⋅ e ∣ |(x-z)\cdot e| ∣ ( x − z ) ⋅ e ∣ (Lemma 2.1 ). Because z ∈ Z z\in\mathcal{Z} z ∈ Z is fixed, the neck terminal N ∗ N_* N ∗ containing z z z has a strictly positive radius ℓ ∗ > 0 \ell_*>0 ℓ ∗ > 0 that persists at every scale r > 0 r>0 r > 0 around z z z .
To bound the neck content Q ( s ) Q(s) Q ( s ) as we zoom in toward z z z , we project the gradient ∇ u \nabla u ∇ u onto an ( n − 2 ) (n-2) ( n − 2 ) -plane M M M nearly orthogonal to both approximate sheet normals. The components of F = P M ∇ u F=P_M\nabla u F = P M ∇ u are translation Jacobi fields, so they obey a Sternberg–Zumbrun-type stability inequality [22 ] , in the vector-valued Bernoulli formulation of [16 Lemma 6.3] . This controls the directional curvature density A M 2 = ∣ F ∣ 2 ∣ D ( F / ∣ F ∣ ) ∣ 2 A_M^2=|F|^2|D(F/|F|)|^2 A M 2 = ∣ F ∣ 2 ∣ D ( F /∣ F ∣ ) ∣ 2 . In Lemma 2.2 , we prove a uniform coercivity gap on each doubled intrinsic ball B 2 r ∗ ( ζ ) ( ζ ) B_{2r_*(\zeta)}(\zeta) B 2 r ∗ ( ζ ) ( ζ ) : if A M 2 A_M^2 A M 2 vanished in the blow-up limit, harmonic unique continuation would produce n − 3 n-3 n − 3 independent translation-invariant directions inside M M M , splitting off an R n − 3 \mathbb{R}^{n-3} R n − 3 factor and yielding a classical stable quotient in R 3 \mathbb{R}^3 R 3 , which must have vanishing Hessian by the three-dimensional theorem of [10 ] . Summing this local gap against the Sternberg–Zumbrun cutoff estimate bounds Q ( θ r ) Q(\theta r) Q ( θ r ) quadratically in terms of the projected Dirichlet energy. In Lemma 2.3 , we couple the two-sided L 2 L^2 L 2 affine excess on U ± ∩ B r U_\pm\cap B_r U ± ∩ B r with κ r 2 − n Q ( r ) \kappa r^{2-n}Q(r) κ r 2 − n Q ( r ) : a divergence-free Rellich stress-energy identity controls free-boundary normal deviation and rules out L 2 L^2 L 2 concentration in the vanishing nodal slab via the Euclidean B V BV B V –Sobolev inequality, allowing harmonic linearization with homogeneous Neumann data across the nodal hyperplane. Iterating the resulting excess contraction E ( θ j r 0 , L j ) ≤ 2 − j E ( r 0 , L 0 ) \mathcal{E}(\theta^j r_0,L_j)\le2^{-j}\mathcal{E}(r_0,L_0) E ( θ j r 0 , L j ) ≤ 2 − j E ( r 0 , L 0 ) on a single sufficiently large tree forces Q ( θ j r 0 ) → 0 Q(\theta^j r_0)\to0 Q ( θ j r 0 ) → 0 , contradicting Q ( θ j r 0 ) ≥ ℓ ∗ n − 2 > 0 Q(\theta^j r_0)\ge\ell_*^{n-2}>0 Q ( θ j r 0 ) ≥ ℓ ∗ n − 2 > 0 .
2. Weiss density pinning, annular splitting, and Voronoi bubble trees. For a sequence { u j } \{u_j\} { u j } with ind ( u j ; B 2 ) ≤ I \operatorname{ind}(u_j;B_2)\le I ind ( u j ; B 2 ) ≤ I , repeatedly applying the weighted point-picking argument of Lemma 3.8 extracts at most I I I scale-separated bubbles V 1 , … , V N V^1,\ldots,V^N V 1 , … , V N and leaves a pointwise curvature bound inversely proportional to the distance d j ( x ) d_j(x) d j ( x ) to the bubble centers. The central difficulty in proving the exact measure identity (2) is to rule out diffuse Hessian mass across intermediate annuli λ j i ≪ s j ≪ 1 \lambda_j^i \ll s_j \ll1 λ j i ≪ s j ≪ 1 and across interaction scales ∣ z j i − z j k ∣ |z_j^i-z_j^k| ∣ z j i − z j k ∣ when several bubbles converge to the same point p p p .
Our key observation (i.e., Lemma 4.1 and Proposition 4.2 ) is that for both possible bubble types—radial exterior balls and axially symmetric catenoid-type solutions—the asymptotic Weiss energy at infinity equals the full unit-ball volume ω n = ∣ B 1 ∣ \omega_n=|B_1| ω n = ∣ B 1 ∣ . On the other hand, integrating the Weiss monotonicity differential inequality from any interior scale up to a fixed macroscopic ball caps W ( u j , r ) W(u_j,r) W ( u j , r ) from above by ω n + O ( 1 / log ( R / r ) ) \omega_n+O(1/\log(R/r)) ω n + O ( 1/ log ( R / r )) . Sandwiching the Weiss functional between the bubble scale and the macroscopic scale pins W ( w j , t ) → ω n W(w_j,t)\to\omega_n W ( w j , t ) → ω n at every intermediate scale s j s_j s j . This simultaneously forces one-homogeneity of any intermediate limit w w w and pins the weak-* limit of the positivity indicators to χ ≡ 1 \chi\equiv1 χ ≡ 1 almost everywhere. Outside at most I I I unstable concentration points, our local curvature estimate (Corollary 1.3) and local multi-sheet analysis (Lemmas 3.5–3.6) then rule out one-sided sheets (which would leave an open zero phase contradicting χ ≡ 1 \chi\equiv1 χ ≡ 1 ) and restrict w w w to be either 0 0 0 or a centered unit wedge ∣ x ⋅ e ∣ |x\cdot e| ∣ x ⋅ e ∣ . Across each scale gap, a topological connectedness dichotomy (Lemmas 4.3–4.4) shows that u j u_j u j is either globally harmonic on the gap annulus—where multipole harmonic decay applies—or splits globally into two disjoint classical branches v ± v_\pm v ± to which the annular improvement-of-flatness theorem of [15 ] applies with polynomial decay rate 1 / 2 1/2 1/2 . Finally, partitioning space into Voronoi nearest-center cells E j i E_j^i E j i around colliding centers isolates every transition scale away from neighboring singularities, completing the no-neck energy accounting.
Organization of the paper. Section 2 establishes the localized signed tree decomposition, the projected Sternberg–Zumbrun neck estimate, and the fixed-center excess contraction, proving Theorem 1.1, Corollary 1.2, and the sharp seven-dimensional threshold (Corollary 2.4). Section 3 develops local gradient bounds, multi-sheet graphical compactness, and the point-picking lemma, deducing Corollary 1.3 and the finite-index classification (Corollary 3.9). Section 4 proves the logarithmic Weiss bound, intermediate-scale density pinning, annular energy vanishing, and Theorem 1.4. Finally, Section 5 computes the exterior-ball energy and constructs the seven-dimensional example with logarithmic lower growth.
Global rigidity below the critical cone dimension
Throughout Sections 2.1–2.3, we argue by contradiction to prove Theorem 1.1. Suppose that in some dimension 4 ≤ n < n c o n e ∗ 4\le n<n_{\mathrm{cone}}^* 4 ≤ n < n cone ∗ there exists a nonflat entire classical stable solution to (1) . By the global gradient bound and normalization procedure of [16 Lemmas 3.5 and 5.2] , we may pass to an entire classical stable solution u u u satisfying ∣ ∇ u ∣ ≤ 1 |\nabla u|\le1 ∣∇ u ∣ ≤ 1 and ∣ D 2 u ∣ ≤ 1 |D^2u|\le1 ∣ D 2 u ∣ ≤ 1 everywhere in Ω \Omega Ω , together with 0 ∈ F B ( u ) 0\in\mathrm{FB}(u) 0 ∈ FB ( u ) and ∣ D 2 u ( 0 ) ∣ = 1 |D^2u(0)|=1 ∣ D 2 u ( 0 ) ∣ = 1 (understood as the norm of the positive-side trace).
Following [16 Definition 5.4] , let Z ⊂ F B ( u ) \mathcal{Z}\subset\mathrm{FB}(u) Z ⊂ FB ( u ) denote the collection of intrinsic neck centers of u u u , with associated regularity scales r ∗ ( ζ ) r_*(\zeta) r ∗ ( ζ ) characterized by the threshold Hessian mass
∫ Ω ∩ B r ∗ ( ζ ) ( ζ ) ∣ D 2 u ∣ n d x = η 0 n > 0 , inf ζ ∈ Z r ∗ ( ζ ) > 0. \int_{\Omega\cap B_{r_*(\zeta)}(\zeta)} |D^2u|^n\,dx=\eta_0^n>0,\qquad\inf_{\zeta\in\mathcal{Z}}r_*(\zeta)>0. ∫ Ω ∩ B r ∗ ( ζ ) ( ζ ) ∣ D 2 u ∣ n d x = η 0 n > 0 , ζ ∈ Z inf r ∗ ( ζ ) > 0.
Because ∣ D 2 u ( 0 ) ∣ = 1 |D^2u(0)|=1 ∣ D 2 u ( 0 ) ∣ = 1 and ∥ D 2 u ∥ L ∞ ( Ω ) ≤ 1 \|D^2u\|_{L^\infty(\Omega)}\le1 ∥ D 2 u ∥ L ∞ ( Ω ) ≤ 1 , small-Hessian ε \varepsilon ε -regularity ensures that Z \mathcal{Z} Z is nonempty and that r ∗ ( ζ ) r_*(\zeta) r ∗ ( ζ ) has a uniform positive lower bound [16 Section 5.3] . Moreover, all local neck and wedge approximation properties established in [16 Section 5] hold verbatim under the dimensional hypothesis n < n c o n e ∗ n<n_{\mathrm{cone}}^* n < n cone ∗ (see [16 Remark 5.1 and Section 5.4] ).
Localized signed tree decomposition at a fixed neck center
Fix once and for all a neck center z ∈ Z z \in\mathcal{Z} z ∈ Z , and abbreviate B r = B r ( z ) B_r = B_r(z) B r = B r ( z ) . Rather than letting the tree root vary dynamically, we verify that for every sufficiently large root radius R R R , the dyadic stopping-time construction of [16 Section 9] initialized at B R B_R B R produces two disjoint Caccioppoli sheets U ± ⊂ Ω ∩ B R / 2 U_{\pm} \subset\Omega\cap B_{R/2} U ± ⊂ Ω ∩ B R /2 whose boundary defects are controlled by the ( n − 2 ) (n-2) ( n − 2 ) -dimensional neck content, while retaining a persistent terminal ball around z z z .
Lemma 2.1. Fix z ∈ Z z \in\mathcal{Z} z ∈ Z and write B r = B r ( z ) B_r = B_r(z) B r = B r ( z ) . For a sufficiently small fixed aperture parameter δ > 0 \delta> 0 δ > 0 and every sufficiently large root radius R R R , there exists a finite Calderón–Zygmund-type signed tree rooted at B R B_R B R yielding disjoint open sets U + , U − ⊂ Ω ∩ B R / 2 U_+, U_- \subset\Omega\cap B_{R/2} U + , U − ⊂ Ω ∩ B R /2 of locally finite perimeter and a finite family of neck terminal balls N = { B ℓ N ( x N ) } \mathcal{N} = \{B_{\ell_N}(x_N)\} N = { B ℓ N ( x N )} , one of which, denoted N ∗ N_* N ∗ , contains z z z . With constants independent of R R R , the decomposition satisfies:
(i) Each neck terminal N N N admits an associated center ζ N ∈ Z ∩ 4 N \zeta_N \in\mathcal{Z} \cap4N ζ N ∈ Z ∩ 4 N with comparable scale r ∗ ( ζ N ) ≍ δ ℓ N r_*(\zeta_N) \asymp_{\delta} \ell_N r ∗ ( ζ N ) ≍ δ ℓ N . The map N ↦ ζ N N \mapsto\zeta_N N ↦ ζ N has bounded multiplicity, the doubled intrinsic balls B 2 r ∗ ( ζ ) ( ζ ) B_{2r_*(\zeta)}(\zeta) B 2 r ∗ ( ζ ) ( ζ ) of distinct centers in Z \mathcal{Z} Z have bounded overlap, and max N ℓ N = o ( R ) \max_N \ell_N = o(R) max N ℓ N = o ( R ) as R → ∞ R \to\infty R → ∞ .
(ii) Whenever 0 < r ≤ R / 32 0 < r \le R/32 0 < r ≤ R /32 and the scale ratio r ∗ ( z ) / r r_*(z)/r r ∗ ( z ) / r is sufficiently small, there exists a unit vector e ∈ S n − 1 e \in S^{n-1} e ∈ S n − 1 such that, in terms of the signed normal coordinate t = ( x − z ) ⋅ e t = (x-z) \cdot e t = ( x − z ) ⋅ e ,
B r ∩ { ± t > δ r r } ⊂ U ± ∩ B r ⊂ B r ∩ { ± t > − δ r r } , ∥ u − ∣ t ∣ ∥ L ∞ ( B r ) = o ( r ) , (3) B_r \cap\{\pm t > \delta_r r\} \subset U_{\pm} \cap B_r \subset B_r \cap\{\pm t > -\delta_r r\}, \qquad\lVert u - |t| \rVert_{L^{\infty}(B_r)} = o(r), \tag*{(3)} B r ∩ { ± t > δ r r } ⊂ U ± ∩ B r ⊂ B r ∩ { ± t > − δ r r } , ∥ u − ∣ t ∣ ∥ L ∞ ( B r ) = o ( r ) , (3) where both δ r \delta_r δ r and r − 1 ∥ u − ∣ t ∣ ∥ L ∞ ( B r ) r^{-1}\lVert u - |t| \rVert_{L^{\infty}(B_r)} r − 1 ∥ u − ∣ t ∣ ∥ L ∞ ( B r ) tend to zero as r ∗ ( z ) / r → 0 r_*(z)/r \to0 r ∗ ( z ) / r → 0 , uniformly in R R R .
(iii) Let Y = Ω ∖ ( U + ∪ U − ) Y = \Omega\setminus(U_+ \cup U_-) Y = Ω ∖ ( U + ∪ U − ) denote the residual positive phase and G ± = ∂ ∗ U ± ∩ Ω G_{\pm} = \partial^* U_{\pm} \cap\Omega G ± = ∂ ∗ U ± ∩ Ω the interior reduced boundaries of the signed sheets. Inside B R / 8 B_{R/8} B R /8 , both Y Y Y and G ± G_{\pm} G ± are covered by the union of closed neck terminals N ‾ \overline{N} N . Defining for 0 < r ≤ R / 32 0 < r \le R/32 0 < r ≤ R /32 the ( n − 2 ) (n-2) ( n − 2 ) -dimensional neck content and maximal neck radius
Q ( r ) = ∑ N ‾ ∩ B r ≠ ∅ ℓ N n − 2 , ℓ ( r ) = max N ‾ ∩ B r ≠ ∅ ℓ N , Q(r) = \sum_{\overline{N}\cap B_r\ne\varnothing} \ell_N^{n-2}, \qquad\ell(r) = \max_{\overline{N}\cap B_r\ne\varnothing} \ell_N, Q ( r ) = N ∩ B r = ∅ ∑ ℓ N n − 2 , ℓ ( r ) = N ∩ B r = ∅ max ℓ N , the index set is nonempty (as N ∗ ∩ B r ≠ ∅ N_* \cap B_r \ne\varnothing N ∗ ∩ B r = ∅ ), and one has the perimeter and volume bounds
S ( r ) : = ∑ ± H n − 1 ( G ± ∩ B r ) ≤ C ℓ ( r ) Q ( r ) , ∣ Y ∩ B r ∣ ≤ C ℓ ( r ) 2 Q ( r ) . (4) S(r) := \sum_{\pm} \mathcal{H}^{n-1}(G_{\pm} \cap B_r) \le C\ell(r)Q(r), \qquad|Y \cap B_r| \le C\ell(r)^2Q(r). \tag*{(4)} S ( r ) := ± ∑ H n − 1 ( G ± ∩ B r ) ≤ C ℓ ( r ) Q ( r ) , ∣ Y ∩ B r ∣ ≤ C ℓ ( r ) 2 Q ( r ) . (4)
Proof. We divide the verification into four steps.
Step 1: Stopping-time tree initialization and propagation of intrinsic scales. We run the dyadic free-boundary net, local-child generation, and stopping rules of [16 Definitions 9.1 and 9.3] starting from the root ball B R = B R ( z ) B_R = B_R(z) B R = B R ( z ) . For any active ball B = B r B ( x B ) B = B_{r_B}(x_B) B = B r B ( x B ) , let h B h_B h B denote its normalized L 2 L^2 L 2 wedge excess (measuring L 2 L^2 L 2 proximity to symmetric unit wedges on 16 B 16B 16 B , as in [16 (9.2)] ). Recall that B B B branches precisely when 4 B ∩ Z ≠ ∅ 4B \cap\mathcal{Z} \ne\varnothing 4 B ∩ Z = ∅ and h B < δ h_B < \delta h B < δ ; it stops as a regular terminal if 4 B ∩ Z = ∅ 4B \cap\mathcal{Z} = \varnothing 4 B ∩ Z = ∅ , and as a neck terminal if 4 B ∩ Z ≠ ∅ 4B \cap\mathcal{Z} \ne\varnothing 4 B ∩ Z = ∅ while h B ≥ δ h_B \ge\delta h B ≥ δ . Because z ∈ Z z \in\mathcal{Z} z ∈ Z is fixed, [16 Lemma 5.6] implies h B R → 0 h_{B_R} \to0 h B R → 0 as R → ∞ R \to\infty R → ∞ , so the root ball branches for all large R R R . Recentering a parent’s optimal wedge at each child’s free-boundary center yields the inherited flatness bound h B ≤ C i n h δ h_B \le C_{\mathrm{inh}}\delta h B ≤ C inh δ for every active ball in the tree [16 Lemma 9.4] .
Fix ρ ∈ ( 0 , 1 / 16 ] \rho\in(0,1/16] ρ ∈ ( 0 , 1/16 ] . By the conditional step of [16 Lemma 9.9] , whenever an active ball B B B and a center ζ ∈ Z ∩ 8 B \zeta\in\mathcal{Z} \cap8B ζ ∈ Z ∩ 8 B satisfy r ∗ ( ζ ) ≤ ρ r B r_*(\zeta) \le\rho r_B r ∗ ( ζ ) ≤ ρ r B , one has the refined upper bound
r ∗ ( ζ ) ≤ C n h B r B . (5) r_*(\zeta) \le C_n h_B r_B. \tag*{(5)} r ∗ ( ζ ) ≤ C n h B r B . (5)
(Indeed, the one-sided barrier of [16 Corollary 9.2] permits recentering a lower wedge barrier at ζ \zeta ζ on B 4 r ∗ ( ζ ) ( ζ ) ⊂ 12 B B_{4r_*(\zeta)}(\zeta) \subset12B B 4 r ∗ ( ζ ) ( ζ ) ⊂ 12 B , where [16 Lemma 9.8] applies.) At the root B R B_R B R , no center ζ ∈ Z ∩ B 8 R \zeta\in\mathcal{Z} \cap B_{8R} ζ ∈ Z ∩ B 8 R can have r ∗ ( ζ ) ≥ ρ R r_*(\zeta) \ge\rho R r ∗ ( ζ ) ≥ ρR once R R R is large: otherwise, scale comparison [16 Lemma 5.7] based at ζ \zeta ζ would force r ∗ ( z ) ≥ r ∗ ( ζ ) / C 8 / ρ ≥ ρ R / C 8 / ρ r_*(z) \ge r_*(\zeta)/C_{8/\rho} \ge\rho R/C_{8/\rho} r ∗ ( z ) ≥ r ∗ ( ζ ) / C 8/ ρ ≥ ρR / C 8/ ρ , contradicting the finiteness of r ∗ ( z ) r_*(z) r ∗ ( z ) . Thus (5) holds at the root. Inductively, if B ′ B' B ′ is a child of a branching ball B B B , then 8 B ′ ⊂ ( 21 / 4 ) B 8B' \subset(21/4)B 8 B ′ ⊂ ( 21/4 ) B and r B = 2 r B ′ r_B = 2r_{B'} r B = 2 r B ′ ; since h B < δ h_B < \delta h B < δ , (5) at B B B gives r ∗ ( ζ ) ≤ 2 C n r δ r B ′ r_*(\zeta) \le2C_n r\delta r_{B'} r ∗ ( ζ ) ≤ 2 C n r δ r B ′ for all ζ ∈ Z ∩ 8 B ′ \zeta\in\mathcal{Z} \cap8B' ζ ∈ Z ∩ 8 B ′ . Choosing δ > 0 \delta> 0 δ > 0 small enough that 2 C n r δ ≤ ρ 2C_n r\delta\le\rho 2 C n r δ ≤ ρ closes the induction and establishes (5) across the entire tree. Because every branching ball meets Z \mathcal{Z} Z in 4 B 4B 4 B and inf Z r ∗ > 0 \inf_{\mathcal{Z}} r_* > 0 inf Z r ∗ > 0 , branching radii are bounded away from zero, and since each dyadic generation has finitely many net points, the tree terminates in finitely many steps.
Step 2: Sheet orientations and terminal scale comparability. We propagate normal orientations from parents to children via [16 Lemma 9.6] . To check same-generation compatibility, let D , E D, E D , E be two active balls of equal radius s s s with 8 D ∩ 8 E ≠ ∅ 8D \cap8E \ne\varnothing 8 D ∩ 8 E = ∅ . Their parent centers lie at mutual distance strictly less than 16 s + 5 s = 21 s 16s + 5s = 21s 16 s + 5 s = 21 s , so the eightfold dilations of their parents already intersect. The root has a single chosen orientation. At the next generation, compare the two absolute affine representatives through their L 2 L^2 L 2 errors with u u u on a ball of radius comparable to s s s in 16 D ∩ 16 E 16D \cap16E 16 D ∩ 16 E . Such a ball can be centered at the midpoint of the two centers. Applying [16 Lemma 9.5] on that ball gives an O ( δ ) O(\delta) O ( δ ) gradient difference and an O ( δ s ) O(\delta s) O ( δ s ) affine difference on 8 D ∩ 8 E 8D \cap8E 8 D ∩ 8 E , up to a sign. The induction hypothesis on the parents and the parent–child gradient bounds rule out the negative sign. This gives the same uniform bounds at the new generation as in [16 Lemma 9.11(ii)] . With this compatibility in place, the absence of holes [16 Lemma 9.11(i)] , the two-sheet splitting over regular terminals [16 Lemma 9.10] , and the assembly of signed cores [16 Definition 9.12 and Lemma 9.13] rely purely on local ball-net geometry and therefore carry over unchanged, producing disjoint open sets U ± ⊂ Ω ∩ B R / 2 U_\pm\subset\Omega\cap B_{R/2} U ± ⊂ Ω ∩ B R /2 that cover Ω ∩ B R / 2 \Omega\cap B_{R/2} Ω ∩ B R /2 away from the closed neck terminals.
Now let N = B ℓ N ( x N ) N = B_{\ell_N}(x_N) N = B ℓ N ( x N ) be any neck terminal and pick ζ N ∈ Z ∩ 4 N \zeta_N \in\mathcal{Z} \cap4N ζ N ∈ Z ∩ 4 N . By (5) and the inherited bound h N ≤ C i n h δ h_N \le C_{\mathrm{in}}h\delta h N ≤ C in h δ , we have r ∗ ( ζ N ) ≤ C δ ℓ N r_*(\zeta_N) \le C\delta\ell_N r ∗ ( ζ N ) ≤ C δ ℓ N . Conversely, if ℓ N / r ∗ ( ζ N ) \ell_N/r_*(\zeta_N) ℓ N / r ∗ ( ζ N ) were arbitrarily large relative to δ \delta δ , then [16 Lemma 5.6] would render u u u uniformly o ( ℓ N ) o(\ell_N) o ( ℓ N ) -close to a centered unit wedge on B 32 ℓ N ( ζ N ) B_{32\ell_N}(\zeta_N) B 32 ℓ N ( ζ N ) ; recentering at x N ∈ F B ( u ) x_N \in\mathrm{FB}(u) x N ∈ FB ( u ) would yield a competitor on 16 N 16N 16 N with h N < δ h_N < \delta h N < δ , violating the neck stopping rule. Hence r ∗ ( ζ N ) ≍ δ ℓ N r_*(\zeta_N) \asymp_\delta\ell_N r ∗ ( ζ N ) ≍ δ ℓ N . Since ζ N ∈ 4 N ⊂ B 5 R \zeta_N \in4N \subset B_{5R} ζ N ∈ 4 N ⊂ B 5 R , applying (5) at the root gives ℓ N ≤ C δ h B R R = o ( R ) \ell_N \le C\delta h_{B_R}R = o(R) ℓ N ≤ C δ h B R R = o ( R ) . Moreover, for any fixed center in Z \mathcal{Z} Z , comparability of ℓ N \ell_N ℓ N with r ∗ ( ζ N ) r_*(\zeta_N) r ∗ ( ζ N ) restricts associated terminals to boundedly many dyadic scales, and net separation bounds the number of terminals per scale. Bounded overlap of the doubled intrinsic balls B 2 r ∗ ( ζ ) ( ζ ) B_{2r_*(\zeta)(\zeta)} B 2 r ∗ ( ζ ) ( ζ ) holds within each scale class of [16 Definition 5.4] by disjointness of the undoubled balls, whereas doubled balls from distinct classes are mutually disjoint because larger-scale centers eliminate smaller-scale candidates within four radii. This proves (i).
