AI Disclosure
The initial draft of § 4 , including the proof extending the smooth-domain assertion reported in [34 ] to bounded nonsmooth simply connected domains, was generated entirely by ChatGPT (6.1 Sol) in response to a few human prompts. Independently and concurrently, K. Burdzy used ChatGPT to obtain an almost identical theorem and proof, as reported in [15 Human Commentary, item (ii)] .
The initial drafts of § 5 –§ 9 , which develop the heat-kernel approach, the nonsmooth examples, the unbounded examples, and the discussion of simplicity, were also generated by ChatGPT (6.1 Sol), with substantial input, guidance, and scrutiny from the author. The present version incorporates a further AI-assisted review and revision of the arguments, references, and exposition. The author’s main interest was to identify additional examples of domains with nonsmooth boundaries and compact Neumann resolvent, including examples satisfying polynomial heat-kernel bounds, for which the first positive Neumann eigenvalue is simple. The aim was to apply the result in § 4 to these domains. The mathematical ideas in these sections, including the introduction of heat-kernel bounds and cross-cut domains and techniques, are due to the author. The author reviewed and edited all AI-assisted material and checked the arguments. The smooth-domain assertion enters only through the explicit Hypothesis 1.1 . Its proof is not assessed here.
The author does not intend to submit these notes for journal publication.
Contents
Introduction
1 Hot spots under compact Sobolev embedding
1.1 Setting and statements
1.2 Compactness and the extension operator
1.3 Spectral convergence under compact embedding
1.4 Passage to the hot-spots conclusion
2 Heat-kernel bounds along the conformal exhaustion
2.1 Uniform bounds along the conformal exhaustion
2.2 A second proof via spectral stability
3 Domains with polynomial heat-kernel bounds
3.1 The Sobolev criterion
3.2 Lipschitz domains, quasidisks, and Hölder graph domains
3.3 Power cusps
4 Cross-cut domains and simplicity of the first positive eigenvalue
4.1 Cross-cut domains
4.2 The cross-cut comparison principle
4.3 Thin profiles
4.4 Example I: Lipschitz tubes
4.5 Example II: quadratic cusps
4.6 Example III: an exponential cusp without a polynomial heat-kernel bound
4.7 Summary of the examples
5 Unbounded domains of finite area
5.1 The Gaussian horn
5.2 An unbounded rooms-and-passages domain
6 On removing the simplicity assumption
Further classes with polynomial heat-kernel bounds
A.1 Quasihyperbolic Hölder domains
A.2 Quasihyperbolic order of the power cusps
A.3 Finite decompositions and a slit domain
References
Introduction
The hot spots conjecture of J. Rauch asserts that a first positive Neumann eigenfunction of a planar domain attains its maximum and minimum on the boundary. See [8 ] for a precise formulation. The conjecture has inspired a vast literature, with connections to probability, analysis, partial differential equations, and geometry. These notes make no attempt to survey this literature, and results closely related to the questions considered here may have been overlooked. These notes take an assertion about smooth domains as an external hypothesis and derive a weaker conclusion for domains with irregular boundary.
The hypothesis. The external input, reported in [34 ] and stated as Hypothesis 1.1, is that on a bounded smooth simply connected planar domain no first positive Neumann eigenfunction has an interior local maximum or minimum. The assertion in [34 ] is stronger, namely that such eigenfunctions have no interior critical points. Its proof is not examined here.
The main result. Let Ω ⊂ R 2 \Omega\subset\mathbb{R}^{2} Ω ⊂ R 2 be a simply connected domain of finite area, and suppose that the embedding H 1 ( Ω ) ↪ L 2 ( Ω ) H^{1}(\Omega) \hookrightarrow L^{2}(\Omega) H 1 ( Ω ) ↪ L 2 ( Ω ) is compact and that the first positive Neumann eigenvalue λ 1 ( Ω ) \lambda_{1}(\Omega) λ 1 ( Ω ) is simple. Theorem 1.2 states that, under Hypothesis 1.1, every real eigenfunction u u u for λ 1 ( Ω ) \lambda_{1}(\Omega) λ 1 ( Ω ) satisfies
sup Ω u = lim δ ↓ 0 sup { u ( x ) : x ∈ Ω , dist ( x , ∂ Ω ) < δ } . \sup_{\Omega} u = \lim_{\delta\downarrow0} \sup\{u(x) : x \in\Omega,\ \operatorname{dist}(x,\partial\Omega) < \delta\}. Ω sup u = δ ↓ 0 lim sup { u ( x ) : x ∈ Ω , dist ( x , ∂ Ω ) < δ } .
and the same identity for the infimum. In words: the supremum and the infimum are approached in arbitrarily thin boundary layers. If Ω \Omega Ω is bounded and u u u is continuous on Ω ‾ \overline{\Omega} Ω , this says that the global extrema of u u u are attained on the boundary. The conclusion is nonstrict. It does not exclude an interior point at which an extremal value is also attained. No regularity of ∂ Ω \partial\Omega ∂ Ω is assumed.
The proof is by conformal exhaustion. Given a conformal bijection F : D → Ω F : \mathbb{D} \to\Omega F : D → Ω , the domains Ω r = F ( r D ) \Omega_{r} = F(r\mathbb{D}) Ω r = F ( r D ) , 0 < r < 1 0 < r < 1 0 < r < 1 , are bounded by analytic Jordan curves and increase to Ω \Omega Ω . Reflection in the parameter disk extends functions from Ω r \Omega_{r} Ω r to Ω \Omega Ω with at most twice the energy. Compact embedding then gives convergence of the first positive eigenvalues and locally uniform convergence of the eigenfunctions. If u u u had a compact superlevel set in Ω \Omega Ω above its boundary-layer value, the approximating eigenfunctions would have interior local maxima on the Ω r \Omega_{r} Ω r , contradicting Hypothesis 1.1. Burdzy’s note [15 ] , obtained independently, uses the same mechanism. It is stated for bounded domains but the proof is essentially the same for domains of finite area.
Two recent preprints give further developments on the hot spots problem. Deng, Jiang, and Yang [22 ] prove that the maximum modulus of every first positive Neumann eigenfunction on a bounded convex planar domain is attained only on the boundary. Their result requires neither boundary smoothness nor simplicity of the eigenvalue. For centrally symmetric convex planar domains, they obtain the strict hot spots conclusion for both extrema. Atar and Burdzy [6 ] construct a bounded starlike Lipschitz domain in three dimensions with simple first positive Neumann eigenvalue whose corresponding eigenfunction has its unique global maximum at an interior point.
Two routes, and the role of heat-kernel bounds. Section 1 proves Theorem 1.2 under compact embedding alone. Section 2 concerns the stronger assumption of a polynomial Neumann heat-kernel bound p Ω ( t , x , y ) ≤ C t − N / 2 p_{\Omega}(t,x,y) \leq C t^{-N/2} p Ω ( t , x , y ) ≤ C t − N /2 for 0 < t ≤ 1 0 < t \leq1 0 < t ≤ 1 , where N > 2 N > 2 N > 2 is an exponent, not the dimension. Such a bound implies compact embedding, so the hot-spots conclusion for these domains is a special case of Theorem 1.2. Section 2 is included for Proposition 2.1, which shows that a bound on the target transfers to every Ω r \Omega_{r} Ω r with r r r close to 1, with the same exponent and a constant independent of r r r .
This is the input required by the stability theory of Bañuelos–Pang [12 ] , and it gives a second proof of the hot-spots conclusion for this class.
Organization. Section 3 lists classes of domains with polynomial heat-kernel bounds: Lipschitz domains, quasidisks, Hölder graph domains, and power cusps. Section 4 compares the Neumann eigenvalues of a domain lying between two graphs with those of a weighted one-dimensional problem obtained by averaging over vertical cross-cuts. This gives explicit simplicity criteria and explicit examples: Lipschitz tubes, quadratic cusps, and an exponential cusp that has compact embedding and a simple first positive eigenvalue but no polynomial heat-kernel bound. Section 5 treats two unbounded examples of finite area: the Gaussian horn and a rooms-and-passages domain with infinitely many bottlenecks and corners. Section 6 shows that, without simplicity, the exhaustion argument still yields at least one first positive eigenfunction satisfying the boundary-layer conclusion. Appendix A collects further classes with polynomial heat-kernel bounds (quasihyperbolic Hölder domains, finite decompositions, a slit disk) that are not needed for the main line.
Hot spots under compact Sobolev embedding
Setting and statements
Throughout, Sobolev spaces and variational arguments are taken over the real scalars. A domain is a nonempty connected open set, and H 1 ( Ω ) H^{1}(\Omega) H 1 ( Ω ) is the Sobolev space of square-integrable functions on Ω \Omega Ω with square-integrable weak first derivatives. For a domain Ω ⊂ R 2 \Omega\subset\mathbb{R}^{2} Ω ⊂ R 2 of finite area, let A Ω A_{\Omega} A Ω be the nonnegative self-adjoint Neumann Laplacian associated with the closed form
E Ω ( v , w ) = ∫ Ω ∇ v ⋅ ∇ w , D ( E Ω ) = H 1 ( Ω ) . \mathcal{E}_{\Omega}(v,w)=\int_{\Omega}\nabla v\cdot\nabla w,\qquad D(\mathcal{E}_{\Omega})=H^{1}(\Omega). E Ω ( v , w ) = ∫ Ω ∇ v ⋅ ∇ w , D ( E Ω ) = H 1 ( Ω ) .
Its semigroup is T t Ω = e − t A Ω T_{t}^{\Omega}=e^{-tA_{\Omega}} T t Ω = e − t A Ω , so the associated heat equation is ∂ t v = Δ v \partial_{t}v=\Delta v ∂ t v = Δ v (Brownian motion with generator 1 2 Δ \frac{1}{2}\Delta 2 1 Δ corresponds to a change of time). This definition requires neither a trace nor a pointwise normal derivative on ∂ Ω \partial\Omega ∂ Ω . We write λ 0 ( Ω ) ≤ λ 1 ( Ω ) ≤ λ 2 ( Ω ) ≤ ⋯ \lambda_{0}(\Omega)\leq\lambda_{1}(\Omega)\leq\lambda_{2}(\Omega)\leq\cdots λ 0 ( Ω ) ≤ λ 1 ( Ω ) ≤ λ 2 ( Ω ) ≤ ⋯ for the min–max values of E Ω \mathcal{E}_{\Omega} E Ω relative to the L 2 ( Ω ) L^{2}(\Omega) L 2 ( Ω ) norm,
λ k ( Ω ) = inf W ⊂ H 1 ( Ω ) dim W = k + 1 sup w ∈ W ∖ { 0 } ∫ Ω ∣ ∇ w ∣ 2 ∫ Ω w 2 . \lambda_{k}(\Omega)=\inf_{\substack{W\subset H^{1}(\Omega)\\ \dim W=k+1}}\ \sup_{w\in W\setminus\{0\}}\frac{\int_{\Omega}|\nabla w|^{2}}{\int_{\Omega}w^{2}}. λ k ( Ω ) = W ⊂ H 1 ( Ω ) d i m W = k + 1 inf w ∈ W ∖ { 0 } sup ∫ Ω w 2 ∫ Ω ∣∇ w ∣ 2 .
the infimum running over all ( k + 1 ) (k+1) ( k + 1 ) -dimensional subspaces W ⊂ H 1 ( Ω ) W\subset H^{1}(\Omega) W ⊂ H 1 ( Ω ) . Compactness of the embedding H 1 ( Ω ) ↪ L 2 ( Ω ) H^{1}(\Omega)\hookrightarrow L^{2}(\Omega) H 1 ( Ω ) ↪ L 2 ( Ω ) means that every bounded sequence in H 1 ( Ω ) H^{1}(\Omega) H 1 ( Ω ) has an L 2 ( Ω ) L^{2}(\Omega) L 2 ( Ω ) -convergent subsequence. When the embedding is compact, the λ k ( Ω ) \lambda_{k}(\Omega) λ k ( Ω ) are the eigenvalues of A Ω A_{\Omega} A Ω counted with multiplicity; λ 0 ( Ω ) = 0 \lambda_{0}(\Omega)=0 λ 0 ( Ω ) = 0 because constants lie in L 2 ( Ω ) L^{2}(\Omega) L 2 ( Ω ) , λ 1 ( Ω ) > 0 \lambda_{1}(\Omega)>0 λ 1 ( Ω ) > 0 by connectedness, finite area, and compact embedding (Lemma 1.4 ), and λ 1 ( Ω ) \lambda_{1}(\Omega) λ 1 ( Ω ) is the first positive eigenvalue . It is simple if its eigenspace is one-dimensional, equivalently if λ 1 ( Ω ) < λ 2 ( Ω ) \lambda_{1}(\Omega)<\lambda_{2}(\Omega) λ 1 ( Ω ) < λ 2 ( Ω ) . A first positive Neumann eigenfunction is a nonzero real u ∈ H 1 ( Ω ) u\in H^{1}(\Omega) u ∈ H 1 ( Ω ) with
∫ Ω ∇ u ⋅ ∇ φ = λ 1 ( Ω ) ∫ Ω u φ for all φ ∈ H 1 ( Ω ) . \int_{\Omega}\nabla u\cdot\nabla\varphi=\lambda_{1}(\Omega)\int_{\Omega}u\varphi\qquad\text{for all }\varphi\in H^{1}(\Omega). ∫ Ω ∇ u ⋅ ∇ φ = λ 1 ( Ω ) ∫ Ω u φ for all φ ∈ H 1 ( Ω ) .
By interior elliptic regularity it is smooth in Ω \Omega Ω , and we always use this smooth representative. For δ > 0 \delta>0 δ > 0 , the boundary layer of Ω \Omega Ω is
B δ ( Ω ) = { x ∈ Ω : dist ( x , ∂ Ω ) < δ } . B_{\delta}(\Omega)=\{x\in\Omega:\operatorname{dist}(x,\partial\Omega)<\delta\}. B δ ( Ω ) = { x ∈ Ω : dist ( x , ∂ Ω ) < δ } .
Hypothesis 1.1 (Smooth-domain input). For every bounded smooth simply connected planar domain G G G , every nonzero real eigenfunction for its first positive Neumann eigenvalue has no interior local maximum or minimum.
Here smooth may be read as C ∞ C^\infty C ∞ . The hypothesis is applied below only to domains bounded by analytic Jordan curves. The assertion reported in [34 ] is that every first positive Neumann eigenfunction on a bounded smooth simply connected planar domain has no interior critical points, which implies Hypothesis 1.1. The source is a model-generated preprint, accompanied by a Lean statement of the result, in a collection whose review status is listed as unchecked. Neither the proof nor the Lean development is assessed here. The weaker statement that global extrema are attained on the boundary would not suffice for the argument below.
Theorem 1.2. Assume Hypothesis 1.1. Let Ω ⊂ R 2 \Omega\subset\mathbb{R}^2 Ω ⊂ R 2 be a simply connected domain with ∣ Ω ∣ < ∞ |\Omega| < \infty ∣Ω∣ < ∞ , and suppose that
H 1 ( Ω ) ↪ L 2 ( Ω ) is compact, (1) H^1(\Omega) \hookrightarrow L^2(\Omega) \quad\text{is compact,} \tag*{(1)} H 1 ( Ω ) ↪ L 2 ( Ω ) is compact, (1) and that its first positive Neumann eigenvalue λ 1 ( Ω ) \lambda_1(\Omega) λ 1 ( Ω ) is simple. Then every real eigenfunction u u u for λ 1 ( Ω ) \lambda_1(\Omega) λ 1 ( Ω ) satisfies
sup Ω u = lim δ ↓ 0 sup B δ ( Ω ) u , inf Ω u = lim δ ↓ 0 inf B δ ( Ω ) u , (2) \sup_{\Omega} u = \lim_{\delta\downarrow0}\sup_{B_\delta(\Omega)} u,\qquad
\inf_{\Omega} u = \lim_{\delta\downarrow0}\inf_{B_\delta(\Omega)} u, \tag*{(2)} Ω sup u = δ ↓ 0 lim B δ ( Ω ) sup u , Ω inf u = δ ↓ 0 lim B δ ( Ω ) inf u , (2) where infinite extremal values are allowed. If Ω \Omega Ω is bounded and u ∈ C ( Ω ‾ ) u \in C(\overline{\Omega}) u ∈ C ( Ω ) , then
max Ω ‾ u = max ∂ Ω u , min Ω ‾ u = min ∂ Ω u . (3) \max_{\overline{\Omega}} u = \max_{\partial\Omega} u,\qquad
\min_{\overline{\Omega}} u = \min_{\partial\Omega} u. \tag*{(3)} Ω max u = ∂ Ω max u , Ω min u = ∂ Ω min u . (3)
Compactness and the extension operator
Fix a conformal bijection F : D → Ω F : \mathbb{D} \to\Omega F : D → Ω , where D = { z ∈ C : ∣ z ∣ < 1 } \mathbb{D} = \{z \in\mathbb{C} : |z| < 1\} D = { z ∈ C : ∣ z ∣ < 1 } , and put
Ω r = F ( r D ) ( 0 < r < 1 ) . \Omega_r = F(r\mathbb{D}) \qquad(0 < r < 1). Ω r = F ( r D ) ( 0 < r < 1 ) .
Each Ω r \Omega_r Ω r is bounded by an analytic Jordan curve, Ω r ⋐ Ω \Omega_r \Subset\Omega Ω r ⋐ Ω , and Ω r ↑ Ω \Omega_r \uparrow\Omega Ω r ↑ Ω as r ↑ 1 r \uparrow1 r ↑ 1 . Every compact subset of Ω \Omega Ω lies in Ω r \Omega_r Ω r for all sufficiently large r r r , and Ω 1 / 2 ‾ ⊂ Ω r \overline{\Omega_{1/2}} \subset\Omega_r Ω 1/2 ⊂ Ω r whenever r > 1 / 2 r > 1/2 r > 1/2 . The domain Ω 1 / 2 \Omega_{1/2} Ω 1/2 serves as an anchor in the Poincaré inequality below.
Lemma 1.4 (Anchored Poincaré inequality). Assume ∣ Ω ∣ < ∞ |\Omega| < \infty ∣Ω∣ < ∞ and (1) . Then there is C 1 < ∞ C_1 < \infty C 1 < ∞ such that
∥ w ∥ L 2 ( Ω ) 2 ≤ C 1 ( ∥ ∇ w ∥ L 2 ( Ω ) 2 + ∥ w ∥ L 2 ( Ω 1 / 2 ) 2 ) ( w ∈ H 1 ( Ω ) ) . (4) \|w\|_{L^2(\Omega)}^2 \le C_1\left(\|\nabla w\|_{L^2(\Omega)}^2 + \|w\|_{L^2(\Omega_{1/2})}^2\right) \qquad(w \in H^1(\Omega)). \tag*{(4)} ∥ w ∥ L 2 ( Ω ) 2 ≤ C 1 ( ∥∇ w ∥ L 2 ( Ω ) 2 + ∥ w ∥ L 2 ( Ω 1/2 ) 2 ) ( w ∈ H 1 ( Ω )) . (4) Moreover,
λ 1 ( Ω ) = min { ∫ Ω ∣ ∇ w ∣ 2 : w ∈ H 1 ( Ω ) , ∫ Ω w = 0 , ∫ Ω w 2 = 1 } > 0. (5) \lambda_1(\Omega) = \min\left\{\int_\Omega|\nabla w|^2 : w \in H^1(\Omega),\ \int_\Omega w = 0,\ \int_\Omega w^2 = 1\right\} > 0. \tag*{(5)} λ 1 ( Ω ) = min { ∫ Ω ∣∇ w ∣ 2 : w ∈ H 1 ( Ω ) , ∫ Ω w = 0 , ∫ Ω w 2 = 1 } > 0. (5)
Proof. Suppose (4) fails. Then there are w j ∈ H 1 ( Ω ) w_j \in H^1(\Omega) w j ∈ H 1 ( Ω ) with ∥ w j ∥ L 2 ( Ω ) = 1 \|w_j\|_{L^2(\Omega)} = 1 ∥ w j ∥ L 2 ( Ω ) = 1 and ∥ ∇ w j ∥ L 2 ( Ω ) + ∥ w j ∥ L 2 ( Ω 1 / 2 ) → 0 \|\nabla w_j\|_{L^2(\Omega)} + \|w_j\|_{L^2(\Omega_{1/2})} \to0 ∥∇ w j ∥ L 2 ( Ω ) + ∥ w j ∥ L 2 ( Ω 1/2 ) → 0 . By compactness, a subsequence converges strongly in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) to some w w w with ∥ w ∥ L 2 ( Ω ) = 1 \|w\|_{L^2(\Omega)} = 1 ∥ w ∥ L 2 ( Ω ) = 1 . Its weak gradient vanishes, so w w w is constant because Ω \Omega Ω is connected, while w = 0 w = 0 w = 0 on Ω 1 / 2 \Omega_{1/2} Ω 1/2 . Hence w = 0 w = 0 w = 0 , a contradiction. This part uses only connectedness and compactness.
Finite area enters in (5) : constants belong to L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) , the mean ∫ Ω w \int_\Omega w ∫ Ω w is defined for w ∈ L 2 ( Ω ) w \in L^2(\Omega) w ∈ L 2 ( Ω ) by the Cauchy–Schwarz inequality, and ∣ Ω ∣ − 1 ∫ Ω h |\Omega|^{-1}\int_\Omega h ∣Ω ∣ − 1 ∫ Ω h is finite for h ∈ H 1 ( Ω ) h \in H^1(\Omega) h ∈ H 1 ( Ω ) . The argument of the first paragraph, with the constraint ∫ Ω w = 0 \int_\Omega w = 0 ∫ Ω w = 0 in place of the anchor, gives a positive Poincaré constant. The direct method, which uses compactness of the embedding and weak lower semicontinuity of the Dirichlet integral, then yields a minimizer in (5) . The Euler–Lagrange equation holds for mean-zero variations. Applying it to h − ∣ Ω ∣ − 1 ∫ Ω h h - |\Omega|^{-1}\int_\Omega h h − ∣Ω ∣ − 1 ∫ Ω h for an arbitrary test function h ∈ H 1 ( Ω ) h \in H^1(\Omega) h ∈ H 1 ( Ω ) gives the full weak Neumann equation, with λ 1 ( Ω ) \lambda_1(\Omega) λ 1 ( Ω ) the minimal value.
Lemma 1.5 (Inversion extension). Let ∣ Ω ∣ < ∞ |\Omega| < \infty ∣Ω∣ < ∞ . For 0 < r < 1 0 < r < 1 0 < r < 1 and v ∈ C ∞ ( Ω r ‾ ) v \in C^\infty(\overline{\Omega_r}) v ∈ C ∞ ( Ω r ) , there is a bounded function Φ r v ∈ H 1 ( Ω ) \Phi_r v \in H^1(\Omega) Φ r v ∈ H 1 ( Ω ) , equal to v v v on Ω r \Omega_r Ω r , such that
∫ Ω ∣ ∇ Φ r v ∣ 2 ≤ 2 ∫ Ω r ∣ ∇ v ∣ 2 . (6) \int_\Omega|\nabla\Phi_r v|^2 \le2 \int_{\Omega_r} |\nabla v|^2. \tag*{(6)} ∫ Ω ∣∇ Φ r v ∣ 2 ≤ 2 ∫ Ω r ∣∇ v ∣ 2 . (6) This statement does not use (1) . If (1) holds and r > 1 / 2 r > 1/2 r > 1/2 , then, with a constant C C C independent of r r r and v v v ,
∥ Φ r v ∥ H 1 ( Ω ) 2 ≤ C ( ∫ Ω r ∣ ∇ v ∣ 2 + ∫ Ω 1 / 2 ∣ v ∣ 2 ) . (7) \|\Phi_r v\|_{H^1(\Omega)}^2 \le C\left(\int_{\Omega_r} |\nabla v|^2 + \int_{\Omega_{1/2}} |v|^2\right). \tag*{(7)} ∥ Φ r v ∥ H 1 ( Ω ) 2 ≤ C ( ∫ Ω r ∣∇ v ∣ 2 + ∫ Ω 1/2 ∣ v ∣ 2 ) . (7)
Proof. Write g = v ∘ F g = v \circ F g = v ∘ F on r D ‾ \overline{r\mathbb{D}} r D . It is smooth there because F F F is holomorphic with nonvanishing derivative in a neighborhood of r D ‾ \overline{r\mathbb{D}} r D . Define
g ~ ( z ) = { g ( z ) , ∣ z ∣ ≤ r , g ( I r ( z ) ) , r < ∣ z ∣ < 1 , I r ( z ) = r 2 z ‾ . \widetilde{g}(z) =
\begin{cases}
g(z), & |z| \le r,\\
g(I_r(z)), & r < |z| < 1,
\end{cases}
\qquad I_r(z) = \frac{r^2}{\overline{z}}. g ( z ) = { g ( z ) , g ( I r ( z )) , ∣ z ∣ ≤ r , r < ∣ z ∣ < 1 , I r ( z ) = z r 2 .
The two definitions agree on ∣ z ∣ = r |z| = r ∣ z ∣ = r . The map I r I_r I r is anticonformal, maps { r < ∣ z ∣ < 1 } \{r < |z| < 1\} { r < ∣ z ∣ < 1 } bijectively onto { r 2 < ∣ z ∣ < r } \{r^2 < |z| < r\} { r 2 < ∣ z ∣ < r } , and scales lengths at z z z by the factor r 2 / ∣ z ∣ 2 ≤ 1 r^2/|z|^2 \le1 r 2 /∣ z ∣ 2 ≤ 1 . Hence g ~ \widetilde{g} g is continuous on D \mathbb{D} D and piecewise smooth with bounded gradient, so g ~ ∈ W 1 , ∞ ( D ) ⊂ H 1 ( D ) \widetilde{g} \in W^{1,\infty}(\mathbb{D}) \subset H^1(\mathbb{D}) g ∈ W 1 , ∞ ( D ) ⊂ H 1 ( D ) . No singular part of the gradient arises on the circle ∣ z ∣ = r |z| = r ∣ z ∣ = r . Since the planar Dirichlet integral is invariant under conformal and anticonformal maps,
∫ D ∖ r D ∣ ∇ g ~ ∣ 2 = ∫ { r 2 < ∣ z ∣ < r } ∣ ∇ g ∣ 2 ≤ ∫ r D ∣ ∇ g ∣ 2 . \int_{\mathbb{D}\setminus r\mathbb{D}} |\nabla\widetilde{g}|^2
= \int_{\{r^2<|z|<r\}} |\nabla g|^2
\le\int_{r\mathbb{D}} |\nabla g|^2. ∫ D ∖ r D ∣∇ g ∣ 2 = ∫ { r 2 < ∣ z ∣ < r } ∣∇ g ∣ 2 ≤ ∫ r D ∣∇ g ∣ 2 .
Set Φ r v = g ~ ∘ F − 1 \Phi_r v = \widetilde{g} \circ F^{-1} Φ r v = g ∘ F − 1 . By the chain rule this function belongs to H l o c 1 ( Ω ) H^1_{\mathrm{loc}}(\Omega) H loc 1 ( Ω ) , and a further change of variables gives (6) . It is bounded by ∥ v ∥ L ∞ ( Ω r ) \|v\|_{L^\infty(\Omega_r)} ∥ v ∥ L ∞ ( Ω r ) , hence lies in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) since ∣ Ω ∣ < ∞ |\Omega| < \infty ∣Ω∣ < ∞ . Its weak derivatives have finite global L 2 L^2 L 2 norm by (6) , so Φ r v ∈ H 1 ( Ω ) \Phi_r v \in H^1(\Omega) Φ r v ∈ H 1 ( Ω ) .
