Introduction and statement
Rauch’s hot spots conjecture [1 ] asserts that an eigenfunction of the first positive Neumann eigenvalue of a planar domain attains its maximum and minimum only on the boundary; see [2 ] for precise formulations. It fails for some domains with holes [3 ] , and in high dimensions even convex domains can have interior hot spots [4 ] . In the plane it is known, among other cases, for domains with two axes of symmetry [5 ] , for lip domains [6 ] , for triangles [7 , 8 , 9 , 10 ] , for isosceles trapezoids and kites [11 ] , for L-shaped domains and several classes of domains tiled by them [12 ] , and for the union of a regular polygon with its reflection across an edge [13 ] ; the last two families include domains with reentrant corners.
A recent preprint [14 ] asserts the strict form, absence of interior critical points, for every smooth simply connected planar domain. Assuming that assertion, Burdzy [15 ] and Bañuelos [16 ] derive, by approximation, the non-strict conclusion on nonsmooth simply connected domains whose first positive eigenvalue is simple, and Bañuelos asks whether it holds on every simply connected polygon, the eigenvalue being possibly multiple [16 Question 1] . Independently of [14 ] , Deng, Jiang and Yang [17 ] prove that on every bounded convex planar domain the maximum of ∣ u ∣ |u| ∣ u ∣ is attained only on the boundary. This note treats domains with corners. It does not assume the assertion of [14 ] , and for convex polygons it answers the question of [16 ] in the strict form.
All functions and Hilbert spaces in this note are real unless complex coordinates are explicitly used for conformal maps. We use the standard Sobolev spaces W k , p W^{k,p} W k , p , write H k = W k , 2 H^k = W^{k,2} H k = W k , 2 , and let H 0 1 H^1_0 H 0 1 denote the closure of compactly supported smooth functions in H 1 H^1 H 1 . The boundary space H 1 / 2 H^{1/2} H 1/2 is the Sobolev trace space; Tr \operatorname{Tr} Tr denotes the trace operator. The notation C k , γ C^{k,\gamma} C k , γ means that the derivatives of order k k k are Hölder continuous with exponent γ \gamma γ . These conventions apply componentwise to vector fields. The symbol B ( x , r ) B(x,r) B ( x , r ) denotes an open Euclidean disk, δ y \delta_y δ y the unit point mass at y y y . The constants C C C in estimates may change from line to line; f ≍ g f \asymp g f ≍ g means two-sided comparison by positive constants in the indicated neighborhood.
A bounded Jordan domain means the bounded component enclosed by a simple closed curve. We use the following precise meaning of a piecewise smooth domain.
Definition 1.1. An admissible domain is a bounded planar Jordan domain Ω \Omega Ω with a finite set of corners { v 1 , … , v m } \{v_{1},\ldots,v_{m}\} { v 1 , … , v m } , where m ≥ 0 m \ge0 m ≥ 0 . Each component of the boundary with the corners removed is a regular C ∞ C^{\infty} C ∞ arc, smooth with nonzero one-sided tangent up to its endpoints. At v j v_{j} v j the interior angle is α j π \alpha_{j}\pi α j π , where 0 < α j < 1 0 < \alpha_{j} < 1 0 < α j < 1 . The case of a C ∞ C^{\infty} C ∞ Jordan boundary with no corners is also allowed. Global convexity is not required. The open smooth part of the boundary is denoted by Γ \Gamma Γ .
In particular Ω \Omega Ω is simply connected and Lipschitz; ∣ Ω ∣ |\Omega| ∣Ω∣ denotes its area. Its outward unit normal on Γ \Gamma Γ is ν \nu ν , boundary arclength is d σ d\sigma d σ , and τ = ( − ν 2 , ν 1 ) \tau=(-\nu_{2},\nu_{1}) τ = ( − ν 2 , ν 1 ) gives the counterclockwise orientation. A boundary vector is tangent when its scalar product with ν \nu ν is zero; for Sobolev fields this condition is imposed on the trace almost everywhere on Γ \Gamma Γ . Signed curvature is fixed by ∂ τ τ = − κ ν \partial_{\tau}\tau=-\kappa\nu ∂ τ τ = − κ ν ; it is positive on a circle bounding a disk. It is bounded on the finite arcs, but may have either sign. Put
μ = inf { ∫ Ω ∣ ∇ v ∣ 2 ∫ Ω v 2 : 0 ≠ v ∈ H 1 ( Ω ) , ∫ Ω v = 0 } . \mu=\inf\left\{\frac{\int_{\Omega}|\nabla v|^{2}}{\int_{\Omega}v^{2}}:0\ne v\in H^{1}(\Omega),\ \int_{\Omega}v=0\right\}. μ = inf { ∫ Ω v 2 ∫ Ω ∣∇ v ∣ 2 : 0 = v ∈ H 1 ( Ω ) , ∫ Ω v = 0 } .
and define its eigenspace by the weak formulation
V = { v ∈ H 1 ( Ω ) : ∫ Ω ∇ v ⋅ ∇ ψ = μ ∫ Ω v ψ for every ψ ∈ H 1 ( Ω ) } . V=\left\{v\in H^{1}(\Omega):\int_{\Omega}\nabla v\cdot\nabla\psi=\mu\int_{\Omega}v\psi\ \text{for every }\psi\in H^{1}(\Omega)\right\}. V = { v ∈ H 1 ( Ω ) : ∫ Ω ∇ v ⋅ ∇ ψ = μ ∫ Ω v ψ for every ψ ∈ H 1 ( Ω ) } .
Write λ D \lambda_{D} λ D for the first Dirichlet eigenvalue, defined by the same Rayleigh quotient over 0 ≠ v ∈ H 0 1 ( Ω ) 0\ne v\in H_{0}^{1}(\Omega) 0 = v ∈ H 0 1 ( Ω ) , without the mean-zero constraint. Poincaré’s inequalities and compact Sobolev embedding on this connected bounded Lipschitz domain give 0 < μ , λ D < ∞ 0<\mu,\lambda_{D}<\infty 0 < μ , λ D < ∞ and a finite-dimensional eigenspace V V V . The operator − Δ D -\Delta_{D} − Δ D below is the Dirichlet realization of the Laplacian on L 2 ( Ω ) L^{2}(\Omega) L 2 ( Ω ) . Unsubscripted L 2 L^{2} L 2 norms in physical-domain arguments are taken over Ω \Omega Ω . Euclidean vector products are denoted by a dot, and ∇ X : ∇ Y = ∑ i , j ( ∂ i X j ) ( ∂ i Y j ) \nabla X:\nabla Y=\sum_{i,j}(\partial_{i}X_{j})(\partial_{i}Y_{j}) ∇ X : ∇ Y = ∑ i , j ( ∂ i X j ) ( ∂ i Y j ) denotes the scalar product of the two Jacobian matrices.
Disk kernels
Let D = { z ∈ C : ∣ z ∣ < 1 } \mathbb{D}=\{z\in\mathbb{C}:|z|<1\} D = { z ∈ C : ∣ z ∣ < 1 } be the unit disk, identified with a subset of R 2 \mathbb{R}^{2} R 2 , and let S 1 = ∂ D S^{1}=\partial\mathbb{D} S 1 = ∂ D with arclength d s ds d s . Its outward normal is ν D ( s ) = s \nu_{\mathbb{D}}(s)=s ν D ( s ) = s . Set
G ( x , y ) = 1 4 π log ( 1 + ( 1 − ∣ x ∣ 2 ) ( 1 − ∣ y ∣ 2 ) ∣ x − y ∣ 2 ) , P s ( x ) = 1 − ∣ x ∣ 2 2 π ∣ s − x ∣ 2 , N 0 ( s , t ) = 1 π ∣ s − t ∣ 2 G(x,y)=\frac{1}{4\pi}\log\left(1+\frac{(1-|x|^{2})(1-|y|^{2})}{|x-y|^{2}}\right),\qquad P_{s}(x)=\frac{1-|x|^{2}}{2\pi|s-x|^{2}},\qquad N_{0}(s,t)=\frac{1}{\pi|s-t|^{2}} G ( x , y ) = 4 π 1 log ( 1 + ∣ x − y ∣ 2 ( 1 − ∣ x ∣ 2 ) ( 1 − ∣ y ∣ 2 ) ) , P s ( x ) = 2 π ∣ s − x ∣ 2 1 − ∣ x ∣ 2 , N 0 ( s , t ) = π ∣ s − t ∣ 2 1
for x , y ∈ D x,y\in\mathbb{D} x , y ∈ D , x ≠ y x\ne y x = y , and s , t ∈ S 1 s,t\in S^{1} s , t ∈ S 1 , s ≠ t s\ne t s = t . These are respectively the Dirichlet Green function, the Poisson kernel and the kernel of the Dirichlet integral. In nonnegative kernel integrals use G ( x , x ) = + ∞ G(x,x)=+\infty G ( x , x ) = + ∞ ; its diagonal is a null set for the absolutely continuous measures considered here. Diagonal values of boundary kernels in double arclength integrals are immaterial and may be set to zero. For a measure d ρ = a ρ d x d\rho=a_{\rho}\,dx d ρ = a ρ d x on D \mathbb{D} D with a ρ ≥ 0 a_{\rho}\ge0 a ρ ≥ 0 measurable, and a nonnegative measurable f f f , write ( T ρ f ) ( x ) = ∫ D G ( x , z ) f ( z ) d ρ ( z ) ∈ [ 0 , ∞ ] (T_{\rho}f)(x)=\int_{\mathbb{D}}G(x,z)f(z)\,d\rho(z)\in[0,\infty] ( T ρ f ) ( x ) = ∫ D G ( x , z ) f ( z ) d ρ ( z ) ∈ [ 0 , ∞ ] , and define, for p ∈ D p\in\mathbb{D} p ∈ D and s ≠ t s\ne t s = t in S 1 S^{1} S 1 (with T ρ 0 = I T_{\rho}^{0}=I T ρ 0 = I , the identity operator),
K ρ ( p , s ) = ∑ n ≥ 0 ( T ρ n P s ) ( p ) , R ρ ( s , t ) = ∑ n ≥ 0 ∫ D P s T ρ n P t d ρ , N ρ = N 0 + R ρ . (1) K^{\rho}(p,s)=\sum_{n\ge0}(T_{\rho}^{n}P_{s})(p),\qquad R^{\rho}(s,t)=\sum_{n\ge0}\int_{\mathbb{D}}P_{s}T_{\rho}^{n}P_{t}\,d\rho,\qquad N^{\rho}=N_{0}+R^{\rho}. \tag*{(1)} K ρ ( p , s ) = n ≥ 0 ∑ ( T ρ n P s ) ( p ) , R ρ ( s , t ) = n ≥ 0 ∑ ∫ D P s T ρ n P t d ρ , N ρ = N 0 + R ρ . (1)
Here T ρ n P s T_{\rho}^{n}P_{s} T ρ n P s denotes an iterated nonnegative kernel integral, even when P s P_{s} P s does not belong to L 2 ( ρ ) L^{2}(\rho) L 2 ( ρ ) . Products involving an infinite value and a zero nonnegative factor are assigned value zero. These are sums of nonnegative integrals: K ρ , N ρ K^{\rho},N^{\rho} K ρ , N ρ have values in ( 0 , ∞ ] (0,\infty] ( 0 , ∞ ] and R ρ R^{\rho} R ρ in [ 0 , ∞ ] [0,\infty] [ 0 , ∞ ] ; for bounded a ρ a_{\rho} a ρ they are the kernels (4.6)–(4.8) of [14 ] . A symmetric kernel D D D with zero diagonal on a set S S S is conditionally negative semidefinite if ∑ i , j ω i ω j D ( s i , s j ) ≤ 0 \sum_{i,j}\omega_{i}\omega_{j}D(s_{i},s_{j})\le0 ∑ i , j ω i ω j D ( s i , s j ) ≤ 0 for all finite families s i ∈ S s_{i}\in S s i ∈ S and real ω i \omega_{i} ω i with ∑ i ω i = 0 \sum_{i}\omega_{i}=0 ∑ i ω i = 0 .
The reciprocal quotient in these kernels is the central analytic object. The kernel assertion needed here concerns bounded nonnegative densities with compact support in D \mathbb{D} D and a strict Dirichlet form bound. We prove that assertion in the kernel section before applying it to the conformal truncations; the main statement has no additional reciprocal-kernel hypothesis.
Theorem A. Let Ω \Omega Ω be an admissible domain in Definition 1.1 . Then for every 0 ≠ u ∈ V 0 \ne u \in V 0 = u ∈ V ,
∇ u ( x ) ≠ 0 ( x ∈ Ω ) , and hence min ∂ Ω u < u ( x ) < max ∂ Ω u ( x ∈ Ω ) . \nabla u(x) \ne0 \qquad(x \in\Omega), \qquad\text{and hence} \qquad\min_{\partial\Omega} u < u(x) < \max_{\partial\Omega} u \qquad(x \in\Omega). ∇ u ( x ) = 0 ( x ∈ Ω ) , and hence ∂ Ω min u < u ( x ) < ∂ Ω max u ( x ∈ Ω ) . A prevertex is a boundary point of the disk mapped to a corner by a conformal bijection Φ : D → Ω \Phi: \mathbb{D} \to\Omega Φ : D → Ω .
Remark 1.2. If m = 0 m = 0 m = 0 , that is, for a C ∞ C^\infty C ∞ Jordan boundary, Theorem A is [14 Theorem 1.1] , and the proof below follows that source except for the finite preservation step; no priority is claimed in that case. The extension addressed here concerns corners: the conformal potential is singular at the prevertices, the boundary gradient extends continuously and vanishes at the corners, and the Schoenberg map is injective only off the prevertices. Lipschitz regularity of the gradient is not assumed. The signed curvature term is retained until it cancels.
Strategy of the proof
Strategy of the proof. Fix 0 ≠ u ∈ V 0 \ne u \in V 0 = u ∈ V and suppose that ∇ u ( ξ 0 ) = 0 \nabla u(\xi_0) = 0 ∇ u ( ξ 0 ) = 0 at an interior point ξ 0 = Φ ( p ) \xi_0 = \Phi(p) ξ 0 = Φ ( p ) , where Φ : D → Ω \Phi: \mathbb{D} \to\Omega Φ : D → Ω is a conformal bijection. The argument has four steps.
A nonnegative boundary energy with known null vectors. For a tangent vector field X X X on Ω \Omega Ω , Rohleder’s principle (Proposition 3.1) gives Q ( X ) ≥ μ ∥ X ∥ 2 2 Q(X) \ge\mu\lVert X \rVert_2^2 Q ( X ) ≥ μ ∥ X ∥ 2 2 , with equality exactly for gradients of eigenfunctions. Pulling back by Φ \Phi Φ and extending boundary data by solutions of − Δ W = q W -\Delta W = qW − Δ W = q W with q = μ ∣ Φ ′ ∣ 2 q = \mu|\Phi'|^2 q = μ ∣ Φ ′ ∣ 2 turns Q − μ ∥ ⋅ ∥ 2 2 Q - \mu\lVert\cdot\rVert_2^2 Q − μ ∥ ⋅ ∥ 2 2 into a nonnegative quadratic form E E E on tangent boundary data on S 1 \mathbb{S}^1 S 1 (§ 5 ). The boundary gradient g = ( ∇ u ) ∘ Φ g = (\nabla u) \circ\Phi g = ( ∇ u ) ∘ Φ satisfies E ( g ) = 0 E(g) = 0 E ( g ) = 0 .
A multiplier identity. For every Lipschitz scalar function b b b on S 1 \mathbb{S}^1 S 1 ,
E ( b g ) = 1 2 ∬ S 1 × S 1 N ( s , t ) ( b ( s ) − b ( t ) ) 2 g ( s ) ⋅ g ( t ) d s d t E(bg) = \frac{1}{2} \iint_{\mathbb{S}^1 \times\mathbb{S}^1} N(s,t)\bigl(b(s)-b(t)\bigr)^2 g(s)\cdot g(t)\,ds\,dt E ( b g ) = 2 1 ∬ S 1 × S 1 N ( s , t ) ( b ( s ) − b ( t ) ) 2 g ( s ) ⋅ g ( t ) d s d t
(Proposition 5.6 ); the signed curvature terms cancel.
A squared-distance kernel. For each p ∈ D p \in\mathbb{D} p ∈ D the kernel D p ( s , t ) = K ( p , s ) K ( p , t ) / N ( s , t ) D_p(s,t) = K(p,s)K(p,t)/N(s,t) D p ( s , t ) = K ( p , s ) K ( p , t ) / N ( s , t ) is conditionally negative semidefinite (Propositions 4.7 and 4.8 ). Hence D p ( s , t ) = ∥ B ( s ) − B ( t ) ∥ 2 D_p(s,t) = \lVert B(s)-B(t)\rVert^2 D p ( s , t ) = ∥ B ( s ) − B ( t ) ∥ 2 for a Lipschitz map B B B into a Hilbert space (Lemma 6.1 ). Applying step 2 to the coordinates b j b_j b j of B B B and summing gives
∑ j E ( b j g ) = 1 2 ∣ ∫ S 1 K ( p , s ) g ( s ) d s ∣ 2 = 1 2 ∣ ∇ u ( ξ 0 ) ∣ 2 = 0. \sum_j E(b_jg) = \frac{1}{2}\left|\int_{\mathbb{S}^1} K(p,s)g(s)\,ds\right|^2 = \frac{1}{2}|\nabla u(\xi_0)|^2 = 0. j ∑ E ( b j g ) = 2 1 ∫ S 1 K ( p , s ) g ( s ) d s 2 = 2 1 ∣∇ u ( ξ 0 ) ∣ 2 = 0.
Contradiction. Every E ( b j g ) E(b_jg) E ( b j g ) vanishes, so every b j g b_jg b j g is the boundary gradient of an eigenfunction. Since B B B is injective off the prevertices, this produces either three linearly independent eigenfunctions, contradicting dim V ≤ 2 \dim V \le2 dim V ≤ 2 (Proposition 2.5 ), or a nonzero eigenfunction with vanishing Cauchy data on a boundary arc, contradicting unique continuation.
Where the corners enter. For a smooth boundary these are the steps of [14 ] . At a corner of angle α j π \alpha_j\pi α j π the potential q q q blows up like ∣ z − s j ∣ 2 α j − 2 |z-s_j|^{2\alpha_j-2} ∣ z − s j ∣ 2 α j − 2 at the prevertex. The kernel theorem is therefore proved first for bounded compactly supported potentials and then passed to q q q by monotone truncation; this requires K ( p , ⋅ ) K(p,\cdot) K ( p , ⋅ ) bounded and N N N finite off the prevertices (Corollary 2.4 and Lemma 4.3). The boundary gradient g g g is continuous and vanishes at the prevertices, but it need not be Lipschitz (Lemma 2.2 ); this is why the energy is defined for bounded H 1 / 2 H^{1/2} H 1/2 data. The curvature identity must hold without vertex terms (Lemma 3.2 ), and the topological step must allow B B B to fail to be injective at finitely many points (Lemma 6.2 ). The condition α j < 1 \alpha_j < 1 α j < 1 is used exactly in these estimates.
Organization. Section 2 2 2 proves the corner estimates and dim V ≤ 2 \dim V \le2 dim V ≤ 2 ; Section 3 3 3 the variational principle and the curvature identity; Section 4 4 4 the kernel theorem and the singular limit; Section 5 5 5 the energy formula and the multiplier identity; Section 6 6 6 the conclusion.
Relation to previous work
The strategy, multiplicity argument, vector variational principle, boundary kernels and Hilbert-space contradiction are adapted from [14 ] . The reciprocal-kernel argument is supplied for bounded compactly supported densities. Its initial regularization and limiting framework follow [14 Section 5] , with proofs included. The finite preservation step is proved by a matrix flow and a cone-invariance argument. This is an alternative to the rank-one update lemma [14 Lemma 5.4] , which preserves the reciprocal condition for a single column update and, when iterated, gives the finite resolvent assertion; it is the inclusion of a proof, not the choice of the flow, that removes [14 Theorem 5.1] as an assumption. The regularization and limiting framework are shared with that source. Their proofs are included here and must be checked as part of the present argument; no assertion of Section 5 is used without proof. Attribution of the statement and of the adapted constructions is retained. Monotone truncation also occurs in that source; its use here needs the corner bounds proved below.
Straight convex polygons are special cases. Compared with that case, three changes are essential: sector separation of variables is replaced by estimates on the disk, tangency-preserving smoothing is carried out in curved coordinates, and a signed curvature integral is retained until it cancels. The variational principle and the treatment of curvature on such domains are already present in [18 Theorem 1.1 and Section 3] ; we include the proofs needed here and make no novelty claim for that framework. Boundary uniqueness on a smooth arc uses extension by zero, not reflection in a line. These changes are proved rather than inferred by analogy. The finite-exceptional-set topology argument is still required.