Step 3: Slab trapping and deep positive caps around z z z . Let 0 < r ≤ R / 32 0 < r \le R/32 0 < r ≤ R /32 with r ∗ ( z ) / r r_*(z)/r r ∗ ( z ) / r small. By [16 Lemma 5.6] , there exists e ∈ S n − 1 e \in S^{n-1} e ∈ S n − 1 such that ∣ t ∣ = ∣ ( x − z ) ⋅ e ∣ |t| = |(x-z)\cdot e| ∣ t ∣ = ∣ ( x − z ) ⋅ e ∣ satisfies ∥ u − t ∥ L ∞ ( B 32 r ) ≤ η r \|u-t\|_{L^\infty(B_{32r})} \le\eta r ∥ u − t ∥ L ∞ ( B 32 r ) ≤ η r , where η → 0 \eta\to0 η → 0 as r ∗ ( z ) / r → 0 r_*(z)/r \to0 r ∗ ( z ) / r → 0 . Suppose N ⊂ B r N \subset B_r N ⊂ B r . If ℓ N ≥ r \ell_N \ge r ℓ N ≥ r , then ∣ ζ N − z ∣ ≤ 6 ℓ N |\zeta_N-z| \le6\ell_N ∣ ζ N − z ∣ ≤ 6 ℓ N , and combining Step 2 with [16 Lemma 5.7] at ζ N \zeta_N ζ N would force r ∗ ( z ) ≥ c δ ℓ N ≥ c δ r r_*(z) \ge c\delta\ell_N \ge c\delta r r ∗ ( z ) ≥ cδ ℓ N ≥ cδ r , impossible once r ∗ ( z ) / r r_*(z)/r r ∗ ( z ) / r is small. Thus ℓ N < r \ell_N < r ℓ N < r and ∣ ζ N − z ∣ ≤ 6 r |\zeta_N-z| \le6r ∣ ζ N − z ∣ ≤ 6 r , whence [16 Lemma 5.8] at scale 16 r 16r 16 r gives r ∗ ( ζ N ) ≤ 2 r r_*(\zeta_N) \le2r r ∗ ( ζ N ) ≤ 2 r . Recentering the lower wedge barrier at ζ N \zeta_N ζ N on B 4 r ∗ ( ζ N ) ( ζ N ) ⊂ B 32 r B_{4r_*(\zeta_N)}(\zeta_N) \subset B_{32r} B 4 r ∗ ( ζ N ) ( ζ N ) ⊂ B 32 r incurs an error at most 2 η r 2\eta r 2 η r , so [16 Lemma 9.8] yields ℓ N ≤ C δ η r \ell_N \le C\delta\eta r ℓ N ≤ C δ η r . Since x N ∈ B 2 r ∩ F B ( u ) x_N \in B_{2r} \cap\mathrm{FB}(u) x N ∈ B 2 r ∩ FB ( u ) also satisfies ∣ ( x N − z ) ⋅ e ∣ ≤ η r |(x_N-z)\cdot e| \le\eta r ∣ ( x N − z ) ⋅ e ∣ ≤ η r , both deep caps B r ∩ { ± t > C ′ δ r } B_r \cap\{ \pm t > C'\delta r\} B r ∩ { ± t > C ′ δ r } lie strictly inside Ω \Omega Ω and avoid every closed neck terminal.
Following the local-child nested cover [16 (9.5)–(9.6)] , there is a descending chain of tree balls keeping z z z inside their quarter-balls that must end at a neck terminal N ∗ ∋ z N_* \ni z N ∗ ∋ z (as z ∈ Z z \in\mathcal{Z} z ∈ Z prevents regular termination). Since ℓ N ∗ ≤ C δ η r ≪ r \ell_{N_*} \le C\delta\eta r \ll r ℓ N ∗ ≤ C δ η r ≪ r , this chain contains a branching ball B B B with r / 8 < r B ≤ r / 4 r/8 < r_B \le r/4 r /8 < r B ≤ r /4 and z ∈ 1 4 B z \in\frac14 B z ∈ 4 1 B . Recentering ∣ t ∣ |t| ∣ t ∣ at x B ∈ F B ( u ) x_B \in\mathrm{FB}(u) x B ∈ FB ( u ) on $16B and invoking affine stability of optimal wedges [16 Lemma 9.5] aligns t t t (after choosing the sign of e e e ) with the signed affine representative ℓ B \ell_B ℓ B of B B B , so that ∥ ℓ B − t ∥ L ∞ ( B ) ≤ C η r \|\ell_B-t\|_{L^\infty(B)} \le C\eta r ∥ ℓ B − t ∥ L ∞ ( B ) ≤ C η r . The two test points x B ± 1 2 r B e x_B \pm\frac{1}{2}r_B e x B ± 2 1 r B e belong to the respective signed cores of B B B [16 (9.17)] and have signed heights at least r / 16 − η r r/16-\eta r r /16 − η r . Because each deep cap in B r B_r B r is connected, avoids all neck terminals, and is covered by U + ∪ U − U_+ \cup U_- U + ∪ U − , meeting the corresponding core places the entire cap inside U ± U_\pm U ± . Disjointness of U + U_+ U + and U − U_- U − then implies the complementary upper slab inclusions, proving (3) .
Step 4: Perimeter and volume estimates on the residual neck set. Because the tree is finite and U ± U_\pm U ± are constructed from Ω \Omega Ω (which has smooth boundary) via finitely many spherical and planar cutoffs, U ± U_\pm U ± have locally finite perimeter. Inside B R / 8 B_{R/8} B R /8 , the union of closed neck terminals N ‾ \overline{N} N covers Y Y Y by [16 (9.20)] ; any point of Ω \Omega Ω outside that closed union possesses an open neighborhood in Ω \Omega Ω partitioned into the disjoint open sets U + U_+ U + and U − U_- U − , and so cannot belong to ∂ ∗ U ± ∩ Ω = G ± \partial^*U_\pm\cap\Omega= G_\pm ∂ ∗ U ± ∩ Ω = G ± .
To bound H n − 1 ( G ± ∩ N ‾ ∩ B R / 8 ) \mathcal{H}^{n-1}(G_\pm\cap\overline{N} \cap B_{R/8}) H n − 1 ( G ± ∩ N ∩ B R /8 ) , we localize the face-counting argument of [16 Lemma 10.3] . Suppose a cutoff boundary face of the signed core of a tree ball B B B intersects N ‾ \overline{N} N . If B B B is a branching ball, any point y y y on its artificial cutoff face satisfies ∣ ℓ B ( y ) ∣ ≥ r B / 16 |\ell_B(y)| \ge r_B/16 ∣ ℓ B ( y ) ∣ ≥ r B /16 ; if ℓ N < r B / 4 \ell_N < r_B/4 ℓ N < r B /4 , evaluating the slab bound at x N ∈ F B ( u ) x_N \in F_B(u) x N ∈ F B ( u ) yields r B / 16 ≤ C δ r B + ℓ N r_B/16 \le C\delta r_B+\ell_N r B /16 ≤ C δ r B + ℓ N , forcing r B ≲ δ ℓ N r_B \lesssim_\delta\ell_N r B ≲ δ ℓ N . If instead B B B is a regular terminal, then 4 B ∩ Z = ∅ 4B \cap Z = \varnothing 4 B ∩ Z = ∅ while ∣ ζ N − x B ∣ ≤ r B + 5 ℓ N |\zeta_N-x_B| \le r_B+5\ell_N ∣ ζ N − x B ∣ ≤ r B + 5 ℓ N , which forces r B + 5 ℓ N ≥ 4 r B r_B+5\ell_N \ge4r_B r B + 5 ℓ N ≥ 4 r B and hence ℓ N ≥ 3 5 r B \ell_N \ge\frac{3}{5}r_B ℓ N ≥ 5 3 r B . Conversely, when r B ≤ ℓ N r_B \le\ell_N r B ≤ ℓ N , let A A A be B B B (if branching) or its parent (if regular), and pick ζ A ∈ Z ∩ 4 A \zeta_A \in Z \cap4A ζ A ∈ Z ∩ 4 A ; then ∣ ζ A − ζ N ∣ ≤ C ℓ N |\zeta_A-\zeta_N| \le C\ell_N ∣ ζ A − ζ N ∣ ≤ C ℓ N and r ∗ ( ζ A ) ≤ C δ r A r_*(\zeta_A) \le C\delta r_A r ∗ ( ζ A ) ≤ C δ r A , whereas scale comparison [16 Lemma 5.7] at ζ N \zeta_N ζ N gives r ∗ ( ζ A ) ≥ c δ ℓ N r_*(\zeta_A) \ge c_\delta\ell_N r ∗ ( ζ A ) ≥ c δ ℓ N , so r B ≳ δ ℓ N r_B \gtrsim_\delta\ell_N r B ≳ δ ℓ N . Thus r B ≃ δ ℓ N r_B \simeq_\delta\ell_N r B ≃ δ ℓ N for every tree ball contributing a face to N ‾ \overline{N} N . Net separation bounds the number of such balls to O δ ( 1 ) O_\delta(1) O δ ( 1 ) , and each spherical or affine cutoff face has ( n − 1 ) (n-1) ( n − 1 ) -measure O ( r B n − 1 ) O(r_B^{n-1}) O ( r B n − 1 ) , yielding
H n − 1 ( G ± ∩ N ‾ ∩ B R / 8 ) ≤ C δ ℓ N n − 1 . \mathcal{H}^{n-1}(G_\pm\cap\overline{N} \cap B_{R/8}) \le C\delta\ell_N^{n-1}. H n − 1 ( G ± ∩ N ∩ B R /8 ) ≤ C δ ℓ N n − 1 .
including any portion of G ± G_\pm G ± lying on ∂ N \partial N ∂ N . Summing this bound and ∣ N ‾ ∣ = C n ℓ N n |\overline{N}|=C_n\ell_N^n ∣ N ∣ = C n ℓ N n over all terminals meeting B r B_r B r , and factoring out one or two powers of ℓ ( r ) \ell(r) ℓ ( r ) , gives (4) .
Sternberg–Zumbrun projection and neck content control
To control the ( n − 2 ) (n-2) ( n − 2 ) -dimensional neck content Q ( s ) Q(s) Q ( s ) in terms of how much ∇ u \nabla u ∇ u deviates from a two-dimensional plane, we project the translation Jacobi fields onto an ( n − 2 ) (n-2) ( n − 2 ) -dimensional subspace M ⊂ R n M \subset\mathbb{R}^n M ⊂ R n and apply a Sternberg–Zumbrun-type stability estimate. The key geometric input is that vanishing of the projected directional curvature at a neck scale splits off an R n − 3 \mathbb{R}^{n-3} R n − 3 Euclidean factor, reducing the limit to a three-dimensional stable solution ruled out by [10 ] .
Lemma 2.2. Let M ⊂ R n M \subset\mathbb{R}^n M ⊂ R n be an ( n − 2 ) (n-2) ( n − 2 ) -dimensional linear subspace, denote by P M P_M P M the orthogonal projection onto M M M , put F = P M ∇ u F=P_M\nabla u F = P M ∇ u , and define the projected directional curvature density
A M 2 = ∣ F ∣ 2 ∣ D ( F ∣ F ∣ ) ∣ 2 where F ≠ 0 , A M 2 = 0 where F = 0. A_M^2=|F|^2\left|D\left(\frac{F}{|F|}\right)\right|^2 \quad\text{where } F\ne0,\qquad A_M^2=0 \quad\text{where } F=0. A M 2 = ∣ F ∣ 2 D ( ∣ F ∣ F ) 2 where F = 0 , A M 2 = 0 where F = 0. Then there exists a constant c = c ( n , η 0 ) > 0 c=c(n,\eta_0)>0 c = c ( n , η 0 ) > 0 , independent of M M M , such that for every intrinsic center ζ ∈ Z \zeta\in Z ζ ∈ Z ,
∫ B 2 r ∗ ( ζ ) ( ζ ) ∩ Ω A M 2 d x ≥ c r ∗ ( ζ ) n − 2 . (6) \int_{B_{2r_*(\zeta)}(\zeta)\cap\Omega} A_M^2\,\mathrm{d}x \ge c r_*(\zeta)^{n-2}. \tag*{(6)} ∫ B 2 r ∗ ( ζ ) ( ζ ) ∩ Ω A M 2 d x ≥ c r ∗ ( ζ ) n − 2 . (6) Moreover, for every smooth cutoff η ∈ C c ∞ ( R n ) \eta\in C^\infty_c(\mathbb{R}^n) η ∈ C c ∞ ( R n ) ,
∫ Ω A M 2 η 2 d x ≤ ∫ Ω ∣ P M ∇ u ∣ 2 ∣ ∇ η ∣ 2 d x . (7) \int_\Omega A_M^2\eta^2\,\mathrm{d}x \le\int_\Omega|P_M\nabla u|^2|\nabla\eta|^2\,\mathrm{d}x. \tag*{(7)} ∫ Ω A M 2 η 2 d x ≤ ∫ Ω ∣ P M ∇ u ∣ 2 ∣∇ η ∣ 2 d x . (7)
Consequently, in the setting of Lemma 2.1, whenever 0 < s ≤ R / 32 0 < s \le R/32 0 < s ≤ R /32 and s < r ≤ R s < r \le R s < r ≤ R are such that all doubled intrinsic balls B 2 r ∗ ( ζ N ) ( ζ N ) B_{2r_*(\zeta_N)}(\zeta_N) B 2 r ∗ ( ζ N ) ( ζ N ) associated with terminals whose closures meet B s B_s B s lie inside B r / 4 B_{r/4} B r /4 , one has
Q ( s ) ≤ C r − 2 ∫ B r / 2 ∩ Ω ∣ P M ∇ u ∣ 2 d x . (8) Q(s) \le Cr^{-2} \int_{B_{r/2}\cap\Omega} |P_M\nabla u|^2\,dx. \tag*{(8)} Q ( s ) ≤ C r − 2 ∫ B r /2 ∩ Ω ∣ P M ∇ u ∣ 2 d x . (8)
Proof. With respect to any orthonormal basis of M M M , the n − 2 n-2 n − 2 components of F = P M ∇ u F=P_M\nabla u F = P M ∇ u are directional derivatives of u u u , hence classical Jacobi fields annihilated by the linearized operator inside Ω \Omega Ω and satisfying the linearized Robin condition on F B ( u ) F\!B(u) F B ( u ) . Applying the vector-valued Sternberg–Zumbrun stability inequality for finite families of Jacobi fields [16 Lemma 6.3] directly gives (2.5).
To establish the uniform lower bound (2.4), suppose by contradiction that along a sequence of normalized solutions u k u_k u k , neck centers ζ k \zeta_k ζ k , and ( n − 2 ) (n-2) ( n − 2 ) -planes M k M_k M k , the rescaled integral r ∗ ( ζ k ) 2 − n ∫ B 2 r ∗ ( ζ k ) ( ζ k ) ∩ Ω k A M k 2 r_*(\zeta_k)^{2-n}\int_{B_{2r_*(\zeta_k)}(\zeta_k)\cap\Omega_k} A^2_{M_k} r ∗ ( ζ k ) 2 − n ∫ B 2 r ∗ ( ζ k ) ( ζ k ) ∩ Ω k A M k 2 tends to zero. Set u ~ k ( x ) = r ∗ ( ζ k ) − 1 u k ( ζ k + r ∗ ( ζ k ) x ) \tilde{u}_k(x)=r_*(\zeta_k)^{-1}u_k(\zeta_k+r_*(\zeta_k)x) u ~ k ( x ) = r ∗ ( ζ k ) − 1 u k ( ζ k + r ∗ ( ζ k ) x ) . By definition of r ∗ ( ζ k ) r_*(\zeta_k) r ∗ ( ζ k ) , each u ~ k \tilde{u}_k u ~ k has L n L^n L n -Hessian mass η 0 n \eta_0^n η 0 n in B 1 B_1 B 1 , while scale comparison [16 Lemma 5.7] provides uniform Hessian bounds on every fixed ball. Extracting a diagonal subsequence via neck compactness [16 Lemmas 4.6 and 6.8] and compactness of the Grassmannian, we obtain an ( n − 2 ) (n-2) ( n − 2 ) -plane M M M and an entire classical stable limit v v v satisfying
∫ B 1 ∩ { v > 0 } ∣ D 2 v ∣ n d x = η 0 n > 0. \int_{B_1\cap\{v>0\}} |D^2v|^n\,dx=\eta_0^n>0. ∫ B 1 ∩ { v > 0 } ∣ D 2 v ∣ n d x = η 0 n > 0.
On the open subset of B 2 ∩ { v > 0 } B_2\cap\{v>0\} B 2 ∩ { v > 0 } where F = P M ∇ v ≠ 0 F=P_M\nabla v\ne0 F = P M ∇ v = 0 , smooth interior convergence implies pointwise convergence of A M k 2 ( u ~ k ) A^2_{M_k}(\tilde{u}_k) A M k 2 ( u ~ k ) to A M 2 ( v ) = ∣ ( I − F ^ ⊗ F ^ ) D F ∣ 2 A^2_M(v)=\left|(I-\widehat{F}\otimes\widehat{F})DF\right|^2 A M 2 ( v ) = ( I − F ⊗ F ) D F 2 , where F ^ = F / ∣ F ∣ \widehat{F}=F/|F| F = F /∣ F ∣ . Fatou’s lemma gives vanishing almost everywhere on { F ≠ 0 } ∩ B 2 ∩ { v > 0 } \{F\ne0\}\cap B_2\cap\{v>0\} { F = 0 } ∩ B 2 ∩ { v > 0 } . Continuity on this set and the convention A M 2 = 0 A^2_M=0 A M 2 = 0 on { F = 0 } \{F=0\} { F = 0 } then give pointwise vanishing on B 2 ∩ { v > 0 } B_2\cap\{v>0\} B 2 ∩ { v > 0 } .
Now let D D D be any connected component of { v > 0 } \{v>0\} { v > 0 } intersecting B 1 B_1 B 1 . If F F F is not identically zero on D ∩ B 2 D\cap B_2 D ∩ B 2 , pick a small interior ball B ′ ⋐ D ∩ B 2 B'\Subset D\cap B_2 B ′ ⋐ D ∩ B 2 on which F ≠ 0 F\ne0 F = 0 ; vanishing of A M 2 ( v ) A^2_M(v) A M 2 ( v ) implies that F ^ \widehat{F} F is constant on B ′ B' B ′ , so all directional derivatives of v v v along the ( n − 3 ) (n-3) ( n − 3 ) -dimensional subspace W = M ∩ F ^ ⊥ W=M\cap\widehat{F}^{\perp} W = M ∩ F ⊥ vanish on B ′ B' B ′ , and hence throughout D D D by unique continuation for harmonic functions. If instead F ≡ 0 F\equiv0 F ≡ 0 on D ∩ B 2 D\cap B_2 D ∩ B 2 , the same unique continuation holds for any ( n − 3 ) (n-3) ( n − 3 ) -dimensional subspace W ⊂ M W\subset M W ⊂ M . Because v v v is constant and strictly positive along every line parallel to W W W passing through a point of D D D , no such line can meet ∂ D \partial D ∂ D . In orthogonal coordinates adapted to R n = R 3 × W \mathbb{R}^n=\mathbb{R}^3\times W R n = R 3 × W , the component splits metrically as a cylinder D = D 0 × R n − 3 D=D_0\times\mathbb{R}^{n-3} D = D 0 × R n − 3 with v ( x , y ) = v 0 ( x ) v(x,y)=v_0(x) v ( x , y ) = v 0 ( x ) . Since the fibers { x } × R n − 3 \{x\}\times\mathbb{R}^{n-3} { x } × R n − 3 are tangent to ∂ D \partial D ∂ D , smooth local one-sidedness descends to D 0 ⊂ R 3 D_0\subset\mathbb{R}^3 D 0 ⊂ R 3 , so extending v 0 v_0 v 0 by zero outside D 0 D_0 D 0 produces an entire classical solution to (1.1) in R 3 \mathbb{R}^3 R 3 .
Furthermore, v ∣ D v|_D v ∣ D inherits stability on its own closure: every point of D ‾ \overline{D} D is either interior to D D D or admits a smooth one-sided free-boundary chart disjoint from all other components of { v > 0 } \{v>0\} { v > 0 } , allowing any compactly supported test function on D ‾ \overline{D} D to be extended by zero to the remaining components without changing Q v Q_v Q v . Given an admissible test function ϕ \phi ϕ on D 0 ‾ ⊂ R 3 \overline{D_0}\subset\mathbb{R}^3 D 0 ⊂ R 3 and a nontrivial cutoff η ∈ C c ∞ ( R n ) \eta\in C^\infty_c(\mathbb{R}^n) η ∈ C c ∞ ( R n ) , testing with the product Φ A ( x , y ) = ϕ ( x ) η ( y / A ) \Phi_A(x,y)=\phi(x)\eta(y/A) Φ A ( x , y ) = ϕ ( x ) η ( y / A ) yields, since the Euclidean factor carries zero principal curvatures,
0 ≤ Q v ∣ D [ Φ A ; R n ] ∥ η ( y / A ) ∥ L 2 2 = Q v 0 [ ϕ ; R 3 ] + A − 2 ∥ ∇ η ∥ L 2 2 ∥ η ∥ L 2 2 ∫ D 0 ϕ 2 d x . 0\le\frac{Q_{v|D}[\Phi_A;\mathbb{R}^n]}{\|\eta(y/A)\|_{L^2}^2}=Q_{v_0}[\phi;\mathbb{R}^3]+A^{-2}\frac{\|\nabla\eta\|_{L^2}^2}{\|\eta\|_{L^2}^2}\int_{D_0}\phi^2\,dx. 0 ≤ ∥ η ( y / A ) ∥ L 2 2 Q v ∣ D [ Φ A ; R n ] = Q v 0 [ ϕ ; R 3 ] + A − 2 ∥ η ∥ L 2 2 ∥∇ η ∥ L 2 2 ∫ D 0 ϕ 2 d x .
Letting A → ∞ A\to\infty A → ∞ shows that v 0 v_0 v 0 is a classical stable solution in R 3 \mathbb{R}^3 R 3 . By three-dimensional stable rigidity [10 Theorem 1.5] , D 2 v 0 ≡ 0 D^2v_0\equiv0 D 2 v 0 ≡ 0 on D 0 D_0 D 0 , and therefore D 2 v ≡ 0 D^2v\equiv0 D 2 v ≡ 0 on every component of { v > 0 } \{v>0\} { v > 0 } meeting B 1 B_1 B 1 , contradicting ∫ B 1 ∩ { v > 0 } ∣ D 2 v ∣ n d x = η 0 n > 0 \int_{B_1\cap\{v>0\}} |D^2v|^n\,dx=\eta_0^n>0 ∫ B 1 ∩ { v > 0 } ∣ D 2 v ∣ n d x = η 0 n > 0 . This proves (6) .
Finally, under the containment hypothesis of (8) , we sum (6) over the distinct centers ζ N \zeta_N ζ N associated with terminals whose closures meet B s B_s B s . Using ℓ N ≲ δ r ∗ ( ζ N ) \ell_N\lesssim_\delta r_*(\zeta_N) ℓ N ≲ δ r ∗ ( ζ N ) , the bounded multiplicity of N ↦ ζ N N\mapsto\zeta_N N ↦ ζ N , and the bounded overlap of B 2 r ∗ ( ζ N ) ( ζ N ) B_{2r_*(\zeta_N)}(\zeta_N) B 2 r ∗ ( ζ N ) ( ζ N ) from Lemma 2.1(i), we bound Q ( s ) Q(s) Q ( s ) by a constant multiple of ∫ B r / 4 ∩ Ω A M 2 d x \int_{B_{r/4}\cap\Omega} A_M^2\,dx ∫ B r /4 ∩ Ω A M 2 d x , and applying (7) with a smooth cutoff η ≡ 1 \eta\equiv1 η ≡ 1 on B r / 4 B_{r/4} B r /4 supported in B r / 2 B_{r/2} B r /2 with ∣ ∇ η ∣ ≤ C / r |\nabla\eta|\le C/r ∣∇ η ∣ ≤ C / r yields (8) . □ \square □
Fixed-center excess decay and proof of Theorem 1.1
Retaining the localized decomposition of Lemma 2.1 around z ∈ Z z\in\mathcal{Z} z ∈ Z , for any radius 0 < r ≤ R / 32 0<r\le R/32 0 < r ≤ R /32 and any pair L = ( L + , L − ) L=(L_+,L_-) L = ( L + , L − ) of affine functions L ± ( x ) = b ± ⋅ ( x − z ) + d ± L_\pm(x)=b_\pm\cdot(x-z)+d_\pm L ± ( x ) = b ± ⋅ ( x − z ) + d ± with unit slopes ∣ b ± ∣ = 1 |b_\pm|=1 ∣ b ± ∣ = 1 , we introduce the coupled quadratic-neck excess
E ( r , L ) 2 = r − n − 2 ∑ ± ∫ U ± ∩ B r ( u − L ± ) 2 d x + κ r 2 − n Q ( r ) , (9) \mathcal{E}(r,L)^2=r^{-n-2}\sum_\pm\int_{U_\pm\cap B_r}(u-L_\pm)^2\,dx+\kappa r^{2-n}Q(r), \tag*{(9)} E ( r , L ) 2 = r − n − 2 ± ∑ ∫ U ± ∩ B r ( u − L ± ) 2 d x + κ r 2 − n Q ( r ) , (9)
where κ > 0 \kappa>0 κ > 0 is a dimensional coupling weight to be fixed below. The heart of the rigidity proof is the following excess-decay contraction.
