2020 Mathematics Subject Classification. Primary 35P15; Secondary 35J40, 49Q10, 74K20
Introduction and main result
In 1955–1956, Payne, Pólya and Weinberger [26 ] conjectured that the disk maximizes the ratio of the first two Dirichlet eigenvalues of a planar membrane. Its Euclidean extension states that, for a bounded domain Ω ⊂ R d \Omega\subset\mathbb{R}^{d} Ω ⊂ R d , d ≥ 2 d \ge2 d ≥ 2 ,
λ 2 ( Ω ) λ 1 ( Ω ) ≤ λ 2 ( B ) B λ 1 ( B ) = j d / 2 , 1 2 j d / 2 − 1 , 1 2 . \frac{\lambda_{2}(\Omega)}{\lambda_{1}(\Omega)} \le\frac{\lambda_{2(B)}{B}}{\lambda_{1}(B)} = \frac{j_{d/2,1}^{2}}{j_{d/2-1,1}^{2}}. λ 1 ( Ω ) λ 2 ( Ω ) ≤ λ 1 ( B ) λ 2 ( B ) B = j d /2 − 1 , 1 2 j d /2 , 1 2 .
Here R d \mathbb{R}^{d} R d is Euclidean space, B = { x ∈ R d : ∣ x ∣ < 1 } B = \{x \in\mathbb{R}^{d} : |x| < 1\} B = { x ∈ R d : ∣ x ∣ < 1 } is the unit ball in R d \mathbb{R}^{d} R d , λ k ( Ω ) \lambda_k(\Omega) λ k ( Ω ) are the eigenvalues of − Δ u = λ u -\Delta u = \lambda u − Δ u = λ u with u = 0 u = 0 u = 0 on ∂ Ω \partial\Omega ∂ Ω , ordered with multiplicity, and j ν , 1 j_{\nu,1} j ν , 1 is the first positive zero of the Bessel function J ν J_\nu J ν . The Laplacian is Δ = ∑ i = 1 d ∂ x i 2 \Delta= \sum_{i=1}^{d} \partial_{x_i}^{2} Δ = ∑ i = 1 d ∂ x i 2 . Ashbaugh and Benguria proved this sharp inequality in every dimension, with equality for balls [4 , 5 ] . The planar ratio question also appears as Problem 77 in Yau’s problem section in [27 ] .
The membrane problem connects spectral geometry with acoustics: the vibration frequencies are proportional to λ k ( Ω ) \sqrt{\lambda_k(\Omega)} λ k ( Ω ) . Kac’s question of whether one can hear the shape of a drum asks whether the full Dirichlet spectrum determines the domain [21 ] . The PPW problem instead asks what geometry can be inferred from an extremal relation between just two frequencies. For a homogeneous isotropic thin elastic plate, bending leads to a fourth-order equation. In the linear model w t t + c Δ 2 w = 0 w_{tt} + c\Delta^{2}w = 0 w tt + c Δ 2 w = 0 , where w w w is the transverse displacement and c > 0 c > 0 c > 0 depends on the material and thickness, the squared frequencies are c c c times the clamped eigenvalues [17 Chapter 1] . Their ratio is independent of this material factor. Interior clamps provide a means of changing the low vibration frequencies; the small-patch analysis in [23 ] provides a mathematical model for this effect. The higher-dimensional problem is the corresponding elliptic spectral problem.
Motivated by the membrane inequality, Ashbaugh formulated in 1999 the analogous ball-ratio conjecture for the vibrating clamped plate [3 Section 4, equation (4.6)] . To state it, let H 2 ( Ω ) H^{2}(\Omega) H 2 ( Ω ) be the Sobolev space of functions with square integrable weak derivatives through order two, and let H 0 2 ( Ω ) H_{0}^{2}(\Omega) H 0 2 ( Ω ) be the closure in that space of the smooth compactly supported functions C c ∞ ( Ω ) C_{c}^{\infty}(\Omega) C c ∞ ( Ω ) . The clamped eigenvalues Γ k ( Ω ) \Gamma_k(\Omega) Γ k ( Ω ) , ordered increasingly and repeated according to multiplicity, are defined by
∫ Ω Δ u Δ v d x = Γ ∫ Ω u v d x for all v ∈ H 0 2 ( Ω ) , 0 ≠ u ∈ H 0 2 ( Ω ) . \int_{\Omega} \Delta u \Delta v \, dx = \Gamma\int_{\Omega} uv \, dx \qquad\text{for all } v \in H_{0}^{2}(\Omega), \quad0 \ne u \in H_{0}^{2}(\Omega). ∫ Ω Δ u Δ v d x = Γ ∫ Ω uv d x for all v ∈ H 0 2 ( Ω ) , 0 = u ∈ H 0 2 ( Ω ) .
On a smooth domain this is equivalent to
Δ 2 u = Γ u in Ω , u = ∂ ν u = 0 on ∂ Ω , \Delta^{2}u = \Gamma u \quad\text{in } \Omega, \qquad u = \partial_{\nu}u = 0 \quad\text{on } \partial\Omega, Δ 2 u = Γ u in Ω , u = ∂ ν u = 0 on ∂ Ω ,
where ν \nu ν is the outward unit normal and ∂ ν \partial_{\nu} ∂ ν denotes normal differentiation. Put
R ( Ω ) = Γ 2 ( Ω ) Γ 1 ( Ω ) , ρ d = R ( B ) . R(\Omega) = \frac{\Gamma_{2}(\Omega)}{\Gamma_{1}(\Omega)}, \qquad\rho_d = R(B). R ( Ω ) = Γ 1 ( Ω ) Γ 2 ( Ω ) , ρ d = R ( B ) .
The conjectured inequality is R ( Ω ) ≤ ρ d R(\Omega) \le\rho_d R ( Ω ) ≤ ρ d . The identity Γ k ( t Ω ) = t − 4 Γ k ( Ω ) \Gamma_k(t\Omega) = t^{-4}\Gamma_k(\Omega) Γ k ( t Ω ) = t − 4 Γ k ( Ω ) for t > 0 t > 0 t > 0 makes R R R invariant under similarities, namely translations, orthogonal transformations and dilations.
Two related plate problems must be distinguished from this ratio question. Rayleigh’s conjecture minimizes the first vibrating-plate eigenvalue at fixed volume; it was proved in dimensions two and three by Nadirashvili and Ashbaugh–Benguria [6 , 24 ] . Plate buckling instead has the equation Δ 2 u = − Λ Δ u \Delta^{2}u = -\Lambda\Delta u Δ 2 u = − ΛΔ u , with spectral parameter Λ \Lambda Λ , and a Rayleigh quotient involving the Dirichlet energy in the denominator. Universal inequalities for clamped vibration eigenvalues, including [13 , 20 ] , estimate higher eigenvalues in terms of lower ones by constants independent of the domain. Such estimates do not establish the particular constant ρ d \rho_d ρ d . Related inequalities and ball spectral formulas for polyharmonic operators are developed in [10 ] .
The membrane analogy is delicate because fourth-order Dirichlet inverses need not preserve positivity. Boggio’s formula gives a positive Green function on a ball [9 ] , whereas the behavior on general domains is substantially different [18 , 19 ] . A second difficulty is that small clamped inclusions can leave a nonzero constraint in the limiting energy space. Kropinski, Lindsay and Ward [22 ] , and Lindsay, Hao and Sommese [23 ] , analyzed planar small-clamp limits and the surviving point-value conditions. The latter work also studied ring configurations and their low eigenvalues. Felli and Romani [15 ] obtained capacitary eigenvalue asymptotics for polyharmonic problems with small holes. These works supply the small-support perspective underlying our construction. Our contribution is to produce a strict violation of the ball comparison and to realize it by connected smooth domains in every dimension.
Dai, Sun and Zhang [14 ] disproved the corresponding buckling eigenvalue-ratio conjecture in every dimension. The spherical constraints, codimension-two holes and logarithmic cutoff mechanism are shared by the two constructions; we use their geometric arguments by reference. The spectral calculation is different. The denominator is ∫ u 2 d x \int u^2\,dx ∫ u 2 d x here, instead of ∫ ∣ ∇ u ∣ 2 d x \int\lvert\nabla u\rvert^2\,dx ∫ ∣ ∇ u ∣ 2 d x for buckling, and the ball modes involve both J ν J_\nu J ν and I ν I_\nu I ν . In the buckling construction the relative first-order shift is 3 ϵ / 2 3\epsilon/2 3 ϵ /2 and the relative second-order gain in the ratio is ϵ 2 / 8 \epsilon^2/8 ϵ 2 /8 . Here the first-order shift is 3 ϵ 3\epsilon 3 ϵ , while the second-order coefficient depends on h ν ( κ ) = κ I ν ′ ( κ ) / I ν ( κ ) h_\nu(\kappa)=\kappa I_\nu'(\kappa)/I_\nu(\kappa) h ν ( κ ) = κ I ν ′ ( κ ) / I ν ( κ ) ; its strict comparison between the first and second ball modes is the essential spectral step. Moreover, we transfer the strict inequality through two spectral limits, identifying the limiting trace spaces and using compactness in L 2 L^2 L 2 . This differs from the quantitative surface-penalty and trial-function estimates in [14 ] .
A codimension-two submanifold has dimension d − 2 d-2 d − 2 and a two-dimensional normal space. A closed tubular neighborhood of radius η \eta η consists of points at distance at most η \eta η from it. We write S m \mathbb{S}^m S m for the unit sphere in R m + 1 \mathbb{R}^{m+1} R m + 1 . Our main theorem is as follows.
Theorem 1.1. For every integer d ≥ 2 d\ge2 d ≥ 2 , there is a connected bounded domain Ω ⊂ R d \Omega\subset\mathbb{R}^d Ω ⊂ R d with C ∞ C^\infty C ∞ boundary such that
R ( Ω ) > ρ d . R(\Omega)>\rho_d. R ( Ω ) > ρ d . It can be obtained by removing from B B B finitely many disjoint closed tubular neighborhoods of smooth embedded copies of S d − 2 \mathbb{S}^{d-2} S d − 2 . In dimension two these are pairs of small disk holes; in dimension three they are solid torus holes (see following Figure 1).
Both eigenvalues increase when extra constraints are imposed, yet their ratio can exceed the ball value. Indeed, domain inclusion gives
Ω ⊂ B ⟹ Γ 1 ( Ω ) ≥ Γ 1 ( B ) , Γ 2 ( Ω ) ≥ Γ 2 ( B ) . \Omega\subset B \quad\Longrightarrow\quad\Gamma_1(\Omega)\ge\Gamma_1(B),\qquad\Gamma_2(\Omega)\ge\Gamma_2(B). Ω ⊂ B ⟹ Γ 1 ( Ω ) ≥ Γ 1 ( B ) , Γ 2 ( Ω ) ≥ Γ 2 ( B ) .
but it does not compare their relative increases. Our construction separates these increases only at second order. Thus the first-order calculation along this family does not detect the failure of the conjecture. The sign of the second-order term must therefore be determined separately; the membrane comparison gives no such information.
The proof has three steps. First, require u = 0 u=0 u = 0 on an interior sphere of radius 1 − ϵ 1-\epsilon 1 − ϵ , leaving its normal derivative free. This is an intermediate constraint, not yet the boundary of a hole. For each fixed ball mode, Lemma 3.1 gives
constrained eigenvalue ball eigenvalue = 1 + 3 ϵ + 23 + h ν ( κ ) 4 ϵ 2 + O ( ϵ 3 ) . \frac{\text{constrained eigenvalue}}{\text{ball eigenvalue}}=1+3\epsilon+\frac{23+h_\nu(\kappa)}{4}\epsilon^2+O(\epsilon^3). ball eigenvalue constrained eigenvalue = 1 + 3 ϵ + 4 23 + h ν ( κ ) ϵ 2 + O ( ϵ 3 ) .
The common term 3 ϵ 3\epsilon 3 ϵ cancels in the ratio. The function h ν ( κ ) h_\nu(\kappa) h ν ( κ ) is strictly larger for every mode belonging to the second ball eigenvalue than for the first mode. Proposition 3.2 therefore proves a positive second-order gain. Two checks are essential: the eigenvalues constructed by the radial calculation must actually occupy the first two positions, and the remainder must be smaller than the positive quadratic term. The proof checks both, including the possible multiplicity of the second eigenvalue.
Second, replace the sphere by finitely many latitudes. In dimension 2 a latitude consists of two points; in dimension 3 it is a circle. Adding more latitudes imposes more value conditions. The corresponding eigenvalues increase to those of the spherical constraint, as proved in § 4 . Once the strict gain on the sphere is fixed, a sufficiently large but finite collection retains it.
Third, enlarge each latitude into a small smooth hole. There is a boundary-condition issue here: a hole requires both u = 0 u=0 u = 0 and ∂ ν u = 0 \partial_\nu u=0 ∂ ν u = 0 , whereas the latitude requires only u = 0 u=0 u = 0 . Section 5 proves that the additional derivative condition can be imposed with an error tending to 0 in H 2 H^2 H 2 as the holes shrink. The reason is the two-dimensional space normal to each latitude: a cutoff varying with the logarithm of the normal distance has sufficiently small derivative cost. The same estimate works in every dimension. We use the geometry in [14 ] to keep the complement connected.
The order of these choices matters. We first fix a sphere with a positive spectral gain, then fix finitely many latitudes, and only then choose positive hole radii. Section 6 bounds the error at each step by less than half the available gain. The resulting counterexample is a smooth connected domain. The remainder estimates and the two approximation limits are both needed to preserve the strict inequality. The identification of the limiting spaces complements general higher-order domain stability results [2 , 12 ] .
Section 2 collects the standard variational, trace and ball spectral facts. Section 3 proves the strict increase under a spherical value constraint. Section 4 treats approximation by finitely many latitudes, and Section 5 treats their replacement by smooth holes. The choices preserving the strict gap are combined in Section 6, which contains the proof of Theorem 1.1 .
Variational tools and spectral data on the ball
We collect the standard facts used below in three lemmas. The first fixes the direction of the variational comparisons, the second specifies the traces and regularity needed at an interior interface, and the third records the ball formulas. The separation of variables and the defining series of the Bessel functions are classical and are not rederived. Throughout, B B B is the unit ball, d x dx d x is Lebesgue measure, and surface integrals use the induced measure. The form spaces and coefficient vectors are taken over the real numbers.
Closed form spaces and traces
Write
Q ( u , v ) = ∫ B Δ u Δ v d x , ( u , v ) = ∫ B u v d x . Q(u,v)=\int_B \Delta u\Delta v\,dx,\qquad(u,v)=\int_B uv\,dx. Q ( u , v ) = ∫ B Δ u Δ v d x , ( u , v ) = ∫ B uv d x .
Functions in H 0 2 ( Ω ) H_0^2(\Omega) H 0 2 ( Ω ) , with Ω ⊂ B \Omega\subset B Ω ⊂ B , are identified with their zero extensions. All form subspaces below are closed in H 0 2 ( B ) H_0^2(B) H 0 2 ( B ) and infinite dimensional.
Lemma 2.1 (Variational comparisons). The norm ∥ Δ u ∥ 2 \lVert\Delta u\rVert_2 ∥ Δ u ∥ 2 is equivalent to ∥ u ∥ H 2 ( B ) \lVert u\rVert_{H^2(B)} ∥ u ∥ H 2 ( B ) on H 0 2 ( B ) H_0^2(B) H 0 2 ( B ) , and the inclusion H 0 2 ( B ) ↪ L 2 ( B ) H_0^2(B)\hookrightarrow L^2(B) H 0 2 ( B ) ↪ L 2 ( B ) is compact. For every closed form subspace V V V , its eigenvalues satisfy
λ k ( V ) = min E ⊂ V dim E = k max 0 ≠ u ∈ E Q ( u , u ) ∥ u ∥ 2 2 , V 1 ⊂ V 2 ⟹ λ k ( V 2 ) ≤ λ k ( V 1 ) . (1) \lambda_k(V)=\min_{\substack{E\subset V\\ \dim E=k}}\ \max_{0\ne u\in E}\frac{Q(u,u)}{\lVert u\rVert_2^2},\qquad V_1\subset V_2\ \Longrightarrow\ \lambda_k(V_2)\leq\lambda_k(V_1). \tag*{(1)} λ k ( V ) = E ⊂ V d i m E = k min 0 = u ∈ E max ∥ u ∥ 2 2 Q ( u , u ) , V 1 ⊂ V 2 ⟹ λ k ( V 2 ) ≤ λ k ( V 1 ) . (1)
If u 1 , … , u k u_1,\ldots,u_k u 1 , … , u k are L 2 L^2 L 2 -orthonormal eigenvectors on V V V , then
∥ ∑ i = 1 k c i u i ∥ 2 2 = ∑ i = 1 k ∣ c i ∣ 2 , Q ( ∑ i c i u i , ∑ i c i u i ) = ∑ i λ i ( V ) ∣ c i ∣ 2 ≤ λ k ( V ) ∑ i ∣ c i ∣ 2 . (2) \left\|\sum_{i=1}^{k} c_i u_i\right\|_2^2=\sum_{i=1}^{k}|c_i|^2,\qquad Q\left(\sum_i c_i u_i,\sum_i c_i u_i\right)=\sum_i\lambda_i(V)|c_i|^2\leq\lambda_k(V)\sum_i|c_i|^2. \tag*{(2)} i = 1 ∑ k c i u i 2 2 = i = 1 ∑ k ∣ c i ∣ 2 , Q ( i ∑ c i u i , i ∑ c i u i ) = i ∑ λ i ( V ) ∣ c i ∣ 2 ≤ λ k ( V ) i ∑ ∣ c i ∣ 2 . (2)
Proof. These are the standard consequences of compact embedding and the min–max principle; see [17 Chapter 2] and [1 Chapters 4 and 6] . For clarity, the norm comparison follows from
∥ D 2 u ∥ 2 2 = ∥ Δ u ∥ 2 2 , ∥ ∇ u ∥ 2 ≤ C ∥ D 2 u ∥ 2 , ∥ u ∥ 2 ≤ C ∥ ∇ u ∥ 2 . \|D^2u\|_2^2=\|\Delta u\|_2^2,\qquad\|\nabla u\|_2\leq C\|D^2u\|_2,\qquad\|u\|_2\leq C\|\nabla u\|_2. ∥ D 2 u ∥ 2 2 = ∥Δ u ∥ 2 2 , ∥∇ u ∥ 2 ≤ C ∥ D 2 u ∥ 2 , ∥ u ∥ 2 ≤ C ∥∇ u ∥ 2 .
The first identity holds by integration by parts for C c ∞ ( B ) C^\infty_c(B) C c ∞ ( B ) and then by density; the other two are Poincaré inequalities. The weak equations give Q ( u i , u j ) = λ i ( V ) δ i j Q(u_i,u_j)=\lambda_i(V)\delta_{ij} Q ( u i , u j ) = λ i ( V ) δ ij , which proves (2.2). Formula (2.1) then gives the stated inclusion comparison.