The classical inputs are the spectral inequality of Filonov, homogeneous Dirichlet and Neumann H 2 H^2 H 2 regularity for curvilinear polygons without reentrant corners, boundary regularity of conformal maps, smooth-boundary elliptic regularity, trace and Sobolev theorems, and Schoenberg’s theorem. We spell out their uses. In particular, the regularity input is local in the corner angles; it does not require the whole domain to be convex.
For context, Rohleder [19 Theorem 1.1 and Remark 2.2] gives unconditional restrictions on interior critical points in convex domains, including absence under the additional condition μ diam ( Ω ) 2 ≤ j 1 , 1 2 \mu\operatorname{diam}(\Omega)^2 \le j_{1,1}^2 μ diam ( Ω ) 2 ≤ j 1 , 1 2 , where j 1 , 1 j_{1,1} j 1 , 1 is the first positive zero of the Bessel function J 1 J_1 J 1 of the first kind and diam ( Ω ) = sup x , y ∈ Ω ∣ x − y ∣ \operatorname{diam}(\Omega)=\sup_{x,y\in\Omega}|x-y| diam ( Ω ) = sup x , y ∈ Ω ∣ x − y ∣ . Bañuelos and Pang [20 Theorem 1.1] prove finiteness of the connected components of positive level sets under smoothness and strictly positive curvature assumptions. These are comparisons of scope, not inputs to the present proof.
Reentrant corners, cusps, infinitely many corners and rough arcs are outside Definition 1.1 . Nor is an approximation argument used to infer strict nonvanishing from a non-strict limiting conclusion.
Corner geometry and analytical estimates
We first separate the classical regularity facts from two estimates that will replace the explicit Bessel expansions available for straight corners.
Lemma 2.1 (classical inputs). (a) μ < λ D \mu< \lambda_D μ < λ D .
(b) For f ∈ L 2 ( Ω ) f \in L^2(\Omega) f ∈ L 2 ( Ω ) , the weak Dirichlet solution of − Δ φ D = f -\Delta\varphi_D=f − Δ φ D = f , φ D ∈ H 0 1 ( Ω ) \varphi_D \in H_0^1(\Omega) φ D ∈ H 0 1 ( Ω ) , and the weak homogeneous Neumann solution of − Δ φ N + φ N = f -\Delta\varphi_N+\varphi_N=f − Δ φ N + φ N = f belonging to H 2 ( Ω ) H^2(\Omega) H 2 ( Ω ) . In particular V ⊂ H 2 ( Ω ) V \subset H^2(\Omega) V ⊂ H 2 ( Ω ) , and its elements are bounded and continuous on Ω ‾ \overline{\Omega} Ω .
(c) A conformal bijection Φ : D → Ω \Phi:\mathbb{D}\to\Omega Φ : D → Ω extends to a homeomorphism of the closures. Put s j = Φ − 1 ( v j ) s_j=\Phi^{-1}(v_j) s j = Φ − 1 ( v j ) and A = S 1 ∖ { s 1 , … , s m } A=S^1\setminus\{s_1,\ldots,s_m\} A = S 1 ∖ { s 1 , … , s m } . On every compact subarc of A A A , Φ \Phi Φ is smooth up to the
boundary and Φ ′ ≠ 0 \Phi' \ne0 Φ ′ = 0 . Near a prevertex,
∣ Φ ′ ( z ) ∣ ≍ ∣ z − s j ∣ α j − 1 , ∣ Φ ( z ) − v j ∣ ≍ ∣ z − s j ∣ α j . (2) |\Phi'(z)| \asymp|z-s_j|^{\alpha_j-1}, \qquad|\Phi(z)-v_j| \asymp|z-s_j|^{\alpha_j}. \tag*{(2)} ∣ Φ ′ ( z ) ∣ ≍ ∣ z − s j ∣ α j − 1 , ∣Φ ( z ) − v j ∣ ≍ ∣ z − s j ∣ α j . (2)
The comparisons hold for approach through D ‾ ∖ { s j } \overline{\mathbb{D}} \setminus\{s_j\} D ∖ { s j } .
(d) The Dirichlet operator − Δ − μ -\Delta-\mu − Δ − μ on Ω \Omega Ω has a positive symmetric Green function G \mathcal{G} G . With G 0 Ω \mathcal{G}_0^\Omega G 0 Ω the Laplace Green function,
G ( ξ , η ) = G 0 Ω ( ξ , η ) + μ ∫ Ω G 0 Ω ( ξ , ζ ) G ( ζ , η ) d ζ . (3) \mathcal{G}(\xi,\eta)=\mathcal{G}_0^\Omega(\xi,\eta)+\mu\int_\Omega\mathcal{G}_0^\Omega(\xi,\zeta)\mathcal{G}(\zeta,\eta)\,d\zeta. \tag*{(3)} G ( ξ , η ) = G 0 Ω ( ξ , η ) + μ ∫ Ω G 0 Ω ( ξ , ζ ) G ( ζ , η ) d ζ . (3)
Its L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) sections are bounded uniformly in the pole. Away from the diagonal it is jointly smooth up to the open smooth arcs in either variable, and vanishes there in each variable.
References and explanations. Part (a) is Filonov’s inequality [21 ] for bounded Lipschitz domains. For (b), use the homogeneous Dirichlet and Neumann regularity theory of [22 ] for curvilinear polygons whose individual arcs are of class C 1 , 1 C^{1,1} C 1 , 1 up to their endpoints, as in Definition 1.1. The whole boundary is not asserted to be C 1 , 1 C^{1,1} C 1 , 1 across a corner. A compact statement for such polygons, taken from [22 ] , is collected in [23 Appendix A.1, Theorem A.2] . For either pure boundary condition the first nonconstant corner exponent is π / ( α j π ) = 1 / α j > 1 \pi/(\alpha_j\pi)=1/\alpha_j>1 π / ( α j π ) = 1/ α j > 1 . Thus no singular term below the H 2 H^2 H 2 threshold occurs. Smooth local boundary charts handle the rest of ∂ Ω \partial\Omega ∂ Ω . This is an angular regularity criterion, not an appeal to global convexity. The Sobolev embedding H 2 ( Ω ) ⊂ C 0 , γ ( Ω ‾ ) H^2(\Omega)\subset C^{0,\gamma}(\overline{\Omega}) H 2 ( Ω ) ⊂ C 0 , γ ( Ω ) for some γ > 0 \gamma>0 γ > 0 gives the last assertion.
Part (c) uses Carathéodory’s theorem and local Kellogg–Warschawski regularity; see [24 Chapters 2–3] . The corner derivative estimates are the usual local power-map consequence of this regularity; see also the proof of [25 Lemma 4.3, equations (8)–(10)] . Specifically, the power ( w − v j ) 1 / α j (w-v_j)^{1/\alpha_j} ( w − v j ) 1/ α j , with a local branch on the interior side of the corner, opens its angle to π \pi π . The two smooth arcs become a C 1 , δ C^{1,\delta} C 1 , δ boundary at the image point, for some δ > 0 \delta>0 δ > 0 . The normalized local conformal map then has a nonzero continuous derivative there. Factoring out the power gives (2) . No holomorphic extension through a merely smooth arc is asserted.
For (d), the Dirichlet resolvent exists by (a), and ( − Δ D − μ ) − 1 = ∑ k ≥ 0 μ k ( − Δ D ) − k − 1 (-\Delta_D-\mu)^{-1}=\sum_{k\geq0}\mu^k(-\Delta_D)^{-k-1} ( − Δ D − μ ) − 1 = ∑ k ≥ 0 μ k ( − Δ D ) − k − 1 is positivity preserving. Set G ( ⋅ , η ) = G 0 Ω ( ⋅ , η ) + μ ( − Δ D − μ ) − 1 G 0 Ω ( ⋅ , η ) \mathcal{G}(\cdot,\eta)=\mathcal{G}_0^\Omega(\cdot,\eta)+\mu(-\Delta_D-\mu)^{-1}\mathcal{G}_0^\Omega(\cdot,\eta) G ( ⋅ , η ) = G 0 Ω ( ⋅ , η ) + μ ( − Δ D − μ ) − 1 G 0 Ω ( ⋅ , η ) . For any fixed R ∗ > diam ( Ω ) R_*>\operatorname{diam}(\Omega) R ∗ > diam ( Ω ) , domain monotonicity and the Green function of the disk B ( η , R ∗ ) B(\eta,R_*) B ( η , R ∗ ) give 0 ≤ G 0 Ω ( ξ , η ) ≤ ( 2 π ) − 1 log ( R ∗ / ∣ ξ − η ∣ ) 0\leq\mathcal{G}_0^\Omega(\xi,\eta)\leq(2\pi)^{-1}\log(R_*/|\xi-\eta|) 0 ≤ G 0 Ω ( ξ , η ) ≤ ( 2 π ) − 1 log ( R ∗ /∣ ξ − η ∣ ) . Its square has a uniformly bounded integral over that disk. Together with the bounded resolvent L 2 → H 2 → L ∞ L^2\to H^2\to L^\infty L 2 → H 2 → L ∞ , this proves uniform L 2 L^2 L 2 sections and 0 ≤ G − G 0 Ω ≤ C 0\leq\mathcal{G}-\mathcal{G}_0^\Omega\leq C 0 ≤ G − G 0 Ω ≤ C . The resolvent identity gives (3) . Local smooth-boundary elliptic regularity, applied in each variable on disjoint neighborhoods, gives the asserted joint regularity off the diagonal.
Fix once and for all a map Φ \Phi Φ as in Lemma 2.1(c), together with its prevertices s j s_j s j and the regular boundary set A A A . Write a = μ ∣ Φ ′ ∣ 2 a=\mu|\Phi'|^2 a = μ ∣ Φ ′ ∣ 2 . By (2) , for some 1 < p 0 < ∞ 1<p_0<\infty 1 < p 0 < ∞ ,
a ∈ L p 0 ( D ) , 0 ≤ a ( z ) ≤ C ( 1 + ∑ j = 1 m ∣ z − s j ∣ 2 α j − 2 ) . (4) a\in L^{p_0}(\mathbb{D}),\qquad0\leq a(z)\leq C\left(1+\sum_{j=1}^{m}|z-s_j|^{2\alpha_j-2}\right). \tag*{(4)} a ∈ L p 0 ( D ) , 0 ≤ a ( z ) ≤ C ( 1 + j = 1 ∑ m ∣ z − s j ∣ 2 α j − 2 ) . (4)
When m ≥ 1 m\geq1 m ≥ 1 , one may choose 1 < p 0 < min j ( 1 − α j ) − 1 1<p_0<\min_j(1-\alpha_j)^{-1} 1 < p 0 < min j ( 1 − α j ) − 1 . If there are no corners, a a a is bounded and all the following boundary estimates are immediate from smooth-boundary regularity.
We record the Sobolev changes of variables used below. Composition with Φ \Phi Φ is an isometry for the Dirichlet seminorm on compactly supported smooth functions. Completing these spaces in that seminorm, which is equivalent to the H 0 1 H^1_0 H 0 1 norm by Poincaré’s inequality on each bounded domain, gives an isomorphism between H 0 1 ( Ω ) H^1_0(\Omega) H 0 1 ( Ω ) and H 0 1 ( D ) H^1_0(\mathbb{D}) H 0 1 ( D ) . The limits agree with ordinary composition on compact subsets, where Φ \Phi Φ is a smooth diffeomorphism. For a general $\psi\in H^1(\mathbb{D}), put p 0 ′ = p 0 / ( p 0 − 1 ) p'_0 = p_0/(p_0 - 1) p 0 ′ = p 0 / ( p 0 − 1 ) . Hölder’s inequality gives
∫ Ω ∣ ψ ∘ Φ − 1 ∣ 2 = μ − 1 ∫ D a ∣ ψ ∣ 2 ≤ μ − 1 ∥ a ∥ L p 0 ∥ ψ ∥ L 2 p 0 ′ 2 < ∞ . \int_{\Omega} \lvert\psi\circ\Phi^{-1}\rvert^2 = \mu^{-1}\int_D a\lvert\psi\rvert^2 \le\mu^{-1}\lVert a\rVert_{L^{p_0}}\lVert\psi\rVert_{L^{2p'_0}}^2 < \infty. ∫ Ω ∣ ψ ∘ Φ − 1 ∣ 2 = μ − 1 ∫ D a ∣ ψ ∣ 2 ≤ μ − 1 ∥ a ∥ L p 0 ∥ ψ ∥ L 2 p 0 ′ 2 < ∞.
The local chain rule and the invariant Dirichlet integral thus place ψ ∘ Φ − 1 \psi\circ\Phi^{-1} ψ ∘ Φ − 1 in H 1 ( Ω ) H^1(\Omega) H 1 ( Ω ) . Conversely a bounded physical H 1 H^1 H 1 function pulls back to H 1 ( D ) H^1(D) H 1 ( D ) , by the same energy identity and boundedness on the disk. These facts do not use any boundary kernel.
Lemma 2.2 (vanishing of the physical gradient). For u ∈ V u\in V u ∈ V , ∇ u ∈ H 1 ( Ω ; R 2 ) ∩ C ( Ω ‾ ; R 2 ) \nabla u\in H^1(\Omega;\mathbb{R}^2)\cap C(\overline{\Omega};\mathbb{R}^2) ∇ u ∈ H 1 ( Ω ; R 2 ) ∩ C ( Ω ; R 2 ) and ∇ u ( v j ) = 0 \nabla u(v_j)=0 ∇ u ( v j ) = 0 . More precisely , with r = ∣ z − s j ∣ r=\lvert z-s_j\rvert r = ∣ z − s j ∣ small,
∣ ∇ u ( Φ ( z ) ) ∣ ≤ C { r α j , 0 < α j < 1 2 , r 1 / 2 ( 1 + ∣ log r ∣ ) , α j = 1 2 , r 1 − α j , 1 2 < α j < 1. (5) \lvert\nabla u(\Phi(z))\rvert\le C
\begin{cases}
r^{\alpha_j}, & 0<\alpha_j<\frac{1}{2},\\
r^{1/2}(1+\lvert\log r\rvert), & \alpha_j=\frac{1}{2},\\
r^{1-\alpha_j}, & \frac{1}{2}<\alpha_j<1.
\end{cases} \tag*{(5)} ∣ ∇ u ( Φ ( z ))∣ ≤ C ⎩ ⎨ ⎧ r α j , r 1/2 ( 1 + ∣ log r ∣) , r 1 − α j , 0 < α j < 2 1 , α j = 2 1 , 2 1 < α j < 1. (5) On Γ \Gamma Γ , the gradient is smooth and tangent.
Proof. By Lemma 2.1, u ∈ H 2 u\in H^2 u ∈ H 2 is bounded and smooth up to Γ \Gamma Γ . Put U = u ∘ Φ U=u\circ\Phi U = u ∘ Φ . Conformal invariance gives U ∈ H 1 ( D ) ∩ L ∞ ( D ) U\in H^1(D)\cap L^\infty(D) U ∈ H 1 ( D ) ∩ L ∞ ( D ) and
∫ D ∇ U ⋅ ∇ ψ = ∫ D a U ψ ( ψ ∈ H 1 ( D ) ) , ∫ D a U = 0. (6) \int_D \nabla U\cdot\nabla\psi=\int_D aU\psi\quad(\psi\in H^1(D)),\qquad\int_D aU=0. \tag*{(6)} ∫ D ∇ U ⋅ ∇ ψ = ∫ D a U ψ ( ψ ∈ H 1 ( D )) , ∫ D a U = 0. (6)
To justify the full test space, ψ ∘ Φ − 1 \psi\circ\Phi^{-1} ψ ∘ Φ − 1 has finite Dirichlet integral and belongs to L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) : by (4) and the planar Sobolev embedding, ∫ D a ψ 2 < ∞ \int_D a\psi^2<\infty ∫ D a ψ 2 < ∞ . It is consequently an admissible test in the physical Neumann problem.
The normalized Neumann Green function of the disk is
G N ( z , y ) = − log ∣ z − y ∣ + log ∣ 1 − z y ‾ ∣ 2 π + ∣ z ∣ 2 + ∣ y ∣ 2 4 π − 3 8 π . G_N(z,y)=-\frac{\log\lvert z-y\rvert+\log\lvert1-z\overline{y}\rvert}{2\pi}+\frac{\lvert z\rvert^2+\lvert y\rvert^2}{4\pi}-\frac{3}{8\pi}. G N ( z , y ) = − 2 π log ∣ z − y ∣ + log ∣ 1 − z y ∣ + 4 π ∣ z ∣ 2 + ∣ y ∣ 2 − 8 π 3 .
Direct differentiation and integration verify − Δ z G N = δ y − 1 / π -\Delta_zG_N=\delta_y-1/\pi − Δ z G N = δ y − 1/ π , ∂ ν D G N = 0 \partial_{\nu_D}G_N=0 ∂ ν D G N = 0 and ∫ D G N ( z , y ) d z = 0 \int_D G_N(z,y)\,dz=0 ∫ D G N ( z , y ) d z = 0 . Furthermore,
∣ 1 − z y ‾ ∣ 2 − ∣ z − y ∣ 2 = ( 1 − ∣ z ∣ 2 ) ( 1 − ∣ y ∣ 2 ) ≥ 0 , ∣ ∇ z G N ( z , y ) ∣ ≤ C ( 1 + ∣ z − y ∣ − 1 ) . \lvert1-z\overline{y}\rvert^2-\lvert z-y\rvert^2=(1-\lvert z\rvert^2)(1-\lvert y\rvert^2)\ge0,\qquad\lvert\nabla_zG_N(z,y)\rvert\le C(1+\lvert z-y\rvert^{-1}). ∣ 1 − z y ∣ 2 − ∣ z − y ∣ 2 = ( 1 − ∣ z ∣ 2 ) ( 1 − ∣ y ∣ 2 ) ≥ 0 , ∣ ∇ z G N ( z , y )∣ ≤ C ( 1 + ∣ z − y ∣ − 1 ) .
The representation of the mean-zero part of U U U with source a U ∈ L p 0 aU\in L^{p_0} a U ∈ L p 0 therefore gives, at interior points,
∣ ∇ U ( z ) ∣ ≤ C ∫ D ( 1 + ∣ z − y ∣ − 1 ) a ( y ) d y . (7) \lvert\nabla U(z)\rvert\le C\int_D(1+\lvert z-y\rvert^{-1})a(y)\,dy. \tag*{(7)} ∣ ∇ U ( z )∣ ≤ C ∫ D ( 1 + ∣ z − y ∣ − 1 ) a ( y ) d y . (7)
One can obtain the representation first for smooth mean-zero sources and then pass to L p 0 L^{p_0} L p 0 using the Neumann solution operator on the smooth disk. For the gradient formula at a fixed interior point, the source is smooth locally; the remaining integral is differentiated away from its singularity.
Fix 0 < r 0 < 1 0<r_0<1 0 < r 0 < 1 . For β > 0 \beta>0 β > 0 , s ∈ S 1 s\in S^1 s ∈ S 1 and r = ∣ z − s ∣ < r 0 / 4 r=\lvert z-s\rvert<r_0/4 r = ∣ z − s ∣ < r 0 /4 , a direct splitting of the integral gives
∫ D ∩ B ( s , r 0 ) ∣ y − s ∣ β − 2 ∣ z − y ∣ d y ≤ C { r β − 1 , 0 < β < 1 , 1 + ∣ log r ∣ , β = 1 , 1 , β > 1. (8) \int_{D\cap B(s,r_0)}\frac{\lvert y-s\rvert^{\beta-2}}{\lvert z-y\rvert}\,dy\le C
\begin{cases}
r^{\beta-1}, & 0<\beta<1,\\
1+\lvert\log r\rvert, & \beta=1,\\
1, & \beta>1.