Now assume (1) and r > 1 / 2 r > 1/2 r > 1/2 . Applying (4) to Φ r v \Phi_r v Φ r v , and using Φ r v = v \Phi_r v = v Φ r v = v on Ω 1 / 2 ⊂ Ω r \Omega_{1/2} \subset\Omega_r Ω 1/2 ⊂ Ω r together with (6) , gives
∥ Φ r v ∥ H 1 ( Ω ) 2 ≤ 2 ( 1 + C 1 ) ∫ Ω r ∣ ∇ v ∣ 2 + C 1 ∫ Ω 1 / 2 ∣ v ∣ 2 , \|\Phi_r v\|_{H^1(\Omega)}^2 \le2(1+C_1)\int_{\Omega_r}|\nabla v|^2 + C_1\int_{\Omega_{1/2}}|v|^2, ∥ Φ r v ∥ H 1 ( Ω ) 2 ≤ 2 ( 1 + C 1 ) ∫ Ω r ∣∇ v ∣ 2 + C 1 ∫ Ω 1/2 ∣ v ∣ 2 ,
which is (7) . By density on the smooth domain Ω r \Omega_r Ω r , this linear extension map extends continuously to all of H 1 ( Ω r ) H^1(\Omega_r) H 1 ( Ω r ) when (1) holds and r > 1 / 2 r > 1/2 r > 1/2 . The extended map still agrees with the original function on Ω r \Omega_r Ω r .
Spectral convergence under compact embedding
A first positive eigenfunction u u u on a domain of finite area is normalized if ∫ u = 0 \int u = 0 ∫ u = 0 and ∫ u 2 = 1 \int u^2 = 1 ∫ u 2 = 1 . The first condition is automatic, since eigenfunctions are orthogonal to constants.
Proposition 1.6. Let Ω \Omega Ω be simply connected with ∣ Ω ∣ < ∞ |\Omega| < \infty ∣Ω∣ < ∞ and satisfy (1) . Then
λ 1 ( Ω r ) ⟶ λ 1 ( Ω ) ( r ↑ 1 ) . (8) \lambda_1(\Omega_r) \longrightarrow\lambda_1(\Omega) \qquad(r \uparrow1). \tag*{(8)} λ 1 ( Ω r ) ⟶ λ 1 ( Ω ) ( r ↑ 1 ) . (8) For every sequence r j ↑ 1 r_j \uparrow1 r j ↑ 1 and every choice of normalized real first positive Neumann eigenfunctions u j u_j u j on Ω r j \Omega_{r_j} Ω r j , there are a subsequence and a normalized first positive Neumann eigenfunction v v v on Ω \Omega Ω such that u j → v u_j \to v u j → v in C l o c ∞ ( Ω ) C^\infty_{\mathrm{loc}}(\Omega) C loc ∞ ( Ω ) along the subsequence. If λ 1 ( Ω ) \lambda_1(\Omega) λ 1 ( Ω ) is simple, the signs of the u j u_j u j can be chosen so that the limit is any prescribed normalized eigenfunction u u u .
Proof. Put λ = λ 1 ( Ω ) \lambda= \lambda_1(\Omega) λ = λ 1 ( Ω ) and λ r = λ 1 ( Ω r ) \lambda_r = \lambda_1(\Omega_r) λ r = λ 1 ( Ω r ) , and let u u u be a normalized first positive eigenfunction on Ω \Omega Ω .
Upper bound. Set c r = ∣ Ω r ∣ − 1 ∫ Ω r u c_r = |\Omega_r|^{-1}\int_{\Omega_r}u c r = ∣ Ω r ∣ − 1 ∫ Ω r u . Since u ∈ L 2 ( Ω ) ⊂ L 1 ( Ω ) u \in L^2(\Omega) \subset L^1(\Omega) u ∈ L 2 ( Ω ) ⊂ L 1 ( Ω ) by finite area, ∫ Ω r u → ∫ Ω u = 0 \int_{\Omega_r} u \to\int_\Omega u = 0 ∫ Ω r u → ∫ Ω u = 0 and ∣ Ω r ∣ ↑ ∣ Ω ∣ |\Omega_r| \uparrow|\Omega| ∣ Ω r ∣ ↑ ∣Ω∣ as r ↑ 1 r \uparrow1 r ↑ 1 . Hence c r → 0 c_r \to0 c r → 0 , and by monotone convergence
∫ Ω r ( u − c r ) 2 → 1 , ∫ Ω r ∣ ∇ u ∣ 2 → λ . \int_{\Omega_r}(u-c_r)^2 \to1, \qquad\int_{\Omega_r}|\nabla u|^2 \to\lambda. ∫ Ω r ( u − c r ) 2 → 1 , ∫ Ω r ∣∇ u ∣ 2 → λ .
Since u − c r u-c_r u − c r has mean zero on Ω r \Omega_r Ω r and is nonzero for large r r r , it is an admissible trial function on Ω r \Omega_r Ω r , and the variational principle yields
lim sup r ↑ 1 λ r ≤ λ . (9) \limsup_{r \uparrow1}\lambda_r \le\lambda. \tag*{(9)} r ↑ 1 lim sup λ r ≤ λ . (9)
Compact extensions. Choose r j ↑ 1 r_j \uparrow1 r j ↑ 1 with r j > 1 / 2 r_j > 1/2 r j > 1/2 and normalized first positive eigenfunctions u j u_j u j on Ω r j \Omega_{r_j} Ω r j , so that
∫ Ω r j u j = 0 , ∫ Ω r j u j 2 = 1 , ∫ Ω r j ∣ ∇ u j ∣ 2 = λ r j . \int_{\Omega_{r_j}}u_j = 0, \qquad\int_{\Omega_{r_j}}u_j^2 = 1, \qquad\int_{\Omega_{r_j}}|\nabla u_j|^2 = \lambda_{r_j}. ∫ Ω r j u j = 0 , ∫ Ω r j u j 2 = 1 , ∫ Ω r j ∣∇ u j ∣ 2 = λ r j .
Elliptic regularity for the smooth Neumann problem gives u j ∈ C ∞ ( Ω r j ‾ ) u_j \in C^\infty(\overline{\Omega_{r_j}}) u j ∈ C ∞ ( Ω r j ) . By Lemma 1.5 and (9) , the functions U j = Φ r j u j U_j = \Phi_{r_j}u_j U j = Φ r j u j are bounded in H 1 ( Ω ) H^1(\Omega) H 1 ( Ω ) . Along a subsequence,
U j ⇀ v weakly in H 1 ( Ω ) , U j → v in L 2 ( Ω ) . U_j \rightharpoonup v \text{ weakly in } H^1(\Omega), \qquad U_j \to v \text{ in } L^2(\Omega). U j ⇀ v weakly in H 1 ( Ω ) , U j → v in L 2 ( Ω ) .
No mass is lost in the omitted region, since
∫ Ω ∖ Ω r j ∣ U j ∣ 2 ≤ 2 ∥ U j − v ∥ L 2 ( Ω ) 2 + 2 ∫ Ω ∖ Ω r j ∣ v ∣ 2 → 0 , \int_{\Omega\setminus\Omega_{r_j}}|U_j|^2 \le2\|U_j-v\|_{L^2(\Omega)}^2 + 2\int_{\Omega\setminus\Omega_{r_j}}|v|^2 \to0, ∫ Ω ∖ Ω r j ∣ U j ∣ 2 ≤ 2∥ U j − v ∥ L 2 ( Ω ) 2 + 2 ∫ Ω ∖ Ω r j ∣ v ∣ 2 → 0 ,
and ∥ U j ∥ L 2 ( Ω ) 2 = 1 + ∫ Ω ∖ Ω r j ∣ U j ∣ 2 \|U_j\|_{L^2(\Omega)}^2 = 1 + \int_{\Omega\setminus\Omega_{r_j}}|U_j|^2 ∥ U j ∥ L 2 ( Ω ) 2 = 1 + ∫ Ω ∖ Ω r j ∣ U j ∣ 2 . Hence ∫ Ω v 2 = 1 \int_\Omega v^2 = 1 ∫ Ω v 2 = 1 . Moreover, ∫ Ω U j = ∫ Ω ∖ Ω r j U j → 0 \int_\Omega U_j = \int_{\Omega\setminus\Omega_{r_j}} U_j \to0 ∫ Ω U j = ∫ Ω ∖ Ω r j U j → 0 by the Cauchy–Schwarz inequality, because ∣ Ω ∖ Ω r j ∣ → 0 |\Omega\setminus\Omega_{r_j}| \to0 ∣Ω ∖ Ω r j ∣ → 0 , while ∫ Ω U j → ∫ Ω v \int_\Omega U_j \to\int_\Omega v ∫ Ω U j → ∫ Ω v since L 2 ( Ω ) ⊂ L 1 ( Ω ) L^2(\Omega) \subset L^1(\Omega) L 2 ( Ω ) ⊂ L 1 ( Ω ) . So ∫ Ω v = 0 \int_\Omega v = 0 ∫ Ω v = 0 .
Lower bound and identification. Fix s < 1 s < 1 s < 1 . For large j j j we have U j = u j U_j = u_j U j = u j on Ω s \Omega_s Ω s , so weak lower semicontinuity gives
∫ Ω s ∣ ∇ v ∣ 2 ≤ lim inf j ∫ Ω s ∣ ∇ U j ∣ 2 ≤ lim inf j λ r j . \int_{\Omega_s}|\nabla v|^2 \le\liminf_j \int_{\Omega_s}|\nabla U_j|^2 \le\liminf_j \lambda_{r_j}. ∫ Ω s ∣∇ v ∣ 2 ≤ j lim inf ∫ Ω s ∣∇ U j ∣ 2 ≤ j lim inf λ r j .
Letting s ↑ 1 s \uparrow1 s ↑ 1 and using (5) ,
λ ≤ ∫ Ω ∣ ∇ v ∣ 2 ≤ lim inf j λ r j . \lambda\le\int_\Omega|\nabla v|^2 \le\liminf_j \lambda_{r_j}. λ ≤ ∫ Ω ∣∇ v ∣ 2 ≤ j lim inf λ r j .
Applying this to a sequence realizing lim inf r ↑ 1 λ r \liminf_{r \uparrow1}\lambda_r lim inf r ↑ 1 λ r , and combining with (9) , proves (8) . For an arbitrary sequence the same argument shows that v v v is a normalized minimizer in (5) , hence
∫ Ω ∇ v ⋅ ∇ h = λ ∫ Ω v h ( h ∈ H 1 ( Ω ) ) . \int_{\Omega}\nabla v\cdot\nabla h=\lambda\int_{\Omega}vh \qquad(h\in H^1(\Omega)). ∫ Ω ∇ v ⋅ ∇ h = λ ∫ Ω v h ( h ∈ H 1 ( Ω )) .
Local smooth convergence. On every relatively compact subdomain, − Δ u j = λ r j u j -\Delta u_j=\lambda_{r_j}u_j − Δ u j = λ r j u j holds for all large j j j , with bounded local L 2 L^2 L 2 norms and bounded eigenvalues. Interior elliptic estimates of every order, Sobolev embedding, and a diagonal subsequence give convergence in C l o c ∞ ( Ω ) C^\infty_{\mathrm{loc}}(\Omega) C loc ∞ ( Ω ) . The limit coincides with the L 2 L^2 L 2 limit v v v . If λ \lambda λ is simple, then v = ± u v=\pm u v = ± u , and after changing signs of the u j u_j u j the limit is u u u .
Passage to the hot-spots conclusion
The following lemma isolates the step that converts a failure of (2) into an interior local maximum of an approximating function. It is used again in Sections 2 and 6.
Lemma 1.7 (Transfer of a compact superlevel set). Let Ω ⊂ R 2 \Omega\subset\mathbb{R}^2 Ω ⊂ R 2 be a domain with ∣ Ω ∣ < ∞ |\Omega|<\infty ∣Ω∣ < ∞ , let u ∈ C ( Ω ) u\in C(\Omega) u ∈ C ( Ω ) be real, and suppose that
sup Ω u > lim δ ↓ 0 sup B δ ( Ω ) u . \sup_{\Omega}u>\lim_{\delta\downarrow0}\sup_{B_\delta(\Omega)}u. Ω sup u > δ ↓ 0 lim B δ ( Ω ) sup u . Let Ω j ⊂ Ω \Omega_j\subset\Omega Ω j ⊂ Ω be open sets such that every compact subset of Ω \Omega Ω lies in Ω j \Omega_j Ω j for all large j j j , and let u j ∈ C ( Ω j ) u_j\in C(\Omega_j) u j ∈ C ( Ω j ) converge to u u u locally uniformly in Ω \Omega Ω . Then, for all sufficiently large j j j , u j u_j u j has a local maximum at some point of Ω j \Omega_j Ω j .
Proof. Write L = lim δ ↓ 0 sup B δ ( Ω ) u L=\lim_{\delta\downarrow0}\sup_{B_\delta(\Omega)}u L = lim δ ↓ 0 sup B δ ( Ω ) u and M = sup Ω u M=\sup_{\Omega}u M = sup Ω u . The limit exists in the extended real numbers because B δ ( Ω ) B_\delta(\Omega) B δ ( Ω ) decreases as δ \delta δ decreases, and L < M L<M L < M by assumption. Choose finite numbers a , t a,t a , t with L < a < t < M L<a<t<M L < a < t < M , and δ > 0 \delta>0 δ > 0 with u ≤ a u\leq a u ≤ a on B δ ( Ω ) B_\delta(\Omega) B δ ( Ω ) . Choose x 0 ∈ Ω x_0\in\Omega x 0 ∈ Ω with u ( x 0 ) > t u(x_0)>t u ( x 0 ) > t , and let V V V be the connected component of { u > t } \{u>t\} { u > t } containing x 0 x_0 x 0 . Every point of V V V satisfies u > a u>a u > a , hence lies outside B δ ( Ω ) B_\delta(\Omega) B δ ( Ω ) , so V ⊂ K δ V\subset K_\delta V ⊂ K δ , where
K δ = { x ∈ Ω : dist ( x , ∂ Ω ) ≥ δ } . K_\delta=\{x\in\Omega:\operatorname{dist}(x,\partial\Omega)\geq\delta\}. K δ = { x ∈ Ω : dist ( x , ∂ Ω ) ≥ δ } . The set K δ K_\delta K δ is compact. It is closed in R 2 \mathbb{R}^2 R 2 . Indeed, a limit of points of K δ K_\delta K δ is at distance at least δ \delta δ from ∂ Ω \partial\Omega ∂ Ω and lies in Ω ‾ \overline{\Omega} Ω . Since Ω ‾ ∖ Ω = ∂ Ω \overline{\Omega}\setminus\Omega=\partial\Omega Ω ∖ Ω = ∂ Ω , it lies in Ω \Omega Ω and hence in K δ K_\delta K δ . It is also bounded. If it were unbounded, it would contain a sequence of points with mutual distances at least δ \delta δ , and the disks of radius δ / 2 \delta/2 δ /2 about these points would be pairwise disjoint and contained in Ω \Omega Ω , contradicting ∣ Ω ∣ < ∞ |\Omega|<\infty ∣Ω∣ < ∞ . Hence V ‾ ⊂ K δ \overline{V}\subset K_\delta V ⊂ K δ is a compact subset of Ω \Omega Ω .
We claim that u = t u=t u = t on ∂ V \partial V ∂ V . By continuity, u ≥ t u\geq t u ≥ t on ∂ V \partial V ∂ V . If u ( y ) > t u(y)>t u ( y ) > t at some y ∈ ∂ V y\in\partial V y ∈ ∂ V , a small ball around y y y would lie in { u > t } \{u>t\} { u > t } and meet V V V , hence lie in V V V . This contradicts y ∈ ∂ V y\in\partial V y ∈ ∂ V because V V V is open.
Put η = u ( x 0 ) − t > 0 \eta=u(x_0)-t>0 η = u ( x 0 ) − t > 0 . For all large j j j we have V ‾ ⊂ Ω j \overline{V}\subset\Omega_j V ⊂ Ω j and, by local uniform convergence, sup V ‾ ∣ u j − u ∣ < η / 3 \sup_{\overline{V}}|u_j-u|<\eta/3 sup V ∣ u j − u ∣ < η /3 . Consequently
u j ( x 0 ) > t + 2 η / 3 , max ∂ V u j < t + η / 3. u_j(x_0)>t+2\eta/3,\qquad\max_{\partial V}u_j<t+\eta/3. u j ( x 0 ) > t + 2 η /3 , ∂ V max u j < t + η /3. The maximum of u j u_j u j on the compact set V ‾ \overline{V} V is therefore attained at a point of the open set V ⊂ Ω j V\subset\Omega_j V ⊂ Ω j , and that point is a local maximum of u j u_j u j in Ω j \Omega_j Ω j .
Proof of Theorem 1.2. Eigenfunctions are nonzero by definition, so we may take u u u normalized. Proposition 1.6 supplies r j ↑ 1 r_j\uparrow1 r j ↑ 1 and first positive eigenfunctions u j u_j u j on Ω r j \Omega_{r_j} Ω r j with u j → u u_j\to u u j → u locally uniformly in Ω \Omega Ω . If the first identity in (2) failed, Lemma 1.7, applied with Ω j = Ω r j \Omega_j=\Omega_{r_j} Ω j = Ω r j , would give an interior local maximum of a first positive Neumann eigenfunction on the bounded smooth simply connected domain Ω r j \Omega_{r_j} Ω r j , contradicting Hypothesis 1.1. Applying the same argument to − u -u − u proves the identity for the infimum.
Now let Ω \Omega Ω be bounded and u ∈ C ( Ω ‾ ) u \in C(\overline{\Omega}) u ∈ C ( Ω ) . By uniform continuity on Ω ‾ \overline{\Omega} Ω , lim sup δ ↓ 0 sup B δ ( Ω ) u ≤ max ∂ Ω u \limsup_{\delta\downarrow0}\sup_{B_\delta(\Omega)} u \le\max_{\partial\Omega} u lim sup δ ↓ 0 sup B δ ( Ω ) u ≤ max ∂ Ω u : each point of B δ ( Ω ) B_\delta(\Omega) B δ ( Ω ) lies within δ \delta δ of a boundary point. Conversely, a maximizing boundary point can be approached from inside Ω \Omega Ω , which gives the opposite inequality. Together with sup Ω u = max Ω ‾ u \sup_\Omega u = \max_{\overline{\Omega}}u sup Ω u = max Ω u , this yields the first identity in (3) . The second is analogous.
Heat-kernel bounds along the conformal exhaustion
Write p Ω ( t , x , y ) p_\Omega(t,x,y) p Ω ( t , x , y ) for the Neumann heat kernel of Ω \Omega Ω , with heat equation ∂ t u = Δ u \partial_t u = \Delta u ∂ t u = Δ u . A polynomial heat-kernel bound is a spatially uniform small-time estimate
0 ≤ p Ω ( t , x , y ) ≤ C 0 t − N / 2 ( 0 < t ≤ 1 ) (10) 0 \le p_\Omega(t,x,y) \le C_0 t^{-N/2} \qquad(0 < t \le1) \tag*{(10)} 0 ≤ p Ω ( t , x , y ) ≤ C 0 t − N /2 ( 0 < t ≤ 1 ) (10)
for almost every x , y ∈ Ω x,y \in\Omega x , y ∈ Ω and some constants N > 2 N > 2 N > 2 and C 0 < ∞ C_0 < \infty C 0 < ∞ . Here N N N is an exponent parameter, used in the Sobolev formulas below and in the stability theorems of [12 ] . The ambient dimension remains two. Bounded smooth and Lipschitz planar domains satisfy the sharp planar estimate C t − 1 Ct^{-1} C t − 1 , that is, (10) with N = 2 N = 2 N = 2 . Since t − 1 ≤ t − N / 2 t^{-1} \le t^{-N/2} t − 1 ≤ t − N /2 for 0 < t ≤ 1 0 < t \le1 0 < t ≤ 1 , they satisfy (10) for every N > 2 N > 2 N > 2 as well, and it is in this weakened form that the results of this section apply to them. Simple connectivity alone does not ensure any polynomial bound. Power cusps require exponents N > 2 N > 2 N > 2 (§ 6 ), and the exponential cusp of § 7.6 has no polynomial bound at all.
Uniform bounds along the conformal exhaustion
The transfer result in this subsection is independent of Hypothesis 1.1 and of simplicity of the eigenvalue. It supplies the uniform estimates that allow the use of the stability theorems in [12 ] .
We use the following standard fact (Sobolev–ultracontractivity). Let T t T_t T t be a symmetric sub-Markov semigroup associated with a nonnegative closed form E \mathcal{E} E with form domain D ( E ) D(\mathcal{E}) D ( E ) , and let N > 2 N > 2 N > 2 . If ∥ T t ∥ L 1 → L ∞ ≤ C t − N / 2 \|T_t\|_{L^1\to L^\infty} \le Ct^{-N/2} ∥ T t ∥ L 1 → L ∞ ≤ C t − N /2 for 0 < t ≤ 1 0 < t \le1 0 < t ≤ 1 , then
∥ w ∥ L q 2 ≤ S ( E ( w , w ) + ∥ w ∥ L 2 2 ) , q = 2 N N − 2 , w ∈ D ( E ) , (11) \|w\|_{L^q}^2 \le S\left(\mathcal{E}(w,w) + \|w\|_{L^2}^2\right), \qquad q = \frac{2N}{N-2}, \qquad w \in D(\mathcal{E}), \tag*{(11)} ∥ w ∥ L q 2 ≤ S ( E ( w , w ) + ∥ w ∥ L 2 2 ) , q = N − 2 2 N , w ∈ D ( E ) , (11)
with the constant S S S depending only on C C C and N N N . See [20 Theorem 2.4.2] . That theorem is stated for bounds valid for all t > 0 t > 0 t > 0 . To apply it, replace T t T_t T t by e − t T t e^{-t}T_t e − t T t , whose form is E + ∥ ⋅ ∥ L 2 2 \mathcal{E} + \|\cdot\|_{L^2}^2 E + ∥ ⋅ ∥ L 2 2 . For t ≥ 1 t \ge1 t ≥ 1 the semigroup property and the L 1 L^1 L 1 contraction property give ∥ e − t T t ∥ L 1 → L ∞ ≤ e − t ∥ T 1 ∥ L 1 → L ∞ ≤ C ′ t − N / 2 \|e^{-t}T_t\|_{L^1\to L^\infty} \le e^{-t}\|T_1\|_{L^1\to L^\infty} \le C't^{-N/2} ∥ e − t T t ∥ L 1 → L ∞ ≤ e − t ∥ T 1 ∥ L 1 → L ∞ ≤ C ′ t − N /2 . The converse implication needed below is proved explicitly by the Nash argument.
Proposition 2.1. Let Ω ⊂ R 2 \Omega\subset\mathbb{R}^2 Ω ⊂ R 2 be bounded and simply connected, and suppose that its Neumann heat kernel satisfies (10) for some constants N > 2 N > 2 N > 2 and C 0 < ∞ C_0 < \infty C 0 < ∞ . Let F : D → Ω F:\mathbb{D} \to\Omega F : D → Ω be a conformal bijection and Ω r = F ( r D ) \Omega_r = F(r\mathbb{D}) Ω r = F ( r D ) . Then there exist r 0 < 1 r_0 < 1 r 0 < 1 and C ∗ < ∞ C_* < \infty C ∗ < ∞ such that
0 ≤ p Ω r ( t , x , y ) ≤ C ∗ t − N / 2 ( r 0 < r < 1 , 0 < t ≤ 1 , x , y ∈ Ω r ) . (12) 0 \le p_{\Omega_r}(t,x,y) \le C_*t^{-N/2} \qquad(r_0 < r < 1,\ 0 < t \le1,\ x,y \in\Omega_r). \tag*{(12)} 0 ≤ p Ω r ( t , x , y ) ≤ C ∗ t − N /2 ( r 0 < r < 1 , 0 < t ≤ 1 , x , y ∈ Ω r ) . (12) The exponent is unchanged, and C ∗ C_* C ∗ is independent of r r r .
Proof. By (11) , the bound (10) gives
∥ w ∥ L q ( Ω ) 2 ≤ S ( ∫ Ω ∣ ∇ w ∣ 2 + ∫ Ω ∣ w ∣ 2 ) , q = 2 N N − 2 , w ∈ H 1 ( Ω ) . (13) \|w\|_{L^q(\Omega)}^2 \le S\left(\int_\Omega|\nabla w|^2 + \int_\Omega|w|^2\right), \qquad q = \frac{2N}{N-2}, \qquad w \in H^1(\Omega). \tag*{(13)} ∥ w ∥ L q ( Ω ) 2 ≤ S ( ∫ Ω ∣∇ w ∣ 2 + ∫ Ω ∣ w ∣ 2 ) , q = N − 2 2 N , w ∈ H 1 ( Ω ) . (13)
We transfer this inequality to Ω r \Omega_r Ω r without comparing heat kernels under inclusion. Step 1: absorption. Take v ∈ C ∞ ( Ω r ) v \in C^\infty(\Omega_r) v ∈ C ∞ ( Ω r ) and let V = Φ r v V=\Phi_rv V = Φ r v be the bounded extension of Lemma 1.5. Its construction and (6) do not use compactness. Put
m r = ∣ Ω ∖ Ω r ∣ , X = ∥ V ∥ L q ( Ω ) 2 , A r = ∥ v ∥ L 2 ( Ω r ) 2 , B r = ∥ ∇ v ∥ L 2 ( Ω r ) 2 . m_r=|\Omega\setminus\Omega_r|,\quad X=\|V\|_{L^q(\Omega)}^2,\quad A_r=\|v\|_{L^2(\Omega_r)}^2,\quad B_r=\|\nabla v\|_{L^2(\Omega_r)}^2. m r = ∣Ω ∖ Ω r ∣ , X = ∥ V ∥ L q ( Ω ) 2 , A r = ∥ v ∥ L 2 ( Ω r ) 2 , B r = ∥∇ v ∥ L 2 ( Ω r ) 2 .
All these quantities are finite since V V V is bounded and ∣ Ω ∣ < ∞ |\Omega|<\infty ∣Ω∣ < ∞ . Hölder’s inequality with 1 − 2 / q = 2 / N 1-2/q=2/N 1 − 2/ q = 2/ N gives
∫ Ω ∖ Ω r ∣ V ∣ 2 ≤ m r 2 / N X . (14) \int_{\Omega\setminus\Omega_r}|V|^2\leq m_r^{2/N}X. \tag*{(14)} ∫ Ω ∖ Ω r ∣ V ∣ 2 ≤ m r 2/ N X . (14)
Combining (13) , (6) and (14) , we obtain
X ≤ S ( 2 B r + A r + m r 2 / N X ) . X\leq S(2B_r+A_r+m_r^{2/N}X). X ≤ S ( 2 B r + A r + m r 2/ N X ) .