Lemma 2.2 (Traces and interface regularity). Let S ⋐ B S\Subset B S ⋐ B be a compact smooth hypersurface and let Σ ⋐ B \Sigma\Subset B Σ ⋐ B be a compact smooth submanifold of codimension two. The value traces are bounded maps
Tr S : H 2 ( B ) ⟶ H 3 / 2 ( S ) , Tr Σ : H 2 ( B ) ⟶ H 1 ( Σ ) . \operatorname{Tr}_S:H^2(B)\longrightarrow H^{3/2}(S),\qquad\operatorname{Tr}_\Sigma:H^2(B)\longrightarrow H^1(\Sigma). Tr S : H 2 ( B ) ⟶ H 3/2 ( S ) , Tr Σ : H 2 ( B ) ⟶ H 1 ( Σ ) . For d = 2 d=2 d = 2 , the second trace means evaluation at finitely many points. On every bounded interval I I I ,
H 2 ( I ) ↪ C 1 , 1 / 2 ( I ‾ ) , ∣ f ′ ( s ) − f ′ ( t ) ∣ ≤ ∣ s − t ∣ 1 / 2 ∥ f ′ ′ ∥ L 2 ( I ) . (3) H^2(I)\hookrightarrow C^{1,1/2}(\overline{I}),\qquad|f'(s)-f'(t)|\leq|s-t|^{1/2}\|f''\|_{L^2(I)}. \tag*{(3)} H 2 ( I ) ↪ C 1 , 1/2 ( I ) , ∣ f ′ ( s ) − f ′ ( t ) ∣ ≤ ∣ s − t ∣ 1/2 ∥ f ′′ ∥ L 2 ( I ) . (3) If u u u is smooth on each side of S S S , with square-integrable derivatives through order two up to S S S , and
[ u ] S = 0 , [ ∇ u ] S = 0 , [u]_S=0,\qquad[\nabla u]_S=0, [ u ] S = 0 , [ ∇ u ] S = 0 , then u ∈ H 2 ( B ) u\in H^2(B) u ∈ H 2 ( B ) . On a smooth bounded domain Ω \Omega Ω ,
H 0 2 ( Ω ) = { u ∈ H 2 ( Ω ) : Tr ∂ Ω u = 0 , Tr ∂ Ω ∂ ν u = 0 } . (4) H^2_0(\Omega)=\{u\in H^2(\Omega):\operatorname{Tr}_{\partial\Omega}u=0,\operatorname{Tr}_{\partial\Omega}\partial_\nu u=0\}. \tag*{(4)} H 0 2 ( Ω ) = { u ∈ H 2 ( Ω ) : Tr ∂ Ω u = 0 , Tr ∂ Ω ∂ ν u = 0 } . (4) Proof. The trace and embedding statements are classical; see [1 Chapters 4 and 7] and [17 Chapter 2] . The loss of smoothness for restriction to codimension q q q is q / 2 q/2 q /2 , so here 2 − 1 / 2 = 3 / 2 2-1/2=3/2 2 − 1/2 = 3/2 and 2 − 2 / 2 = 1 2-2/2=1 2 − 2/2 = 1 . In a flat chart this also follows from Fourier Cauchy–Schwarz and
∫ R q ( Q 2 + ∣ η ∣ 2 ) − 2 d η = C q Q q − 4 , Q = ( 1 + ∣ ξ ∣ 2 ) 1 / 2 , q = 1 , 2. \int_{\mathbb{R}^q}(Q^2+|\eta|^2)^{-2}\,d\eta=C_qQ^{q-4},\qquad Q=(1+|\xi|^2)^{1/2},\qquad q=1,2. ∫ R q ( Q 2 + ∣ η ∣ 2 ) − 2 d η = C q Q q − 4 , Q = ( 1 + ∣ ξ ∣ 2 ) 1/2 , q = 1 , 2. Smooth charts and a partition of unity give the displayed surface traces. Formula (2.3) follows from f ′ ( s ) − f ′ ( t ) = ∫ t s f ′ ′ ( x ) d x f'(s)-f'(t)=\int_t^s f''(x)\,dx f ′ ( s ) − f ′ ( t ) = ∫ t s f ′′ ( x ) d x and Cauchy–Schwarz. For the interface assertion, the distributional derivative of a piecewise smooth function is its piecewise derivative plus its jump times the corresponding normal component of surface measure. Thus
[ u ] S = 0 ⟹ ∂ i u = ( ∂ i u ) p w , [ ∂ i u ] S = 0 ⟹ ∂ j ∂ i u = ( ∂ j ∂ i u ) p w ∈ L 2 . [u]_S=0\Longrightarrow\partial_i u=(\partial_i u)_{\mathrm{pw}},\qquad[\partial_i u]_S=0\Longrightarrow\partial_j\partial_i u=(\partial_j\partial_i u)_{\mathrm{pw}}\in L^2. [ u ] S = 0 ⟹ ∂ i u = ( ∂ i u ) pw , [ ∂ i u ] S = 0 ⟹ ∂ j ∂ i u = ( ∂ j ∂ i u ) pw ∈ L 2 . Here the subscript pw means that the derivative is taken separately on the two sides. The characterization (2.4) is the standard kernel theorem for the two boundary traces.
Ball modes and the Bessel comparison
Put a = ( d − 2 ) / 2 a=(d-2)/2 a = ( d − 2 ) /2 . We use J ν J_\nu J ν and I ν I_\nu I ν for the ordinary and modified Bessel functions of the first kind; see [25 Sections 10.2, 10.25 and 10.29] . Define
W ν ( s ) = J ν ( s ) I ν ′ ( s ) − J ν ′ ( s ) I ν ( s ) , h ν ( s ) = s I ν ′ ( s ) I ν ( s ) . W_\nu(s)=J_\nu(s)I_\nu'(s)-J_\nu'(s)I_\nu(s),\qquad h_\nu(s)=\frac{sI_\nu'(s)}{I_\nu(s)}. W ν ( s ) = J ν ( s ) I ν ′ ( s ) − J ν ′ ( s ) I ν ( s ) , h ν ( s ) = I ν ( s ) s I ν ′ ( s ) .
Lemma 2.3 (Spectral data on the ball). Let Y l Y_l Y l be a spherical harmonic of degree l l l and put ν = l + a \nu=l+a ν = l + a . The radial operator and regular separated solutions are
Δ ( F ( x ) Y l ( θ ) ) = ( D l F ) ( x ) Y l ( θ ) , D l F = F ′ ′ + d − 1 x F ′ − l ( l + d − 2 ) x 2 F , \Delta(F(x)Y_l(\theta))=(D_lF)(x)Y_l(\theta),\qquad D_lF=F''+\frac{d-1}{x}F'-\frac{l(l+d-2)}{x^2}F, Δ ( F ( x ) Y l ( θ )) = ( D l F ) ( x ) Y l ( θ ) , D l F = F ′′ + x d − 1 F ′ − x 2 l ( l + d − 2 ) F , D l 2 F = k 4 F ⟹ F ( x ) = x − a { A J ν ( k x ) + B I ν ( k x ) } . (5) D_l^2F=k^4F\quad\Longrightarrow\quad F(x)=x^{-a}\{AJ_\nu(kx)+BI_\nu(kx)\}. \tag*{(5)} D l 2 F = k 4 F ⟹ F ( x ) = x − a { A J ν ( k x ) + B I ν ( k x )} . (5) The ball eigenvalues in this sector are k 4 k^4 k 4 with W ν ( k ) = 0 W_\nu(k)=0 W ν ( k ) = 0 . Every positive zero κ \kappa κ of W ν W_\nu W ν is simple, and
W ν ′ ( s ) = 2 J ν ( s ) I ν ( s ) − W ν ( s ) s , W ν ′ ′ ( κ ) W ν ′ ( κ ) = 2 I ν ′ ( κ ) I ν ( κ ) − 1 κ . (6) W_\nu'(s)=2J_\nu(s)I_\nu(s)-\frac{W_\nu(s)}{s},\qquad\frac{W_\nu''(\kappa)}{W_\nu'(\kappa)}=2\frac{I_\nu'(\kappa)}{I_\nu(\kappa)}-\frac{1}{\kappa}. \tag*{(6)} W ν ′ ( s ) = 2 J ν ( s ) I ν ( s ) − s W ν ( s ) , W ν ′ ( κ ) W ν ′′ ( κ ) = 2 I ν ( κ ) I ν ′ ( κ ) − κ 1 . (6) The first ball eigenvalue is simple and has a positive radial eigenfunction. Each ball eigenspace is a finite sum of separated harmonic sectors.
Proof. The separated formulas are standard; see [10 Appendix A] and [8 ] . The clamped boundary determinant is k W ν ( k ) kW_\nu(k) k W ν ( k ) . We record only the identities used in the perturbation calculation. The two Bessel equations give
W ν ′ = J ν I ν ′ ′ − J ν ′ ′ I ν = 2 J ν I ν − W ν s . W_\nu'=J_\nu I_\nu''-J_\nu''I_\nu=2J_\nu I_\nu-\frac{W_\nu}{s}. W ν ′ = J ν I ν ′′ − J ν ′′ I ν = 2 J ν I ν − s W ν .
Since I ν ( κ ) > 0 I_\nu(\kappa)>0 I ν ( κ ) > 0 , the equalities J ν ( κ ) = W ν ( κ ) = 0 J_\nu(\kappa)=W_\nu(\kappa)=0 J ν ( κ ) = W ν ( κ ) = 0 would imply J ν ′ ( κ ) = 0 J_\nu'(\kappa)=0 J ν ′ ( κ ) = 0 , contrary to uniqueness for the Bessel equation. Hence W ν ′ ( κ ) = 2 J ν ( κ ) I ν ( κ ) ≠ 0 W_\nu'(\kappa)=2J_\nu(\kappa)I_\nu(\kappa)\ne0 W ν ′ ( κ ) = 2 J ν ( κ ) I ν ( κ ) = 0 . Differentiating once more and using J ν ′ ( κ ) / J ν ( κ ) = I ν ′ ( κ ) / I ν ( κ ) J_\nu'(\kappa)/J_\nu(\kappa)=I_\nu'(\kappa)/I_\nu(\kappa) J ν ′ ( κ ) / J ν ( κ ) = I ν ′ ( κ ) / I ν ( κ ) gives the second identity in (6) . The assertion about the first eigenfunction follows from Boggio’s positive Green function and rotational invariance [9 , 17 ] . Finite multiplicity follows from compactness. □
We shall use the following standard modified-Bessel comparison. For fixed ν ≥ 0 \nu\ge0 ν ≥ 0 , s ↦ h ν ( s ) s\mapsto h_\nu(s) s ↦ h ν ( s ) is strictly increasing on ( 0 , + ∞ ) (0,+\infty) ( 0 , + ∞ ) by [7 Theorem 1(b)] . For clarity, comparison of two orders has the short identity
h μ ( s ) − h ν ( s ) = ( μ 2 − ν 2 ) ∫ 0 1 I μ ( s x ) I ν ( s x ) I μ ( s ) I ν ( s ) d x x > 0 , μ > ν ≥ 0. (7) h_\mu(s)-h_\nu(s)=(\mu^2-\nu^2)\int_0^1\frac{I_\mu(sx)I_\nu(sx)}{I_\mu(s)I_\nu(s)}\frac{\mathrm{d}x}{x}>0,\qquad\mu>\nu\ge0. \tag*{(7)} h μ ( s ) − h ν ( s ) = ( μ 2 − ν 2 ) ∫ 0 1 I μ ( s ) I ν ( s ) I μ ( s x ) I ν ( s x ) x d x > 0 , μ > ν ≥ 0. (7)
Indeed, for v α ( x ) = I α ( s x ) / I α ( s ) v_\alpha(x)=I_\alpha(sx)/I_\alpha(s) v α ( x ) = I α ( s x ) / I α ( s ) the modified Bessel equation implies
d d x { x ( v μ ′ v ν − v ν ′ v μ ) } = μ 2 − ν 2 x v μ v ν . \frac{\mathrm{d}}{\mathrm{d}x}\left\{x(v_\mu'v_\nu-v_\nu'v_\mu)\right\}=\frac{\mu^2-\nu^2}{x}v_\mu v_\nu. d x d { x ( v μ ′ v ν − v ν ′ v μ ) } = x μ 2 − ν 2 v μ v ν .
Its boundary value at 1 1 1 is h μ ( s ) − h ν ( s ) h_{\mu}(s)-h_{\nu}(s) h μ ( s ) − h ν ( s ) , whereas the boundary value at 0 0 0 is 0 0 0 , since it is O ( x μ + ν ) O(x^{\mu+\nu}) O ( x μ + ν ) . The integral converges because μ + ν > 0 \mu+\nu>0 μ + ν > 0 and is strictly positive. Thus this comparison, including the endpoint ν = 0 \nu=0 ν = 0 , needs no separate variational minimization argument.
Finally, each fixed clamped eigenvalue is continuous under C 2 C^{2} C 2 diffeomorphisms of a fixed smooth domain [11 , 12 ] . This stability result is used only after the holes have been fixed; their singular shrinking limit is proved separately in § 5 .
An increased eigenvalue ratio under a spherical constraint
We establish a strict increase of the ratio by imposing a zero-value constraint on an interior sphere close to the boundary of the ball. The dimension is fixed throughout the construction. Figure 2 shows how this auxiliary problem fits into the three steps of the construction. The approximations by finite supports and smooth holes are carried out in § 4 and § 5 , respectively.
Figure 2. The three form spaces used in the construction, illustrated in dimension two. The arrows indicate the order of choosing the geometry, whereas the labels indicate the spectral approximation limits. First fix r r r to obtain a strict gap, then choose finite N N N , and finally choose the hole radius η \eta η .
Let d ≥ 2 d\ge2 d ≥ 2 , a = ( d − 2 ) / 2 a=(d-2)/2 a = ( d − 2 ) /2 , and S r = { x ∈ R d : ∣ x ∣ = r } S_r=\{x\in\mathbb{R}^{d}: |x|=r\} S r = { x ∈ R d : ∣ x ∣ = r } . For 0 < r < 1 0<r<1 0 < r < 1 use the closed form space
V r = { u ∈ H 0 2 ( B ) : Tr S r u = 0 } . (8) V_r=\{u\in H_0^2(B): \operatorname{Tr}_{S_r}u=0\}. \tag*{(8)} V r = { u ∈ H 0 2 ( B ) : Tr S r u = 0 } . (8)
Here Tr S r \operatorname{Tr}_{S_r} Tr S r denotes restriction in the Sobolev trace sense, the continuous extension of restriction of smooth functions. Only the value trace is constrained on S r S_r S r . Let its ordered eigenvalues be μ k ( r ) \mu_k(r) μ k ( r ) . The trace map is bounded, so its kernel is closed in H 0 2 ( B ) H_0^2(B) H 0 2 ( B ) . The space V r V_r V r contains C ∞ ( r B ) C^\infty(rB) C ∞ ( r B ) , where r B rB r B is the concentric ball of radius r r r , and hence has the discrete spectrum described in § 2 . Rotations preserve this kernel. The spherical harmonic decomposition therefore separates both its energy and its constraints, just as for the ball. More explicitly, the trace of a separated function f ( x ) Y l ( θ ) f(x)Y_l(\theta) f ( x ) Y l ( θ ) is f ( r ) Y l ( θ ) f(r)Y_l(\theta) f ( r ) Y l ( θ ) . Orthogonality of the angular harmonics makes the trace condition u = 0 u=0 u = 0 equivalent to value 0 for each radial coefficient. It leaves the radial derivatives at r r r unconstrained. Fix a degree l l l , put ν = l + a \nu= l + a ν = l + a , and let κ > 0 \kappa> 0 κ > 0 be a zero of W ν W_\nu W ν . The corresponding ball mode has radial factor
x − a { J ν ( κ x ) − J ν ( κ ) I ν ( κ ) I ν ( κ x ) } . x^{-a}\left\{J_\nu(\kappa x)-\frac{J_\nu(\kappa)}{I_\nu(\kappa)}I_\nu(\kappa x)\right\}. x − a { J ν ( κ x ) − I ν ( κ ) J ν ( κ ) I ν ( κ x ) } .
The denominator is nonzero because I ν I_\nu I ν is positive on the positive real axis. Set
γ = κ 4 , h ν ( κ ) = κ I ν ′ ( κ ) I ν ( κ ) . \gamma=\kappa^4,\qquad h_\nu(\kappa)=\frac{\kappa I_\nu'(\kappa)}{I_\nu(\kappa)}. γ = κ 4 , h ν ( κ ) = I ν ( κ ) κ I ν ′ ( κ ) .
We compute the supported eigenvalue through second order in the distance ϵ = 1 − r \epsilon=1-r ϵ = 1 − r to the outer boundary. The first-order coefficient will be the same for all fixed modes; the dependence on the mode first enters the second-order coefficient.
Lemma 3.1. For each fixed separated ball eigenmode, a real eigenvalue branch of (8) , for r = 1 − ϵ r=1-\epsilon r = 1 − ϵ , satisfies
λ ν , κ ( 1 − ϵ ) = γ [ 1 + 3 ϵ + 23 + h ν ( κ ) 4 ϵ 2 + O ( ϵ 3 ) ] . (9) \lambda_{\nu,\kappa}(1-\epsilon)=\gamma\left[1+3\epsilon+\frac{23+h_\nu(\kappa)}{4}\epsilon^2+O(\epsilon^3)\right]. \tag*{(9)} λ ν , κ ( 1 − ϵ ) = γ [ 1 + 3 ϵ + 4 23 + h ν ( κ ) ϵ 2 + O ( ϵ 3 ) ] . (9) The remainder is uniform on every fixed finite collection of ball modes. In particular it is uniform on the first and second ball eigenspaces, at each fixed dimension.
The calculation proceeds from the boundary conditions. We solve the radial equation inside the sphere and in the thin outer annulus, normalizing both radial derivatives at the sphere to 1. We then match the remaining quantity D l f D_lf D l f , which integration by parts identifies below. The inner solution is given by the standard ball formulas; the annular solution is controlled by an explicit inverse on the fixed interval [ 0 , 1 ] [0,1] [ 0 , 1 ] . Matching them gives an exact equation for the eigenvalue. Expanding this equation, with the remainder bounded throughout, yields (9) .
Proof. Step 1. The conditions at the interior sphere.
For 0 < δ < min { r , 1 − r } 0<\delta<\min\{r,1-r\} 0 < δ < min { r , 1 − r } , the radial coordinate belongs to I = ( r − δ , r + δ ) ⋐ ( 0 , 1 ) I=(r-\delta,r+\delta)\Subset(0,1) I = ( r − δ , r + δ ) ⋐ ( 0 , 1 ) . On this interval the weight x d − 1 x^{d-1} x d − 1 is bounded above and below. A separated H 2 H^2 H 2 function therefore has f ∈ H 2 ( I ) f\in H^2(I) f ∈ H 2 ( I ) , and Lemma 2.2 gives the common traces f ( r ) f(r) f ( r ) and f ′ ( r ) f'(r) f ′ ( r ) . Here we use the one-dimensional embedding (3) .
Write b = d − 1 b=d-1 b = d − 1 and L = l ( l + d − 2 ) L=l(l+d-2) L = l ( l + d − 2 ) . Recall the radial operator D l f = f ′ ′ + b x − 1 f ′ − L x − 2 f D_lf=f''+bx^{-1}f'-Lx^{-2}f D l f = f ′′ + b x − 1 f ′ − L x − 2 f . The radial energy is ∫ ( D l f ) ( D l g ) x b d x \int(D_lf)(D_lg)x^b\,dx ∫ ( D l f ) ( D l g ) x b d x , after choosing an angular harmonic of unit L 2 L^2 L 2 norm. To keep track of the interface sign, use D l g = x − b ( x b g ′ ) ′ − L x − 2 g D_lg=x^{-b}(x^bg')'-Lx^{-2}g D l g = x − b ( x b g ′ ) ′ − L x − 2 g . On an interval ( α , β ) (\alpha,\beta) ( α , β ) where f f f is smooth, put q = D l f q=D_lf q = D l f . The two integrations by parts are
∫ α β q D l g x b d x = ∫ α β q ( x b g ′ ) ′ d x − L ∫ α β q g x b − 2 d x = [ x b q g ′ ] α β − ∫ α β q ′ g ′ x b d x − L ∫ α β q g x b − 2 d x = [ x b ( q g ′ − q ′ g ) ] α β + ∫ α β D l q g x b d x . \begin{aligned}
\int_\alpha^\beta qD_lg\,x^b\,dx
=\int_\alpha^\beta q(x^bg')'\,dx-L\int_\alpha^\beta qgx^{b-2}\,dx \\
&=\left[x^bqg'\right]_\alpha^\beta-\int_\alpha^\beta q'g'x^b\,dx-L\int_\alpha^\beta qgx^{b-2}\,dx \\
&=\left[x^b(qg'-q'g)\right]_\alpha^\beta+\int_\alpha^\beta D_lq\,gx^b\,dx.
\end{aligned} ∫ α β q D l g x b d x = ∫ α β q ( x b g ′ ) ′ d x − L ∫ α β q g x b − 2 d x = [ x b q g ′ ] α β − ∫ α β q ′ g ′ x b d x − L ∫ α β q g x b − 2 d x = [ x b ( q g ′ − q ′ g ) ] α β + ∫ α β D l q g x b d x .