\end{cases} \tag*{(8)} ∫ D ∩ B ( s , r 0 ) ∣ z − y ∣ ∣ y − s ∣ β − 2 d y ≤ C ⎩ ⎨ ⎧ r β − 1 , 1 + ∣ log r ∣ , 1 , 0 < β < 1 , β = 1 , β > 1. (8)
Here is the estimate explicitly. On ∣ y − s ∣ < r / 2 \lvert y-s\rvert<r/2 ∣ y − s ∣ < r /2 , the denominator is at least r / 2 r/2 r /2 , giving C r β − 1 Cr^{\beta-1} C r β − 1 . On r / 2 ≤ ∣ y − s ∣ ≤ 2 r r/2\le\lvert y-s\rvert\le2r r /2 ≤ ∣ y − s ∣ ≤ 2 r , the weight is at most C r β − 2 Cr^{\beta-2} C r β − 2 and the integral of ∣ z − y ∣ − 1 \lvert z-y\rvert^{-1} ∣ z − y ∣ − 1 over that set is at most C r Cr C r . On 2 r < ∣ y − s ∣ < r 0 2r<\lvert y-s\rvert<r_0 2 r < ∣ y − s ∣ < r 0 , the denominator is at least ∣ y − s ∣ / 2 \lvert y-s\rvert/2 ∣ y − s ∣ /2 , giving C ∫ 2 r r 0 t β − 2 d t C\int_{2r}^{r_0}t^{\beta-2}\,dt C ∫ 2 r r 0 t β − 2 d t . These are the three cases in (8) ; the other prevertices contribute a bounded amount.
Apply this with β = 2 α j \beta= 2\alpha_j β = 2 α j in (7) , and use ∣ ∇ u ( Φ ( z ) ) ∣ = ∣ ∇ U ( z ) ∣ / ∣ Φ ′ ( z ) ∣ |\nabla u(\Phi(z))| = |\nabla U(z)|/|\Phi'(z)| ∣∇ u ( Φ ( z )) ∣ = ∣∇ U ( z ) ∣/∣ Φ ′ ( z ) ∣ and (2) . This gives (5) , whose right side tends to zero in every case. Together with smoothness up to each open arc this gives the continuous extension of the gradient, while its H 1 H^1 H 1 membership follows from u ∈ H 2 u \in H^2 u ∈ H 2 .
Lemma 2.3 (a Dirichlet regularity iteration). Suppose a a a obeys (4) . If v ∈ H 0 1 ( D ) ∩ L ∞ ( D ) v \in H^1_0(\mathbb{D}) \cap L^\infty(\mathbb{D}) v ∈ H 0 1 ( D ) ∩ L ∞ ( D ) satisfies − Δ v = a v + f -\Delta v = av + f − Δ v = a v + f with f ∈ L ∞ ( D ) f \in L^\infty(\mathbb{D}) f ∈ L ∞ ( D ) , then v ∈ W 2 , p ( D ) v \in W^{2,p}(\mathbb{D}) v ∈ W 2 , p ( D ) for some p > 2 p > 2 p > 2 . In particular v ∈ C 1 , γ ( D ‾ ) v \in C^{1,\gamma}(\overline{\mathbb{D}}) v ∈ C 1 , γ ( D ) for some γ > 0 \gamma> 0 γ > 0 .
Proof. The case m = 0 m = 0 m = 0 follows from smooth-disk regularity with bounded a a a . Otherwise let α ∗ = min 1 ≤ j ≤ m α j > 0 \alpha_* = \min_{1 \le j \le m} \alpha_j > 0 α ∗ = min 1 ≤ j ≤ m α j > 0 . Initially a v + f ∈ L p 0 av + f \in L^{p_0} a v + f ∈ L p 0 for any 1 < p 0 < min j ( 1 − α j ) − 1 1 < p_0 < \min_j(1-\alpha_j)^{-1} 1 < p 0 < min j ( 1 − α j ) − 1 . Varying p 0 p_0 p 0 , Dirichlet W 2 , p 0 W^{2,p_0} W 2 , p 0 regularity on the disk and Sobolev embedding give v ∈ C 0 , γ 0 ( D ‾ ) v \in C^{0,\gamma_0}(\overline{\mathbb{D}}) v ∈ C 0 , γ 0 ( D ) for every 0 < γ 0 < min ( 1 , 2 α ∗ ) 0 < \gamma_0 < \min(1,2\alpha_*) 0 < γ 0 < min ( 1 , 2 α ∗ ) ; only such subunit exponents will be used. Since the boundary trace is zero, ∣ v ( z ) ∣ ≤ C ∣ z − s j ∣ γ 0 |v(z)| \le C|z-s_j|^{\gamma_0} ∣ v ( z ) ∣ ≤ C ∣ z − s j ∣ γ 0 near every s j s_j s j . Consequently, if such an exponent γ 0 ∈ [ 0 , 1 ) \gamma_0 \in[0,1) γ 0 ∈ [ 0 , 1 ) is available, then
∣ a v + f ∣ ≤ C ( 1 + ∑ j ∣ z − s j ∣ 2 α j − 2 + γ 0 ) . |av+f| \le C\left(1+\sum_j |z-s_j|^{2\alpha_j-2+\gamma_0}\right). ∣ a v + f ∣ ≤ C ( 1 + j ∑ ∣ z − s j ∣ 2 α j − 2 + γ 0 ) .
This belongs to L p L^p L p whenever p > 1 p > 1 p > 1 and p ( 2 − 2 α ∗ − γ 0 ) < 2 p(2-2\alpha_*-\gamma_0) < 2 p ( 2 − 2 α ∗ − γ 0 ) < 2 if the expression in parentheses is positive; if it is nonpositive there is no finite upper restriction on p p p . Another W 2 , p W^{2,p} W 2 , p estimate therefore yields every exponent
0 < γ 1 < min ( 1 , γ 0 + 2 α ∗ ) . 0 < \gamma_1 < \min(1,\gamma_0+2\alpha_*). 0 < γ 1 < min ( 1 , γ 0 + 2 α ∗ ) .
Starting with boundedness, finitely many iterations give an exponent 0 < γ 0 < 1 0 < \gamma_0 < 1 0 < γ 0 < 1 with γ 0 > 1 − 2 α ∗ \gamma_0 > 1-2\alpha_* γ 0 > 1 − 2 α ∗ . For example one can increase the exponent in steps smaller than α ∗ \alpha_* α ∗ until crossing that threshold, choosing a last step below 1. The displayed integrability condition then allows some p > 2 p > 2 p > 2 . A final disk Dirichlet estimate proves the assertion. All exponent choices are strict, so no endpoint embedding is used. □
Corollary 2.4 (boundary control of the Green function). For fixed p ∈ D p \in\mathbb{D} p ∈ D , the function U p ( z ) = G ( Φ ( p ) , Φ ( z ) ) U_p(z) = \mathcal{G}(\Phi(p),\Phi(z)) U p ( z ) = G ( Φ ( p ) , Φ ( z )) has bounded first derivatives in a neighborhood of S 1 \mathbb{S}^1 S 1 and extends there with class C 1 , γ C^{1,\gamma} C 1 , γ for some γ > 0 \gamma> 0 γ > 0 .
Proof. Choose a smooth cutoff χ \chi χ equal to zero near p p p and equal to one near S 1 \mathbb{S}^1 S 1 . The decomposition of G \mathcal{G} G in Lemma 2.1 and conformal invariance show that v = χ U p ∈ H 0 1 ( D ) ∩ L ∞ ( D ) v = \chi U_p \in H^1_0(\mathbb{D}) \cap L^\infty(\mathbb{D}) v = χ U p ∈ H 0 1 ( D ) ∩ L ∞ ( D ) : the Laplace Green singularity has been removed, and the resolvent correction is a bounded physical H 0 1 H^1_0 H 0 1 function. Away from p p p , − Δ U p = a U p -\Delta U_p = aU_p − Δ U p = a U p , so − Δ v = a v + f -\Delta v = av + f − Δ v = a v + f with f = − 2 ∇ χ ⋅ ∇ U p − ( Δ χ ) U p f = -2\nabla\chi\cdot\nabla U_p - (\Delta\chi)U_p f = − 2∇ χ ⋅ ∇ U p − ( Δ χ ) U p smooth and compactly supported in D ∖ { p } \mathbb{D}\setminus\{p\} D ∖ { p } . Apply Lemma 2.3 . □
The two preceding estimates do different jobs: Lemma 2.2 controls Neumann eigenfunctions in physical coordinates, whereas Corollary 2.4 controls a Dirichlet Green function in disk coordinates. Confusing these two boundary problems would not justify either use.
Proposition 2.5 (after [26 ] and [14 Proposition 2.5] ). dim V ≤ 2 \dim V \le2 dim V ≤ 2 .
Proof. We follow the boundary-sign proof of [14 Lemmas 2.3, 2.4 and Proposition 2.5] and indicate the changes. Let L ∈ H 0 1 ( Ω ) L \in H^1_0(\Omega) L ∈ H 0 1 ( Ω ) solve ( − Δ − μ ) L = 1 (-\Delta-\mu)L=1 ( − Δ − μ ) L = 1 . Then L ∈ H 2 ( Ω ) L \in H^2(\Omega) L ∈ H 2 ( Ω ) by Lemma 2.1 (b), L ≥ 0 L \ge0 L ≥ 0 (test with the negative part and use μ < λ D \mu< \lambda_D μ < λ D ), − Δ L = 1 + μ L > 0 -\Delta L = 1+\mu L > 0 − Δ L = 1 + μL > 0 , so L > 0 L > 0 L > 0 in Ω \Omega Ω , and L L L is smooth up to the open smooth arcs. Hopf’s lemma gives − ∂ ν L > 0 -\partial_\nu L > 0 − ∂ ν L > 0 on every open smooth arc. For v ∈ V v \in V v ∈ V , Green’s identity for H 2 H^2 H 2 functions on a Lipschitz domain gives
0 = ∫ Ω v = ∫ Ω v ( − Δ − μ ) L = − ∫ ∂ Ω v ∂ ν L d σ = ∫ S 1 ω ‾ ( v ∘ Φ ) d s , ω ‾ = ∣ Φ ′ ∣ ( − ∂ ν L ) ∘ Φ . 0=\int_\Omega v=\int_\Omega v(-\Delta-\mu)L=-\int_{\partial\Omega}v\,\partial_\nu L\,d\sigma=\int_{\mathbb{S}^1}\overline{\omega}(v\circ\Phi)\,ds,\qquad\overline{\omega}=|\Phi'|(-\partial_\nu L)\circ\Phi. 0 = ∫ Ω v = ∫ Ω v ( − Δ − μ ) L = − ∫ ∂ Ω v ∂ ν L d σ = ∫ S 1 ω ( v ∘ Φ ) d s , ω = ∣ Φ ′ ∣ ( − ∂ ν L ) ∘ Φ.
Here ϖ ∈ L 1 ( S 1 ) \varpi\in L^{1}(\mathbb{S}^{1}) ϖ ∈ L 1 ( S 1 ) and ϖ > 0 \varpi> 0 ϖ > 0 on A A A . The trace h v = v ∘ Φ ∣ S 1 h_v = v \circ\Phi|_{\mathbb{S}^{1}} h v = v ∘ Φ ∣ S 1 is continuous. If h v = 0 h_v = 0 h v = 0 then v ∈ H 0 1 ( Ω ) v \in H^1_0(\Omega) v ∈ H 0 1 ( Ω ) and μ < λ D \mu< \lambda_D μ < λ D forces v = 0 v = 0 v = 0 ; otherwise h v h_v h v takes both signs. The sets { v > 0 } \{v > 0\} { v > 0 } and { v < 0 } \{v < 0\} { v < 0 } are connected. Indeed, if there were at least three nodal domains, restrict v v v to each of two of them and extend these restrictions by zero to the rest of Ω \Omega Ω . These functions belong to H 1 ( Ω ) H^1(\Omega) H 1 ( Ω ) and retain Rayleigh quotient μ \mu μ , by truncation at the zero level and the weak eigenfunction identity. They can be combined with nonzero coefficients to have mean zero. Their Rayleigh quotient would still be μ \mu μ , so the combination would be a μ \mu μ -eigenfunction vanishing on the third nodal domain. Interior unique continuation rules this out. This is the usual variational proof of the first-positive-eigenvalue case of Courant’s argument. A boundary point where v > 0 v > 0 v > 0 has a neighborhood in Ω ‾ \overline{\Omega} Ω on which v > 0 v > 0 v > 0 , and the local Lipschitz boundary chart joins it to nearby interior points through the same sign region; hence four cyclically ordered boundary points with strictly alternating signs would give two disjoint crosscuts of D ‾ \overline{D} D (after applying the homeomorphism Φ − 1 \Phi^{-1} Φ − 1 ) with alternating endpoints, which is impossible. For completeness, we spell out the last step of the boundary-sign proof in [14 Proposition 2.5] . Parametrize the circle by s = e i θ s = e^{i\theta} s = e i θ . If dim V ≥ 3 \dim V \ge3 dim V ≥ 3 , choose 0 ≠ v ∈ V 0 \ne v \in V 0 = v ∈ V with ∫ ϖ h v cos θ = ∫ ϖ h v sin θ = 0 \int\varpi h_v \cos\theta= \int\varpi h_v \sin\theta= 0 ∫ ϖ h v cos θ = ∫ ϖ h v sin θ = 0 and write h = h v h = h_v h = h v . Cut the circle at a strictly negative point of angle θ 0 \theta_0 θ 0 . Let θ − \theta_- θ − , θ + \theta_+ θ + be the infimum and supremum of the angles where h > 0 h > 0 h > 0 in ( θ 0 , θ 0 + 2 π ) (\theta_0,\theta_0 + 2\pi) ( θ 0 , θ 0 + 2 π ) . Continuity and the signs give θ 0 < θ − < θ + < θ 0 + 2 π \theta_0 < \theta_- < \theta_+ < \theta_0 + 2\pi θ 0 < θ − < θ + < θ 0 + 2 π . There is no positive value outside [ θ − , θ + ] [\theta_-,\theta_+] [ θ − , θ + ] , and no negative value inside ( θ − , θ + ) (\theta_-,\theta_+) ( θ − , θ + ) : otherwise two positive points on either side, together with the negative cut point, give four alternating signs. Set
F ( θ ) = cos ( θ − θ − + θ + 2 ) − cos ( θ + − θ − 2 ) . F(\theta) = \cos\left(\theta- \frac{\theta_- + \theta_+}{2}\right) - \cos\left(\frac{\theta_+ - \theta_-}{2}\right). F ( θ ) = cos ( θ − 2 θ − + θ + ) − cos ( 2 θ + − θ − ) .
This function is positive on ( θ − , θ + ) (\theta_-,\theta_+) ( θ − , θ + ) and negative outside [ θ − , θ + ] [\theta_-,\theta_+] [ θ − , θ + ] in the cut interval. Thus F h ≥ 0 Fh \ge0 F h ≥ 0 everywhere and F h > 0 Fh > 0 F h > 0 on a nonempty open set. Since ϖ > 0 \varpi> 0 ϖ > 0 almost everywhere, ∫ ϖ F h > 0 \int\varpi Fh > 0 ∫ ϖ F h > 0 , contradicting the orthogonality of ϖ h \varpi h ϖ h to 1 1 1 , cos θ \cos\theta cos θ and sin θ \sin\theta sin θ . □
The vector principle and the curvature identity
A field X ∈ H 1 ( Ω ; R 2 ) X \in H^1(\Omega;\mathbb{R}^2) X ∈ H 1 ( Ω ; R 2 ) is tangent if the trace of X ⋅ ν X \cdot\nu X ⋅ ν vanishes on every open smooth arc. Write div X = ∂ 1 X 1 + ∂ 2 X 2 \operatorname{div} X = \partial_1 X_1 + \partial_2 X_2 div X = ∂ 1 X 1 + ∂ 2 X 2 and curl X = ∂ 1 X 2 − ∂ 2 X 1 \operatorname{curl} X = \partial_1 X_2 - \partial_2 X_1 curl X = ∂ 1 X 2 − ∂ 2 X 1 , and put Q ( X ) = ∫ Ω ( ( div X ) 2 + ( curl X ) 2 ) Q(X) = \int_{\Omega}((\operatorname{div} X)^2 + (\operatorname{curl} X)^2) Q ( X ) = ∫ Ω (( div X ) 2 + ( curl X ) 2 ) .
Proposition 3.1 (Rohleder’s principle [18 ] ). For every tangent X ∈ H 1 ( Ω ; R 2 ) X \in H^1(\Omega;\mathbb{R}^2) X ∈ H 1 ( Ω ; R 2 ) ,
Q ( X ) ≥ μ ∫ Ω ∣ X ∣ 2 , Q(X) \ge\mu\int_{\Omega} |X|^2, Q ( X ) ≥ μ ∫ Ω ∣ X ∣ 2 , with equality if and only if X = ∇ v X = \nabla v X = ∇ v for some v ∈ V v \in V v ∈ V .
Proof. Solve Δ φ N = div X \Delta\varphi_N = \operatorname{div} X Δ φ N = div X with ∂ ν φ N = 0 \partial_\nu\varphi_N = 0 ∂ ν φ N = 0 and ∫ Ω φ N = 0 \int_{\Omega}\varphi_N = 0 ∫ Ω φ N = 0 , and solve Δ φ D = curl X \Delta\varphi_D = \operatorname{curl} X Δ φ D = curl X with φ D ∈ H 0 1 ( Ω ) \varphi_D \in H^1_0(\Omega) φ D ∈ H 0 1 ( Ω ) . Tangency gives the Neumann compatibility condition ∫ Ω div X = 0 \int_{\Omega}\operatorname{div} X = 0 ∫ Ω div X = 0 . Both solutions are in H 2 ( Ω ) H^2(\Omega) H 2 ( Ω ) by Lemma 2.1 (b); for the Neumann solution apply that statement to ( − Δ + 1 ) φ N = − div X + φ N ∈ L 2 (-\Delta+1)\varphi_N = -\operatorname{div} X + \varphi_N \in L^2 ( − Δ + 1 ) φ N = − div X + φ N ∈ L 2 . For ∇ ⊥ φ D = ( − ∂ 2 φ D , ∂ 1 φ D ) \nabla^\perp\varphi_D = (-\partial_2\varphi_D,\partial_1\varphi_D) ∇ ⊥ φ D = ( − ∂ 2 φ D , ∂ 1 φ D ) , the orientation convention gives
( ∇ ⊥ φ D ) ⋅ ν = − ∂ τ φ D = 0 on Γ . (\nabla^\perp\varphi_D)\cdot\nu= -\partial_\tau\varphi_D = 0 \quad\text{on }\Gamma. ( ∇ ⊥ φ D ) ⋅ ν = − ∂ τ φ D = 0 on Γ.
Thus Z = X − ∇ φ N − ∇ ⊥ φ D Z = X - \nabla\varphi_N - \nabla^\perp\varphi_D Z = X − ∇ φ N − ∇ ⊥ φ D is in H 1 H^1 H 1 , tangent, divergence free and curl free. On the simply connected Ω \Omega Ω it has a potential Z = ∇ φ H Z = \nabla\varphi_H Z = ∇ φ H , with φ H ∈ H 1 ( Ω ) \varphi_H \in H^1(\Omega) φ H ∈ H 1 ( Ω ) after adding a constant. The weak integration-by-parts identity gives ∫ Ω ∣ ∇ φ H ∣ 2 = 0 \int_{\Omega}|\nabla\varphi_H|^2 = 0 ∫ Ω ∣∇ φ H ∣ 2 = 0 . Hence X = ∇ φ N + ∇ ⊥ φ D X = \nabla\varphi_N + \nabla^\perp\varphi_D X = ∇ φ N + ∇ ⊥ φ D . The two terms are L 2 L^2 L 2 -orthogonal, because div ( ∇ ⊥ φ D ) = 0 \operatorname{div}(\nabla^\perp\varphi_D) = 0 div ( ∇ ⊥ φ D ) = 0 and its normal trace vanishes. Consequently Q ( X ) = ∥ Δ φ N ∥ 2 2 + ∥ Δ φ D ∥ 2 2 Q(X) = \|\Delta\varphi_N\|_2^2 + \|\Delta\varphi_D\|_2^2 Q ( X ) = ∥Δ φ N ∥ 2 2 + ∥Δ φ D ∥ 2 2 . Integration by parts, Cauchy–Schwarz and the Neumann Poincaré inequality give
∥ ∇ φ N ∥ 2 2 = − ∫ Ω φ N Δ φ N ≤ μ − 1 / 2 ∥ ∇ φ N ∥ 2 ∥ Δ φ N ∥ 2 . \|\nabla\varphi_N\|_2^2 = -\int_{\Omega}\varphi_N\Delta\varphi_N \le\mu^{-1/2}\|\nabla\varphi_N\|_2\|\Delta\varphi_N\|_2. ∥∇ φ N ∥ 2 2 = − ∫ Ω φ N Δ φ N ≤ μ − 1/2 ∥∇ φ N ∥ 2 ∥Δ φ N ∥ 2 .