Since m r ↓ 0 m_r\downarrow0 m r ↓ 0 as r ↑ 1 r\uparrow1 r ↑ 1 , choose r 0 r_0 r 0 such that S m r 2 / N ≤ 1 2 Sm_r^{2/N}\leq\frac12 S m r 2/ N ≤ 2 1 for r > r 0 r>r_0 r > r 0 . Absorbing the last term,
∥ v ∥ L q ( Ω r ) 2 ≤ X ≤ 2 S ( 2 B r + A r ) ≤ 4 S ( ∥ ∇ v ∥ L 2 ( Ω r ) 2 + ∥ v ∥ L 2 ( Ω r ) 2 ) . (15) \|v\|_{L^q(\Omega_r)}^2\leq X\leq2S(2B_r+A_r)\leq4S\left(\|\nabla v\|_{L^2(\Omega_r)}^2+\|v\|_{L^2(\Omega_r)}^2\right). \tag*{(15)} ∥ v ∥ L q ( Ω r ) 2 ≤ X ≤ 2 S ( 2 B r + A r ) ≤ 4 S ( ∥∇ v ∥ L 2 ( Ω r ) 2 + ∥ v ∥ L 2 ( Ω r ) 2 ) . (15)
Because Ω r \Omega_r Ω r has smooth boundary, C ∞ ( Ω r ‾ ) C^\infty(\overline{\Omega_r}) C ∞ ( Ω r ) is dense in H 1 ( Ω r ) H^1(\Omega_r) H 1 ( Ω r ) . Applying (15) to differences of smooth approximants shows that they are Cauchy in L q L^q L q , and their L q L^q L q limit agrees with their L 2 L^2 L 2 limit. Hence (15) holds for all v ∈ H 1 ( Ω r ) v\in H^1(\Omega_r) v ∈ H 1 ( Ω r ) with the same constant K = 4 S K=4S K = 4 S , independent of r r r . Step 2: from Sobolev to the heat-kernel bound. Interpolation between L 1 L^1 L 1 and L q L^q L q gives
∥ v ∥ L 2 ≤ ∥ v ∥ L 1 2 / ( N + 2 ) ∥ v ∥ L q N / ( N + 2 ) , \|v\|_{L^2}\leq\|v\|_{L^1}^{2/(N+2)}\|v\|_{L^q}^{N/(N+2)}, ∥ v ∥ L 2 ≤ ∥ v ∥ L 1 2/ ( N + 2 ) ∥ v ∥ L q N / ( N + 2 ) ,
so (15) implies the Nash inequality
∥ v ∥ L 2 2 + 4 / N ≤ K ( ∥ ∇ v ∥ L 2 2 + ∥ v ∥ L 2 2 ) ∥ v ∥ L 1 4 / N . (16) \|v\|_{L^2}^{2+4/N}\leq K\left(\|\nabla v\|_{L^2}^2+\|v\|_{L^2}^2\right)\|v\|_{L^1}^{4/N}. \tag*{(16)} ∥ v ∥ L 2 2 + 4/ N ≤ K ( ∥∇ v ∥ L 2 2 + ∥ v ∥ L 2 2 ) ∥ v ∥ L 1 4/ N . (16)
All norms below are taken on Ω r \Omega_r Ω r . Let S t = e − t ( A Ω r + I ) S_t=e^{-t(A_{\Omega_r}+I)} S t = e − t ( A Ω r + I ) , an L 1 L^1 L 1 contraction, and let f ∈ L 1 ∩ L 2 f\in L^1\cap L^2 f ∈ L 1 ∩ L 2 with f ≠ 0 f\ne0 f = 0 . For y ( t ) = ∥ S t f ∥ L 2 2 y(t)=\|S_tf\|_{L^2}^2 y ( t ) = ∥ S t f ∥ L 2 2 , the spectral theorem and (16) give, for t > 0 t>0 t > 0 ,
y ′ ( t ) = − 2 ( ∥ ∇ S t f ∥ L 2 2 + ∥ S t f ∥ L 2 2 ) ≤ − 2 K ∥ f ∥ L 1 4 / N y ( t ) 1 + 2 / N . y'(t)=-2\left(\|\nabla S_tf\|_{L^2}^2+\|S_tf\|_{L^2}^2\right)\leq-\frac{2}{K\|f\|_{L^1}^{4/N}}y(t)^{1+2/N}. y ′ ( t ) = − 2 ( ∥∇ S t f ∥ L 2 2 + ∥ S t f ∥ L 2 2 ) ≤ − K ∥ f ∥ L 1 4/ N 2 y ( t ) 1 + 2/ N .
If y y y vanishes at some t 0 t_0 t 0 , it vanishes for all t ≥ t 0 t\geq t_0 t ≥ t 0 by the semigroup property and the bound below is trivial. Otherwise integration yields
y ( t ) ≤ ( N K 4 t ) N / 2 ∥ f ∥ L 1 2 . y(t)\leq\left(\frac{NK}{4t}\right)^{N/2}\|f\|_{L^1}^2. y ( t ) ≤ ( 4 t N K ) N /2 ∥ f ∥ L 1 2 .
By density, duality, and S t = S t / 2 S t / 2 S_t=S_{t/2}S_{t/2} S t = S t /2 S t /2 ,
∥ S t ∥ L 1 → L ∞ ≤ ( N K 2 t ) N / 2 . \|S_t\|_{L^1\to L^\infty}\leq\left(\frac{NK}{2t}\right)^{N/2}. ∥ S t ∥ L 1 → L ∞ ≤ ( 2 t N K ) N /2 .
Since T t Ω r = e t S t T_t^{\Omega_r}=e^tS_t T t Ω r = e t S t , for 0 < t ≤ 1 0<t\leq1 0 < t ≤ 1 we obtain
∥ T t Ω r ∥ L 1 → L ∞ ≤ e ( N K 2 ) N / 2 t − N / 2 . \|T_t^{\Omega_r}\|_{L^1\to L^\infty}\leq e\left(\frac{NK}{2}\right)^{N/2}t^{-N/2}. ∥ T t Ω r ∥ L 1 → L ∞ ≤ e ( 2 N K ) N /2 t − N /2 .
This is the kernel bound almost everywhere with C ∗ = e ( N K / 2 ) N / 2 C_*=e(NK/2)^{N/2} C ∗ = e ( N K /2 ) N /2 . Interior continuity of the heat kernel of the smooth domain Ω r \Omega_r Ω r extends it to every x , y ∈ Ω r x,y\in\Omega_r x , y ∈ Ω r .
A second proof via spectral stability
Corollary 2.4. Assume Hypothesis 1.1. Let Ω \Omega Ω be bounded and simply connected and satisfy (10) for some N > 2 N > 2 N > 2 . If its first positive Neumann eigenvalue is simple, then (2) holds for every real eigenfunction belonging to that eigenvalue, and (3) holds for those eigenfunctions that extend continuously to Ω ‾ \overline{\Omega} Ω .
By Remark 2.3, Corollary 2.4 is a special case of Theorem 1.2. It includes every target satisfying C t − 1 Ct^{-1} C t − 1 and having a simple first positive eigenvalue, by the choice of any fixed N > 2 N > 2 N > 2 . We give a second proof, which uses Proposition 2.1 together with the stability theory of [12 ] in place of Proposition 1.6. Its additional content is that the uniform bounds required by [12 ] follow from a bound on the target alone, for conformal exhaustions, by the extension and absorption argument rather than by monotonicity of Neumann kernels. Pang’s sequel [35 ] treats multiplicity and verifies uniform bounds for the usual polygonal approximation of the Koch snowflake.
Second proof of Corollary 2.4, via [12 ] . Proposition 2.1 supplies the bound p Ω r ( t , x , y ) ≤ C ∗ t − N / 2 p_{\Omega_r}(t,x,y) \le C_*t^{-N/2} p Ω r ( t , x , y ) ≤ C ∗ t − N /2 for r 0 < r < 1 r_0 < r < 1 r 0 < r < 1 and 0 < t ≤ 1 0 < t \le1 0 < t ≤ 1 , and (10) supplies the bound on Ω \Omega Ω . Put C 0 = max { 1 , C 0 , C ∗ } C_0 = \max\{1,C_0,C_*\} C 0 = max { 1 , C 0 , C ∗ } so that one constant serves Ω \Omega Ω and all approximants with r > r 0 r > r_0 r > r 0 , as [12 ] , Hypothesis 1.1 requires. We use the interior-continuous versions of these kernels, so the bounds hold at every interior pair of points, as in [18 Proposition 3] . The geometric condition of [12 ] is that the boundary of the approximant labeled by δ \delta δ lies in the δ \delta δ -layer of ∂ Ω \partial\Omega ∂ Ω . Put
d r = sup x ∈ ∂ Ω r dist ( x , ∂ Ω ) . d_r = \sup_{x\in\partial\Omega_r}\operatorname{dist}(x,\partial\Omega). d r = x ∈ ∂ Ω r sup dist ( x , ∂ Ω ) .
Then d r → 0 d_r \to0 d r → 0 : for ε > 0 \varepsilon> 0 ε > 0 the set K ε = { x ∈ Ω : dist ( x , ∂ Ω ) ≥ ε } K_\varepsilon= \{x \in\Omega: \operatorname{dist}(x, \partial\Omega) \ge\varepsilon\} K ε = { x ∈ Ω : dist ( x , ∂ Ω ) ≥ ε } is compact (see the proof of Lemma 1.7 ), so K ε ⊂ Ω r K_\varepsilon\subset\Omega_r K ε ⊂ Ω r for large r r r , whence ∂ Ω r ∩ K ε = ∅ \partial\Omega_r \cap K_\varepsilon= \varnothing ∂ Ω r ∩ K ε = ∅ and d r ≤ ε d_r \le\varepsilon d r ≤ ε . Choose r j ↑ 1 r_j \uparrow1 r j ↑ 1 with r j > r 0 r_j > r_0 r j > r 0 and d r j ≤ 1 / j d_{r_j} \le1/j d r j ≤ 1/ j , and label the approximants by Ω δ = Ω r j \Omega_\delta= \Omega_{r_j} Ω δ = Ω r j for 1 / j ≤ δ < 1 / ( j − 1 ) 1/j \le\delta< 1/(j-1) 1/ j ≤ δ < 1/ ( j − 1 ) , j ≥ 2 j \ge2 j ≥ 2 .
Theorems 1.3, 1.4 and 1.5 of [12 ] then give convergence of the first positive eigenvalues (indexed μ 2 \mu_2 μ 2 there), eventual simplicity of the first positive eigenvalue on the approximants, and uniform convergence on compact subsets of Ω \Omega Ω of the L 2 L^2 L 2 -normalized first eigenfunctions u j u_j u j of Ω r j \Omega_{r_j} Ω r j to the prescribed normalized eigenfunction u u u , once the sign of each u j u_j u j is fixed by ∫ Ω r j u j u > 0 \int_{\Omega_{r_j}} u_j u > 0 ∫ Ω r j u j u > 0 , as in the proof of [12 Theorem 1.5] . Two remarks on the scope of these theorems are in order. First, [12 Theorem 1.5] states uniform convergence on subdomains Ω 0 \Omega_0 Ω 0 of Ω \Omega Ω , and its proof uses that Ω ‾ 0 \overline{\Omega}_0 Ω 0 is a compact subset of Ω \Omega Ω . This is the form used here. Second, [12 Theorem 1.3] rests on its Lemma 2.2, a semigroup-convergence statement taken from Burdzy–Chen [16 ] ; Theorem 2 of [16 ] is proved for bounded targets with continuous boundary and increasing approximants with continuous boundaries, and its extension to targets without boundary regularity is described in [16 Section 3] . For a target with no boundary regularity this second proof therefore depends on that extension, whereas the first proof, through Theorem 1.2 , does not.
With the locally uniform convergence in hand, Lemma 1.7 applies exactly as in the proof of Theorem 1.2 .
Remark 2.5 (Relation to reflecting Brownian motion). Burdzy–Chen [16 ] prove weak convergence of reflecting Brownian motions along increasing-domain approximations, with the definitions and qualifications specified there. This supplies a semigroup-convergence input in the Bañuelos–Pang theory. Weak convergence alone does not provide the uniform heat-kernel bounds used here, nor does it by itself establish compact Sobolev embedding. § 4 instead uses energy estimates and compactness directly.
Domains with polynomial heat-kernel bounds
This section lists classes of domains satisfying (10) . The bounds concern 0 < t ≤ 1 0 < t \le1 0 < t ≤ 1 and allow constants depending on the domain. By Proposition 2.1 , a target bound with exponent N > 2 N > 2 N > 2 yields uniform bounds on its conformal exhaustion, and by Remark 2.3 it implies compact embedding. Simplicity of the first positive Neumann eigenvalue is a separate matter, treated in § 7 . Further classes (quasihyperbolic Hölder domains, finite decompositions, a slit disk) are collected in Appendix § 10 .
The Sobolev criterion
For a planar domain Ω \Omega Ω and 2 < q < ∞ 2 < q < \infty 2 < q < ∞ , the continuous embedding estimate
∥ v ∥ L q ( Ω ) 2 ≤ C ( ∥ ∇ v ∥ L 2 ( Ω ) 2 + ∥ v ∥ L 2 ( Ω ) 2 ) ( v ∈ H 1 ( Ω ) ) (17) \|v\|_{L^q(\Omega)}^2 \le C\left(\|\nabla v\|_{L^2(\Omega)}^2 + \|v\|_{L^2(\Omega)}^2\right) \qquad(v \in H^1(\Omega)) \tag*{(17)} ∥ v ∥ L q ( Ω ) 2 ≤ C ( ∥∇ v ∥ L 2 ( Ω ) 2 + ∥ v ∥ L 2 ( Ω ) 2 ) ( v ∈ H 1 ( Ω )) (17)
is equivalent to a polynomial Neumann heat-kernel bound with exponent N = 2 q / ( q − 2 ) > 2 N = 2q/(q - 2) > 2 N = 2 q / ( q − 2 ) > 2 . See [18 Proposition 3] , where this exponent is denoted M M M and N N N is the dimension. The condition is a continuous embedding into L q L^q L q , not merely compactness into L 2 L^2 L 2 . It is used in § 7.6 and in Remark 5.5 to show that certain domains have no polynomial bound.
Lipschitz domains, quasidisks, and Hölder graph domains
Bounded Lipschitz domains, including nonconvex polygons, satisfy p Ω ( t , x , y ) ≤ C t − 1 p_\Omega(t,x,y) \le Ct^{-1} p Ω ( t , x , y ) ≤ C t − 1 for 0 < t ≤ 1 0 < t \le1 0 < t ≤ 1 . See [12 Remark 1.2] .
A further class consists of bounded quasidisks, that is, images of the disk under quasiconformal homeomorphisms of the plane. They include the Koch snowflake domain. Quasidisks are uniform domains, and the Neumann heat kernel of a bounded uniform planar domain satisfies C t − 1 Ct^{-1} C t − 1 by the two-sided Gaussian estimates of Gyrya–Saloff-Coste [24 ] and Murugan [33 ] , although the boundary need not be rectifiable or locally representable as a graph. Here the small-scale volume bound ∣ B ( x , r ) ∩ Ω ∣ ≥ c r 2 |B(x,r)\cap\Omega|\ge cr^{2} ∣ B ( x , r ) ∩ Ω∣ ≥ c r 2 on a uniform planar domain converts the volume denominator in the Gaussian estimate into C t − 1 Ct^{-1} C t − 1 . The snowflake and its polygonal approximations are also treated in [35 Theorem 1.3] . For the weaker polynomial bound needed here it suffices that uniform domains have the Sobolev extension property [30 ] , which gives (17) for every finite q > 2 q>2 q > 2 .
Suppose Ω \Omega Ω is a bounded planar domain whose boundary is locally, after a rotation and translation, the graph of a Hölder continuous function of exponent 0 < γ < 1 0<\gamma<1 0 < γ < 1 , with Ω \Omega Ω lying locally on one side of the graph (the class Lip γ \operatorname{Lip}\gamma Lip γ of [18 ] ). Then [18 Proposition 3] gives
q = 2 ( 1 + γ ) 1 − γ , N = 1 + 1 γ , p Ω ( t , x , y ) ≤ C t − ( 1 + 1 / γ ) / 2 . (18) q=\frac{2(1+\gamma)}{1-\gamma},\qquad N=1+\frac{1}{\gamma},\qquad p_{\Omega}(t,x,y)\le Ct^{-(1+1/\gamma)/2}. \tag*{(18)} q = 1 − γ 2 ( 1 + γ ) , N = 1 + γ 1 , p Ω ( t , x , y ) ≤ C t − ( 1 + 1/ γ ) /2 . (18)
Power cusps
For α > 1 \alpha>1 α > 1 , consider the bounded simply connected domain
Ω α = { ( x , y ) : 0 < x < 1 , ∣ y ∣ < x α } . \Omega_{\alpha}=\{(x,y):0<x<1,\ |y|<x^{\alpha}\}. Ω α = {( x , y ) : 0 < x < 1 , ∣ y ∣ < x α } .
Near the tip its boundary is the graph x = ∣ y ∣ 1 / α x=|y|^{1/\alpha} x = ∣ y ∣ 1/ α , with the domain on the side x > ∣ y ∣ 1 / α x>|y|^{1/\alpha} x > ∣ y ∣ 1/ α . The remaining boundary is Lipschitz. Figure 3.1 compares the profile curves y = x α y=x^{\alpha} y = x α for several α \alpha α with the exponential profile treated in § 7 . Applying (18) with γ = 1 / α \gamma=1/\alpha γ = 1/ α gives
Figure 3.1. Profile curves y = w ( x ) y=w(x) y = w ( x ) , 0 < x < 1 0<x<1 0 < x < 1 , bounding the domains { 0 < x < 1 , ∣ y ∣ < w ( x ) } \{0<x<1,\ |y|<w(x)\} { 0 < x < 1 , ∣ y ∣ < w ( x )} . The power profiles w ( x ) = x α w(x)=x^{\alpha} w ( x ) = x α give the cusps Ω α \Omega_{\alpha} Ω α (α = 1 \alpha=1 α = 1 is a Lipschitz wedge). The profile e − 1 / x e^{-1/x} e − 1/ x is flatter than every power at the tip and leads to the exponential cusp of Theorem 4.10 .
p Ω α ( t , z , w ) ≤ C α t − ( α + 1 ) / 2 . (19) p_{\Omega_{\alpha}}(t,z,w)\le C_{\alpha}t^{-(\alpha+1)/2}. \tag*{(19)} p Ω α ( t , z , w ) ≤ C α t − ( α + 1 ) /2 . (19)
This is the planar specialization of [18 Example 6] , and the exponent is optimal for this family. The domain is not Lipschitz at the tip. These cusps also have positive quasihyperbolic order. See Appendix § 10 .
Cross-cut domains and simplicity of the first positive eigenvalue
All geometric, compactness, and spectral estimates in this section and § 8 are independent of Hypothesis 1.1. That hypothesis is used only for the hot-spots conclusions.
When the Sobolev embedding is compact, simplicity of λ 1 \lambda_1 λ 1 is precisely the inequality λ 1 < λ 2 \lambda_1 < \lambda_2 λ 1 < λ 2 . Heat-kernel bounds alone do not imply it. The disk and the square are counterexamples. In this section we verify it for a family of nonsmooth simply connected domains, all of which lie between two graphs. For such cross-cut domains the Neumann eigenvalues can be compared with those of an explicit weighted one-dimensional problem. This comparison replaces the compactness-based thin-domain arguments and gives explicit thresholds. Figure 4.1 shows the three families treated below, and Table 4.1 in § 7.9 summarizes the results.
Figure 4.1. The explicit domains of Section 4. (a) A Lipschitz tube T h T_h T h (Proposition 4.8) with a reentrant corner at ( 1 , 1 ) (1,1) ( 1 , 1 ) . (b) The quadratic cusp Q ε Q_\varepsilon Q ε (Theorem 4.9 ) and (c) the exponential cusp E ε E_\varepsilon E ε (Theorem 4.10 ), drawn with ε = 0.4 \varepsilon= 0.4 ε = 0.4 and ε = 1 \varepsilon= 1 ε = 1 so that the shapes are visible. The theorems concern the thinner domains ε ≤ 1 / 5 \varepsilon\le1/5 ε ≤ 1/5 and ε ≤ 1 / 10 \varepsilon\le1/10 ε ≤ 1/10 . Dashed lines mark y = 0 y = 0 y = 0 .
Table 4.1. The explicit examples of Section 4. All three are bounded, simply connected, have compact Sobolev embedding, and have a simple first positive Neumann eigenvalue.
Cross-cut domains
Definition 4.1. Let L > 0 L > 0 L > 0 and let a < b a < b a < b be locally Lipschitz functions on ( 0 , L ) (0,L) ( 0 , L ) . The cross-cut domain is
G a , b = { ( x , y ) : 0 < x < L , a ( x ) < y < b ( x ) } , G_{a,b} = \{(x,y) : 0 < x < L,\ a(x) < y < b(x)\}, G a , b = {( x , y ) : 0 < x < L , a ( x ) < y < b ( x )} ,
and ℓ = b − a \ell= b-a ℓ = b − a is its cross-cut length . The vertical segment over x x x is the cross-cut at x x x (Figure 4.2 ). We assume
Figure 4.2. Cross-cuts (dashed) of a cross-cut domain, here the profile h ( x ) = e − 1 / x h(x)=e^{-1/x} h ( x ) = e − 1/ x with a = − h a=-h a = − h , b = h b=h b = h . The marked cross-cut is at x 0 = 0.8 x_0=0.8 x 0 = 0.8 . The cross-cut over x x x is an interval of length ℓ ( x ) \ell(x) ℓ ( x ) . Its Neumann spectrum is ( n π / ℓ ( x ) ) 2 (n\pi/\ell(x))^2 ( nπ / ℓ ( x ) ) 2 , n ≥ 0 n\ge0 n ≥ 0 .
ℓ ∗ = sup 0 < x < L ℓ ( x ) < ∞ , κ ∗ 2 = ess sup 0 < x < L a ′ ( x ) 2 + a ′ ( x ) b ′ ( x ) + b ′ ( x ) 2 3 < ∞ . \ell_* = \sup_{0<x<L} \ell(x) < \infty,\qquad\kappa_*^2 = \operatorname*{ess\,sup}_{0<x<L}\frac{a'(x)^2+a'(x)b'(x)+b'(x)^2}{3} < \infty. ℓ ∗ = 0 < x < L sup ℓ ( x ) < ∞ , κ ∗ 2 = 0 < x < L ess sup 3 a ′ ( x ) 2 + a ′ ( x ) b ′ ( x ) + b ′ ( x ) 2 < ∞.
The map ( x , t ) ↦ ( x , a ( x ) + ℓ ( x ) t ) (x,t) \mapsto(x,a(x)+\ell(x)t) ( x , t ) ↦ ( x , a ( x ) + ℓ ( x ) t ) is a homeomorphism of ( 0 , L ) × ( 0 , 1 ) (0,L)\times(0,1) ( 0 , L ) × ( 0 , 1 ) onto G a , b G_{a,b} G a , b , so G a , b G_{a,b} G a , b is simply connected. The bound on κ ∗ \kappa_* κ ∗ implies global bounds on a ′ a' a ′ and b ′ b' b ′ , since a ′ 2 + a ′ b ′ + b ′ 2 ≥ ( a ′ 2 + b ′ 2 ) / 2 a'^2+a'b'+b'^2 \ge(a'^2+b'^2)/2 a ′2 + a ′ b ′ + b ′2 ≥ ( a ′2 + b ′2 ) /2 . Thus both functions extend to Lipschitz functions on [ 0 , L ] [0,L] [ 0 , L ] , and G a , b G_{a,b} G a , b is bounded. Note that κ ∗ ≤ max { ∥ a ′ ∥ ∞ , ∥ b ′ ∥ ∞ } \kappa_* \le\max\{\lVert a'\rVert_\infty,\lVert b'\rVert_\infty\} κ ∗ ≤ max {∥ a ′ ∥ ∞ , ∥ b ′ ∥ ∞ } . In the symmetric case a = − h a=-h a = − h , b = h b=h b = h one has ℓ = 2 h \ell=2h ℓ = 2 h and κ ∗ = ∥ h ′ ∥ ∞ / 3 \kappa_*=\lVert h'\rVert_\infty/\sqrt{3} κ ∗ = ∥ h ′ ∥ ∞ / 3 . For a sheared strip a = h a=h a = h , b = 1 + h b=1+h b = 1 + h one has ℓ = 1 \ell=1 ℓ = 1 and κ ∗ = ∥ h ′ ∥ ∞ \kappa_*=\lVert h'\rVert_\infty κ ∗ = ∥ h ′ ∥ ∞ .
The longitudinal problem is the following weighted one-dimensional problem. Let
V ℓ = { f ∈ H l o c 1 ( 0 , L ) : ∫ 0 L ℓ ( ∣ f ∣ 2 + ∣ f ′ ∣ 2 ) d x < ∞ } , \mathcal{V}_{\ell}=\left\{f\in H^1_{\mathrm{loc}}(0,L):\int_0^L \ell\left(|f|^2+|f'|^2\right)\,dx<\infty\right\}, V ℓ = { f ∈ H loc 1 ( 0 , L ) : ∫ 0 L ℓ ( ∣ f ∣ 2 + ∣ f ′ ∣ 2 ) d x < ∞ } ,
and let 0 = μ 0 ≤ μ 1 ≤ μ 2 ≤ ⋯ 0=\mu_0\le\mu_1\le\mu_2\le\cdots 0 = μ 0 ≤ μ 1 ≤ μ 2 ≤ ⋯ be the min–max values of
∫ 0 L ℓ ∣ f ′ ∣ 2 d x ∫ 0 L ℓ ∣ f ∣ 2 d x , f ∈ V ℓ , (20) \frac{\int_0^L \ell|f'|^2\,dx}{\int_0^L \ell|f|^2\,dx},\qquad f\in\mathcal{V}_{\ell}, \tag*{(20)} ∫ 0 L ℓ ∣ f ∣ 2 d x ∫ 0 L ℓ ∣ f ′ ∣ 2 d x , f ∈ V ℓ , (20)
defined over ( k + 1 ) (k+1) ( k + 1 ) -dimensional subspaces of V ℓ V_\ell V ℓ exactly as the λ k \lambda_k λ k in § 4 . No compactness is assumed in this definition. A function u ( x , y ) = f ( x ) u(x,y)=f(x) u ( x , y ) = f ( x ) belongs to H 1 ( G a , b ) H^1(G_{a,b}) H 1 ( G a , b ) if and only if f ∈ V ℓ f\in V_\ell f ∈ V ℓ , and then ∫ G a , b u 2 = ∫ 0 L ℓ f 2 d x \int_{G_{a,b}}u^2=\int_0^L\ell f^2\,dx ∫ G a , b u 2 = ∫ 0 L ℓ f 2 d x and ∫ G a , b ∣ ∇ u ∣ 2 = ∫ 0 L ℓ ∣ f ′ ∣ 2 d x \int_{G_{a,b}}|\nabla u|^2=\int_0^L\ell|f'|^2\,dx ∫ G a , b ∣∇ u ∣ 2 = ∫ 0 L ℓ ∣ f ′ ∣ 2 d x . If H 1 ( G a , b ) ↪ L 2 ( G a , b ) H^1(G_{a,b})\hookrightarrow L^2(G_{a,b}) H 1 ( G a , b ) ↪ L 2 ( G a , b ) is compact, then so is V ℓ ↪ L 2 ( ℓ d x ) V_\ell\hookrightarrow L^2(\ell\,dx) V ℓ ↪ L 2 ( ℓ d x ) , since y y y -independent functions form a closed subspace. In that case the μ k \mu_k μ k are eigenvalues of a self-adjoint operator.