Lemma 2.3. There exist constants θ ∈ ( 0 , 1 / 32 ) \theta\in(0,1/32) θ ∈ ( 0 , 1/32 ) , κ > 0 \kappa>0 κ > 0 , and λ ∗ > 0 \lambda_*>0 λ ∗ > 0 , depending only on n n n and the fixed tree aperture δ \delta δ , such that whenever 0 < r ≤ R / 32 0<r\le R/32 0 < r ≤ R /32 and E ( r , L ) ≤ λ ∗ \mathcal{E}(r,L)\le\lambda_* E ( r , L ) ≤ λ ∗ , there exists a new pair L ′ = ( L + ′ , L − ′ ) L'=(L'_+,L'_-) L ′ = ( L + ′ , L − ′ ) of unit-slope affine functions satisfying
E ( θ r , L ′ ) ≤ 1 2 E ( r , L ) . \mathcal{E}(\theta r,L')\le\frac{1}{2}\mathcal{E}(r,L). E ( θ r , L ′ ) ≤ 2 1 E ( r , L ) .
Proof. To make it transparent that no circularity occurs among the constants, we note at the outset that δ \delta δ has already been fixed in Lemma 2.1; below we will first select θ ∈ ( 0 , 1 / 32 ) \theta\in(0,1/32) θ ∈ ( 0 , 1/32 ) solely from interior harmonic Taylor decay, next fix κ > 0 \kappa>0 κ > 0 small relative to θ n − 2 \theta^{n-2} θ n − 2 , and finally choose λ ∗ > 0 \lambda_*>0 λ ∗ > 0 small enough for all compactness and higher-order absorption steps at the fixed pair ( θ , κ ) (\theta,\kappa) ( θ , κ ) .
Step 1: Rescaling, neck smallness, and approximate normal alignment. Translating z z z to the origin and rescaling from B r ( z ) B_r(z) B r ( z ) to B 1 ( 0 ) B_1(0) B 1 ( 0 ) , put λ = E ( 1 , L ) ≤ λ ∗ \lambda=\mathcal{E}(1,L)\le\lambda_* λ = E ( 1 , L ) ≤ λ ∗ , and let Q Q Q , S S S , ℓ \ell ℓ refer to the rescaled tree. Directly from (9) and (4) , setting q = λ 2 / κ q=\lambda^2/\kappa q = λ 2 / κ , we have
Q ( 1 ) ≤ q , ℓ ( 1 ) ≤ q 1 / ( n − 2 ) , S ( 1 ) ≤ C q 1 + 1 / ( n − 2 ) , ∣ Y ∩ B 1 ∣ ≤ C q 1 + 2 / ( n − 2 ) . (10) Q(1)\le q,\qquad\ell(1)\le q^{1/(n-2)},\qquad S(1)\le Cq^{1+1/(n-2)},\qquad|Y\cap B_1|\le Cq^{1+2/(n-2)}. \tag*{(10)} Q ( 1 ) ≤ q , ℓ ( 1 ) ≤ q 1/ ( n − 2 ) , S ( 1 ) ≤ C q 1 + 1/ ( n − 2 ) , ∣ Y ∩ B 1 ∣ ≤ C q 1 + 2/ ( n − 2 ) . (10)
For each fixed κ > 0 \kappa>0 κ > 0 , these bounds give
S ( 1 ) + ∣ Y ∩ B 1 ∣ λ 2 ≤ C κ ( q 1 / ( n − 2 ) + q 2 / ( n − 2 ) ) ⟶ 0 ( λ → 0 ) . \frac{S(1)+|Y\cap B_1|}{\lambda^2}\le\frac{C}{\kappa}\left(q^{1/(n-2)}+q^{2/(n-2)}\right)\longrightarrow0\qquad(\lambda\to0). λ 2 S ( 1 ) + ∣ Y ∩ B 1 ∣ ≤ κ C ( q 1/ ( n − 2 ) + q 2/ ( n − 2 ) ) ⟶ 0 ( λ → 0 ) .
We may therefore impose S ( 1 ) + ∣ Y ∩ B 1 ∣ ≤ λ 2 S(1)+|Y\cap B_1|\le\lambda^2 S ( 1 ) + ∣ Y ∩ B 1 ∣ ≤ λ 2 by shrinking the smallness threshold after fixing κ \kappa κ . Moreover, the terminal N ∗ N_* N ∗ containing the origin satisfies ∣ ζ N ∗ ∣ ≤ 5 ℓ N ∗ |\zeta_{N_*}|\le5\ell_{N_*} ∣ ζ N ∗ ∣ ≤ 5 ℓ N ∗ and r ∗ ( ζ N ∗ ) ≍ ℓ N ∗ ≤ ℓ ( 1 ) r_*(\zeta_{N_*})\asymp\ell_{N_*}\le\ell(1) r ∗ ( ζ N ∗ ) ≍ ℓ N ∗ ≤ ℓ ( 1 ) , so scale comparison [16 Lemma 5.7] gives r ∗ ( 0 ) ≤ C ℓ ( 1 ) → 0 r_*(0)\le C\ell(1)\to0 r ∗ ( 0 ) ≤ C ℓ ( 1 ) → 0 in the rescaled variables. Lemma 2.1(ii) therefore applies on B 1 B_1 B 1 ; rotating coordinates so that e = e n e=e_n e = e n , the slab inclusions (3) hold with t = x n t=x_n t = x n , and L 2 L^2 L 2 comparison of L ± ( x ) = b ± ⋅ x + d ± L_\pm(x)=b_\pm\cdot x+d_\pm L ± ( x ) = b ± ⋅ x + d ± against ± x n \pm x_n ± x n on fixed interior balls inside the deep caps gives b ± = ± e n + o ( 1 ) b_\pm=\pm e_n+o(1) b ± = ± e n + o ( 1 ) and d ± = o ( 1 ) d_\pm=o(1) d ± = o ( 1 ) as λ → 0 \lambda\to0 λ → 0 .
Step 2: One-sided lower barrier and Caccioppoli energy bound. Fix a sign σ ∈ { + , − } \sigma\in\{+,-\} σ ∈ { + , − } and write w = u − L σ w=u-L_\sigma w = u − L σ and h = ∇ w h=\nabla w h = ∇ w on U σ U_\sigma U σ . We first show that w ≥ − C λ w\ge-C\lambda w ≥ − C λ throughout B 2 / 3 B_{2/3} B 2/3 (with w w w defined globally via u ≥ 0 u\ge0 u ≥ 0 ). Indeed, at any point x ∈ B 2 / 3 x\in B_{2/3} x ∈ B 2/3 where w ( x ) < 0 w(x)<0 w ( x ) < 0 , necessarily L σ ( x ) > 0 L_\sigma(x)>0 L σ ( x ) > 0 , which by Step 1 forces σ x n ≥ − o ( 1 ) \sigma x_n\ge-o(1) σ x n ≥ − o ( 1 ) . Shifting along the unit slope to x ′ = x + 1 4 b σ x' = x + \frac{1}{4}b_\sigma x ′ = x + 4 1 b σ , the global 1-Lipschitz bound ∣ ∇ u ∣ ≤ 1 |\nabla u| \le1 ∣∇ u ∣ ≤ 1 implies u ( x ) − u ( x ′ ) ≥ − ∣ x − x ′ ∣ = − 1 4 = L σ ( x ) − L σ ( x ′ ) u(x) - u(x') \ge-|x-x'| = -\frac{1}{4} = L_\sigma(x) - L_\sigma(x') u ( x ) − u ( x ′ ) ≥ − ∣ x − x ′ ∣ = − 4 1 = L σ ( x ) − L σ ( x ′ ) , so w ( x ) ≥ w ( x ′ ) w(x) \ge w(x') w ( x ) ≥ w ( x ′ ) . Because ∣ x ′ ∣ ≤ 2 3 + 1 4 = 11 12 |x'| \le\frac{2}{3} + \frac{1}{4} = \frac{11}{12} ∣ x ′ ∣ ≤ 3 2 + 4 1 = 12 11 and σ x n ′ ≥ 1 4 − o ( 1 ) \sigma x'_n \ge\frac{1}{4} - o(1) σ x n ′ ≥ 4 1 − o ( 1 ) , a ball of fixed radius around x ′ x' x ′ is contained in the deep cap U σ ∩ B 1 U_\sigma\cap B_1 U σ ∩ B 1 by (3) , where w w w is harmonic and the mean-value inequality yields ∣ w ( x ′ ) ∣ ≤ C λ |w(x')| \le C\lambda ∣ w ( x ′ ) ∣ ≤ C λ . Thus w + C λ ≥ 0 w+C\lambda\ge0 w + C λ ≥ 0 on B 2 / 3 B_{2/3} B 2/3 .
Along the physical free boundary ∂ ∗ U σ ∩ F B ( u ) \partial^*U_\sigma\cap\mathrm{FB}(u) ∂ ∗ U σ ∩ FB ( u ) , the outward unit normal to U σ ⊂ Ω U_\sigma\subset\Omega U σ ⊂ Ω coincides H n − 1 \mathcal{H}^{n-1} H n − 1 -almost everywhere with ν = − ∇ u \nu=-\nabla u ν = − ∇ u , and since ∣ b σ ∣ = ∣ ∇ u ∣ = 1 |b_\sigma|=|\nabla u|=1 ∣ b σ ∣ = ∣∇ u ∣ = 1 , the outward normal derivative satisfies ∂ ν w = − 1 + b σ ⋅ ∇ u ≤ 0 \partial_\nu w=-1+b_\sigma\cdot\nabla u\le0 ∂ ν w = − 1 + b σ ⋅ ∇ u ≤ 0 . Choose a smooth cutoff ζ ∈ C ∞ ( B 2 / 3 ) \zeta\in C^\infty(B_{2/3}) ζ ∈ C ∞ ( B 2/3 ) with ζ ≡ 1 \zeta\equiv1 ζ ≡ 1 on B 3 / 5 B_{3/5} B 3/5 . Because u u u is classical, its positive-phase derivatives extend smoothly across F B ( u ) \mathrm{FB}(u) FB ( u ) on supp ζ \operatorname{supp}\zeta supp ζ , permitting application of the Gauss–Green–De Giorgi formula to ζ 2 ( w + C λ ) ∇ w \zeta^2(w+C\lambda)\nabla w ζ 2 ( w + C λ ) ∇ w on the Caccioppoli set U σ U_\sigma U σ . The boundary integral along ∂ ∗ U σ ∩ F B ( u ) \partial^*U_\sigma\cap\mathrm{FB}(u) ∂ ∗ U σ ∩ FB ( u ) is nonpositive since w + C λ ≥ 0 w+C\lambda\ge0 w + C λ ≥ 0 and ∂ ν w ≤ 0 \partial_\nu w\le0 ∂ ν w ≤ 0 , while along the interior partition boundary G σ G_\sigma G σ we have ∣ ∇ w ∣ ≤ 2 |\nabla w|\le2 ∣∇ w ∣ ≤ 2 and ∣ w + C λ ∣ ≤ C |w+C\lambda|\le C ∣ w + C λ ∣ ≤ C , contributing at most C S ( 1 ) CS(1) C S ( 1 ) . Absorbing the interior ∇ ζ \nabla\zeta ∇ ζ cross-term via Young’s inequality and invoking (10) gives the Caccioppoli bound
∫ U σ ∩ B 3 / 5 ∣ h ∣ 2 d x ≤ C ( λ 2 + S ( 1 ) ) ≤ C λ 2 . (11) \int_{U_\sigma\cap B_{3/5}} |h|^2\,dx \le C(\lambda^2+S(1))\le C\lambda^2. \tag*{(11)} ∫ U σ ∩ B 3/5 ∣ h ∣ 2 d x ≤ C ( λ 2 + S ( 1 )) ≤ C λ 2 . (11)
The constants in this bound depend only on n n n and the fixed aperture δ \delta δ . The smallness threshold may depend on κ \kappa κ .
Step 3: Pohozaev–Rellich flux identity and BV trace control. Still fixing a side σ \sigma σ with b = b σ b=b_\sigma b = b σ , consider the Rellich stress-energy vector field
K b ( h ) = 1 2 ∣ h ∣ 2 b − ( b ⋅ h ) h , K_b(h)=\tfrac{1}{2}|h|^2b-(b\cdot h)h, K b ( h ) = 2 1 ∣ h ∣ 2 b − ( b ⋅ h ) h ,
which satisfies div K b ( h ) = 0 \operatorname{div}K_b(h)=0 div K b ( h ) = 0 in U σ U_\sigma U σ because w w w is harmonic and D 2 w D^2w D 2 w is symmetric. On F B ( u ) \mathrm{FB}(u) FB ( u ) , writing g = ∇ u g=\nabla u g = ∇ u and setting d = 1 − b ⋅ g = 1 + b ⋅ ν ≥ 0 d=1-b\cdot g=1+b\cdot\nu\ge0 d = 1 − b ⋅ g = 1 + b ⋅ ν ≥ 0 , the unit-length identities ∣ g ∣ = ∣ b ∣ = 1 |g|=|b|=1 ∣ g ∣ = ∣ b ∣ = 1 give 1 2 ∣ h ∣ 2 = d \tfrac{1}{2}|h|^2=d 2 1 ∣ h ∣ 2 = d and b ⋅ h = − d b\cdot h=-d b ⋅ h = − d , so K b ( h ) = d g K_b(h)=dg K b ( h ) = d g and hence
K b ( h ) ⋅ ν = ∂ ν w = − d ≤ 0 on F B ( u ) . (12) K_b(h)\cdot\nu=\partial_\nu w=-d\le0\quad\text{on }\mathrm{FB}(u). \tag*{(12)} K b ( h ) ⋅ ν = ∂ ν w = − d ≤ 0 on FB ( u ) . (12)
Testing div K b ( h ) = 0 \operatorname{div}K_b(h)=0 div K b ( h ) = 0 on U σ U_\sigma U σ against a nonnegative cutoff equal to 1 1 1 on B 1 / 2 B_{1/2} B 1/2 and supported in B 3 / 5 B_{3/5} B 3/5 , and bounding the interior gradient term by (11) and the co-normal flux across G σ G_\sigma G σ by C S ( 1 ) CS(1) C S ( 1 ) , we obtain the boundary alignment estimate
∫ ∂ ∗ U σ ∩ F B ( u ) ∩ B 1 / 2 d d H n − 1 ≤ C ( λ 2 + S ( 1 ) ) ≤ C λ 2 . (13) \int_{\partial^*U_\sigma\cap\mathrm{FB}(u)\cap B_{1/2}} d\,d\mathcal{H}^{n-1}\le C(\lambda^2+S(1))\le C\lambda^2. \tag*{(13)} ∫ ∂ ∗ U σ ∩ FB ( u ) ∩ B 1/2 d d H n − 1 ≤ C ( λ 2 + S ( 1 )) ≤ C λ 2 . (13)
Applying Gauss–Green next to φ ∇ w \varphi\nabla w φ ∇ w for arbitrary φ ∈ C ∞ ( B 1 / 2 ) \varphi\in C^\infty(B_{1/2}) φ ∈ C ∞ ( B 1/2 ) , and using ∣ ∂ ν w ∣ = d |\partial_\nu w|=d ∣ ∂ ν w ∣ = d on ∂ ∗ U σ ∩ F B ( u ) \partial^*U_\sigma\cap\mathrm{FB}(u) ∂ ∗ U σ ∩ FB ( u ) together with ∣ h ∣ ≤ 2 |h|\le2 ∣ h ∣ ≤ 2 on G σ G_\sigma G σ , (13) yields the weak homogeneous Neumann bound
∣ ∫ U σ ∇ ( w / λ ) ⋅ ∇ φ d x ∣ ≤ C λ ∥ φ ∥ L ∞ ∀ φ ∈ C ∞ ( B 1 / 2 ) . (14) \left|\int_{U_\sigma}\nabla(w/\lambda)\cdot\nabla\varphi\,dx\right|\le C\lambda\|\varphi\|_{L^\infty}\qquad\forall\ \varphi\in C^\infty(B_{1/2}). \tag*{(14)} ∫ U σ ∇ ( w / λ ) ⋅ ∇ φ d x ≤ C λ ∥ φ ∥ L ∞ ∀ φ ∈ C ∞ ( B 1/2 ) . (14)
Crucially, (13) also prevents L 2 L^2 L 2 energy from concentrating inside the vanishing slab { ∣ x n ∣ ≤ δ 1 } \{|x_n|\le\delta_1\} { ∣ x n ∣ ≤ δ 1 } . Pick χ ∈ C ∞ ( B 1 / 2 ) \chi\in C^\infty(B_{1/2}) χ ∈ C ∞ ( B 1/2 ) with 0 ≤ χ ≤ 1 0\le\chi\le1 0 ≤ χ ≤ 1 and χ ≡ 1 \chi\equiv1 χ ≡ 1 on B 3 / 8 B_{3/8} B 3/8 , and set f = χ 2 w 2 f=\chi^2w^2 f = χ 2 w 2 . By Cauchy–Schwarz, (9) , and (11) , ∫ U σ ∣ ∇ f ∣ d x ≤ C λ 2 \int_{U_\sigma}|\nabla f|\,dx\le C\lambda^2 ∫ U σ ∣∇ f ∣ d x ≤ C λ 2 . On ∂ ∗ U σ ∩ F B ( u ) \partial^*U_\sigma\cap\mathrm{FB}(u) ∂ ∗ U σ ∩ FB ( u ) , decompose 1 = − b ⋅ ν + d 1=-b\cdot\nu+d 1 = − b ⋅ ν + d : applying Gauss–Green to the vector field f b fb f b bounds the integral of f ( − b ⋅ ν ) f(-b\cdot\nu) f ( − b ⋅ ν ) by ∫ U σ ∣ ∇ f ∣ d x + C S ( 1 ) \int_{U_\sigma}|\nabla f|\,dx+CS(1) ∫ U σ ∣∇ f ∣ d x + C S ( 1 ) , while boundedness of f f f and (13) bound the integral of f d fd f d by C λ 2 C\lambda^2 C λ 2 , and the trace of f f f on G σ G_\sigma G σ contributes at most C S ( 1 ) CS(1) C S ( 1 ) . Consequently, the zero extension f 1 U σ f1_{U_\sigma} f 1 U σ belongs to B V ( R n ) BV(\mathbb{R}^n) B V ( R n ) with total variation O ( λ 2 ) O(\lambda^2) O ( λ 2 ) , and the Euclidean B V BV B V –Sobolev embedding yields the improved integrability bound
∥ w ∥ L 2 n / ( n − 1 ) ( U σ ∩ B 3 / 8 ) ≤ C λ . (15) \|w\|_{L^{2n/(n-1)}(U_\sigma\cap B_{3/8})}\le C\lambda. \tag*{(15)} ∥ w ∥ L 2 n / ( n − 1 ) ( U σ ∩ B 3/8 ) ≤ C λ . (15)
Step 4: Harmonic linearization via Neumann reflection. We now establish the affine excess improvement by compactness as λ → 0 \lambda\to0 λ → 0 . Along any sequence with λ → 0 \lambda\to0 λ → 0 , the slab width in (3) shrinks to zero, so U ± ∩ B 1 U_{\pm} \cap B_1 U ± ∩ B 1 converge to the upper and lower half-balls B 1 ± = B 1 ∩ { ± x n > 0 } B_1^{\pm} = B_1 \cap\{\pm x_n > 0\} B 1 ± = B 1 ∩ { ± x n > 0 } . For the normalized defects v ± = ( u − L ± ) / λ v_{\pm} = (u-L_{\pm})/\lambda v ± = ( u − L ± ) / λ , Hölder’s inequality and (15) show that on any thin neighborhood T ε = B 3 / 8 ∩ { ∣ x n ∣ < ε } T_{\varepsilon} = B_{3/8} \cap\{|x_n| < \varepsilon\} T ε = B 3/8 ∩ { ∣ x n ∣ < ε } ,
∫ T ε ∩ U ± ∣ v ± ∣ 2 d x ≤ C ∣ T ε ∣ 1 / n = O ( ε 1 / n ) , \int_{T_{\varepsilon}\cap U_{\pm}} |v_{\pm}|^2 \, dx \le C|T_{\varepsilon}|^{1/n} = O(\varepsilon^{1/n}), ∫ T ε ∩ U ± ∣ v ± ∣ 2 d x ≤ C ∣ T ε ∣ 1/ n = O ( ε 1/ n ) ,
uniformly in λ \lambda λ . Away from { x n = 0 } \{x_n = 0\} { x n = 0 } , standard interior harmonic estimates give compact convergence on compact subsets of B 3 / 8 ± B_{3/8}^{\pm} B 3/8 ± , so letting ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 upgrades convergence of the zero extensions v ± 1 U ± v_{\pm}\mathbf{1}_{U_{\pm}} v ± 1 U ± to strong convergence in L 2 ( B 1 / 3 ) L^2(B_{1/3}) L 2 ( B 1/3 ) toward harmonic limits v ˉ ± \bar{v}_{\pm} v ˉ ± on B 1 / 3 ± B_{1/3}^{\pm} B 1/3 ± satisfying ∥ v ˉ ± ∥ L 2 ( B 1 / 3 ± ) ≤ 1 \|\bar{v}_{\pm}\|_{L^2(B_{1/3}^{\pm})} \le1 ∥ v ˉ ± ∥ L 2 ( B 1/3 ± ) ≤ 1 . Simultaneously, (11) provides weak L 2 L^2 L 2 convergence of ( ∇ v ± ) 1 U ± (\nabla v_{\pm})\mathbf{1}_{U_{\pm}} ( ∇ v ± ) 1 U ± to ( ∇ v ˉ ± ) 1 B 1 / 3 ± (\nabla\bar{v}_{\pm})\mathbf{1}_{B_{1/3}^{\pm}} ( ∇ v ˉ ± ) 1 B 1/3 ± , and passing to the limit in (14) shows that each v ˉ ± \bar{v}_{\pm} v ˉ ± satisfies the weak homogeneous Neumann boundary condition ∂ n v ˉ ± = 0 \partial_n\bar{v}_{\pm} = 0 ∂ n v ˉ ± = 0 along B 1 / 3 ∩ { x n = 0 } B_{1/3} \cap\{x_n = 0\} B 1/3 ∩ { x n = 0 } . Even reflection across { x n = 0 } \{x_n = 0\} { x n = 0 } therefore extends v ˉ ± \bar{v}_{\pm} v ˉ ± harmonically across the nodal hyperplane to all of B 1 / 3 B_{1/3} B 1/3 .
Let p ± ( x ) p_{\pm}(x) p ± ( x ) be the first-order Taylor polynomials of these even-reflected harmonic functions at the origin; by evenness in x n x_n x n , ∂ n p ± = 0 \partial_n p_{\pm} = 0 ∂ n p ± = 0 , so ∇ p ± \nabla p_{\pm} ∇ p ± are purely tangential and bounded by a dimensional constant, while interior harmonic Taylor estimates give ∑ ± ∥ v ˉ ± − p ± ∥ L 2 ( B θ ± ) 2 ≤ C 0 θ n + 4 \sum_{\pm}\|\bar{v}_{\pm}-p_{\pm}\|_{L^2(B_{\theta}^{\pm})}^2 \le C_0\theta^{n+4} ∑ ± ∥ v ˉ ± − p ± ∥ L 2 ( B θ ± ) 2 ≤ C 0 θ n + 4 for all θ ∈ ( 0 , 1 / 32 ) \theta\in(0,1/32) θ ∈ ( 0 , 1/32 ) . Define updated unit-slope affine maps L ± ′ L_{\pm}' L ± ′ by normalizing the slope of L ± + λ p ± L_{\pm}+\lambda p_{\pm} L ± + λ p ± while keeping its value at the origin: because b ± → ± e n b_{\pm}\to\pm e_n b ± → ± e n and ∂ n p ± = 0 \partial_n p_{\pm}=0 ∂ n p ± = 0 , we have
∣ b ± + λ ∇ p ± ∣ − 1 = λ b ± ⋅ ∇ p ± + O ( λ 2 ) = o ( λ ) , |b_{\pm}+\lambda\nabla p_{\pm}|-1=\lambda b_{\pm}\cdot\nabla p_{\pm}+O(\lambda^2)=o(\lambda), ∣ b ± + λ ∇ p ± ∣ − 1 = λ b ± ⋅ ∇ p ± + O ( λ 2 ) = o ( λ ) ,
so slope normalization perturbs L ± + λ p ± L_{\pm}+\lambda p_{\pm} L ± + λ p ± on B θ B_{\theta} B θ by at most o ( λ ) θ o(\lambda)\theta o ( λ ) θ . Fix θ ∈ ( 0 , 1 / 32 ) \theta\in(0,1/32) θ ∈ ( 0 , 1/32 ) small enough that C 0 θ 2 < 1 / 16 C_0\theta^2<1/16 C 0 θ 2 < 1/16 . By strong L 2 L^2 L 2 convergence (via a contradiction sequence if the bound failed for arbitrarily small λ \lambda λ ), for all sufficiently small λ \lambda λ we obtain
θ − n − 2 ∑ ± ∫ U ± ∩ B θ ( u − L ± ′ ) 2 d x ≤ 1 8 λ 2 . (16) \theta^{-n-2}\sum_{\pm}\int_{U_{\pm}\cap B_{\theta}}(u-L_{\pm}')^2\,dx\le\frac{1}{8}\lambda^2. \tag*{(16)} θ − n − 2 ± ∑ ∫ U ± ∩ B θ ( u − L ± ′ ) 2 d x ≤ 8 1 λ 2 . (16)
Step 5: Controlling the terminal neck content and closing the contraction. Choose an ( n − 2 ) (n-2) ( n − 2 ) -dimensional subspace M ⊂ b + ⊥ ∩ b − ⊥ M\subset b_{+}^{\perp}\cap b_{-}^{\perp} M ⊂ b + ⊥ ∩ b − ⊥ . On U ± U_{\pm} U ± we have P M ∇ u = P M ( ∇ u − b ± ) P_M\nabla u=P_M(\nabla u-b_{\pm}) P M ∇ u = P M ( ∇ u − b ± ) , so ∣ P M ∇ u ∣ ≤ ∣ h ∣ |P_M\nabla u|\le|h| ∣ P M ∇ u ∣ ≤ ∣ h ∣ , whereas on the residual set Y Y Y we simply use ∣ P M ∇ u ∣ ≤ ∣ ∇ u ∣ ≤ 1 |P_M\nabla u|\le|\nabla u|\le1 ∣ P M ∇ u ∣ ≤ ∣∇ u ∣ ≤ 1 . Combining (11) with the volume bound ∣ Y ∩ B 1 ∣ = o ( λ 2 ) |Y\cap B_1|=o(\lambda^2) ∣ Y ∩ B 1 ∣ = o ( λ 2 ) from (10) yields
∫ B 1 / 2 ∩ Ω ∣ P M ∇ u ∣ 2 d x ≤ C λ 2 . \int_{B_{1/2}\cap\Omega}|P_M\nabla u|^2\,dx\le C\lambda^2. ∫ B 1/2 ∩ Ω ∣ P M ∇ u ∣ 2 d x ≤ C λ 2 .