Together with the analogous Dirichlet inequality this yields
Q ( X ) ≥ μ ∥ ∇ φ N ∥ 2 2 + λ D ∥ ∇ φ D ∥ 2 2 = μ ∥ X ∥ 2 2 + ( λ D − μ ) ∥ ∇ φ D ∥ 2 2 . Q(X) \ge\mu\lVert\nabla\varphi_N\rVert_2^2+\lambda_D\lVert\nabla\varphi_D\rVert_2^2=\mu\lVert X\rVert_2^2+(\lambda_D-\mu)\lVert\nabla\varphi_D\rVert_2^2. Q ( X ) ≥ μ ∥ ∇ φ N ∥ 2 2 + λ D ∥ ∇ φ D ∥ 2 2 = μ ∥ X ∥ 2 2 + ( λ D − μ ) ∥ ∇ φ D ∥ 2 2 .
Equality forces φ D = 0 \varphi_D=0 φ D = 0 and equality in the Neumann Rayleigh quotient, so φ N ∈ V \varphi_N\in V φ N ∈ V (including φ N = 0 \varphi_N=0 φ N = 0 ). Conversely, for v ∈ V v\in V v ∈ V , ∇ v \nabla v ∇ v is an H 1 H^1 H 1 tangent field and Q ( ∇ v ) = μ 2 ∥ v ∥ 2 2 = μ ∥ ∇ v ∥ 2 2 Q(\nabla v)=\mu^2\lVert v\rVert_2^2=\mu\lVert\nabla v\rVert_2^2 Q ( ∇ v ) = μ 2 ∥ v ∥ 2 2 = μ ∥ ∇ v ∥ 2 2 .
The spectral principle is expressed using divergence and curl. To transfer it conformally we need its gradient form, including the signed curvature integral. This is the curvature form underlying [18 Section 3] ; the curvature convention there has the opposite sign to ours. We give the density argument for all tangent H 1 H^1 H 1 fields explicitly.
Lemma 3.2 (curvature identity without a vertex term). For every tangent field X ∈ H 1 ( Ω ; R 2 ) X\in H^1(\Omega;\mathbb{R}^2) X ∈ H 1 ( Ω ; R 2 ) ,
Q ( X ) = ∫ Ω ∣ ∇ X ∣ 2 + ∫ Γ κ ∣ Tr X ∣ 2 d σ . (9) Q(X)=\int_{\Omega}\lvert\nabla X\rvert^2+\int_{\Gamma}\kappa\lvert\operatorname{Tr}X\rvert^2\,d\sigma. \tag*{(9)} Q ( X ) = ∫ Ω ∣ ∇ X ∣ 2 + ∫ Γ κ ∣ Tr X ∣ 2 d σ . (9) The trace integral is finite and there is no additional term supported at the vertices.
Proof. Pointwise, ( div X ) 2 + ( curl X ) 2 − ∣ ∇ X ∣ 2 = 2 det ∇ X (\operatorname{div}X)^2+(\operatorname{curl}X)^2-\lvert\nabla X\rvert^2=2\det\nabla X ( div X ) 2 + ( curl X ) 2 − ∣ ∇ X ∣ 2 = 2 det ∇ X . For a smooth tangent field that vanishes near the corners, the divergence theorem and integration by parts along the boundary give
2 ∫ Ω det ∇ X = ∫ Γ ( X 1 ∂ τ X 2 − X 2 ∂ τ X 1 ) d σ . 2\int_{\Omega}\det\nabla X=\int_{\Gamma}(X_1\partial_\tau X_2-X_2\partial_\tau X_1)\,d\sigma. 2 ∫ Ω det ∇ X = ∫ Γ ( X 1 ∂ τ X 2 − X 2 ∂ τ X 1 ) d σ . On an arc write X = f τ τ X=f_\tau\tau X = f τ τ . The integrand is f τ 2 det ( τ , ∂ τ τ ) = κ f τ 2 f_\tau^2\det(\tau,\partial_\tau\tau)=\kappa f_\tau^2 f τ 2 det ( τ , ∂ τ τ ) = κ f τ 2 , with the sign convention fixed above. This proves (9) for these fields, without endpoint terms.
We justify density in the tangent H 1 H^1 H 1 space. First truncate radially: X Λ = X min ( 1 , Λ / ∣ X ∣ ) X^\Lambda=X\min(1,\Lambda/\lvert X\rvert) X Λ = X min ( 1 , Λ/ ∣ X ∣) for X ≠ 0 X\ne0 X = 0 , and X Λ = 0 X^\Lambda=0 X Λ = 0 where X = 0 X=0 X = 0 . The radial retraction is Lipschitz, ∣ ∇ X Λ ∣ ≤ ∣ ∇ X ∣ \lvert\nabla X^\Lambda\rvert\le\lvert\nabla X\rvert ∣ ∇ X Λ ∣ ≤ ∣ ∇ X ∣ , and X Λ → X X^\Lambda\to X X Λ → X in H 1 H^1 H 1 as Λ → ∞ \Lambda\to\infty Λ → ∞ . Its trace is still tangent. For fixed Λ \Lambda Λ , use smooth logarithmic cutoffs 0 ≤ η ε ≤ 1 0\le\eta_\varepsilon\le1 0 ≤ η ε ≤ 1 that vanish near the finite corner set, tend to one almost everywhere, and satisfy ∥ ∇ η ε ∥ 2 → 0 \lVert\nabla\eta_\varepsilon\rVert_2\to0 ∥ ∇ η ε ∥ 2 → 0 . Such cutoffs follow explicitly from the planar logarithmic radial profile: its squared Dirichlet norm between radii ε 2 \varepsilon^2 ε 2 and ε \varepsilon ε is 2 π / ∣ log ε ∣ 2\pi/\lvert\log\varepsilon\rvert 2 π / ∣ log ε ∣ , and smoothing preserves the vanishing limit. Then
∥ ∇ ( η ε X Λ − X Λ ) ∥ 2 ≤ ∥ ( 1 − η ε ) ∇ X Λ ∥ 2 + Λ ∥ ∇ η ε ∥ 2 ⟶ 0 , \lVert\nabla(\eta_\varepsilon X^\Lambda-X^\Lambda)\rVert_2\le\lVert(1-\eta_\varepsilon)\nabla X^\Lambda\rVert_2+\Lambda\lVert\nabla\eta_\varepsilon\rVert_2\longrightarrow0, ∥ ∇ ( η ε X Λ − X Λ ) ∥ 2 ≤ ∥( 1 − η ε ) ∇ X Λ ∥ 2 + Λ ∥ ∇ η ε ∥ 2 ⟶ 0 , and convergence in L 2 L^2 L 2 is immediate.
A field vanishing near the corners can now be smoothed while preserving tangency. Use a partition of unity by interior patches and tubular patches of open arcs. In a tubular patch with coordinates ( s , t ) (s,t) ( s , t ) , express the field as f τ ( s , t ) τ ( s ) + f ν ( s , t ) ν ( s ) f_\tau(s,t)\tau(s)+f_\nu(s,t)\nu(s) f τ ( s , t ) τ ( s ) + f ν ( s , t ) ν ( s ) . The normal coefficient has trace f ν ( s , 0 ) = 0 f_\nu(s,0)=0 f ν ( s , 0 ) = 0 . Extend f τ f_\tau f τ evenly and f ν f_\nu f ν oddly in t t t , mollify with an even kernel, and then return to the smooth moving frame. The resulting field is smooth and tangent on t = 0 t=0 t = 0 , and converges in H 1 H^1 H 1 under the smooth coordinate change. Supports can be kept inside the patches and away from the corners. Interior patches use ordinary mollification. Their sum proves the required density.
Both sides of (9) are continuous under H 1 H^1 H 1 convergence: for the boundary term use the boundedness of κ \kappa κ and the continuous trace map H 1 ( Ω ) → L 2 ( ∂ Ω ) H^1(\Omega)\to L^2(\partial\Omega) H 1 ( Ω ) → L 2 ( ∂ Ω ) . Thus the identity passes to every tangent H 1 H^1 H 1 field.
Remark 3.3. The turning-angle atoms of the geometric curvature measure are not inserted in (9) . An arbitrary H 1 H^1 H 1 field has no canonical point values at vertices, and the density proof determines the closed energy identity directly. For eigenfunction gradients the continuous corner values are in fact zero, but that stronger fact is not needed to prove this identity for all tangent fields.
Kernels for the singular conformal potential
Keep the previously fixed Φ \Phi Φ and a a a , and introduce the measure
d q = a d x , a = μ ∣ Φ ′ ∣ 2 . dq=a\,dx,\qquad a=\mu|\Phi'|^2. d q = a d x , a = μ ∣ Φ ′ ∣ 2 .
By (2) , a a a is locally bounded in D \mathbb{D} D and near every point of A A A , and has the integrable corner singularities in (4) ; q ( D ) = μ ∣ Ω ∣ < ∞ q(\mathbb{D})=\mu|\Omega|<\infty q ( D ) = μ ∣Ω∣ < ∞ . Conformal invariance of the Dirichlet integral and the Dirichlet inequality on Ω \Omega Ω give
∫ D ψ 2 d q ≤ d ∫ D ∣ ∇ ψ ∣ 2 ( ψ ∈ H 0 1 ( D ) ) , d = μ / λ D < 1. (10) \int_{\mathbb{D}}\psi^2\,dq\le d\int_{\mathbb{D}}|\nabla\psi|^2\qquad(\psi\in H_0^1(\mathbb{D})),\qquad d=\mu/\lambda_D<1. \tag*{(10)} ∫ D ψ 2 d q ≤ d ∫ D ∣∇ ψ ∣ 2 ( ψ ∈ H 0 1 ( D )) , d = μ / λ D < 1. (10)
We write T = T q T=T_q T = T q , K = K q K=K^q K = K q , R = R q R=R^q R = R q , N = N q N=N^q N = N q for the objects in (1) , and ⟨ f , h ⟩ q = ∫ D f h d q \langle f,h\rangle_q=\int_{\mathbb{D}}fh\,dq ⟨ f , h ⟩ q = ∫ D f h d q and ∥ f ∥ q = ⟨ f , f ⟩ q 1 / 2 \|f\|_q=\langle f,f\rangle_q^{1/2} ∥ f ∥ q = ⟨ f , f ⟩ q 1/2 . For any nonnegative measure ρ \rho ρ used below, the analogous inner product and norm on L 2 ( ρ ) L^2(\rho) L 2 ( ρ ) use d ρ d\rho d ρ instead of d q dq d q ; for vector functions the integrand is the Euclidean scalar product.
Lemma 4.1. For f ∈ L 2 ( q ) f\in L^2(q) f ∈ L 2 ( q ) there is a unique T f ∈ H 0 1 ( D ) Tf\in H_0^1(\mathbb{D}) T f ∈ H 0 1 ( D ) with ∫ D ∇ T f ⋅ ∇ ψ = ⟨ ψ , f ⟩ q \int_{\mathbb{D}}\nabla Tf\cdot\nabla\psi=\langle\psi,f\rangle_q ∫ D ∇ T f ⋅ ∇ ψ = ⟨ ψ , f ⟩ q for all ψ ∈ H 0 1 ( D ) \psi\in H_0^1(\mathbb{D}) ψ ∈ H 0 1 ( D ) . The operator T T T is self-adjoint on L 2 ( q ) L^2(q) L 2 ( q ) with ∥ T ∥ ≤ d \|T\|\le d ∥ T ∥ ≤ d , and T f ( x ) = ∫ D G ( x , z ) f ( z ) d q ( z ) Tf(x)=\int_{\mathbb{D}}G(x,z)f(z)\,dq(z) T f ( x ) = ∫ D G ( x , z ) f ( z ) d q ( z ) for almost every x x x ; for f ≥ 0 f\ge0 f ≥ 0 this is the kernel integral of (1) . The same holds for every measure 0 ≤ ρ ≤ q 0\le\rho\le q 0 ≤ ρ ≤ q , with the same d d d .
Proof. By (10) , ∣ ⟨ ψ , f ⟩ q ∣ ≤ d ∥ ∇ ψ ∥ 2 ∥ f ∥ q |\langle\psi,f\rangle_q|\le\sqrt{d}\|\nabla\psi\|_2\|f\|_q ∣ ⟨ ψ , f ⟩ q ∣ ≤ d ∥∇ ψ ∥ 2 ∥ f ∥ q , so the Riesz theorem gives T f Tf T f . Testing with ψ = T f \psi=Tf ψ = T f and using (10) again, ∥ T f ∥ q 2 ≤ d ∥ ∇ T f ∥ 2 2 = d ⟨ T f , f ⟩ q ≤ d ∥ T f ∥ q ∥ f ∥ q \|Tf\|_q^2\le d\|\nabla Tf\|_2^2=d\langle Tf,f\rangle_q\le d\|Tf\|_q\|f\|_q ∥ T f ∥ q 2 ≤ d ∥∇ T f ∥ 2 2 = d ⟨ T f , f ⟩ q ≤ d ∥ T f ∥ q ∥ f ∥ q . Symmetry is clear from the weak formulation. For f ≥ 0 f\ge0 f ≥ 0 bounded with compact support in D \mathbb{D} D the source a f af a f is bounded and the Green representation is classical; a general f ≥ 0 f\ge0 f ≥ 0 is the increasing limit of such functions, the potentials increase to the kernel integral, and they converge to T f Tf T f in H 0 1 H_0^1 H 0 1 .
In particular, for each 0 ≤ ρ ≤ q 0\le\rho\le q 0 ≤ ρ ≤ q the bounded inverse on L 2 ( ρ ) L^2(\rho) L 2 ( ρ ) is
( I − T ρ ) − 1 = ∑ n = 0 ∞ T ρ n , ∥ ( I − T ρ ) − 1 ∥ ≤ ( 1 − d ) − 1 . (I-T_\rho)^{-1}=\sum_{n=0}^{\infty}T_\rho^n,\qquad\|(I-T_\rho)^{-1}\|\le(1-d)^{-1}. ( I − T ρ ) − 1 = n = 0 ∑ ∞ T ρ n , ∥ ( I − T ρ ) − 1 ∥ ≤ ( 1 − d ) − 1 .
This follows from the preceding norm bound, before any solution formula uses the inverse. The integral definition of T ρ T_\rho T ρ initially applies to nonnegative functions; the lemma identifies its bounded linear extension to signed L 2 ( ρ ) L^2(\rho) L 2 ( ρ ) data.
The interior kernel is the Green function of Ω \Omega Ω
For x ≠ y x\ne y x = y in D \mathbb{D} D let M ( x , y ) = ∑ n ≥ 0 ( T n G ( ⋅ , y ) ) ( x ) ∈ ( 0 , ∞ ] M(x,y)=\sum_{n\ge0}(T^nG(\mathord\cdot,y))(x)\in(0,\infty] M ( x , y ) = ∑ n ≥ 0 ( T n G ( ⋅ , y )) ( x ) ∈ ( 0 , ∞ ] , the sum over all chains G ( x , z 1 ) G ( z 1 , z 2 ) ⋯ G ( z n , y ) G(x,z_1)G(z_1,z_2)\cdots G(z_n,y) G ( x , z 1 ) G ( z 1 , z 2 ) ⋯ G ( z n , y ) integrated against q ⊗ n q^{\otimes n} q ⊗ n , and
G q ( x , y ) = G ( Φ ( x ) , Φ ( y ) ) . G_q(x,y)=G(\Phi(x),\Phi(y)). G q ( x , y ) = G ( Φ ( x ) , Φ ( y )) .
Lemma 4.2. M = G q M=G_q M = G q off the diagonal. In particular M ρ ≤ G q M^\rho\le G_q M ρ ≤ G q for every measure 0 ≤ ρ ≤ q 0\le\rho\le q 0 ≤ ρ ≤ q , where M ρ M^\rho M ρ is the chain sum with ρ \rho ρ in place of q q q .
Proof. The Green function of the Laplacian is conformally invariant, G Ω 0 ( Φ ( x ) , Φ ( z ) ) = G ( x , z ) G_\Omega^0(\Phi(x),\Phi(z))=G(x,z) G Ω 0 ( Φ ( x ) , Φ ( z )) = G ( x , z ) , and d ζ = ∣ Φ ′ ( z ) ∣ 2 d z d\zeta=|\Phi'(z)|^2dz d ζ = ∣ Φ ′ ( z ) ∣ 2 d z . So (3) becomes
G q ( x , y ) = G ( x , y ) + ∫ D G ( x , z ) G q ( z , y ) d q ( z ) ( x ≠ y ) . (11) G_q(x,y)=G(x,y)+\int_{\mathbb{D}}G(x,z)G_q(z,y)\,dq(z)\qquad(x\ne y). \tag*{(11)} G q ( x , y ) = G ( x , y ) + ∫ D G ( x , z ) G q ( z , y ) d q ( z ) ( x = y ) . (11)
For each integer ℓ ≥ 0 \ell\ge0 ℓ ≥ 0 , ℓ + 1 \ell+1 ℓ + 1 substitutions of (11) give
G q ( x , y ) = ∑ n = 0 ℓ ( T n G ( ⋅ , y ) ) ( x ) + ⟨ G ( x , ⋅ ) , T ℓ G q ( ⋅ , y ) ⟩ q . G_q(x,y)=\sum_{n=0}^{\ell}(T^nG(\mathord\cdot,y))(x)+\langle G(x,\mathord\cdot),T^\ell G_q(\mathord\cdot,y)\rangle_q. G q ( x , y ) = n = 0 ∑ ℓ ( T n G ( ⋅ , y )) ( x ) + ⟨ G ( x , ⋅ ) , T ℓ G q ( ⋅ , y ) ⟩ q .
The remainder is nonnegative, so M ≤ G q M \le G_q M ≤ G q . Both G ( x , ⋅ ) G(x,\cdot) G ( x , ⋅ ) and G q ( ⋅ , y ) G_q(\cdot,y) G q ( ⋅ , y ) belong to L 2 ( q ) L^2(q) L 2 ( q ) : each has a logarithmic singularity at an interior point, where a a a is bounded, the first is bounded elsewhere, and ∥ G q ( ⋅ , y ) ∥ q 2 = μ ∥ G ( ⋅ , Φ ( y ) ) ∥ L 2 ( Ω ) 2 \lVert G_q(\cdot,y)\rVert_q^2=\mu\lVert G(\cdot,\Phi(y))\rVert_{L^2(\Omega)}^2 ∥ G q ( ⋅ , y ) ∥ q 2 = μ ∥ G ( ⋅ , Φ ( y )) ∥ L 2 ( Ω ) 2 . By Lemma 4.1 the remainder is at most d ℓ ∥ G ( x , ⋅ ) ∥ q ∥ G q ( ⋅ , y ) ∥ q → 0 d^\ell\lVert G(x,\cdot)\rVert_q\lVert G_q(\cdot,y)\rVert_q\to0 d ℓ ∥ G ( x , ⋅ ) ∥ q ∥ G q ( ⋅ , y ) ∥ q → 0 as ℓ → ∞ \ell\to\infty ℓ → ∞ .
The boundary kernels
Let P ( ξ , η ) = − ∂ ν ( η ) G ( ξ , η ) \mathcal{P}(\xi,\eta)=-\partial_{\nu(\eta)}\mathcal{G}(\xi,\eta) P ( ξ , η ) = − ∂ ν ( η ) G ( ξ , η ) for ξ ∈ Ω \xi\in\Omega ξ ∈ Ω , η ∈ Γ \eta\in\Gamma η ∈ Γ be the Poisson kernel of − Δ − μ -\Delta-\mu − Δ − μ on Ω \Omega Ω . For p ∈ D p\in\mathbb{D} p ∈ D and s ≠ t s\ne t s = t in A A A , put ξ s = Φ ( s ) \xi_s=\Phi(s) ξ s = Φ ( s ) , ξ t = Φ ( t ) \xi_t=\Phi(t) ξ t = Φ ( t ) and define
K q ( p , s ) = ∣ Φ ′ ( s ) ∣ P ( Φ ( p ) , ξ s ) , N q ( s , t ) = ∣ Φ ′ ( s ) ∣ ∣ Φ ′ ( t ) ∣ ∂ ν ( ξ s ) ∂ ν ( ξ t ) G ( ξ s , ξ t ) . K_q(p,s)=|\Phi'(s)|\mathcal{P}(\Phi(p),\xi_s),\qquad N_q(s,t)=|\Phi'(s)||\Phi'(t)|\partial_{\nu(\xi_s)}\partial_{\nu(\xi_t)}\mathcal{G}(\xi_s,\xi_t). K q ( p , s ) = ∣ Φ ′ ( s ) ∣ P ( Φ ( p ) , ξ s ) , N q ( s , t ) = ∣ Φ ′ ( s ) ∣∣ Φ ′ ( t ) ∣ ∂ ν ( ξ s ) ∂ ν ( ξ t ) G ( ξ s , ξ t ) .
Lemma 4.3. Let p ∈ D p\in\mathbb{D} p ∈ D and s ≠ t s\ne t s = t in A A A .