Here g ( r ) = 0 g(r)=0 g ( r ) = 0 and its common slope g ′ ( r ) g'(r) g ′ ( r ) is arbitrary. The upper endpoint of the inner interval and the lower endpoint of the outer interval have opposite signs. Their sum is
r b { D l f ( r − ) − D l f ( r + ) } g ′ ( r ) . r^{b}\{D_lf(r-) - D_lf(r+)\}g'(r). r b { D l f ( r − ) − D l f ( r + )} g ′ ( r ) .
The common slope can be prescribed arbitrarily by taking a test function supported in a small annulus about r r r and equal to a multiple of x − r x-r x − r near the interface. Thus its coefficient must vanish. In addition to the clamped data at 1 1 1 and regularity at 0 0 0 , we must therefore arrange
f ( r − ) = f ( r + ) = 0 , f ′ ( r − ) = f ′ ( r + ) , f(r-) = f(r+) = 0,\qquad f'(r-) = f'(r+), f ( r − ) = f ( r + ) = 0 , f ′ ( r − ) = f ′ ( r + ) ,
D l f ( r − ) = D l f ( r + ) . D_lf(r-) = D_lf(r+). D l f ( r − ) = D l f ( r + ) .
The first line provides the value and gradient continuity needed for global H 2 H^2 H 2 regularity; the second line matches the moments with the same sign. The coefficient of g ( r ) g(r) g ( r ) imposes no condition on ( D l f ) ′ (D_lf)' ( D l f ) ′ , since the test value already vanishes. At the end of the proof we verify that these conditions also suffice for the full weak eigenvalue equation.
Step 2. The inner value of D l f D_lf D l f .
A regular solution with interface value 0 0 0 is
F ( x ) = x − a { J ν ( k x ) I ν ( k r ) − I ν ( k x ) J ν ( k r ) } , F ′ ( r ) = − k r − a W ν ( k r ) , k 4 = λ . F(x) = x^{-a}\{J_\nu(kx)I_\nu(kr)-I_\nu(kx)J_\nu(kr)\},\qquad F'(r)=-kr^{-a}W_\nu(kr),\qquad k^4=\lambda. F ( x ) = x − a { J ν ( k x ) I ν ( k r ) − I ν ( k x ) J ν ( k r )} , F ′ ( r ) = − k r − a W ν ( k r ) , k 4 = λ .
To check the slope formula, differentiating x − a x^{-a} x − a contributes 0 0 0 at x = r x=r x = r , because the expression in braces vanishes there. The remaining derivative is
k r − a { J ν ′ ( k r ) I ν ( k r ) − I ν ′ ( k r ) J ν ( k r ) } = − k r − a W ν ( k r ) . kr^{-a}\{J_\nu'(kr)I_\nu(kr)-I_\nu'(kr)J_\nu(kr)\}=-kr^{-a}W_\nu(kr). k r − a { J ν ′ ( k r ) I ν ( k r ) − I ν ′ ( k r ) J ν ( k r )} = − k r − a W ν ( k r ) .
Whenever this is nonzero, the normalized inner solution is f i n ( x ) = F ( x ) / F ′ ( r ) f_{\mathrm{in}}(x)=F(x)/F'(r) f in ( x ) = F ( x ) / F ′ ( r ) . It has value 0 0 0 and slope 1 1 1 at r r r . The Bessel equations give
D l ( x − a J ν ( k x ) ) = − k 2 x − a J ν ( k x ) , D_l\left(x^{-a}J_\nu(kx)\right)=-k^2x^{-a}J_\nu(kx), D l ( x − a J ν ( k x ) ) = − k 2 x − a J ν ( k x ) ,
D l ( x − a I ν ( k x ) ) = k 2 x − a I ν ( k x ) . D_l\left(x^{-a}I_\nu(kx)\right)=k^2x^{-a}I_\nu(kx). D l ( x − a I ν ( k x ) ) = k 2 x − a I ν ( k x ) .
Consequently D l F ( r ) = − 2 k 2 r − a J ν ( k r ) I ν ( k r ) D_lF(r)=-2k^2r^{-a}J_\nu(kr)I_\nu(kr) D l F ( r ) = − 2 k 2 r − a J ν ( k r ) I ν ( k r ) . The quotient determining the inner moment is therefore
D l F ( r ) F ′ ( r ) = 2 k J ν ( k r ) I ν ( k r ) W ν ( k r ) = k W ν ′ ( k r ) W ν ( k r ) + 1 r . \frac{D_lF(r)}{F'(r)}=\frac{2kJ_\nu(kr)I_\nu(kr)}{W_\nu(kr)}=k\frac{W_\nu'(kr)}{W_\nu(kr)}+\frac{1}{r}. F ′ ( r ) D l F ( r ) = W ν ( k r ) 2 k J ν ( k r ) I ν ( k r ) = k W ν ( k r ) W ν ′ ( k r ) + r 1 .
Here we used 2 J ν I ν = W ν ′ + W ν / x 2J_\nu I_\nu=W_\nu'+W_\nu/x 2 J ν I ν = W ν ′ + W ν / x at x = k r x=kr x = k r . Equivalently,
D l f ( r ) = T ν ( k r ) + 1 r , T ν ( x ) = x W ν ′ ( x ) W ν ( x ) . (10) D_lf(r)=\frac{T_\nu(kr)+1}{r},\qquad T_\nu(x)=\frac{xW_\nu'(x)}{W_\nu(x)}. \tag*{(10)} D l f ( r ) = r T ν ( k r ) + 1 , T ν ( x ) = W ν ( x ) x W ν ′ ( x ) . (10)
More explicitly, F ′ ′ ( r ) / F ′ ( r ) = { T ν ( k r ) − 2 a } / r F''(r)/F'(r)=\{T_\nu(kr)-2a\}/r F ′′ ( r ) / F ′ ( r ) = { T ν ( k r ) − 2 a } / r , and b − 2 a = 1 b-2a=1 b − 2 a = 1 . The solution found below has ( k r ) 4 < γ (kr)^4<\gamma ( k r ) 4 < γ near the simple zero γ \gamma γ , so this normalization is legitimate.
Step 3. The annulus inverse and its remainder.
Impose f ( r ) = 0 f(r)=0 f ( r ) = 0 , f ′ ( r ) = 1 f'(r)=1 f ′ ( r ) = 1 , f ( 1 ) = f ′ ( 1 ) = 0 f(1)=f'(1)=0 f ( 1 ) = f ′ ( 1 ) = 0 on the outer annulus. To compute its fourth-order equation, first differentiate q = D l f q=D_lf q = D l f :
q ′ = f ′ ′ ′ + b x f ′ ′ − b + L x 2 f ′ + 2 L x 3 f , q' = f''' + \frac{b}{x}f'' - \frac{b+L}{x^2}f' + \frac{2L}{x^3}f, q ′ = f ′′′ + x b f ′′ − x 2 b + L f ′ + x 3 2 L f ,
q ′ ′ = f ′ ′ ′ ′ + b x f ′ ′ ′ − 2 b + L x 2 f ′ ′ + 2 b + 4 L x 3 f ′ − 6 L x 4 f . q'' = f'''' + \frac{b}{x}f''' - \frac{2b+L}{x^2}f'' + \frac{2b+4L}{x^3}f' - \frac{6L}{x^4}f. q ′′ = f ′′′′ + x b f ′′′ − x 2 2 b + L f ′′ + x 3 2 b + 4 L f ′ − x 4 6 L f .
In D l q = q ′ ′ + b x − 1 q ′ − L x − 2 q D_lq=q''+bx^{-1}q'-Lx^{-2}q D l q = q ′′ + b x − 1 q ′ − L x − 2 q , the coefficients of f ′ ′ f'' f ′′ , f ′ f' f ′ and f f f are obtained respectively from
− ( 2 b + L ) + b 2 − L = b ( b − 2 ) − 2 L , -(2b+L)+b^2-L=b(b-2)-2L, − ( 2 b + L ) + b 2 − L = b ( b − 2 ) − 2 L ,
( 2 b + 4 L ) − b ( b + L ) − b L = ( 2 − b ) ( b + 2 L ) , (2b+4L)-b(b+L)-bL=(2-b)(b+2L), ( 2 b + 4 L ) − b ( b + L ) − b L = ( 2 − b ) ( b + 2 L ) ,
− 6 L + 2 b L + L 2 = L ( L + 2 b − 6 ) . -6L+2bL+L^2=L(L+2b-6). − 6 L + 2 b L + L 2 = L ( L + 2 b − 6 ) .
Thus
D l 2 f = f ′ ′ ′ ′ + 2 b x f ′ ′ ′ + b ( b − 2 ) − 2 L x 2 f ′ ′ + ( 2 − b ) ( b + 2 L ) x 3 f ′ + L ( L + 2 b − 6 ) x 4 f . D_l^2f=f''''+\frac{2b}{x}f'''+\frac{b(b-2)-2L}{x^2}f''+\frac{(2-b)(b+2L)}{x^3}f'+\frac{L(L+2b-6)}{x^4}f. D l 2 f = f ′′′′ + x 2 b f ′′′ + x 2 b ( b − 2 ) − 2 L f ′′ + x 3 ( 2 − b ) ( b + 2 L ) f ′ + x 4 L ( L + 2 b − 6 ) f .
For x ∈ [ r , 1 ] x\in[r,1] x ∈ [ r , 1 ] , put z = ( x − r ) / ϵ ∈ [ 0 , 1 ] z=(x-r)/\epsilon\in[0,1] z = ( x − r ) / ϵ ∈ [ 0 , 1 ] and f ( x ) = ϵ Y ( z ) f(x)=\epsilon Y(z) f ( x ) = ϵ Y ( z ) . The chain rule gives f ( j ) = ϵ 1 − j Y ( j ) f^{(j)}=\epsilon^{1-j}Y^{(j)} f ( j ) = ϵ 1 − j Y ( j ) for 0 ≤ j ≤ 4 0\le j\le4 0 ≤ j ≤ 4 . Multiplying D l 2 f − λ f = 0 D_l^2f-\lambda f=0 D l 2 f − λ f = 0 by ϵ 3 \epsilon^3 ϵ 3 gives the exact equation
L ϵ , λ Y = Y ′ ′ ′ ′ + 2 b ϵ r + ϵ z Y ′ ′ ′ + { b ( b − 2 ) − 2 L } ϵ 2 ( r + ϵ z ) 2 Y ′ ′ + ( 2 − b ) ( b + 2 L ) ϵ 3 ( r + ϵ z ) 3 Y ′ + ϵ 4 { L ( L + 2 b − 6 ) ( r + ϵ z ) 4 − λ } Y = 0. (11) \mathcal{L}_{\epsilon,\lambda}Y=Y''''+\frac{2b\epsilon}{r+\epsilon z}Y'''+\frac{\{b(b-2)-2L\}\epsilon^2}{(r+\epsilon z)^2}Y''+\frac{(2-b)(b+2L)\epsilon^3}{(r+\epsilon z)^3}Y'+\epsilon^4\left\{\frac{L(L+2b-6)}{(r+\epsilon z)^4}-\lambda\right\}Y=0. \tag*{(11)} L ϵ , λ Y = Y ′′′′ + r + ϵz 2 b ϵ Y ′′′ + ( r + ϵz ) 2 { b ( b − 2 ) − 2 L } ϵ 2 Y ′′ + ( r + ϵz ) 3 ( 2 − b ) ( b + 2 L ) ϵ 3 Y ′ + ϵ 4 { ( r + ϵz ) 4 L ( L + 2 b − 6 ) − λ } Y = 0. (11)
In particular, the eigenvalue parameter first enters the annulus equation at order ϵ 4 \epsilon^4 ϵ 4 . Because f ′ = Y ′ f'=Y' f ′ = Y ′ , the four transformed data are Y ( 0 ) = Y ( 1 ) = Y ′ ( 1 ) = 0 Y(0)=Y(1)=Y'(1)=0 Y ( 0 ) = Y ( 1 ) = Y ′ ( 1 ) = 0 , Y ′ ( 0 ) = 1 Y'(0)=1 Y ′ ( 0 ) = 1 . Fix 0 < α < 1 0<\alpha<1 0 < α < 1 . Here C k , α C^{k,\alpha} C k , α consists of functions with k k k continuous derivatives whose k k k th derivative is Hölder continuous with exponent α \alpha α . Let X X X be C 4 , α [ 0 , 1 ] C^{4,\alpha}[0,1] C 4 , α [ 0 , 1 ] with all four homogeneous clamped data. The inverse B \mathcal{B} B of D 4 D^4 D 4 on X X X can be written explicitly. For g ∈ C 0 , α g\in C^{0,\alpha} g ∈ C 0 , α define
H j = ∫ 0 1 ( 1 − t ) j j ! g ( t ) d t ( j = 2 , 3 ) , H_j=\int_0^1\frac{(1-t)^j}{j!}g(t)\,dt\qquad(j=2,3), H j = ∫ 0 1 j ! ( 1 − t ) j g ( t ) d t ( j = 2 , 3 ) ,
( B g ) ( z ) = ∫ 0 z ( z − t ) 3 6 g ( t ) d t + ( H 2 − 3 H 3 ) z 2 + ( 2 H 3 − H 2 ) z 3 . (12) (\mathcal{B}g)(z)=\int_0^z\frac{(z-t)^3}{6}g(t)\,dt+(H_2-3H_3)z^2+(2H_3-H_2)z^3. \tag*{(12)} ( B g ) ( z ) = ∫ 0 z 6 ( z − t ) 3 g ( t ) d t + ( H 2 − 3 H 3 ) z 2 + ( 2 H 3 − H 2 ) z 3 . (12)
Here is the boundary-value calculation for (3.5). Integrating w ′ ′ ′ ′ = g w''''=g w ′′′′ = g and imposing w ( 0 ) = w ′ ( 0 ) = 0 w(0)=w'(0)=0 w ( 0 ) = w ′ ( 0 ) = 0 gives
w ( z ) = ∫ 0 z ( z − t ) 3 6 g ( t ) d t + A z 2 + B z 3 , w(z)=\int_{0}^{z}\frac{(z-t)^3}{6}g(t)\,\mathrm{d}t+Az^2+Bz^3, w ( z ) = ∫ 0 z 6 ( z − t ) 3 g ( t ) d t + A z 2 + B z 3 ,
w ′ ( z ) = ∫ 0 z ( z − t ) 2 2 g ( t ) d t + 2 A z + 3 B z 2 , w'(z)=\int_{0}^{z}\frac{(z-t)^2}{2}g(t)\,\mathrm{d}t+2Az+3Bz^2, w ′ ( z ) = ∫ 0 z 2 ( z − t ) 2 g ( t ) d t + 2 A z + 3 B z 2 ,
w ′ ′ ( z ) = ∫ 0 z ( z − t ) g ( t ) d t + 2 A + 6 B z , w''(z)=\int_{0}^{z}(z-t)g(t)\,\mathrm{d}t+2A+6Bz, w ′′ ( z ) = ∫ 0 z ( z − t ) g ( t ) d t + 2 A + 6 B z ,
w ′ ′ ′ ( z ) = ∫ 0 z g ( t ) d t + 6 B , w ′ ′ ′ ′ ( z ) = g ( z ) . w'''(z)=\int_{0}^{z}g(t)\,\mathrm{d}t+6B,\qquad w''''(z)=g(z). w ′′′ ( z ) = ∫ 0 z g ( t ) d t + 6 B , w ′′′′ ( z ) = g ( z ) .
Consequently the remaining two boundary conditions are equivalent to
( 1 1 2 3 ) ( A B ) = − ( H 3 H 2 ) , ( A B ) = ( 3 − 1 − 2 1 ) ( − H 3 − H 2 ) = ( H 2 − 3 H 3 2 H 3 − H 2 ) . \begin{pmatrix}1&1\\2&3\end{pmatrix}
\begin{pmatrix}A\\B\end{pmatrix}
=-\begin{pmatrix}H_3\\H_2\end{pmatrix},
\qquad
\begin{pmatrix}A\\B\end{pmatrix}
=\begin{pmatrix}3&-1\\-2&1\end{pmatrix}
\begin{pmatrix}-H_3\\-H_2\end{pmatrix}
=\begin{pmatrix}H_2-3H_3\\2H_3-H_2\end{pmatrix}. ( 1 2 1 3 ) ( A B ) = − ( H 3 H 2 ) , ( A B ) = ( 3 − 2 − 1 1 ) ( − H 3 − H 2 ) = ( H 2 − 3 H 3 2 H 3 − H 2 ) .
The determinant of this matrix is 1. Thus the solution is unique, and substitution gives (12) , including all four clamped boundary data. With the norm given by the sum of derivative supremum norms through order four and the order-four Hölder seminorm, we obtain
∥ B ∥ C 0 , α → C 4 , α ≤ 12. \|\mathcal{B}\|_{C^{0,\alpha}\to C^{4,\alpha}}\le12. ∥ B ∥ C 0 , α → C 4 , α ≤ 12.
Indeed ∣ H 2 ∣ ≤ ∥ g ∥ ∞ / 6 |H_2|\le\|g\|_\infty/6 ∣ H 2 ∣ ≤ ∥ g ∥ ∞ /6 and ∣ H 3 ∣ ≤ ∥ g ∥ ∞ / 24 |H_3|\le\|g\|_\infty/24 ∣ H 3 ∣ ≤ ∥ g ∥ ∞ /24 , whence ∣ A ∣ ≤ 7 ∥ g ∥ ∞ / 24 |A|\le7\|g\|_\infty/24 ∣ A ∣ ≤ 7∥ g ∥ ∞ /24 and ∣ B ∣ ≤ ∥ g ∥ ∞ / 4 |B|\le\|g\|_\infty/4 ∣ B ∣ ≤ ∥ g ∥ ∞ /4 . For example, the integral parts of w w w , w ′ w' w ′ , w ′ ′ w'' w ′′ and w ′ ′ ′ w''' w ′′′ are bounded by ∥ g ∥ ∞ / 24 \|g\|_\infty/24 ∥ g ∥ ∞ /24 , ∥ g ∥ ∞ / 6 \|g\|_\infty/6 ∥ g ∥ ∞ /6 , ∥ g ∥ ∞ / 2 \|g\|_\infty/2 ∥ g ∥ ∞ /2 and ∥ g ∥ ∞ \|g\|_\infty ∥ g ∥ ∞ , respectively. Adding the differentiated polynomial terms in (3.5) bounds the five supremum norms, in increasing derivative order, by the entries of
( 7 12 , 3 2 , 31 12 , 5 2 , 1 ) ∥ g ∥ ∞ . \left(\frac{7}{12},\frac{3}{2},\frac{31}{12},\frac{5}{2},1\right)\|g\|_\infty. ( 12 7 , 2 3 , 12 31 , 2 5 , 1 ) ∥ g ∥ ∞ .
The fourth-derivative Hölder seminorm is exactly [ g ] α [g]_\alpha [ g ] α . The sum of all these bounds is at most
12 ( ∥ g ∥ ∞ + [ g ] α ) . 12\left(\|g\|_\infty+[g]_\alpha\right). 12 ( ∥ g ∥ ∞ + [ g ] α ) .