Remark 4.2 (Cross-cut techniques in the literature). The cross-cut approach below is motivated by the classical use of cross-cuts in geometric function theory and probabilistic potential theory. Decomposing a planar domain along arcs joining boundary points plays an important role in estimates for conformal maps and harmonic measure. See [2 , 23 ] . Related geometric ideas enter estimates for Brownian exit times and the Dirichlet fundamental frequency. See for example [9 , 10 , 13 , 26 ] . For the Neumann problem considered here, the comparison uses the cross-cut decomposition directly, without a conformal map. On each vertical cross-cut, the transverse Neumann spectrum is explicit, while the slopes of the bounding graphs control the interaction between the longitudinal and transverse directions. This mechanism is closely related to dimension reduction in thin domains [25 , 29 ] and the analysis of Neumann Laplacians on cuspidal domains [27 ] .
The lip domains of Atar and Burdzy [4 ] form a subclass of the cross-cut domains considered here. Their theorem establishes simplicity of the first positive Neumann eigenvalue for every nonsquare lip domain, whereas the sufficient condition in Corollary 4.4 does not recover that full conclusion. Conversely, their theorem requires the domain itself to be Lipschitz and therefore does not cover the quadratic and exponential cusp domains treated below. For these non-Lipschitz examples, our comparison gives explicit spectral-gap bounds and sufficient parameter ranges for simplicity. See Theorems 4.9 and 4.10 . The results therefore overlap on lip domains satisfying Corollary 4.4, where simplicity is already known from [4 ] . For an earlier probabilistic approach to the hot spots conjecture, see [8 ] .
The cross-cut comparison principle
Proposition 4.3 (Cross-cut comparison). Let G = G a , b G=G_{a,b} G = G a , b be a cross-cut domain as in Definition 4.1, let λ k ( G ) \lambda_k(G) λ k ( G ) be its Neumann min–max values as in § 4 , and put c = ( ℓ ∗ / π ) 2 c=(\ell_*/\pi)^2 c = ( ℓ ∗ / π ) 2 . Then for every k ≥ 0 k\ge0 k ≥ 0 ,
μ k 1 + κ k 2 κ ∗ 2 + c μ k ≤ λ k ( G ) ≤ μ k . (21) \frac{\mu_k}{1+\frac{\kappa_k^2}{\kappa_*^2}+c\mu_k}\le\lambda_k(G)\le\mu_k. \tag*{(21)} 1 + κ ∗ 2 κ k 2 + c μ k μ k ≤ λ k ( G ) ≤ μ k . (21) No compactness is assumed.
Proof. Upper bound. The functions u ( x , y ) = f ( x ) u(x,y)=f(x) u ( x , y ) = f ( x ) , f ∈ V ℓ f\in V_\ell f ∈ V ℓ , lie in H 1 ( G ) H^1(G) H 1 ( G ) with the same Rayleigh quotient as f f f . Min–max over this subspace gives λ k ( G ) ≤ μ k \lambda_k(G)\le\mu_k λ k ( G ) ≤ μ k .
Lower bound. Fix k k k and η > 0 \eta> 0 η > 0 , and put Λ = λ k ( G ) + η \Lambda= \lambda_k(G) + \eta Λ = λ k ( G ) + η . By the min–max characterization there is a ( k + 1 ) (k + 1) ( k + 1 ) -dimensional subspace W ⊂ H 1 ( G ) W \subset H^1(G) W ⊂ H 1 ( G ) of real functions with ∫ G ∣ ∇ u ∣ 2 ≤ Λ ∫ G u 2 \int_G |\nabla u|^2 \leq\Lambda\int_G u^2 ∫ G ∣∇ u ∣ 2 ≤ Λ ∫ G u 2 for all u ∈ W u \in W u ∈ W . For u ∈ W u \in W u ∈ W write A = ∥ u x ∥ L 2 ( G ) A = \|u_x\|_{L^2(G)} A = ∥ u x ∥ L 2 ( G ) and B = ∥ u y ∥ L 2 ( G ) B = \|u_y\|_{L^2(G)} B = ∥ u y ∥ L 2 ( G ) , and define the cross-cut average and fluctuation
u ˉ ( x ) = 1 ℓ ( x ) ∫ a ( x ) b ( x ) u ( x , y ) d y , u ~ = u − u ˉ . \bar{u}(x) = \frac{1}{\ell(x)} \int_{a(x)}^{b(x)} u(x,y)\,\mathrm{d}y, \qquad\widetilde{u} = u - \bar{u}. u ˉ ( x ) = ℓ ( x ) 1 ∫ a ( x ) b ( x ) u ( x , y ) d y , u = u − u ˉ .
(1) Fluctuation. For almost every x x x , u ( x , ⋅ ) ∈ H 1 ( a ( x ) , b ( x ) ) u(x,\cdot) \in H^1(a(x),b(x)) u ( x , ⋅ ) ∈ H 1 ( a ( x ) , b ( x )) and u ~ ( x , ⋅ ) \widetilde{u}(x,\cdot) u ( x , ⋅ ) has mean zero. The first nonzero Neumann eigenvalue of an interval of length ℓ \ell ℓ is ( π / ℓ ) 2 (\pi/\ell)^2 ( π / ℓ ) 2 , so the Poincaré–Wirtinger inequality on each cross-cut gives
∫ a b ∣ u ~ ∣ 2 d y ≤ ( ℓ π ) 2 ∫ a b ∣ u y ∣ 2 d y , hence ∥ u ~ ∥ L 2 ( G ) 2 ≤ c B . \int_a^b |\widetilde{u}|^2\,\mathrm{d}y \leq\left(\frac{\ell}{\pi}\right)^2 \int_a^b |u_y|^2\,\mathrm{d}y, \qquad\text{hence} \qquad\|\widetilde{u}\|_{L^2(G)}^2 \leq cB. ∫ a b ∣ u ∣ 2 d y ≤ ( π ℓ ) 2 ∫ a b ∣ u y ∣ 2 d y , hence ∥ u ∥ L 2 ( G ) 2 ≤ c B .
Since u ˉ \bar{u} u ˉ and u ~ \widetilde{u} u are orthogonal on each cross-cut,
∥ u ∥ L 2 ( G ) 2 = ∫ 0 L ℓ u ˉ 2 d x + ∥ u ~ ∥ L 2 ( G ) 2 . (22) \|u\|_{L^2(G)}^2 = \int_0^L \ell\,\bar{u}^2\,\mathrm{d}x + \|\widetilde{u}\|_{L^2(G)}^2. \tag*{(22)} ∥ u ∥ L 2 ( G ) 2 = ∫ 0 L ℓ u ˉ 2 d x + ∥ u ∥ L 2 ( G ) 2 . (22)
(2) Averaged energy. Fix a compact interval J ⋐ ( 0 , L ) J \Subset(0,L) J ⋐ ( 0 , L ) . The map ( x , t ) ↦ ( x , a ( x ) + ℓ ( x ) t ) (x,t) \mapsto(x,a(x) + \ell(x)t) ( x , t ) ↦ ( x , a ( x ) + ℓ ( x ) t ) is bi-Lipschitz from J × ( 0 , 1 ) J \times(0,1) J × ( 0 , 1 ) onto the corresponding portion of G G G , because ℓ \ell ℓ is bounded away from zero on J J J . Thus, for u ∈ H 1 ( G ) u \in H^1(G) u ∈ H 1 ( G ) , its pullback v v v belongs to H 1 ( J × ( 0 , 1 ) ) H^1(J \times(0,1)) H 1 ( J × ( 0 , 1 )) . Sobolev slicing and the weak chain rule give u ˉ ( x ) = ∫ 0 1 v ( x , t ) d t \bar{u}(x) = \int_0^1 v(x,t)\,\mathrm{d}t u ˉ ( x ) = ∫ 0 1 v ( x , t ) d t and, almost everywhere,
v x = u x + ( a ′ + ℓ ′ t ) u y , v_x = u_x + (a' + \ell't)u_y, v x = u x + ( a ′ + ℓ ′ t ) u y ,
evaluated at ( x , a + ℓ t ) (x,a + \ell t) ( x , a + ℓ t ) .
Taking the weak derivative of this average,
u ˉ ′ ( x ) = ∫ 0 1 u x d t + ∫ 0 1 ( a ′ + ℓ ′ t ) u y d t . \bar{u}'(x) = \int_0^1 u_x\,\mathrm{d}t + \int_0^1 (a' + \ell't)u_y\,\mathrm{d}t. u ˉ ′ ( x ) = ∫ 0 1 u x d t + ∫ 0 1 ( a ′ + ℓ ′ t ) u y d t .
By the Cauchy–Schwarz inequality, ∣ ∫ 0 1 u x d t ∣ ≤ ℓ − 1 / 2 ∥ u x ( x , ⋅ ) ∥ L 2 ( a , b ) \left|\int_0^1 u_x\,\mathrm{d}t\right| \leq\ell^{-1/2}\|u_x(x,\cdot)\|_{L^2(a,b)} ∫ 0 1 u x d t ≤ ℓ − 1/2 ∥ u x ( x , ⋅ ) ∥ L 2 ( a , b ) . Since a ′ + ℓ ′ t a' + \ell't a ′ + ℓ ′ t is affine in t t t with end values a ′ a' a ′ and b ′ b' b ′ , we have ∫ 0 1 ( a ′ + ℓ ′ t ) 2 d t = ( a ′ 2 + a ′ b ′ + b ′ 2 ) / 3 ≤ κ ∗ 2 \int_0^1 (a' + \ell't)^2\,\mathrm{d}t = (a'^2 + a'b' + b'^2)/3 \leq\kappa_*^2 ∫ 0 1 ( a ′ + ℓ ′ t ) 2 d t = ( a ′2 + a ′ b ′ + b ′2 ) /3 ≤ κ ∗ 2 , and therefore
∣ ∫ 0 1 ( a ′ + ℓ ′ t ) u y d t ∣ ≤ κ ∗ ℓ − 1 / 2 ∥ u y ( x , ⋅ ) ∥ L 2 ( a , b ) . \left|\int_0^1 (a' + \ell't)u_y\,\mathrm{d}t\right| \leq\kappa_* \ell^{-1/2}\|u_y(x,\cdot)\|_{L^2(a,b)}. ∫ 0 1 ( a ′ + ℓ ′ t ) u y d t ≤ κ ∗ ℓ − 1/2 ∥ u y ( x , ⋅ ) ∥ L 2 ( a , b ) .
Consequently
ℓ ∣ u ˉ ′ ∣ ≤ ∥ u x ( x , ⋅ ) ∥ L 2 ( a , b ) + κ ∗ ∥ u y ( x , ⋅ ) ∥ L 2 ( a , b ) , \sqrt{\ell}\,|\bar{u}'| \leq\|u_x(x,\cdot)\|_{L^2(a,b)} + \kappa_*\|u_y(x,\cdot)\|_{L^2(a,b)}, ℓ ∣ u ˉ ′ ∣ ≤ ∥ u x ( x , ⋅ ) ∥ L 2 ( a , b ) + κ ∗ ∥ u y ( x , ⋅ ) ∥ L 2 ( a , b ) ,
and Minkowski’s inequality in L 2 ( d x ) L^2(\mathrm{d}x) L 2 ( d x ) together with the Cauchy–Schwarz inequality give
∫ 0 L ℓ ∣ u ˉ ′ ∣ 2 d x ≤ ( A + κ ∗ B ) 2 ≤ ( 1 + κ ∗ 2 ) ( A + B ) . (23) \int_0^L \ell|\bar{u}'|^2\,\mathrm{d}x \leq(\sqrt{A} + \kappa_*\sqrt{B})^2 \leq(1+\kappa_*^2)(A+B). \tag*{(23)} ∫ 0 L ℓ ∣ u ˉ ′ ∣ 2 d x ≤ ( A + κ ∗ B ) 2 ≤ ( 1 + κ ∗ 2 ) ( A + B ) . (23)
Since J J J was arbitrary, the derivative identity holds locally throughout ( 0 , L ) (0,L) ( 0 , L ) . Integration and monotone exhaustion by such intervals give (23) globally. Jensen’s inequality also gives ∫ 0 L ℓ u ˉ 2 ≤ ∥ u ∥ L 2 ( G ) 2 \int_0^L \ell\bar{u}^2 \leq\|u\|_{L^2(G)}^2 ∫ 0 L ℓ u ˉ 2 ≤ ∥ u ∥ L 2 ( G ) 2 . Hence u ˉ ∈ V ℓ \bar{u} \in V_\ell u ˉ ∈ V ℓ .
(3) Min–max. Suppose c Λ < 1 c\Lambda< 1 c Λ < 1 . For u ∈ W u \in W u ∈ W , (22) and B ≤ Λ ∥ u ∥ 2 B \leq\Lambda\|u\|^2 B ≤ Λ∥ u ∥ 2 give
∫ 0 L ℓ u ˉ 2 d x ≥ ( 1 − c Λ ) ∥ u ∥ L 2 ( G ) 2 > 0 ( u ≠ 0 ) , \int_0^L \ell\bar{u}^2\,\mathrm{d}x \geq(1-c\Lambda)\|u\|_{L^2(G)}^2 > 0 \qquad(u \ne0), ∫ 0 L ℓ u ˉ 2 d x ≥ ( 1 − c Λ ) ∥ u ∥ L 2 ( G ) 2 > 0 ( u = 0 ) ,
so u ↦ u ˉ u \mapsto\bar{u} u ↦ u ˉ is injective on W W W , and its image is a ( k + 1 ) (k+1) ( k + 1 ) -dimensional subspace of V ℓ V_\ell V ℓ . By (23) and A + B ≤ Λ ∥ u ∥ 2 A+B\leq\Lambda\lVert u\rVert^2 A + B ≤ Λ ∥ u ∥ 2 , every element of the image has Rayleigh quotient at most
( 1 + κ ∗ 2 ) Λ 1 − c Λ . \frac{(1+\kappa_*^2)\Lambda}{1-c\Lambda}. 1 − c Λ ( 1 + κ ∗ 2 ) Λ .
By min–max for the longitudinal problem, μ k ( 1 − c Λ ) ≤ ( 1 + κ ∗ 2 ) Λ \mu_k(1-c\Lambda)\leq(1+\kappa_*^2)\Lambda μ k ( 1 − c Λ ) ≤ ( 1 + κ ∗ 2 ) Λ , that is, Λ ≥ μ k / ( 1 + κ ∗ 2 + c μ k ) \Lambda\geq\mu_k/(1+\kappa_*^2+c\mu_k) Λ ≥ μ k / ( 1 + κ ∗ 2 + c μ k ) . If c Λ ≥ 1 c\Lambda\geq1 c Λ ≥ 1 , the same lower bound holds trivially because μ k / ( 1 + κ ∗ 2 + c μ k ) < 1 / c \mu_k/(1+\kappa_*^2+c\mu_k)<1/c μ k / ( 1 + κ ∗ 2 + c μ k ) < 1/ c . Letting η ↓ 0 \eta\downarrow0 η ↓ 0 proves (4.1).
Corollary 4.4 (Simplicity criterion). Assume that H 1 ( G a , b ) ↪ L 2 ( G a , b ) H^1(G_{a,b})\hookrightarrow L^2(G_{a,b}) H 1 ( G a , b ) ↪ L 2 ( G a , b ) is compact. If ( 1 + κ ∗ 2 ) μ 1 + c μ 1 μ 2 < μ 2 (1+\kappa_*^2)\mu_1+c\mu_1\mu_2<\mu_2 ( 1 + κ ∗ 2 ) μ 1 + c μ 1 μ 2 < μ 2 , then λ 1 ( G a , b ) < λ 2 ( G a , b ) \lambda_1(G_{a,b})<\lambda_2(G_{a,b}) λ 1 ( G a , b ) < λ 2 ( G a , b ) .
Proof. By (4.1), λ 1 ≤ μ 1 < μ 2 / ( 1 + κ ∗ 2 + c μ 2 ) ≤ λ 2 \lambda_1\leq\mu_1<\mu_2/(1+\kappa_*^2+c\mu_2)\leq\lambda_2 λ 1 ≤ μ 1 < μ 2 / ( 1 + κ ∗ 2 + c μ 2 ) ≤ λ 2 , where the middle inequality is the hypothesis.
Note that μ ↦ μ / ( 1 + κ ∗ 2 + c μ ) \mu\mapsto\mu/(1+\kappa_*^2+c\mu) μ ↦ μ / ( 1 + κ ∗ 2 + c μ ) is increasing, so in applications one may insert a lower bound for μ 2 \mu_2 μ 2 and an upper bound for μ 1 \mu_1 μ 1 .
Thin profiles
Let w : ( 0 , 1 ) → ( 0 , ∞ ) w:(0,1)\to(0,\infty) w : ( 0 , 1 ) → ( 0 , ∞ ) be locally Lipschitz with W 0 = sup w < ∞ W_0=\sup w<\infty W 0 = sup w < ∞ and W 1 = ess sup ∣ w ′ ∣ < ∞ W_1=\operatorname{ess\,sup}|w'|<\infty W 1 = ess sup ∣ w ′ ∣ < ∞ , and let
G ε w = { ( x , y ) : 0 < x < 1 , ∣ y ∣ < ε w ( x ) } , ε > 0. G_\varepsilon^w=\{(x,y):0<x<1,\ |y|<\varepsilon w(x)\},\qquad\varepsilon>0. G ε w = {( x , y ) : 0 < x < 1 , ∣ y ∣ < εw ( x )} , ε > 0.
This is the cross-cut domain with a = − ε w a=-\varepsilon w a = − εw , b = ε w b=\varepsilon w b = εw , so ℓ = 2 ε w \ell=2\varepsilon w ℓ = 2 εw , ℓ ∗ = 2 ε W 0 \ell_*=2\varepsilon W_0 ℓ ∗ = 2 ε W 0 and κ ∗ = ε W 1 / 3 \kappa_*=\varepsilon W_1/\sqrt{3} κ ∗ = ε W 1 / 3 . The longitudinal problem has the weight 2 ε w 2\varepsilon w 2 εw , and the constant 2 ε 2\varepsilon 2 ε cancels in the Rayleigh quotient. Hence its min–max values μ k = μ k ( w ) \mu_k=\mu_k(w) μ k = μ k ( w ) are independent of ε \varepsilon ε and are those of ∫ 0 1 w ∣ f ′ ∣ 2 \int_0^1w|f'|^2 ∫ 0 1 w ∣ f ′ ∣ 2 relative to ∫ 0 1 w f 2 \int_0^1wf^2 ∫ 0 1 w f 2 .
Corollary 4.5 (Thin profiles). With the notation above, for all ε > 0 \varepsilon>0 ε > 0 and k ≥ 0 k\geq0 k ≥ 0 ,
μ k 1 + ε 2 ( W 1 2 / 3 + ( 4 W 0 2 / π 2 ) μ k ) ≤ λ k ( G ε w ) ≤ μ k . (24) \frac{\mu_k}{1+\varepsilon^2\left(W_1^2/3+(4W_0^2/\pi^2)\mu_k\right)}\leq\lambda_k(G_\varepsilon^w)\leq\mu_k. \tag*{(24)} 1 + ε 2 ( W 1 2 /3 + ( 4 W 0 2 / π 2 ) μ k ) μ k ≤ λ k ( G ε w ) ≤ μ k . (24)
In particular λ k ( G ε w ) → μ k \lambda_k(G_\varepsilon^w)\to\mu_k λ k ( G ε w ) → μ k as ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 (Figure 4.3 ), and, provided H 1 ( G ε w ) ↪ L 2 ( G ε w ) H^1(G_\varepsilon^w)\hookrightarrow L^2(G_\varepsilon^w) H 1 ( G ε w ) ↪ L 2 ( G ε w ) is compact, Corollary 4.4 shows that its first positive eigenvalue is simple whenever
Figure 4.3. The vertical compression G 1 w → G ε w G^w_1 \to G^w_\varepsilon G 1 w → G ε w for w ( x ) = x 2 w(x)=x^2 w ( x ) = x 2 . As ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 the domain collapses onto the segment ( 0 , 1 ) (0,1) ( 0 , 1 ) , and by Corollary 4.5 each fixed-index Neumann eigenvalue converges to the corresponding weighted one-dimensional eigenvalue.
μ 1 ( 1 + ε 2 ( W 1 2 / 3 + ( 4 W 0 2 / π 2 ) μ 2 ) ) < μ 2 . (25) \mu_1\left(1+\varepsilon^2\left(W_1^2/3+(4W_0^2/\pi^2)\mu_2\right)\right)<\mu_2. \tag*{(25)} μ 1 ( 1 + ε 2 ( W 1 2 /3 + ( 4 W 0 2 / π 2 ) μ 2 ) ) < μ 2 . (25)
Proof. This is Proposition 4.3 with κ ∗ 2 = ε 2 W 1 2 / 3 \kappa_*^2=\varepsilon^2W_1^2/3 κ ∗ 2 = ε 2 W 1 2 /3 and c = ( 2 ε W 0 / π ) 2 c=(2\varepsilon W_0/\pi)^2 c = ( 2 ε W 0 / π ) 2 .
Lemma 4.6 (Simplicity of the longitudinal spectrum). Suppose in addition that w w w is positive and C 1 C^1 C 1 on ( 0 , 1 ] (0,1] ( 0 , 1 ] and that H 1 ( G 1 w ) ↪ L 2 ( G 1 w ) H^1(G_1^w)\hookrightarrow L^2(G_1^w) H 1 ( G 1 w ) ↪ L 2 ( G 1 w ) is compact. Then the longitudinal eigenvalues satisfy 0 = μ 0 < μ 1 < μ 2 < ⋯ 0=\mu_0<\mu_1<\mu_2<\cdots 0 = μ 0 < μ 1 < μ 2 < ⋯ , that is, every longitudinal eigenvalue is simple.
Proof. By the remark after Definition 4.1, the longitudinal problem has compact resolvent, so the μ k \mu_k μ k are eigenvalues. An eigenfunction f f f satisfies − ( w f ′ ) ′ = μ w f -(wf')'=\mu wf − ( w f ′ ) ′ = μ w f weakly in ( 0 , 1 ) (0,1) ( 0 , 1 ) . By local regularity it is a classical C 2 C^2 C 2 solution. Near x = 1 x=1 x = 1 one has w ≥ c > 0 w\geq c>0 w ≥ c > 0 , so f ∈ H 1 ( 1 − θ , 1 ) f\in H^1(1-\theta,1) f ∈ H 1 ( 1 − θ , 1 ) for some θ > 0 \theta>0 θ > 0 , and the equation ( w f ′ ) ′ = − μ w f ∈ L 2 ( 1 − θ , 1 ) (wf')'=-\mu wf\in L^2(1-\theta,1) ( w f ′ ) ′ = − μ w f ∈ L 2 ( 1 − θ , 1 ) gives w f ′ ∈ H 1 ( 1 − θ , 1 ) ⊂ C [ 1 − θ , 1 ] wf'\in H^1(1-\theta,1)\subset C[1-\theta,1] w f ′ ∈ H 1 ( 1 − θ , 1 ) ⊂ C [ 1 − θ , 1 ] . Hence f ∈ C 1 [ 1 − θ , 1 ] f\in C^1[1-\theta,1] f ∈ C 1 [ 1 − θ , 1 ] . Testing the weak equation against functions that are smooth up to x = 1 x=1 x = 1 and vanish near 0 0 0 , and integrating by parts, gives the natural condition f ′ ( 1 ) = 0 f'(1)=0 f ′ ( 1 ) = 0 , because w ( 1 ) > 0 w(1)>0 w ( 1 ) > 0 . Since w ∈ C 1 w\in C^1 w ∈ C 1 and w > 0 w>0 w > 0 near x = 1 x=1 x = 1 , the ODE is regular there, so ( f ( 1 ) , f ′ ( 1 ) ) (f(1),f'(1)) ( f ( 1 ) , f ′ ( 1 )) determines f f f . The solutions with f ′ ( 1 ) = 0 f'(1)=0 f ′ ( 1 ) = 0 therefore form a space of dimension at most one, and every eigenvalue is simple. Constants give the simple eigenvalue 0 0 0 , and μ 1 > 0 \mu_1>0 μ 1 > 0 since ( 0 , 1 ) (0,1) ( 0 , 1 ) is connected.
Theorem 4.7 (Simplicity for thin profiles). Under the hypotheses of Lemma 4.6, there is ε 0 > 0 \varepsilon_0>0 ε 0 > 0 such that for every 0 < ε < ε 0 0<\varepsilon<\varepsilon_0 0 < ε < ε 0 the domain G ε w G^w_\varepsilon G ε w is bounded and simply connected, has compact embedding H 1 ( G ε w ) ↪ L 2 ( G ε w ) H^1(G^w_\varepsilon)\hookrightarrow L^2(G^w_\varepsilon) H 1 ( G ε w ) ↪ L 2 ( G ε w ) , and has a simple first positive Neumann eigenvalue. One may take any ε 0 \varepsilon_0 ε 0 for which (25) holds on ( 0 , ε 0 ) (0,\varepsilon_0) ( 0 , ε 0 ) .
Proof. Compactness for G ε w G^w_\varepsilon G ε w follows from that of G 1 w G^w_1 G 1 w , because G ε w G^w_\varepsilon G ε w is the image of G 1 w G^w_1 G 1 w under the linear map ( x , y ) ↦ ( x , ε y ) (x,y)\mapsto(x,\varepsilon y) ( x , y ) ↦ ( x , ε y ) . By Lemma 4.6, μ 1 < μ 2 \mu_1<\mu_2 μ 1 < μ 2 , so (25) holds for all sufficiently small ε \varepsilon ε . Corollary 4.5 then gives λ 1 ( G ε w ) < λ 2 ( G ε w ) \lambda_1(G^w_\varepsilon)<\lambda_2(G^w_\varepsilon) λ 1 ( G ε w ) < λ 2 ( G ε w ) .
Example I: Lipschitz tubes
Proposition 4.8. Let h : [ 0 , 2 ] → R h:[0,2]\to\mathbb{R} h : [ 0 , 2 ] → R be Lipschitz with κ = ∥ h ′ ∥ ∞ < 2 \kappa=\lVert h'\rVert_\infty<\sqrt{2} κ = ∥ h ′ ∥ ∞ < 2 . Then
T h = { ( x , y ) : 0 < x < 2 , h ( x ) < y < 1 + h ( x ) } T_h=\{(x,y):0<x<2,\ h(x)<y<1+h(x)\} T h = {( x , y ) : 0 < x < 2 , h ( x ) < y < 1 + h ( x )}
is a bounded simply connected Lipschitz domain, with a polynomial Neumann heat-kernel bound C t − 1 Ct^{-1} C t − 1 and compact embedding, and its first positive Neumann eigenvalue is simple. More precisely,
λ 1 ( T h ) ≤ π 2 4 < π 2 2 + κ 2 ≤ λ 2 ( T h ) . \lambda_1(T_h)\le\frac{\pi^2}{4}<\frac{\pi^2}{2+\kappa^2}\le\lambda_2(T_h). λ 1 ( T h ) ≤ 4 π 2 < 2 + κ 2 π 2 ≤ λ 2 ( T h ) .