(a) K ( p , s ) = K q ( p , s ) K(p,s)=K_q(p,s) K ( p , s ) = K q ( p , s ) ; in particular 0 < K ( p , s ) < ∞ 0<K(p,s)<\infty 0 < K ( p , s ) < ∞ .
(b) sup s ∈ A K ( p , s ) < ∞ \sup_{s\in A}K(p,s)<\infty sup s ∈ A K ( p , s ) < ∞ .
(c) R ( s , t ) = ∫ D P s ( z ) K ( z , t ) d q ( z ) R(s,t)=\int_{\mathbb{D}}P_s(z)K(z,t)\,dq(z) R ( s , t ) = ∫ D P s ( z ) K ( z , t ) d q ( z ) and N 0 ( s , t ) ≤ N ( s , t ) ≤ N q ( s , t ) < ∞ N_0(s,t)\le N(s,t)\le N_q(s,t)<\infty N 0 ( s , t ) ≤ N ( s , t ) ≤ N q ( s , t ) < ∞ .
(d) R R R is symmetric and ∬ S 1 × S 1 R d s d t ≤ q ( D ) / ( 1 − d ) \iint_{S^1\times S^1}R\,ds\,dt\le q(\mathbb{D})/(1-d) ∬ S 1 × S 1 R d s d t ≤ q ( D ) / ( 1 − d ) .
Proof. (a) By Tonelli and Lemma 4.2 ,
K ( p , s ) = P s ( p ) + ∫ D M ( p , z ) P s ( z ) d q ( z ) = P s ( p ) + ∫ D G q ( p , z ) P s ( z ) d q ( z ) . (12) K(p,s)=P_s(p)+\int_{\mathbb{D}}M(p,z)P_s(z)\,dq(z)=P_s(p)+\int_{\mathbb{D}}G_q(p,z)P_s(z)\,dq(z). \tag*{(12)} K ( p , s ) = P s ( p ) + ∫ D M ( p , z ) P s ( z ) d q ( z ) = P s ( p ) + ∫ D G q ( p , z ) P s ( z ) d q ( z ) . (12)
Apply (11) , with the roles of the variables exchanged by symmetry, at y = r s y=rs y = r s , r ↑ 1 r\uparrow1 r ↑ 1 :
G q ( p , r s ) 1 − r = G ( p , r s ) 1 − r + ∫ D G q ( p , z ) G ( z , r s ) 1 − r d q ( z ) . \frac{G_q(p,rs)}{1-r}=\frac{G(p,rs)}{1-r}+\int_{\mathbb{D}}G_q(p,z)\frac{G(z,rs)}{1-r}\,dq(z). 1 − r G q ( p , r s ) = 1 − r G ( p , r s ) + ∫ D G q ( p , z ) 1 − r G ( z , r s ) d q ( z ) .
The explicit formula gives G ( z , r s ) / ( 1 − r ) → P s ( z ) G(z,rs)/(1-r)\to P_s(z) G ( z , r s ) / ( 1 − r ) → P s ( z ) for every z ∈ D z\in\mathbb{D} z ∈ D . On the left, Φ \Phi Φ is smooth and conformal up to s s s , and s Φ ′ ( s ) = ∣ Φ ′ ( s ) ∣ ν ( ξ s ) s\Phi'(s)=|\Phi'(s)|\nu(\xi_s) s Φ ′ ( s ) = ∣ Φ ′ ( s ) ∣ ν ( ξ s ) . Thus Φ ( r s ) = ξ s − ( 1 − r ) ∣ Φ ′ ( s ) ∣ ν ( ξ s ) + O ( ( 1 − r ) 2 ) \Phi(rs)=\xi_s-(1-r)|\Phi'(s)|\nu(\xi_s)+O((1-r)^2) Φ ( r s ) = ξ s − ( 1 − r ) ∣ Φ ′ ( s ) ∣ ν ( ξ s ) + O (( 1 − r ) 2 ) ; since G ( Φ ( p ) , ⋅ ) \mathcal{G}(\Phi(p),\cdot) G ( Φ ( p ) , ⋅ ) is smooth up to Γ \Gamma Γ and vanishes there, the left side tends to K q ( p , s ) K_q(p,s) K q ( p , s ) . Fatou's lemma and (4.3) give K ( p , s ) ≤ K q ( p , s ) < ∞ K(p,s)\le K_q(p,s)<\infty K ( p , s ) ≤ K q ( p , s ) < ∞ . For the equality let B r B_r B r be the disk with center r s rs r s and radius ( 1 − r ) / 2 (1-r)/2 ( 1 − r ) /2 . If z ∈ D ∖ B r z\in\mathbb{D}\setminus B_r z ∈ D ∖ B r then ∣ z − s ∣ ≤ 3 ∣ z − r s ∣ |z-s|\le3|z-rs| ∣ z − s ∣ ≤ 3∣ z − r s ∣ and
G ( z , r s ) 1 − r ≤ ( 1 − ∣ z ∣ 2 ) ( 1 + r ) 4 π ∣ z − r s ∣ 2 ≤ 9 P s ( z ) , \frac{G(z,rs)}{1-r}\le\frac{(1-|z|^2)(1+r)}{4\pi|z-rs|^2}\le9P_s(z), 1 − r G ( z , r s ) ≤ 4 π ∣ z − r s ∣ 2 ( 1 − ∣ z ∣ 2 ) ( 1 + r ) ≤ 9 P s ( z ) ,
and G q ( p , ⋅ ) P s G_q(p,\cdot)P_s G q ( p , ⋅ ) P s is q q q -integrable by what was just proved. If z ∈ B r z\in B_r z ∈ B r then G ( z , r s ) ≤ 1 4 π log ( 1 + 6 ( 1 − r ) 2 / ∣ z − r s ∣ 2 ) G(z,rs)\le\frac{1}{4\pi}\log(1+6(1-r)^2/|z-rs|^2) G ( z , r s ) ≤ 4 π 1 log ( 1 + 6 ( 1 − r ) 2 /∣ z − r s ∣ 2 ) , whose integral over B r B_r B r is O ( ( 1 − r ) 2 ) O((1-r)^2) O (( 1 − r ) 2 ) , while a a a and G q ( p , ⋅ ) G_q(p,\cdot) G q ( p , ⋅ ) are bounded near s s s ; this part contributes O ( 1 − r ) O(1-r) O ( 1 − r ) . Dominated convergence gives
K q ( p , s ) = P s ( p ) + ∫ D G q ( p , z ) P s ( z ) d q ( z ) = K ( p , s ) . K_q(p,s)=P_s(p)+\int_{\mathbb{D}}G_q(p,z)P_s(z)\,dq(z)=K(p,s). K q ( p , s ) = P s ( p ) + ∫ D G q ( p , z ) P s ( z ) d q ( z ) = K ( p , s ) .
(b) By (a), on A A A the kernel is the inward disk normal derivative of U p ( z ) = G ( Φ ( p ) , Φ ( z ) ) U_p(z)=\mathcal{G}(\Phi(p),\Phi(z)) U p ( z ) = G ( Φ ( p ) , Φ ( z )) :
K ( p , s ) = − ∂ ν D U p ( s ) . K(p,s)=-\partial_{\nu_{\mathbb{D}}}U_p(s). K ( p , s ) = − ∂ ν D U p ( s ) .
Corollary 2.4 bounds these derivatives uniformly, including as s s s approaches a prevertex. This also provides a continuous extension of K ( p , ⋅ ) ∣ A K(p,\cdot)|_A K ( p , ⋅ ) ∣ A to the circle, although only its boundedness on A A A is used. No continuity or finiteness of N N N at the prevertices is asserted.
(c) The first formula is Tonelli's theorem applied to (1) . Apply (11) at ( r s , r t ) (rs,rt) ( r s , r t ) :
G q ( r s , r t ) − G ( r s , r t ) ( 1 − r ) 2 = ∫ D G ( r s , z ) 1 − r ⋅ G q ( z , r t ) 1 − r d q ( z ) . \frac{G_q(rs,rt)-G(rs,rt)}{(1-r)^2}
=\int_{\mathbb{D}}\frac{G(rs,z)}{1-r}\cdot\frac{G_q(z,rt)}{1-r}\,dq(z). ( 1 − r ) 2 G q ( r s , r t ) − G ( r s , r t ) = ∫ D 1 − r G ( r s , z ) ⋅ 1 − r G q ( z , r t ) d q ( z ) .
By the computation in (a), the integrand tends to P s ( z ) K q ( z , t ) = P s ( z ) K ( z , t ) P_s(z)K_q(z,t)=P_s(z)K(z,t) P s ( z ) K q ( z , t ) = P s ( z ) K ( z , t ) for every z z z . On the left, G ( r s , r t ) / ( 1 − r ) 2 → N 0 ( s , t ) G(rs,rt)/(1-r)^2\to N_0(s,t) G ( r s , r t ) / ( 1 − r ) 2 → N 0 ( s , t ) , and by the joint smoothness in Lemma 2.1 (d) and the vanishing of G \mathcal{G} G on Γ \Gamma Γ in each variable, G q ( r s , r t ) / ( 1 − r ) 2 → N q ( s , t ) G_q(rs,rt)/(1-r)^2\to N_q(s,t) G q ( r s , r t ) / ( 1 − r ) 2 → N q ( s , t ) . Fatou's lemma gives R ≤ N q − N 0 R\le N_q-N_0 R ≤ N q − N 0 .
(d) Symmetry follows by reversing the chains. Since ∫ S 1 P s ( z ) d s = 1 \int_{S^1}P_s(z)\,ds=1 ∫ S 1 P s ( z ) d s = 1 , Tonelli gives ∬ R = ∑ n ⟨ 1 , T n 1 ⟩ q ≤ ∑ n d n ∥ 1 ∥ q 2 \iint R=\sum_n\langle1,T^n1\rangle_q\le\sum_n d^n\lVert1\rVert_q^2 ∬ R = ∑ n ⟨ 1 , T n 1 ⟩ q ≤ ∑ n d n ∥ 1 ∥ q 2 .
A global reciprocal-matrix flow
We now prove the compact-support kernel assertion needed in the singular limit. The initial regularization and the reciprocal-matrix strategy come from [14 Section 5] ; their proofs are included. The preservation step is proved through a differential equation and invariance of the negative semidefinite cone. This is an alternative proof of the finite resolvent assertion obtained by iterating the column-update lemma [14, Lemma 5.4]. No assertion from [14 Theorem 5.1] is assumed.
For real symmetric matrices, J ⪰ 0 J \succeq0 J ⪰ 0 and J ⪯ 0 J \preceq0 J ⪯ 0 mean positive and negative semidefiniteness. Write P ( J ) \mathrm{P}(J) P ( J ) when J J J has at most one positive eigenvalue, with multiplicity. This property is preserved by congruence, principal submatrices, subtraction of a positive semidefinite matrix, and limits in a fixed finite dimension. The notation M ∘ − 1 M^{\circ-1} M ∘− 1 means the entrywise reciprocal, not the matrix inverse. Matrix norms without a subscript are Euclidean operator norms. The entrywise product is denoted by ⊙ \odot ⊙ . If U ⪰ 0 U \succeq0 U ⪰ 0 , then U ⊙ U ⪰ 0 U \odot U \succeq0 U ⊙ U ⪰ 0 : represent U i j U_{ij} U ij by scalar products and take tensor products of the representing vectors.
Lemma 4.4 (global preservation). Let W = diag ( w 1 , … , w n ) W=\operatorname{diag}(w_1,\ldots,w_n) W = diag ( w 1 , … , w n ) with w i ≥ 0 w_i\geq0 w i ≥ 0 . Suppose M ( θ ) M(\theta) M ( θ ) is a real symmetric C 1 C^1 C 1 matrix on [ 0 , 1 ] [0,1] [ 0 , 1 ] , all its entries are strictly positive, and
d M d θ = M W M . \frac{\mathrm{d}M}{\mathrm{d}\theta}=MWM. d θ d M = M W M . If P ( M ( 0 ) ∘ − 1 ) \mathrm{P}(M(0)^{\circ-1}) P ( M ( 0 ) ∘− 1 ) holds, then P ( M ( θ ) ∘ − 1 ) \mathrm{P}(M(\theta)^{\circ-1}) P ( M ( θ ) ∘− 1 ) holds for every θ ∈ [ 0 , 1 ] \theta\in[0,1] θ ∈ [ 0 , 1 ] .
Proof. Put J = M ∘ − 1 J=M^{\circ-1} J = M ∘− 1 ; a dot denotes differentiation in θ \theta θ in this proof. Direct differentiation gives the locally Lipschitz vector field
J ˙ i j = − ∑ k = 1 n w k J i j 2 J i k J j k (13) \dot{J}_{ij}=-\sum_{k=1}^{n}w_k\frac{J_{ij}^{2}}{J_{ik}J_{jk}} \tag*{(13)} J ˙ ij = − k = 1 ∑ n w k J ik J j k J ij 2 (13)
on the open set of symmetric matrices with strictly positive entries.
First suppose P ( J ) \mathrm{P}(J) P ( J ) holds, and take v ∈ ker J v\in\ker J v ∈ ker J . For each k k k let
D k = diag ( J 1 k , … , J n k ) , A k = D k − 1 J D k − 1 , σ k = J k k − 1 . D_k=\operatorname{diag}(J_{1k},\ldots,J_{nk}),\qquad A_k=D_k^{-1}JD_k^{-1},\qquad\sigma_k=J_{kk}^{-1}. D k = diag ( J 1 k , … , J nk ) , A k = D k − 1 J D k − 1 , σ k = J k k − 1 .
The k k k -th row and column of A k A_k A k are constant, equal to σ k > 0 \sigma_k>0 σ k > 0 . The Schur complement at that entry, together with P ( A k ) \mathrm{P}(A_k) P ( A k ) , gives
U k = σ k 11 T − A k ⪰ 0 , U_k=\sigma_k 11^{T}-A_k\succeq0, U k = σ k 1 1 T − A k ⪰ 0 ,
where 1 1 1 is the vector of ones. More explicitly, the k k k -th row of U k U_k U k is zero and its remaining block is the negative of that Schur complement. Set y k = D k v y_k=D_kv y k = D k v . Then
1 T y k = ( J v ) k = 0 , A k y k = 0 , U k y k = 0. 1^{T}y_k=(Jv)_k=0,\qquad A_ky_k=0,\qquad U_ky_k=0. 1 T y k = ( J v ) k = 0 , A k y k = 0 , U k y k = 0.
Expanding A k ⊙ A k A_k\odot A_k A k ⊙ A k and using these three identities yields
v T J ˙ v = − ∑ k w k y k T ( U k ⊙ U k ) y k ≤ 0. (14) v^{T}\dot{J}v=-\sum_k w_k y_k^{T}(U_k\odot U_k)y_k\leq0. \tag*{(14)} v T J ˙ v = − k ∑ w k y k T ( U k ⊙ U k ) y k ≤ 0. (14)
This identity applies at every matrix in the specified open set having P \mathrm{P} P , not only at the initial matrix.
The passage from this sign condition at boundary points of the cone to forward invariance is the tangency criterion for flow-invariant closed sets of Nagumo [27 ] and Brezis [28 ] . We include the short argument for the cone at hand. The case n = 1 n=1 n = 1 is immediate. For n ≥ 2 n\geq2 n ≥ 2 , fix the last index as a pivot and write
J = ( J 00 b b T h ) , h > 0 , S = J 00 − b b T / h . J=\begin{pmatrix}J_{00}&b\\ b^{T}&h\end{pmatrix},\qquad h>0,\qquad S=J_{00}-bb^{T}/h. J = ( J 00 b T b h ) , h > 0 , S = J 00 − b b T / h .
Here the index 00 denotes the block of the first n − 1 n-1 n − 1 indices. Congruence gives J ∼ diag ( S , h ) J \sim\operatorname{diag}(S,h) J ∼ diag ( S , h ) , so P ( J ) P(J) P ( J ) is equivalent to S ⪯ 0 S \preceq0 S ⪯ 0 . The variables ( S , b , h ) (S,b,h) ( S , b , h ) are smooth coordinates on the open set under consideration. Let F ( S , b , h ) \mathcal{F}(S,b,h) F ( S , b , h ) be the S S S component of the vector field (13) . If S ⪯ 0 S \preceq0 S ⪯ 0 and x ∈ ker S x \in\ker S x ∈ ker S , the vector v = ( x , − b T x / h ) v=(x,-b^{T}x/h) v = ( x , − b T x / h ) belongs to ker J \ker J ker J , and differentiating the Schur complement gives
x T F ( S , b , h ) x = v T J ˙ v ≤ 0. x^{T}\mathcal{F}(S,b,h)x=v^{T}\dot{J}v\leq0. x T F ( S , b , h ) x = v T J ˙ v ≤ 0.
For a symmetric matrix S S S , let S + S_{+} S + retain its positive eigenvalues and replace the others by zero, and put S − = S − S + S_{-}=S-S_{+} S − = S − S + . Thus S − ⪯ 0 S_{-}\preceq0 S − ⪯ 0 , S + ⪰ 0 S_{+}\succeq0 S + ⪰ 0 and S − S + = 0 S_{-}S_{+}=0 S − S + = 0 . Near any point with S ⪯ 0 S\preceq0 S ⪯ 0 in our open coordinate domain, the projected point ( S − , b , h ) (S_{-},b,h) ( S − , b , h ) remains in that domain. The vector field is Lipschitz on a sufficiently small compact neighborhood containing these points. The squared distance to the negative semidefinite cone is ∥ S + ∥ F 2 \lVert S_{+}\rVert_{F}^{2} ∥ S + ∥ F 2 , where ∥ B ∥ F 2 = tr ( B T B ) \lVert B\rVert_{F}^{2}=\operatorname{tr}(B^{T}B) ∥ B ∥ F 2 = tr ( B T B ) is the Frobenius norm. Its derivative is
1 2 d d θ ∥ S + ∥ F 2 = tr ( S + F ( S , b , h ) ) . \frac{1}{2}\frac{d}{d\theta}\lVert S_{+}\rVert_{F}^{2}=\operatorname{tr}(S_{+}\mathcal{F}(S,b,h)). 2 1 d θ d ∥ S + ∥ F 2 = tr ( S + F ( S , b , h )) .
This derivative formula also holds at zero eigenvalues: the scalar function r ↦ ( max ( r , 0 ) ) 2 r\mapsto(\max(r,0))^{2} r ↦ ( max ( r , 0 ) ) 2 is C 1 C^{1} C 1 , or equivalently one may use the differentiability of squared distance to a closed convex set. The range of S + S_{+} S + is contained in ker S − \ker S_{-} ker S − , so the null-vector inequality at the projected point gives tr ( S + F ( S − , b , h ) ) ≤ 0 \operatorname{tr}(S_{+}\mathcal{F}(S_{-},b,h))\leq0 tr ( S + F ( S − , b , h )) ≤ 0 . Consequently, for a local Lipschitz constant L ∗ L_{*} L ∗ ,
1 2 d d θ ∥ S + ∥ F 2 ≤ L ∗ ∥ S + ∥ F 2 . \frac{1}{2}\frac{d}{d\theta}\lVert S_{+}\rVert_{F}^{2}\leq L_{*}\lVert S_{+}\rVert_{F}^{2}. 2 1 d θ d ∥ S + ∥ F 2 ≤ L ∗ ∥ S + ∥ F 2 .
Gronwall’s inequality shows that this distance remains zero if it starts at zero. This proves local forward invariance at every cone point in the open domain. A first-exit argument extends it along the whole trajectory: at any alleged first exit the trajectory is still in the open domain and the same local argument applies. Positivity and continuity of the entries on the given compact time interval ensure that this domain is never left. Thus S ⪯ 0 S\preceq0 S ⪯ 0 throughout, proving the lemma. □
Lemma 4.5 (a finite resolvent). Let C C C be real symmetric with strictly positive entries, P ( C ∘ − 1 ) P(C^{\circ-1}) P ( C ∘− 1 ) , and let W W W be diagonal with nonnegative entries. If ∥ W 1 / 2 C W 1 / 2 ∥ < 1 \lVert W^{1/2}CW^{1/2}\rVert<1 ∥ W 1/2 C W 1/2 ∥ < 1 , then the matrix
C + C W 1 / 2 ( I − W 1 / 2 C W 1 / 2 ) − 1 W 1 / 2 C C+CW^{1/2}(I-W^{1/2}CW^{1/2})^{-1}W^{1/2}C C + C W 1/2 ( I − W 1/2 C W 1/2 ) − 1 W 1/2 C has an entrywise reciprocal with property P P P .