Moreover, for small λ \lambda λ , every neck terminal meeting B θ B_{\theta} B θ has doubled intrinsic ball contained in B θ + C ℓ ( 1 ) ⊂ B 1 / 4 B_{\theta+C\ell(1)}\subset B_{1/4} B θ + C ℓ ( 1 ) ⊂ B 1/4 . Applying (8) from Lemma 2.2 (at s = θ s=\theta s = θ and r = 1 r=1 r = 1 in the rescaled variables) gives Q ( θ ) ≤ C 1 λ 2 Q(\theta)\le C_1\lambda^2 Q ( θ ) ≤ C 1 λ 2 , with C 1 C_1 C 1 depending only on n n n and the fixed aperture δ \delta δ .
We now fix κ > 0 \kappa>0 κ > 0 so that κ C 1 θ 2 − n ≤ 1 / 8 \kappa C_1\theta^{2-n}\le1/8 κ C 1 θ 2 − n ≤ 1/8 , and then fix λ ∗ > 0 \lambda_*>0 λ ∗ > 0 small enough for all smallness steps above at this fixed ( θ , κ ) (\theta,\kappa) ( θ , κ ) . Adding κ θ 2 − n Q ( θ ) ≤ 1 8 λ 2 \kappa\theta^{2-n}Q(\theta)\le\frac18\lambda^2 κ θ 2 − n Q ( θ ) ≤ 8 1 λ 2 to (16) gives E ( θ , L ′ ) 2 ≤ 1 4 λ 2 = 1 4 E ( 1 , L ) 2 \mathcal{E}(\theta,L')^2\le\frac14\lambda^2=\frac14\mathcal{E}(1,L)^2 E ( θ , L ′ ) 2 ≤ 4 1 λ 2 = 4 1 E ( 1 , L ) 2 , completing the proof. □
Proof of Theorem 1.1. Fix z ∈ Z z \in\mathcal{Z} z ∈ Z , build the localized tree of Lemma 2.1 at root radius R R R , and set r 0 = R / 32 r_0=R/32 r 0 = R /32 . By Lemma 2.1 (i), max N ℓ N = o ( R ) \max_N \ell_N=o(R) max N ℓ N = o ( R ) , so for all large R R R the doubled intrinsic balls associated with terminals meeting B r 0 B_{r_0} B r 0 lie inside B R / 4 B_{R/4} B R /4 . Let e R ∈ S n − 1 e_R\in\mathbb{S}^{n-1} e R ∈ S n − 1 be an approximate wedge direction at scale R R R given by [16 Lemma 5.6] , and choose any ( n − 2 ) (n-2) ( n − 2 ) -plane M R ⊂ e R ⊥ M_R\subset e_R^\perp M R ⊂ e R ⊥ . Rescaling B R ( z ) B_R(z) B R ( z ) to B 1 ( 0 ) B_1(0) B 1 ( 0 ) and rotating e R e_R e R to e n e_n e n , the rescaled functions converge uniformly to ∣ x n ∣ |x_n| ∣ x n ∣ on B 1 B_1 B 1 , so interior harmonic derivative convergence off { x n = 0 } \{x_n=0\} { x n = 0 } combined with ∣ ∇ u ∣ ≤ 1 |\nabla u|\leq1 ∣∇ u ∣ ≤ 1 gives R − n ∫ B R / 2 ∩ Ω ∣ P M R ∇ u ∣ 2 d x = o ( 1 ) R^{-n}\int_{B_{R/2}\cap\Omega}|P_{M_R}\nabla u|^2\,dx=o(1) R − n ∫ B R /2 ∩ Ω ∣ P M R ∇ u ∣ 2 d x = o ( 1 ) as R → ∞ R\to\infty R → ∞ . Applying (8) with s = r 0 s=r_0 s = r 0 and r = R r=R r = R therefore yields the initial neck bound
r 0 2 − n Q ( r 0 ) = 32 n − 2 R 2 − n Q ( r 0 ) = o ( 1 ) ( R → ∞ ) . (17) r_0^{2-n}Q(r_0)=32^{n-2}R^{2-n}Q(r_0)=o(1)\qquad(R\to\infty). \tag*{(17)} r 0 2 − n Q ( r 0 ) = 3 2 n − 2 R 2 − n Q ( r 0 ) = o ( 1 ) ( R → ∞ ) . (17)
Take L 0 , ± = ± t L_{0,\pm}=\pm t L 0 , ± = ± t for the wedge coordinate t = ( x − z ) ⋅ e t=(x-z)\cdot e t = ( x − z ) ⋅ e from (3) at radius r 0 r_0 r 0 . Since ± t > − δ r 0 r 0 \pm t>-\delta_{r_0}r_0 ± t > − δ r 0 r 0 on U ± ∩ B r 0 U_\pm\cap B_{r_0} U ± ∩ B r 0 , we have
sup U ± ∩ B r 0 ∣ u − L 0 , ± ∣ ≤ ∥ u − ∣ t ∣ ∥ L ∞ ( B r 0 ) + 2 δ r 0 r 0 = o ( r 0 ) . \sup_{U_\pm\cap B_{r_0}}|u-L_{0,\pm}|\leq\|u-|t|\|_{L^\infty(B_{r_0})}+2\delta_{r_0}r_0=o(r_0). U ± ∩ B r 0 sup ∣ u − L 0 , ± ∣ ≤ ∥ u − ∣ t ∣ ∥ L ∞ ( B r 0 ) + 2 δ r 0 r 0 = o ( r 0 ) .
The normalized affine term in (9) thus tends to zero. Together with (17) , this gives E ( r 0 , L 0 ) ≤ λ ∗ \mathcal{E}(r_0,L_0)\leq\lambda_* E ( r 0 , L 0 ) ≤ λ ∗ once we fix a sufficiently large root radius R R R .
Iterating Lemma 2.3 on this fixed tree along the geometric scales r j = θ j r 0 r_j=\theta^j r_0 r j = θ j r 0 produces unit-slope pairs L j L_j L j satisfying E ( r j , L j ) ≤ 2 − j E ( r 0 , L 0 ) \mathcal{E}(r_j,L_j)\leq2^{-j}\mathcal{E}(r_0,L_0) E ( r j , L j ) ≤ 2 − j E ( r 0 , L 0 ) . Because the fixed neck terminal N ∗ N_* N ∗ containing z z z has a strictly positive radius ℓ ∗ > 0 \ell_*>0 ℓ ∗ > 0 and meets B r j B_{r_j} B r j for every j ≥ 0 j\geq0 j ≥ 0 , we arrive at
0 < ℓ ∗ n − 2 ≤ Q ( r j ) ≤ κ − 1 r j n − 2 E ( r j , L j ) 2 ≤ κ − 1 r j n − 2 4 − j E ( r 0 , L 0 ) 2 ⟶ 0 ( j → ∞ ) , 0<\ell_*^{n-2}\leq Q(r_j)\leq\kappa^{-1}r_j^{n-2}\mathcal{E}(r_j,L_j)^2\leq\kappa^{-1}r_j^{n-2}4^{-j}\mathcal{E}(r_0,L_0)^2\longrightarrow0\qquad(j\to\infty), 0 < ℓ ∗ n − 2 ≤ Q ( r j ) ≤ κ − 1 r j n − 2 E ( r j , L j ) 2 ≤ κ − 1 r j n − 2 4 − j E ( r 0 , L 0 ) 2 ⟶ 0 ( j → ∞ ) ,
a contradiction. Thus no nonflat entire classical stable solution exists, so every entire classical stable solution satisfies D 2 u ≡ 0 D^2u\equiv0 D 2 u ≡ 0 in Ω \Omega Ω .
Finally, when D 2 u = 0 D^2u=0 D 2 u = 0 in Ω \Omega Ω , ∇ u \nabla u ∇ u is constant on each connected component of Ω \Omega Ω . If ∇ u ≡ 0 \nabla u\equiv0 ∇ u ≡ 0 on a component, that component has empty boundary inside R n \mathbb{R}^n R n , so Ω = R n \Omega=\mathbb{R}^n Ω = R n and u u u is a positive constant. Otherwise, each component has constant nonzero gradient—of unit length whenever FB ( u ) ≠ ∅ \operatorname{FB}(u)\neq\varnothing FB ( u ) = ∅ —and since u = 0 u=0 u = 0 on its finite boundary, the component must be an open affine half-space. At most two disjoint open half-spaces can exist in R n \mathbb{R}^n R n , and if there are two, their normals must be opposite. Thus u u u is one-dimensional in every case. □
Classification up to dimension six and sharpness in dimension seven
Proof of Corollary 1.2. By the smooth stable cone classification [14 Theorem 3 and Remark 4] , which incorporates [21 ] in dimensions five and six, and the minimizing cone of De Silva and Jerison [12 ] , one has n c o n e ∗ > 6 n_{\mathrm{cone}}^*>6 n cone ∗ > 6 . Theorem 1.1 therefore yields D 2 u ≡ 0 D^2u\equiv0 D 2 u ≡ 0 and one-dimensionality in dimensions 4 ≤ n ≤ 6 4\leq n\leq6 4 ≤ n ≤ 6 , while dimensions n = 2 , 3 n=2,3 n = 2 , 3 are established in [10 , 14 ] . The affine component analysis at the end of the proof of Theorem 1.1 , together with ∣ ∇ u ∣ = 1 |\nabla u|=1 ∣∇ u ∣ = 1 on FB ( u ) \operatorname{FB}(u) FB ( u ) and smooth local one-sidedness of Ω \Omega Ω (which precludes contact of two parallel free-boundary hyperplanes and forces b < a b<a b < a ), yields the stated family of profiles. Conversely, every listed profile has flat free boundary with H = 0 H=0 H = 0 , so Q u [ ϕ ; R n ] = ∫ Ω ∣ ∇ ϕ ∣ 2 d x ≥ 0 Q_u[\phi;\mathbb{R}^n]=\int_{\Omega}|\nabla\phi|^2\,dx\geq0 Q u [ ϕ ; R n ] = ∫ Ω ∣∇ ϕ ∣ 2 d x ≥ 0 and u u u is stable. □
Corollary 2.4. The smallest dimension admitting a nonflat entire classical stable solution to (1) is n e n t i r e ∗ = 7 n_{\mathrm{entire}}^*=7 n entire ∗ = 7 .
Proof. Corollary 1.2 rules out dimensions 2 ≤ n ≤ 6 2 \le n \le6 2 ≤ n ≤ 6 , and in dimension n = 1 n=1 n = 1 harmonicity trivially gives u ′ ′ = 0 u''=0 u ′′ = 0 on every positive interval. In dimension n = 7 n=7 n = 7 , let C 0 C_0 C 0 be a nonflat one-homogeneous global minimizer constructed by De Silva and Jerison [12 ] . Applying the Hardt–Simon-type foliation theorem of Edelen, Spolaor and Velichkov [13 Theorem 1.5] to C 0 C_0 C 0 yields an entire global minimizer V V V in R 7 \mathbb{R}^7 R 7 with empty singular set whose blow-down at infinity is C 0 C_0 C 0 . Smoothness of FB ( V ) \operatorname{FB}(V) FB ( V ) and global minimality imply that V V V is an entire classical stable solution. Since the affine-component classification of solutions with D 2 u ≡ 0 D^2u \equiv0 D 2 u ≡ 0 in the proof of Theorem 1.1 holds in arbitrary dimension, any solution with vanishing positive-phase Hessian has blow-down equal to 0 0 0 , a half-space profile, or a unit wedge; because C 0 C_0 C 0 is none of these, V V V is genuinely nonflat. □
Local curvature estimates and compactness of classical solutions
In this section we assemble the local regularity, multi-sheet graphical convergence, and point-picking tools needed both to deduce the interior curvature bound (Corollary 1.3 ) and to extract finite-Morse-index bubbles in § 4 . Throughout what follows, a function of the form ∣ ( x − p ) ⋅ e ∣ \lvert(x-p)\cdot e\rvert ∣( x − p ) ⋅ e ∣ with e ∈ S n − 1 e\in\mathbb{S}^{n-1} e ∈ S n − 1 is called a unit wedge (and a centered unit wedge when p = 0 p=0 p = 0 ); note that its classical Hessian vanishes identically on its positive phase R n ∖ { x : ( x − p ) ⋅ e = 0 } \mathbb{R}^n \setminus\{x:(x-p)\cdot e=0\} R n ∖ { x : ( x − p ) ⋅ e = 0 } .
Interior Lipschitz bounds and graphical sheet convergence
We begin with a universal interior Lipschitz estimate for classical solutions whose free boundary meets an inner ball, independent of stability or index.
Lemma 3.1. Let u u u be a classical solution to (1.1) in B R ⊂ R n B_R\subset\mathbb{R}^n B R ⊂ R n such that FB ( u ) ∩ B R / 2 ≠ ∅ \operatorname{FB}(u)\cap B_{R/2}\ne\varnothing FB ( u ) ∩ B R /2 = ∅ . For every θ ∈ ( 0 , 1 ) \theta\in(0,1) θ ∈ ( 0 , 1 ) ,
R − 1 ∥ u ∥ L ∞ ( B θ R ) + ∥ ∇ u ∥ L ∞ ( B θ R ∩ { u > 0 } ) ≤ C ( n , θ ) . R^{-1}\lVert u\rVert_{L^\infty(B_{\theta R})}+\lVert\nabla u\rVert_{L^\infty(B_{\theta R}\cap\{u>0\})}\le C(n,\theta). R − 1 ∥ u ∥ L ∞ ( B θ R ) + ∥ ∇ u ∥ L ∞ ( B θ R ∩ { u > 0 }) ≤ C ( n , θ ) . Proof. Rescale to R = 1 R=1 R = 1 and write d ( x ) = dist ( x , { u = 0 } ) d(x)=\operatorname{dist}(x,\{u=0\}) d ( x ) = dist ( x , { u = 0 }) . Fix θ ∈ [ 1 / 2 , 1 ) \theta\in[1/2,1) θ ∈ [ 1/2 , 1 ) and set δ 0 = 1 4 ( 1 − θ ) \delta_0=\frac14(1-\theta) δ 0 = 4 1 ( 1 − θ ) . At any point x ∈ B θ ∩ { u > 0 } x\in B_\theta\cap\{u>0\} x ∈ B θ ∩ { u > 0 } with d ( x ) ≤ δ 0 d(x)\le\delta_0 d ( x ) ≤ δ 0 , the ball B d ( x ) ( x ) B_{d(x)}(x) B d ( x ) ( x ) lies in B 1 ∩ { u > 0 } B_1\cap\{u>0\} B 1 ∩ { u > 0 } and touches FB ( u ) \operatorname{FB}(u) FB ( u ) at some contact point q q q . By Harnack’s inequality, u ≥ c n u ( x ) u\ge c_nu(x) u ≥ c n u ( x ) on ∂ B d ( x ) / 2 ( x ) \partial B_{d(x)/2}(x) ∂ B d ( x ) /2 ( x ) ; comparing u u u on the annulus B d ( x ) ( x ) ∖ B d ( x ) / 2 ( x ) B_{d(x)}(x)\setminus B_{d(x)/2}(x) B d ( x ) ( x ) ∖ B d ( x ) /2 ( x ) with the radial harmonic barrier taking values c n u ( x ) c_nu(x) c n u ( x ) on the inner sphere and 0 0 0 on the outer sphere, the Hopf boundary condition ∣ ∇ u ( q ) ∣ = 1 \lvert\nabla u(q)\rvert=1 ∣ ∇ u ( q )∣ = 1 yields 1 ≥ c n ′ u ( x ) / d ( x ) 1\ge c'_nu(x)/d(x) 1 ≥ c n ′ u ( x ) / d ( x ) , and interior harmonic gradient estimates on B d ( x ) / 2 ( x ) B_{d(x)/2}(x) B d ( x ) /2 ( x ) then give ∣ ∇ u ( x ) ∣ ≤ C n \lvert\nabla u(x)\rvert\le C_n ∣ ∇ u ( x )∣ ≤ C n . When d ( x ) > δ 0 d(x)>\delta_0 d ( x ) > δ 0 , join a point of FB ( u ) ∩ B 1 / 2 \operatorname{FB}(u)\cap B_{1/2} FB ( u ) ∩ B 1/2 to x x x by a line segment: after its last exit from { d ≤ δ 0 } \{d\le\delta_0\} { d ≤ δ 0 } , the remaining segment stays at distance at least δ 0 \delta_0 δ 0 from { u = 0 } ∪ ∂ B 1 \{u=0\}\cup\partial B_1 { u = 0 } ∪ ∂ B 1 and is covered by a Harnack chain of length O θ ( 1 ) O_\theta(1) O θ ( 1 ) , bounding u ( x ) u(x) u ( x ) and hence ∣ ∇ u ( x ) ∣ \lvert\nabla u(x)\rvert ∣ ∇ u ( x )∣ . □
along positive subintervals of arbitrary line segments extends this gradient bound to a uniform ambient Lipschitz constant.
Lemma 3.3. Suppose classical solutions v j v_j v j converge locally uniformly on an open set G ⊂ R n G \subset\mathbb{R}^n G ⊂ R n to a function v v v , and suppose their positive-phase Hessians D 2 v j D^2v_j D 2 v j are uniformly bounded on every compact subset of G G G . Then for every 1 ≤ p < ∞ 1 \le p < \infty 1 ≤ p < ∞ ,
D 2 v j 1 { v j > 0 } ⟶ D 2 v 1 { v > 0 } strongly in L l o c p ( G ) . D^2v_j\mathbf{1}_{\{v_j>0\}} \longrightarrow D^2v\mathbf{1}_{\{v>0\}} \quad\text{strongly in } L^p_{\mathrm{loc}}(G). D 2 v j 1 { v j > 0 } ⟶ D 2 v 1 { v > 0 } strongly in L loc p ( G ) . Proof. Let F ⊂ G F \subset G F ⊂ G be the set of subsequential accumulation points in G G G of the free boundaries FB ( v j ) \operatorname{FB}(v_j) FB ( v j ) . Clearly F F F is relatively closed in G G G and contained in { v = 0 } \{v=0\} { v = 0 } . Given q ∈ F q \in F q ∈ F , pass to a subsequence with q j ∈ FB ( v j ) q_j \in\operatorname{FB}(v_j) q j ∈ FB ( v j ) converging to q q q and inward unit normals converging to ν i n \nu_{\mathrm{in}} ν in . For every sufficiently small r > 0 r>0 r > 0 , passing to the limit in (18) at t = r / 2 t=r/2 t = r /2 gives v ( q + r 2 ν i n ) ≥ r / 4 v(q+\frac{r}{2}\nu_{\mathrm{in}}) \ge r/4 v ( q + 2 r ν in ) ≥ r /4 ; combined with the uniform local Lipschitz bound from Remark 3.2, this produces a ball of radius c r cr cr inside B r ( q ) ∩ { v > 0 } B_r(q) \cap\{v>0\} B r ( q ) ∩ { v > 0 } . Hence F F F is locally porous and has Lebesgue measure zero. On any small ball B ′ ⋐ G ∖ F B' \Subset G \setminus F B ′ ⋐ G ∖ F , FB ( v j ) ∩ B ′ \operatorname{FB}(v_j) \cap B' FB ( v j ) ∩ B ′ is empty for all large j j j , so each v j v_j v j is harmonic on B ′ B' B ′ (possibly identically zero), and interior harmonic convergence shows that D 2 v j 1 { v j > 0 } → D 2 v 1 { v > 0 } D^2v_j\mathbf{1}_{\{v_j>0\}} \to D^2v\mathbf{1}_{\{v>0\}} D 2 v j 1 { v j > 0 } → D 2 v 1 { v > 0 } uniformly on compact subsets of G ∖ F G \setminus F G ∖ F (using the minimum principle to distinguish v > 0 v>0 v > 0 from v ≡ 0 v\equiv0 v ≡ 0 on B ′ B' B ′ ). Since ∣ F ∣ = 0 |F|=0 ∣ F ∣ = 0 , dominated convergence yields strong convergence in L l o c p ( G ) L^p_{\mathrm{loc}}(G) L loc p ( G ) .□
Lemma 3.4. Let v j v_j v j be classical solutions on expanding balls B R j B_{R_j} B R j (R j → ∞ R_j \to\infty R j → ∞ ) with a uniform Lipschitz bound ∥ ∇ v j ∥ L ∞ ( B R j ) ≤ C 0 \|\nabla v_j\|_{L^\infty(B_{R_j})} \le C_0 ∥∇ v j ∥ L ∞ ( B R j ) ≤ C 0 . If v j → v v_j \to v v j → v locally uniformly on R n \mathbb{R}^n R n , then ∣ ∇ v ∣ ≤ 1 |\nabla v| \le1 ∣∇ v ∣ ≤ 1 throughout { v > 0 } \{v>0\} { v > 0 } .
Proof. By interior harmonic convergence, v v v is harmonic on { v > 0 } \{v>0\} { v > 0 } . If { v > 0 } = R n \{v>0\}=\mathbb{R}^n { v > 0 } = R n , Liouville’s theorem for nonnegative harmonic functions implies that v v v is constant and ∇ v ≡ 0 \nabla v \equiv0 ∇ v ≡ 0 . Suppose instead that { v = 0 } ≠ ∅ \{v=0\}\ne\varnothing { v = 0 } = ∅ and L : = sup { v > 0 } ∣ ∇ v ∣ > 1 L:=\sup_{\{v>0\}}|\nabla v|>1 L := sup { v > 0 } ∣∇ v ∣ > 1 . Following the nearest-zero rescaling of [18 proof of Proposition A.5] , choose x k ∈ { v > 0 } x_k \in\{v>0\} x k ∈ { v > 0 } with ∣ ∇ v ( x k ) ∣ → L |\nabla v(x_k)| \to L ∣∇ v ( x k ) ∣ → L and set d k = dist ( x k , { v = 0 } ) > 0 d_k=\operatorname{dist}(x_k,\{v=0\})>0 d k = dist ( x k , { v = 0 }) > 0 . The rescaled functions w k ( x ) = d k − 1 v ( x k + d k x ) w_k(x)=d_k^{-1}v(x_k+d_kx) w k ( x ) = d k − 1 v ( x k + d k x ) are uniformly L L L -Lipschitz, vanish at a point of ∂ B 1 \partial B_1 ∂ B 1 , and are harmonic on B 1 B_1 B 1 . Along a subsequence, w k → w w_k \to w w k → w locally uniformly on R n \mathbb{R}^n R n and smoothly on B 1 B_1 B 1 , with ∣ ∇ w ( 0 ) ∣ = L > 0 |\nabla w(0)|=L>0 ∣∇ w ( 0 ) ∣ = L > 0 (so w > 0 w>0 w > 0 on B 1 B_1 B 1 by the minimum principle) and ∣ ∇ w ∣ ≤ L |\nabla w|\le L ∣∇ w ∣ ≤ L on { w > 0 } \{w>0\} { w > 0 } . Since Δ ∣ ∇ w ∣ 2 = 2 ∣ D 2 w ∣ 2 ≥ 0 \Delta|\nabla w|^2=2|D^2w|^2\ge0 Δ∣∇ w ∣ 2 = 2∣ D 2 w ∣ 2 ≥ 0 on the connected component of { w > 0 } \{w>0\} { w > 0 } containing 0 0 0 , the strong maximum principle forces D 2 w ≡ 0 D^2w\equiv0 D 2 w ≡ 0 there, making that component an open half-space on which w w w has slope L L L . On the other hand, selecting a diagonal sequence v j ( k ) v_{j(k)} v j ( k ) approximating v v v within d k / k d_k/k d k / k on x k + d k B k x_k+d_kB_k x k + d k B k exhibits w w w as a locally uniform limit of rescaled classical solutions, so w w w is a viscosity supersolution in the sense of [18 Definition A.1 and Lemma A.3] ; testing its half-space boundary with an interior tangent ball forces L ≤ 1 L\le1 L ≤ 1 , a contradiction.□
We next show that under a positive-phase Hessian bound, approximating free boundaries decompose locally into either a single smooth graph or two ordered oppositely oriented graphs enclosing a collapsing zero-phase layer (compare [10 proof of Lemma 5.1] ).