Set P = L − D 4 \mathcal{P}=\mathcal{L}-D^4 P = L − D 4 . We use the rescaled operator also at ϵ = 0 \epsilon=0 ϵ = 0 , where it equals D 4 D^4 D 4 . Since r + ϵ z = 1 − ϵ ( 1 − z ) r+\epsilon z=1-\epsilon(1-z) r + ϵz = 1 − ϵ ( 1 − z ) , for ∣ ϵ ∣ < 1 / 4 |\epsilon|<1/4 ∣ ϵ ∣ < 1/4 and 0 ≤ z ≤ 1 0\le z\le1 0 ≤ z ≤ 1 this denominator is at least 3 / 4 3/4 3/4 . Each reciprocal power has a bounded C 0 , α C^{0,\alpha} C 0 , α norm, uniformly in ϵ \epsilon ϵ , as follows by differentiating it once with respect to z z z . For a coefficient function c c c , the product estimate
∥ c v ∥ C 0 , α ≤ ∥ c ∥ ∞ ∥ v ∥ ∞ + ∥ c ∥ ∞ [ v ] α + [ c ] α ∥ v ∥ ∞ \|cv\|_{C^{0,\alpha}}\le\|c\|_\infty\|v\|_\infty+\|c\|_\infty[v]_\alpha+[c]_\alpha\|v\|_\infty ∥ c v ∥ C 0 , α ≤ ∥ c ∥ ∞ ∥ v ∥ ∞ + ∥ c ∥ ∞ [ v ] α + [ c ] α ∥ v ∥ ∞
then controls each term of P \mathcal{P} P . For fixed b b b , L L L and λ \lambda λ in a fixed bounded neighborhood of γ \gamma γ , it gives
∥ P ∥ X → C 0 , α ≤ C { ∣ ϵ ∣ + ∣ ϵ ∣ 2 + ∣ ϵ ∣ 3 + ∣ ϵ ∣ 4 ( 1 + ∣ λ ∣ ) } ≤ C ′ ∣ ϵ ∣ . \|\mathcal{P}\|_{X\to C^{0,\alpha}}
\le C\left\{|\epsilon|+|\epsilon|^2+|\epsilon|^3+|\epsilon|^4(1+|\lambda|)\right\}
\le C'|\epsilon|. ∥ P ∥ X → C 0 , α ≤ C { ∣ ϵ ∣ + ∣ ϵ ∣ 2 + ∣ ϵ ∣ 3 + ∣ ϵ ∣ 4 ( 1 + ∣ λ ∣ ) } ≤ C ′ ∣ ϵ ∣.
Here derivatives of order at most three have C 0 , α C^{0,\alpha} C 0 , α norm controlled by the C 4 , α C^{4,\alpha} C 4 , α norm. Renaming the last constant as C C C , we obtain the bound used below. The uniformly convergent power series of { 1 − ϵ ( 1 − z ) } − j \{1-\epsilon(1-z)\}^{-j} { 1 − ϵ ( 1 − z ) } − j show that these coefficients are analytic in ϵ \epsilon ϵ with values in C 0 , α C^{0,\alpha} C 0 , α ; their dependence on λ \lambda λ is polynomial. For 12 C ∣ ϵ ∣ < 1 / 2 12C|\epsilon|<1/2 12 C ∣ ϵ ∣ < 1/2 the convergent operator series
L − 1 = ∑ j = 0 + ∞ ( − B P ) j B \mathcal{L}^{-1}=\sum_{j=0}^{+\infty}(-\mathcal{B}P)^j\mathcal{B} L − 1 = j = 0 ∑ + ∞ ( − B P ) j B
has norm at most 12 ∑ j = 0 + ∞ ( 12 C ∣ ϵ ∣ ) j ≤ 24 12\sum_{j=0}^{+\infty}(12C|\epsilon|)^j\le24 12 ∑ j = 0 + ∞ ( 12 C ∣ ϵ ∣ ) j ≤ 24 . The convergence is uniform on a smaller parameter neighborhood, so the inverse is analytic jointly in ϵ \epsilon ϵ and λ \lambda λ . Here L = D 4 ( I + B P ) \mathcal{L}=D^4(I+\mathcal{B}P) L = D 4 ( I + B P ) on X X X , which explains the order of the operators in the series. The solution with the prescribed nonzero slope is obtained by writing Y = Y 0 + w Y=Y_0+w Y = Y 0 + w , where Y 0 = z ( 1 − z ) 2 Y_0=z(1-z)^2 Y 0 = z ( 1 − z ) 2 and w ∈ X w\in X w ∈ X , and solving w = − L − 1 L Y 0 w=-\mathcal{L}^{-1}\mathcal{L}Y_0 w = − L − 1 L Y 0 . This is legitimate because Y 0 Y_0 Y 0 already has all four prescribed data, so w w w must have homogeneous data. Conversely, the displayed solution belongs to X X X and makes L ( Y 0 + w ) = 0 \mathcal{L}(Y_0+w)=0 L ( Y 0 + w ) = 0 , proving existence and uniqueness of the annulus solution. Since Y 0 ′ ′ ′ = 0 Y_0'''=0 Y 0 ′′′ = 0 and P Y 0 = O ( ϵ ) PY_0=O(\epsilon) P Y 0 = O ( ϵ ) in C 0 , α C^{0,\alpha} C 0 , α , it also gives ∥ Y − Y 0 ∥ C 4 , α = O ( ϵ ) \|Y-Y_0\|_{C^{4,\alpha}}=O(\epsilon) ∥ Y − Y 0 ∥ C 4 , α = O ( ϵ ) . The leading profile is shown in Figure 3 .
Figure 3. The leading rescaled annulus profile, with z = ( x − r ) / ϵ z=(x-r)/\epsilon z = ( x − r ) / ϵ and ϵ = 1 − r \epsilon=1-r ϵ = 1 − r . The slope at the support is normalized. Its curvature Y 0 ′ ′ ( 0 ) = − 4 Y_0''(0)=-4 Y 0 ′′ ( 0 ) = − 4 produces the leading outer moment − 4 / ϵ -4/\epsilon − 4/ ϵ .
To determine the next term, put Y 1 = − b z 2 ( 1 − z ) 2 / 2 Y_1=-bz^2(1-z)^2/2 Y 1 = − b z 2 ( 1 − z ) 2 /2 . Its factors give the four homogeneous boundary data, and direct differentiation gives
Y 1 ′ ′ ′ = 6 , Y 1 ′ ′ = − b + 6 b z − 6 b z 2 , Y 1 ′ ′ ′ = 6 b − 12 b z , Y 1 ′ ′ ′ ′ = − 12 b . Y_1'''=6,\qquad Y_1''=-b+6bz-6bz^2,\qquad Y_1'''=6b-12bz,\qquad Y_1''''=-12b. Y 1 ′′′ = 6 , Y 1 ′′ = − b + 6 b z − 6 b z 2 , Y 1 ′′′ = 6 b − 12 b z , Y 1 ′′′′ = − 12 b .
Set Y a p p = Y 0 + ( ϵ / r ) Y 1 Y_{\mathrm{app}}=Y_0+(\epsilon/r)Y_1 Y app = Y 0 + ( ϵ / r ) Y 1 . In the residual calculation, use the same radial variable x = r + ϵ z x=r+\epsilon z x = r + ϵz as in (11) . The terms of order 1 1 1 in ϵ \epsilon ϵ cancel as follows:
ϵ r Y 1 ′ ′ ′ ′ + 2 b ϵ x Y 0 ′ ′ ′ = 12 b ϵ ( 1 x − 1 r ) = − 12 b ϵ 2 z r x . \frac{\epsilon}{r}Y_1''''+\frac{2b\epsilon}{x}Y_0'''=12b\epsilon\left(\frac{1}{x}-\frac{1}{r}\right)=-\frac{12b\epsilon^2z}{rx}. r ϵ Y 1 ′′′′ + x 2 b ϵ Y 0 ′′′ = 12 b ϵ ( x 1 − r 1 ) = − r x 12 b ϵ 2 z .
The remaining terms of (11) all contain at least two factors of ϵ \epsilon ϵ . More precisely, L Y a p p = \mathcal{L}Y_{\mathrm{app}}= L Y app = ϵ 2 Q \epsilon^{2}Q ϵ 2 Q , where
Q = − 12 b z r x + 2 b r x Y 1 ′ ′ ′ + [ b ( b − 2 ) − 2 L ] x 2 ( Y 0 ′ ′ + ϵ r Y 1 ′ ′ ) + ( 2 − b ) ( b + 2 L ) ϵ x 3 ( Y 0 ′ + ϵ r Y 1 ′ ) + ϵ 2 ( L ( L + 2 b − 6 ) x 4 − λ ) ( Y 0 + ϵ r Y 1 ) . Q=-\frac{12bz}{rx}+\frac{2b}{rx}Y_{1}^{\prime\prime\prime}+\frac{[b(b-2)-2L]}{x^{2}}\left(Y_{0}^{\prime\prime}+\frac{\epsilon}{r}Y_{1}^{\prime\prime}\right)
+\frac{(2-b)(b+2L)\epsilon}{x^{3}}\left(Y_{0}^{\prime}+\frac{\epsilon}{r}Y_{1}^{\prime}\right)
+\epsilon^{2}\left(\frac{L(L+2b-6)}{x^{4}}-\lambda\right)\left(Y_{0}+\frac{\epsilon}{r}Y_{1}\right). Q = − r x 12 b z + r x 2 b Y 1 ′′′ + x 2 [ b ( b − 2 ) − 2 L ] ( Y 0 ′′ + r ϵ Y 1 ′′ ) + x 3 ( 2 − b ) ( b + 2 L ) ϵ ( Y 0 ′ + r ϵ Y 1 ′ ) + ϵ 2 ( x 4 L ( L + 2 b − 6 ) − λ ) ( Y 0 + r ϵ Y 1 ) .
This expression is analytic in ϵ \epsilon ϵ , λ \lambda λ and bounded in C 0 , α C^{0,\alpha} C 0 , α on a fixed neighborhood of ( 0 , γ ) (0,\gamma) ( 0 , γ ) . Both Y Y Y and Y a p p Y_{\mathrm{app}} Y app have the same boundary data. Their difference is therefore in X X X and satisfies L ( Y − Y a p p ) = − ϵ 2 Q \mathcal{L}(Y-Y_{\mathrm{app}})=-\epsilon^{2}Q L ( Y − Y app ) = − ϵ 2 Q . Applying the inverse just constructed proves the exact identity
Y = Y 0 + ϵ r Y 1 + ϵ 2 R Y ( ϵ , λ ) , R Y = − L − 1 Q . Y=Y_{0}+\frac{\epsilon}{r}Y_{1}+\epsilon^{2}R_{Y}(\epsilon,\lambda),\qquad R_{Y}=-\mathcal{L}^{-1}Q. Y = Y 0 + r ϵ Y 1 + ϵ 2 R Y ( ϵ , λ ) , R Y = − L − 1 Q .
In particular, R Y R_{Y} R Y is analytic and ∥ R Y ∥ C 4 , α ≤ 24 ∥ Q ∥ C 0 , α \|R_{Y}\|_{C^{4,\alpha}}\leq24\|Q\|_{C^{0,\alpha}} ∥ R Y ∥ C 4 , α ≤ 24∥ Q ∥ C 0 , α . Evaluation of its second derivative at 0 0 0 is a bounded linear functional on C 4 , α C^{4,\alpha} C 4 , α . Since Y 0 ′ ′ ( 0 ) = − 4 Y_{0}^{\prime\prime}(0)=-4 Y 0 ′′ ( 0 ) = − 4 and Y 1 ′ ′ ( 0 ) = − b Y_{1}^{\prime\prime}(0)=-b Y 1 ′′ ( 0 ) = − b , we obtain
f ′ ′ ( r ) = ϵ − 1 Y ′ ′ ( 0 ) = − 4 ϵ − b r + ϵ ∂ z 2 R Y ( ϵ , λ ) ( 0 ) . f^{\prime\prime}(r)=\epsilon^{-1}Y^{\prime\prime}(0)=-\frac{4}{\epsilon}-\frac{b}{r}+\epsilon\partial_{z}^{2}R_{Y}(\epsilon,\lambda)(0). f ′′ ( r ) = ϵ − 1 Y ′′ ( 0 ) = − ϵ 4 − r b + ϵ ∂ z 2 R Y ( ϵ , λ ) ( 0 ) .
At the interface f ( r ) = 0 f(r)=0 f ( r ) = 0 and f ′ ( r ) = 1 f^{\prime}(r)=1 f ′ ( r ) = 1 . Thus the term − L r − 2 f ( r ) -Lr^{-2}f(r) − L r − 2 f ( r ) vanishes, and b r − 1 f ′ ( r ) = b / r br^{-1}f^{\prime}(r)=b/r b r − 1 f ′ ( r ) = b / r cancels the constant term above. The outer moment consequently has the exact form
D l f ( r ) = − 4 ϵ + ϵ R o ( ϵ , λ ) , (13) D_{l}f(r)=-\frac{4}{\epsilon}+\epsilon R_{o}(\epsilon,\lambda), \tag*{(13)} D l f ( r ) = − ϵ 4 + ϵ R o ( ϵ , λ ) , (13)
where R o ( ϵ , λ ) = ∂ z 2 R Y ( ϵ , λ ) ( 0 ) R_{o}(\epsilon,\lambda)=\partial_{z}^{2}R_{Y}(\epsilon,\lambda)(0) R o ( ϵ , λ ) = ∂ z 2 R Y ( ϵ , λ ) ( 0 ) is analytic and bounded. The absence of a constant term in this moment is what permits the second-order eigenvalue coefficient to be read from the inner Laurent expansion.
Step 4. The constant term after subtracting the pole.
At x = κ x=\kappa x = κ , the identities for W ν W_{\nu} W ν give
W ν ′ ′ ( κ ) W ν ′ ( κ ) = 2 I ν ′ ( κ ) I ν ( κ ) − 1 κ . \frac{W_{\nu}^{\prime\prime}(\kappa)}{W_{\nu}^{\prime}(\kappa)}=2\frac{I_{\nu}^{\prime}(\kappa)}{I_{\nu}(\kappa)}-\frac{1}{\kappa}. W ν ′ ( κ ) W ν ′′ ( κ ) = 2 I ν ( κ ) I ν ′ ( κ ) − κ 1 .
Indeed differentiation of W ν ′ = 2 J ν I ν − W ν / x W_{\nu}^{\prime}=2J_{\nu}I_{\nu}-W_{\nu}/x W ν ′ = 2 J ν I ν − W ν / x gives
W ν ′ ′ = 2 J ν ′ I ν + 2 J ν I ν ′ − W ν ′ x + W ν x 2 . W_{\nu}^{\prime\prime}=2J_{\nu}^{\prime}I_{\nu}+2J_{\nu}I_{\nu}^{\prime}-\frac{W_{\nu}^{\prime}}{x}+\frac{W_{\nu}}{x^{2}}. W ν ′′ = 2 J ν ′ I ν + 2 J ν I ν ′ − x W ν ′ + x 2 W ν .
At κ \kappa κ , one has W ν = 0 W_{\nu}=0 W ν = 0 , W ν ′ = 2 J ν I ν ≠ 0 W_{\nu}^{\prime}=2J_{\nu}I_{\nu}\ne0 W ν ′ = 2 J ν I ν = 0 and J ν ′ / J ν = I ν ′ / I ν J_{\nu}^{\prime}/J_{\nu}=I_{\nu}^{\prime}/I_{\nu} J ν ′ / J ν = I ν ′ / I ν . Division by 2 J ν I ν 2J_{\nu}I_{\nu} 2 J ν I ν gives the stated identity. For δ = x − κ \delta=x-\kappa δ = x − κ , write A = W ν ′ ( κ ) ≠ 0 A=W_{\nu}^{\prime}(\kappa)\ne0 A = W ν ′ ( κ ) = 0 and B = W ν ′ ′ ( κ ) B=W_{\nu}^{\prime\prime}(\kappa) B = W ν ′′ ( κ ) . In the quotient below, factor A δ A\delta A δ from the denominator and use
( 1 + B 2 A δ + O ( δ 2 ) ) − 1 = 1 − B 2 A δ + O ( δ 2 ) . \left(1+\frac{B}{2A}\delta+O(\delta^{2})\right)^{-1}=1-\frac{B}{2A}\delta+O(\delta^{2}). ( 1 + 2 A B δ + O ( δ 2 ) ) − 1 = 1 − 2 A B δ + O ( δ 2 ) .
Multiplying this reciprocal by 1 + ( B / A ) δ + O ( δ 2 ) 1 + (B/A)\,\delta+ O(\delta^2) 1 + ( B / A ) δ + O ( δ 2 ) produces 1 + ( B / ( 2 A ) ) δ + O ( δ 2 ) 1 + (B/(2A))\,\delta+ O(\delta^2) 1 + ( B / ( 2 A )) δ + O ( δ 2 ) . Thus Taylor division gives
W ν ( x ) = A δ + 1 2 B δ 2 + O ( δ 3 ) , W_\nu(x) = A\delta+ \tfrac{1}{2}B\delta^2 + O(\delta^3), W ν ( x ) = A δ + 2 1 B δ 2 + O ( δ 3 ) ,
W ν ′ ( x ) W ν ( x ) = 1 δ + B 2 A + O ( δ ) , \frac{W_\nu'(x)}{W_\nu(x)} = \frac{1}{\delta} + \frac{B}{2A} + O(\delta), W ν ( x ) W ν ′ ( x ) = δ 1 + 2 A B + O ( δ ) ,
T ν ( x ) = κ δ + 1 + κ B 2 A + O ( δ ) = κ δ + 1 2 + h ν ( κ ) + O ( δ ) . T_\nu(x) = \frac{\kappa}{\delta} + 1 + \frac{\kappa B}{2A} + O(\delta) = \frac{\kappa}{\delta} + \frac{1}{2} + h_\nu(\kappa) + O(\delta). T ν ( x ) = δ κ + 1 + 2 A κ B + O ( δ ) = δ κ + 2 1 + h ν ( κ ) + O ( δ ) .
On the other hand, setting y = x 4 y=x^4 y = x 4 gives y − γ = 4 κ 3 δ + 6 κ 2 δ 2 + O ( δ 3 ) y-\gamma=4\kappa^3\delta+6\kappa^2\delta^2+O(\delta^3) y − γ = 4 κ 3 δ + 6 κ 2 δ 2 + O ( δ 3 ) , whence
4 γ y − γ = κ δ ( 1 + 3 δ 2 κ + O ( δ 2 ) ) − 1 = κ δ − 3 2 + O ( δ ) . \frac{4\gamma}{y-\gamma} = \frac{\kappa}{\delta}\left(1+\frac{3\delta}{2\kappa}+O(\delta^2)\right)^{-1} = \frac{\kappa}{\delta}-\frac{3}{2}+O(\delta). y − γ 4 γ = δ κ ( 1 + 2 κ 3 δ + O ( δ 2 ) ) − 1 = δ κ − 2 3 + O ( δ ) .
The change of variable y = x 4 y=x^4 y = x 4 is analytic and locally invertible at κ > 0 \kappa>0 κ > 0 , since its derivative there is 4 κ 3 ≠ 0 4\kappa^3\ne0 4 κ 3 = 0 . Both functions have the same simple pole in this coordinate. After subtracting the pole, the difference is analytic, with constant term ( 1 / 2 + h ν ( κ ) ) + 3 / 2 = 2 + h ν ( κ ) (1/2+h_\nu(\kappa))+3/2=2+h_\nu(\kappa) ( 1/2 + h ν ( κ )) + 3/2 = 2 + h ν ( κ ) . Subtracting that constant leaves an analytic function vanishing at y = γ y=\gamma y = γ , which can be divided by y − γ y-\gamma y − γ . This proves the exact representation
T ν ( x ) = 4 γ y − γ + 2 + h ν ( κ ) + ( y − γ ) A ν ( y ) , (14) T_\nu(x) = \frac{4\gamma}{y-\gamma} + 2 + h_\nu(\kappa) + (y-\gamma)A_\nu(y), \tag*{(14)} T ν ( x ) = y − γ 4 γ + 2 + h ν ( κ ) + ( y − γ ) A ν ( y ) , (14)
where A ν A_\nu A ν is analytic near γ \gamma γ .
Step 5. Solving the exact matching equation.
Write y = ( κ r ) 4 = γ ( 1 − ϵ Z ) y=(\kappa r)^4=\gamma(1-\epsilon Z) y = ( κ r ) 4 = γ ( 1 − ϵ Z ) and λ = y / r 4 \lambda=y/r^4 λ = y / r 4 . We solve for Z Z Z ; the required solution will have Z Z Z close to 1. Combining (10) , (13) and (14) , equality of the inner and outer moments becomes
1 r { 4 γ y − γ + 3 + h ν ( κ ) + ( y − γ ) A ν ( y ) } = 4 ϵ + ϵ R O ( ϵ , y / r 4 ) . \frac{1}{r}\left\{\frac{4\gamma}{y-\gamma}+3+h_\nu(\kappa)+(y-\gamma)A_\nu(y)\right\}=\frac{4}{\epsilon}+\epsilon R_O(\epsilon,y/r^4). r 1 { y − γ 4 γ + 3 + h ν ( κ ) + ( y − γ ) A ν ( y ) } = ϵ 4 + ϵ R O ( ϵ , y / r 4 ) .