Proof. Put B = W 1 / 2 C W 1 / 2 B=W^{1/2}CW^{1/2} B = W 1/2 C W 1/2 and
M ( θ ) = C + θ C W 1 / 2 ( I − θ B ) − 1 W 1 / 2 C , 0 ≤ θ ≤ 1. M(\theta)=C+\theta CW^{1/2}(I-\theta B)^{-1}W^{1/2}C,\qquad0\leq\theta\leq1. M ( θ ) = C + θ C W 1/2 ( I − θ B ) − 1 W 1/2 C , 0 ≤ θ ≤ 1.
The inverse exists on this whole interval. Its geometric series shows that M M M has strictly positive entries, since B B B has nonnegative entries. Differentiation, or the identity M ( θ ) W 1 / 2 = C W 1 / 2 ( I − θ B ) − 1 M(\theta)W^{1/2}=CW^{1/2}(I-\theta B)^{-1} M ( θ ) W 1/2 = C W 1/2 ( I − θ B ) − 1 , gives M ˙ = M W M \dot{M}=MWM M ˙ = M W M . Apply Lemma 4.4 at θ = 1 \theta=1 θ = 1 . Neither invertibility nor positive definiteness of C C C is needed. □
An initial regularization, with its proof
The following construction is adapted from [14 Section 5] . It is reproved to avoid taking its reciprocal-kernel assertion as an input.
Lemma 4.6 (regularization from [14 ] ). There are bounded continuous strictly positive kernels C η C_{\eta} C η on D × D \mathbb{D}\times\mathbb{D} D × D , 0 < η ≤ 1 0<\eta\leq1 0 < η ≤ 1 , such that C η ↑ G C_{\eta}\mathrel{\uparrow}G C η ↑ G as η ↓ 0 \eta\downarrow0 η ↓ 0 , including the infinite diagonal, and every finite entrywise reciprocal matrix of C η C_{\eta} C η has property P P P .
Proof. First define
L ( r ) = 1 log ( 1 + 1 / r ) ( r > 0 ) , L ( 0 ) = 0. \mathcal{L}(r)=\frac{1}{\log(1+1/r)}\quad(r>0),\qquad\mathcal{L}(0)=0. L ( r ) = log ( 1 + 1/ r ) 1 ( r > 0 ) , L ( 0 ) = 0.
For x , y ∈ D x,y\in\mathbb{D} x , y ∈ D and r ≥ 0 r\geq0 r ≥ 0 put
h ( x , y ) = ∣ x − y ∣ 2 ( 1 − ∣ x ∣ 2 ) ( 1 − ∣ y ∣ 2 ) , f ∗ ( r ) = r + 1 2 − L ( r ) . h(x,y)=\frac{|x-y|^2}{(1-|x|^2)(1-|y|^2)},\qquad f_*(r)=r+\frac{1}{2}-\mathcal{L}(r). h ( x , y ) = ( 1 − ∣ x ∣ 2 ) ( 1 − ∣ y ∣ 2 ) ∣ x − y ∣ 2 , f ∗ ( r ) = r + 2 1 − L ( r ) .
Positive diagonal congruence with entries 1 − ∣ x ∣ 2 1-|x|^2 1 − ∣ x ∣ 2 turns the kernel h + 1 2 h+\frac{1}{2} h + 2 1 into 1 2 ( 1 + ∣ x ∣ 2 ) ( 1 + ∣ y ∣ 2 ) − 2 x ⋅ y \frac{1}{2}(1+|x|^2)(1+|y|^2)-2x\cdot y 2 1 ( 1 + ∣ x ∣ 2 ) ( 1 + ∣ y ∣ 2 ) − 2 x ⋅ y . This is a positive rank-one kernel minus a positive semidefinite kernel and hence has P on every finite set.
We next show that f ∗ ( h ) f_*(h) f ∗ ( h ) is positive semidefinite. For 0 < α < 1 0<\alpha<1 0 < α < 1 , normalize v α − 1 ( 1 + v ) − 1 d v v^{\alpha-1}(1+v)^{-1}\,dv v α − 1 ( 1 + v ) − 1 d v to a probability measure on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) and denote expectation by E α \mathbb{E}_{\alpha} E α . Write r v = ( 1 + v ) − 1 r_v=(1+v)^{-1} r v = ( 1 + v ) − 1 . Integration by parts gives E α ( 1 − r v ) = α \mathbb{E}_{\alpha}(1-r_v)=\alpha E α ( 1 − r v ) = α . The elementary identities
L ( r ) = ∫ 0 1 r 1 − α ( r + 1 ) α d α , r 1 − α ( r + 1 ) α = E α r ( r + 1 ) r + r v \mathcal{L}(r)=\int_0^1 r^{1-\alpha}(r+1)^\alpha\,d\alpha,\qquad r^{1-\alpha}(r+1)^\alpha=\mathbb{E}_{\alpha}\frac{r(r+1)}{r+r_v} L ( r ) = ∫ 0 1 r 1 − α ( r + 1 ) α d α , r 1 − α ( r + 1 ) α = E α r + r v r ( r + 1 )
hold for r > 0 r>0 r > 0 . The second follows by scaling v v v in the defining integral: the normalization cancels ∫ 0 ∞ v α − 1 / ( 1 + v ) d v \int_0^\infty v^{\alpha-1}/(1+v)\,dv ∫ 0 ∞ v α − 1 / ( 1 + v ) d v . Polynomial division then gives
f ∗ ( r ) = ∫ 0 1 E α r v ( 1 − r v ) r + r v d α , 0 ≤ f ∗ ( r ) ≤ f ∗ ( 0 ) = 1 2 . f_*(r)=\int_0^1\mathbb{E}_{\alpha}\frac{r_v(1-r_v)}{r+r_v}\,d\alpha,\qquad0\leq f_*(r)\leq f_*(0)=\frac{1}{2}. f ∗ ( r ) = ∫ 0 1 E α r + r v r v ( 1 − r v ) d α , 0 ≤ f ∗ ( r ) ≤ f ∗ ( 0 ) = 2 1 .
The formula extends to r = 0 r=0 r = 0 by dominated convergence. Now
k ∗ ( x , y ) = 1 1 + h ( x , y ) = ( 1 − ∣ x ∣ 2 ) ( 1 − ∣ y ∣ 2 ) ∣ 1 − x y ‾ ∣ 2 k_*(x,y)=\frac{1}{1+h(x,y)}=\frac{(1-|x|^2)(1-|y|^2)}{|1-x\overline{y}|^2} k ∗ ( x , y ) = 1 + h ( x , y ) 1 = ∣1 − x y ∣ 2 ( 1 − ∣ x ∣ 2 ) ( 1 − ∣ y ∣ 2 )
is positive semidefinite: the geometric power series makes ( 1 − x y ‾ ) − 1 (1-x\overline{y})^{-1} ( 1 − x y ) − 1 a complex Gram kernel, and multiplication by its complex conjugate and the diagonal factors preserves that property. For every 0 < r v < 1 0<r_v<1 0 < r v < 1 ,
1 h + r v = ∑ j = 0 ∞ ( 1 − r v ) j k ∗ j + 1 \frac{1}{h+r_v}=\sum_{j=0}^{\infty}(1-r_v)^j k_*^{j+1} h + r v 1 = j = 0 ∑ ∞ ( 1 − r v ) j k ∗ j + 1
is a sum of positive semidefinite kernels, where powers are entrywise. It converges also on the diagonal. The positive mixture above proves positive semidefiniteness of f ∗ ( h ) f_*(h) f ∗ ( h ) .
Define
1 C η ( x , y ) = 4 π [ h + 1 2 − ( 1 − η ) f ∗ ( h ) ] = 4 π [ L ( h ) + η f ∗ ( h ) ] . \frac{1}{C_\eta(x,y)}=4\pi\left[h+\frac{1}{2}-(1-\eta)f_*(h)\right]=4\pi\left[\mathcal{L}(h)+\eta f_*(h)\right]. C η ( x , y ) 1 = 4 π [ h + 2 1 − ( 1 − η ) f ∗ ( h ) ] = 4 π [ L ( h ) + η f ∗ ( h ) ] .
The first expression has P. The bracket is at least η / 2 \eta/2 η /2 , since it equals η ( h + 1 2 ) + ( 1 − η ) L ( h ) \eta\left(h+\frac{1}{2}\right)+(1-\eta)\mathcal{L}(h) η ( h + 2 1 ) + ( 1 − η ) L ( h ) . Thus C η C_\eta C η is bounded, continuous and strictly positive. Since f ∗ ≥ 0 f_*\geq0 f ∗ ≥ 0 , it increases to G G G as η ↓ 0 \eta\downarrow0 η ↓ 0 , and C η ( x , x ) = 1 / ( 2 π η ) C_\eta(x,x)=1/(2\pi\eta) C η ( x , x ) = 1/ ( 2 π η ) . □
From the flow to the compact-support boundary theorem
Proposition 4.7 (compact-support reciprocal theorem). Let d ρ = a ρ d x d\rho=a_\rho\,dx d ρ = a ρ d x , where a ρ ≥ 0 a_\rho\geq0 a ρ ≥ 0 is bounded and supported in a compact subset of D \mathbb{D} D . Suppose
∫ D ψ 2 d ρ ≤ d ρ ∫ D ∣ ∇ ψ ∣ 2 ( ψ ∈ H 0 1 ( D ) ) , 0 < d ρ < 1. \int_{\mathbb{D}}\psi^2\,d\rho\leq d_\rho\int_{\mathbb{D}}|\nabla\psi|^2\quad(\psi\in H_0^1(\mathbb{D})),\qquad0<d_\rho<1. ∫ D ψ 2 d ρ ≤ d ρ ∫ D ∣∇ ψ ∣ 2 ( ψ ∈ H 0 1 ( D )) , 0 < d ρ < 1. For p ∈ D p\in\mathbb{D} p ∈ D define
D p ρ ( s , t ) = K ρ ( p , s ) K ρ ( p , t ) N ρ ( s , t ) ( s ≠ t ) , D p ρ ( s , s ) = 0. D_p^\rho(s,t)=\frac{K^\rho(p,s)K^\rho(p,t)}{N^\rho(s,t)}\quad(s\neq t),\qquad D_p^\rho(s,s)=0. D p ρ ( s , t ) = N ρ ( s , t ) K ρ ( p , s ) K ρ ( p , t ) ( s = t ) , D p ρ ( s , s ) = 0. Then D p ρ D_p^\rho D p ρ is conditionally negative semidefinite on S 1 \mathbb{S}^1 S 1 .
Proof. The proof of Lemma 4.1, with this form inequality in place of (10) , gives a self-adjoint T ρ T_\rho T ρ on L 2 ( ρ ) L^2(\rho) L 2 ( ρ ) with ∥ T ρ ∥ ≤ d ρ < 1 \lVert T_\rho\rVert\le d_\rho< 1 ∥ T ρ ∥ ≤ d ρ < 1 . For distinct interior points define
M ρ ( x , y ) = G ( x , y ) + ⟨ G ( x , ⋅ ) , ( I − T ρ ) − 1 G ( y , ⋅ ) ⟩ ρ . M^\rho(x,y)=G(x,y)+\langle G(x,\cdot),(I-T_\rho)^{-1}G(y,\cdot)\rangle_\rho. M ρ ( x , y ) = G ( x , y ) + ⟨ G ( x , ⋅ ) , ( I − T ρ ) − 1 G ( y , ⋅ ) ⟩ ρ .
Its correction is finite: logarithmic singularities belong to L 2 ( ρ ) L^2(\rho) L 2 ( ρ ) . Set M ρ ( x , x ) = + ∞ M^\rho(x,x)=+\infty M ρ ( x , x ) = + ∞ . We first prove P for its finite reciprocal matrices with zero diagonal.
Fix 0 < η ≤ 1 0<\eta\le1 0 < η ≤ 1 and replace G G G by C η C_\eta C η from Lemma 4.6. The resulting integral operator T η T_\eta T η satisfies ∣ T η f ∣ ≤ T ρ ∣ f ∣ \lvert T_\eta f\rvert\le T_\rho\lvert f\rvert ∣ T η f ∣ ≤ T ρ ∣ f ∣ , and hence ∥ T η ∥ ≤ d ρ \lVert T_\eta\rVert\le d_\rho ∥ T η ∥ ≤ d ρ . This estimate does not assume positive semidefiniteness of T η T_\eta T η . Choose a compact integration set Q ρ ⋐ D Q_\rho\Subset D Q ρ ⋐ D containing the support and partition it into finitely many measurable cells whose maximal diameter tends to zero. Choose a representative z α z_\alpha z α in every nonempty cell and let w α w_\alpha w α be its mass. Cells of zero ρ \rho ρ -mass are omitted only from the finite matrices and normalized indicators below. Let Π c e l l \Pi_{\mathrm{cell}} Π cell map each cell to its representative. Uniform continuity on this compact set gives
∥ C η ( Π c e l l ( ⋅ ) , Π c e l l ( ⋅ ) ) − C η ∥ ∞ ⟶ 0. \lVert C_\eta(\Pi_{\mathrm{cell}}(\cdot),\Pi_{\mathrm{cell}}(\cdot))-C_\eta\rVert_\infty\longrightarrow0. ∥ C η ( Π cell ( ⋅ ) , Π cell ( ⋅ )) − C η ∥ ∞ ⟶ 0.
The corresponding operators on L 2 ( ρ ) L^2(\rho) L 2 ( ρ ) therefore converge to T η T_\eta T η in norm, with error bounded by ρ ( D ) \rho(D) ρ ( D ) times that uniform error. Their restrictions to the span of the normalized cell indicators are the symmetric matrices B α β = w α C η ( z α , z β ) w β B_{\alpha\beta}=\sqrt{w_\alpha}C_\eta(z_\alpha,z_\beta)\sqrt{w_\beta} B α β = w α C η ( z α , z β ) w β ; they vanish on the orthogonal complement. In particular ∥ B ∥ < 1 \lVert B\rVert<1 ∥ B ∥ < 1 for all sufficiently fine partitions. More precisely, their operator norms are eventually at most ( 1 + d ρ ) / 2 (1+d_\rho)/2 ( 1 + d ρ ) /2 , so their resolvent norms are at most 2 / ( 1 − d ρ ) 2/(1-d_\rho) 2/ ( 1 − d ρ ) .
Fix any finite set of evaluation points and join it to the integration nodes, merging coincident points. Give evaluation-only points weight zero. The matrix C C C of C η C_\eta C η on this union has strictly positive entries and P ( C − 1 ) P(C^{-1}) P ( C − 1 ) by Lemma 4.6. After removing the rows and columns belonging to evaluation-only points, the remaining block of W 1 / 2 C W 1 / 2 W^{1/2}CW^{1/2} W 1/2 C W 1/2 is B B B ; the removed rows and columns are zero. Lemma 4.5 therefore proves P for the reciprocal of the discrete resolvent, on the whole union and then on the evaluation points.
For fixed x x x , the source C η ( x , Π c e l l ( ⋅ ) ) C_\eta(x,\Pi_{\mathrm{cell}}(\cdot)) C η ( x , Π cell ( ⋅ )) converges uniformly on the integration set, hence in L 2 ( ρ ) L^2(\rho) L 2 ( ρ ) , to C η ( x , ⋅ ) C_\eta(x,\cdot) C η ( x , ⋅ ) . Norm convergence of the operators, their common strict norm bound for fine partitions, and the identity for the difference of inverses show that the discrete resolvent entries converge to
M η ρ ( x , y ) = C η ( x , y ) + ⟨ C η ( x , ⋅ ) , ( I − T η ) − 1 C η ( y , ⋅ ) ⟩ ρ . M^\rho_\eta(x,y)=C_\eta(x,y)+\langle C_\eta(x,\cdot),(I-T_\eta)^{-1}C_\eta(y,\cdot)\rangle_\rho. M η ρ ( x , y ) = C η ( x , y ) + ⟨ C η ( x , ⋅ ) , ( I − T η ) − 1 C η ( y , ⋅ ) ⟩ ρ .
Thus its reciprocal matrices have P. These partitions are an analytic limiting argument; no finite computed partition is used as evidence.
Let η ↓ 0 \eta\downarrow0 η ↓ 0 . The geometric resolvent series is the sum of nonnegative chain integrals. Monotone convergence in each product integral and then in the series gives M η ρ ( x , y ) ↑ M ρ ( x , y ) M^\rho_\eta(x,y)\uparrow M^\rho(x,y) M η ρ ( x , y ) ↑ M ρ ( x , y ) for x ≠ y x\ne y x = y . The correction in the limit is bounded in absolute value by
∥ G ( x , ⋅ ) ∥ L 2 ( ρ ) ∥ G ( y , ⋅ ) ∥ L 2 ( ρ ) 1 − d ρ . \frac{\lVert G(x,\cdot)\rVert_{L^2(\rho)}\lVert G(y,\cdot)\rVert_{L^2(\rho)}}{1-d_\rho}. 1 − d ρ ∥ G ( x , ⋅ ) ∥ L 2 ( ρ ) ∥ G ( y , ⋅ ) ∥ L 2 ( ρ ) .
On the diagonal the direct term tends to infinity. Taking reciprocal entries and fixed-size limits proves P for M ρ M^\rho M ρ with reciprocal diagonal zero. If ρ = 0 \rho=0 ρ = 0 , the same conclusion follows directly from the regularization without partitions.
On the compact integration support the Poisson sections are bounded, so Tonelli’s theorem and the resolvent series give
K ρ ( p , s ) = P s ( p ) + ⟨ G ( p , ⋅ ) , ( I − T ρ ) − 1 P s ⟩ ρ , K^\rho(p,s)=P_s(p)+\langle G(p,\cdot),(I-T_\rho)^{-1}P_s\rangle_\rho, K ρ ( p , s ) = P s ( p ) + ⟨ G ( p , ⋅ ) , ( I − T ρ ) − 1 P s ⟩ ρ ,
N ρ ( s , t ) = N 0 ( s , t ) + ⟨ P s , ( I − T ρ ) − 1 P t ⟩ ρ ( s ≠ t ) . N^\rho(s,t)=N_0(s,t)+\langle P_s,(I-T_\rho)^{-1}P_t\rangle_\rho\qquad(s\ne t). N ρ ( s , t ) = N 0 ( s , t ) + ⟨ P s , ( I − T ρ ) − 1 P t ⟩ ρ ( s = t ) .
Now fix an integer n b ≥ 1 n_b \ge1 n b ≥ 1 , take distinct t 1 , … , t n b ∈ S 1 t_1,\ldots,t_{n_b} \in\mathbb{S}^1 t 1 , … , t n b ∈ S 1 , keep p ∈ D p \in\mathbb{D} p ∈ D fixed, and let r ↑ 1 r \uparrow1 r ↑ 1 with r > ∣ p ∣ r > |p| r > ∣ p ∣ . The explicit Green formula gives
G ( r s , x ) 1 − r ⟶ P s ( x ) , G ( r s , r t ) ( 1 − r ) 2 ⟶ N 0 ( s , t ) ( s ≠ t ) . \frac{G(rs,x)}{1-r} \longrightarrow P_s(x), \qquad\frac{G(rs,rt)}{(1-r)^2} \longrightarrow N_0(s,t) \quad(s \ne t). 1 − r G ( r s , x ) ⟶ P s ( x ) , ( 1 − r ) 2 G ( r s , r t ) ⟶ N 0 ( s , t ) ( s = t ) .
The first convergence is uniform on the compact integration support. Using the bounded resolvent and the L 2 ( ρ ) L^2(\rho) L 2 ( ρ ) source at p p p gives
M ρ ( p , r t i ) 1 − r ⟶ K ρ ( p , t i ) , M ρ ( r t i , r t j ) ( 1 − r ) 2 ⟶ N ρ ( t i , t j ) ( i ≠ j ) . \frac{M^\rho(p,rt_i)}{1-r} \longrightarrow K^\rho(p,t_i), \qquad\frac{M^\rho(rt_i,rt_j)}{(1-r)^2} \longrightarrow N^\rho(t_i,t_j) \quad(i \ne j). 1 − r M ρ ( p , r t i ) ⟶ K ρ ( p , t i ) , ( 1 − r ) 2 M ρ ( r t i , r t j ) ⟶ N ρ ( t i , t j ) ( i = j ) .