Lemma 3.5. Suppose classical solutions v j v_j v j converge locally uniformly near a point q q q with uniformly bounded positive-phase Hessians there, and assume q q q is an accumulation point of FB ( v j ) \operatorname{FB}(v_j) FB ( v j ) . After passing to a subsequence, in a fixed coordinate cylinder around q q q the free boundaries FB ( v j ) \operatorname{FB}(v_j) FB ( v j ) consist either of a single smooth graph or of two ordered smooth graphs with opposite inward normals separated by the zero phase. In smaller graph coordinates,
the graphs and their positive-side solutions converge smoothly; in the single-graph case v v v has a nonempty open zero phase in the cylinder. In the two-graph case the two limiting graphs meet at q q q .
Proof. For q j ∈ F B ( v j ) q_j \in\mathrm{FB}(v_j) q j ∈ FB ( v j ) near q q q , choose coordinates with vertical axis along the inward unit normal e j = ∇ v j ( q j ) e_j = \nabla v_j(q_j) e j = ∇ v j ( q j ) . Differentiating v j = 0 v_j = 0 v j = 0 along F B ( v j ) \mathrm{FB}(v_j) FB ( v j ) with ∣ ∇ v j ∣ = 1 |\nabla v_j| = 1 ∣∇ v j ∣ = 1 bounds the second fundamental form of F B ( v j ) \mathrm{FB}(v_j) FB ( v j ) by the positive-phase Hessian bound K K K . Lifting radial rays from the tangent disk of radius r r r via orthogonal projection keeps the unit normal within angle π / 4 \pi/4 π /4 of e j e_j e j up to a uniform radius r ≪ K − 1 r \ll K^{-1} r ≪ K − 1 , yielding a smooth free-boundary graph over the r r r -disk with slope at most C K r CKr C K r . For a point a a a on the graph, ∂ e j v j ( a ) ≥ 3 / 4 \partial_{e_j}v_j(a) \ge3/4 ∂ e j v j ( a ) ≥ 3/4 . Integrating in the fixed vertical direction, up to any first exit, gives
v j ( a + t e j ) ≥ 3 t / 4 − K t 2 / 2 > 0 v_j(a+te_j) \ge3t/4 - Kt^2/2 > 0 v j ( a + t e j ) ≥ 3 t /4 − K t 2 /2 > 0
for 0 < t ≤ c r 0 < t \le cr 0 < t ≤ cr and r r r small. This rules out such an exit and gives a positive strip of height comparable to r r r over a smaller disk. Retaining v j v_j v j above the graph and setting it to zero below defines a classical one-sided solution in a ball B c r ( q j ) B_{cr}(q_j) B cr ( q j ) that deviates from the unit half-space profile by at most C K r 2 CKr^2 C K r 2 . Shrinking r r r allows us to apply one-sided flatness regularity [11 ] in the quantitative form of [10 Lemma 3.7] . Together with implicit differentiation, this gives uniform C k C^k C k bounds for every k ≥ 1 k \ge1 k ≥ 1 on both the graph and its positive-side solution.
To count sheets near q q q , extract a first limiting graph through a sequence q j → q q_j \to q q j → q with limiting inward normal e e e . Any second sequence of free-boundary points tending to q q q off this sheet generates a second graph whose limiting inward normal e ′ e' e ′ cannot be transverse to e e e (otherwise its tangent plane at q q q would enter the open positive strip of the first sheet, forcing free-boundary points of v j v_j v j inside { v j > 0 } \{v_j > 0\} { v j > 0 } ) and cannot equal + e +e + e (otherwise one graph would lie strictly above the other over a common disk and enter its positive strip, since F B ( v j ) \mathrm{FB}(v_j) FB ( v j ) is embedded). Thus at most one additional sheet can approach q q q , necessarily with opposite limiting inward normal e ′ = − e e' = -e e ′ = − e . If no additional sheet approaches q q q , a smaller cylinder contains only the first sheet for large j j j . Otherwise choose the second sheet through a sequence tending to q q q , then shrink the cylinder to exclude any further components by the same argument. In the two-sheet case, non-intersection with each other’s positive strips forces the upper graph to have upward positive phase and the lower graph to have downward positive phase. The region between the graphs lies in { v j = 0 } \{v_j = 0\} { v j = 0 } , and the two limit graphs meet at q q q by construction.
Lemma 3.6. Let v ∈ C ( R n ) v \in C(\mathbb{R}^n) v ∈ C ( R n ) be nonnegative, harmonic on { v > 0 } \{v > 0\} { v > 0 } , and satisfy ∣ ∇ v ∣ ≤ 1 |\nabla v| \le1 ∣∇ v ∣ ≤ 1 on { v > 0 } \{v > 0\} { v > 0 } . If two smooth one-sided Bernoulli free-boundary graphs of v v v with opposite inward normals touch tangentially at a point and bound opposite positive components, then v v v is globally a unit wedge.
Proof. On each adjacent positive branch, the function h = 1 − ∣ ∇ v ∣ 2 ≥ 0 h = 1 - |\nabla v|^2 \ge0 h = 1 − ∣∇ v ∣ 2 ≥ 0 is superharmonic (Δ h = − 2 ∣ D 2 v ∣ 2 ≤ 0 \Delta h = -2|D^2v|^2 \le0 Δ h = − 2∣ D 2 v ∣ 2 ≤ 0 ) and vanishes along its smooth free boundary, where harmonicity and ∣ ∇ v ∣ = 1 |\nabla v| = 1 ∣∇ v ∣ = 1 give ∂ ν i n h = 2 H \partial_{\nu_{\mathrm{in}}}h = 2H ∂ ν in h = 2 H for the inward mean curvature H = div F B ν i n H = \operatorname{div}_{\mathrm{FB}}\nu_{\mathrm{in}} H = div FB ν in . By the Hopf boundary point lemma, both inward mean curvatures H ± H_{\pm} H ± are nonnegative. Writing the ordered graphs in common tangent coordinates at the contact point as x n = f + ( x ′ ) ≥ f − ( x ′ ) x_n = f_+(x') \ge f_-(x') x n = f + ( x ′ ) ≥ f − ( x ′ ) , optimality at contact gives D 2 ( f + − f − ) ≥ 0 D^2(f_+ - f_-) \ge0 D 2 ( f + − f − ) ≥ 0 there, so H + + H − = − Δ x ′ ( f + − f − ) ≤ 0 H_+ + H_- = -\Delta_{x'}(f_+ - f_-) \le0 H + + H − = − Δ x ′ ( f + − f − ) ≤ 0 . Hence H + = H − = 0 H_+ = H_- = 0 H + = H − = 0 at contact; the strict Hopf lemma forces h ≡ 0 h \equiv0 h ≡ 0 , hence D 2 v ≡ 0 D^2v \equiv0 D 2 v ≡ 0 , on each local positive branch. Unique continuation extends affineness across both positive components in R n \mathbb{R}^n R n , which by continuity and vanishing on their common boundary hyperplane must be complementary unit half-spaces.
Lemma 3.7. Let v j v_j v j be classical solutions on expanding balls D j ↑ R n D_j \uparrow\mathbb{R}^n D j ↑ R n with 0 ∈ F B ( v j ) 0 \in\mathrm{FB}(v_j) 0 ∈ FB ( v j ) , a uniform Lipschitz bound ∥ ∇ v j ∥ L ∞ ( D j ) ≤ C 0 \lVert\nabla v_j\rVert_{L^\infty(D_j)} \le C_0 ∥ ∇ v j ∥ L ∞ ( D j ) ≤ C 0 , and uniformly bounded positive-phase Hessians on every compact subset of R n \mathbb{R}^n R n . Assume there exists a fixed ball on which sup ∣ D 2 v j ∣ ≥ c 0 > 0 \sup|D^2v_j| \ge c_0 > 0 sup ∣ D 2 v j ∣ ≥ c 0 > 0 (including positive-side boundary traces). Then, up to a subsequence, v j v_j v j converges locally uniformly on R n \mathbb{R}^n R n to a nonflat entire classical solution v v v , with smooth convergence in local positive-side graph charts, 1 { v j > 0 } → 1 { v > 0 } \mathbf{1}_{\{v_j>0\}} \to\mathbf{1}_{\{v>0\}} 1 { v j > 0 } → 1 { v > 0 } in L l o c 1 ( R n ) L^1_{\mathrm{loc}}(\mathbb{R}^n) L loc 1 ( R n ) , and
ind ( v ) ≤ lim inf j → ∞ ind ( v j ; D j ) . \operatorname{ind}(v) \le\liminf_{j\to\infty} \operatorname{ind}(v_j;D_j). ind ( v ) ≤ j → ∞ lim inf ind ( v j ; D j ) .
Proof. Pass first to a subsequence realizing the index liminf. The Arzelá–Ascoli theorem gives a locally uniform limit v v v after further extraction. It is harmonic on { v > 0 } \{v>0\} { v > 0 } and satisfies ∣ ∇ v ∣ ≤ 1 |\nabla v| \le1 ∣∇ v ∣ ≤ 1 there by Lemma 3.4 .
We first rule out vanishing of the Hessian throughout { v > 0 } \{v>0\} { v > 0 } . If it vanished there, Lemma 3.3 would give ∫ B 2 R ∩ { v j > 0 } ∣ D 2 v j ∣ n d x → 0 \int_{B_{2R}\cap\{v_j>0\}} |D^2v_j|^n\,dx \to0 ∫ B 2 R ∩ { v j > 0 } ∣ D 2 v j ∣ n d x → 0 for any fixed R R R . Choose B R B_R B R to contain the fixed ball from the peak hypothesis. Hessian ε \varepsilon ε -regularity [16 Lemma 3.4] , rescaled from B 2 R B_{2R} B 2 R , would then give ∥ D 2 v j ∥ L ∞ ( B R ∩ { v j > 0 } ) → 0 \lVert D^2v_j\rVert_{L^\infty(B_R\cap\{v_j>0\})} \to0 ∥ D 2 v j ∥ L ∞ ( B R ∩ { v j > 0 }) → 0 . The same bound holds for the one-sided traces, contradicting the peak hypothesis.
Every q ∈ F B ( v ) q \in\mathrm{FB}(v) q ∈ FB ( v ) is approached by points of F B ( v j ) \mathrm{FB}(v_j) FB ( v j ) . Otherwise a fixed ball about q q q would contain no approximating free boundary along a subsequence. In that ball each v j v_j v j is harmonic, including the identically zero case, so the nonnegative harmonic limit vanishes identically there by the minimum principle. Conversely, an accumulation point of F B ( v j ) \mathrm{FB}(v_j) FB ( v j ) lies in { v = 0 } \{v=0\} { v = 0 } and is approached by positive points of v v v , by passing to the limit in the inward-segment bound (18) after extraction of the normals. Hence it lies in F B ( v ) \mathrm{FB}(v) FB ( v ) .
Apply Lemma 3.5 at a point of F B ( v ) \mathrm{FB}(v) FB ( v ) . Two limiting sheets would meet there and Lemma 3.6 would force a unit wedge, excluded by nonvanishing of the positive-phase Hessian. Thus the limit has one classical sheet and an open zero side at each free boundary point. The local graph convergence follows by the same compactness argument on every subsequence. In particular, v v v is an entire classical solution whose Hessian is nonzero somewhere in { v > 0 } \{v>0\} { v > 0 } .
Positivity persists on compact subsets of { v > 0 } \{v>0\} { v > 0 } . Each component of int { v = 0 } \operatorname{int}\{v=0\} int { v = 0 } meets the zero side of a one-sheeted boundary chart, since it has a boundary point in F B ( v ) \mathrm{FB}(v) FB ( v ) . A point of the component can be joined to such a zero side by a compact path in the component. A small neighborhood of this path avoids F B ( v j ) \mathrm{FB}(v_j) FB ( v j ) for large j j j , by the characterization of its accumulation points, and contains a point below the approximating sheet in the zero phase. The zero phase therefore persists near the original point. Since H n ( F B ( v ) ) = 0 \mathcal{H}^n(\mathrm{FB}(v))=0 H n ( FB ( v )) = 0 , this proves 1 { v j > 0 } → 1 { v > 0 } \mathbf{1}_{\{v_j>0\}} \to\mathbf{1}_{\{v>0\}} 1 { v j > 0 } → 1 { v > 0 } in L l o c 1 ( R n ) L^1_{\mathrm{loc}}(\mathbb{R}^n) L loc 1 ( R n ) .
Finally, given an m m m -dimensional subspace of C c ∞ ( R n ) C_c^\infty(\mathbb{R}^n) C c ∞ ( R n ) on which Q v Q_v Q v is negative definite, restricting a basis ϕ 1 , … , ϕ m \phi_1,\ldots,\phi_m ϕ 1 , … , ϕ m to { v j > 0 } \{v_j>0\} { v j > 0 } yields admissible tests on D j D_j D j for large j j j . Phase convergence in L l o c 1 L^1_{\mathrm{loc}} L loc 1 and smooth single-sheeted free-boundary convergence imply entrywise convergence of the m × m m\times m m × m Gram matrices of Q v j [ ⋅ ; D j ] Q_{v_j}[\cdot;D_j] Q v j [ ⋅ ; D j ] to the negative definite matrix of Q v [ ⋅ ; R n ] Q_v[\cdot;\mathbb{R}^n] Q v [ ⋅ ; R n ] , proving lower semicontinuity of the Morse index. □
Point-picking selection, interior curvature bounds, and finite-index entire profiles
Lemma 3.8. Let u j u_j u j be classical solutions on a bounded open set D ⊂ R n D \subset\mathbb{R}^n D ⊂ R n with uniformly bounded Lipschitz norms on compact subsets of D D D , and let d j d_j d j be nonnegative 1-Lipschitz functions on D D D satisfying d j ≤ dist ( ⋅ , ∂ D ) d_j \leq\operatorname{dist}(\cdot,\partial D) d j ≤ dist ( ⋅ , ∂ D ) . If there exist points x j ∈ K ∩ { u j > 0 } x_j \in K \cap\{u_j > 0\} x j ∈ K ∩ { u j > 0 } in a fixed compact set K ⋐ D K \Subset D K ⋐ D such that d j ( x j ) ∣ D 2 u j ( x j ) ∣ → ∞ d_j(x_j)|D^2u_j(x_j)| \to\infty d j ( x j ) ∣ D 2 u j ( x j ) ∣ → ∞ , then there exist points y j y_j y j in a compact subset of D D D and free-boundary centers z j ∈ FB ( u j ) z_j \in\operatorname{FB}(u_j) z j ∈ FB ( u j ) such that, setting λ j = ∣ D 2 u j ( y j ) ∣ − 1 \lambda_j = |D^2u_j(y_j)|^{-1} λ j = ∣ D 2 u j ( y j ) ∣ − 1 ,
λ j → 0 , ∣ y j − z j ∣ = O ( λ j ) , λ j = o ( d j ( y j ) ) , \lambda_j \to0,\qquad|y_j-z_j|=O(\lambda_j),\qquad\lambda_j=o(d_j(y_j)), λ j → 0 , ∣ y j − z j ∣ = O ( λ j ) , λ j = o ( d j ( y j )) , and the rescalings u ~ j ( x ) = λ j − 1 u j ( z j + λ j x ) \widetilde u_j(x)=\lambda_j^{-1}u_j(z_j+\lambda_j x) u j ( x ) = λ j − 1 u j ( z j + λ j x ) have positive-phase Hessians bounded by 2 on every fixed ball for large j j j , converging along a subsequence to a nonflat entire classical limit as in Lemma 3.7 .
Proof. Set ρ j = 1 4 d j ( x j ) ≤ 1 4 dist ( x j , ∂ D ) \rho_j=\frac{1}{4}d_j(x_j)\leq\frac{1}{4}\operatorname{dist}(x_j,\partial D) ρ j = 4 1 d j ( x j ) ≤ 4 1 dist ( x j , ∂ D ) . For every y ∈ B ρ j ( x j ) ‾ y\in\overline{B_{\rho_j}(x_j)} y ∈ B ρ j ( x j ) ,
dist ( y , ∂ D ) ≥ 3 4 dist ( x j , ∂ D ) ≥ 3 4 dist ( K , ∂ D ) . \operatorname{dist}(y,\partial D)\geq\frac{3}{4}\operatorname{dist}(x_j,\partial D)\geq\frac{3}{4}\operatorname{dist}(K,\partial D). dist ( y , ∂ D ) ≥ 4 3 dist ( x j , ∂ D ) ≥ 4 3 dist ( K , ∂ D ) .
Choose y j ∈ { u j > 0 } ∩ B ρ j ( x j ) ‾ y_j\in\{u_j>0\}\cap\overline{B_{\rho_j}(x_j)} y j ∈ { u j > 0 } ∩ B ρ j ( x j ) maximizing the weighted quantity ( ρ j − ∣ x − x j ∣ ) ∣ D 2 u j ( x ) ∣ (\rho_j-|x-x_j|)|D^2u_j(x)| ( ρ j − ∣ x − x j ∣ ) ∣ D 2 u j ( x ) ∣ (including positive-side boundary traces). Put t j = ρ j − ∣ y j − x j ∣ > 0 t_j=\rho_j-|y_j-x_j|>0 t j = ρ j − ∣ y j − x j ∣ > 0 and a j = ∣ D 2 u j ( y j ) ∣ = λ j − 1 a_j=|D^2u_j(y_j)|=\lambda_j^{-1} a j = ∣ D 2 u j ( y j ) ∣ = λ j − 1 ; comparing at x j x_j x j gives a j t j ≥ ρ j ∣ D 2 u j ( x j ) ∣ → ∞ a_jt_j\geq\rho_j|D^2u_j(x_j)|\to\infty a j t j ≥ ρ j ∣ D 2 u j ( x j ) ∣ → ∞ . Because all selection balls stay uniformly away from ∂ D \partial D ∂ D , a single Lipschitz constant L L L bounds ∣ ∇ u j ∣ |\nabla u_j| ∣∇ u j ∣ on them. If y j ∈ { u j > 0 } y_j\in\{u_j>0\} y j ∈ { u j > 0 } , interior harmonic derivative estimates on the ball around y j y_j y j of radius 1 2 min { t j , dist ( y j , FB ( u j ) ) } \frac{1}{2}\min\{t_j,\operatorname{dist}(y_j,\operatorname{FB}(u_j))\} 2 1 min { t j , dist ( y j , FB ( u j ))} give
a j ≤ C L min { t j , dist ( y j , FB ( u j ) ) } , a_j\leq\frac{CL}{\min\{t_j,\operatorname{dist}(y_j,\operatorname{FB}(u_j))\}}, a j ≤ min { t j , dist ( y j , FB ( u j ))} C L ,
which in view of a j t j → ∞ a_jt_j\to\infty a j t j → ∞ yields a nearest point z j ∈ FB ( u j ) z_j\in\operatorname{FB}(u_j) z j ∈ FB ( u j ) at distance O ( λ j ) O(\lambda_j) O ( λ j ) from y j y_j y j (taking z j = y j z_j=y_j z j = y j if y j ∈ FB ( u j ) y_j\in\operatorname{FB}(u_j) y j ∈ FB ( u j ) ). By choice of y j y_j y j , we have ∣ D 2 u j ∣ ≤ 2 a j |D^2u_j|\leq2a_j ∣ D 2 u j ∣ ≤ 2 a j throughout B t j / 2 ( y j ) ∩ { u j > 0 } B_{t_j/2}(y_j)\cap\{u_j>0\} B t j /2 ( y j ) ∩ { u j > 0 } , while 1-Lipschitz continuity of d j d_j d j gives d j ( y j ) ≥ d j ( x j ) − ρ j = 3 4 d j ( x j ) d_j(y_j)\geq d_j(x_j)-\rho_j=\frac{3}{4}d_j(x_j) d j ( y j ) ≥ d j ( x j ) − ρ j = 4 3 d j ( x j ) , whence λ j / d j ( y j ) ≤ 1 3 a j ρ j → 0 \lambda_j/d_j(y_j)\leq\frac{1}{3a_j\rho_j}\to0 λ j / d j ( y j ) ≤ 3 a j ρ j 1 → 0 and t j / λ j → ∞ t_j/\lambda_j\to\infty t j / λ j → ∞ . Since ∣ y j − z j ∣ = O ( λ j ) |y_j-z_j|=O(\lambda_j) ∣ y j − z j ∣ = O ( λ j ) , for large j j j we have B t j / 4 ( z j ) ⊂ B t j / 2 ( y j ) B_{t_j/4}(z_j)\subset B_{t_j/2}(y_j) B t j /4 ( z j ) ⊂ B t j /2 ( y j ) . Thus the rescalings have uniform Lipschitz and Hessian bounds on B t j / ( 4 λ j ) B_{t_j/(4\lambda_j)} B t j / ( 4 λ j ) , whose radii tend to infinity. The rescaled Hessian peak stays at bounded distance from the origin, so Lemma 3.7 applies. □ \square □
Proof of Corollary 1.3 . By translation and Bernoulli scaling we may assume z = 0 ∈ FB ( u ) z=0\in\operatorname{FB}(u) z = 0 ∈ FB ( u ) and R = 1 R=1 R = 1 . If sup B 1 / 2 ∩ { u j > 0 } ∣ D 2 u j ∣ → ∞ \sup_{B_{1/2}\cap\{u_j>0\}}|D^2u_j|\to\infty sup B 1/2 ∩ { u j > 0 } ∣ D 2 u j ∣ → ∞ along a sequence of classical stable solutions u j u_j u j in B 1 B_1 B 1 with 0 ∈ FB ( u j ) 0\in\operatorname{FB}(u_j) 0 ∈ FB ( u j ) , then Lemma 3.1 supplies uniform interior Lipschitz bounds, and applying Lemma 3.8 with d j ( x ) = dist ( x , ∂ B 1 ) d_j(x)=\operatorname{dist}(x,\partial B_1) d j ( x ) = dist ( x , ∂ B 1 ) on K = B 1 / 2 ‾ K=\overline{B_{1/2}} K = B 1/2 produces a nonflat entire classical limit v v v with ind ( v ) ≤ 0 \operatorname{ind}(v)\leq0 ind ( v ) ≤ 0 , contradicting Corollary 1.2. Finally, differentiating u = 0 u=0 u = 0 twice tangentially along FB ( u ) \operatorname{FB}(u) FB ( u ) and using ∣ ∇ u ∣ = 1 |\nabla u|=1 ∣∇ u ∣ = 1 identifies I I \mathrm{II} II with the tangential restriction of the positive-side trace of D 2 u D^2u D 2 u . □ \square □
Combining Corollary 1.2 with [15, Theorem 1.2]} now yields the complete list of nonflat finite-Morse-index bubbles in dimensions three through six.
Corollary 3.9. For 3 ≤ n ≤ 6 3\leq n\leq6 3 ≤ n ≤ 6 , every entire classical solution u u u to (1) of finite Morse index is axially symmetric up to a rigid motion. If u u u is not one-dimensional, then up to rigid motion and Bernoulli rescaling it is either the radial exterior-ball solution
u ( x ) = 1 − ∣ x ∣ 2 − n n − 2 ( ∣ x ∣ > 1 ) , u(x)=\frac{1-|x|^{2-n}}{n-2}\qquad(|x|>1), u ( x ) = n − 2 1 − ∣ x ∣ 2 − n ( ∣ x ∣ > 1 ) , extended by zero on { ∣ x ∣ ≤ 1 } \{|x|\leq1\} { ∣ x ∣ ≤ 1 } , or an axially symmetric catenoid-type (two-ended) solution as in [15 Theorem 1.2] , which is even in x n x_n x n and satisfies ∂ x n u > 0 \partial_{x_n}u>0 ∂ x n u > 0 on { x n > 0 } ∩ { u > 0 } \{x_n>0\}\cap\{u>0\} { x n > 0 } ∩ { u > 0 }
and ∂ r u < 0 \partial_r u < 0 ∂ r u < 0 on { r > 0 } ∩ { u > 0 } \{r > 0\} \cap\{u > 0\} { r > 0 } ∩ { u > 0 } , where r = ∣ x ′ ∣ r = |x'| r = ∣ x ′ ∣ . Its zero set has the form
{ ( x ′ , x n ) : r ≥ 1 , ∣ x n ∣ ≤ g ( r ) } , g ( 1 ) = 0 , g ′ ( r ) > 0 ( r > 1 ) , lim r → ∞ g ( r ) < ∞ . \{(x',x_n): r \ge1,\ |x_n| \le g(r)\}, \qquad g(1)=0,\quad g'(r)>0\ (r>1),\qquad\lim_{r\to\infty}g(r)<\infty. {( x ′ , x n ) : r ≥ 1 , ∣ x n ∣ ≤ g ( r )} , g ( 1 ) = 0 , g ′ ( r ) > 0 ( r > 1 ) , r → ∞ lim g ( r ) < ∞.