Multiply by r ϵ r\epsilon r ϵ and use y − γ = − γ ϵ Z y-\gamma=-\gamma\epsilon Z y − γ = − γ ϵ Z . The pole term becomes 4 γ ϵ / ( y − γ ) = − 4 / Z 4\gamma\epsilon/(y-\gamma)=-4/Z 4 γ ϵ / ( y − γ ) = − 4/ Z , while the last term on the left becomes − γ ϵ 2 Z A ν ( y ) -\gamma\epsilon^2ZA_\nu(y) − γ ϵ 2 Z A ν ( y ) . Hence the exact matching equation is
− 4 Z + 4 r + ϵ { 3 + h ν ( κ ) } + ϵ 2 R ( ϵ , Z ) = 0 , -\frac{4}{Z}+4r+\epsilon\{3+h_\nu(\kappa)\}+\epsilon^2\mathcal{R}(\epsilon,Z)=0, − Z 4 + 4 r + ϵ { 3 + h ν ( κ )} + ϵ 2 R ( ϵ , Z ) = 0 ,
where the remainder is explicitly
R ( ϵ , Z ) = − γ Z A ν ( y ) − r R O ( ϵ , y / r 4 ) . \mathcal{R}(\epsilon,Z)=-\gamma ZA_\nu(y)-rR_O(\epsilon,y/r^4). R ( ϵ , Z ) = − γ Z A ν ( y ) − r R O ( ϵ , y / r 4 ) .
This is analytic near ( 0 , 1 ) (0,1) ( 0 , 1 ) , including at ϵ = 0 \epsilon=0 ϵ = 0 . We now find a real solution and bound its error. Using r = 1 − ϵ r=1-\epsilon r = 1 − ϵ , rewrite the equation as the fixed-point equation
Z = Φ ϵ ( Z ) , Φ ϵ ( Z ) = 4 4 + ϵ { h ν ( κ ) − 1 } + ϵ 2 R ( ϵ , Z ) . Z=\Phi_\epsilon(Z),\qquad\Phi_\epsilon(Z)=\frac{4}{4+\epsilon\{h_\nu(\kappa)-1\}+\epsilon^2\mathcal{R}(\epsilon,Z)}. Z = Φ ϵ ( Z ) , Φ ϵ ( Z ) = 4 + ϵ { h ν ( κ ) − 1 } + ϵ 2 R ( ϵ , Z ) 4 .
Choose 0 < c ≤ 1 / 2 0<c\le1/2 0 < c ≤ 1/2 and ϵ 0 > 0 \epsilon_0>0 ϵ 0 > 0 so that R \mathcal{R} R and ∂ Z R \partial_Z\mathcal{R} ∂ Z R are bounded by a constant M M M on ∣ Z − 1 ∣ ≤ c |Z-1|\le c ∣ Z − 1∣ ≤ c , ∣ ϵ ∣ ≤ ϵ 0 |\epsilon|\le\epsilon_0 ∣ ϵ ∣ ≤ ϵ 0 . For sufficiently small ∣ ϵ ∣ |\epsilon| ∣ ϵ ∣ , the denominator in Φ ϵ \Phi_\epsilon Φ ϵ is at least three in absolute value. On this rectangle,
∣ Φ ϵ ( Z ) − 1 ∣ ≤ ∣ ϵ ∣ ∣ h ν ( κ ) − 1 ∣ + ϵ 2 M 3 ≤ c , |\Phi_\epsilon(Z)-1|\le\frac{|\epsilon||h_\nu(\kappa)-1|+\epsilon^2M}{3}\le c, ∣ Φ ϵ ( Z ) − 1∣ ≤ 3 ∣ ϵ ∣∣ h ν ( κ ) − 1∣ + ϵ 2 M ≤ c ,
∣ ∂ Z Φ ϵ ( Z ) ∣ ≤ 4 ϵ 2 M 9 ≤ 1 2 . |\partial_Z\Phi_\epsilon(Z)|\le\frac{4\epsilon^2M}{9}\le\frac{1}{2}. ∣ ∂ Z Φ ϵ ( Z ) ∣ ≤ 9 4 ϵ 2 M ≤ 2 1 .
Thus iteration on this closed interval yields a unique real fixed point Z ϵ Z_\epsilon Z ϵ . To check continuity in the parameter, apply the contraction estimate to two such fixed points:
∣ Z ϵ − Z ϵ ′ ∣ ≤ 1 2 ∣ Z ϵ − Z ϵ ′ ∣ + sup ∣ Z − 1 ∣ ≤ c ∣ Φ ϵ ( Z ) − Φ ϵ ′ ( Z ) ∣ . |Z_\epsilon-Z_{\epsilon'}|\le\frac{1}{2}|Z_\epsilon-Z_{\epsilon'}|+\sup_{|Z-1|\le c}|\Phi_\epsilon(Z)-\Phi_{\epsilon'}(Z)|. ∣ Z ϵ − Z ϵ ′ ∣ ≤ 2 1 ∣ Z ϵ − Z ϵ ′ ∣ + ∣ Z − 1∣ ≤ c sup ∣ Φ ϵ ( Z ) − Φ ϵ ′ ( Z ) ∣.
The supremum tends to 0 0 0 as ϵ ′ → ϵ \epsilon'\to\epsilon ϵ ′ → ϵ , by uniform continuity on the fixed rectangle. Moving the first term to the left proves continuity of Z ϵ Z_\epsilon Z ϵ . The first estimate for Φ ϵ \Phi_\epsilon Φ ϵ , applied at the fixed point, gives w = Z − 1 = O ( ϵ ) w=Z-1=O(\epsilon) w = Z − 1 = O ( ϵ ) . To obtain the next term without assuming an expansion for Z Z Z , use the exact algebraic identity
− 4 1 + w + 4 = 4 w − 4 w 2 1 + w . -\frac{4}{1+w}+4=4w-\frac{4w^2}{1+w}. − 1 + w 4 + 4 = 4 w − 1 + w 4 w 2 .
The matching equation therefore gives
w = 1 − h ν ( κ ) 4 ϵ + w 2 1 + w − ϵ 2 4 R ( ϵ , 1 + w ) . w=\frac{1-h_\nu(\kappa)}{4}\epsilon+\frac{w^2}{1+w}-\frac{\epsilon^2}{4}\mathcal{R}(\epsilon,1+w). w = 4 1 − h ν ( κ ) ϵ + 1 + w w 2 − 4 ϵ 2 R ( ϵ , 1 + w ) .
The last two terms are O ( ϵ 2 ) O(\epsilon^2) O ( ϵ 2 ) : indeed, ∣ w ∣ ≤ C ∣ ϵ ∣ |w|\le C|\epsilon| ∣ w ∣ ≤ C ∣ ϵ ∣ , ∣ 1 + w ∣ ≥ 1 / 2 |1+w|\ge1/2 ∣1 + w ∣ ≥ 1/2 , and ∣ R ∣ ≤ M |\mathcal{R}|\le M ∣ R ∣ ≤ M . This proves
Z = 1 + 1 − h ν ( κ ) 4 ϵ + O ( ϵ 2 ) , y γ = 1 − ϵ + h ν ( κ ) − 1 4 ϵ 2 + O ( ϵ 3 ) . Z=1+\frac{1-h_\nu(\kappa)}{4}\epsilon+O(\epsilon^2),\qquad\frac{y}{\gamma}=1-\epsilon+\frac{h_\nu(\kappa)-1}{4}\epsilon^2+O(\epsilon^3). Z = 1 + 4 1 − h ν ( κ ) ϵ + O ( ϵ 2 ) , γ y = 1 − ϵ + 4 h ν ( κ ) − 1 ϵ 2 + O ( ϵ 3 ) .
The error for y / γ y/\gamma y / γ is of order ϵ 3 \epsilon^3 ϵ 3 because the O ( ϵ 2 ) O(\epsilon^2) O ( ϵ 2 ) error in Z Z Z is multiplied by ϵ \epsilon ϵ . Finally, r − 4 = 1 + 4 ϵ + 10 ϵ 2 + O ( ϵ 3 ) r^{-4}=1+4\epsilon+10\epsilon^2+O(\epsilon^3) r − 4 = 1 + 4 ϵ + 10 ϵ 2 + O ( ϵ 3 ) , and multiplication gives
λ γ = { 1 − ϵ + h ν ( κ ) − 1 4 ϵ 2 + O ( ϵ 3 ) } { 1 + 4 ϵ + 10 ϵ 2 + O ( ϵ 3 ) } = 1 + 3 ϵ + { 10 − 4 + h ν ( κ ) − 1 4 } ϵ 2 + O ( ϵ 3 ) . \begin{aligned}
\frac{\lambda}{\gamma}=\left\{1-\epsilon+\frac{h_\nu(\kappa)-1}{4}\epsilon^2+O(\epsilon^3)\right\}\left\{1+4\epsilon+10\epsilon^2+O(\epsilon^3)\right\} \\
&=1+3\epsilon+\left\{10-4+\frac{h_\nu(\kappa)-1}{4}\right\}\epsilon^2+O(\epsilon^3).
\end{aligned} γ λ = { 1 − ϵ + 4 h ν ( κ ) − 1 ϵ 2 + O ( ϵ 3 ) } { 1 + 4 ϵ + 10 ϵ 2 + O ( ϵ 3 ) } = 1 + 3 ϵ + { 10 − 4 + 4 h ν ( κ ) − 1 } ϵ 2 + O ( ϵ 3 ) .
Since the coefficient in braces equals ( 23 + h ν ( κ ) ) / 4 (23+h_\nu(\kappa))/4 ( 23 + h ν ( κ )) /4 , this is (3.2) with a bounded remainder.
Step 6. The weak eigenfunction and uniform remainder.
It remains to verify that the root just found gives an eigenfunction of the constrained form. For sufficiently small positive ϵ \epsilon ϵ , one has Z ≥ 1 / 2 Z\ge1/2 Z ≥ 1/2 and 0 < y = γ ( 1 − ϵ Z ) < γ 0<y=\gamma(1-\epsilon Z)<\gamma 0 < y = γ ( 1 − ϵ Z ) < γ . Since the zero κ \kappa κ of W ν W_\nu W ν is isolated, we may decrease the allowed ϵ \epsilon ϵ so that κ r = y 1 / 4 \kappa r=y^{1/4} κ r = y 1/4 lies below κ \kappa κ in an interval containing no other zero. In particular, W ν ( k r ) ≠ 0 W_\nu(kr) \ne0 W ν ( k r ) = 0 , and the inner normalization is legitimate. Take k = λ 1 / 4 k=\lambda^{1/4} k = λ 1/4 and define
f ( x ) = { F ( x ) / F ′ ( r ) , 0 < x ≤ r , ϵ Y ( ( x − r ) / ϵ ) , r ≤ x ≤ 1. f(x)=
\begin{cases}
F(x)/F'(r), & 0<x\le r,\\
\epsilon Y((x-r)/\epsilon), & r\le x\le1.
\end{cases} f ( x ) = { F ( x ) / F ′ ( r ) , ϵ Y (( x − r ) / ϵ ) , 0 < x ≤ r , r ≤ x ≤ 1.
Both values at r r r are 0, both slopes equal 1, and the moments agree by the exact matching equation. In particular f f f is nonzero. Choose a spherical harmonic Y l Y_l Y l of unit L 2 L^2 L 2 norm and put u ( x , θ ) = f ( x ) Y l ( θ ) u(x,\theta)=f(x)Y_l(\theta) u ( x , θ ) = f ( x ) Y l ( θ ) . The regular ball factors in Lemma 2.3 are smooth at the origin after multiplication by Y l Y_l Y l . On each side of the interface u u u is smooth and satisfies Δ 2 u = λ u \Delta^2u=\lambda u Δ 2 u = λ u .
To check global H 2 H^2 H 2 regularity, the gradient on either side has the polar expression
∇ u = f ′ ( x ) Y l ( θ ) θ + x − 1 f ( x ) ∇ S d − 1 Y l ( θ ) , \nabla u=f'(x)Y_l(\theta)\,\theta+x^{-1}f(x)\nabla_{\mathbb{S}^{d-1}}Y_l(\theta), ∇ u = f ′ ( x ) Y l ( θ ) θ + x − 1 f ( x ) ∇ S d − 1 Y l ( θ ) ,
where ∇ S d − 1 \nabla_{\mathbb{S}^{d-1}} ∇ S d − 1 is the tangential gradient. With [ ⋅ ] S r [\cdot]_{S_r} [ ⋅ ] S r denoting the outer trace minus the inner trace, the matching conditions give
[ u ] S r = 0 , [ ∇ u ] S r = [ f ′ ] r Y l θ + r − 1 [ f ] r ∇ S d − 1 Y l = 0. [u]_{S_r}=0,\qquad[\nabla u]_{S_r}=[f']_rY_l\theta+r^{-1}[f]_r\nabla_{\mathbb{S}^{d-1}}Y_l=0. [ u ] S r = 0 , [ ∇ u ] S r = [ f ′ ] r Y l θ + r − 1 [ f ] r ∇ S d − 1 Y l = 0.
All derivatives through order two are square integrable on the two pieces. Lemma 2.2 and the clamped traces at ∣ x ∣ = 1 |x|=1 ∣ x ∣ = 1 yield
u ∈ H 2 ( B ) , u ∣ ∂ B = ∂ ν u ∣ ∂ B = 0 ⟹ u ∈ H 0 2 ( B ) , Tr S r u = 0 ⟹ u ∈ V r . u\in H^2(B),\quad u|_{\partial B}=\partial_\nu u|_{\partial B}=0\Longrightarrow u\in H_0^2(B),\quad\operatorname{Tr}_{S_r}u=0\Longrightarrow u\in V_r. u ∈ H 2 ( B ) , u ∣ ∂ B = ∂ ν u ∣ ∂ B = 0 ⟹ u ∈ H 0 2 ( B ) , Tr S r u = 0 ⟹ u ∈ V r .
For the weak equation, first take an arbitrary v ∈ C c ∞ ( B ) v\in C^\infty_c(B) v ∈ C c ∞ ( B ) and integrate on the two regions. The outward normal of the inner ball is ∂ r \partial_r ∂ r , whereas the normal of the annulus at its inner boundary is − ∂ r -\partial_r − ∂ r . Their boundary terms sum to
∫ S r { ( Δ u ( r − ) − Δ u ( r + ) ) ∂ r v + ( ∂ r Δ u ( r + ) − ∂ r Δ u ( r − ) ) v } d S . \int_{S_r}\{(\Delta u(r-)-\Delta u(r+))\partial_rv+(\partial_r\Delta u(r+)-\partial_r\Delta u(r-))v\}\,dS. ∫ S r {( Δ u ( r − ) − Δ u ( r + )) ∂ r v + ( ∂ r Δ u ( r + ) − ∂ r Δ u ( r − )) v } d S .
The first term is 0 by moment matching. Hence
Q ( u , v ) − λ ( u , v ) = ∫ S r { ∂ r Δ u ( r + ) − ∂ r Δ u ( r − ) } Tr S r v d S . Q(u,v)-\lambda(u,v)=\int_{S_r}\{\partial_r\Delta u(r+)-\partial_r\Delta u(r-)\}\operatorname{Tr}_{S_r}v\,dS. Q ( u , v ) − λ ( u , v ) = ∫ S r { ∂ r Δ u ( r + ) − ∂ r Δ u ( r − )} Tr S r v d S .
Put J = ∂ r Δ u ( r + ) − ∂ r Δ u ( r − ) J=\partial_r\Delta u(r+)-\partial_r\Delta u(r-) J = ∂ r Δ u ( r + ) − ∂ r Δ u ( r − ) . Since J ∈ C ∞ ( S r ) J\in C^\infty(S_r) J ∈ C ∞ ( S r ) , Lemma 2.2 gives
∣ ∫ S r J Tr S r v d S ∣ ≤ ∥ J ∥ L 2 ( S r ) ∥ Tr S r v ∥ L 2 ( S r ) ≤ C ∥ J ∥ L 2 ( S r ) ∥ v ∥ H 2 ( B ) . \left|\int_{S_r}J\operatorname{Tr}_{S_r}v\,dS\right|\le\|J\|_{L^2(S_r)}\|\operatorname{Tr}_{S_r}v\|_{L^2(S_r)}\le C\|J\|_{L^2(S_r)}\|v\|_{H^2(B)}. ∫ S r J Tr S r v d S ≤ ∥ J ∥ L 2 ( S r ) ∥ Tr S r v ∥ L 2 ( S r ) ≤ C ∥ J ∥ L 2 ( S r ) ∥ v ∥ H 2 ( B ) .
For v m ∈ C c ∞ ( B ) v_m\in C^\infty_c(B) v m ∈ C c ∞ ( B ) with v m → v v_m\to v v m → v in H 2 ( B ) H^2(B) H 2 ( B ) , both sides of the preceding weak identity converge. Hence
Q ( u , v ) − λ ( u , v ) = ∫ S r J Tr S r v d S ( v ∈ H 0 2 ( B ) ) , Q(u,v)-\lambda(u,v)=\int_{S_r}J\operatorname{Tr}_{S_r}v\,dS\qquad(v\in H_0^2(B)), Q ( u , v ) − λ ( u , v ) = ∫ S r J Tr S r v d S ( v ∈ H 0 2 ( B )) ,
and, in particular,
v ∈ V r ⟹ Tr S r v = 0 ⟹ Q ( u , v ) = λ ( u , v ) . v \in V_r \Longrightarrow\operatorname{Tr}_{S_r} v = 0 \Longrightarrow Q(u,v) = \lambda(u,v). v ∈ V r ⟹ Tr S r v = 0 ⟹ Q ( u , v ) = λ ( u , v ) .
This proves the constrained eigenvalue equation without any pointwise regularity assumption on the test function.
For a finite set F \mathcal{F} F of separated modes, the preceding construction supplies constants ϵ ν , κ > 0 \epsilon_{\nu,\kappa} > 0 ϵ ν , κ > 0 and C ν , κ < + ∞ C_{\nu,\kappa} < +\infty C ν , κ < + ∞ such that
∣ λ ν , κ ( 1 − ϵ ) κ 4 − 1 − 3 ϵ − 23 + h ν ( κ ) 4 ϵ 2 ∣ ≤ C ν , κ ϵ 3 ( 0 < ϵ < ϵ ν , κ ) . \left\lvert\frac{\lambda_{\nu,\kappa}(1-\epsilon)}{\kappa^4} - 1 - 3\epsilon- \frac{23+h_{\nu}(\kappa)}{4}\epsilon^2 \right\rvert\le C_{\nu,\kappa}\epsilon^3 \qquad(0 < \epsilon< \epsilon_{\nu,\kappa}). κ 4 λ ν , κ ( 1 − ϵ ) − 1 − 3 ϵ − 4 23 + h ν ( κ ) ϵ 2 ≤ C ν , κ ϵ 3 ( 0 < ϵ < ϵ ν , κ ) .
Taking
ϵ F = min ( ν , κ ) ∈ F ϵ ν , κ > 0 , C F = max ( ν , κ ) ∈ F C ν , κ < + ∞ \epsilon_{\mathcal{F}} = \min_{(\nu,\kappa)\in\mathcal{F}} \epsilon_{\nu,\kappa} > 0, \qquad C_{\mathcal{F}} = \max_{(\nu,\kappa)\in\mathcal{F}} C_{\nu,\kappa} < +\infty ϵ F = ( ν , κ ) ∈ F min ϵ ν , κ > 0 , C F = ( ν , κ ) ∈ F max C ν , κ < + ∞
gives a common remainder bound. The radial equation depends on the degree l l l but not on the choice of harmonic within that degree, so the same constants cover its angular multiplicity.
We now compare the two eigenvalues. We first verify that the constructed branches occupy the first two positions in the ordered spectrum. We then compare their second-order coefficients and bound the remainder.
Proposition 3.2. For every fixed d ≥ 2 d \ge2 d ≥ 2 , there is r < 1 r < 1 r < 1 sufficiently close to 1 1 1 such that μ 2 ( r ) / μ 1 ( r ) > ρ d \mu_2(r)/\mu_1(r) > \rho_d μ 2 ( r ) / μ 1 ( r ) > ρ d .