All limiting entries are finite and positive: the resolvent expressions use bounded Poisson sections on the support. Congruence of the reciprocal matrix by diagonal factors 1 − r 1-r 1 − r at moving points and 1 1 1 at p p p , followed by the limit, gives P \mathrm{P} P for
J = ( B ℓ ℓ T 0 ) , B i i = 0 , B i j = 1 N ρ ( t i , t j ) ( i ≠ j ) , ℓ i = 1 K ρ ( p , t i ) . \mathrm{J}=\begin{pmatrix} B & \ell\\ \ell^T & 0 \end{pmatrix}, \qquad B_{ii}=0,\qquad B_{ij}=\frac{1}{N^\rho(t_i,t_j)}\ (i\ne j),\qquad\ell_i=\frac{1}{K^\rho(p,t_i)}. J = ( B ℓ T ℓ 0 ) , B ii = 0 , B ij = N ρ ( t i , t j ) 1 ( i = j ) , ℓ i = K ρ ( p , t i ) 1 .
For completeness, this implies x T B x ≤ 0 x^T Bx \le0 x T B x ≤ 0 when ℓ T x = 0 \ell^T x=0 ℓ T x = 0 . Otherwise choose y y y with ℓ T y ≠ 0 \ell^T y \ne0 ℓ T y = 0 and subtract ( x T B y ) / ( x T B x ) (x^TBy)/(x^TBx) ( x T B y ) / ( x T B x ) times x x x to arrange x T B y = 0 x^TBy=0 x T B y = 0 . Then ( x , 0 ) (x,0) ( x , 0 ) and ( y , c ) (y,c) ( y , c ) are orthogonal for the form J \mathrm{J} J . Choosing c c c so that y T B y + 2 c ℓ T y > 0 y^TBy+2c\ell^T y>0 y T B y + 2 c ℓ T y > 0 produces a two-dimensional positive subspace, contrary to P \mathrm{P} P . Finally set x i = K ρ ( p , t i ) ω i x_i=K^\rho(p,t_i)\omega_i x i = K ρ ( p , t i ) ω i for coefficients with ∑ i ω i = 0 \sum_i\omega_i=0 ∑ i ω i = 0 . The resulting inequality is exactly the required conditional negativity. Repeated boundary points are handled by combining their coefficients. □
The proof controls every fixed finite configuration at full subcritical amplitude: the only strict bound used is the operator bound d ρ < 1 d_\rho<1 d ρ < 1 , before the flow and in the operator limits.
Conditional negativity for the singular potential
Proposition 4.8. For every p ∈ D p \in\mathbb{D} p ∈ D the kernel
D p ( s , t ) = K ( p , s ) K ( p , t ) N ( s , t ) ( s ≠ t ) , D p ( s , s ) = 0 , D_p(s,t)=\frac{K(p,s)K(p,t)}{N(s,t)}\quad(s\ne t),\qquad D_p(s,s)=0, D p ( s , t ) = N ( s , t ) K ( p , s ) K ( p , t ) ( s = t ) , D p ( s , s ) = 0 , is conditionally negative semidefinite on A A A . Moreover, with C p = sup A K ( p , ⋅ ) C_p=\sup_A K(p,\cdot) C p = sup A K ( p , ⋅ ) ,
0 < D p ( s , t ) ≤ π C p 2 ∣ s − t ∣ 2 ( s ≠ t in A ) . (15) 0<D_p(s,t)\le\pi C_p^2|s-t|^2\quad(s\ne t\text{ in }A). \tag*{(15)} 0 < D p ( s , t ) ≤ π C p 2 ∣ s − t ∣ 2 ( s = t in A ) . (15)
Proof. Let r n ↑ 1 r_n \uparrow1 r n ↑ 1 and ρ n = 1 { ∣ x ∣ ≤ r n } q \rho_n=\mathbf{1}_{\{|x|\le r_n\}}q ρ n = 1 { ∣ x ∣ ≤ r n } q . Each ρ n \rho_n ρ n has a bounded density and satisfies (10) with the same d d d . The finite-valued assertions for these truncations require no additional theorem from [14 ] : on their compact support the Poisson functions P s P_s P s are uniformly bounded for s ∈ S 1 s\in\mathbb{S}^1 s ∈ S 1 . The formulas
K ρ n ( p , s ) = P s ( p ) + ⟨ G ( p , ⋅ ) , ( I − T ρ n ) − 1 P s ⟩ ρ n , R ρ n ( s , t ) = ⟨ P s , ( I − T ρ n ) − 1 P t ⟩ ρ n K^{\rho_n}(p,s)=P_s(p)+\langle G(p,\cdot),(I-T_{\rho_n})^{-1}P_s\rangle_{\rho_n},\qquad R^{\rho_n}(s,t)=\langle P_s,(I-T_{\rho_n})^{-1}P_t\rangle_{\rho_n} K ρ n ( p , s ) = P s ( p ) + ⟨ G ( p , ⋅ ) , ( I − T ρ n ) − 1 P s ⟩ ρ n , R ρ n ( s , t ) = ⟨ P s , ( I − T ρ n ) − 1 P t ⟩ ρ n
are finite by Lemma 4.1 ; G ( p , ⋅ ) ∈ L 2 ( ρ n ) G(p,\cdot)\in L^2(\rho_n) G ( p , ⋅ ) ∈ L 2 ( ρ n ) even if the pole lies in the support. Thus Proposition 4.7 applies and gives conditional negative semidefiniteness of D p ρ n D_p^{\rho_n} D p ρ n on S 1 \mathbb{S}^1 S 1 . Every term in (1) is an integral of a nonnegative function against a power of the measure, so by monotone convergence K ρ n ( p , s ) ↑ K ( p , s ) K^{\rho_n}(p,s)\uparrow K(p,s) K ρ n ( p , s ) ↑ K ( p , s ) and N ρ n ( s , t ) ↑ N ( s , t ) N^{\rho_n}(s,t)\uparrow N(s,t) N ρ n ( s , t ) ↑ N ( s , t ) for all s ≠ t s\ne t s = t . On A A A the limits are finite and positive by Lemma 4.3, hence D p ρ n ( t i , t j ) → D p ( t i , t j ) D_p^{\rho_n}(t_i,t_j)\to D_p(t_i,t_j) D p ρ n ( t i , t j ) → D p ( t i , t j ) for any finite family ( t i ) (t_i) ( t i ) in A A A , and the inequality passes to the limit. The bound (15) follows from N ≥ N 0 N\ge N_0 N ≥ N 0 and Lemma 4.3(b). □
Solution formulas
For a bounded Borel function h h h on S 1 \mathbb{S}^{1} S 1 let H h ( x ) = ∫ S 1 P s ( x ) h ( s ) d s \operatorname{H}h(x)=\int_{\mathbb{S}^{1}}P_s(x)h(s)\,ds H h ( x ) = ∫ S 1 P s ( x ) h ( s ) d s be its bounded harmonic extension. Since q q q is finite, H h ∈ L 2 ( q ) \operatorname{H}h\in L^{2}(q) H h ∈ L 2 ( q ) , and we define
W h = H h + T ( I − T ) − 1 H h . (16) W_h=\operatorname{H}h+T(I-T)^{-1}\operatorname{H}h. \tag*{(16)} W h = H h + T ( I − T ) − 1 H h . (16)
Lemma 4.9. Let h , k h,k h , k be bounded Borel functions on S 1 \mathbb{S}^{1} S 1 .
(a) ζ h = W h − H h ∈ H 0 1 ( D ) \zeta_h=W_h-\operatorname{H}h\in H^1_0(\mathbb{D}) ζ h = W h − H h ∈ H 0 1 ( D ) and ∫ D ∇ ζ h ⋅ ∇ ψ = ⟨ ψ , W h ⟩ q \int_{\mathbb{D}}\nabla\zeta_h\cdot\nabla\psi=\langle\psi,W_h\rangle_q ∫ D ∇ ζ h ⋅ ∇ ψ = ⟨ ψ , W h ⟩ q for all ψ ∈ H 0 1 ( D ) \psi\in H^1_0(\mathbb{D}) ψ ∈ H 0 1 ( D ) .
(b) W h ( p ) = ∫ S 1 K ( p , s ) h ( s ) d s W_h(p)=\int_{\mathbb{S}^{1}}K(p,s)h(s)\,ds W h ( p ) = ∫ S 1 K ( p , s ) h ( s ) d s for every p ∈ D p\in\mathbb{D} p ∈ D (continuous representative).
(c) ⟨ H h , ( I − T ) − 1 H k ⟩ q = ∫ S 1 × S 1 R ( s , t ) h ( s ) k ( t ) d s d t \langle\operatorname{H}h,(I-T)^{-1}\operatorname{H}k\rangle_q=\int_{\mathbb{S}^{1}\times\mathbb{S}^{1}}R(s,t)h(s)k(t)\,ds\,dt ⟨ H h , ( I − T ) − 1 H k ⟩ q = ∫ S 1 × S 1 R ( s , t ) h ( s ) k ( t ) d s d t .
Proof. (a) holds because W h = ( I − T ) − 1 H h W_h=(I-T)^{-1}\operatorname{H}h W h = ( I − T ) − 1 H h in L 2 ( q ) L^2(q) L 2 ( q ) and ζ h = T ( I − T ) − 1 H h \zeta_h=T(I-T)^{-1}\operatorname{H}h ζ h = T ( I − T ) − 1 H h . For (b) and (c) expand ( I − T ) − 1 = ∑ n T n (I-T)^{-1}=\sum_nT^n ( I − T ) − 1 = ∑ n T n , which converges in operator norm, write T n H h = ∫ S 1 ( T n P s ) h ( s ) d s T^n\operatorname{H}h=\int_{\mathbb{S}^{1}}(T^nP_s)h(s)\,ds T n H h = ∫ S 1 ( T n P s ) h ( s ) d s and exchange sums and integrals. For ∣ h ∣ |h| ∣ h ∣ , ∣ k ∣ |k| ∣ k ∣ all terms are nonnegative and the totals are bounded by ∥ h ∥ ∞ ∫ K ( p , s ) d s \lVert h\rVert_\infty\int K(p,s)\,ds ∥ h ∥ ∞ ∫ K ( p , s ) d s and ∥ h ∥ ∞ ∥ k ∥ ∞ ∬ R \lVert h\rVert_\infty\lVert k\rVert_\infty\iint R ∥ h ∥ ∞ ∥ k ∥ ∞ ∬ R , finite by Lemma 4.3; this justifies the exchanges for signed data. Pointwise evaluation at p p p uses ζ h ( p ) = ⟨ G ( p , ⋅ ) , ( I − T ) − 1 H h ⟩ q \zeta_h(p)=\langle G(p,\cdot),(I-T)^{-1}\operatorname{H}h\rangle_q ζ h ( p ) = ⟨ G ( p , ⋅ ) , ( I − T ) − 1 H h ⟩ q . Here p ↦ G ( p , ⋅ ) p\mapsto G(p,\cdot) p ↦ G ( p , ⋅ ) is continuous from D \mathbb{D} D into L 2 ( q ) L^2(q) L 2 ( q ) (the potential is bounded near p p p , and locally the logarithmic singularity is continuous under translation in L 2 L^2 L 2 , and away from p p p dominated convergence applies because the sections are uniformly bounded while q q q is finite), so this formula is the continuous representative of ζ h \zeta_h ζ h and (b) holds at every point, not only almost everywhere. □
Boundary energy and the multiplier identity
Definition 5.1. A \mathcal{A} A is the vector space of bounded Borel maps h : S 1 → R 2 h:\mathbb{S}^{1}\to\mathbb{R}^{2} h : S 1 → R 2 that belong to H 1 / 2 ( S 1 ; R 2 ) H^{1/2}(\mathbb{S}^{1};\mathbb{R}^{2}) H 1/2 ( S 1 ; R 2 ) and are tangent: h ( s ) ⋅ ν ( Φ ( s ) ) = 0 h(s)\cdot\nu(\Phi(s))=0 h ( s ) ⋅ ν ( Φ ( s )) = 0 for almost every s ∈ A s\in A s ∈ A . For h ∈ A h\in\mathcal{A} h ∈ A let W h W_h W h be defined componentwise by (16) and
X h = W h ∘ Φ − 1 . X_h=W_h\circ\Phi^{-1}. X h = W h ∘ Φ − 1 . Only the almost-everywhere class of h h h affects W h W_h W h , X h X_h X h , or the energy below. All pointwise claims about eigenfunction traces will use their continuous representatives. The next lemma establishes Sobolev regularity before an energy is assigned to X h X_h X h .
Lemma 5.2 (well-defined physical extension). For every h ∈ A h\in\mathcal{A} h ∈ A , the field X h X_h X h belongs to H 1 ( Ω ; R 2 ) H^1(\Omega;\mathbb{R}^{2}) H 1 ( Ω ; R 2 ) and is tangent, with trace h ∘ Φ − 1 h\circ\Phi^{-1} h ∘ Φ − 1 on Γ \Gamma Γ .
Proof. Since h ∈ H 1 / 2 h\in H^{1/2} h ∈ H 1/2 , H h ∈ H 1 ( D ) \operatorname{H}h\in H^1(\mathbb{D}) H h ∈ H 1 ( D ) with trace h h h , and ζ h ∈ H 0 1 ( D ) \zeta_h\in H^1_0(\mathbb{D}) ζ h ∈ H 0 1 ( D ) , so W h ∈ H 1 ( D ) W_h\in H^1(\mathbb{D}) W h ∈ H 1 ( D ) . The Dirichlet integral is conformally invariant and ∫ Ω ∣ X h ∣ 2 = μ − 1 ∥ W h ∥ q 2 < ∞ \int_{\Omega}|X_h|^2=\mu^{-1}\lVert W_h\rVert_q^2<\infty ∫ Ω ∣ X h ∣ 2 = μ − 1 ∥ W h ∥ q 2 < ∞ , so X h ∈ H 1 ( Ω ) X_h\in H^1(\Omega) X h ∈ H 1 ( Ω ) . Near a compact subset of an open smooth arc Φ − 1 \Phi^{-1} Φ − 1 is a diffeomorphism up to the boundary, so traces correspond: the trace of X h X_h X h on Γ \Gamma Γ is h ∘ Φ − 1 h\circ\Phi^{-1} h ∘ Φ − 1 , which is tangent. □
Definition 5.3. For h ∈ A h\in\mathcal{A} h ∈ A , define
E ( h ) = Q ( X h ) − μ ∥ X h ∥ L 2 ( Ω ) 2 , E ( h , k ) = 1 2 ( E ( h + k ) − E ( h ) − E ( k ) ) . E(h)=Q(X_h)-\mu\lVert X_h\rVert_{L^2(\Omega)}^2,\qquad E(h,k)=\frac{1}{2}\bigl(E(h+k)-E(h)-E(k)\bigr). E ( h ) = Q ( X h ) − μ ∥ X h ∥ L 2 ( Ω ) 2 , E ( h , k ) = 2 1 ( E ( h + k ) − E ( h ) − E ( k ) ) . These quantities are finite by Lemma 5.2 . The linearity of h ↦ X h h\mapsto X_h h ↦ X h makes E ( h , k ) E(h,k) E ( h , k ) a symmetric bilinear form, with E ( h , h ) = E ( h ) E(h,h)=E(h) E ( h , h ) = E ( h ) .
In [14 ] the data are Lipschitz. The corner gradient here is continuous and bounded, but Lipschitz regularity is not required. Its Sobolev trace is H 1 / 2 H^{1/2} H 1/2 , which motivates this class. Define also the pulled-back curvature
c ( s ) = κ ( Φ ( s ) ) ∣ Φ ′ ( s ) ∣ ( s ∈ A ) , ∫ S 1 ∣ c ∣ d s = ∫ Γ ∣ κ ∣ d σ < ∞ . c(s)=\kappa(\Phi(s))|\Phi'(s)|\quad(s\in A),\qquad\int_{\mathbb{S}^{1}}|c|\,ds=\int_{\Gamma}|\kappa|\,d\sigma<\infty. c ( s ) = κ ( Φ ( s )) ∣ Φ ′ ( s ) ∣ ( s ∈ A ) , ∫ S 1 ∣ c ∣ d s = ∫ Γ ∣ κ ∣ d σ < ∞.
Values assigned at the finite exceptional set do not affect any integral. Although κ \kappa κ is bounded on the physical arcs, c c c need not be bounded.
Lemma 5.4. Let h , k ∈ A h,k \in\mathcal{A} h , k ∈ A .
(a) E ( h ) ≥ 0 E(h) \ge0 E ( h ) ≥ 0 , and E ( h ) = 0 E(h)=0 E ( h ) = 0 if and only if X h = ∇ v X_h=\nabla v X h = ∇ v for some v ∈ V v \in V v ∈ V .
(b) One has
E ( h , k ) = 1 2 ∬ N 0 ( s , t ) ( h ( s ) − h ( t ) ) ⋅ ( k ( s ) − k ( t ) ) d s d t − ∬ R ( s , t ) h ( s ) ⋅ k ( t ) d s d t + ∫ S 1 c ( s ) h ( s ) ⋅ k ( s ) d s . (17) E(h,k)=\frac{1}{2}\iint N_0(s,t)(h(s)-h(t))\cdot(k(s)-k(t))\,ds\,dt
-\iint R(s,t)h(s)\cdot k(t)\,ds\,dt+\int_{\mathbb{S}^1}c(s)h(s)\cdot k(s)\,ds. \tag*{(17)} E ( h , k ) = 2 1 ∬ N 0 ( s , t ) ( h ( s ) − h ( t )) ⋅ ( k ( s ) − k ( t )) d s d t − ∬ R ( s , t ) h ( s ) ⋅ k ( t ) d s d t + ∫ S 1 c ( s ) h ( s ) ⋅ k ( s ) d s . (17)
Proof. (a) is Proposition 3.1.
(b) By (9) and conformal invariance, E ( h , k ) = ∫ D ∇ W h : ∇ W k − ⟨ W h , W k ⟩ q + ∫ S 1 c h ⋅ k E(h,k)=\int_{\mathbb{D}}\nabla W_h:\nabla W_k-\langle W_h,W_k\rangle_q+\int_{\mathbb{S}^1}ch\cdot k E ( h , k ) = ∫ D ∇ W h : ∇ W k − ⟨ W h , W k ⟩ q + ∫ S 1 c h ⋅ k . The last term is absolutely integrable because h , k h,k h , k are bounded and c ∈ L 1 c\in L^1 c ∈ L 1 . For the volume terms, by Lemma 4.9(a) for k k k , tested with ζ h \zeta_h ζ h , the function W h W_h W h may be replaced by H h Hh H h ; since H h Hh H h is harmonic and ζ k ∈ H 0 1 \zeta_k\in H^1_0 ζ k ∈ H 0 1 , ∫ ∇ H h : ∇ W k = ∫ ∇ H h : ∇ H k \int\nabla Hh:\nabla W_k=\int\nabla Hh:\nabla Hk ∫ ∇ H h : ∇ W k = ∫ ∇ H h : ∇ H k . Thus
E ( h , k ) = ∫ D ∇ H h : ∇ H k − ⟨ H h , ( I − T ) − 1 H k ⟩ q + ∫ S 1 c h ⋅ k . E(h,k)=\int_{\mathbb{D}}\nabla Hh:\nabla Hk-\langle Hh,(I-T)^{-1}Hk\rangle_q+\int_{\mathbb{S}^1}ch\cdot k. E ( h , k ) = ∫ D ∇ H h : ∇ H k − ⟨ H h , ( I − T ) − 1 H k ⟩ q + ∫ S 1 c h ⋅ k .
The first term is the Douglas integral, valid for H 1 / 2 H^{1/2} H 1/2 data (both sides diagonalize in Fourier modes), and the second is Lemma 4.9(c). The first double integral is absolutely convergent by Cauchy–Schwarz in the H 1 / 2 H^{1/2} H 1/2 seminorm; the second is absolutely convergent because R ∈ L 1 R\in L^1 R ∈ L 1 and h , k h,k h , k are bounded. The last term is precisely the transformed curvature integral. □ \square □
Fix 0 ≠ u ∈ V 0\ne u\in V 0 = u ∈ V and put g = ( ∇ u ) ∘ Φ ∣ S 1 g=(\nabla u)\circ\Phi|_{\mathbb{S}^1} g = ( ∇ u ) ∘ Φ ∣ S 1 .
Lemma 5.5. g ∈ A g\in\mathcal{A} g ∈ A , g g g is continuous, g ( s j ) = 0 g(s_j)=0 g ( s j ) = 0 for every prevertex, W g = ( ∇ u ) ∘ Φ W_g=(\nabla u)\circ\Phi W g = ( ∇ u ) ∘ Φ , and
∫ S 1 K ( p , s ) g ( s ) d s = ∇ u ( Φ ( p ) ) ( p ∈ D ) , E ( g ) = 0. (18) \int_{\mathbb{S}^1}K(p,s)g(s)\,ds=\nabla u(\Phi(p))\quad(p\in\mathbb{D}),\qquad E(g)=0. \tag*{(18)} ∫ S 1 K ( p , s ) g ( s ) d s = ∇ u ( Φ ( p )) ( p ∈ D ) , E ( g ) = 0. (18) If b : S 1 → R b:\mathbb{S}^1\to\mathbb{R} b : S 1 → R is Lipschitz, then b g ∈ A bg\in\mathcal{A} b g ∈ A .