Proof. T h T_h T h is the cross-cut domain with a = h a=h a = h , b = 1 + h b=1+h b = 1 + h , L = 2 L=2 L = 2 , so ℓ ≡ 1 \ell\equiv1 ℓ ≡ 1 , ℓ ∗ = 1 \ell_*=1 ℓ ∗ = 1 and κ ∗ = κ \kappa_*=\kappa κ ∗ = κ . The weighted problem is the Neumann problem for − f ′ ′ -f'' − f ′′ on ( 0 , 2 ) (0,2) ( 0 , 2 ) , with μ k = ( k π / 2 ) 2 \mu_k=(k\pi/2)^2 μ k = ( k π /2 ) 2 . Thus μ 1 = π 2 / 4 \mu_1=\pi^2/4 μ 1 = π 2 /4 and μ 2 = π 2 \mu_2=\pi^2 μ 2 = π 2 . With c = 1 / π 2 c=1/\pi^2 c = 1/ π 2 , Proposition 4.3 gives
λ 1 ≤ π 2 4 , λ 2 ≥ π 2 1 + κ 2 + c π 2 = π 2 2 + κ 2 . \lambda_1\le\frac{\pi^2}{4},\qquad\lambda_2\ge\frac{\pi^2}{1+\kappa^2+c\pi^2}=\frac{\pi^2}{2+\kappa^2}. λ 1 ≤ 4 π 2 , λ 2 ≥ 1 + κ 2 + c π 2 π 2 = 2 + κ 2 π 2 .
and π 2 / ( 2 + κ 2 ) > π 2 / 4 \pi^2/(2+\kappa^2)>\pi^2/4 π 2 / ( 2 + κ 2 ) > π 2 /4 if and only if κ 2 < 2 \kappa^2<2 κ 2 < 2 . The domain T h T_h T h is Lipschitz for every finite κ \kappa κ : at each of the four corners a vertical side meets a graph of slope at most κ \kappa κ transversally, and in a suitably rotated frame the boundary near the corner is a Lipschitz graph. Hence the heat-kernel bound C t − 1 Ct^{-1} C t − 1 and compactness follow from § 6 . The restriction κ < 2 \kappa<\sqrt{2} κ < 2 is used only for the spectral gap.
For h ( x ) = ∣ x − 1 ∣ / 10 h(x)=|x-1|/10 h ( x ) = ∣ x − 1∣/10 one obtains a nonconvex polygon with a reentrant corner at ( 1 , 1 ) (1,1) ( 1 , 1 ) (Figure 4.1 (a)). Slope bounds strictly below 2 \sqrt{2} 2 are allowed, and no differentiability of h h h beyond Lipschitz continuity is required.
Example II: quadratic cusps
Theorem 4.9. For every 0 < ε ≤ 1 / 5 0 < \varepsilon\le1/5 0 < ε ≤ 1/5 , the domain
Q ε = { ( x , y ) : 0 < x < 1 , ∣ y ∣ < ε x 2 } Q_{\varepsilon} = \{(x,y): 0 < x < 1,\ |y| < \varepsilon x^{2}\} Q ε = {( x , y ) : 0 < x < 1 , ∣ y ∣ < ε x 2 } is bounded and simply connected, has a non-Lipschitz cusp (of quasihyperbolic order 1 / 2 1/2 1/2 in the sense of Appendix § 10 ), satisfies p Q ε ( t , z , w ) ≤ C ε t − 3 / 2 p_{Q_{\varepsilon}}(t,z,w) \le C_{\varepsilon}t^{-3/2} p Q ε ( t , z , w ) ≤ C ε t − 3/2 for 0 < t ≤ 1 0 < t \le1 0 < t ≤ 1 , has compact embedding, and has a simple first positive Neumann eigenvalue. More precisely,
λ 1 ( Q ε ) < 22.21 < 23.3 < λ 2 ( Q ε ) . \lambda_{1}(Q_{\varepsilon}) < 22.21 < 23.3 < \lambda_{2}(Q_{\varepsilon}). λ 1 ( Q ε ) < 22.21 < 23.3 < λ 2 ( Q ε ) .
Proof. The geometric statements and the heat-kernel bound are those of the power cusp Ω α \Omega_{\alpha} Ω α with α = 2 \alpha= 2 α = 2 in § 6.3 , since Q ε Q_{\varepsilon} Q ε is the image of Ω 2 \Omega_{2} Ω 2 under the bi-Lipschitz map ( x , y ) ↦ ( x , ε y ) (x,y) \mapsto(x,\varepsilon y) ( x , y ) ↦ ( x , ε y ) . Compactness follows from Remark 2.3 . For the spectrum, Q ε = G ε w Q_{\varepsilon} = G_{\varepsilon}^{w} Q ε = G ε w with w ( x ) = x 2 w(x) = x^{2} w ( x ) = x 2 , W 0 = 1 W_{0} = 1 W 0 = 1 and W 1 = 2 W_{1} = 2 W 1 = 2 .
The longitudinal problem. It is the problem for ∫ 0 1 x 2 ∣ f ′ ∣ 2 \int_{0}^{1}x^{2}|f'|^{2} ∫ 0 1 x 2 ∣ f ′ ∣ 2 relative to ∫ 0 1 x 2 f 2 \int_{0}^{1}x^{2}f^{2} ∫ 0 1 x 2 f 2 , that is, the radial Neumann problem on the unit ball B ⊂ R 3 \mathbb{B} \subset\mathbb{R}^{3} B ⊂ R 3 : for Ψ ( z ) = f ( ∣ z ∣ ) \Psi(z) = f(|z|) Ψ ( z ) = f ( ∣ z ∣ ) one has ∫ B ∣ ∇ Ψ ∣ 2 = 4 π ∫ 0 1 x 2 f ′ 2 d x \int_{\mathbb{B}}|\nabla\Psi|^{2} = 4\pi\int_{0}^{1}x^{2}f'^{2}\,dx ∫ B ∣∇Ψ ∣ 2 = 4 π ∫ 0 1 x 2 f ′ 2 d x and ∫ B Ψ 2 = 4 π ∫ 0 1 x 2 f 2 d x \int_{\mathbb{B}}\Psi^{2} = 4\pi\int_{0}^{1}x^{2}f^{2}\,dx ∫ B Ψ 2 = 4 π ∫ 0 1 x 2 f 2 d x . By Lemma 4.6 its eigenvalues are simple. An eigenfunction solves − ( x 2 f ′ ) ′ = μ x 2 f -(x^{2}f')' = \mu x^{2}f − ( x 2 f ′ ) ′ = μ x 2 f with f ′ ( 1 ) = 0 f'(1) = 0 f ′ ( 1 ) = 0 . For μ = k 2 > 0 \mu= k^{2} > 0 μ = k 2 > 0 the solutions are A sin ( k x ) / x + B cos ( k x ) / x A\sin(kx)/x + B\cos(kx)/x A sin ( k x ) / x + B cos ( k x ) / x , and finiteness of ∫ x 2 ∣ f ′ ∣ 2 \int x^{2}|f'|^{2} ∫ x 2 ∣ f ′ ∣ 2 forces B = 0 B = 0 B = 0 . The condition f ′ ( 1 ) = 0 f'(1) = 0 f ′ ( 1 ) = 0 for f = sin ( k x ) / x f = \sin(kx)/x f = sin ( k x ) / x reads k cos k = sin k k\cos k = \sin k k cos k = sin k , that is, tan k = k \tan k = k tan k = k . The function tan k − k \tan k - k tan k − k increases from − ∞ -\infty − ∞ to + ∞ +\infty + ∞ on each interval ( j π − π / 2 , j π + π / 2 ) (j\pi-\pi/2,j\pi+\pi/2) ( j π − π /2 , j π + π /2 ) and is negative on ( j π − π / 2 , j π ] (j\pi-\pi/2,j\pi] ( j π − π /2 , j π ] for j ≥ 1 j \ge1 j ≥ 1 , so the positive roots are k j ∈ ( j π , j π + π / 2 ) k_{j} \in(j\pi,j\pi+\pi/2) k j ∈ ( j π , j π + π /2 ) , j = 1 , 2 , … j = 1,2,\ldots j = 1 , 2 , … . Hence
μ 1 = k 1 2 < ( 3 π / 2 ) 2 < 22.21 , μ 2 = k 2 2 > ( 2 π ) 2 > 39.47. \mu_{1} = k_{1}^{2} < (3\pi/2)^{2} < 22.21,\qquad\mu_{2} = k_{2}^{2} > (2\pi)^{2} > 39.47. μ 1 = k 1 2 < ( 3 π /2 ) 2 < 22.21 , μ 2 = k 2 2 > ( 2 π ) 2 > 39.47.
The comparison. By Corollary 4.5, with W 1 2 / 3 = 4 / 3 W_{1}^{2}/3 = 4/3 W 1 2 /3 = 4/3 and 4 W 0 2 / π 2 = 4 / π 2 4W_{0}^{2}/\pi^{2} = 4/\pi^{2} 4 W 0 2 / π 2 = 4/ π 2 ,
λ 1 ≤ μ 1 < 22.21 , λ 2 ≥ μ 2 1 + ε 2 ( 4 / 3 + ( 4 / π 2 ) μ 2 ) ≥ 4 π 2 1 + ε 2 ( 4 / 3 + 16 ) = 4 π 2 1 + ( 52 / 3 ) ε 2 . \lambda_{1} \le\mu_{1} < 22.21,\qquad
\lambda_{2} \ge\frac{\mu_{2}}{1+\varepsilon^{2}(4/3+(4/\pi^{2})\mu_{2})} \ge\frac{4\pi^{2}}{1+\varepsilon^{2}(4/3+16)} = \frac{4\pi^{2}}{1+(52/3)\varepsilon^{2}}. λ 1 ≤ μ 1 < 22.21 , λ 2 ≥ 1 + ε 2 ( 4/3 + ( 4/ π 2 ) μ 2 ) μ 2 ≥ 1 + ε 2 ( 4/3 + 16 ) 4 π 2 = 1 + ( 52/3 ) ε 2 4 π 2 .
where we used that the left-hand function is increasing in μ 2 \mu_{2} μ 2 and inserted μ 2 > 4 π 2 \mu_{2} > 4\pi^{2} μ 2 > 4 π 2 . For ε ≤ 1 / 5 \varepsilon\le1/5 ε ≤ 1/5 the right-hand side is at least 4 π 2 / 1.6934 > 23.3 4\pi^{2}/1.6934 > 23.3 4 π 2 /1.6934 > 23.3 . Hence λ 1 < 22.21 < 23.3 < λ 2 \lambda_{1} < 22.21 < 23.3 < \lambda_{2} λ 1 < 22.21 < 23.3 < λ 2 . □
Numerically, the roots of tan k = k \tan k = k tan k = k give μ 1 ≈ 20.19 \mu_{1} \approx20.19 μ 1 ≈ 20.19 and μ 2 ≈ 59.68 \mu_{2} \approx59.68 μ 2 ≈ 59.68 . These values are not used in the proof.
Example III: an exponential cusp without a polynomial heat-kernel bound
Theorem 4.10. For every 0 < ε ≤ 1 / 10 0 < \varepsilon\le1/10 0 < ε ≤ 1/10 , the domain
E ε = { ( x , y ) : 0 < x < 1 , ∣ y ∣ < ε e − 1 / x } E_{\varepsilon} = \{(x,y): 0 < x < 1,\ |y| < \varepsilon e^{-1/x}\} E ε = {( x , y ) : 0 < x < 1 , ∣ y ∣ < ε e − 1/ x } is bounded and simply connected, has a non-Lipschitz cusp, and has compact embedding H 1 ( E ε ) ↪ L 2 ( E ε ) H^{1}(E_{\varepsilon}) \hookrightarrow L^{2}(E_{\varepsilon}) H 1 ( E ε ) ↪ L 2 ( E ε ) . Its Neumann eigenvalues, counted with multiplicity from λ 0 = 0 \lambda_{0} = 0 λ 0 = 0 , satisfy
λ 1 ( E ε ) < 31.3 < 32 < λ 2 ( E ε ) . \lambda_{1}(E_{\varepsilon}) < 31.3 < 32 < \lambda_{2}(E_{\varepsilon}). λ 1 ( E ε ) < 31.3 < 32 < λ 2 ( E ε ) . In particular, its first positive Neumann eigenvalue is simple. Nevertheless, there are no constants C , N < ∞ C,N < \infty C , N < ∞ such that p E ε ( t , z , w ) ≤ C t − N / 2 p_{E_{\varepsilon}}(t,z,w) \le Ct^{-N/2} p E ε ( t , z , w ) ≤ C t − N /2 for 0 < t ≤ 1 0 < t \le1 0 < t ≤ 1 and almost every z , w z,w z , w .
The proof combines a longitudinal spectral estimate with the cross-cut comparison. Compactness follows from the continuous-graph boundary criterion, while test functions near the tip rule out every polynomial heat-kernel bound.
Geometry and compact embedding
Geometry. Near the tip, the domain lies to the right of the continuous graph
x = g ε ( y ) , g ε ( y ) = 1 log ( ε / ∣ y ∣ ) ( y ≠ 0 ) , g ε ( 0 ) = 0 , x=g_{\varepsilon}(y),\qquad g_{\varepsilon}(y)=\frac{1}{\log(\varepsilon/\lvert y\rvert)}\quad(y\ne0),\qquad g_{\varepsilon}(0)=0, x = g ε ( y ) , g ε ( y ) = log ( ε / ∣ y ∣) 1 ( y = 0 ) , g ε ( 0 ) = 0 ,
and away from the tip the boundary is Lipschitz. The opening angle at the tip tends to zero, so the cusp is not Lipschitz. The map ( x , t ) ↦ ( x , ε e − 1 / x t ) (x,t)\mapsto(x,\varepsilon e^{-1/x}t) ( x , t ) ↦ ( x , ε e − 1/ x t ) is a homeomorphism of ( 0 , 1 ) × ( − 1 , 1 ) (0,1)\times(-1,1) ( 0 , 1 ) × ( − 1 , 1 ) onto E ε E_{\varepsilon} E ε , hence E ε E_{\varepsilon} E ε is bounded and simply connected.
Compact embedding. E ε E_{\varepsilon} E ε is a bounded domain whose boundary is locally the graph of a continuous function, and for such domains the embedding H 1 ↪ L 2 H^1\hookrightarrow L^2 H 1 ↪ L 2 is compact. See Burenkov [17 ] and the introduction of [18 ] .
The longitudinal spectral gap
Set w ( x ) = e − 1 / x w(x)=e^{-1/x} w ( x ) = e − 1/ x and consider the closed form
a [ f ] = ∫ 0 1 w ∣ f ′ ∣ 2 d x in L 2 ( ( 0 , 1 ) , w d x ) , a[f]=\int_0^1w\lvert f'\rvert^2\,dx\quad\text{in }L^2((0,1),w\,dx), a [ f ] = ∫ 0 1 w ∣ f ′ ∣ 2 d x in L 2 (( 0 , 1 ) , w d x ) ,
with domain V = { f ∈ H l o c 1 ( 0 , 1 ) : ∫ 0 1 w ( ∣ f ∣ 2 + ∣ f ′ ∣ 2 ) d x < ∞ } \mathcal{V}=\{f\in H^1_{\mathrm{loc}}(0,1):\int_0^1w(\lvert f\rvert^2+\lvert f'\rvert^2)\,dx<\infty\} V = { f ∈ H loc 1 ( 0 , 1 ) : ∫ 0 1 w (∣ f ∣ 2 + ∣ f ′ ∣ 2 ) d x < ∞ } . This is the longitudinal problem of § 7.1 for the profile w w w . Its compactness follows from the compact embedding for E 1 E_1 E 1 just established. Denote its eigenvalues by 0 = μ 0 ≤ μ 1 ≤ μ 2 ≤ ⋯ 0=\mu_0\le\mu_1\le\mu_2\le\cdots 0 = μ 0 ≤ μ 1 ≤ μ 2 ≤ ⋯ .
Lemma 4.11. For the profile w ( x ) = e − 1 / x w(x)=e^{-1/x} w ( x ) = e − 1/ x , the longitudinal eigenvalues satisfy
μ 1 < 31.3 , μ 2 > 4 π 2 − 27 4 > 32.7. \mu_1<31.3,\qquad\mu_2>4\pi^2-\frac{27}{4}>32.7. μ 1 < 31.3 , μ 2 > 4 π 2 − 4 27 > 32.7.
Proof. Upper bound. Write I j = ∫ 0 1 x j e − 1 / x d x I_j=\int_0^1x^j e^{-1/x}\,dx I j = ∫ 0 1 x j e − 1/ x d x . The trial function f ( x ) = x − I 1 / I 0 f(x)=x-I_1/I_0 f ( x ) = x − I 1 / I 0 is orthogonal to constants in L 2 ( w d x ) L^2(w\,dx) L 2 ( w d x ) , and ∫ 0 1 w ∣ f ′ ∣ 2 = I 0 \int_0^1w\lvert f'\rvert^2=I_0 ∫ 0 1 w ∣ f ′ ∣ 2 = I 0 , ∫ 0 1 w f 2 = I 2 − I 1 2 / I 0 \int_0^1wf^2=I_2-I_1^2/I_0 ∫ 0 1 w f 2 = I 2 − I 1 2 / I 0 , so
μ 1 ≤ I 0 I 2 − I 1 2 / I 0 . \mu_1\le\frac{I_0}{I_2-I_1^2/I_0}. μ 1 ≤ I 2 − I 1 2 / I 0 I 0 .
The following enclosures suffice:
0.14840 < I 0 < 0.1487 , 0.1096 < I 1 < 0.1098 , 0.0860 < I 2 < 0.0861. 0.14840<I_0<0.1487,\qquad0.1096<I_1<0.1098,\qquad0.0860<I_2<0.0861. 0.14840 < I 0 < 0.1487 , 0.1096 < I 1 < 0.1098 , 0.0860 < I 2 < 0.0861.
They imply
μ 1 ≤ 0.1487 0.0860 − ( 0.1098 ) 2 / 0.14840 < 31.3. \mu_1\le\frac{0.1487}{0.0860-(0.1098)^2/0.14840}<31.3. μ 1 ≤ 0.0860 − ( 0.1098 ) 2 /0.14840 0.1487 < 31.3.
For a reproducible evaluation of the enclosures, integration by parts gives
I 0 = e − 1 − E 1 ( 1 ) , I 1 = 1 2 E 1 ( 1 ) , I 2 = 1 3 ( e − 1 − I 1 ) , I_0=e^{-1}-E_1(1),\qquad I_1=\frac{1}{2}E_1(1),\qquad I_2=\frac{1}{3}(e^{-1}-I_1), I 0 = e − 1 − E 1 ( 1 ) , I 1 = 2 1 E 1 ( 1 ) , I 2 = 3 1 ( e − 1 − I 1 ) ,
where
E 1 ( 1 ) = ∫ 1 ∞ e − t t d t = − γ − ∑ m = 1 ∞ ( − 1 ) m m m ! . E_1(1)=\int_1^\infty\frac{e^{-t}}{t}\,dt=-\gamma-\sum_{m=1}^{\infty}\frac{(-1)^m}{m\,m!}. E 1 ( 1 ) = ∫ 1 ∞ t e − t d t = − γ − m = 1 ∑ ∞ m m ! ( − 1 ) m .
and γ \gamma γ is the Euler–Mascheroni constant. Only the bounds
0.5772 < γ < 0.5773 , 0.3678 < e − 1 < 0.3679 0.5772 < \gamma< 0.5773,\qquad0.3678 < e^{-1} < 0.3679 0.5772 < γ < 0.5773 , 0.3678 < e − 1 < 0.3679
are needed. To check the enclosures using rational arithmetic, put
S 10 = − ∑ m = 1 10 ( − 1 ) m m m ! , R = 1 11 ⋅ 11 ! . S_{10}=-\sum_{m=1}^{10}\frac{(-1)^m}{m\,m!},\qquad R=\frac{1}{11\cdot11!}. S 10 = − m = 1 ∑ 10 m m ! ( − 1 ) m , R = 11 ⋅ 11 ! 1 .
The alternating-series estimate gives
S 10 − 0.5773 < E 1 ( 1 ) < S 10 + R − 0.5772. S_{10}-0.5773<E_1(1)<S_{10}+R-0.5772. S 10 − 0.5773 < E 1 ( 1 ) < S 10 + R − 0.5772.
Substitution into the formulas for I 0 , I 1 , I 2 I_0,I_1,I_2 I 0 , I 1 , I 2 , together with the displayed bounds for e − 1 e^{-1} e − 1 yields the enclosures above. The tightest of them is the lower bound I 0 > 0.3678 − ( S 10 + R − 0.5772 ) > 0.14840 I_0>0.3678-(S_{10}+R-0.5772)>0.14840 I 0 > 0.3678 − ( S 10 + R − 0.5772 ) > 0.14840 .
Transformation. For the lower bound, put g = w f g=\sqrt{w}f g = w f . This map is unitary from L 2 ( w d x ) L^2(w\,dx) L 2 ( w d x ) to L 2 ( d x ) L^2(dx) L 2 ( d x ) . Since w ′ / w = x − 2 w'/w=x^{-2} w ′ / w = x − 2 , one has f ′ = w − 1 / 2 ( g ′ − g / ( 2 x 2 ) ) f'=w^{-1/2}(g'-g/(2x^2)) f ′ = w − 1/2 ( g ′ − g / ( 2 x 2 )) , so for real g g g smooth up to x = 1 x=1 x = 1 and vanishing near 0 0 0 ,
w ∣ f ′ ∣ 2 = ∣ g ′ − g 2 x 2 ∣ 2 = ∣ g ′ ∣ 2 − g g ′ x 2 + ∣ g ∣ 2 4 x 4 , ∫ 0 1 g g ′ x 2 d x = ∣ g ( 1 ) ∣ 2 2 + ∫ 0 1 ∣ g ∣ 2 x 3 d x , w|f'|^2=\left|g'-\frac{g}{2x^2}\right|^2=|g'|^2-\frac{gg'}{x^2}+\frac{|g|^2}{4x^4},\qquad\int_0^1\frac{gg'}{x^2}\,dx=\frac{|g(1)|^2}{2}+\int_0^1\frac{|g|^2}{x^3}\,dx, w ∣ f ′ ∣ 2 = g ′ − 2 x 2 g 2 = ∣ g ′ ∣ 2 − x 2 g g ′ + 4 x 4 ∣ g ∣ 2 , ∫ 0 1 x 2 g g ′ d x = 2 ∣ g ( 1 ) ∣ 2 + ∫ 0 1 x 3 ∣ g ∣ 2 d x ,
the second identity by integration by parts, since ( 1 / ( 2 x 2 ) ) ′ = − x − 3 (1/(2x^2))'=-x^{-3} ( 1/ ( 2 x 2 ) ) ′ = − x − 3 and the boundary term at 0 0 0 vanishes. Integrating the first identity gives
a [ f ] = ∫ 0 1 ( ∣ g ′ ∣ 2 + V ∣ g ∣ 2 ) d x − 1 2 ∣ g ( 1 ) ∣ 2 , V ( x ) = 1 4 x 4 − 1 x 3 ≥ − 27 4 . (26) a[f]=\int_0^1\left(|g'|^2+V|g|^2\right)\,dx-\frac{1}{2}|g(1)|^2,\qquad V(x)=\frac{1}{4x^4}-\frac{1}{x^3}\geq-\frac{27}{4}. \tag*{(26)} a [ f ] = ∫ 0 1 ( ∣ g ′ ∣ 2 + V ∣ g ∣ 2 ) d x − 2 1 ∣ g ( 1 ) ∣ 2 , V ( x ) = 4 x 4 1 − x 3 1 ≥ − 4 27 . (26)
The last bound follows by minimizing t 4 / 4 − t 3 t^4/4-t^3 t 4 /4 − t 3 over t = 1 / x ≥ 1 t=1/x\geq1 t = 1/ x ≥ 1 . The minimum occurs at t = 3 t=3 t = 3 .
Form domain. We justify (26) , as an inequality, on the whole form domain. First, truncating f ∈ V f\in\mathcal{V} f ∈ V to bounded functions gives convergence in the weighted form norm. A bounded function can then be cut off on ( 0 , r ) (0,r) ( 0 , r ) , with transition on ( r , 2 r ) (r,2r) ( r , 2 r ) : the added derivative cost is bounded by C ∥ f ∥ ∞ 2 r − 2 ∫ r 2 r w ( x ) d x → 0 C\|f\|_\infty^2r^{-2}\int_r^{2r}w(x)\,dx\to0 C ∥ f ∥ ∞ 2 r − 2 ∫ r 2 r w ( x ) d x → 0 , and the contribution of ( 0 , 2 r ) (0,2r) ( 0 , 2 r ) to ∫ w ( ∣ f ∣ 2 + ∣ f ′ ∣ 2 ) \int w(|f|^2+|f'|^2) ∫ w ( ∣ f ∣ 2 + ∣ f ′ ∣ 2 ) tends to zero by integrability. Smooth approximation away from zero therefore produces a form core of functions that are smooth up to x = 1 x=1 x = 1 and vanish near zero, on which (26) holds. The trace inequality
∣ g ( 1 ) ∣ 2 ≤ η ∥ g ′ ∥ 2 2 + C η ∥ g ∥ 2 2 |g(1)|^2\leq\eta\|g'\|_2^2+C_\eta\|g\|_2^2 ∣ g ( 1 ) ∣ 2 ≤ η ∥ g ′ ∥ 2 2 + C η ∥ g ∥ 2 2
and the lower bound for V V V show that the weighted form norm controls ∥ g ∥ H 1 \|g\|_{H^1} ∥ g ∥ H 1 . Passing to the closure, every transformed form-domain function belongs to H 1 ( 0 , 1 ) H^1(0,1) H 1 ( 0 , 1 ) with g ( 0 ) = 0 g(0)=0 g ( 0 ) = 0 , and
a [ f ] ≥ ∫ 0 1 ∣ g ′ ∣ 2 d x − 1 2 ∣ g ( 1 ) ∣ 2 − 27 4 ∥ g ∥ 2 2 . a[f]\geq\int_0^1|g'|^2\,dx-\frac{1}{2}|g(1)|^2-\frac{27}{4}\|g\|_2^2. a [ f ] ≥ ∫ 0 1 ∣ g ′ ∣ 2 d x − 2 1 ∣ g ( 1 ) ∣ 2 − 4 27 ∥ g ∥ 2 2 .
This inequality is all that is needed. No additional endpoint condition at zero is imposed on f f f .