Proof. By Boggio’s positivity theorem, the first ball eigenvalue γ 0 = k 0 4 \gamma_0 = k_0^4 γ 0 = k 0 4 is simple and radial, of order a = ( d − 2 ) / 2 a=(d-2)/2 a = ( d − 2 ) /2 . Write γ ∗ = k ∗ 4 = Γ 2 ( B ) \gamma_* = k_*^4 = \Gamma_2(B) γ ∗ = k ∗ 4 = Γ 2 ( B ) and let m m m be its multiplicity. Its separated orders form a finite set, all at least a a a . Simplicity of the first eigenvalue gives k ∗ > k 0 k_* > k_0 k ∗ > k 0 , and the definition of m m m gives Γ m + 2 ( B ) > γ ∗ \Gamma_{m+2}(B) > \gamma_* Γ m + 2 ( B ) > γ ∗ . These are the two spectral gaps used below. Define
h 0 = h a ( k 0 ) , h ∗ = min { h ν ( k ∗ ) : ν occurs at γ ∗ } . h_0 = h_a(k_0), \qquad h_* = \min\{h_\nu(k_*) : \nu\text{ occurs at } \gamma_*\}. h 0 = h a ( k 0 ) , h ∗ = min { h ν ( k ∗ ) : ν occurs at γ ∗ } .
The argument monotonicity cited in § 2 and (7) give, for every occurring ν ≥ a \nu\ge a ν ≥ a ,
h ν ( k ∗ ) ≥ h a ( k ∗ ) > h a ( k 0 ) = h 0 , h ∗ − h 0 > 0. h_\nu(k_*) \ge h_a(k_*) > h_a(k_0) = h_0, \qquad h_* - h_0 > 0. h ν ( k ∗ ) ≥ h a ( k ∗ ) > h a ( k 0 ) = h 0 , h ∗ − h 0 > 0.
We check explicitly that the constructed branches are the ordered first two eigenvalues. Subspace inclusion gives μ j ( r ) ≥ Γ j ( B ) \mu_j(r) \ge\Gamma_j(B) μ j ( r ) ≥ Γ j ( B ) for every j j j . Lemma 3.1 constructs a ground branch tending to γ 0 \gamma_0 γ 0 and m m m branches, counted with their spherical-harmonic multiplicities, tending to γ ∗ \gamma_* γ ∗ . They give m + 1 m+1 m + 1 independent eigenvectors: distinct harmonic sectors are orthogonal, and within each sector the angular basis is orthogonal. To make their separation explicit, put
g = 1 4 min { γ ∗ − γ 0 , Γ m + 2 ( B ) − γ ∗ } > 0. g = \frac{1}{4}\min\{\gamma_*-\gamma_0,\Gamma_{m+2}(B)-\gamma_*\} > 0. g = 4 1 min { γ ∗ − γ 0 , Γ m + 2 ( B ) − γ ∗ } > 0.
There are only finitely many constructed branches. Their convergence therefore permits a common choice of ϵ \epsilon ϵ such that the ground branch is within g g g of γ 0 \gamma_0 γ 0 and each of the other branches is within g g g of γ ∗ \gamma_* γ ∗ . In particular, the ground branch lies strictly below all the others and below γ ∗ \gamma_* γ ∗ , while all the constructed branches lie below Γ m + 2 ( B ) \Gamma_{m+2}(B) Γ m + 2 ( B ) . If a radial mode occurs at γ ∗ \gamma_* γ ∗ , its orthogonality to the ground branch follows from the weak equations: for eigenvectors u , v u,v u , v with distinct eigenvalues λ , μ \lambda,\mu λ , μ ,
λ ( u , v ) = Q ( u , v ) = Q ( v , u ) = μ ( u , v ) , \lambda(u,v) = Q(u,v) = Q(v,u) = \mu(u,v), λ ( u , v ) = Q ( u , v ) = Q ( v , u ) = μ ( u , v ) ,
so ( u , v ) = 0 (u,v)=0 ( u , v ) = 0 . Together with angular orthogonality, this proves independence of all m + 1 m+1 m + 1 constructed eigenvectors.
There cannot be another eigenvalue below γ ∗ \gamma_* γ ∗ , because μ 2 ≥ γ ∗ \mu_2 \ge\gamma_* μ 2 ≥ γ ∗ . The ground branch must therefore be μ 1 \mu_1 μ 1 . Set β = Γ m + 2 ( B ) \beta=\Gamma_{m+2}(B) β = Γ m + 2 ( B ) and let E < β E_{<\beta} E < β be the sum of constrained eigenspaces with eigenvalues below β \beta β . The constructed branches and the inclusion comparison give
dim E < β ≥ m + 1 , μ m + 2 ( r ) ≥ β . \dim E_{<\beta} \ge m+1,\qquad\mu_{m+2}(r) \ge\beta. dim E < β ≥ m + 1 , μ m + 2 ( r ) ≥ β .
If dim E < β ≥ m + 2 \dim E_{<\beta} \ge m+2 dim E < β ≥ m + 2 , choose m + 2 m+2 m + 2 orthonormal eigenvectors there. By (2) , their span E E E satisfies
μ m + 2 ( r ) ≤ max 0 ≠ u ∈ E Q ( u , u ) ∥ u ∥ 2 = max 1 ≤ j ≤ m + 2 μ j < β , \mu_{m+2}(r) \le\max_{0\ne u\in E} \frac{Q(u,u)}{\lVert u\rVert^2} = \max_{1\le j\le m+2} \mu_j < \beta, μ m + 2 ( r ) ≤ 0 = u ∈ E max ∥ u ∥ 2 Q ( u , u ) = 1 ≤ j ≤ m + 2 max μ j < β ,
where μ j \mu_j μ j denotes the eigenvalue attached to the chosen j j j th vector. This is a contradiction. Hence dim E < β = m + 1 \dim E_{<\beta}=m+1 dim E < β = m + 1 . Thus the m m m constructed branches near γ ∗ \gamma_* γ ∗ fill exactly positions 2 , … , m + 1 2,\ldots,m+1 2 , … , m + 1 . The argument counts the dimensions of eigenspaces and hence also applies when some branches coincide.
Write c 0 = ( 23 + h 0 ) / 4 c_0=(23+h_0)/4 c 0 = ( 23 + h 0 ) /4 and c ∗ = ( 23 + h ∗ ) / 4 c_*=(23+h_*)/4 c ∗ = ( 23 + h ∗ ) /4 . Taking the minimum over the branches tending to the second ball eigenvalue gives
μ 2 ( 1 − ϵ ) = γ ∗ { 1 + 3 ϵ + c ∗ ϵ 2 + O ( ϵ 3 ) } . \mu_2(1-\epsilon)=\gamma_*\left\{1+3\epsilon+c_*\epsilon^2+O\left(\epsilon^3\right)\right\}. μ 2 ( 1 − ϵ ) = γ ∗ { 1 + 3 ϵ + c ∗ ϵ 2 + O ( ϵ 3 ) } .
Indeed every branch has its quadratic coefficient at least c ∗ c_* c ∗ , and at least one has coefficient exactly c ∗ c_* c ∗ . A common bound C ϵ 3 C\epsilon^3 C ϵ 3 on their remainders bounds their minimum between the same quadratic polynomial minus and plus C ϵ 3 C\epsilon^3 C ϵ 3 . For the ground branch, write t = 3 ϵ + c 0 ϵ 2 + O ( ϵ 3 ) t=3\epsilon+c_0\epsilon^2+O\left(\epsilon^3\right) t = 3 ϵ + c 0 ϵ 2 + O ( ϵ 3 ) and decrease ϵ \epsilon ϵ so that ∣ t ∣ ≤ 1 / 2 |t|\le1/2 ∣ t ∣ ≤ 1/2 . The exact identity
1 1 + t = 1 − t + t 2 − t 3 1 + t \frac{1}{1+t}=1-t+t^2-\frac{t^3}{1+t} 1 + t 1 = 1 − t + t 2 − 1 + t t 3
has a remainder of absolute value at most 2 ∣ t ∣ 3 2|t|^3 2∣ t ∣ 3 . Since t 2 = 9 ϵ 2 + O ( ϵ 3 ) t^2=9\epsilon^2+O\left(\epsilon^3\right) t 2 = 9 ϵ 2 + O ( ϵ 3 ) , it follows that
( 1 + 3 ϵ + c 0 ϵ 2 + O ( ϵ 3 ) ) − 1 = 1 − 3 ϵ + ( 9 − c 0 ) ϵ 2 + O ( ϵ 3 ) . \left(1+3\epsilon+c_0\epsilon^2+O\left(\epsilon^3\right)\right)^{-1}=1-3\epsilon+(9-c_0)\epsilon^2+O\left(\epsilon^3\right). ( 1 + 3 ϵ + c 0 ϵ 2 + O ( ϵ 3 ) ) − 1 = 1 − 3 ϵ + ( 9 − c 0 ) ϵ 2 + O ( ϵ 3 ) .
Multiplying this expression by the normalized second eigenvalue gives linear coefficient 0 0 0 , and its quadratic coefficient is
c ∗ + ( 9 − c 0 ) + 3 ( − 3 ) = c ∗ − c 0 = h ∗ − h 0 4 > 0. c_*+(9-c_0)+3(-3)=c_*-c_0=\frac{h_*-h_0}{4}>0. c ∗ + ( 9 − c 0 ) + 3 ( − 3 ) = c ∗ − c 0 = 4 h ∗ − h 0 > 0.
The bounded reciprocal also preserves the order of the remainder. Therefore
μ 2 ( 1 − ϵ ) μ 1 ( 1 − ϵ ) = ρ d [ 1 + h ∗ − h 0 4 ϵ 2 + O ( ϵ 3 ) ] > ρ d \frac{\mu_2(1-\epsilon)}{\mu_1(1-\epsilon)}=\rho_d\left[1+\frac{h_* - h_0}{4}\epsilon^2+O\left(\epsilon^3\right)\right]>\rho_d μ 1 ( 1 − ϵ ) μ 2 ( 1 − ϵ ) = ρ d [ 1 + 4 h ∗ − h 0 ϵ 2 + O ( ϵ 3 ) ] > ρ d
for sufficiently small positive ϵ \epsilon ϵ . More precisely, if the absolute remainder inside the brackets is bounded by C ϵ 3 C\epsilon^3 C ϵ 3 , choose ϵ \epsilon ϵ small enough that C ϵ < ( h ∗ − h 0 ) / 8 C\epsilon< (h_* - h_0)/8 C ϵ < ( h ∗ − h 0 ) /8 , in addition to the earlier finite set of restrictions. Then
μ 2 ( 1 − ϵ ) μ 1 ( 1 − ϵ ) − ρ d > ρ d ( h ∗ − h 0 ) 8 ϵ 2 > 0. \frac{\mu_2(1-\epsilon)}{\mu_1(1-\epsilon)}-\rho_d>\frac{\rho_d(h_* - h_0)}{8}\epsilon^2>0. μ 1 ( 1 − ϵ ) μ 2 ( 1 − ϵ ) − ρ d > 8 ρ d ( h ∗ − h 0 ) ϵ 2 > 0.
This supplies a definite strict margin once r = 1 − ϵ r=1-\epsilon r = 1 − ϵ is fixed. Figure 4 illustrates the comparison between the positive quadratic term and the cubic error. Only the orders actually occurring in the second eigenspace enter the minimum defining h ∗ h_* h ∗ ; no classification of their angular degrees is needed.
Figure 4. Why the second-order calculation gives a strict inequality. The relative gain is ( h ∗ − h 0 ) ϵ 2 / 4 + O ( ϵ 3 ) (h_* - h_0)\epsilon^2/4 + O(\epsilon^3) ( h ∗ − h 0 ) ϵ 2 /4 + O ( ϵ 3 ) . The solid curve represents its positive quadratic term; the dashed curves represent upper and lower bounds after allowing for the cubic remainder. For sufficiently small positive ϵ \epsilon ϵ , even the lower bound is positive. The axes are rescaled for this schematic diagram.
Approximation by finitely many latitudes
Fix the radius r r r obtained in Proposition 3.2 . This section replaces the spherical value constraint by finitely many latitude constraints. The main point is continuity with respect to the height of a latitude. The restriction of an H 2 H^2 H 2 function to the sphere need not be pointwise continuous in every dimension; instead, its restrictions to nearby latitudes are close in L 2 L^2 L 2 , after the latitudes have been parametrized by the same unit sphere. Vanishing on a dense set of latitudes therefore forces the entire spherical trace to vanish. Compactness then gives convergence of the eigenvalues. The supports will be thickened only after their number has been fixed, in § 5 .
Lemma 4.1. Fix 0 < r < 1 0<r<1 0 < r < 1 , choose distinct a j a_j a j dense in ( − r , r ) (-r,r) ( − r , r ) , and set
M j = { x : ∣ x ∣ = r , x d = a j } , V N = { u ∈ H 0 2 ( B ) : Tr M j u = 0 , 1 ≤ j ≤ N } . M_j=\{x:|x|=r,\ x_d=a_j\},\qquad V_N=\{u\in H_0^2(B):\operatorname{Tr}_{M_j}u=0,\ 1\leq j\leq N\}. M j = { x : ∣ x ∣ = r , x d = a j } , V N = { u ∈ H 0 2 ( B ) : Tr M j u = 0 , 1 ≤ j ≤ N } . Then V N V_N V N are closed decreasing spaces, ⋂ N V N = V r \bigcap_N V_N=V_r ⋂ N V N = V r , and for every fixed k k k ,
λ k ( V N ) ↑ μ k ( r ) . (15) \lambda_k(V_N)\mathrel{\uparrow}\mu_k(r). \tag*{(15)} λ k ( V N ) ↑ μ k ( r ) . (15)
Proof. Step 1. The intersection. Each latitude is a compact smooth codimension-two submanifold, so its value trace is bounded by Lemma 2.2. Thus V N V_N V N is an intersection of finitely many closed kernels, and V N + 1 ⊂ V N V_{N+1}\subset V_N V N + 1 ⊂ V N . These spaces are infinite dimensional: every function in C c ∞ ( ( r / 2 ) B ) C_c^\infty((r/2)B) C c ∞ (( r /2 ) B ) vanishes on every latitude. The eigenvalues in the statement are therefore defined by the variational framework of § 2 .
By Lemma 2.2,
U = Tr S r u ∈ H 3 / 2 ( S r ) ⊂ H 1 ( S r ) , ∥ U ∥ H 3 / 2 ( S r ) ≤ C r ∥ u ∥ H 2 ( B ) . U=\operatorname{Tr}_{S_r}u\in H^{3/2}(S_r)\subset H^1(S_r),\qquad\lVert U\rVert_{H^{3/2}(S_r)}\leq C_r\lVert u\rVert_{H^2(B)}. U = Tr S r u ∈ H 3/2 ( S r ) ⊂ H 1 ( S r ) , ∥ U ∥ H 3/2 ( S r ) ≤ C r ∥ u ∥ H 2 ( B ) .
We now use this standard trace estimate to identify the intersection of the latitude kernels.
Fix a closed latitude band with − r < a − < a + < r -r<a_-<a_+<r − r < a − < a + < r , and put I = ( a − , a + ) I=(a_-,a_+) I = ( a − , a + ) . Writing ϱ ( a ) = r 2 − a 2 \varrho(a)=\sqrt{r^2-a^2} ϱ ( a ) = r 2 − a 2 , parametrize the band by
F ( a , ω ) = ( ϱ ( a ) ω , a ) , a ∈ I , ω ∈ S d − 2 . F(a,\omega)=(\varrho(a)\omega,a),\qquad a\in I,\quad\omega\in S^{d-2}. F ( a , ω ) = ( ϱ ( a ) ω , a ) , a ∈ I , ω ∈ S d − 2 .
The derivative in a a a has length r / ϱ ( a ) r/\varrho(a) r / ϱ ( a ) , whereas tangent vectors in the sphere directions are multiplied by ϱ ( a ) \varrho(a) ϱ ( a ) . These directions are orthogonal. Thus
d S = r ϱ ( a ) d − 3 d a d ω , dS=r\varrho(a)^{d-3}\,da\,d\omega, d S = r ϱ ( a ) d − 3 d a d ω ,
where d ω d\omega d ω is surface measure on S d − 2 S^{d-2} S d − 2 and counting measure when d = 2 d=2 d = 2 . On the chosen band ϱ ( a ) \varrho(a) ϱ ( a ) is bounded away from 0 0 0 , so the coordinate metric, its inverse and this density have uniform bounds. Since H 3 / 2 ⊂ H 1 H^{3/2}\subset H^1 H 3/2 ⊂ H 1 , the pullback
U ~ ( a ) ( ω ) = U ( F ( a , ω ) ) \widetilde{U}(a)(\omega)=U(F(a,\omega)) U ( a ) ( ω ) = U ( F ( a , ω ))
belongs to H 1 ( I ; H ) H^1(I;\mathcal{H}) H 1 ( I ; H ) , where H = L 2 ( S d − 2 ) \mathcal{H}=L^2(S^{d-2}) H = L 2 ( S d − 2 ) . This means that U ~ \widetilde{U} U and its weak a a a derivative are square integrable with values in the Hilbert space H \mathcal{H} H .
For a smooth H \mathcal{H} H -valued function, the fundamental theorem of calculus and Cauchy–Schwarz give
∥ U ~ ( a ) − U ~ ( b ) ∥ H ≤ ∫ min { a , b } max { a , b } ∥ ∂ a U ~ ( t ) ∥ H d t ≤ ∣ a − b ∣ 1 / 2 ∥ ∂ a U ~ ∥ L 2 ( I ; H ) . \lVert\widetilde{U}(a)-\widetilde{U}(b)\rVert_{\mathcal{H}}
\leq\int_{\min\{a,b\}}^{\max\{a,b\}}\lVert\partial_a\widetilde{U}(t)\rVert_{\mathcal{H}}\,dt
\leq|a-b|^{1/2}\lVert\partial_a\widetilde{U}\rVert_{L^2(I;\mathcal{H})}. ∥ U ( a ) − U ( b ) ∥ H ≤ ∫ m i n { a , b } m a x { a , b } ∥ ∂ a U ( t ) ∥ H d t ≤ ∣ a − b ∣ 1/2 ∥ ∂ a U ∥ L 2 ( I ; H ) .
Average the inequality ∥ U ~ ( a ) ∥ H ≤ ∥ U ~ ( b ) ∥ H + ∣ I ∣ 1 / 2 ∥ ∂ a U ~ ∥ L 2 ( I ; H ) \|\widetilde{U}(a)\|_{\mathcal{H}} \le\|\widetilde{U}(b)\|_{\mathcal{H}} + |I|^{1/2}\|\partial_a\widetilde{U}\|_{L^2(I;\mathcal{H})} ∥ U ( a ) ∥ H ≤ ∥ U ( b ) ∥ H + ∣ I ∣ 1/2 ∥ ∂ a U ∥ L 2 ( I ; H ) over b ∈ I b \in I b ∈ I . A further Cauchy–Schwarz estimate gives
sup a ∈ I ∥ U ~ ( a ) ∥ H ≤ ∣ I ∣ − 1 / 2 ∥ U ~ ∥ L 2 ( I ; H ) + ∣ I ∣ 1 / 2 ∥ ∂ a U ~ ∥ L 2 ( I ; H ) . \sup_{a\in I}\|\widetilde{U}(a)\|_{\mathcal{H}} \le|I|^{-1/2}\|\widetilde{U}\|_{L^2(I;\mathcal{H})} + |I|^{1/2}\|\partial_a\widetilde{U}\|_{L^2(I;\mathcal{H})}. a ∈ I sup ∥ U ( a ) ∥ H ≤ ∣ I ∣ − 1/2 ∥ U ∥ L 2 ( I ; H ) + ∣ I ∣ 1/2 ∥ ∂ a U ∥ L 2 ( I ; H ) .
Approximation by smooth functions in H 1 ( I ; H ) H^1(I;\mathcal{H}) H 1 ( I ; H ) now gives a uniformly convergent sequence of continuous representatives. Its limit is a continuous representative of U ~ \widetilde{U} U , and both displayed estimates remain valid for it. In particular, evaluation at any fixed height is bounded.