Proof. By Lemma 2.2 each w = ∂ i u w=\partial_i u w = ∂ i u is in C ( Ω ‾ ) ∩ H 1 ( Ω ) C(\overline{\Omega})\cap H^1(\Omega) C ( Ω ) ∩ H 1 ( Ω ) , vanishes at the vertices and solves ( Δ + μ ) w = 0 (\Delta+\mu)w=0 ( Δ + μ ) w = 0 ; ∇ u \nabla u ∇ u is tangent on the open smooth arcs. Then w ∘ Φ ∈ C ( D ‾ ) ∩ H 1 ( D ) w\circ\Phi\in C(\overline{\mathbb{D}})\cap H^1(\mathbb{D}) w ∘ Φ ∈ C ( D ) ∩ H 1 ( D ) , so its restriction g i g_i g i to S 1 \mathbb{S}^1 S 1 is in H 1 / 2 H^{1/2} H 1/2 ; hence g ∈ A g\in\mathcal{A} g ∈ A . For ψ ∈ C c ∞ ( D ) \psi\in C^\infty_c(\mathbb{D}) ψ ∈ C c ∞ ( D ) , a change of variables gives ∫ D ∇ ( w ∘ Φ ) ⋅ ∇ ψ = ⟨ ψ , w ∘ Φ ⟩ q \int_{\mathbb{D}}\nabla(w\circ\Phi)\cdot\nabla\psi=\langle\psi,w\circ\Phi\rangle_q ∫ D ∇ ( w ∘ Φ ) ⋅ ∇ ψ = ⟨ ψ , w ∘ Φ ⟩ q , and this extends to ψ ∈ H 0 1 ( D ) \psi\in H^1_0(\mathbb{D}) ψ ∈ H 0 1 ( D ) by (10) . By Lemma 4.9(a), W g i W_{g_i} W g i satisfies the same identity and has the same trace. The difference δ ∈ H 0 1 ( D ) \delta\in H^1_0(\mathbb{D}) δ ∈ H 0 1 ( D ) satisfies ∫ ∣ ∇ δ ∣ 2 = ∥ δ ∥ q 2 ≤ d ∫ ∣ ∇ δ ∣ 2 \int|\nabla\delta|^2=\|\delta\|_q^2\le d\int|\nabla\delta|^2 ∫ ∣∇ δ ∣ 2 = ∥ δ ∥ q 2 ≤ d ∫ ∣∇ δ ∣ 2 , so δ = 0 \delta=0 δ = 0 . Both W g W_g W g (in the representative of Lemma 4.9(b)) and ( ∇ u ) ∘ Φ (\nabla u)\circ\Phi ( ∇ u ) ∘ Φ are continuous in D \mathbb{D} D , so they agree at every p ∈ D p\in\mathbb{D} p ∈ D ; this is the first part of (18) . The second part follows from X g = ∇ u X_g=\nabla u X g = ∇ u and Proposition 3.1. Finally, if b b b has Lipschitz constant ℓ \ell ℓ , then ∣ b ( s ) g ( s ) − b ( t ) g ( t ) ∣ 2 ≤ 2 ∥ b ∥ ∞ 2 ∣ g ( s ) − g ( t ) ∣ 2 + 2 ℓ 2 ∥ g ∥ ∞ 2 ∣ s − t ∣ 2 |b(s)g(s)-b(t)g(t)|^2\le2\|b\|_\infty^2|g(s)-g(t)|^2+2\ell^2\|g\|_\infty^2|s-t|^2 ∣ b ( s ) g ( s ) − b ( t ) g ( t ) ∣ 2 ≤ 2∥ b ∥ ∞ 2 ∣ g ( s ) − g ( t ) ∣ 2 + 2 ℓ 2 ∥ g ∥ ∞ 2 ∣ s − t ∣ 2 , and N 0 ( s , t ) ∣ s − t ∣ 2 N_0(s,t)|s-t|^2 N 0 ( s , t ) ∣ s − t ∣ 2 is bounded, so b g ∈ H 1 / 2 bg\in H^{1/2} b g ∈ H 1/2 ; it is bounded and tangent. □ \square □
Proposition 5.6 (multiplier identity). For every Lipschitz b : S 1 → R b:\mathbb{S}^1\to\mathbb{R} b : S 1 → R ,
E ( b g ) = 1 2 ∬ S 1 × S 1 N ( s , t ) ( b ( s ) − b ( t ) ) 2 g ( s ) ⋅ g ( t ) d s d t , (19) E(bg)=\frac{1}{2}\iint_{\mathbb{S}^1\times\mathbb{S}^1}N(s,t)(b(s)-b(t))^2g(s)\cdot g(t)\,ds\,dt, \tag*{(19)} E ( b g ) = 2 1 ∬ S 1 × S 1 N ( s , t ) ( b ( s ) − b ( t ) ) 2 g ( s ) ⋅ g ( t ) d s d t , (19) and the integral converges absolutely.
Proof. E E E is a nonnegative quadratic form on A \mathcal{A} A and E ( g ) = 0 E(g)=0 E ( g ) = 0 , hence E ( g , k ) = 0 E(g,k)=0 E ( g , k ) = 0 for all k ∈ A k\in\mathcal{A} k ∈ A ; in particular E ( g , b 2 g ) = 0 E(g,b^2g)=0 E ( g , b 2 g ) = 0 . Subtract (17) for the pair ( g , b 2 g ) (g,b^2g) ( g , b 2 g ) from (17) for ( b g , b g ) (bg,bg) ( b g , b g ) . The curvature contribution cancels exactly:
∫ S 1 c ( s ) ∣ b ( s ) g ( s ) ∣ 2 d s − ∫ S 1 c ( s ) g ( s ) ⋅ b ( s ) 2 g ( s ) d s = 0. (20) \int_{\mathbb{S}^1}c(s)|b(s)g(s)|^2\,ds-\int_{\mathbb{S}^1}c(s)g(s)\cdot b(s)^2g(s)\,ds=0. \tag*{(20)} ∫ S 1 c ( s ) ∣ b ( s ) g ( s ) ∣ 2 d s − ∫ S 1 c ( s ) g ( s ) ⋅ b ( s ) 2 g ( s ) d s = 0. (20)
Both integrals are absolutely convergent, and this uses no sign condition on c c c or κ \kappa κ . Both N 0 N_0 N 0 -integrals converge absolutely (the data are in H 1 / 2 H^{1/2} H 1/2 ), and the difference of the integrands, writing b s = b ( s ) b_s=b(s) b s = b ( s ) , b t = b ( t ) b_t=b(t) b t = b ( t ) , g s = g ( s ) g_s=g(s) g s = g ( s ) and g t = g ( t ) g_t=g(t) g t = g ( t ) , is
∣ b s g s − b t g t ∣ 2 − ( g s − g t ) ⋅ ( b s 2 g s − b t 2 g t ) = ( b s − b t ) 2 g s ⋅ g t . \lvert b_sg_s-b_tg_t\rvert^2-(g_s-g_t)\cdot(b_s^2g_s-b_t^2g_t)=(b_s-b_t)^2g_s\cdot g_t. ∣ b s g s − b t g t ∣ 2 − ( g s − g t ) ⋅ ( b s 2 g s − b t 2 g t ) = ( b s − b t ) 2 g s ⋅ g t .
For the second terms, the symmetry of R R R gives − ∬ R ( b s b t − b s 2 ) g s ⋅ g t = 1 2 ∬ R ( b s − b t ) 2 g s ⋅ g t -\iint R(b_sb_t-b_s^2)g_s\cdot g_t=\frac{1}{2}\iint R(b_s-b_t)^2g_s\cdot g_t − ∬ R ( b s b t − b s 2 ) g s ⋅ g t = 2 1 ∬ R ( b s − b t ) 2 g s ⋅ g t . Absolute convergence: N 0 ( b s − b t ) 2 N_0(b_s-b_t)^2 N 0 ( b s − b t ) 2 is bounded and R R R is integrable.
Proof of Theorem A
Lemma 6.1. Let p ∈ D p\in\mathbb{D} p ∈ D . There are a separable real Hilbert space H \mathcal{H} H and a Lipschitz map B ‾ : S 1 → H \overline{B}:S^1\to\mathcal{H} B : S 1 → H , injective on A A A , with ∥ B ‾ ( s ) − B ‾ ( t ) ∥ 2 = D p ( s , t ) \lVert\overline{B}(s)-\overline{B}(t)\rVert^2=D_p(s,t) ∥ B ( s ) − B ( t ) ∥ 2 = D p ( s , t ) for s , t ∈ A s,t\in A s , t ∈ A .
Proof. We recall the Gram-kernel construction in Schoenberg’s theorem [29 ] . Fix s ∗ ∈ A s_*\in A s ∗ ∈ A and set
H p ( s , t ) = 1 2 ( D p ( s , s ∗ ) + D p ( t , s ∗ ) − D p ( s , t ) ) . H_p(s,t)=\frac{1}{2}\left(D_p(s,s_*)+D_p(t,s_*)-D_p(s,t)\right). H p ( s , t ) = 2 1 ( D p ( s , s ∗ ) + D p ( t , s ∗ ) − D p ( s , t ) ) .
For any finite points t i ∈ A t_i\in A t i ∈ A and coefficients c i c_i c i , append the coefficient − ∑ i c i -\sum_i c_i − ∑ i c i at s ∗ s_* s ∗ in Proposition 4.8. The resulting inequality is ∑ i , j c i c j H p ( t i , t j ) ≥ 0 \sum_{i,j}c_ic_jH_p(t_i,t_j)\ge0 ∑ i , j c i c j H p ( t i , t j ) ≥ 0 . Thus H p H_p H p is a positive semidefinite Gram kernel. Quotient the real vector space spanned by symbols e s e_s e s by the null space of the bilinear form ⟨ e s , e t ⟩ = H p ( s , t ) \langle e_s,e_t\rangle=H_p(s,t) ⟨ e s , e t ⟩ = H p ( s , t ) , and complete it to obtain H \mathcal{H} H . The images B ( s ) B(s) B ( s ) of e s e_s e s satisfy ∥ B ( s ) − B ( t ) ∥ 2 = D p ( s , t ) \lVert B(s)-B(t)\rVert^2=D_p(s,t) ∥ B ( s ) − B ( t ) ∥ 2 = D p ( s , t ) . By (15) the resulting map B B B is injective and Lipschitz on A A A . Since A A A is dense in S 1 S^1 S 1 , it extends uniquely to a Lipschitz map B ‾ \overline{B} B on S 1 S^1 S 1 . The image is compact and hence separable; its closed linear span is a separable Hilbert space and contains the whole construction. Restrict H \mathcal{H} H to this span.
The extension need not be injective at the prevertices. The next elementary lemma replaces the statement “an injective continuous image of a circle is not contained in a line”.
Lemma 6.2. No continuous f : S 1 → R f:S^1\to\mathbb{R} f : S 1 → R is injective on the complement of a finite set F F F .
Proof. Such an f f f is not constant; let f ( s − ) = min f < max f = f ( s + ) f(s_-)=\min f<\max f=f(s_+) f ( s − ) = min f < max f = f ( s + ) . Each of the two open arcs of S 1 ∖ { s − , s + } S^1\setminus\{s_-,s_+\} S 1 ∖ { s − , s + } is mapped onto a set containing ( min f , max f ) (\min f,\max f) ( min f , max f ) . Choose t 0 t_0 t 0 in this interval outside the finite set f ( F ) f(F) f ( F ) . Then t 0 t_0 t 0 is attained in both arcs, at two distinct points outside F F F .
Proof of Theorem A. Let 0 ≠ u ∈ V 0\ne u\in V 0 = u ∈ V , g g g as above, and suppose ∇ u ( Φ ( p ) ) = 0 \nabla u(\Phi(p))=0 ∇ u ( Φ ( p )) = 0 for some p ∈ D p\in\mathbb{D} p ∈ D . Take B ‾ \overline{B} B from Lemma 6.1 and let b j b_j b j be its coordinates in an orthonormal basis of H \mathcal{H} H ; they are Lipschitz. By Proposition 5.6, for partial sums the absolute integrand is at most
N ( s , t ) D p ( s , t ) ∣ g ( s ) ∣ ∣ g ( t ) ∣ = K ( p , s ) K ( p , t ) ∣ g ( s ) ∣ ∣ g ( t ) ∣ ≤ C p 2 ∥ g ∥ ∞ 2 N(s,t)D_p(s,t)\lvert g(s)\rvert\lvert g(t)\rvert=K(p,s)K(p,t)\lvert g(s)\rvert\lvert g(t)\rvert\le C_p^2\lVert g\rVert_\infty^2 N ( s , t ) D p ( s , t ) ∣ g ( s )∣ ∣ g ( t )∣ = K ( p , s ) K ( p , t ) ∣ g ( s )∣ ∣ g ( t )∣ ≤ C p 2 ∥ g ∥ ∞ 2
for almost every ( s , t ) (s,t) ( s , t ) : the diagonal and pairs involving a prevertex form an arclength-product null set. Thus by dominated convergence and (18)
∑ j E ( b j g ) = 1 2 ∬ K ( p , s ) K ( p , t ) g ( s ) ⋅ g ( t ) d s d t − 1 2 ∣ ∫ S 1 K ( p , s ) g ( s ) d s ∣ 2 − 1 2 ∣ ∇ u ( Φ ( p ) ) ∣ 2 = 0. \sum_j E(b_jg)=\frac{1}{2}\iint K(p,s)K(p,t)g(s)\cdot g(t)\,ds\,dt-\frac{1}{2}\left\lvert\int_{S^1}K(p,s)g(s)\,ds\right\rvert^2-\frac{1}{2}\lvert\nabla u(\Phi(p))\rvert^2=0. j ∑ E ( b j g ) = 2 1 ∬ K ( p , s ) K ( p , t ) g ( s ) ⋅ g ( t ) d s d t − 2 1 ∫ S 1 K ( p , s ) g ( s ) d s 2 − 2 1 ∣ ∇ u ( Φ ( p )) ∣ 2 = 0.
Every term is nonnegative, so E ( b j g ) = 0 E(b_jg)=0 E ( b j g ) = 0 for all j j j , and by Lemma 5.4 there are u j ∈ V u_j\in V u j ∈ V with X b j g = ∇ u j X_{b_jg}=\nabla u_j X b j g = ∇ u j . The traces agree almost everywhere on ∂ Ω \partial\Omega ∂ Ω and the boundary functions are continuous. On every compact subarc of A A A , Φ \Phi Φ is a smooth boundary diffeomorphism, so equality holds everywhere there. Continuity extends it to the prevertices as well; both sides vanish there by Lemma 2.2. Thus
( ∇ u j ) ∘ Φ = b j g on S 1 . (21) (\nabla u_j)\circ\Phi=b_jg\qquad\text{on }S^1. \tag*{(21)} ( ∇ u j ) ∘ Φ = b j g on S 1 . (21)
Let S g = { s ∈ S 1 : g ( s ) ≠ 0 } S_g=\{s\in\mathbb{S}^1:g(s)\ne0\} S g = { s ∈ S 1 : g ( s ) = 0 } . It is open, and contained in A A A . It is nonempty: otherwise each ∂ i u ∈ C ( Ω ‾ ) ∩ H 1 ( Ω ) \partial_i u\in C(\overline{\Omega})\cap H^1(\Omega) ∂ i u ∈ C ( Ω ) ∩ H 1 ( Ω ) vanishes on ∂ Ω \partial\Omega ∂ Ω , so ∂ i u ∈ H 0 1 ( Ω ) \partial_i u\in H^1_0(\Omega) ∂ i u ∈ H 0 1 ( Ω ) solves ( Δ + μ ) ∂ i u = 0 (\Delta+\mu)\partial_i u=0 ( Δ + μ ) ∂ i u = 0 , and μ < λ D \mu<\lambda_D μ < λ D gives ∇ u = 0 \nabla u=0 ∇ u = 0 , hence u = 0 u=0 u = 0 .
Case 1: S g S_g S g is dense in S 1 \mathbb{S}^1 S 1 . If B ‾ ( S g ) \overline{B}(S_g) B ( S g ) were contained in an affine line, so would be B ‾ ( S 1 ) \overline{B}(\mathbb{S}^1) B ( S 1 ) by continuity, and a coordinate on that line would contradict Lemma 6.2 with F = S 1 ∖ A F=\mathbb{S}^1\setminus A F = S 1 ∖ A . So there are three points of S g S_g S g with affinely independent images, hence two indices j , k j,k j , k such that 1 , b j , b k 1,b_j,b_k 1 , b j , b k are linearly independent on these three points: the two independent difference vectors have a nonzero two-coordinate minor in an orthonormal basis. As g ≠ 0 g\ne0 g = 0 there, the functions g , b j g , b k g g,b_jg,b_kg g , b j g , b k g are linearly independent, and by (21) so are u , u j , u k u,u_j,u_k u , u j , u k . This contradicts Proposition 2.5 .
Case 2: S g S_g S g is not dense. Then g = 0 g=0 g = 0 on an open arc I 0 ⊂ A I_0\subset A I 0 ⊂ A , and Φ ( I 0 ) \Phi(I_0) Φ ( I 0 ) is a regular open boundary arc. Since B ‾ \overline{B} B is injective on the nonempty open set S g ⊂ A S_g\subset A S g ⊂ A , some b j b_j b j is not constant on S g S_g S g ; then g g g and b j g b_jg b j g , and hence u u u and u j u_j u j , are linearly independent. Their gradients vanish on Φ ( I 0 ) \Phi(I_0) Φ ( I 0 ) by (21) , so both are constant there. A nonzero combination w ∈ V w\in V w ∈ V has w = 0 w=0 w = 0 and ∇ w = 0 \nabla w=0 ∇ w = 0 on Φ ( I 0 ) \Phi(I_0) Φ ( I 0 ) . Choose a ball meeting that arc but no corners or other boundary pieces, and extend w w w by zero to the exterior portion of the ball. Its zero Dirichlet trace makes the extension H 1 H^1 H 1 , and its zero Neumann trace makes the extension a weak solution of ( Δ + μ ) w = 0 (\Delta+\mu)w=0 ( Δ + μ ) w = 0 throughout the ball: integration by parts produces no interface term. Interior elliptic regularity then makes this extension real analytic. It vanishes on an open exterior set, so it vanishes in the ball and, by interior unique continuation, throughout Ω \Omega Ω . This contradicts the choice of a nonzero w w w . No reflection through a curved arc, and no analyticity of the boundary, is used.
Hence ∇ u ≠ 0 \nabla u\ne0 ∇ u = 0 in Ω \Omega Ω . Continuity on the compact closure guarantees a global maximum and minimum. Neither can occur in the interior, since an interior extremum would be a critical point. Both are therefore boundary values, and no interior point can attain either, proving the strict inequalities.
Concluding remarks
The reciprocal-kernel input is proved in Proposition 4.7 for the bounded compactly supported densities used by the truncation argument, and Proposition 4.8 passes to the singular conformal potential. Thus Theorem A does not assume [14 Theorem 5.1] . The method, including the regularization and the limiting framework of [14 Section 5] , is adapted from that preprint, and the variational framework with curvature is due to Rohleder [18 ] . The contribution of this paper is the extension to corners, with the corner and trace estimates it requires, together with a differential proof of the finite preservation step, which is an alternative to [14 Lemma 5.4] .
The condition α j < 1 \alpha_j<1 α j < 1 is what makes every exponent in (5) positive. At a reentrant corner the first singular Neumann mode can have an unbounded gradient and the H 2 H^2 H 2 input may fail, and the method gives no conclusion there. Some domains with a reentrant corner, such as L-shaped polygons, are treated by a different argument in [12 ] , which makes positive use of that singularity. Cusps, infinitely many corners, rough boundaries and general convex domains are not treated. For the last class one would need to pass to the limit without losing the strict conclusion, which the present argument does not provide.
Acknowledgments
Acknowledgments. Large language models were used as assistants in developing and checking the arguments and in preparing the text. The authors have verified the content and take full responsibility for it.
References
Universidad Nacional de Colombia, Medellín, Colombia Email address: sapinedamo@unal.edu.co
Universidad EAFIT, Medellín, Colombia Email address: egalleg8@eafit.edu.co
Universidad Nacional de Colombia, Leticia, Amazonas, Colombia Email address: jporras@unal.edu.co
Universidad de Antioquia, Medellín, Colombia Email address: heriberto.rua@udea.edu.co