Lower bound. The comparison form b [ g ] = ∫ 0 1 ∣ g ′ ∣ 2 − 1 2 ∣ g ( 1 ) ∣ 2 b[g]=\int_0^1|g'|^2-\frac{1}{2}|g(1)|^2 b [ g ] = ∫ 0 1 ∣ g ′ ∣ 2 − 2 1 ∣ g ( 1 ) ∣ 2 on { g ∈ H 1 ( 0 , 1 ) : g ( 0 ) = 0 } \{g\in H^1(0,1):g(0)=0\} { g ∈ H 1 ( 0 , 1 ) : g ( 0 ) = 0 } has the boundary conditions g ( 0 ) = 0 g(0)=0 g ( 0 ) = 0 , g ′ ( 1 ) = g ( 1 ) / 2 g'(1)=g(1)/2 g ′ ( 1 ) = g ( 1 ) /2 . It is strictly positive, since ∣ g ( 1 ) ∣ 2 ≤ ∫ 0 1 ∣ g ′ ∣ 2 |g(1)|^2\leq\int_0^1|g'|^2 ∣ g ( 1 ) ∣ 2 ≤ ∫ 0 1 ∣ g ′ ∣ 2 . Write its eigenvalues as 0 < ν 0 < ν 1 < ⋯ 0<\nu_0<\nu_1<\cdots 0 < ν 0 < ν 1 < ⋯ . Its eigenfunctions are sin ( k x ) \sin(kx) sin ( k x ) with tan k = 2 k \tan k=2k tan k = 2 k and eigenvalues ν j = k j 2 \nu_j=k_j^2 ν j = k j 2 , where k j k_j k j is the unique root of tan k = 2 k \tan k=2k tan k = 2 k in ( j π , j π + π / 2 ) (j\pi,j\pi+\pi/2) ( j π , j π + π /2 ) , j = 0 , 1 , 2 , … j=0,1,2,\ldots j = 0 , 1 , 2 , … . To verify uniqueness, note that tan k − 2 k \tan k-2k tan k − 2 k is strictly convex on each ( j π , j π + π / 2 ) (j\pi,j\pi+\pi/2) ( j π , j π + π /2 ) . For j ≥ 1 j\geq1 j ≥ 1 it is negative at the left endpoint and tends to + ∞ +\infty + ∞ at the right endpoint, so it has exactly one root. For j = 0 j=0 j = 0 it vanishes at zero and has derivative − 1 -1 − 1 there, so it again has exactly one root in the open interval. The value k 0 k_0 k 0 yields no nonzero eigenfunction. On ( j π − π / 2 , j π ) (j\pi-\pi/2,j\pi) ( j π − π /2 , j π ) it is negative for j ≥ 1 j\geq1 j ≥ 1 . In particular ν 2 > 4 π 2 \nu_2>4\pi^2 ν 2 > 4 π 2 . The min–max principle, which remains valid for the possibly smaller transformed form domain, gives μ 2 ≥ ν 2 − 27 / 4 > 4 π 2 − 27 / 4 \mu_2\geq\nu_2-27/4>4\pi^2-27/4 μ 2 ≥ ν 2 − 27/4 > 4 π 2 − 27/4 .
Numerically, μ 1 ≈ 24.37 \mu_1\approx24.37 μ 1 ≈ 24.37 and μ 2 ≈ 76.19 \mu_2\approx76.19 μ 2 ≈ 76.19 . These values are not used in the proof.
Proof of Theorem 4.10
Proof of Theorem 4.10 . Step 1: no polynomial heat-kernel bound. Fix ε > 0 \varepsilon>0 ε > 0 . Let u r u_r u r depend only on x x x , equal one for x ≤ r x\le r x ≤ r , decrease linearly to zero on ( r , r + r 2 ) (r,r+r^2) ( r , r + r 2 ) , and vanish beyond. The substitution s = 1 / x s=1/x s = 1/ x gives
∫ 0 r e − 1 / x d x = ∫ 1 / r ∞ e − s s − 2 d s ≍ r 2 e − 1 / r . (27) \int_0^r e^{-1/x}\,dx=\int_{1/r}^{\infty}e^{-s}s^{-2}\,ds\asymp r^2e^{-1/r}. \tag*{(27)} ∫ 0 r e − 1/ x d x = ∫ 1/ r ∞ e − s s − 2 d s ≍ r 2 e − 1/ r . (27)
The upper bound follows from s − 2 ≤ r 2 s^{-2}\le r^2 s − 2 ≤ r 2 on s ≥ 1 / r s\ge1/r s ≥ 1/ r , and the lower bound from integration over [ 1 / r , 1 / r + 1 ] [1/r,1/r+1] [ 1/ r , 1/ r + 1 ] . On ( r , r + r 2 ) (r,r+r^2) ( r , r + r 2 ) the weight e − 1 / x e^{-1/x} e − 1/ x lies between e − 1 / r e^{-1/r} e − 1/ r and e e − 1 / r e e^{-1/r} e e − 1/ r , and the squared gradient of u r u_r u r equals r − 4 r^{-4} r − 4 on this strip, whose area is O ε ( r 2 e − 1 / r ) O_{\varepsilon}(r^2e^{-1/r}) O ε ( r 2 e − 1/ r ) . Thus, for every q > 2 q>2 q > 2 ,
∥ u r ∥ L q ( E ε ) 2 ≥ c ε , q r 4 / q e − 2 / ( q r ) , ∥ u r ∥ H 1 ( E ε ) 2 ≤ C ε r − 2 e − 1 / r , \|u_r\|_{L^q(E_\varepsilon)}^2\ge c_{\varepsilon,q}r^{4/q}e^{-2/(qr)},\qquad\|u_r\|_{H^1(E_\varepsilon)}^2\le C_\varepsilon r^{-2}e^{-1/r}, ∥ u r ∥ L q ( E ε ) 2 ≥ c ε , q r 4/ q e − 2/ ( q r ) , ∥ u r ∥ H 1 ( E ε ) 2 ≤ C ε r − 2 e − 1/ r ,
and therefore
∥ u r ∥ L q 2 ∥ u r ∥ H 1 2 ≥ c ε , q r 2 + 4 / q e ( 1 − 2 / q ) / r ⟶ ∞ ( r ↓ 0 ) . \frac{\|u_r\|_{L^q}^2}{\|u_r\|_{H^1}^2}\ge c_{\varepsilon,q}r^{2+4/q}e^{(1-2/q)/r}\longrightarrow\infty\qquad(r\downarrow0). ∥ u r ∥ H 1 2 ∥ u r ∥ L q 2 ≥ c ε , q r 2 + 4/ q e ( 1 − 2/ q ) / r ⟶ ∞ ( r ↓ 0 ) .
Hence no continuous embedding of H 1 ( E ε ) H^1(E_\varepsilon) H 1 ( E ε ) into any L q L^q L q , q > 2 q>2 q > 2 , holds, so (17) fails for every q > 2 q>2 q > 2 . By the equivalence recorded in § 6 (see [18 Proposition 3] ) and by (11) , every polynomial heat-kernel bound with N > 2 N>2 N > 2 would imply (17) , so all such bounds are excluded. A bound with exponent at most two would imply one with a larger exponent and is excluded as well.
Step 2: the upper bound λ 1 ( E ε ) < 31.3 \lambda_1(E_\varepsilon)<31.3 λ 1 ( E ε ) < 31.3 . The domain E ε E_\varepsilon E ε is the thin-profile domain G ε w G_\varepsilon^w G ε w with w ( x ) = e − 1 / x w(x)=e^{-1/x} w ( x ) = e − 1/ x , W 0 = e − 1 W_0=e^{-1} W 0 = e − 1 and W 1 = sup 0 < x < 1 e − 1 / x x − 2 = 4 e − 2 W_1=\sup_{0<x<1}e^{-1/x}x^{-2}=4e^{-2} W 1 = sup 0 < x < 1 e − 1/ x x − 2 = 4 e − 2 (attained at x = 1 / 2 x=1/2 x = 1/2 ). By the upper bound in (24) and Lemma 4.11 , λ 1 ( E ε ) ≤ μ 1 < 31.3 \lambda_1(E_\varepsilon)\le\mu_1<31.3 λ 1 ( E ε ) ≤ μ 1 < 31.3 .
Step 3: the lower bound λ 2 ( E ε ) > 32 \lambda_2(E_\varepsilon)>32 λ 2 ( E ε ) > 32 . The profile constants satisfy
W 1 2 3 = 16 e − 4 3 < 0.1 , 4 W 0 2 π 2 = 4 e 2 π 2 < 0.055. \frac{W_1^2}{3}=\frac{16e^{-4}}{3}<0.1,\qquad\frac{4W_0^2}{\pi^2}=\frac{4}{e^2\pi^2}<0.055. 3 W 1 2 = 3 16 e − 4 < 0.1 , π 2 4 W 0 2 = e 2 π 2 4 < 0.055.
For fixed ε > 0 \varepsilon>0 ε > 0 , the function
μ ⟼ μ 1 + ε 2 ( W 1 2 / 3 + ( 4 W 0 2 / π 2 ) μ ) \mu\longmapsto\frac{\mu}{1+\varepsilon^2(W_1^2/3+(4W_0^2/\pi^2)\mu)} μ ⟼ 1 + ε 2 ( W 1 2 /3 + ( 4 W 0 2 / π 2 ) μ ) μ
is increasing. Hence (24) and Lemma 4.11 give
λ 2 ( E ε ) > 32.7 1 + ε 2 ( 0.1 + 0.055 ⋅ 32.7 ) . \lambda_2(E_\varepsilon)>\frac{32.7}{1+\varepsilon^2(0.1+0.055\cdot32.7)}. λ 2 ( E ε ) > 1 + ε 2 ( 0.1 + 0.055 ⋅ 32.7 ) 32.7 .
If 0 < ε ≤ 1 / 10 0<\varepsilon\le1/10 0 < ε ≤ 1/10 , the denominator is less than 1.02 1.02 1.02 , so
λ 2 ( E ε ) > 32.7 1.02 > 32. \lambda_2(E_\varepsilon)>\frac{32.7}{1.02}>32. λ 2 ( E ε ) > 1.02 32.7 > 32.
Combining this with Step 2 gives
λ 1 ( E ε ) < 31.3 < 32 < λ 2 ( E ε ) , 0 < ε ≤ 1 / 10 , \lambda_1(E_\varepsilon)<31.3<32<\lambda_2(E_\varepsilon),\qquad0<\varepsilon\le1/10, λ 1 ( E ε ) < 31.3 < 32 < λ 2 ( E ε ) , 0 < ε ≤ 1/10 ,
which establishes simplicity and completes the proof. □
Remark 4.12 (Scope of the numerical bounds). The proofs of Theorems 4.9 and 4.10 use explicit scalar enclosures, not numerical eigenvalue approximations. The ranges ε ≤ 1 / 5 \varepsilon\le1/5 ε ≤ 1/5 and ε ≤ 1 / 10 \varepsilon\le1/10 ε ≤ 1/10 are chosen to keep the enclosures and the final spectral gaps simple. They are sufficient and are not asserted to be optimal.
Summary of the examples
Table 4.1 collects the three families and the bounds proved for them.
Corollary 4.13. Assume Hypothesis 1.1. The hot-spots conclusions (2) and, for eigenfunctions continuous up to the boundary, (3) hold for the Lipschitz tubes T h T_h T h with ∥ h ′ ∥ ∞ < 2 \lVert h'\rVert_\infty< \sqrt{2} ∥ h ′ ∥ ∞ < 2 , for the quadratic cusps Q ε Q_\varepsilon Q ε with 0 < ε ≤ 1 / 5 0 < \varepsilon\le1/5 0 < ε ≤ 1/5 , and for the exponential cusps E ε E_\varepsilon E ε with 0 < ε ≤ 1 / 10 0 < \varepsilon\le1/10 0 < ε ≤ 1/10 , for example for
{ ( x , y ) : 0 < x < 1 , ∣ y ∣ < 10 − 2 e − 1 / x } . \{(x,y): 0 < x < 1,\ \lvert y\rvert< 10^{-2}e^{-1/x}\}. {( x , y ) : 0 < x < 1 , ∣ y ∣ < 1 0 − 2 e − 1/ x } . The last family satisfies compact embedding and simplicity but lies outside every class defined by a polynomial Neumann heat-kernel upper bound. Hence for it Theorem 1.2 applies, whereas Corollary 2.4 does not.
Proof. Each family consists of bounded simply connected domains with compact embedding and a simple first positive eigenvalue (Proposition 4.8 and Theorems 4.9 and 4.10 ), so Theorem 1.2 applies.
Unbounded domains of finite area
Theorem 1.2 allows unbounded domains of finite area. This permits examples whose boundary recedes to infinity while the cross-cuts shrink rapidly. The Gaussian horn below has compact Sobolev embedding and, for an explicit range of widths, a simple first positive Neumann eigenvalue. Sufficiently thin horns with profile e − x α e^{-x^\alpha} e − x α , α ≥ 2 \alpha\ge2 α ≥ 2 , have the same properties (Remark 5.4). None of these domains admits a polynomial Neumann heat-kernel bound (Remark 5.5). The final subsection gives a different unbounded example, consisting of square rooms joined by rectangular passages, for which compactness and simplicity follow from elementary estimates.
The Gaussian horn
Figure 5.1 shows the domain considered in the following theorem.
Theorem 5.1 (The Gaussian horn). For ε > 0 \varepsilon> 0 ε > 0 , let
G ε = { ( x , y ) : x > 0 , ∣ y ∣ < ε e − x 2 } . G_\varepsilon= \{(x,y): x > 0,\ \lvert y\rvert< \varepsilon e^{-x^2}\}. G ε = {( x , y ) : x > 0 , ∣ y ∣ < ε e − x 2 } . Then G ε G_\varepsilon G ε is unbounded and simply connected, ∣ G ε ∣ = ε π \lvert G_\varepsilon\rvert= \varepsilon\sqrt{\pi} ∣ G ε ∣ = ε π , and H 1 ( G ε ) ↪ L 2 ( G ε ) H^1(G_\varepsilon) \hookrightarrow L^2(G_\varepsilon) H 1 ( G ε ) ↪ L 2 ( G ε ) is compact. Its first two positive Neumann eigenvalues satisfy
λ 1 ( G ε ) ≤ 4 , λ 2 ( G ε ) ≥ 8 1 + ε 2 ( 2 / ( 3 e ) + 32 / π 2 ) . (28) \lambda_1(G_\varepsilon) \le4,\qquad\lambda_2(G_\varepsilon) \ge\frac{8}{1+\varepsilon^2(2/(3e)+32/\pi^2)}. \tag*{(28)} λ 1 ( G ε ) ≤ 4 , λ 2 ( G ε ) ≥ 1 + ε 2 ( 2/ ( 3 e ) + 32/ π 2 ) 8 . (28) In particular, λ 1 ( G ε ) \lambda_1(G_\varepsilon) λ 1 ( G ε ) is simple whenever 0 < ε ≤ 1 2 0 < \varepsilon\le\frac{1}{2} 0 < ε ≤ 2 1 .
Figure 5.1. The Gaussian horn G ε = { ( x , y ) : x > 0 , ∣ y ∣ < ε e − x 2 } G_\varepsilon=\{(x,y):x>0,\ |y|<\varepsilon e^{-x^2}\} G ε = {( x , y ) : x > 0 , ∣ y ∣ < ε e − x 2 } . Dashed curves indicate continuation to infinity.
Proof of Theorem 5.1 . Write w ( x ) = e − x 2 w(x)=e^{-x^2} w ( x ) = e − x 2 . Note that w w w is decreasing on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) . The map ( x , t ) ↦ ( x , ε w ( x ) t ) (x,t)\mapsto(x,\varepsilon w(x)t) ( x , t ) ↦ ( x , εw ( x ) t ) is a homeomorphism from ( 0 , ∞ ) × ( − 1 , 1 ) (0,\infty)\times(-1,1) ( 0 , ∞ ) × ( − 1 , 1 ) onto G ε G_\varepsilon G ε , proving simple connectivity. Integration gives the stated area.
Cross-cut estimates. For u ∈ H 1 ( G ε ) u\in H^1(G_\varepsilon) u ∈ H 1 ( G ε ) , define its cross-cut average, as in § 7 , by
u ˉ ( x ) = 1 2 ∫ − 1 1 u ( x , ε w ( x ) t ) d t . \bar{u}(x)=\frac{1}{2}\int_{-1}^{1}u(x,\varepsilon w(x)t)\,dt. u ˉ ( x ) = 2 1 ∫ − 1 1 u ( x , εw ( x ) t ) d t .
Put A = ∫ G ε ∣ u x ∣ 2 A=\int_{G_\varepsilon}|u_x|^2 A = ∫ G ε ∣ u x ∣ 2 , B = ∫ G ε ∣ u y ∣ 2 B=\int_{G_\varepsilon}|u_y|^2 B = ∫ G ε ∣ u y ∣ 2 , W 0 = sup w = 1 W_0=\sup w=1 W 0 = sup w = 1 , and W 1 = sup ∣ w ′ ∣ = 2 / e W_1=\sup|w'|=\sqrt{2/e} W 1 = sup ∣ w ′ ∣ = 2/ e . Jensen’s inequality gives (29) , and differentiation of the average with the Cauchy–Schwarz inequality gives (30) :
2 ε ∫ 0 ∞ w ∣ u ˉ ∣ 2 d x ≤ ∥ u ∥ 2 2 , (29) 2\varepsilon\int_0^\infty w|\bar{u}|^2\,dx\leq\|u\|_2^2, \tag*{(29)} 2 ε ∫ 0 ∞ w ∣ u ˉ ∣ 2 d x ≤ ∥ u ∥ 2 2 , (29)
2 ε ∫ 0 ∞ w ∣ u ˉ ′ ∣ 2 d x ≤ ( A + ε W 1 3 B ) 2 ≤ ( 1 + ε 2 W 1 2 / 3 ) ( A + B ) . (30) 2\varepsilon\int_0^\infty w|\bar{u}'|^2\,dx\leq\left(\sqrt{A}+\frac{\varepsilon W_1}{\sqrt{3}}\sqrt{B}\right)^2\leq(1+\varepsilon^2W_1^2/3)(A+B). \tag*{(30)} 2 ε ∫ 0 ∞ w ∣ u ˉ ′ ∣ 2 d x ≤ ( A + 3 ε W 1 B ) 2 ≤ ( 1 + ε 2 W 1 2 /3 ) ( A + B ) . (30)
The constant 1 / 3 1/3 1/3 is the average of t 2 t^2 t 2 on ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) . The Neumann Poincaré inequality on each cross-cut, whose length 2 ε w ( x ) 2\varepsilon w(x) 2 εw ( x ) is at most 2 ε w ( R ) 2\varepsilon w(R) 2 εw ( R ) for x > R x>R x > R because w w w is decreasing, gives
∫ { x > R } ∩ G ε ∣ u − u ˉ ( x ) ∣ 2 ≤ 4 ε 2 w ( R ) 2 π 2 ∫ { x > R } ∩ G ε ∣ u y ∣ 2 , R ≥ 0. (31) \int_{\{x>R\}\cap G_\varepsilon}|u-\bar{u}(x)|^2\leq\frac{4\varepsilon^2w(R)^2}{\pi^2}\int_{\{x>R\}\cap G_\varepsilon}|u_y|^2,\qquad R\geq0. \tag*{(31)} ∫ { x > R } ∩ G ε ∣ u − u ˉ ( x ) ∣ 2 ≤ π 2 4 ε 2 w ( R ) 2 ∫ { x > R } ∩ G ε ∣ u y ∣ 2 , R ≥ 0. (31)
These statements follow first on finite truncations, where the change of coordinates is regular, and then by Sobolev approximation and monotone convergence. On each cross-cut, the average and fluctuation are orthogonal in L 2 L^2 L 2 .
Compact embedding. Let V w = { f ∈ H l o c 1 ( 0 , ∞ ) : ∫ 0 ∞ w ( ∣ f ∣ 2 + ∣ f ′ ∣ 2 ) d x < ∞ } V_w=\{f\in H^1_{\mathrm{loc}}(0,\infty):\int_0^\infty w(|f|^2+|f'|^2)\,dx<\infty\} V w = { f ∈ H loc 1 ( 0 , ∞ ) : ∫ 0 ∞ w ( ∣ f ∣ 2 + ∣ f ′ ∣ 2 ) d x < ∞ } be the weighted form domain. The unitary substitution g = w f = e − x 2 / 2 f g=\sqrt{w}f=e^{-x^2/2}f g = w f = e − x 2 /2 f gives
∫ 0 ∞ w ∣ f ′ ∣ 2 d x = ∫ 0 ∞ ( ∣ g ′ ∣ 2 + ( x 2 − 1 ) ∣ g ∣ 2 ) d x . (32) \int_0^\infty w|f'|^2\,dx=\int_0^\infty(|g'|^2+(x^2-1)|g|^2)\,dx. \tag*{(32)} ∫ 0 ∞ w ∣ f ′ ∣ 2 d x = ∫ 0 ∞ ( ∣ g ′ ∣ 2 + ( x 2 − 1 ) ∣ g ∣ 2 ) d x . (32)
For smooth functions of bounded support this is integration by parts, since w f ′ 2 = ( g ′ + x g ) 2 = g ′ 2 + x ( g 2 ) ′ + x 2 g 2 wf'^2=(g'+xg)^2=g'^2+x(g^2)'+x^2g^2 w f ′2 = ( g ′ + xg ) 2 = g ′2 + x ( g 2 ) ′ + x 2 g 2 and the boundary term x g 2 xg^2 x g 2 vanishes at 0 0 0 and beyond the support. Here C c ∞ ( [ 0 , ∞ ) ) C^\infty_c([0,\infty)) C c ∞ ([ 0 , ∞ )) denotes functions smooth up to zero and of compact support in the half-line. No vanishing condition is imposed at zero. These functions are dense in V w V_w V w (multiply by χ ( x / n ) \chi(x/n) χ ( x / n ) for a cutoff χ \chi χ and mollify) and in { g ∈ H 1 ( 0 , ∞ ) : x g ∈ L 2 } \{g\in H^1(0,\infty):xg\in L^2\} { g ∈ H 1 ( 0 , ∞ ) : xg ∈ L 2 } , and the map f ↦ w f f\mapsto\sqrt{w}f f ↦ w f carries the first space onto the second, so the identity extends to V w V_w V w . Equivalently, the identity with ∫ ∣ g ∣ 2 \int|g|^2 ∫ ∣ g ∣ 2 added to both sides identifies the two completed form norms. In particular, since x 2 ≥ R 2 x^2\geq R^2 x 2 ≥ R 2 on ( R , ∞ ) (R,\infty) ( R , ∞ ) ,
∫ R ∞ w ∣ f ∣ 2 d x ≤ 1 R 2 ∫ 0 ∞ x 2 ∣ g ∣ 2 d x ≤ 1 R 2 ∫ 0 ∞ w ( ∣ f ′ ∣ 2 + ∣ f ∣ 2 ) d x . \int_R^\infty w|f|^2\,dx\leq\frac{1}{R^2}\int_0^\infty x^2|g|^2\,dx\leq\frac{1}{R^2}\int_0^\infty w(|f'|^2+|f|^2)\,dx. ∫ R ∞ w ∣ f ∣ 2 d x ≤ R 2 1 ∫ 0 ∞ x 2 ∣ g ∣ 2 d x ≤ R 2 1 ∫ 0 ∞ w ( ∣ f ′ ∣ 2 + ∣ f ∣ 2 ) d x .
By orthogonality, (29) –(31) imply
∫ { x > R } ∩ G ε ∣ u ∣ 2 ≤ ∥ u ∥ 2 2 + ( 1 + 2 ε 2 3 e ) ∥ ∇ u ∥ 2 2 R 2 + 4 ε 2 e − 2 R 2 π 2 ∥ u y ∥ 2 2 , R > 0. (33) \int_{\{x>R\}\cap\mathcal{G}_{\varepsilon}} |u|^2 \le\frac{\|u\|_2^2+\left(1+\frac{2\varepsilon^2}{3e}\right)\|\nabla u\|_2^2}{R^2}+\frac{4\varepsilon^2e^{-2R^2}}{\pi^2}\|u_y\|_2^2,\qquad R>0. \tag*{(33)} ∫ { x > R } ∩ G ε ∣ u ∣ 2 ≤ R 2 ∥ u ∥ 2 2 + ( 1 + 3 e 2 ε 2 ) ∥∇ u ∥ 2 2 + π 2 4 ε 2 e − 2 R 2 ∥ u y ∥ 2 2 , R > 0. (33)
Thus the L 2 L^2 L 2 tails tend to zero uniformly on bounded subsets of H 1 ( G ε ) H^1(\mathcal{G}_{\varepsilon}) H 1 ( G ε ) . Every finite truncation is a bounded Lipschitz domain. Local Rellich compactness, a diagonal subsequence, and (33) prove global compactness.
The longitudinal spectrum. Consider the form ∫ 0 ∞ w ∣ f ′ ∣ 2 \int_0^\infty w|f'|^2 ∫ 0 ∞ w ∣ f ′ ∣ 2 in L 2 ( w d x ) L^2(w\,dx) L 2 ( w d x ) with form domain V w V_w V w . No trace restriction is imposed on the form domain at x = 0 x=0 x = 0 . The associated operator has the natural Neumann condition there. Its differential expression is
L f = − f ′ ′ + 2 x f ′ . Lf=-f''+2xf'. L f = − f ′′ + 2 x f ′ .
Under (32) , it is the half-line harmonic oscillator − g ′ ′ + ( x 2 − 1 ) g -g''+(x^2-1)g − g ′′ + ( x 2 − 1 ) g with Neumann condition at 0 0 0 : neither form domain imposes a condition at 0 0 0 , and for smooth f f f one has g ′ ( 0 ) = f ′ ( 0 ) g'(0)=f'(0) g ′ ( 0 ) = f ′ ( 0 ) because ( w ) ′ ( 0 ) = 0 (\sqrt{w})'(0)=0 ( w ) ′ ( 0 ) = 0 . Even reflection identifies this with the even subspace of the full-line harmonic oscillator. The Hermite spectrum is therefore
μ k = 4 k , k = 0 , 1 , 2 , … , \mu_k=4k,\qquad k=0,1,2,\ldots, μ k = 4 k , k = 0 , 1 , 2 , … ,
with eigenfunctions the even Hermite polynomials. In particular, 1 1 1 , x 2 − 1 2 x^2-\frac{1}{2} x 2 − 2 1 , and x 4 − 3 x 2 + 3 4 x^4-3x^2+\frac{3}{4} x 4 − 3 x 2 + 4 3 are the first three weighted eigenfunctions, as one checks by applying L L L .
Spectral comparison and simplicity. Here the cross-cut comparison can be justified directly on the half-line. Put
a = 1 + ε 2 W 1 2 3 , c = 4 ε 2 W 0 2 π 2 . a=1+\frac{\varepsilon^2W_1^2}{3},\qquad c=\frac{4\varepsilon^2W_0^2}{\pi^2}. a = 1 + 3 ε 2 W 1 2 , c = π 2 4 ε 2 W 0 2 .
Trial functions depending only on x x x give λ k ≤ μ k \lambda_k\le\mu_k λ k ≤ μ k . For the lower bound, suppose first that c λ k < 1 c\lambda_k<1 c λ k < 1 and average the span of the first k + 1 k+1 k + 1 planar eigenfunctions. For every u u u in that span, orthogonality and (31) give
2 ε ∫ w ∣ u ˉ ∣ 2 = ∥ u ∥ 2 2 − ∥ u − u ˉ ∥ 2 2 ≥ ( 1 − c λ k ) ∥ u ∥ 2 2 . 2\varepsilon\int w|\bar{u}|^2=\|u\|_2^2-\|u-\bar{u}\|_2^2\ge(1-c\lambda_k)\|u\|_2^2. 2 ε ∫ w ∣ u ˉ ∣ 2 = ∥ u ∥ 2 2 − ∥ u − u ˉ ∥ 2 2 ≥ ( 1 − c λ k ) ∥ u ∥ 2 2 .