At a selected height a j ∈ I a_j \in I a j ∈ I , evaluation of U ~ \widetilde{U} U agrees with the trace Tr M j u \operatorname{Tr}_{M_j}u Tr M j u expressed in the variable ω \omega ω . To verify this identity, take u m ∈ C c ∞ ( B ) u_m \in C^\infty_c(B) u m ∈ C c ∞ ( B ) converging to u u u in H 2 H^2 H 2 . The spherical trace estimate followed by the evaluation bound shows that the evaluated traces of u m u_m u m converge in H \mathcal{H} H . The codimension-two trace estimate shows convergence of their traces on M j M_j M j in L 2 ( M j ) L^2(M_j) L 2 ( M j ) . The two norms are related by
∥ f ∥ L 2 ( M j ) 2 = ϱ ( a j ) d − 2 ∥ f ( F ( a j , ⋅ ) ) ∥ H 2 . \|f\|_{L^2(M_j)}^2=\varrho(a_j)^{d-2}\|f(F(a_j,\cdot))\|_{\mathcal{H}}^2. ∥ f ∥ L 2 ( M j ) 2 = ϱ ( a j ) d − 2 ∥ f ( F ( a j , ⋅ )) ∥ H 2 .
The surface factor is fixed and positive, so convergence in either norm implies convergence in the other. The two restrictions agree for every smooth u m u_m u m and hence agree in the limit.
If u ∈ ⋂ N V N u \in\bigcap_N V_N u ∈ ⋂ N V N , take any a ∈ I a \in I a ∈ I and selected heights a j m ∈ I a_{j_m} \in I a j m ∈ I with a j m → a a_{j_m} \to a a j m → a . The trace consistency and the preceding continuity estimate give
U ~ ( a j m ) = 0 , ∥ U ~ ( a ) ∥ H ≤ ∣ a − a j m ∣ 1 / 2 ∥ ∂ a U ~ ∥ L 2 ( I ; H ) ⟶ 0. \widetilde{U}(a_{j_m})=0,\qquad\|\widetilde{U}(a)\|_{\mathcal{H}}\le|a-a_{j_m}|^{1/2}\|\partial_a\widetilde{U}\|_{L^2(I;\mathcal{H})}\longrightarrow0. U ( a j m ) = 0 , ∥ U ( a ) ∥ H ≤ ∣ a − a j m ∣ 1/2 ∥ ∂ a U ∥ L 2 ( I ; H ) ⟶ 0.
The sphere minus its two poles is covered by countably many of these bands, so U = 0 U=0 U = 0 almost everywhere on S r S_r S r . The poles have surface measure 0. Conversely, if U = 0 U=0 U = 0 , its pullback is 0 almost everywhere on each band, and its continuous representative is identically 0 there. The trace consistency just proved implies Tr M j u = 0 \operatorname{Tr}_{M_j}u=0 Tr M j u = 0 for every j j j . This proves ⋂ N V N = V r \bigcap_N V_N=V_r ⋂ N V N = V r . In dimension two, H = L 2 ( S 0 ) = R 2 \mathcal{H}=L^2(S^0)=\mathbb{R}^2 H = L 2 ( S 0 ) = R 2 simply records the values on the two arcs, so the same argument applies.
Step 2. The spectral limit.
Write λ i , N = λ i ( V N ) \lambda_{i,N}=\lambda_i(V_N) λ i , N = λ i ( V N ) . The inclusions V r ⊂ V N + 1 ⊂ V N V_r\subset V_{N+1}\subset V_N V r ⊂ V N + 1 ⊂ V N imply
λ k , N ≤ λ k , N + 1 ≤ μ k ( r ) . \lambda_{k,N}\le\lambda_{k,N+1}\le\mu_k(r). λ k , N ≤ λ k , N + 1 ≤ μ k ( r ) .
Thus L k = lim N → + ∞ λ k , N L_k=\lim_{N\to+\infty}\lambda_{k,N} L k = lim N → + ∞ λ k , N exists and is at most μ k ( r ) \mu_k(r) μ k ( r ) . We prove the reverse inequality by retaining the whole first k k k -dimensional spectral subspace in the limit. Choose real L 2 L^2 L 2 -orthonormal eigenvectors u 1 , N , … , u k , N u_{1,N},\ldots,u_{k,N} u 1 , N , … , u k , N on V N V_N V N . Their weak eigenvalue equations give
Q ( u i , N , u j , N ) = λ i , N ( u i , N , u j , N ) = λ i , N δ i j , Q ( u i , N , u i , N ) ≤ μ k ( r ) . Q(u_{i,N},u_{j,N})=\lambda_{i,N}(u_{i,N},u_{j,N})=\lambda_{i,N}\delta_{ij},\qquad Q(u_{i,N},u_{i,N})\le\mu_k(r). Q ( u i , N , u j , N ) = λ i , N ( u i , N , u j , N ) = λ i , N δ ij , Q ( u i , N , u i , N ) ≤ μ k ( r ) .
The energy norm equivalence in Section 2 bounds their H 2 H^2 H 2 norms. Weak compactness and the compact embedding into L 2 L^2 L 2 yield one common subsequence N m → + ∞ N_m \to+\infty N m → + ∞ such that, for 1 ≤ i ≤ k 1 \le i \le k 1 ≤ i ≤ k ,
u i , N m ⇀ u i in H 2 ( B ) , u i , N m ⟶ u i in L 2 ( B ) . u_{i,N_m} \rightharpoonup u_i \text{ in } H^2(B), \qquad u_{i,N_m} \longrightarrow u_i \text{ in } L^2(B). u i , N m ⇀ u i in H 2 ( B ) , u i , N m ⟶ u i in L 2 ( B ) .
Only finitely many vectors are involved, so repeated subsequence selection gives these convergences simultaneously. For every fixed M M M , the tail with N m ≥ M N_m \ge M N m ≥ M belongs to V M V_M V M . A closed linear subspace is weakly closed; hence each u i u_i u i lies in every V M V_M V M , and therefore in V r V_r V r .
Strong L 2 L^2 L 2 convergence preserves all inner products. For example,
∣ ( u i , u j ) − ( u i , N m , u j , N m ) ∣ ≤ ∥ u i − u i , N m ∥ 2 ∥ u j ∥ 2 + ∥ u i , N m ∥ 2 ∥ u j − u j , N m ∥ 2 ⟶ 0. \left|(u_i,u_j)-(u_{i,N_m},u_{j,N_m})\right|
\le\|u_i-u_{i,N_m}\|_2\|u_j\|_2
+\|u_{i,N_m}\|_2\|u_j-u_{j,N_m}\|_2 \longrightarrow0. ∣ ( u i , u j ) − ( u i , N m , u j , N m ) ∣ ≤ ∥ u i − u i , N m ∥ 2 ∥ u j ∥ 2 + ∥ u i , N m ∥ 2 ∥ u j − u j , N m ∥ 2 ⟶ 0.
Consequently ( u i , u j ) = δ i j (u_i,u_j)=\delta_{ij} ( u i , u j ) = δ ij , and E = span { u 1 , … , u k } E=\operatorname{span}\{u_1,\ldots,u_k\} E = span { u 1 , … , u k } has dimension k k k . For any fixed c ∈ R k c\in\mathbb{R}^k c ∈ R k , linearity of weak convergence and weak lower semicontinuity of the L 2 L^2 L 2 norm of the Laplacian give
Q ( ∑ i c i u i , ∑ i c i u i ) ≤ lim inf m → + ∞ Q ( ∑ i c i u i , N m , ∑ i c i u i , N m ) = lim inf m → + ∞ ∑ i c i 2 λ i , N m ≤ L k ∑ i c i 2 . Q\left(\sum_i c_i u_i,\sum_i c_i u_i\right)
\le\liminf_{m\to+\infty}Q\left(\sum_i c_i u_{i,N_m},\sum_i c_i u_{i,N_m}\right)
= \liminf_{m\to+\infty}\sum_i c_i^2\lambda_{i,N_m}
\le L_k\sum_i c_i^2. Q ( i ∑ c i u i , i ∑ c i u i ) ≤ m → + ∞ lim inf Q ( i ∑ c i u i , N m , i ∑ c i u i , N m ) = m → + ∞ lim inf i ∑ c i 2 λ i , N m ≤ L k i ∑ c i 2 .
The subsequence was selected for the basis vectors, so no further subsequence depending on c c c is needed. Since ∥ ∑ i c i u i ∥ 2 2 = ∑ i c i 2 \left\|\sum_i c_i u_i\right\|_2^2=\sum_i c_i^2 ∥ ∑ i c i u i ∥ 2 2 = ∑ i c i 2 , every nonzero vector of E E E has Rayleigh quotient at most L k L_k L k . Using this single k k k -dimensional space in the min–max formula yields
μ k ( r ) ≤ max 0 ≠ u ∈ E Q ( u , u ) ∥ u ∥ 2 2 ≤ L k . \mu_k(r)\le\max_{0\ne u\in E}\frac{Q(u,u)}{\|u\|_2^2}\le L_k. μ k ( r ) ≤ 0 = u ∈ E max ∥ u ∥ 2 2 Q ( u , u ) ≤ L k .
Together with the opposite inequality already established, this proves (15) .
From latitude supports to smooth holes
The finite family of supports is now fixed. The issue is to replace value conditions on sets of dimension d − 2 d-2 d − 2 by both clamped conditions on actual hole boundaries. We first prove the needed approximation of functions in H 2 H^2 H 2 , and then apply it to finite-dimensional trial spaces to obtain eigenvalue convergence.
The limiting space and logarithmic cutoff
The ingredients are the classical trace and extension estimates [1 Chapter 7] and the logarithmic cutoff in two normal dimensions, also used in [14 Section 6] . We record their consequence for the present form space. The density argument is included because its conclusion determines which boundary condition survives when the holes shrink.
Lemma 5.1. Let Σ \Sigma Σ be a finite disjoint union of compact smooth embedded codimension-two submanifolds without boundary in B B B . With zero extension to B B B understood.
H 0 2 ( B ∖ Σ ) = { u ∈ H 0 2 ( B ) : Tr Σ u = 0 } . (16) H^2_0(B \setminus\Sigma) = \{u \in H^2_0(B) : \operatorname{Tr}_{\Sigma} u = 0\}. \tag*{(16)} H 0 2 ( B ∖ Σ ) = { u ∈ H 0 2 ( B ) : Tr Σ u = 0 } . (16) Here Tr Σ : H 2 ( B ) → H 1 ( Σ ) \operatorname{Tr}_{\Sigma}: H^2(B) \to H^1(\Sigma) Tr Σ : H 2 ( B ) → H 1 ( Σ ) is a bounded value trace. For d = 2 d=2 d = 2 this is the finite vector of point values.
Proof. We must approximate every H 2 H^2 H 2 function with value trace 0 0 0 on Σ \Sigma Σ by smooth functions vanishing in a neighborhood of Σ \Sigma Σ . Ordinary smooth approximation need not preserve that trace. Steps 1 and 2 correct this first; Steps 3 and 4 then remove a neighborhood with an error tending to 0 0 0 in H 2 H^2 H 2 .
Step 1. Correcting the trace of a smooth approximation.
Lemma 2.2 supplies the bounded value trace. We also need a bounded extension operator E : H 1 ( Σ ) → H 0 2 ( B ) E:H^1(\Sigma)\to H^2_0(B) E : H 1 ( Σ ) → H 0 2 ( B ) such that Tr Σ ( E f ) = f \operatorname{Tr}_{\Sigma}(Ef)=f Tr Σ ( E f ) = f and smooth data give functions in C c ∞ ( B ) C^\infty_c(B) C c ∞ ( B ) . Here is the local formula and estimate, fixing the Fourier convention explicitly. Write ( y , z ) ∈ R d − 2 × R 2 (y,z)\in\mathbb{R}^{d-2}\times\mathbb{R}^2 ( y , z ) ∈ R d − 2 × R 2 , let q = ( 1 + ∣ ξ ∣ 2 ) 1 / 2 q=(1+|\xi|^2)^{1/2} q = ( 1 + ∣ ξ ∣ 2 ) 1/2 , and choose χ ∈ C ∞ ( R 2 ) \chi\in C^\infty(\mathbb{R}^2) χ ∈ C ∞ ( R 2 ) equal to 1 1 1 near 0 0 0 . Set
E f ^ ( ξ , z ) = χ ( z ) e − q 2 ∣ z ∣ 2 f ^ ( ξ ) , \widehat{Ef}(\xi,z)=\chi(z)e^{-q^2|z|^2}\widehat{f}(\xi), E f ( ξ , z ) = χ ( z ) e − q 2 ∣ z ∣ 2 f ( ξ ) ,
f ^ ( ξ ) = ( 2 π ) − ( d − 2 ) / 2 ∫ R d − 2 e − i y ⋅ ξ f ( y ) d y . (17) \widehat{f}(\xi)=(2\pi)^{-(d-2)/2}\int_{\mathbb{R}^{d-2}}e^{-iy\cdot\xi}f(y)\,dy. \tag*{(17)} f ( ξ ) = ( 2 π ) − ( d − 2 ) /2 ∫ R d − 2 e − i y ⋅ ξ f ( y ) d y . (17)
For this unitary transform, Plancherel’s theorem states
∫ R d − 2 ∣ g ( y ) ∣ 2 d y = ∫ R d − 2 ∣ g ^ ( ξ ) ∣ 2 d ξ , D y β g ^ = ( i ξ ) β g ^ ; \int_{\mathbb{R}^{d-2}}|g(y)|^2\,dy=\int_{\mathbb{R}^{d-2}}|\widehat{g}(\xi)|^2\,d\xi,\qquad\widehat{D^\beta_y g}=(i\xi)^\beta\widehat{g}; ∫ R d − 2 ∣ g ( y ) ∣ 2 d y = ∫ R d − 2 ∣ g ( ξ ) ∣ 2 d ξ , D y β g = ( i ξ ) β g ;
see [16 Chapter 8] . Applied for each fixed z z z , and then integrated in z z z by Tonelli’s theorem, it gives
∥ D y β D z γ E f ∥ 2 2 = ∫ R d − 2 ∣ ξ β ∣ 2 ∣ f ^ ( ξ ) ∣ 2 ∫ R 2 ∣ D z γ ( χ ( z ) e − q 2 ∣ z ∣ 2 ) ∣ 2 d z d ξ ≤ C ∫ R d − 2 q 2 ∣ f ^ ( ξ ) ∣ 2 d ξ = C ∥ f ∥ H 1 2 , ∣ β ∣ + ∣ γ ∣ ≤ 2. \begin{aligned}
\left\|D^\beta_yD^\gamma_zEf\right\|_2^2
=\int_{\mathbb{R}^{d-2}}|\xi^\beta|^2|\widehat{f}(\xi)|^2
\int_{\mathbb{R}^2}\left|D^\gamma_z\left(\chi(z)e^{-q^2|z|^2}\right)\right|^2\,dz\,d\xi\\
&\leq C\int_{\mathbb{R}^{d-2}}q^2|\widehat{f}(\xi)|^2\,d\xi
=C\|f\|_{H^1}^2,\qquad|\beta|+|\gamma|\leq2.
\end{aligned} D y β D z γ E f 2 2 = ∫ R d − 2 ∣ ξ β ∣ 2 ∣ f ( ξ ) ∣ 2 ∫ R 2 D z γ ( χ ( z ) e − q 2 ∣ z ∣ 2 ) 2 d z d ξ ≤ C ∫ R d − 2 q 2 ∣ f ( ξ ) ∣ 2 d ξ = C ∥ f ∥ H 1 2 , ∣ β ∣ + ∣ γ ∣ ≤ 2.
To verify the inequality, write D z μ e − q 2 ∣ z ∣ 2 = q ∣ μ ∣ P μ ( q z ) e − q 2 ∣ z ∣ 2 D^\mu_z e^{-q^2|z|^2}=q^{|\mu|}P_\mu(qz)e^{-q^2|z|^2} D z μ e − q 2 ∣ z ∣ 2 = q ∣ μ ∣ P μ ( q z ) e − q 2 ∣ z ∣ 2 . Each product-rule term with μ ≤ γ \mu\leq\gamma μ ≤ γ is bounded using
∣ ξ ∣ 2 ∣ β ∣ ∫ R 2 ∣ D z μ e − q 2 ∣ z ∣ 2 ∣ 2 d z ≤ C q 2 ∣ β ∣ + 2 ∣ μ ∣ − 2 ≤ C q 2 . |\xi|^{2|\beta|}\int_{\mathbb{R}^2}\left|D^\mu_z e^{-q^2|z|^2}\right|^2\,dz
\leq Cq^{2|\beta|+2|\mu|-2}\leq Cq^2. ∣ ξ ∣ 2∣ β ∣ ∫ R 2 D z μ e − q 2 ∣ z ∣ 2 2 d z ≤ C q 2∣ β ∣ + 2∣ μ ∣ − 2 ≤ C q 2 .
The exponent − 2 -2 − 2 comes from the change of normal variables w = q z w=qz w = q z . Summation over derivatives gives ∥ E f ∥ H 2 ≤ C ∥ f ∥ H 1 \|Ef\|_{H^2}\leq C\|f\|_{H^1} ∥ E f ∥ H 2 ≤ C ∥ f ∥ H 1 , and evaluation at z = 0 z=0 z = 0 gives Tr E f = f \operatorname{Tr}Ef=f Tr E f = f . For smooth compactly supported local data, rapid Fourier decay permits differentiation to every order. Localizing in smooth normal charts, multiplying by cutoffs equal to 1 1 1 on the trace data, and summing with a partition of unity therefore gives
Tr Σ E = Id , ∥ E f ∥ H 2 ( B ) ≤ C ∥ f ∥ H 1 ( Σ ) , E ( C ∞ ( Σ ) ) ⊂ C c ∞ ( B ) . \operatorname{Tr}_{\Sigma}E=\operatorname{Id},\qquad
\|Ef\|_{H^2(B)}\leq C\|f\|_{H^1(\Sigma)},\qquad
E(C^\infty(\Sigma))\subset C^\infty_c(B). Tr Σ E = Id , ∥ E f ∥ H 2 ( B ) ≤ C ∥ f ∥ H 1 ( Σ ) , E ( C ∞ ( Σ )) ⊂ C c ∞ ( B ) .
All chart cutoffs have compact support in their charts, so the extensions by zero create no jumps. When d = 2 d=2 d = 2 , choose disjoint smooth bumps b p b_p b p with b p ( p ) = 1 b_p(p)=1 b p ( p ) = 1 and use E f = ∑ p ∈ Σ f ( p ) b p Ef=\sum_{p\in\Sigma}f(p)b_p E f = ∑ p ∈ Σ f ( p ) b p instead.
Smooth approximation within the zero-trace kernel.
Suppose u ∈ H 0 2 ( B ) u\in H^2_0(B) u ∈ H 0 2 ( B ) and Tr Σ u = 0 \operatorname{Tr}_{\Sigma}u=0 Tr Σ u = 0 . By the definition of H 0 2 H^2_0 H 0 2 , choose w j ∈ C c ∞ ( B ) w_j\in C^\infty_c(B) w j ∈ C c ∞ ( B ) with ∥ w j − u ∥ H 2 < 1 / j \lVert w_j-u\rVert_{H^2}<1/j ∥ w j − u ∥ H 2 < 1/ j . Correct their traces by setting
v j = w j − E ( Tr Σ w j ) . v_j=w_j-E\left(\operatorname{Tr}_{\Sigma}w_j\right). v j = w j − E ( Tr Σ w j ) .
The preceding properties of E E E show that v j v_j v j is smooth and compactly supported in B B B , and that Tr Σ v j = Tr Σ w j − Tr Σ w j = 0 \operatorname{Tr}_{\Sigma}v_j=\operatorname{Tr}_{\Sigma}w_j-\operatorname{Tr}_{\Sigma}w_j=0 Tr Σ v j = Tr Σ w j − Tr Σ w j = 0 . For a smooth function this is pointwise vanishing on Σ \Sigma Σ . Furthermore,
∥ v j − u ∥ H 2 ≤ ∥ w j − u ∥ H 2 + C ∥ Tr Σ ( w j − u ) ∥ H 1 ( Σ ) ≤ C ′ ∥ w j − u ∥ H 2 ≤ C ′ / j . \lVert v_j-u\rVert_{H^2}\leq\lVert w_j-u\rVert_{H^2}+C\left\lVert\operatorname{Tr}_{\Sigma}(w_j-u)\right\rVert_{H^1(\Sigma)}
\leq C'\lVert w_j-u\rVert_{H^2}\leq C'/j. ∥ v j − u ∥ H 2 ≤ ∥ w j − u ∥ H 2 + C ∥ Tr Σ ( w j − u ) ∥ H 1 ( Σ ) ≤ C ′ ∥ w j − u ∥ H 2 ≤ C ′ / j .