Proof. Theorem 1.2 of [15 ] applies for 3 ≤ n < min { n e n t i r e ∗ , n c o n e ∗ } 3 \le n < \min\{n^*_{\mathrm{entire}},n^*_{\mathrm{cone}}\} 3 ≤ n < min { n entire ∗ , n cone ∗ } (see [15 (1.3)–(1.5)] ). Since n e n t i r e ∗ = 7 n^*_{\mathrm{entire}}=7 n entire ∗ = 7 by Corollary 2.4 and n c o n e ∗ = 7 n^*_{\mathrm{cone}}=7 n cone ∗ = 7 by the cone classification [14 Theorem 3 and Remark 4] and the seven-dimensional example [12 ] , both thresholds equal 7, so the classification holds for all 3 ≤ n ≤ 6 3 \le n \le6 3 ≤ n ≤ 6 .□ \square □
Hessian energy identity at bounded Morse index
To prove Theorem 1.4, extracting scale-separated bubbles via Lemma 3.8 is only the first step: one must still show that no diffuse L n L^n L n -Hessian energy remains in the annular neck regions separating bubble scales from macroscopic scales or from colliding companion bubbles. We accomplish this in three steps: first pinning the Weiss density across intermediate scales to force full asymptotic positive-phase measure and classify intermediate limits; next establishing integrable Hessian decay on intermediate annuli; and finally organizing multi-bubble collisions via a Voronoi nearest-center scale tree.
Logarithmic Weiss bound and classification of intermediate blow-ups
Throughout this section we write A ( a , b ) = B b ∖ B a ‾ A(a,b)=B_b\setminus\overline{B_a} A ( a , b ) = B b ∖ B a for open concentric annuli and ω n = ∣ B 1 ∣ \omega_n=|B_1| ω n = ∣ B 1 ∣ for the volume of the Euclidean unit ball. For a function u u u on B R B_R B R with 0 ∈ F B ( u ) 0\in\mathrm{FB}(u) 0 ∈ FB ( u ) , recall the scale-invariant Weiss energy functional
W ( u , r ) = r − n ∫ B r ( ∣ ∇ u ∣ 2 + 1 { u > 0 } ) d x − r − n − 1 ∫ ∂ B r u 2 d H n − 1 . W(u,r)=r^{-n}\int_{B_r}\bigl(|\nabla u|^2+\mathbf{1}_{\{u>0\}}\bigr)\,dx-r^{-n-1}\int_{\partial B_r}u^2\,d\mathcal{H}^{n-1}. W ( u , r ) = r − n ∫ B r ( ∣∇ u ∣ 2 + 1 { u > 0 } ) d x − r − n − 1 ∫ ∂ B r u 2 d H n − 1 .
Our starting point is a quantitative logarithmic upper bound showing that any Lipschitz classical solution cannot exceed the full-density threshold ω n \omega_n ω n by more than O ( 1 / log ( R / r ) ) O(1/\log(R/r)) O ( 1/ log ( R / r )) .
Lemma 4.1. Let u u u be a classical solution to (1.1) in B R ⊂ R n B_R\subset\mathbb{R}^n B R ⊂ R n satisfying u ( 0 ) = 0 u(0)=0 u ( 0 ) = 0 and ∥ ∇ u ∥ L ∞ ( B R ) ≤ L \|\nabla u\|_{L^\infty(B_R)}\le L ∥∇ u ∥ L ∞ ( B R ) ≤ L . Then
W ( u , r ) ≤ ω n + n ω n L 2 2 log ( R / r ) ∀ 0 < r < R . W(u,r)\le\omega_n+\frac{n\omega_nL^2}{2\log(R/r)}
\qquad\forall\,0<r<R. W ( u , r ) ≤ ω n + 2 log ( R / r ) n ω n L 2 ∀ 0 < r < R .
In particular, every entire classical solution with 0 ∈ F B ( u ) 0\in\mathrm{FB}(u) 0 ∈ FB ( u ) satisfies W ( u , r ) ≤ ω n W(u,r)\le\omega_n W ( u , r ) ≤ ω n for all r > 0 r>0 r > 0 .
Proof. Integrating Δ u = 0 \Delta u=0 Δ u = 0 against u u u by parts on B r ∩ { u > 0 } B_r\cap\{u>0\} B r ∩ { u > 0 } and using Weiss’s monotonicity identity for stationary solutions [23 Theorem 3.1] (see also [19 Proposition 3.4] ) give, for almost every r ∈ ( 0 , R ) r\in(0,R) r ∈ ( 0 , R ) ,
W ( u , r ) = r − n ∣ B r ∩ { u > 0 } ∣ + r − n ∫ ∂ B r u ( ∂ r u − u / r ) d H n − 1 , W(u,r)=r^{-n}|B_r\cap\{u>0\}|+r^{-n}\int_{\partial B_r}u(\partial_r u-u/r)\,d\mathcal{H}^{n-1}, W ( u , r ) = r − n ∣ B r ∩ { u > 0 } ∣ + r − n ∫ ∂ B r u ( ∂ r u − u / r ) d H n − 1 ,
∂ r W ( u , r ) = 2 r − n ∫ ∂ B r ( ∂ r u − u / r ) 2 d H n − 1 . \partial_rW(u,r)=2r^{-n}\int_{\partial B_r}(\partial_r u-u/r)^2\,d\mathcal{H}^{n-1}. ∂ r W ( u , r ) = 2 r − n ∫ ∂ B r ( ∂ r u − u / r ) 2 d H n − 1 .
Bounding the first term in W ( u , r ) W(u,r) W ( u , r ) by ω n \omega_n ω n , using 0 ≤ u ≤ L r 0\le u\le Lr 0 ≤ u ≤ L r on ∂ B r \partial B_r ∂ B r with H n − 1 ( ∂ B r ) = n ω n r n − 1 \mathcal{H}^{n-1}(\partial B_r)=n\omega_nr^{n-1} H n − 1 ( ∂ B r ) = n ω n r n − 1 , and applying the Cauchy–Schwarz inequality yield the Riccati-type differential inequality
( W ( u , r ) − ω n ) 2 ≤ 1 2 n ω n L 2 r ∂ r W ( u , r ) . \bigl(W(u,r)-\omega_n\bigr)^2\le\frac{1}{2}n\omega_nL^2r\,\partial_rW(u,r). ( W ( u , r ) − ω n ) 2 ≤ 2 1 n ω n L 2 r ∂ r W ( u , r ) .
Whenever W ( u , r ) > ω n W(u,r)>\omega_n W ( u , r ) > ω n , monotonicity guarantees W ( u , s ) ≥ W ( u , r ) > ω n W(u,s)\ge W(u,r)>\omega_n W ( u , s ) ≥ W ( u , r ) > ω n for all s ∈ [ r , R ) s\in[r,R) s ∈ [ r , R ) ; integrating − ∂ s ( W ( u , s ) − ω n ) − 1 -\partial_s(W(u,s)-\omega_n)^{-1} − ∂ s ( W ( u , s ) − ω n ) − 1 over [ r , R ′ ] [r,R'] [ r , R ′ ] and letting R ′ ↑ R R'\uparrow R R ′ ↑ R gives the stated bound. For an entire solution, ∣ ∇ u ∣ ≤ 1 |\nabla u|\le1 ∣∇ u ∣ ≤ 1 globally by [16 Lemma 3.5] , so letting R → ∞ R\to\infty R → ∞ gives W ( u , r ) ≤ ω n W(u,r)\le\omega_n W ( u , r ) ≤ ω n .□ \square □
Proposition 4.2. Let 3 ≤ n ≤ 6 3 \le n \le6 3 ≤ n ≤ 6 , and let u j u_j u j be classical solutions in a fixed neighborhood of an interior point p p p with uniformly bounded Lipschitz norms and Morse indices at most I I I . Suppose z j ∈ F B ( u j ) z_j \in\mathrm{FB}(u_j) z j ∈ FB ( u j ) converge to p p p and, for scales λ j → 0 \lambda_j \to0 λ j → 0 , the bubble rescalings λ j − 1 u j ( z j + λ j x ) \lambda_j^{-1}u_j(z_j+\lambda_j x) λ j − 1 u j ( z j + λ j x ) converge as in Lemma 3.7 to a nonflat entire finite-index solution V V V . Then for every intermediate scale sequence s j → 0 s_j \to0 s j → 0 satisfying λ j / s j → 0 \lambda_j/s_j \to0 λ j / s j → 0 , every locally uniform subsequential limit w w w of w j ( x ) = s j − 1 u j ( z j + s j x ) w_j(x)=s_j^{-1}u_j(z_j+s_jx) w j ( x ) = s j − 1 u j ( z j + s j x ) is either identically zero or a centered unit wedge ∣ x ⋅ e ∣ |x\cdot e| ∣ x ⋅ e ∣ , and along the same subsequence 1 { w j > 0 } → 1 \mathbf{1}_{\{w_j>0\}}\to1 1 { w j > 0 } → 1 strongly in L l o c 1 ( R n ) L^1_{\mathrm{loc}}(\mathbb{R}^n) L loc 1 ( R n ) .
Proof. Pass to a subsequence along which w j → w w_j \to w w j → w locally uniformly on R n \mathbb{R}^n R n . We first note that for any locally uniform limit f f f of uniformly Lipschitz classical solutions, f f f is nonnegative, locally Lipschitz, and harmonic on { f > 0 } \{f>0\} { f > 0 } ; testing Δ f = 0 \Delta f=0 Δ f = 0 against ψ ( f − ε ) + \psi(f-\varepsilon)_+ ψ ( f − ε ) + for ψ ∈ C c ∞ ( R n ) \psi\in C^\infty_c(\mathbb{R}^n) ψ ∈ C c ∞ ( R n ) and letting ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 (using ∇ f = 0 \nabla f=0 ∇ f = 0 a.e. on { f = 0 } \{f=0\} { f = 0 } ) gives the distributional identity
∫ R n ψ ∣ ∇ f ∣ 2 d x = − ∫ R n f ∇ f ⋅ ∇ ψ d x = 1 2 ∫ R n f 2 Δ ψ d x . \int_{\mathbb{R}^n}\psi|\nabla f|^2\,dx=-\int_{\mathbb{R}^n}f\nabla f\cdot\nabla\psi\,dx=\frac{1}{2}\int_{\mathbb{R}^n}f^2\Delta\psi\,dx. ∫ R n ψ ∣∇ f ∣ 2 d x = − ∫ R n f ∇ f ⋅ ∇ ψ d x = 2 1 ∫ R n f 2 Δ ψ d x .
Because the same identity holds for each w j w_j w j , uniform convergence implies convergence of localized L 2 L^2 L 2 gradient norms, upgrading weak H l o c 1 H^1_{\mathrm{loc}} H loc 1 convergence to strong convergence w j → w w_j\to w w j → w in H l o c 1 ( R n ) H^1_{\mathrm{loc}}(\mathbb{R}^n) H loc 1 ( R n ) .
Next we compute lim R → ∞ W ( V , R ) \lim_{R\to\infty}W(V,R) lim R → ∞ W ( V , R ) . By Corollary 3.9, the blow-down V R ( x ) = R − 1 V ( R x ) V_R(x)=R^{-1}V(Rx) V R ( x ) = R − 1 V ( R x ) as R → ∞ R\to\infty R → ∞ converges locally uniformly (and hence strongly in H l o c 1 H^1_{\mathrm{loc}} H loc 1 ) to 0 0 0 when V V V is an exterior-ball solution, and to a centered unit wedge ∣ x ⋅ e ∣ |x\cdot e| ∣ x ⋅ e ∣ when V V V is a catenoid-type solution (via the asymptotic end expansion of [15 Remark 1.4] ). In both cases 1 { V R > 0 } → 1 \mathbf{1}_{\{V_R>0\}}\to1 1 { V R > 0 } → 1 in L l o c 1 ( R n ) L^1_{\mathrm{loc}}(\mathbb{R}^n) L loc 1 ( R n ) , and the Dirichlet and boundary terms in W ( V , R ) = W ( V R , 1 ) W(V,R)=W(V_R,1) W ( V , R ) = W ( V R , 1 ) cancel in the limit (trivially for 0 0 0 , and by integration by parts on the two complementary half-spaces for ∣ x ⋅ e ∣ |x\cdot e| ∣ x ⋅ e ∣ ), giving lim R → ∞ W ( V , R ) = ω n \lim_{R\to\infty}W(V,R)=\omega_n lim R → ∞ W ( V , R ) = ω n .
Now fix any t > 0 t>0 t > 0 and R > 0 R>0 R > 0 . Since s j t > λ j R s_jt>\lambda_jR s j t > λ j R for large j j j , Weiss monotonicity and smooth/phase convergence at the bubble scale give
lim inf j → ∞ W ( w j , t ) = lim inf j → ∞ W ( u j ( z j + ⋅ ) , s j t ) ≥ lim j → ∞ W ( u j ( z j + ⋅ ) , λ j R ) = W ( V , R ) . \liminf_{j\to\infty}W(w_j,t)=\liminf_{j\to\infty}W(u_j(z_j+\cdot),s_jt)\ge\lim_{j\to\infty}W(u_j(z_j+\cdot),\lambda_jR)=W(V,R). j → ∞ lim inf W ( w j , t ) = j → ∞ lim inf W ( u j ( z j + ⋅ ) , s j t ) ≥ j → ∞ lim W ( u j ( z j + ⋅ ) , λ j R ) = W ( V , R ) .
Letting R → ∞ R\to\infty R → ∞ yields lim inf j → ∞ W ( w j , t ) ≥ ω n \liminf_{j\to\infty}W(w_j,t)\ge\omega_n lim inf j → ∞ W ( w j , t ) ≥ ω n , while applying Lemma 4.1 from radius s j t s_jt s j t up to a fixed macroscopic ball around p p p gives lim sup j → ∞ W ( w j , t ) ≤ ω n \limsup_{j\to\infty}W(w_j,t)\le\omega_n lim sup j → ∞ W ( w j , t ) ≤ ω n . Hence W ( w j , t ) → ω n W(w_j,t)\to\omega_n W ( w j , t ) → ω n for every t > 0 t>0 t > 0 . Passing to the limit via strong H l o c 1 H^1_{\mathrm{loc}} H loc 1 convergence in the integrated Weiss identity
W ( w j , b ) − W ( w j , a ) = 2 ∫ A ( a , b ) ∣ x ∣ − n ( ∂ r w j − w j / ∣ x ∣ ) 2 d x ( 0 < a < b < ∞ ) W(w_j,b)-W(w_j,a)=2\int_{A(a,b)}|x|^{-n}(\partial_rw_j-w_j/|x|)^2\,dx \qquad(0<a<b<\infty) W ( w j , b ) − W ( w j , a ) = 2 ∫ A ( a , b ) ∣ x ∣ − n ( ∂ r w j − w j /∣ x ∣ ) 2 d x ( 0 < a < b < ∞ )
shows that w w w is homogeneous of degree one. Furthermore, if χ ∈ L ∞ ( R n ) \chi\in L^\infty(\mathbb{R}^n) χ ∈ L ∞ ( R n ) is any weak-* subsequential limit of 1 { w j > 0 } \mathbf{1}_{\{w_j>0\}} 1 { w j > 0 } , then passing to the limit in W ( w j , t ) W(w_j,t) W ( w j , t ) gives
ω n = t − n ∫ B t ( ∣ ∇ w ∣ 2 + χ ) d x − t − n − 1 ∫ ∂ B t w 2 d H n − 1 . \omega_n=t^{-n}\int_{B_t}(|\nabla w|^2+\chi)\,dx-t^{-n-1}\int_{\partial B_t}w^2\,d\mathcal{H}^{n-1}. ω n = t − n ∫ B t ( ∣∇ w ∣ 2 + χ ) d x − t − n − 1 ∫ ∂ B t w 2 d H n − 1 .
Because w w w is 1 1 1 -homogeneous and Δ ( w 2 ) = 2 ∣ ∇ w ∣ 2 \Delta(w^2)=2|\nabla w|^2 Δ ( w 2 ) = 2∣∇ w ∣ 2 distributionally, ∫ B t ∣ ∇ w ∣ 2 d x = t − 1 ∫ ∂ B t w 2 d H n − 1 \int_{B_t}|\nabla w|^2\,dx=t^{-1}\int_{\partial B_t}w^2\,d\mathcal{H}^{n-1} ∫ B t ∣∇ w ∣ 2 d x = t − 1 ∫ ∂ B t w 2 d H n − 1 for almost every t > 0 t>0 t > 0 , leaving ∫ B t χ d x = ω n t n = ∣ B t ∣ \int_{B_t}\chi\,dx=\omega_nt^n=|B_t| ∫ B t χ d x = ω n t n = ∣ B t ∣ . Since 0 ≤ χ ≤ 1 0\le\chi\le1 0 ≤ χ ≤ 1 , we conclude χ = 1 \chi=1 χ = 1 and therefore 1 { w j > 0 } → 1 \mathbf{1}_{\{w_j>0\}}\to1 1 { w j > 0 } → 1 strongly in L l o c 1 ( R n ) L^1_{\mathrm{loc}}(\mathbb{R}^n) L loc 1 ( R n ) .
It remains to classify the 1 1 1 -homogeneous limit w w w . Diagonalizing over all balls with rational centers and rational radii, pass to a further subsequence along which each rational ball is either eventually stable or eventually unstable for w j w_j w j . The complement Σ \Sigma Σ of the union O \mathcal{O} O of eventually stable rational balls contains at most I I I points (as I + 1 I+1 I + 1 distinct points in Σ \Sigma Σ could be separated by I + 1 I+1 I + 1 disjoint eventually unstable rational balls, producing I + 1 I+1 I + 1 mutually orthogonal negative-definite directions). On every compact subset of O \mathcal{O} O , Corollary 1.3 and the interior harmonic estimate give uniform positive-phase Hessian bounds for w j w_j w j . Indeed, if B 3 r ( x ) B_{3r}(x) B 3 r ( x ) is contained in an eventually stable rational ball, use the curvature estimate centered at a point of FB ( w j ) ∩ B r / 2 ( x ) \operatorname{FB}(w_j) \cap B_{r/2}(x) FB ( w j ) ∩ B r /2 ( x ) , when such a point exists, with radius 2 r 2r 2 r . Otherwise w j w_j w j is harmonic in B r / 2 ( x ) B_{r/2}(x) B r /2 ( x ) , including the identically zero case. This bounds the Hessian in B r / 4 ( x ) B_{r/4}(x) B r /4 ( x ) in either case. Lemma 3.4 gives ∣ ∇ w ∣ ≤ 1 |\nabla w| \le1 ∣∇ w ∣ ≤ 1 on { w > 0 } \{w > 0\} { w > 0 } .
Suppose w ≢ 0 w \not\equiv0 w ≡ 0 . Then FB ( w ) ∖ { 0 } \operatorname{FB}(w) \setminus\{0\} FB ( w ) ∖ { 0 } must be nonempty: otherwise w > 0 w > 0 w > 0 would be harmonic on R n ∖ { 0 } \mathbb{R}^n \setminus\{0\} R n ∖ { 0 } , which for n ≥ 3 n \ge3 n ≥ 3 has a removable singularity at 0 0 0 and would force w w w to be a positive constant, incompatible with 1 1 1 -homogeneity. By 1 1 1 -homogeneity, FB ( w ) ∖ { 0 } \operatorname{FB}(w) \setminus\{0\} FB ( w ) ∖ { 0 } contains an entire open ray, which necessarily meets O = R n ∖ Σ \mathcal{O} = \mathbb{R}^n \setminus\Sigma O = R n ∖ Σ since # Σ ≤ I \#\Sigma\le I #Σ ≤ I . Choose q ∈ FB ( w ) ∩ O q \in\operatorname{FB}(w) \cap\mathcal{O} q ∈ FB ( w ) ∩ O on that ray. If a ball about q q q contained no approximating free boundary along a subsequence, its nonnegative harmonic limit would vanish identically there by the minimum principle, contrary to q ∈ FB ( w ) q \in\operatorname{FB}(w) q ∈ FB ( w ) . Thus Lemma 3.5 applies at q q q . A single limiting sheet would leave an open zero side. Two sheets separated at q q q would leave an open zero layer. In either case a smaller ball would remain in the zero phase of w j w_j w j , contrary to 1 { w j > 0 } → 1 \mathbf{1}_{\{w_j>0\}} \to1 1 { w j > 0 } → 1 in L l o c 1 L^1_{\mathrm{loc}} L loc 1 . The two limit sheets therefore meet at q q q , and Lemma 3.6 identifies w w w with a unit wedge, centered at the origin by 1 1 1 -homogeneity.
Annular branch splitting and vanishing of neck energy
When an intermediate profile is a centered unit wedge rather than zero, the positivity set on a long dyadic annulus surrounds a thin collapsing zero-phase sheet. We first show that local two-graph descriptions across consecutive scales glue coherently into two globally disjoint classical branches trapped between parallel shifted half-spaces.
Lemma 4.3. Let 0 < s 0 < s 1 / 16 0 < s_0 < s_1/16 0 < s 0 < s 1 /16 , and let u u u be a classical solution to (1) on A ( s 0 / 4 , 16 s 1 ) A(s_0/4,16s_1) A ( s 0 /4 , 16 s 1 ) . Assume that for every s ∈ [ s 0 , s 1 ] s \in[s_0,s_1] s ∈ [ s 0 , s 1 ] there exists e s ∈ S n − 1 e_s \in\mathbb{S}^{n-1} e s ∈ S n − 1 satisfying
∣ u ( x ) − ∣ e s ⋅ x ∣ ∣ ≤ η s ∀ x ∈ A ( s / 4 , 16 s ) , |u(x) - |e_s \cdot x|| \le\eta s \qquad\forall x \in A(s/4,16s), ∣ u ( x ) − ∣ e s ⋅ x ∣∣ ≤ η s ∀ x ∈ A ( s /4 , 16 s ) , such that in orthogonal coordinates x = ξ + τ e s x = \xi+ \tau e_s x = ξ + τ e s (ξ ⊥ e s \xi\perp e_s ξ ⊥ e s ), for every ξ \xi ξ with 3 8 s < ∣ ξ ∣ < 9 s \frac{3}{8}s < |\xi| < 9s 8 3 s < ∣ ξ ∣ < 9 s the intersection of { u = 0 } \{u = 0\} { u = 0 } with { ξ + τ e s : ∣ τ ∣ < s / 16 } \{\xi+ \tau e_s : |\tau| < s/16\} { ξ + τ e s : ∣ τ ∣ < s /16 } is a closed interval whose two endpoints vary smoothly with ξ \xi ξ and have absolute height at most η s \eta s η s . Provided η > 0 \eta> 0 η > 0 is smaller than a dimensional constant, on the trimmed annulus A ( ρ , R ) A(\rho,R) A ( ρ , R ) with ρ = 4 s 0 \rho= 4s_0 ρ = 4 s 0 and R = s 1 / 4 R = s_1/4 R = s 1 /4 one can decompose u = v + + v − u = v_+ + v_- u = v + + v − into two classical solutions with disjoint positivity sets, and for every 2 ρ < s < R / 2 2\rho< s < R/2 2 ρ < s < R /2 orient e s e_s e s so that
( ± e s ⋅ x − C η s ) + ≤ v ± ( x ) ≤ ( ± e s ⋅ x + C η s ) + ∀ x ∈ A ( s / 2 , 2 s ) , (\pm e_s \cdot x - C\eta s)_+ \le v_\pm(x) \le(\pm e_s \cdot x + C\eta s)_+ \qquad\forall x \in A(s/2,2s), ( ± e s ⋅ x − C η s ) + ≤ v ± ( x ) ≤ ( ± e s ⋅ x + C η s ) + ∀ x ∈ A ( s /2 , 2 s ) , with C > 0 C > 0 C > 0 independent of the scale ratio s 1 / s 0 s_1/s_0 s 1 / s 0 .
Proof. At each scale s ∈ [ s 0 , s 1 ] s \in[s_0,s_1] s ∈ [ s 0 , s 1 ] , the two ordered endpoint graphs partition A ( 7 16 s , 8 s ) ∩ { u > 0 } A(\frac{7}{16}s,8s)\cap\{u > 0\} A ( 16 7 s , 8 s ) ∩ { u > 0 } into two locally constant sign classes (determined by sgn τ \operatorname{sgn}\tau sgn τ outside ∣ τ ∣ ≤ η s |\tau| \le\eta s ∣ τ ∣ ≤ η s , and above or below the zero-phase interval inside ∣ τ ∣ ≤ η s |\tau| \le\eta s ∣ τ ∣ ≤ η s , where ∣ ξ ∣ > 3 8 s |\xi| > \frac{3}{8}s ∣ ξ ∣ > 8 3 s holds automatically). Consider the dyadic ladder t = 2 k s 0 ≤ s 1 t = 2^k s_0 \le s_1 t = 2 k s 0 ≤ s 1 with retained shells A ( t , 4 t ) A(t,4t) A ( t , 4 t ) . Evaluating the wedge approximations at scales t t t and 2 t 2t 2 t along the equator { x ⊥ e t : ∣ x ∣ = 3 t } \{x \perp e_t : |x| = 3t\} { x ⊥ e t : ∣ x ∣ = 3 t } gives ∣ P e t ⊥ e 2 t ∣ ≤ C η |P_{e_t^\perp}e_{2t}| \le C\eta ∣ P e t ⊥ e 2 t ∣ ≤ C η , allowing successive orientation of the dyadic normals so that ∣ e t − e 2 t ∣ ≤ C η |e_t-e_{2t}| \le C\eta ∣ e t − e 2 t ∣ ≤ C η . On the overlap A ( 2 t , 4 t ) A(2t,4t) A ( 2 t , 4 t ) , the two scale labels manifestly agree in the deep caps { ± e t ⋅ x > C 0 η t } \{\pm e_t \cdot x > C_0\eta t\} { ± e t ⋅ x > C 0 η t } ; moreover, any positive point of A ( 2 t , 4 t ) A(2t,4t) A ( 2 t , 4 t ) outside those caps can be translated by at most C 1 η t C_1\eta t C 1 η t along ± e t \pm e_t ± e t away from the zero interval into its own deep cap along a segment that remains strictly positive and inside both extended shells A ( 7 16 t , 8 t ) A(\frac{7}{16}t,8t) A ( 16 7 t , 8 t ) and A ( 7 8 t , 16 t ) A(\frac{7}{8}t,16t) A ( 8 7 t , 16 t ) , forcing the t t t - and 2 t 2t 2 t -labels to agree throughout A ( 2 t , 4 t ) ∩ { u > 0 } A(2t,4t)\cap\{u > 0\} A ( 2 t , 4 t ) ∩ { u > 0 } .