The averaging map is consequently injective on that span. Equation (30) and the weighted min–max principle yield μ k ≤ a λ k / ( 1 − c λ k ) \mu_k\le a\lambda_k/(1-c\lambda_k) μ k ≤ a λ k / ( 1 − c λ k ) , or
μ k 1 + ε 2 ( W 1 2 / 3 + 4 W 0 2 μ k / π 2 ) ≤ λ k ≤ μ k . (34) \frac{\mu_k}{1+\varepsilon^2(W_1^2/3+4W_0^2\mu_k/\pi^2)}\le\lambda_k\le\mu_k. \tag*{(34)} 1 + ε 2 ( W 1 2 /3 + 4 W 0 2 μ k / π 2 ) μ k ≤ λ k ≤ μ k . (34)
If c λ k ≥ 1 c\lambda_k\ge1 c λ k ≥ 1 , the same lower bound is immediate since its left-hand side is less than 1 / c 1/c 1/ c . Substituting μ 1 = 4 \mu_1=4 μ 1 = 4 , μ 2 = 8 \mu_2=8 μ 2 = 8 proves (28) . Finally,
2 3 e + 32 π 2 < 4. \frac{2}{3e}+\frac{32}{\pi^2}<4. 3 e 2 + π 2 32 < 4.
For example, e > 2 e>2 e > 2 and π > 3 \pi>3 π > 3 bound the left-hand side by 1 / 3 + 32 / 9 = 35 / 9 < 4 1/3+32/9=35/9<4 1/3 + 32/9 = 35/9 < 4 . For ε ≤ 1 / 2 \varepsilon\le1/2 ε ≤ 1/2 the denominator in (28) is strictly less than 2 2 2 , so λ 1 ≤ 4 < λ 2 \lambda_1\le4<\lambda_2 λ 1 ≤ 4 < λ 2 .
Corollary 5.2. Assume Hypothesis 1.1 . For 0 < ε ≤ 1 2 0<\varepsilon\le\frac{1}{2} 0 < ε ≤ 2 1 , every real first positive Neumann eigenfunction of G ε \mathcal{G}_{\varepsilon} G ε satisfies (2) .
Proof. By Theorem 5.1, G ε \mathcal{G}_{\varepsilon} G ε satisfies the hypotheses of Theorem 1.2.
Remark 5.3 (Extrema approached at infinity). For every δ > 0 \delta> 0 δ > 0 the layer B δ ( G ε ) B_{\delta}(G_{\varepsilon}) B δ ( G ε ) contains the whole tail { x > R } ∩ G ε \{x > R\} \cap G_{\varepsilon} { x > R } ∩ G ε for R R R sufficiently large, because the cross-cuts have length 2 ε e − x 2 2\varepsilon e^{-x^{2}} 2 ε e − x 2 . If a supremum or infimum is approached along that tail, the corresponding identity in (2) is automatic. No assertion about growth or boundedness of the planar eigenfunction at infinity follows from the longitudinal comparison used here.
Remark 5.4 (Other horns). The same argument applies to
G ε , α = { ( x , y ) : x > 0 , ∣ y ∣ < ε e − x α } , α ≥ 2. G_{\varepsilon,\alpha}=\{(x,y):x>0,\lvert y\rvert<\varepsilon e^{-x^{\alpha}}\},\qquad\alpha\ge2. G ε , α = {( x , y ) : x > 0 , ∣ y ∣ < ε e − x α } , α ≥ 2.
The longitudinal substitution g = e − x α / 2 f g=e^{-x^{\alpha}/2}f g = e − x α /2 f gives the potential
V α ( x ) = α 2 4 x 2 α − 2 − α ( α − 1 ) 2 x α − 2 . V_{\alpha}(x)=\frac{\alpha^{2}}{4}x^{2\alpha-2}-\frac{\alpha(\alpha-1)}{2}x^{\alpha-2}. V α ( x ) = 4 α 2 x 2 α − 2 − 2 α ( α − 1 ) x α − 2 .
which is bounded below and tends to infinity (for 1 < α < 2 1<\alpha<2 1 < α < 2 it would tend to − ∞ -\infty − ∞ at 0 0 0 , which is why α ≥ 2 \alpha\ge2 α ≥ 2 is assumed). The resulting weighted tail estimate and the transverse Poincaré inequality give compact embedding. The longitudinal eigenvalues are discrete, because V α → ∞ V_{\alpha}\to\infty V α → ∞ gives a compact resolvent, and simple, because the equation is regular at 0 0 0 for α ≥ 2 \alpha\ge2 α ≥ 2 , so that the Neumann condition g ′ ( 0 ) = 0 g'(0)=0 g ′ ( 0 ) = 0 determines the eigenfunction up to a scalar. For each fixed α \alpha α , the comparison in (34) , with W 0 = 1 W_{0}=1 W 0 = 1 and W 1 = sup ∣ ( e − x α ) ′ ∣ < ∞ W_{1}=\sup\lvert(e^{-x^{\alpha}})'\rvert<\infty W 1 = sup ∣( e − x α ) ′ ∣ < ∞ , therefore gives λ 1 < λ 2 \lambda_{1}<\lambda_{2} λ 1 < λ 2 for all sufficiently small ε > 0 \varepsilon>0 ε > 0 . These horns also satisfy the hypotheses of Theorem 1.2 .
Remark 5.5 (Failure of polynomial heat-kernel bounds). On any unbounded open set D D D of finite area, the continuous embedding H 1 ( D ) ↪ L q ( D ) H^{1}(D)\hookrightarrow L^{q}(D) H 1 ( D ) ↪ L q ( D ) fails for every q > 2 q>2 q > 2 . See Adams [1 ] . By the Sobolev–heat-kernel equivalence [18 Proposition 3] , this excludes every spatially uniform bound p D ( t , z , z ′ ) ≤ C t − N / 2 p_{D}(t,z,z')\le Ct^{-N/2} p D ( t , z , z ′ ) ≤ C t − N /2 for 0 < t ≤ 1 0<t\le1 0 < t ≤ 1 with finite N > 0 N>0 N > 0 (a bound with N ≤ 2 N\le2 N ≤ 2 would imply one with N > 2 N>2 N > 2 on 0 < t ≤ 1 0<t\le1 0 < t ≤ 1 ). For the horns the failure can be seen directly. On G ε , α G_{\varepsilon,\alpha} G ε , α , the function u = e s x α u=e^{sx^{\alpha}} u = e s x α has
∫ u 2 = 2 ε ∫ 0 ∞ e ( 2 s − 1 ) x α d x , ∫ ∣ ∇ u ∣ 2 = 2 ε s 2 α 2 ∫ 0 ∞ x 2 α − 2 e ( 2 s − 1 ) x α d x , \int u^{2}=2\varepsilon\int_{0}^{\infty}e^{(2s-1)x^{\alpha}}\,dx,\qquad\int\lvert\nabla u\rvert^{2}=2\varepsilon s^{2}\alpha^{2}\int_{0}^{\infty}x^{2\alpha-2}e^{(2s-1)x^{\alpha}}\,dx, ∫ u 2 = 2 ε ∫ 0 ∞ e ( 2 s − 1 ) x α d x , ∫ ∣ ∇ u ∣ 2 = 2 ε s 2 α 2 ∫ 0 ∞ x 2 α − 2 e ( 2 s − 1 ) x α d x ,
both finite for s < 1 / 2 s<1/2 s < 1/2 , while
∫ u q = 2 ε ∫ 0 ∞ e ( q s − 1 ) x α d x = ∞ \int u^{q}=2\varepsilon\int_{0}^{\infty}e^{(qs-1)x^{\alpha}}\,dx=\infty ∫ u q = 2 ε ∫ 0 ∞ e ( q s − 1 ) x α d x = ∞
for s ≥ 1 / q s\ge1/q s ≥ 1/ q . For q > 2 q>2 q > 2 take s = 1 / q s=1/q s = 1/ q . This does not preclude compact embedding into L 2 ( D ) L^{2}(D) L 2 ( D ) . The horns therefore illustrate the greater scope of the compactness hypothesis in Theorem 1.2 .
An unbounded rooms-and-passages domain
The following construction is motivated by the example of Bañuelos and Davis [11 Section 4] . It consists of shrinking disks joined by thin passages. The example distinguishes convergence of the normalized Dirichlet heat kernel that is uniform in one spatial variable from convergence that is uniform in both variables. This distinction is related to intrinsic ultracontractivity (IU) and Half-IU. Here we replace the disks with squares and choose the dimensions to ensure compactness of the Neumann resolvent and simplicity of the first positive eigenvalue. Unlike the Gaussian horn, this domain has repeated bottlenecks and infinitely many corners, although every finite truncation is Lipschitz. An additional advantage is that both spectral properties follow from elementary estimates on squares and rectangles, which simplify the analysis compared with the disks in the original construction.
Rooms-and-passages domains also appear in the analysis of Neumann essential spectra [27 ] and comparisons with one-dimensional operators on their skeletons [14 ] . The probabilistic background includes exit times and lifetimes of conditioned Brownian motion [19 , 11 ] , and their connections. with intrinsic ultracontractivity [21 , 7 ] . These connections motivate the geometry. No assertion about Dirichlet intrinsic ultracontractivity is needed below.
Let L n L_n L n denote passage length and h n h_n h n passage width. These differ from the cross-cut width ℓ \ell ℓ in § 7 . For n ≥ 1 n \ge1 n ≥ 1 , set
r n = e − n 2 , L n = 1 n , h n = n 2 e − 2 ( n + 1 ) 2 = n 2 r n + 1 2 . (35) r_n=e^{-n^2},\qquad L_n=\frac{1}{n},\qquad h_n=n^2e^{-2(n+1)^2}=n^2r_{n+1}^2. \tag*{(35)} r n = e − n 2 , L n = n 1 , h n = n 2 e − 2 ( n + 1 ) 2 = n 2 r n + 1 2 . (35)
Define x 1 = 0 x_1=0 x 1 = 0 and x n + 1 = x n + r n + L n x_{n+1}=x_n+r_n+L_n x n + 1 = x n + r n + L n , and let
Q n = ( x n , x n + r n ) × ( − r n / 2 , r n / 2 ) , Q_n=(x_n,x_n+r_n)\times(-r_n/2,r_n/2), Q n = ( x n , x n + r n ) × ( − r n /2 , r n /2 ) ,
P n = ( x n + r n , x n + 1 ) × ( − h n / 2 , h n / 2 ) . P_n=(x_n+r_n,x_{n+1})\times(-h_n/2,h_n/2). P n = ( x n + r n , x n + 1 ) × ( − h n /2 , h n /2 ) .
The domain is
Ω = int ( ⋃ n ≥ 1 ( Q n ‾ ∪ P n ‾ ) ) . (36) \Omega=\operatorname{int}\left(\bigcup_{n\ge1}(\overline{Q_n}\cup\overline{P_n})\right). \tag*{(36)} Ω = int ( n ≥ 1 ⋃ ( Q n ∪ P n ) ) . (36)
Figure 5.2 shows the arrangement schematically. The open interfaces between adjacent pieces belong to Ω \Omega Ω . Since h n < r n + 1 < r n h_n<r_{n+1}<r_n h n < r n + 1 < r n , each passage is narrower than both adjacent rooms. Every vertical section is an interval containing the horizontal axis. Vertical contraction onto that axis shows that Ω \Omega Ω is contractible. Moreover, ∑ n L n = ∞ \sum_n L_n=\infty ∑ n L n = ∞ , whereas
Figure 5.2. Square rooms Q n Q_n Q n of side r n r_n r n , joined by rectangular passages P n P_n P n of length L n L_n L n and width h n h_n h n . The interfaces are open inside the domain. The drawing is schematic. Room sizes and passage widths are enlarged for visibility. Dashed lines indicate continuation to infinity.
∣ Ω ∣ = ∑ n ≥ 1 ( r n 2 + h n L n ) = ∑ n ≥ 1 ( e − 2 n 2 + n e − 2 ( n + 1 ) 2 ) < ∞ . (37) |\Omega|=\sum_{n\ge1}(r_n^2+h_nL_n)=\sum_{n\ge1}\left(e^{-2n^2}+ne^{-2(n+1)^2}\right)<\infty. \tag*{(37)} ∣Ω∣ = n ≥ 1 ∑ ( r n 2 + h n L n ) = n ≥ 1 ∑ ( e − 2 n 2 + n e − 2 ( n + 1 ) 2 ) < ∞. (37)
Every bounded horizontal interval meets only finitely many pieces. The boundary is locally Lipschitz at every finite boundary point, but the shrinking widths prevent a uniform positive scale for boundary charts. Write Ω ≥ M = Ω ∩ { ( x , y ) : x > x M } \Omega_{\ge M}=\Omega\cap\{(x,y):x>x_M\} Ω ≥ M = Ω ∩ {( x , y ) : x > x M } . Up to sets of area zero, this is the union of Q n Q_n Q n and P n P_n P n for n ≥ M n\ge M n ≥ M .
Theorem 5.6. The domain (36) is an unbounded simply connected planar domain of finite area with locally Lipschitz boundary. The embedding H 1 ( Ω ) ↪ L 2 ( Ω ) H^1(\Omega)\hookrightarrow L^2(\Omega) H 1 ( Ω ) ↪ L 2 ( Ω ) is compact. More precisely, there exists C < ∞ C<\infty C < ∞ such that
∫ Ω ≥ M ∣ u ∣ 2 ≤ C M 2 ( ∫ Ω ∣ u ∣ 2 + ∫ Ω ∣ ∇ u ∣ 2 ) , M ≥ 2 , u ∈ H 1 ( Ω ) . (38) \int_{\Omega_{\ge M}}|u|^2\le\frac{C}{M^2}\left(\int_\Omega|u|^2+\int_\Omega|\nabla u|^2\right),\qquad M\ge2,\quad u\in H^1(\Omega). \tag*{(38)} ∫ Ω ≥ M ∣ u ∣ 2 ≤ M 2 C ( ∫ Ω ∣ u ∣ 2 + ∫ Ω ∣∇ u ∣ 2 ) , M ≥ 2 , u ∈ H 1 ( Ω ) . (38)
Its Neumann eigenvalues, counted with multiplicity, satisfy
0 = λ 0 < λ 1 < 4 5 < 10 9 ≤ λ 2 . 0=\lambda_0<\lambda_1<\frac{4}{5}<\frac{10}{9}\le\lambda_2. 0 = λ 0 < λ 1 < 5 4 < 9 10 ≤ λ 2 .
In particular, the first positive eigenvalue is simple. Under Hypothesis 1.1 1.1 1.1 , every real eigenfunction associated with λ 1 \lambda_1 λ 1 satisfies the boundary-layer identities (2) .
Proof. Compactness. First record explicit local estimates that will also be used for simplicity. For a square Q Q Q of side r r r , an opening I I I of length h ≤ r h\le r h ≤ r in one side, and the respective averages a = ∣ Q ∣ − 1 ∫ Q u a=|Q|^{-1}\int_Q u a = ∣ Q ∣ − 1 ∫ Q u and b = h − 1 ∫ I u b=h^{-1}\int_Iu b = h − 1 ∫ I u , one has
∫ Q ∣ u − a ∣ 2 ≤ r 2 π 2 ∫ Q ∣ ∇ u ∣ 2 , ∣ b − a ∣ 2 ≤ r 2 h ∫ Q ∣ ∇ u ∣ 2 . (39) \int_Q|u-a|^2\le\frac{r^2}{\pi^2}\int_Q|\nabla u|^2,\qquad|b-a|^2\le\frac{r}{2h}\int_Q|\nabla u|^2. \tag*{(39)} ∫ Q ∣ u − a ∣ 2 ≤ π 2 r 2 ∫ Q ∣∇ u ∣ 2 , ∣ b − a ∣ 2 ≤ 2 h r ∫ Q ∣∇ u ∣ 2 . (39)
The first estimate is the Poincaré inequality on a square. To verify the second with its stated constant, take Q = ( 0 , r ) 2 Q=(0,r)^2 Q = ( 0 , r ) 2 and the relevant side to be J = { 0 } × ( 0 , r ) J=\{0\}\times(0,r) J = { 0 } × ( 0 , r ) . For g ( y ) = r − 1 ∫ 0 r u ( x , y ) d x g(y)=r^{-1}\int_0^r u(x,y)\,dx g ( y ) = r − 1 ∫ 0 r u ( x , y ) d x , the one-dimensional Poincaré inequality gives
∥ g − a ∥ L 2 ( 0 , r ) 2 ≤ r π 2 ∫ Q ∣ ∂ y u ∣ 2 . \|g-a\|_{L^2(0,r)}^2\le\frac{r}{\pi^2}\int_Q|\partial_yu|^2. ∥ g − a ∥ L 2 ( 0 , r ) 2 ≤ π 2 r ∫ Q ∣ ∂ y u ∣ 2 .
Also,
u ( 0 , y ) − g ( y ) = − 1 r ∫ 0 r ( r − x ) ∂ x u ( x , y ) d x , ∥ u ( 0 , ⋅ ) − g ∥ L 2 ( 0 , r ) 2 ≤ r 3 ∫ Q ∣ ∂ x u ∣ 2 . u(0,y)-g(y)=-\frac{1}{r}\int_0^r(r-x)\partial_xu(x,y)\,dx,\qquad\|u(0,\cdot)-g\|_{L^2(0,r)}^2\le\frac{r}{3}\int_Q|\partial_xu|^2. u ( 0 , y ) − g ( y ) = − r 1 ∫ 0 r ( r − x ) ∂ x u ( x , y ) d x , ∥ u ( 0 , ⋅ ) − g ∥ L 2 ( 0 , r ) 2 ≤ 3 r ∫ Q ∣ ∂ x u ∣ 2 .
The triangle inequality followed by Cauchy–Schwarz yields
∫ J ∣ u − a ∣ 2 ≤ r ( 1 3 + 1 π 2 ) ∫ Q ∣ ∇ u ∣ 2 ≤ r 2 ∫ Q ∣ ∇ u ∣ 2 . \int_J|u-a|^2\le r\left(\frac{1}{3}+\frac{1}{\pi^2}\right)\int_Q|\nabla u|^2\le\frac{r}{2}\int_Q|\nabla u|^2. ∫ J ∣ u − a ∣ 2 ≤ r ( 3 1 + π 2 1 ) ∫ Q ∣∇ u ∣ 2 ≤ 2 r ∫ Q ∣∇ u ∣ 2 .
Averaging over I ⊂ J I\subset J I ⊂ J proves the second estimate in (39) . Density and continuity of the trace justify the calculation for H 1 ( Q ) H^1(Q) H 1 ( Q ) functions.
Let a n a_n a n be the average of u u u on Q n Q_n Q n . The matching traces at the openings and the elementary passage-end estimate
∣ b r i g h t − b l e f t ∣ 2 ≤ L n h n ∫ P n ∣ ∇ u ∣ 2 |b_{\mathrm{right}}-b_{\mathrm{left}}|^2\le\frac{L_n}{h_n}\int_{P_n}|\nabla u|^2 ∣ b right − b left ∣ 2 ≤ h n L n ∫ P n ∣∇ u ∣ 2
imply
∣ a n + 1 − a n ∣ 2 ≤ ρ n E n , ρ n = r n / 2 + L n + r n + 1 / 2 h n , (40) |a_{n+1}-a_n|^2\le\rho_nE_n,\qquad\rho_n=\frac{r_n/2+L_n+r_{n+1}/2}{h_n}, \tag*{(40)} ∣ a n + 1 − a n ∣ 2 ≤ ρ n E n , ρ n = h n r n /2 + L n + r n + 1 /2 , (40)
where
E n = ∫ Q n ∣ ∇ u ∣ 2 + ∫ P n ∣ ∇ u ∣ 2 + ∫ Q n + 1 ∣ ∇ u ∣ 2 . E_n=\int_{Q_n}|\nabla u|^2+\int_{P_n}|\nabla u|^2+\int_{Q_{n+1}}|\nabla u|^2. E n = ∫ Q n ∣∇ u ∣ 2 + ∫ P n ∣∇ u ∣ 2 + ∫ Q n + 1 ∣∇ u ∣ 2 .
Indeed, each of the three comparisons is bounded by the square root of its resistance coefficient times the energy in the corresponding piece; Cauchy–Schwarz gives (40) . The mass in a passage satisfies
∫ P n ∣ u ∣ 2 ≤ 4 h n L n ∣ a n ∣ 2 + 2 L n r n ∫ Q n ∣ ∇ u ∣ 2 + 8 L n 2 π 2 ∫ P n ∣ ∇ u ∣ 2 . (41) \int_{P_n}|u|^2\le4h_nL_n|a_n|^2+2L_nr_n\int_{Q_n}|\nabla u|^2+\frac{8L_n^2}{\pi^2}\int_{P_n}|\nabla u|^2. \tag*{(41)} ∫ P n ∣ u ∣ 2 ≤ 4 h n L n ∣ a n ∣ 2 + 2 L n r n ∫ Q n ∣∇ u ∣ 2 + π 2 8 L n 2 ∫ P n ∣∇ u ∣ 2 . (41)
For completeness, let m ( s ) m(s) m ( s ) be the transverse average on P n P_n P n , 0 < s < L n 0<s<L_n 0 < s < L n , and b = m ( 0 ) b=m(0) b = m ( 0 ) . The transverse Poincaré inequality and the one-dimensional Dirichlet–Neumann inequality give
∫ P n ∣ u − m ( s ) ∣ 2 ≤ h n 2 π 2 ∫ P n ∣ ∂ y u ∣ 2 , \int_{P_n} |u-m(s)|^2 \le\frac{h_n^2}{\pi^2}\int_{P_n} |\partial_yu|^2, ∫ P n ∣ u − m ( s ) ∣ 2 ≤ π 2 h n 2 ∫ P n ∣ ∂ y u ∣ 2 ,
h n ∫ 0 L n ∣ m ( s ) ∣ 2 d s ≤ 2 h n L n ∣ b ∣ 2 + 8 L n 2 π 2 ∫ P n ∣ ∂ x u ∣ 2 . h_n\int_0^{L_n}|m(s)|^2\,ds \le2h_nL_n|b|^2+\frac{8L_n^2}{\pi^2}\int_{P_n}|\partial_xu|^2. h n ∫ 0 L n ∣ m ( s ) ∣ 2 d s ≤ 2 h n L n ∣ b ∣ 2 + π 2 8 L n 2 ∫ P n ∣ ∂ x u ∣ 2 .
Use ∣ b ∣ 2 ≤ 2 ∣ a n ∣ 2 + ( r n / h n ) ∫ Q n ∣ ∇ u ∣ 2 |b|^2\le2|a_n|^2+(r_n/h_n)\int_{Q_n}|\nabla u|^2 ∣ b ∣ 2 ≤ 2∣ a n ∣ 2 + ( r n / h n ) ∫ Q n ∣∇ u ∣ 2 , the orthogonal decomposition into transverse average and fluctuation, and h n ≤ L n h_n\le L_n h n ≤ L n to obtain (41) . For the rooms,
∫ Q n ∣ u ∣ 2 ≤ r n 2 ∣ a n ∣ 2 + r n 2 π 2 ∫ Q n ∣ ∇ u ∣ 2 . (42) \int_{Q_n}|u|^2\le r_n^2|a_n|^2+\frac{r_n^2}{\pi^2}\int_{Q_n}|\nabla u|^2. \tag*{(42)} ∫ Q n ∣ u ∣ 2 ≤ r n 2 ∣ a n ∣ 2 + π 2 r n 2 ∫ Q n ∣∇ u ∣ 2 . (42)
Put E = ∫ Ω ∣ ∇ u ∣ 2 E=\int_\Omega|\nabla u|^2 E = ∫ Ω ∣∇ u ∣ 2 , A 1 = 0 A_1=0 A 1 = 0 , and A n = ∑ k = 1 n − 1 ρ k A_n=\sum_{k=1}^{n-1}\rho_k A n = ∑ k = 1 n − 1 ρ k for n ≥ 2 n\ge2 n ≥ 2 . Since ∑ k E k ≤ 2 E \sum_kE_k\le2E ∑ k E k ≤ 2 E , telescoping (40) gives
∣ a n ∣ 2 ≤ 2 ∣ a 1 ∣ 2 + 4 A n E . (43) |a_n|^2\le2|a_1|^2+4A_nE. \tag*{(43)} ∣ a n ∣ 2 ≤ 2∣ a 1 ∣ 2 + 4 A n E . (43)
Set T n = e 2 ( n + 1 ) 2 / n 3 T_n=e^{2(n+1)^2}/n^3 T n = e 2 ( n + 1 ) 2 / n 3 . Because r n , r n + 1 ≤ 1 / n r_n,r_{n+1}\le1/n r n , r n + 1 ≤ 1/ n , one has T n ≤ ρ n ≤ 2 T n T_n\le\rho_n\le2T_n T n ≤ ρ n ≤ 2 T n , and for n ≥ 2 n\ge2 n ≥ 2 ,
T n − 1 T n = ( n n − 1 ) 3 e − 4 n − 2 ≤ 8 e − 10 < 1 100 . (44) \frac{T_{n-1}}{T_n}=\left(\frac{n}{n-1}\right)^3e^{-4n-2}\le8e^{-10}<\frac{1}{100}. \tag*{(44)} T n T n − 1 = ( n − 1 n ) 3 e − 4 n − 2 ≤ 8 e − 10 < 100 1 . (44)
Consequently,
A n ≤ C e 2 n 2 ( n − 1 ) 3 ( n ≥ 2 ) , W n : = r n 2 + 4 h n L n = e − 2 n 2 ( 1 + 4 n e − 4 n − 2 ) ≤ 11 10 e − 2 n 2 . A_n\le C\frac{e^{2n^2}}{(n-1)^3}\quad(n\ge2),\qquad W_n:=r_n^2+4h_nL_n=e^{-2n^2}(1+4ne^{-4n-2})\le\frac{11}{10}e^{-2n^2}. A n ≤ C ( n − 1 ) 3 e 2 n 2 ( n ≥ 2 ) , W n := r n 2 + 4 h n L n = e − 2 n 2 ( 1 + 4 n e − 4 n − 2 ) ≤ 10 11 e − 2 n 2 .
It follows that
∑ n ≥ M W n A n ≤ C M 2 , ∑ n ≥ M W n ≤ C M 2 , M ≥ 2. \sum_{n\ge M}W_nA_n\le\frac{C}{M^2},\qquad\sum_{n\ge M}W_n\le\frac{C}{M^2},\qquad M\ge2. n ≥ M ∑ W n A n ≤ M 2 C , n ≥ M ∑ W n ≤ M 2 C , M ≥ 2.
The fluctuation coefficients in (41) and (42) are bounded by C / n 2 C/n^2 C / n 2 . Summing those estimates and using (43) gives
∫ Ω ≥ M ∣ u ∣ 2 ≤ C M 2 ( ∣ a 1 ∣ 2 + E ) . \int_{\Omega_{\ge M}}|u|^2\le\frac{C}{M^2}(|a_1|^2+E). ∫ Ω ≥ M ∣ u ∣ 2 ≤ M 2 C ( ∣ a 1 ∣ 2 + E ) .
Since ∣ a 1 ∣ 2 ≤ ∣ Q 1 ∣ − 1 ∫ Q 1 ∣ u ∣ 2 |a_1|^2\le|Q_1|^{-1}\int_{Q_1}|u|^2 ∣ a 1 ∣ 2 ≤ ∣ Q 1 ∣ − 1 ∫ Q 1 ∣ u ∣ 2 , this proves (38) . Each initial segment U M = Ω ∩ { ( x , y ) : x < x M } U_M=\Omega\cap\{(x,y):x<x_M\} U M = Ω ∩ {( x , y ) : x < x M } is a bounded Lipschitz polygonal domain. Rellich compactness on these initial segments, a diagonal subsequence, and the uniform tail estimate prove compactness on Ω \Omega Ω . The Neumann resolvent is therefore compact. Connectedness and finite area imply that its zero eigenspace consists precisely of the constants.