Setting v ± = u v_{\pm}=u v ± = u on the resulting global positive components in A ( ρ , R ) A(\rho,R) A ( ρ , R ) and zero elsewhere defines two disjoint classical solutions on A ( ρ , R ) A(\rho,R) A ( ρ , R ) , since locally each inherits one of the two smooth one-sided sheets of u u u . Finally, for any intermediate s ∈ ( 2 ρ , R / 2 ) s\in(2\rho,R/2) s ∈ ( 2 ρ , R /2 ) , choosing the dyadic scale in [ s / 2 , s ] [s/2,s] [ s /2 , s ] orients e s e_s e s within O ( η ) O(\eta) O ( η ) of the overlapping dyadic axes, identifies the local s s s -branches with v ± v_{\pm} v ± on A ( s / 2 , 2 s ) A(s/2,2s) A ( s /2 , 2 s ) , and yields the two-sided half-space trapping bounds both in the caps { ± e s ⋅ x > η s } \{\pm e_s\cdot x>\eta s\} { ± e s ⋅ x > η s } (where v ± = u v_{\pm}=u v ± = u and v ∓ = 0 v_{\mp}=0 v ∓ = 0 ) and in the thin strip { ∣ e s ⋅ x ∣ ≤ η s } \{|e_s\cdot x|\le\eta s\} { ∣ e s ⋅ x ∣ ≤ η s } (where 0 ≤ v ± ≤ 2 η s 0\le v_{\pm}\le2\eta s 0 ≤ v ± ≤ 2 η s ).□
Lemma 4.4. Let 3 ≤ n ≤ 6 3\le n\le6 3 ≤ n ≤ 6 , let a j / b j → 0 a_j/b_j\to0 a j / b j → 0 , and suppose classical solutions u j u_j u j centered at z j ∈ F B ( u j ) z_j\in\mathrm{FB}(u_j) z j ∈ FB ( u j ) have uniformly bounded Lipschitz norms on B 16 b j ( z j ) B_{16b_j}(z_j) B 16 b j ( z j ) . Assume that for every sequence s j s_j s j with a j / s j → 0 a_j/s_j\to0 a j / s j → 0 and s j / b j → 0 s_j/b_j\to0 s j / b j → 0 , the rescalings s j − 1 u j ( z j + s j x ) s_j^{-1}u_j(z_j+s_jx) s j − 1 u j ( z j + s j x ) have uniformly bounded positive-phase Hessians on compact annuli in R n ∖ { 0 } \mathbb{R}^n\setminus\{0\} R n ∖ { 0 } and admit only 0 0 0 or centered unit wedges as locally uniform subsequential limits. Then
lim M → ∞ lim sup j → ∞ μ u j ( B b j / M ( z j ) ∖ B M a j ( z j ) ) = 0. (19) \lim_{M\to\infty}\limsup_{j\to\infty}\mu_{u_j}\bigl(B_{b_j/M}(z_j)\setminus B_{Ma_j}(z_j)\bigr)=0. \tag*{(19)} M → ∞ lim j → ∞ lim sup μ u j ( B b j / M ( z j ) ∖ B M a j ( z j ) ) = 0. (19)
Proof. Translate z j z_j z j to the origin. Choose η > 0 \eta>0 η > 0 small enough for Lemma 4.3 and so that its trapping constant C η C_\eta C η meets the threshold in [15 Theorem 2.2] for exponent 1 / 2 1/2 1/2 . We also choose η \eta η small enough for the one-sided ε \varepsilon ε -regularity estimate below. We claim that there exists a fixed S ≥ 2 S\ge2 S ≥ 2 such that for all large j j j and every s ∈ [ S a j , b j / S ] s\in[S a_j,b_j/S] s ∈ [ S a j , b j / S ] , one of two mutually exclusive regimes holds:
sup A ( s , 2 s ) u j ≤ η s \sup_{A(s,2s)}u_j\le\eta s sup A ( s , 2 s ) u j ≤ η s and F B ( u j ) ∩ A ( s , 2 s ) = ∅ \mathrm{FB}(u_j)\cap A(s,2s)=\varnothing FB ( u j ) ∩ A ( s , 2 s ) = ∅ ;
the two-graph wedge hypotheses of Lemma 4.3 hold at scale s s s .
Were this false, there would exist S k → ∞ S_k\to\infty S k → ∞ , j k → ∞ j_k\to\infty j k → ∞ , and scales s k ∈ [ S k a j k , b j k / S k ] s_k\in[S_k a_{j_k},b_{j_k}/S_k] s k ∈ [ S k a j k , b j k / S k ] violating both (1) and (2). Passing to a locally uniform limit of w k ( x ) = s k − 1 u j k ( s k x ) w_k(x)=s_k^{-1}u_{j_k}(s_kx) w k ( x ) = s k − 1 u j k ( s k x ) , if w k → 0 w_k\to0 w k → 0 , then no free-boundary points can persist in A ( 1 , 2 ) A(1,2) A ( 1 , 2 ) by the inward positivity bound (18) , so (1) holds for large k k k . If instead w k → ∣ e ⋅ x ∣ w_k\to|e\cdot x| w k → ∣ e ⋅ x ∣ , then F B ( w k ) \mathrm{FB}(w_k) FB ( w k ) on compact annuli concentrates inside a vanishing strip around e ⊥ e^\perp e ⊥ . For any convergent sequence of normals with limit ν \nu ν , pass to the limit in the first bound of (18) at a fixed sufficiently small t > 0 t>0 t > 0 to get t ∣ ν ⋅ e ∣ ≥ t − K t 2 / 2 t|\nu\cdot e|\ge t-Kt^2/2 t ∣ ν ⋅ e ∣ ≥ t − K t 2 /2 , with K K K the local uniform Hessian bound. Letting t ↓ 0 t\downarrow0 t ↓ 0 shows ∣ ν k ⋅ e ∣ → 1 |\nu_k\cdot e|\to1 ∣ ν k ⋅ e ∣ → 1 uniformly in the annular region in question. Along each vertical segment { ξ + τ e : ∣ τ ∣ ≤ 1 / 16 } \{\xi+\tau e:|\tau|\le1/16\} { ξ + τ e : ∣ τ ∣ ≤ 1/16 } with 3 / 8 < ∣ ξ ∣ < 9 3/8<|\xi|<9 3/8 < ∣ ξ ∣ < 9 , interior harmonic convergence gives ∂ e w k ( ξ ± 1 16 e ) → ± 1 \partial_e w_k(\xi\pm\frac{1}{16}e)\to\pm1 ∂ e w k ( ξ ± 16 1 e ) → ± 1 ; the uniform Hessian bound rules out both the absence of zeros (which would force ∂ e w k \partial_e w_k ∂ e w k to jump by nearly 2 2 2 across [ − δ , δ ] [-\delta,\delta] [ − δ , δ ] for δ ≪ C − 1 \delta\ll C^{-1} δ ≪ C − 1 ) and the presence of multiple zero intervals (which would trap a vanishingly short positive interval with opposite unit endpoint slopes), while transversality and one-sidedness produce a nondegenerate zero interval whose endpoints form smooth graphs of height at most η \eta η , fulfilling (2) for large k k k .
Because the continuous scale function s ↦ s − 1 sup A ( s , 2 s ) u j s\mapsto s^{-1}\sup_{A(s,2s)}u_j s ↦ s − 1 sup A ( s , 2 s ) u j lies in [ 0 , η ] [0,\eta] [ 0 , η ] under (1) and in [ 1 − η , ∞ ) [1-\eta,\infty) [ 1 − η , ∞ ) under (2), connectedness of [ S a j , b j / S ] [S a_j,b_j/S] [ S a j , b j / S ] forces a single regime to hold across the entire interval for each large j j j . Put ρ = 4 S a j \rho=4S a_j ρ = 4 S a j and R = b j / ( 4 S ) R=b_j/(4S) R = b j / ( 4 S ) .
In regime (1), u j u_j u j is harmonic (possibly identically zero) on A ( ρ , R ) A(\rho,R) A ( ρ , R ) . Decomposing g = ∂ m u j g=\partial_m u_j g = ∂ m u j on A ( 2 ρ , R / 2 ) A(2\rho,R/2) A ( 2 ρ , R /2 ) via the harmonic extension G G G of g ∣ ∂ B R / 2 g|_{\partial B_{R/2}} g ∣ ∂ B R /2 to B R / 2 B_{R/2} B R /2 , the maximum principle gives ∣ g − G ∣ ≤ 2 L ( 2 ρ / ∣ x ∣ ) n − 2 |g-G|\le2L(2\rho/|x|)^{n-2} ∣ g − G ∣ ≤ 2 L ( 2 ρ /∣ x ∣ ) n − 2 , and interior harmonic derivative bounds on balls of radius ≲ s \lesssim s ≲ s around ∣ x ∣ = s ∈ ( 8 ρ , R / 8 ) |x|=s\in(8\rho,R/8) ∣ x ∣ = s ∈ ( 8 ρ , R /8 ) yield
s ∣ D 2 u j ( x ) ∣ ≤ C L [ ( ρ / s ) n − 2 + s / R ] . s|D^2u_j(x)|\le CL\left[(\rho/s)^{n-2}+s/R\right]. s ∣ D 2 u j ( x ) ∣ ≤ C L [ ( ρ / s ) n − 2 + s / R ] .
In regime (2), Lemma 4.3 splits u j = v + + v − u_j=v_++v_- u j = v + + v − on A ( ρ , R ) A(\rho,R) A ( ρ , R ) , and applying the annular improvement-of-flatness theorem [15 Theorem 2.2] with decay exponent 1 / 2 1/2 1/2 to each v ± v_{\pm} v ± supplies unit-slope affine approximations on A ( s / 2 , 2 s ) ∩ { v ± > 0 } A(s/2,2s) \cap\{v_{\pm}>0\} A ( s /2 , 2 s ) ∩ { v ± > 0 } with height error
h ≤ C η s [ ( ρ / s ) n − 3 / 2 + ( s / R ) 1 / 2 ] . h \le C\eta s\left[(\rho/s)^{n-3/2}+(s/R)^{1/2}\right]. h ≤ C η s [ ( ρ / s ) n − 3/2 + ( s / R ) 1/2 ] .
At a point of { ∣ x ∣ = s } ∩ { v ± > 0 } \{|x|=s\}\cap\{v_{\pm}>0\} { ∣ x ∣ = s } ∩ { v ± > 0 } at distance at least s / 32 s/32 s /32 from FB ( v ± ) \operatorname{FB}(v_{\pm}) FB ( v ± ) , apply the interior harmonic estimate to the difference from the affine fit L L L . Otherwise take a nearest free-boundary point q q q . Continuity gives ∣ L ( q ) ∣ ≤ h |L(q)|\le h ∣ L ( q ) ∣ ≤ h , so ∣ v ± − ( L − L ( q ) ) ∣ ≤ 2 h |v_{\pm}-(L-L(q))|\le2h ∣ v ± − ( L − L ( q )) ∣ ≤ 2 h on the positive phase of B s / 8 ( q ) ⊂ A ( s / 2 , 2 s ) B_{s/8}(q)\subset A(s/2,2s) B s /8 ( q ) ⊂ A ( s /2 , 2 s ) . Translate to q q q , rotate the unit slope, and rescale by s / 8 s/8 s /8 to apply [10 Lemma 3.7] . The rescaled flatness is at most 16 h / s = O ( η ) 16h/s=O(\eta) 16 h / s = O ( η ) , within the required threshold by the choice of η \eta η . Both estimates give s ∣ D 2 v ± ( x ) ∣ ≤ C h / s s|D^2v_{\pm}(x)|\le Ch/s s ∣ D 2 v ± ( x ) ∣ ≤ C h / s . Since n ≥ 3 n\ge3 n ≥ 3 , in both regimes we have s ∣ D 2 u j ( x ) ∣ ≤ C ( L + η ) [ ( ρ / s ) 1 / 2 + ( s / R ) 1 / 2 ] s|D^2u_j(x)|\le C(L+\eta)[(\rho/s)^{1/2}+(s/R)^{1/2}] s ∣ D 2 u j ( x ) ∣ ≤ C ( L + η ) [( ρ / s ) 1/2 + ( s / R ) 1/2 ] throughout A ( 8 ρ , R / 8 ) ∩ { u j > 0 } A(8\rho,R/8)\cap\{u_j>0\} A ( 8 ρ , R /8 ) ∩ { u j > 0 } . Raising to the n n n -th power and integrating against s − 1 d s s^{-1}\,ds s − 1 d s over M a j < ∣ x ∣ < b j / M Ma_j<|x|<b_j/M M a j < ∣ x ∣ < b j / M (for M ≫ S M\gg S M ≫ S ) bounds the annular Hessian mass by C ( L + η ) n ( S / M ) n / 2 C(L+\eta)^n(S/M)^{n/2} C ( L + η ) n ( S / M ) n /2 , which vanishes as M → ∞ M\to\infty M → ∞ .
The same exterior affine-fit estimate immediately yields finiteness of the total Hessian energy for every finite-Morse-index bubble.
Corollary 4.5. Let 3 ≤ n ≤ 6 3\le n\le6 3 ≤ n ≤ 6 , and let V V V be an entire classical solution to (1.1) of finite Morse index. Then μ V ( R n ) < ∞ \mu_V(\mathbb{R}^n)<\infty μ V ( R n ) < ∞ .
Proof. If V V V is one-dimensional, μ V ( R n ) = 0 \mu_V(\mathbb{R}^n)=0 μ V ( R n ) = 0 ; if V V V is a radial exterior-ball solution, ∣ D 2 V ( x ) ∣ = O ( ∣ x ∣ − n ) |D^2V(x)|=O(|x|^{-n}) ∣ D 2 V ( x ) ∣ = O ( ∣ x ∣ − n ) as ∣ x ∣ → ∞ |x|\to\infty ∣ x ∣ → ∞ , which is integrable in L n ( R n ∖ B 1 ) L^n(\mathbb{R}^n\setminus B_1) L n ( R n ∖ B 1 ) . For an axially symmetric catenoid-type solution, [15 Proposition 3.1] splits V V V outside a fixed ball into two classical branches with disjoint supports and centered annular heights o ( s ) o(s) o ( s ) as s → ∞ s\to\infty s → ∞ . Choose ρ \rho ρ large enough that both branches meet the smallness hypothesis of [15 Theorem 2.2] on A ( ρ , R ) A(\rho,R) A ( ρ , R ) for every sufficiently large R R R . Applying the theorem with exponent 1 / 2 1/2 1/2 , using the same one-sided estimate as above and letting R → ∞ R\to\infty R → ∞ gives ∣ x ∣ ∣ D 2 V ( x ) ∣ ≤ C ( ρ / ∣ x ∣ ) n − 3 / 2 |x||D^2V(x)|\le C(\rho/|x|)^{n-3/2} ∣ x ∣∣ D 2 V ( x ) ∣ ≤ C ( ρ /∣ x ∣ ) n − 3/2 on the positive phase for ∣ x ∣ > 8 ρ |x|>8\rho ∣ x ∣ > 8 ρ . Thus ∣ D 2 V ∣ n ≤ C 1 ∣ x ∣ − n ( n − 1 / 2 ) |D^2V|^n\le C_1|x|^{-n(n-1/2)} ∣ D 2 V ∣ n ≤ C 1 ∣ x ∣ − n ( n − 1/2 ) there, which is integrable at infinity.
Voronoi scale tree and proof of Theorem 1.4
Proof of Theorem 1.4 . We organize the proof into four stages.
Stage 1: Iterative curvature selection and index accounting. By Lemma 3.1 and FB ( u j ) ∩ B 1 ≠ ∅ \operatorname{FB}(u_j)\cap B_1\ne\varnothing FB ( u j ) ∩ B 1 = ∅ , the sequence { u j } \{u_j\} { u j } is locally equi-Lipschitz in B 2 B_2 B 2 and converges locally uniformly after passing to a subsequence to a locally Lipschitz limit U U U that is harmonic on { U > 0 } \{U>0\} { U > 0 } . Suppose inductively that m ≥ 0 m\ge0 m ≥ 0 bubble sequences ( z j i , λ j i , V i ) 1 ≤ i ≤ m (z_j^i,\lambda_j^i,V^i)_{1\le i\le m} ( z j i , λ j i , V i ) 1 ≤ i ≤ m have already been extracted, and define the 1-Lipschitz distance weight
d j ( x ) = min { dist ( x , ∂ B 2 ) , min 1 ≤ i ≤ m ∣ x − z j i ∣ } . d_j(x)=\min\left\{\operatorname{dist}(x,\partial B_2),\ \min_{1\le i\le m}|x-z_j^i|\right\}. d j ( x ) = min { dist ( x , ∂ B 2 ) , 1 ≤ i ≤ m min ∣ x − z j i ∣ } .
(omitting the inner minimum when m = 0 m=0 m = 0 ). If d j ∣ D 2 u j ∣ d_j|D^2u_j| d j ∣ D 2 u j ∣ is unbounded on some compact subset of B 2 B_2 B 2 , Lemma 3.8 extracts a new center z j m + 1 ∈ FB ( u j ) z_j^{m+1}\in\operatorname{FB}(u_j) z j m + 1 ∈ FB ( u j ) converging to some p m + 1 ∈ B 2 p^{m+1}\in B_2 p m + 1 ∈ B 2 and a scale λ j m + 1 → 0 \lambda_j^{m+1}\to0 λ j m + 1 → 0 along a subsequence such that the rescalings converge to a nonflat entire classical solution V m + 1 V^{m+1} V m + 1 . By Corollary 1.2, ind ( V m + 1 ) ≥ 1 \operatorname{ind}(V^{m+1})\ge1 ind ( V m + 1 ) ≥ 1 , while Lemma 3.7 gives ind ( V m + 1 ) ≤ I \operatorname{ind}(V^{m+1})\le I ind ( V m + 1 ) ≤ I . Moreover, for each 1 ≤ i ≤ m 1\le i\le m 1 ≤ i ≤ m , we have ∣ z j m + 1 − z j i ∣ / λ j m + 1 → ∞ |z_j^{m+1}-z_j^i|/\lambda_j^{m+1}\to\infty ∣ z j m + 1 − z j i ∣/ λ j m + 1 → ∞ because λ j m + 1 = o ( d j ( y j ) ) \lambda_j^{m+1}=o(d_j(y_j)) λ j m + 1 = o ( d j ( y j )) and ∣ y j − z j m + 1 ∣ = O ( λ j m + 1 ) |y_j-z_j^{m+1}|=O(\lambda_j^{m+1}) ∣ y j − z j m + 1 ∣ = O ( λ j m + 1 ) ; conversely, if ∣ z j m + 1 − z j i ∣ / λ j i |z_j^{m+1}-z_j^i|/\lambda_j^i ∣ z j m + 1 − z j i ∣/ λ j i remained bounded, then ∣ y j − z j i ∣ ≤ 2 ∣ z j m + 1 − z j i ∣ = O ( λ j i ) |y_j-z_j^i|\le2|z_j^{m+1}-z_j^i|=O(\lambda_j^i) ∣ y j − z j i ∣ ≤ 2∣ z j m + 1 − z j i ∣ = O ( λ j i ) for large j j j , and the bounded Hessian of the i i i -th bubble rescaling would force d j ( y j ) / λ j m + 1 ≤ ∣ y j − z j i ∣ ∣ D 2 u j ( y j ) ∣ ≤ C d_j(y_j)/\lambda_j^{m+1}\le|y_j-z_j^i||D^2u_j(y_j)|\le C d j ( y j ) / λ j m + 1 ≤ ∣ y j − z j i ∣∣ D 2 u j ( y j ) ∣ ≤ C , a contradiction. Thus all extracted bubbles satisfy the separation in part (ii). Choosing for each 1 ≤ i ≤ m + 1 1 \le i \le m+1 1 ≤ i ≤ m + 1 an ind ( V i ) \operatorname{ind}(V^i) ind ( V i ) -dimensional subspace of admissible test functions supported in a fixed ball B R i ( 0 ) B_{R_i}(0) B R i ( 0 ) on which Q V i Q_{V^i} Q V i is negative definite, mutual scale separation ensures that for all sufficiently large j j j the rescaled support balls B R i λ j i ( z j i ) B_{R_i\lambda^i_j}(z^i_j) B R i λ j i ( z j i ) are pairwise disjoint subsets of B 2 B_2 B 2 . Consequently, the transferred negative-definite subspaces have mutually disjoint supports in B 2 B_2 B 2 , forcing ∑ i = 1 m + 1 ind ( V i ) ≤ I \sum_{i=1}^{m+1}\operatorname{ind}(V^i) \le I ∑ i = 1 m + 1 ind ( V i ) ≤ I . Since each bubble consumes at least one unit of index, the selection terminates at some 0 ≤ N ≤ I 0 \le N \le I 0 ≤ N ≤ I with the uniform residual curvature bound
∣ D 2 u j ( x ) ∣ ≤ C K d j ( x ) ∀ x ∈ K ∩ { u j > 0 } , d j ( x ) > 0 , (20) |D^2u_j(x)| \le\frac{C_K}{d_j(x)} \qquad\forall x \in K \cap\{u_j>0\},\ d_j(x)>0, \tag*{(20)} ∣ D 2 u j ( x ) ∣ ≤ d j ( x ) C K ∀ x ∈ K ∩ { u j > 0 } , d j ( x ) > 0 , (20)
on every compact subset K ⋐ B 2 K \Subset B_2 K ⋐ B 2 . Part (i) now follows from Corollaries 3.9 and 4.5.
Stage 2: Voronoi cell partition and scale hierarchy at a concentration point. Fix a limiting concentration point p ∈ { p 1 , … , p N } p \in\{p^1,\ldots,p^N\} p ∈ { p 1 , … , p N } , put J p = { 1 ≤ i ≤ N : p i = p } J_p=\{1\le i\le N:p^i=p\} J p = { 1 ≤ i ≤ N : p i = p } , and choose b > 0 b>0 b > 0 small enough that B 32 b ( p ) ⋐ B 2 B_{32b}(p)\Subset B_2 B 32 b ( p ) ⋐ B 2 are pairwise disjoint across distinct concentration points. Partition R n \mathbb{R}^n R n into the Voronoi nearest-center cells E j i E^i_j E j i associated with { z j i : i ∈ J p } \{z^i_j:i\in J_p\} { z j i : i ∈ J p } (breaking ties on bisecting hyperplanes arbitrarily), so that
x ∈ E j i ⟹ ∣ x − z j k ∣ ≥ ∣ x − z j i ∣ ∀ k ∈ J p . x \in E^i_j \quad\Longrightarrow\quad|x-z^k_j| \ge|x-z^i_j| \qquad\forall k \in J_p. x ∈ E j i ⟹ ∣ x − z j k ∣ ≥ ∣ x − z j i ∣ ∀ k ∈ J p .
For each fixed i ∈ J p i\in J_p i ∈ J p , consider the finite list of scales { λ j i } ∪ { ∣ z j i − z j k ∣ : k ∈ J p ∖ { i } } ∪ { b } \{\lambda^i_j\}\cup\{|z^i_j-z^k_j|:k\in J_p\setminus\{i\}\}\cup\{b\} { λ j i } ∪ { ∣ z j i − z j k ∣ : k ∈ J p ∖ { i }} ∪ { b } . Passing to a subsequence so that all pairwise scale ratios converge in [ 0 , ∞ ] [0,\infty] [ 0 , ∞ ] , merge scales with finite nonzero limiting ratios and choose one representative from each equivalence class to form a strictly ordered hierarchy
d j , 0 ≪ d j , 1 ≪ ⋯ ≪ d j , m , d j , 0 = λ j i , d j , m = b . d_{j,0}\ll d_{j,1}\ll\cdots\ll d_{j,m},\qquad d_{j,0}=\lambda^i_j,\qquad d_{j,m}=b. d j , 0 ≪ d j , 1 ≪ ⋯ ≪ d j , m , d j , 0 = λ j i , d j , m = b .
where a j ≪ b j a_j\ll b_j a j ≪ b j signifies a j / b j → 0 a_j/b_j\to0 a j / b j → 0 (notice d j , 0 ≪ d j , 1 d_{j,0}\ll d_{j,1} d j , 0 ≪ d j , 1 by part (ii)). Take one further subsequence, common to all finitely many groups and scales, on which every bounded vector sequence ( z j k − z j i ) / d j , h (z^k_j-z^i_j)/d_{j,h} ( z j k − z j i ) / d j , h with i , k ∈ J p i,k\in J_p i , k ∈ J p and 0 < h < m 0<h<m 0 < h < m converges. The norms of the other such vector sequences tend to infinity by the choice of the scale ratios.
Stage 3: Vanishing of gap and transition-band energies. First consider any scale gap d j , h ≪ s j ≪ d j , h + 1 d_{j,h}\ll s_j\ll d_{j,h+1} d j , h ≪ s j ≪ d j , h + 1 (0 ≤ h < m 0\le h<m 0 ≤ h < m ). Rescaling u j u_j u j around z j i z^i_j z j i by s j s_j s j , every other center z j k z^k_j z j k either collapses onto the origin (if ∣ z j k − z j i ∣ ≲ d j , h |z^k_j-z^i_j|\lesssim d_{j,h} ∣ z j k − z j i ∣ ≲ d j , h ) or escapes to infinity (if ∣ z j k − z j i ∣ ≳ d j , h + 1 |z^k_j-z^i_j|\gtrsim d_{j,h+1} ∣ z j k − z j i ∣ ≳ d j , h + 1 ). The residual bound (20) therefore furnishes uniform Hessian bounds on compact annuli in R n ∖ { 0 } \mathbb{R}^n\setminus\{0\} R n ∖ { 0 } , and Proposition 4.2 restricts every subsequential limit to 0 0 0 or a centered unit wedge. Applying Lemma 4.4 shows that the full gap annulus A ( M d j , h , d j , h + 1 / M ) + z j i A(Md_{j,h},d_{j,h+1}/M)+z^i_j A ( M d j , h , d j , h + 1 / M ) + z j i carries asymptotically vanishing μ u j \mu_{u_j} μ u j -mass as j → ∞ j\to\infty j → ∞ followed by M → ∞ M\to\infty M → ∞ .
Next consider an intermediate representative scale d j , h d_{j,h} d j , h with 0 < h < m 0<h<m 0 < h < m , and restrict the transition band A ( d j , h / M , M d j , h ) + z j i A(d_{j,h}/M,Md_{j,h})+z^i_j A ( d j , h / M , M d j , h ) + z j i to the Voronoi cell E j i E^i_j E j i . After rescaling around z j i z^i_j z j i by d j , h d_{j,h} d j , h , every point y y y in the rescaled restricted band satisfies ∣ y ∣ ≥ 1 / M |y|\ge1/M ∣ y ∣ ≥ 1/ M , and by the Voronoi property its distance to every rescaled companion center ( z j k − z j i ) / d j , h (z^k_j-z^i_j)/d_{j,h} ( z j k − z j i ) / d j , h (k ∈ J p k\in J_p k ∈ J p ) is also at least ∣ y ∣ ≥ 1 / M |y|\ge1/M ∣ y ∣ ≥ 1/ M , while centers outside J p J_p J p escape to infinity. Thus, for each fixed M ≥ 2 M\ge2 M ≥ 2 and all large j j j , the rescaled restricted band lies in { 1 / M ≤ ∣ y ∣ ≤ M } \{1/M\le|y|\le M\} { 1/ M ≤ ∣ y ∣ ≤ M } at distance at least 1 / ( 2 M ) 1/(2M) 1/ ( 2 M ) from each of the finitely many limiting centers selected above. The resulting fixed compact set has a neighborhood where (20) bounds the rescaled Hessians uniformly. By Proposition 4.2 , every locally uniform limit is 0 0 0 or a unit wedge, both with vanishing positive-phase Hessian. Lemma 3.3 with p = n p=n p = n therefore gives, at each fixed M M M ,
lim j → ∞ μ u j ( E j i ∩ ( A ( d j , h / M , M d j , h ) + z j i ) ) = 0. \lim_{j\to\infty}\mu_{u_j}\left(E^i_j\cap\left(A(d_{j,h}/M,Md_{j,h})+z^i_j\right)\right)=0. j → ∞ lim μ u j ( E j i ∩ ( A ( d j , h / M , M d j , h ) + z j i ) ) = 0.
Meanwhile, on the core ball B M λ j i ( z j i ) B_{M\lambda_j^i}(z_j^i) B M λ j i ( z j i ) —which lies entirely inside E j i E_j^i E j i for large j j j since λ j i ≪ ∣ z j i − z j ∣ \lambda_j^i \ll|z_j^i-z_j| λ j i ≪ ∣ z j i − z j ∣ —Lemma 3.3 gives lim j → ∞ μ u j ( B M λ j i ( z j i ) ) = μ V i ( B M ) \lim_{j\to\infty}\mu_{u_j}(B_{M\lambda_j^i}(z_j^i))=\mu_{V^i}(B_M) lim j → ∞ μ u j ( B M λ j i ( z j i )) = μ V i ( B M ) .
Stage 4: Measure synthesis. For fixed large M M M and all sufficiently large j j j , decomposing each cell E j i ∩ B b / M ( z j i ) E_j^i\cap B_{b/M}(z_j^i) E j i ∩ B b / M ( z j i ) radially into its core ball, restricted transition bands, and gap annuli covers the union
U j , M = ⋃ i ∈ J p ( E j i ∩ B b / M ( z j i ) ) , \mathcal{U}_{j,M}=\bigcup_{i\in J_p}(E_j^i\cap B_{b/M}(z_j^i)), U j , M = i ∈ J p ⋃ ( E j i ∩ B b / M ( z j i )) ,
yielding two-sided mass bounds
∑ i ∈ J p μ u j ( B M λ j i ( z j i ) ) ≤ μ u j ( U j , M ) ≤ ∑ i ∈ J p μ u j ( B M λ j i ( z j i ) ) + R j , M , lim M → ∞ lim sup j → ∞ R j , M = 0. \sum_{i\in J_p}\mu_{u_j}(B_{M\lambda_j^i}(z_j^i))\leq\mu_{u_j}(\mathcal{U}_{j,M})\leq\sum_{i\in J_p}\mu_{u_j}(B_{M\lambda_j^i}(z_j^i))+\mathcal{R}_{j,M},\qquad\lim_{M\to\infty}\limsup_{j\to\infty}\mathcal{R}_{j,M}=0. i ∈ J p ∑ μ u j ( B M λ j i ( z j i )) ≤ μ u j ( U j , M ) ≤ i ∈ J p ∑ μ u j ( B M λ j i ( z j i )) + R j , M , M → ∞ lim j → ∞ lim sup R j , M = 0.
Since z j i → p z_j^i\to p z j i → p for all i ∈ J p i\in J_p i ∈ J p , we have the geometric sandwich B b / ( 2 M ) ( p ) ⊂ U j , M ⊂ B 2 b / M ( p ) B_{b/(2M)}(p)\subset\mathcal{U}_{j,M}\subset B_{2b/M}(p) B b / ( 2 M ) ( p ) ⊂ U j , M ⊂ B 2 b / M ( p ) for all large j j j . Evaluating the upper bound at a single fixed M M M gives uniform local mass bounds for { μ u j } \{\mu_{u_j}\} { μ u j } near each p p p , and away from the finite set { p 1 , … , p N } \{p^1,\ldots,p^N\} { p 1 , … , p N } , (20) and Lemma 3.3 give strong L l o c 1 L^1_{\mathrm{loc}} L loc 1 convergence of ∣ D 2 u j ∣ n 1 { u j > 0 } |D^2u_j|^{n}1_{\{u_j>0\}} ∣ D 2 u j ∣ n 1 { u j > 0 } to ∣ D 2 U ∣ n 1 { U > 0 } |D^2U|^{n}1_{\{U>0\}} ∣ D 2 U ∣ n 1 { U > 0 } (whose integrability across { p 1 , … , p N } \{p^1,\ldots,p^N\} { p 1 , … , p N } follows from Fatou’s lemma and the uniform local mass bound). Finally, sending j → ∞ j\to\infty j → ∞ and then M → ∞ M\to\infty M → ∞ along continuity radii b / ( 2 M ) b/(2M) b / ( 2 M ) and 2 b / M 2b/M 2 b / M of a weak-* Radon limit identifies the atomic weight at p p p as ∑ i ∈ J p lim M → ∞ μ V i ( B M ) = ∑ i ∈ J p μ V i ( R n ) \sum_{i\in J_p}\lim_{M\to\infty}\mu_{V^i}(B_M)=\sum_{i\in J_p}\mu_{V^i}(\mathbb{R}^n) ∑ i ∈ J p lim M → ∞ μ V i ( B M ) = ∑ i ∈ J p μ V i ( R n ) , establishing (2) .□
Corollary 4.6. Under the hypotheses of Theorem 1.4 , for every 0 < r < 2 0<r<2 0 < r < 2 there exists C = C ( n , I , r ) > 0 C=C(n,I,r)>0 C = C ( n , I , r ) > 0 such that μ u j ( B r ) ≤ C \mu_{u_j}(B_r)\leq C μ u j ( B r ) ≤ C .
Proof. Were μ u j ( B r ) → ∞ \mu_{u_j}(B_r)\to\infty μ u j ( B r ) → ∞ along a sequence satisfying the hypotheses, testing (2) along a convergent subsequence against a nonnegative function χ ∈ C c ( B 2 ) \chi\in C_c(B_2) χ ∈ C c ( B 2 ) with χ ≡ 1 \chi\equiv1 χ ≡ 1 on B r ‾ \overline{B_r} B r would yield a finite limit, a contradiction.□
Corollary 4.7. Under the hypotheses of Theorem 1.4 with I ≥ 1 I\geq1 I ≥ 1 , suppose there exist I I I sequences of positive-phase points x j ℓ → p ℓ x_j^\ell\to p^\ell x j ℓ → p ℓ (ℓ = 1 , … , I \ell=1,\ldots,I ℓ = 1 , … , I ) with pairwise distinct limits p 1 , … , p I ∈ B 2 p^1,\ldots,p^I\in B_2 p 1 , … , p I ∈ B 2 along which ∣ D 2 u j ( x j ℓ ) ∣ → ∞ |D^2u_j(x_j^\ell)|\to\infty ∣ D 2 u j ( x j ℓ ) ∣ → ∞ . Then on the subsequence in Theorem 1.4 precisely one bubble forms at each p ℓ p^\ell p ℓ , each bubble has Morse index one, no other atoms occur, and for I = 1 I=1 I = 1 the unique atom carries the full Hessian energy of its bubble.
Proof. By (20) , curvature blow-up at p ℓ p^\ell p ℓ requires at least one bubble center converging to p ℓ p^\ell p ℓ . Since p 1 , … , p I p^1,\ldots,p^I p 1 , … , p I are distinct and each nonflat bubble satisfies ind ( V i ) ≥ 1 \operatorname{ind}(V^i)\geq1 ind ( V i ) ≥ 1 , the index bound ∑ i = 1 N ind ( V i ) ≤ I \sum_{i=1}^N\operatorname{ind}(V^i)\leq I ∑ i = 1 N ind ( V i ) ≤ I forces N = I N=I N = I and ind ( V i ) = 1 \operatorname{ind}(V^i)=1 ind ( V i ) = 1 for each i i i .□
Model bubble energies and seven-dimensional counterexamples
The radial exterior-ball solution has Morse index one by separation of variables [4 Remark 1.3] . We include the calculation, compute its exact Hessian energy, and illustrate two rescaling regimes. We also show that the uniform bounded-index Hessian energy bound fails in dimension seven.
Proposition 5.1. For every n ≥ 3 n\geq3 n ≥ 3 , the unit exterior-ball solution V ( x ) = ( n − 2 ) − 1 ( 1 − ∣ x ∣ 2 − n ) V(x)=(n-2)^{-1}(1-|x|^{2-n}) V ( x ) = ( n − 2 ) − 1 ( 1 − ∣ x ∣ 2 − n ) on { ∣ x ∣ > 1 } \{|x|>1\} { ∣ x ∣ > 1 } , extended by zero to B 1 B_1 B 1 , satisfies
ind ( V ) = 1 , E n b a l l : = μ V ( R n ) = ∣ S n − 1 ∣ [ n ( n − 1 ) ] n / 2 − 1 , \operatorname{ind}(V)=1,\qquad E_n^{\mathrm{ball}}:=\mu_V(\mathbb{R}^n)=|\mathbb{S}^{n-1}|[n(n-1)]^{n/2-1}, ind ( V ) = 1 , E n ball := μ V ( R n ) = ∣ S n − 1 ∣ [ n ( n − 1 ) ] n /2 − 1 , and both quantities are invariant under rigid motions and Bernoulli rescaling.
Proof. At r = ∣ x ∣ > 1 r=|x|>1 r = ∣ x ∣ > 1 , D 2 V ( x ) D^2V(x) D 2 V ( x ) has radial eigenvalue − ( n − 1 ) r − n -(n-1)r^{-n} − ( n − 1 ) r − n and n − 1 n-1 n − 1 tangential eigenvalues r − n r^{-n} r − n , so ∣ D 2 V ( x ) ∣ 2 = n ( n − 1 ) r − 2 n \lvert D^2V(x)\rvert^2=n(n-1)r^{-2n} ∣ D 2 V ( x ) ∣ 2 = n ( n − 1 ) r − 2 n . Integrating in polar coordinates gives
μ V ( R n ) = ∣ S n − 1 ∣ [ n ( n − 1 ) ] n / 2 ∫ 1 ∞ r n − 1 − n 2 d r = ∣ S n − 1 ∣ [ n ( n − 1 ) ] n / 2 − 1 , \mu_V(\mathbb{R}^n)=\lvert S^{n-1}\rvert[n(n-1)]^{n/2}\int_1^\infty r^{n-1- n^2}\,dr=\lvert S^{n-1}\rvert[n(n-1)]^{n/2-1}, μ V ( R n ) = ∣ S n − 1 ∣ [ n ( n − 1 ) ] n /2 ∫ 1 ∞ r n − 1 − n 2 d r = ∣ S n − 1 ∣ [ n ( n − 1 ) ] n /2 − 1 ,
since ∫ 1 ∞ r n − 1 − n 2 d r = ( n 2 − n ) − 1 \int_1^\infty r^{n-1-n^2}\,dr=(n^2-n)^{-1} ∫ 1 ∞ r n − 1 − n 2 d r = ( n 2 − n ) − 1 .
To compute ind ( V ) \operatorname{ind}(V) ind ( V ) , recall H = n − 1 H=n-1 H = n − 1 on ∂ B 1 \partial B_1 ∂ B 1 and expand the boundary trace ϕ ∣ ∂ B 1 \phi|_{\partial B_1} ϕ ∣ ∂ B 1 of an admissible test function ϕ ∈ C c ∞ ( R n ) \phi\in C^\infty_c(\mathbb{R}^n) ϕ ∈ C c ∞ ( R n ) in an L 2 ( S n − 1 ) L^2(S^{n-1}) L 2 ( S n − 1 ) -orthonormal basis of spherical harmonics { Y ℓ m } \{Y_{\ell m}\} { Y ℓ m } of degree ℓ ≥ 0 \ell\ge0 ℓ ≥ 0 with coefficients c ℓ m c_{\ell m} c ℓ m . By Dirichlet’s principle on R n ∖ B 1 ‾ \mathbb{R}^n\setminus\overline{B_1} R n ∖ B 1 , ∫ ∣ x ∣ > 1 ∣ ∇ ϕ ∣ 2 d x \int_{\lvert x\rvert>1}\lvert\nabla\phi\rvert^2\,dx ∫ ∣ x ∣ > 1 ∣ ∇ ϕ ∣ 2 d x is bounded below by the Dirichlet energy of the decaying harmonic extension ∑ ℓ , m c ℓ m r − ( ℓ + n − 2 ) Y ℓ m \sum_{\ell,m}c_{\ell m}r^{-(\ell+n-2)}Y_{\ell m} ∑ ℓ , m c ℓ m r − ( ℓ + n − 2 ) Y ℓ m , giving
Q V [ ϕ ; R n ] ≥ ∑ ℓ , m ( ℓ + n − 2 ) ∣ c ℓ m ∣ 2 − ( n − 1 ) ∑ ℓ , m ∣ c ℓ m ∣ 2 = ∑ ℓ , m ( ℓ − 1 ) ∣ c ℓ m ∣ 2 . Q_V[\phi;\mathbb{R}^n]\ge\sum_{\ell,m}(\ell+n-2)\lvert c_{\ell m}\rvert^2-(n-1)\sum_{\ell,m}\lvert c_{\ell m}\rvert^2=\sum_{\ell,m}(\ell-1)\lvert c_{\ell m}\rvert^2. Q V [ ϕ ; R n ] ≥ ℓ , m ∑ ( ℓ + n − 2 ) ∣ c ℓ m ∣ 2 − ( n − 1 ) ℓ , m ∑ ∣ c ℓ m ∣ 2 = ℓ , m ∑ ( ℓ − 1 ) ∣ c ℓ m ∣ 2 .
Hence Q V Q_V Q V is nonnegative on the codimension-one subspace of traces with zero spherical average (c 00 = 0 c_{00}=0 c 00 = 0 ), giving ind ( V ) ≤ 1 \operatorname{ind}(V)\le1 ind ( V ) ≤ 1 . Conversely, take χ ∈ C ∞ ( [ 0 , ∞ ) ) \chi\in C^\infty([0,\infty)) χ ∈ C ∞ ([ 0 , ∞ )) , equal to one on [ 0 , 1 ] [0,1] [ 0 , 1 ] and zero on [ 2 , ∞ ) [2,\infty) [ 2 , ∞ ) . The test ϕ R ( x ) = χ ( ∣ x ∣ / R ) ∣ x ∣ 2 − n \phi_R(x)=\chi(\lvert x\rvert/R)\lvert x\rvert^{2-n} ϕ R ( x ) = χ (∣ x ∣ / R ) ∣ x ∣ 2 − n on { ∣ x ∣ ≥ 1 } \{\lvert x\rvert\ge1\} {∣ x ∣ ≥ 1 } has a smooth compactly supported extension to R n \mathbb{R}^n R n , obtained by changing the formula only near the origin. The uncut radial mode has form − ∣ S n − 1 ∣ -\lvert S^{n-1}\rvert − ∣ S n − 1 ∣ , and the tail and cutoff change its Dirichlet energy by O ( R 2 − n ) O(R^{2-n}) O ( R 2 − n ) . Thus Q V [ ϕ R ; R n ] < 0 Q_V[\phi_R;\mathbb{R}^n]<0 Q V [ ϕ R ; R n ] < 0 for large R R R , giving ind ( V ) = 1 \operatorname{ind}(V)=1 ind ( V ) = 1 . Rigid motions preserve both quantities. For the scaled profile λ V ( ( x − z ) / λ ) \lambda V((x-z)/\lambda) λV (( x − z ) / λ ) , the pullback of tests multiplies the quadratic form by λ n − 2 \lambda^{n-2} λ n − 2 , and the Hessian energy is invariant.□
Remark 5.2. In Theorem 1.4 , if m p m_p m p exterior-ball bubbles concentrate at p p p , they contribute atomic mass m p E n b a l l m_pE_n^{\mathrm{ball}} m p E n ball and consume m p m_p m p units of Morse index. Even for a single shrinking exterior ball u λ ( x ) = λ V ( x / λ ) u_\lambda(x)=\lambda V(x/\lambda) u λ ( x ) = λV ( x / λ ) (λ ↓ 0 \lambda\downarrow0 λ ↓ 0 ), one sees both hallmarks of Proposition 4.2 : u λ → 0 u_\lambda\to0 u λ → 0 uniformly, 1 { u λ > 0 } → 1 \mathbf{1}_{\{u_\lambda>0\}}\to1 1 { u λ > 0 } → 1 in L l o c 1 ( R n ) L^1_{\mathrm{loc}}(\mathbb{R}^n) L loc 1 ( R n ) , and μ u λ ⇀ E n b a l l δ 0 \mu_{u_\lambda}\rightharpoonup E_n^{\mathrm{ball}}\delta_0 μ u λ ⇀ E n ball δ 0 with ind ( u λ ; B 2 ) = 1 \operatorname{ind}(u_\lambda;B_2)=1 ind ( u λ ; B 2 ) = 1 for small λ \lambda λ . By contrast, the collapsing stable two-plane family u λ ( x ) = ( x n − λ ) + + ( − x n − λ ) + u_\lambda(x)=(x_n-\lambda)_++(-x_n-\lambda)_+ u λ ( x ) = ( x n − λ ) + + ( − x n − λ ) + converges uniformly to the centered unit wedge ∣ x n ∣ \lvert x_n\rvert ∣ x n ∣ with 1 { u λ > 0 } → 1 \mathbf{1}_{\{u_\lambda>0\}}\to1 1 { u λ > 0 } → 1 in L l o c 1 ( R n ) L^1_{\mathrm{loc}}(\mathbb{R}^n) L loc 1 ( R n ) , while carrying identically zero Hessian energy μ u λ ≡ 0 \mu_{u_\lambda}\equiv0 μ u λ ≡ 0 at every λ > 0 \lambda>0 λ > 0 .
In dimension seven, a smooth minimizing leaf with nonflat conical blow-down gives a family of Morse index zero whose interior Hessian energies grow at least logarithmically. Hence the uniform local mass bound underlying Theorem 1.4 fails in dimension seven.
Proposition 5.3. In R 7 \mathbb{R}^7 R 7 , there exists a one-parameter family of entire classical minimizers { u λ } 0 < λ < 1 / 2 \{u_\lambda\}_{0<\lambda<1/2} { u λ } 0 < λ < 1/2 satisfying 0 ∈ F B ( u λ ) 0\in\mathrm{FB}(u_\lambda) 0 ∈ FB ( u λ ) and ind ( u λ ; B 2 ) = 0 \operatorname{ind}(u_\lambda;B_2)=0 ind ( u λ ; B 2 ) = 0 such that
μ u λ ( B 1 ) ≥ c log ( 1 / λ ) − C ∀ λ ∈ ( 0 , 1 / 2 ) , \mu_{u_\lambda}(B_1)\ge c\log(1/\lambda)-C \qquad\forall\lambda\in(0,1/2), μ u λ ( B 1 ) ≥ c log ( 1/ λ ) − C ∀ λ ∈ ( 0 , 1/2 ) , for positive constants c , C > 0 c,C>0 c , C > 0 independent of $\lambda.
Proof. Let C 0 C_0 C 0 be the seven-dimensional nonflat minimizing cone of De Silva and Jerison [12 ] , and let V V V be a smooth entire minimizing leaf with blow-down C 0 C_0 C 0 provided by [13 Theorem 1.5] . Because C 0 C_0 C 0 is 1 1 1 -homogeneous and nonflat, its positive-phase Hessian does not vanish identically on A ( 1 , 2 ) ∩ { C 0 > 0 } A(1,2)\cap\{C_0>0\} A ( 1 , 2 ) ∩ { C 0 > 0 } , so there exists a compact set K ⊂ A ( 1 , 2 ) ∩ { C 0 > 0 } K\subset A(1,2)\cap\{C_0>0\} K ⊂ A ( 1 , 2 ) ∩ { C 0 > 0 } on which ∫ K ∣ D 2 C 0 ∣ 7 d x > 0 \int_K\lvert D^2C_0\rvert^7\,dx>0 ∫ K ∣ D 2 C 0 ∣ 7 d x > 0 . Since R − 1 V ( R x ) → C 0 ( x ) R^{-1}V(Rx)\to C_0(x) R − 1 V ( R x ) → C 0 ( x ) locally uniformly as R → ∞ R\to\infty R → ∞ , interior harmonic estimates yield smooth convergence on K K K , whence there exist R 0 ≥ 1 R_0\ge1 R 0 ≥ 1 and c 0 > 0 c_0>0 c 0 > 0 such that
μ V ( B 2 R ∖ B R ) ≥ ∫ K ∣ D 2 ( R − 1 V ( R x ) ) ∣ 7 d x ≥ c 0 > 0 ∀ R ≥ R 0 . \mu_V(B_{2R}\setminus B_R)\ge\int_K\lvert D^2(R^{-1}V(Rx))\rvert^7\,dx\ge c_0>0\qquad\forall R\ge R_0. μ V ( B 2 R ∖ B R ) ≥ ∫ K ∣ D 2 ( R − 1 V ( R x )) ∣ 7 d x ≥ c 0 > 0 ∀ R ≥ R 0 .
Fix any point z 0 ∈ F B ( V ) z_0 \in\mathrm{FB}(V) z 0 ∈ FB ( V ) and define the rescaled entire minimizers u λ ( x ) = λ V ( z 0 + x / λ ) u_\lambda(x) = \lambda V(z_0 + x/\lambda) u λ ( x ) = λV ( z 0 + x / λ ) . By scale invariance, μ u λ ( B 1 ) = μ V ( B 1 / λ ( z 0 ) ) \mu_{u_\lambda}(B_1) = \mu_V(B_{1/\lambda}(z_0)) μ u λ ( B 1 ) = μ V ( B 1/ λ ( z 0 )) . For small λ > 0 \lambda> 0 λ > 0 , the ball B 1 / λ ( z 0 ) B_{1/\lambda}(z_0) B 1/ λ ( z 0 ) contains ≍ log ( 1 / λ ) \asymp\log(1/\lambda) ≍ log ( 1/ λ ) pairwise disjoint dyadic annuli B 2 k + 1 R 0 ( 0 ) ∖ B 2 k R 0 ( 0 ) B_{2^{k+1}R_0}(0) \setminus B_{2^kR_0}(0) B 2 k + 1 R 0 ( 0 ) ∖ B 2 k R 0 ( 0 ) , each contributing at least c 0 c_0 c 0 to μ V \mu_V μ V . This proves the lower bound for small λ \lambda λ , and increasing C C C covers the rest of ( 0 , 1 / 2 ) (0, 1/2) ( 0 , 1/2 ) . The functions u λ u_\lambda u λ are minimizers by scaling and translation, so their Morse index on B 2 B_2 B 2 is zero. □ \square □
Remark 5.4. In dimension seven, Proposition 5.3 exhibits conical annuli that retain positive Hessian energy on each dyadic scale even at Morse index zero.
Declarations
Availability of data and materials. No datasets were generated or analysed during the current study.
Conflict of interest. The authors declare no conflict of interest.
Use of artificial intelligence. AI tools were used to assist in the exploration and verification of mathematical proofs, as well as the translation and linguistic polishing of the manuscript. The authors take full responsibility for the mathematical content and the final manuscript.
References