Simplicity. For an upper bound on λ 1 \lambda_1 λ 1 , let f f f vanish on Q 1 Q_1 Q 1 , increase linearly from zero to one along P 1 P_1 P 1 , and equal one on the remaining domain. Then f ∈ H 1 ( Ω ) f\in H^1(\Omega) f ∈ H 1 ( Ω ) and ∫ Ω ∣ ∇ f ∣ 2 = h 1 / L 1 = e − 8 \int_\Omega|\nabla f|^2=h_1/L_1=e^{-8} ∫ Ω ∣∇ f ∣ 2 = h 1 / L 1 = e − 8 . Write m 2 = e − 8 = ∣ Q 2 ∣ = ∣ P 1 ∣ m_2=e^{-8}=|Q_2|=|P_1| m 2 = e − 8 = ∣ Q 2 ∣ = ∣ P 1 ∣ and m 1 = e − 2 = ∣ Q 1 ∣ m_1=e^{-2}=|Q_1| m 1 = e − 2 = ∣ Q 1 ∣ . Minimizing over constants, first on the whole domain and then on the first three pieces, gives
∫ Ω ∣ f − f Ω ∣ 2 = min c ∈ R ∫ Ω ∣ f − c ∣ 2 ≥ min c ∈ R ∫ Q 1 ∪ P 1 ∪ Q 2 ∣ f − c ∣ 2 = 4 m 2 3 − 9 m 2 2 4 ( m 1 + 2 m 2 ) . \begin{aligned}
\int_\Omega|f-f_\Omega|^2=\min_{c\in\mathbb R}\int_\Omega|f-c|^2 \\
&\ge\min_{c\in\mathbb R}\int_{Q_1\cup P_1\cup Q_2}|f-c|^2=\frac{4m_2}{3}-\frac{9m_2^2}{4(m_1+2m_2)}.
\end{aligned} ∫ Ω ∣ f − f Ω ∣ 2 = c ∈ R min ∫ Ω ∣ f − c ∣ 2 ≥ c ∈ R min ∫ Q 1 ∪ P 1 ∪ Q 2 ∣ f − c ∣ 2 = 3 4 m 2 − 4 ( m 1 + 2 m 2 ) 9 m 2 2 .
Here f Ω f_{\Omega} f Ω denotes the average on Ω \Omega Ω . Since m 1 / m 2 = e 6 > 25 m_1/m_2=e^6>25 m 1 / m 2 = e 6 > 25 , the last expression exceeds 5 m 2 / 4 5m_2/4 5 m 2 /4 . The Rayleigh principle yields λ 1 < 4 / 5 \lambda_1<4/5 λ 1 < 4/5 .
To bound λ 2 \lambda_2 λ 2 , consider the subspace
V = { u ∈ H 1 ( Ω ) : ∫ Q 1 u = 0 , ∫ Q 2 u = 0 } . V=\left\{u\in H^1(\Omega):\int_{Q_1}u=0,\ \int_{Q_2}u=0\right\}. V = { u ∈ H 1 ( Ω ) : ∫ Q 1 u = 0 , ∫ Q 2 u = 0 } .
It suffices to prove
∫ Ω ∣ u ∣ 2 ≤ 9 10 ∫ Ω ∣ ∇ u ∣ 2 , u ∈ V . (45) \int_{\Omega}|u|^2\le\frac{9}{10}\int_{\Omega}|\nabla u|^2,\qquad u\in V. \tag*{(45)} ∫ Ω ∣ u ∣ 2 ≤ 10 9 ∫ Ω ∣∇ u ∣ 2 , u ∈ V . (45)
Indeed, these are two independent L 2 L^2 L 2 -orthogonality conditions, so the min–max principle gives λ 2 ≥ 10 / 9 \lambda_2\ge10/9 λ 2 ≥ 10/9 . Fix u ∈ V u\in V u ∈ V , so that a 1 = a 2 = 0 a_1=a_2=0 a 1 = a 2 = 0 . Estimates (41) –(42) give
∫ Q 1 ∪ P 1 ∣ u ∣ 2 ≤ ( e − 2 π 2 + 2 e − 1 ) ∫ Q 1 ∣ ∇ u ∣ 2 + 8 π 2 ∫ P 1 ∣ ∇ u ∣ 2 ≤ 9 10 ∫ Q 1 ∪ P 1 ∣ ∇ u ∣ 2 . (46) \begin{aligned}
\int_{Q_1\cup P_1}|u|^2\le\left(\frac{e^{-2}}{\pi^2}+2e^{-1}\right)\int_{Q_1}|\nabla u|^2+\frac{8}{\pi^2}\int_{P_1}|\nabla u|^2 \\
&\le\frac{9}{10}\int_{Q_1\cup P_1}|\nabla u|^2. \tag*{(46)}
\end{aligned} ∫ Q 1 ∪ P 1 ∣ u ∣ 2 ≤ ( π 2 e − 2 + 2 e − 1 ) ∫ Q 1 ∣∇ u ∣ 2 + π 2 8 ∫ P 1 ∣∇ u ∣ 2 ≤ 10 9 ∫ Q 1 ∪ P 1 ∣∇ u ∣ 2 . (46)
For the tail, put E T = ∫ Ω ≥ 2 ∣ ∇ u ∣ 2 E_T=\int_{\Omega_{\ge2}}|\nabla u|^2 E T = ∫ Ω ≥ 2 ∣∇ u ∣ 2 and S n = ∑ k = 2 n − 1 ρ k S_n=\sum_{k=2}^{n-1}\rho_k S n = ∑ k = 2 n − 1 ρ k for n ≥ 3 n\ge3 n ≥ 3 . Since a 2 = 0 a_2=0 a 2 = 0 , (40) gives
∣ a n ∣ 2 ≤ 2 S n E T . |a_n|^2\le2S_nE_T. ∣ a n ∣ 2 ≤ 2 S n E T .
For k ≥ 2 k\ge2 k ≥ 2 ,
ρ k = ( 1 + k ( r k + r k + 1 ) 2 ) T k ≤ 21 20 T k , \rho_k=\left(1+\frac{k(r_k+r_{k+1})}{2}\right)T_k\le\frac{21}{20}T_k, ρ k = ( 1 + 2 k ( r k + r k + 1 ) ) T k ≤ 20 21 T k ,
because k ( r k + r k + 1 ) / 2 ≤ k e − k 2 ≤ 2 e − 4 < 1 / 20 k(r_k+r_{k+1})/2\le ke^{-k^2}\le2e^{-4}<1/20 k ( r k + r k + 1 ) /2 ≤ k e − k 2 ≤ 2 e − 4 < 1/20 . Together with (44) , this gives
S n ≤ 21 20 100 99 T n − 1 < 11 10 e 2 n 2 ( n − 1 ) 3 . S_n\le\frac{21}{20}\frac{100}{99}T_{n-1}<\frac{11}{10}\frac{e^{2n^2}}{(n-1)^3}. S n ≤ 20 21 99 100 T n − 1 < 10 11 ( n − 1 ) 3 e 2 n 2 .
Using the bound on W n W_n W n from the compactness argument,
∑ n = 3 ∞ W n S n ≤ 121 100 ∑ k = 2 ∞ k − 3 ≤ 121 400 . (47) \sum_{n=3}^{\infty}W_nS_n\le\frac{121}{100}\sum_{k=2}^{\infty}k^{-3}\le\frac{121}{400}. \tag*{(47)} n = 3 ∑ ∞ W n S n ≤ 100 121 k = 2 ∑ ∞ k − 3 ≤ 400 121 . (47)
where ∑ k = 2 ∞ k − 3 ≤ 1 / 8 + ∫ 2 ∞ t − 3 d t = 1 / 4 \sum_{k=2}^{\infty}k^{-3}\le1/8+\int_2^{\infty}t^{-3}\,dt=1/4 ∑ k = 2 ∞ k − 3 ≤ 1/8 + ∫ 2 ∞ t − 3 d t = 1/4 . For n ≥ 2 n\ge2 n ≥ 2 the fluctuation coefficients satisfy
r n 2 π 2 + 2 L n r n ≤ 1 4 , 8 L n 2 π 2 ≤ 1 4 . \frac{r_n^2}{\pi^2}+2L_nr_n\le\frac14,\qquad\frac{8L_n^2}{\pi^2}\le\frac14. π 2 r n 2 + 2 L n r n ≤ 4 1 , π 2 8 L n 2 ≤ 4 1 .
Thus (41) –(42) , a 2 = 0 a_2=0 a 2 = 0 , and (47) imply
∫ Ω ≥ 2 ∣ u ∣ 2 ≤ ∑ n = 3 ∞ W n ∣ a n ∣ 2 + 1 4 E T ≤ ( 121 200 + 1 4 ) E T = 171 200 E T ≤ 9 10 E T . \int_{\Omega_{\ge2}}|u|^2\le\sum_{n=3}^{\infty}W_n|a_n|^2+\frac14E_T\le\left(\frac{121}{200}+\frac14\right)E_T=\frac{171}{200}E_T\le\frac9{10}E_T. ∫ Ω ≥ 2 ∣ u ∣ 2 ≤ n = 3 ∑ ∞ W n ∣ a n ∣ 2 + 4 1 E T ≤ ( 200 121 + 4 1 ) E T = 200 171 E T ≤ 10 9 E T .
Combining this with (46) proves (45) , and hence simplicity. The conditional hot spots conclusion follows from Theorem 1.2 .
As an unbounded domain of finite area, this example also admits no polynomial Neumann heat-kernel bound, by Remark 5.5. The numerical bounds above are deliberately nonoptimal. Their purpose is to separate the first two positive eigenvalues without identifying the nodal set.
On removing the simplicity assumption
The examples given above do not exhaust the domains to which Theorem 1.2 or Corollary 2.4 applies. Several results in the literature establish simplicity of the first positive Neumann eigenvalue for specific classes of domains, often in connection with results on the hot spots conjecture. These classes include sufficiently elongated bounded convex planar domains [8 Proposition 2.4] , nonsquare lip domains [4 Theorem 1(i)] , the class of smooth convex planar domains covered by [5 Theorem 2.6] , and all non-equilateral triangles [36 Theorem 1.2] . Moreover, among simply connected planar polygons with any fixed number of vertices greater than or equal to four, almost every polygon has simple Neumann spectrum. See [28 Theorem 1.1] for the precise statement and the meaning of “almost every.” Convexity is not required. The latter paper establishes generic simplicity of the spectrum, but does not prove a hot spots theorem.
These bounded Lipschitz domains have compact embedding H 1 ↪ L 2 H^1 \hookrightarrow L^2 H 1 ↪ L 2 and satisfy polynomial Neumann heat-kernel bounds. Consequently, under Hypothesis 1.1 , Theorem 1.2 and Corollary 2.4 apply to them. For lip domains and triangles, the hot spots conclusion is already known independently. See [4 Theorem 1(ii)] and [31 erratum, Theorems A and 4.1] , respectively.
The explicit spectral comparisons in these notes concern domains with nonsmooth boundary, particularly non-Lipschitz cusps, and the unbounded Gaussian horn of finite area. The exponential cusp and the Gaussian horn satisfy the compactness and simplicity assumptions but admit no polynomial Neumann heat-kernel bound, illustrating the broader scope of the compactness approach.
Simplicity is not necessary for a domain to satisfy the hot spots conclusion. The square and the disk are bounded simply connected Lipschitz domains whose first positive Neumann eigenvalue has multiplicity two, yet every eigenfunction in the corresponding eigenspace satisfies the nonstrict hot spots conclusion. These examples motivate the question of whether the simplicity assumption can be removed more generally.
The preceding results suggest two questions. Throughout, assume Hypothesis 1.1 , and impose no restriction on the multiplicity of the first positive Neumann eigenvalue.
Question 1: Lipschitz domains. Let Ω ⊂ R 2 \Omega\subset\mathbb{R}^2 Ω ⊂ R 2 be a bounded simply connected Lipschitz domain. Does every nonzero real eigenfunction u u u associated with λ 1 ( Ω ) \lambda_1(\Omega) λ 1 ( Ω ) satisfy
sup Ω u = lim δ ↓ 0 sup B δ ( Ω ) u , inf Ω u = lim δ ↓ 0 inf B δ ( Ω ) u , \sup_{\Omega} u = \lim_{\delta\downarrow0} \sup_{B_\delta(\Omega)} u,\qquad
\inf_{\Omega} u = \lim_{\delta\downarrow0} \inf_{B_\delta(\Omega)} u, Ω sup u = δ ↓ 0 lim B δ ( Ω ) sup u , Ω inf u = δ ↓ 0 lim B δ ( Ω ) inf u ,
where
B δ ( Ω ) = { x ∈ Ω : dist ( x , ∂ Ω ) < δ } ? B_\delta(\Omega) = \{x \in\Omega: \operatorname{dist}(x,\partial\Omega) < \delta\}? B δ ( Ω ) = { x ∈ Ω : dist ( x , ∂ Ω ) < δ }?
In particular, does this hold for every bounded simply connected polygonal domain, including those whose first positive Neumann eigenvalue may be multiple?
The result of Deng, Jiang, and Yang [22 ] does not settle this question, since the polygons considered here need not be convex.
Question 2: Domains with compact embedding. Does the same conclusion hold for every simply connected planar domain Ω \Omega Ω of finite area for which H 1 ( Ω ) ↪ L 2 ( Ω ) H^1(\Omega) \hookrightarrow L^2(\Omega) H 1 ( Ω ) ↪ L 2 ( Ω ) is compact, without assuming boundedness or any regularity of ∂ Ω \partial\Omega ∂ Ω ?
In both questions, eigenfunctions are understood through their smooth interior representatives, and infinite supremal or infimal values are allowed. Question 2 includes Question 1, since bounded Lipschitz domains have compact Sobolev embedding. The distinction separates the multiplicity question for bounded Lipschitz domains from its extension to domains with non-Lipschitz boundaries or unbounded domains of finite area.
The approximation argument gives a partial answer. The boundary-layer identities hold for at least one nonzero real first positive Neumann eigenfunction. The following proposition therefore gives a boundary-layer version of (HS3) in [8 ] . For bounded domains and eigenfunctions continuous on the closure, these identities are equivalent to the boundary inequalities in (HS3). The two questions above ask for the corresponding version of (HS2).
Proposition 6.1. Assume Hypothesis 1.1. Every simply connected planar domain Ω \Omega Ω of finite area with compact embedding H 1 ( Ω ) ↪ L 2 ( Ω ) H^{1}(\Omega) \hookrightarrow L^{2}(\Omega) H 1 ( Ω ) ↪ L 2 ( Ω ) admits a nonzero real first positive Neumann eigenfunction satisfying (2) , regardless of the multiplicity of λ 1 ( Ω ) \lambda_{1}(\Omega) λ 1 ( Ω ) .
Proof. Let F : D → Ω F:\mathbb{D}\to\Omega F : D → Ω be a conformal bijection, set Ω r = F ( r D ) \Omega_{r}=F(r\mathbb{D}) Ω r = F ( r D ) , and choose normalized real first positive Neumann eigenfunctions u j u_{j} u j on Ω r j \Omega_{r_{j}} Ω r j , where r j ↑ 1 r_{j}\uparrow1 r j ↑ 1 . Proposition 1.6, whose compactness and variational arguments do not require simplicity, gives a subsequence converging locally uniformly in Ω \Omega Ω to a normalized first positive Neumann eigenfunction v v v . If the supremum identity in (2) failed for v v v , Lemma 1.7 would give, for large j j j , an interior local maximum of u j u_{j} u j on the bounded smooth simply connected domain Ω r j \Omega_{r_{j}} Ω r j , contradicting Hypothesis 1.1. The infimum identity follows by applying the same argument to − v -v − v .
The remaining difficulty is to approximate each prescribed eigenfunction in the first positive eigenspace by first positive eigenfunctions on smooth approximating domains. When the limiting eigenvalue is multiple, the associated spectral cluster may split. Convergence of the entire cluster does not resolve this difficulty. Its elements need not be first positive eigenfunctions on the approximants, and Hypothesis 1.1 does not apply to their arbitrary linear combinations. The present argument therefore proves the conclusion for at least one limiting eigenfunction, but not for every eigenfunction in the eigenspace.
Remark 6.2. The same obstruction arises in the heat-kernel approach through spectral stability. The convergence of the spectral subspace associated with a multiple limiting eigenvalue does not ensure that each prescribed limiting eigenfunction can be approximated by first positive eigenfunctions. Thus this alternative approach does not by itself remove the simplicity assumption.
Further classes with polynomial heat-kernel bounds
This appendix presents two additional criteria for polynomial Neumann heat-kernel bounds and computes the quasihyperbolic order of the power cusps. These results supplement the examples and are not needed for the main proofs.
Quasihyperbolic Hölder domains
Here “Hölder domain” has the quasihyperbolic meaning used in [7 ] and the references therein, rather than the boundary-graph meaning of § 6 . Write δ ( z ) = dist ( z , ∂ Ω ) \delta(z)=\operatorname{dist}(z,\partial\Omega) δ ( z ) = dist ( z , ∂ Ω ) and define
k Ω ( z , w ) = inf γ ∫ γ ∣ d ζ ∣ δ ( ζ ) k_{\Omega}(z,w)=\inf_{\gamma}\int_{\gamma}\frac{|d\zeta|}{\delta(\zeta)} k Ω ( z , w ) = γ inf ∫ γ δ ( ζ ) ∣ d ζ ∣
where the infimum runs over rectifiable curves γ \gamma γ joining z z z to w w w in Ω \Omega Ω .
Proposition A.2. Every bounded planar domain satisfying (A.2) satisfies (17) for every finite q > 2 q>2 q > 2 . Consequently, for every N > 2 N>2 N > 2 there is a constant C Ω , N C_{\Omega,N} C Ω , N such that
p Ω ( t , z , w ) ≤ C Ω , N t − N / 2 , 0 < t ≤ 1 , p_{\Omega}(t,z,w) \le C_{\Omega,N}t^{-N/2}, \qquad0<t\le1, p Ω ( t , z , w ) ≤ C Ω , N t − N /2 , 0 < t ≤ 1 , for almost every z , w ∈ Ω z,w\in\Omega z , w ∈ Ω . In particular, H 1 ( Ω ) ↪ L 2 ( Ω ) H^1(\Omega)\hookrightarrow L^2(\Omega) H 1 ( Ω ) ↪ L 2 ( Ω ) is compact.
Proof. In (A.2) we may enlarge A A A and adjust B B B : the function log ( δ ( z 0 ) / δ ( z ) ) \log(\delta(z_0)/\delta(z)) log ( δ ( z 0 ) / δ ( z )) is bounded below on a bounded domain, so a larger coefficient is compensated by a larger additive constant. Write the coefficient as 1 / β 1/\beta 1/ β with 0 < β ≤ 1 0<\beta\le1 0 < β ≤ 1 . The theorem of Koskela–Onninen–Tyson [32 ] gives the Sobolev–Poincaré inequality
∥ u − u Ω ∥ L q ( Ω ) ≤ C ∥ ∇ u ∥ L p ( Ω ) , u Ω = ∣ Ω ∣ − 1 ∫ Ω u , \lVert u-u_{\Omega}\rVert_{L^q(\Omega)}\le C\lVert\nabla u\rVert_{L^p(\Omega)}, \qquad u_{\Omega}=|\Omega|^{-1}\int_{\Omega}u, ∥ u − u Ω ∥ L q ( Ω ) ≤ C ∥ ∇ u ∥ L p ( Ω ) , u Ω = ∣Ω ∣ − 1 ∫ Ω u ,
provided max { 1 , 2 − 2 β } < p < 2 \max\{1,2-2\beta\}<p<2 max { 1 , 2 − 2 β } < p < 2 and p ≤ q < 2 β p / ( 2 − p ) p\le q<2\beta p/(2-p) p ≤ q < 2 β p / ( 2 − p ) . Given any finite q > 2 q>2 q > 2 , choose p < 2 p<2 p < 2 sufficiently close to two. Finiteness of the area gives
∥ ∇ u ∥ p ≤ ∣ Ω ∣ 1 / p − 1 / 2 ∥ ∇ u ∥ 2 , ∥ u Ω ∥ q ≤ ∣ Ω ∣ 1 / q − 1 / 2 ∥ u ∥ 2 . \lVert\nabla u\rVert_p\le|\Omega|^{1/p-1/2}\lVert\nabla u\rVert_2, \qquad\lVert u_{\Omega}\rVert_q\le|\Omega|^{1/q-1/2}\lVert u\rVert_2. ∥ ∇ u ∥ p ≤ ∣Ω ∣ 1/ p − 1/2 ∥ ∇ u ∥ 2 , ∥ u Ω ∥ q ≤ ∣Ω ∣ 1/ q − 1/2 ∥ u ∥ 2 .
Together these give (17) . Proposition 3 of [18 ] then yields the kernel bound with N = 2 q / ( q − 2 ) N=2q/(q-2) N = 2 q / ( q − 2 ) , and hence with every N > 2 N>2 N > 2 . Compactness of the embedding follows from Remark 2.3 .
The argument does not yield the endpoint bound C t − 1 Ct^{-1} C t − 1 .
:::
Quasihyperbolic order of the power cusps
The power cusps Ω α \Omega_\alpha Ω α of § 6.3 have positive quasihyperbolic order. To avoid confusing the cusp exponent α > 1 \alpha> 1 α > 1 with the order, put η = 1 − 1 / α ∈ ( 0 , 1 ) \eta= 1 - 1/\alpha\in(0,1) η = 1 − 1/ α ∈ ( 0 , 1 ) and fix z 0 = ( 1 / 2 , 0 ) z_0 = (1/2,0) z 0 = ( 1/2 , 0 ) . Along the axis, as x ↓ 0 x \downarrow0 x ↓ 0 ,
δ ( x , 0 ) ≍ x α , k Ω α ( ( x , 0 ) , z 0 ) ≍ x 1 − α ≍ δ ( x , 0 ) − η \delta(x,0) \asymp x^\alpha,\qquad k_{\Omega_\alpha}((x,0),z_0) \asymp x^{1-\alpha} \asymp\delta(x,0)^{-\eta} δ ( x , 0 ) ≍ x α , k Ω α (( x , 0 ) , z 0 ) ≍ x 1 − α ≍ δ ( x , 0 ) − η
For the upper bound integrate r − α r^{-\alpha} r − α along the axis. For the lower bound, any path from ( x , 0 ) (x,0) ( x , 0 ) to z 0 z_0 z 0 must traverse the horizontal coordinates between x x x and 1 / 2 1/2 1/2 , while its boundary distance at horizontal coordinate r r r is at most r α r^\alpha r α . Near the tip, the bounded slopes of the graphs make boundary distance comparable to the vertical gap from the nearer graph. Thus, for an off-axis point ( x , y ) (x,y) ( x , y ) near the tip, a vertical segment to the axis costs at most C [ 1 + log ( x α / δ ( x , y ) ) ] C[1 + \log(x^\alpha/\delta(x,y))] C [ 1 + log ( x α / δ ( x , y ))] in quasihyperbolic length. Since δ ( x , y ) ≤ x α \delta(x,y) \leq x^\alpha δ ( x , y ) ≤ x α and log ( 1 / d ) ≤ C η d − η \log(1/d) \leq C_\eta d^{-\eta} log ( 1/ d ) ≤ C η d − η for small d d d , this gives (A.3) with η = 1 − 1 / α \eta= 1 - 1/\alpha η = 1 − 1/ α . The remaining Lipschitz boundary contributes only logarithmic growth. The axis estimate excludes order zero and every positive order smaller than η \eta η .
Thus, for every 0 < η < 1 0 < \eta< 1 0 < η < 1 there is a positive-order example with a polynomial Neumann bound, whose exponent is
N = α + 1 = 2 − η 1 − η N = \alpha+ 1 = \frac{2-\eta}{1-\eta} N = α + 1 = 1 − η 2 − η
This exponent is optimal for this cusp family [18 Example 6] . Whether the positive-order quasihyperbolic condition alone implies a polynomial Neumann bound on every domain satisfying it is not addressed here.
Finite decompositions and a slit domain
Proposition A.3. Let Ω \Omega Ω be bounded. Suppose there are finitely many open subdomains Ω 1 , … , Ω m ⊂ Ω \Omega_1,\ldots,\Omega_m \subset\Omega Ω 1 , … , Ω m ⊂ Ω whose union covers Ω \Omega Ω up to a null set. If each Ω j \Omega_j Ω j satisfies (17) with an exponent q j > 2 q_j > 2 q j > 2 , then Ω \Omega Ω satisfies (17) with q = min j q j > 2 q = \min_j q_j > 2 q = min j q j > 2 . Consequently Ω \Omega Ω has a polynomial Neumann heat-kernel bound and compact Sobolev embedding.
Proof. By Hölder’s inequality and ∣ Ω j ∣ < ∞ |\Omega_j| < \infty ∣ Ω j ∣ < ∞ , the estimate on each piece holds with the common exponent q q q . Restriction takes u ∈ H 1 ( Ω ) u \in H^1(\Omega) u ∈ H 1 ( Ω ) into H 1 ( Ω j ) H^1(\Omega_j) H 1 ( Ω j ) . Since 2 / q < 1 2/q < 1 2/ q < 1 ,
∥ u ∥ L q ( Ω ) 2 ≤ ( ∑ j = 1 m ∥ u ∥ L q ( Ω j ) q ) 2 / q ≤ ∑ j = 1 m ∥ u ∥ L q ( Ω j ) 2 ≤ C ∑ j = 1 m ∥ u ∥ H 1 ( Ω j ) 2 ≤ C m ∥ u ∥ H 1 ( Ω ) 2 . \begin{aligned}
\lVert u\rVert_{L^q(\Omega)}^2 \leq\left(\sum_{j=1}^{m}\lVert u\rVert_{L^q(\Omega_j)}^q\right)^{2/q} \leq\sum_{j=1}^{m}\lVert u\rVert_{L^q(\Omega_j)}^2 \\
&\leq C\sum_{j=1}^{m}\lVert u\rVert_{H^1(\Omega_j)}^2 \leq Cm\lVert u\rVert_{H^1(\Omega)}^2.
\end{aligned} ∥ u ∥ L q ( Ω ) 2 ≤ ( j = 1 ∑ m ∥ u ∥ L q ( Ω j ) q ) 2/ q ≤ j = 1 ∑ m ∥ u ∥ L q ( Ω j ) 2 ≤ C j = 1 ∑ m ∥ u ∥ H 1 ( Ω j ) 2 ≤ C m ∥ u ∥ H 1 ( Ω ) 2 .
The equivalence between (17) and the heat-kernel bound, together with Remark 2.3 , completes the proof. No extension or matching of functions across the interfaces is needed.
For example, the slit disk
Ω = D ∖ { ( x , 0 ) : 0 ≤ x < 1 } \Omega= \mathbb{D} \setminus\{(x,0): 0 \leq x < 1\} Ω = D ∖ {( x , 0 ) : 0 ≤ x < 1 }
is bounded and simply connected, but it is neither a Lipschitz domain nor a Jordan domain, and it does not satisfy the one-sided Hölder graph condition of § 6 along the slit. Its upper and lower half-disks are Lipschitz subdomains covering it up to a null set, so Proposition A.3 supplies a polynomial bound with any chosen N > 2 N > 2 N > 2 . More generally, finite decompositions into Hölder graph domains or quasidisks give the same conclusion for some finite N N N .
References