We shall cut off each fixed v j v_j v j near Σ \Sigma Σ . The cutoff estimate may depend on its smooth derivatives; uniform bounds on these derivatives as j → + ∞ j\to+\infty j → + ∞ are not needed.
Normal coordinates and the logarithmic cutoff.
Fix one smooth function v ∈ C c ∞ ( B ) v\in C^\infty_c(B) v ∈ C c ∞ ( B ) vanishing on Σ \Sigma Σ , and put ρ ( x ) = dist ( x , Σ ) \rho(x)=\operatorname{dist}(x,\Sigma) ρ ( x ) = dist ( x , Σ ) . Since Σ \Sigma Σ is smooth, compact and embedded, it has a tubular neighborhood in which every point has a unique nearest point on Σ \Sigma Σ . Shrink its radius ρ 0 \rho_0 ρ 0 so that ρ 0 < 1 \rho_0<1 ρ 0 < 1 , its closure lies in B B B , and the tubular neighborhoods of distinct components are disjoint. In a smooth normal chart F ( y , z ) = p ( y ) + z 1 N 1 ( y ) + z 2 N 2 ( y ) F(y,z)=p(y)+z_1N_1(y)+z_2N_2(y) F ( y , z ) = p ( y ) + z 1 N 1 ( y ) + z 2 N 2 ( y ) , where N 1 , N 2 N_1,N_2 N 1 , N 2 form an orthonormal normal frame, ρ ( F ( y , z ) ) = ∣ z ∣ \rho(F(y,z))=\lvert z\rvert ρ ( F ( y , z )) = ∣ z ∣ . Relative to surface measure on Σ \Sigma Σ and Lebesgue measure in the two normal variables, the volume Jacobian is smooth, equals 1 1 1 at z = 0 z=0 z = 0 , and is bounded above and below by positive constants. Polar coordinates z = ρ ( cos θ , sin θ ) z=\rho(\cos\theta,\sin\theta) z = ρ ( cos θ , sin θ ) therefore make the volume measure comparable to ρ d ρ d θ d Σ \rho\,d\rho\,d\theta\,d\Sigma ρ d ρ d θ d Σ , where d Σ d\Sigma d Σ is the induced measure on the submanifold. For d = 2 d=2 d = 2 it is counting measure on the finite support.
The smooth tubular coordinates give, away from Σ \Sigma Σ ,
∣ ∇ ρ ∣ ≤ C , ∣ D 2 ρ ∣ ≤ C / ρ . \lvert\nabla\rho\rvert\leq C,\qquad\lvert D^2\rho\rvert\leq C/\rho. ∣ ∇ ρ ∣ ≤ C , ∣ D 2 ρ ∣ ≤ C / ρ .
If p p p is the nearest point to x x x , then v ( p ) = 0 v(p)=0 v ( p ) = 0 and
v ( x ) = ∫ 0 1 ∇ v ( p + t ( x − p ) ) ⋅ ( x − p ) d t . v(x)=\int_0^1\nabla v\left(p+t(x-p)\right)\cdot(x-p)\,dt. v ( x ) = ∫ 0 1 ∇ v ( p + t ( x − p ) ) ⋅ ( x − p ) d t .
Consequently, for the fixed smooth v v v ,
∣ v ( x ) ∣ ≤ C v ρ ( x ) , ∣ ∇ v ( x ) ∣ + ∣ D 2 v ( x ) ∣ ≤ C v . \lvert v(x)\rvert\leq C_v\rho(x),\qquad\lvert\nabla v(x)\rvert+\lvert D^2v(x)\rvert\leq C_v. ∣ v ( x )∣ ≤ C v ρ ( x ) , ∣ ∇ v ( x )∣ + ∣ D 2 v ( x )∣ ≤ C v .
Constants denoted by C v C_v C v may increase in the estimates below; they depend on v v v and the fixed tubular neighborhood, but not on its cutoff radius δ \delta δ . Figure 5 illustrates the cutoff.
Figure 5. The logarithmic cutoff in the normal plane (schematic). On the right, 2 log δ < log δ < 0 2\log\delta< \log\delta< 0 2 log δ < log δ < 0 for 0 < δ < 1 0 < \delta< 1 0 < δ < 1 ; the horizontal origin corresponds to ρ = 1 \rho= 1 ρ = 1 . For smooth v v v vanishing on Σ \Sigma Σ , one has v = O ( ρ ) v = O(\rho) v = O ( ρ ) , and the squared H 2 H^2 H 2 cutoff error is bounded by C / ∣ log δ ∣ C/|\log\delta| C /∣ log δ ∣ .
Choose a smooth function ψ \psi ψ with 0 ≤ ψ ≤ 1 0 \le\psi\le1 0 ≤ ψ ≤ 1 , equal to 0 0 0 on ( − ∞ , 0 ] (-\infty,0] ( − ∞ , 0 ] and to 1 1 1 on [ 1 , + ∞ ) [1,+\infty) [ 1 , + ∞ ) . For 0 < δ < min { ρ 0 , e − 1 } 0 < \delta< \min\{\rho_0,e^{-1}\} 0 < δ < min { ρ 0 , e − 1 } , put L δ = ∣ log δ ∣ L_\delta= |\log\delta| L δ = ∣ log δ ∣ and define
η δ ( x ) = ψ ( log ( ρ ( x ) / δ 2 ) L δ ) \eta_\delta(x) = \psi\left(\frac{\log(\rho(x)/\delta^2)}{L_\delta}\right) η δ ( x ) = ψ ( L δ log ( ρ ( x ) / δ 2 ) )
in the tubular neighborhood, with value 0 0 0 on Σ \Sigma Σ and value 1 1 1 outside the tubular neighborhood. Since
log ( δ 2 / δ 2 ) L δ = 0 , log ( δ / δ 2 ) L δ = 1 , \frac{\log(\delta^2/\delta^2)}{L_\delta}=0,\qquad\frac{\log(\delta/\delta^2)}{L_\delta}=1, L δ log ( δ 2 / δ 2 ) = 0 , L δ log ( δ / δ 2 ) = 1 ,
this function vanishes for ρ ≤ δ 2 \rho\le\delta^2 ρ ≤ δ 2 and equals 1 1 1 for ρ ≥ δ \rho\ge\delta ρ ≥ δ . All its derivatives of positive order vanish unless δ 2 < ρ < δ \delta^2 < \rho< \delta δ 2 < ρ < δ . It is smooth across these two transition radii because ψ \psi ψ is smooth and constant on either side of its transition interval. The possible singularity of the distance at Σ \Sigma Σ causes no difficulty, since the cutoff is identically 0 0 0 on a neighborhood of Σ \Sigma Σ .
Radial differentiation gives
d η δ d ρ = ψ ′ ( log ( ρ / δ 2 ) L δ ) ρ L δ , \frac{d\eta_\delta}{d\rho}=\frac{\psi'\left(\frac{\log(\rho/\delta^2)}{L_\delta}\right)}{\rho L_\delta}, d ρ d η δ = ρ L δ ψ ′ ( L δ l o g ( ρ / δ 2 ) ) ,
d 2 η δ d ρ 2 = ψ ′ ′ ( log ( ρ / δ 2 ) L δ ) ρ 2 L δ 2 − ψ ′ ( log ( ρ / δ 2 ) L δ ) ρ 2 L δ . \frac{d^2\eta_\delta}{d\rho^2}=\frac{\psi''\left(\frac{\log(\rho/\delta^2)}{L_\delta}\right)}{\rho^2L_\delta^2}-\frac{\psi'\left(\frac{\log(\rho/\delta^2)}{L_\delta}\right)}{\rho^2L_\delta}. d ρ 2 d 2 η δ = ρ 2 L δ 2 ψ ′′ ( L δ l o g ( ρ / δ 2 ) ) − ρ 2 L δ ψ ′ ( L δ l o g ( ρ / δ 2 ) ) .
The spatial chain rule reads
∇ η δ = d η δ d ρ ∇ ρ , D 2 η δ = d 2 η δ d ρ 2 ∇ ρ ⊗ ∇ ρ + d η δ d ρ D 2 ρ . \nabla\eta_\delta=\frac{d\eta_\delta}{d\rho}\nabla\rho,\qquad D^2\eta_\delta=\frac{d^2\eta_\delta}{d\rho^2}\nabla\rho\otimes\nabla\rho+\frac{d\eta_\delta}{d\rho}D^2\rho. ∇ η δ = d ρ d η δ ∇ ρ , D 2 η δ = d ρ 2 d 2 η δ ∇ ρ ⊗ ∇ ρ + d ρ d η δ D 2 ρ .
Using L δ ≥ 1 L_\delta\ge1 L δ ≥ 1 and the distance bounds gives
∣ ∇ η δ ∣ ≤ C ρ L δ , ∣ D 2 η δ ∣ ≤ C ρ 2 L δ ( δ 2 < ρ < δ ) . |\nabla\eta_\delta|\le\frac{C}{\rho L_\delta},\qquad|D^2\eta_\delta|\le\frac{C}{\rho^2L_\delta}\qquad(\delta^2<\rho<\delta). ∣∇ η δ ∣ ≤ ρ L δ C , ∣ D 2 η δ ∣ ≤ ρ 2 L δ C ( δ 2 < ρ < δ ) .
Step 4. The cutoff error and the two inclusions.
Set w δ = ( 1 − η δ ) v w_\delta=(1-\eta_\delta)v w δ = ( 1 − η δ ) v . This is supported where ρ ≤ δ \rho\le\delta ρ ≤ δ . The zero-order estimate follows from ∣ v ∣ ≤ C v ρ |v|\le C_v\rho ∣ v ∣ ≤ C v ρ and the normal area element:
∥ w δ ∥ 2 2 ≤ C v ∫ 0 δ ρ 2 ρ d ρ ≤ C v δ 4 . \|w_\delta\|_2^2\le C_v\int_0^\delta\rho^2\rho\,\mathrm{d}\rho\le C_v\delta^4. ∥ w δ ∥ 2 2 ≤ C v ∫ 0 δ ρ 2 ρ d ρ ≤ C v δ 4 .
For its gradient, the product rule gives
∇ w δ = ( 1 − η δ ) ∇ v − v ∇ η δ . \nabla w_\delta=(1-\eta_\delta)\nabla v-v\nabla\eta_\delta. ∇ w δ = ( 1 − η δ ) ∇ v − v ∇ η δ .
The first term is bounded on a region with volume at most C δ 2 C\delta^2 C δ 2 . For the second term, the factor ρ \rho ρ in v v v cancels the factor ρ − 1 \rho^{-1} ρ − 1 in the cutoff derivative. Hence
∥ ( 1 − η δ ) ∇ v ∥ 2 2 ≤ C v ∫ 0 δ ρ d ρ ≤ C v δ 2 , \|(1-\eta_\delta)\nabla v\|_2^2\le C_v\int_0^\delta\rho\,\mathrm{d}\rho\le C_v\delta^2, ∥ ( 1 − η δ ) ∇ v ∥ 2 2 ≤ C v ∫ 0 δ ρ d ρ ≤ C v δ 2 ,
∥ v ∇ η δ ∥ 2 2 ≤ C v L δ 2 ∫ δ 2 δ ρ d ρ ≤ C v δ 2 L δ 2 . \|v\nabla\eta_\delta\|_2^2\le\frac{C_v}{L_\delta^2}\int_{\delta^2}^{\delta}\rho\,\mathrm{d}\rho\le\frac{C_v\delta^2}{L_\delta^2}. ∥ v ∇ η δ ∥ 2 2 ≤ L δ 2 C v ∫ δ 2 δ ρ d ρ ≤ L δ 2 C v δ 2 .
The inequality ∣ A + B ∣ 2 ≤ 2 ∣ A ∣ 2 + 2 ∣ B ∣ 2 |A+B|^2\le2|A|^2+2|B|^2 ∣ A + B ∣ 2 ≤ 2∣ A ∣ 2 + 2∣ B ∣ 2 now bounds ∥ ∇ w δ ∥ 2 2 \|\nabla w_\delta\|_2^2 ∥∇ w δ ∥ 2 2 by C v δ 2 ( 1 + L δ − 2 ) C_v\delta^2(1+L_\delta^{-2}) C v δ 2 ( 1 + L δ − 2 ) .
For the Hessian,
D 2 w δ = ( 1 − η δ ) D 2 v − ∇ η δ ⊗ ∇ v − ∇ v ⊗ ∇ η δ − v D 2 η δ . D^2w_\delta=(1-\eta_\delta)D^2v-\nabla\eta_\delta\otimes\nabla v-\nabla v\otimes\nabla\eta_\delta-vD^2\eta_\delta. D 2 w δ = ( 1 − η δ ) D 2 v − ∇ η δ ⊗ ∇ v − ∇ v ⊗ ∇ η δ − v D 2 η δ .
The first term again has squared norm at most C v δ 2 C_v\delta^2 C v δ 2 . Each of the two mixed terms is bounded in absolute value by C v / ( ρ L δ ) C_v/(\rho L_\delta) C v / ( ρ L δ ) . The last term has the same bound, since ∣ v ∣ ≤ C v ρ |v|\le C_v\rho ∣ v ∣ ≤ C v ρ . All three are supported in the transition annulus, so their squared norms are at most
C v L δ 2 ∫ δ 2 δ ρ d ρ ρ 2 = C v L δ 2 ( log δ − log δ 2 ) = C v L δ . \frac{C_v}{L_\delta^2}\int_{\delta^2}^{\delta}\frac{\rho\,\mathrm{d}\rho}{\rho^2}
=\frac{C_v}{L_\delta^2}(\log\delta-\log\delta^2)
=\frac{C_v}{L_\delta}. L δ 2 C v ∫ δ 2 δ ρ 2 ρ d ρ = L δ 2 C v ( log δ − log δ 2 ) = L δ C v .
After applying ∣ A 1 + ⋯ + A 4 ∣ 2 ≤ 4 ∑ i ∣ A i ∣ 2 |A_1+\cdots+A_4|^2\le4\sum_i|A_i|^2 ∣ A 1 + ⋯ + A 4 ∣ 2 ≤ 4 ∑ i ∣ A i ∣ 2 , we obtain
∥ D 2 w δ ∥ 2 2 ≤ C v ( δ 2 + L δ − 1 ) . \|D^2w_\delta\|_2^2\le C_v\left(\delta^2+L_\delta^{-1}\right). ∥ D 2 w δ ∥ 2 2 ≤ C v ( δ 2 + L δ − 1 ) .
Combining all three derivative orders and using δ < 1 \delta<1 δ < 1 and L δ ≥ 1 L_\delta\ge1 L δ ≥ 1 proves
∥ ( 1 − η δ ) v ∥ H 2 2 ≤ C v ( δ 2 + ∣ log δ ∣ − 1 ) ⟶ 0. \|(1-\eta_\delta)v\|_{H^2}^2\le C_v\left(\delta^2+|\log\delta|^{-1}\right)\longrightarrow0. ∥ ( 1 − η δ ) v ∥ H 2 2 ≤ C v ( δ 2 + ∣ log δ ∣ − 1 ) ⟶ 0.
For a finite union of components, disjoint tubular neighborhoods permit the same construction on each component. The corresponding integrals add, and their number is fixed throughout this limit.
Return to the sequence v j v_j v j . For each j j j , choose δ j < min { ρ 0 , e − 1 , 1 / j } \delta_j<\min\{\rho_0,e^{-1},1/j\} δ j < min { ρ 0 , e − 1 , 1/ j } sufficiently small that
∥ ( 1 − η δ j ) v j ∥ H 2 < 1 / j . \|(1-\eta_{\delta_j})v_j\|_{H^2}<1/j. ∥ ( 1 − η δ j ) v j ∥ H 2 < 1/ j .
This is possible separately for every j j j , even if C v j C_{v_j} C v j grows. We may also impose δ j < δ j − 1 / 2 \delta_j < \delta_{j-1}/2 δ j < δ j − 1 /2 for j > 1 j > 1 j > 1 . The function η δ j v j \eta_{\delta_j}v_j η δ j v j is smooth, has compact support in B B B , and vanishes on the open neighborhood { ρ < δ j 2 } \{\rho< \delta_j^2\} { ρ < δ j 2 } of Σ \Sigma Σ . It therefore belongs to C c ∞ ( B ∖ Σ ) C^\infty_c(B \setminus\Sigma) C c ∞ ( B ∖ Σ ) , and
∥ η δ j v j − u ∥ H 2 ≤ ∥ ( 1 − η δ j ) v j ∥ H 2 + ∥ v j − u ∥ H 2 ≤ ( 1 + C ′ ) / j ⟶ 0. \|\eta_{\delta_j}v_j-u\|_{H^2}\leq\|(1-\eta_{\delta_j})v_j\|_{H^2}+\|v_j-u\|_{H^2}\leq(1+C')/j\longrightarrow0. ∥ η δ j v j − u ∥ H 2 ≤ ∥ ( 1 − η δ j ) v j ∥ H 2 + ∥ v j − u ∥ H 2 ≤ ( 1 + C ′ ) / j ⟶ 0.
This proves the inclusion from right to left in (5.1).
Finally, let u ∈ H 0 2 ( B ∖ Σ ) u\in H^2_0(B\setminus\Sigma) u ∈ H 0 2 ( B ∖ Σ ) . By its definition there are ϕ j ∈ C c ∞ ( B ∖ Σ ) \phi_j\in C^\infty_c(B\setminus\Sigma) ϕ j ∈ C c ∞ ( B ∖ Σ ) converging to u u u in that Sobolev space. Each ϕ j \phi_j ϕ j vanishes on a neighborhood of Σ \Sigma Σ , so its zero extension to B B B is smooth. These extensions have the same H 2 H^2 H 2 norms, since Σ \Sigma Σ has volume zero. Thus they converge in H 2 ( B ) H^2(B) H 2 ( B ) to the zero extension of u u u , which belongs to H 0 2 ( B ) H^2_0(B) H 0 2 ( B ) . Their traces vanish, and continuity of Tr Σ \operatorname{Tr}_{\Sigma} Tr Σ gives Tr Σ u = 0 \operatorname{Tr}_{\Sigma}u=0 Tr Σ u = 0 . This proves the opposite inclusion and completes the space identity.
Tubular holes and their spectrum
We use the tubular-hole geometry illustrated after Theorem 1.1 in Figure 1 . For a fixed finite family of the latitudes introduced in Section 4, put R j = r 2 − a j 2 > 0 R_j=\sqrt{r^2-a_j^2}>0 R j = r 2 − a j 2 > 0 and Σ = ⋃ j = 1 N M j \Sigma=\bigcup_{j=1}^{N}M_j Σ = ⋃ j = 1 N M j . We use the tubular coordinates of [14 Section 6, Step 3] , with the notation below. The case N = 1 N=1 N = 1 omits all minima over distinct indices. Choose
Figure 1. Three-dimensional geometry of the holes. The outer sphere is drawn transparently in (a); the shaded solid tori are removed, not part of the domain. The remaining region is Ω N , η \Omega_{N,\eta} Ω N , η . Panel (b) shows a plane through the vertical axis: each torus meets it in two disks. The dashed circle is the interior sphere supporting the latitude circles. The two-hole configuration shows the shape and separation of the holes. In the construction, N N N is chosen first and η \eta η is then taken sufficiently small.
0 < τ < 1 4 min { 1 − r , R 1 , … , R N , ∣ a i − a j ∣ : i ≠ j } . (18) 0<\tau<\frac{1}{4}\min\{1-r,R_1,\ldots,R_N,\lvert a_i-a_j\rvert:i\ne j\}. \tag*{(18)} 0 < τ < 4 1 min { 1 − r , R 1 , … , R N , ∣ a i − a j ∣ : i = j } . (18)
For each latitude the normal coordinates are
F j ( ω , s , t ) = ( ( R j + s ) ω , a j + t ) , ω ∈ S d − 2 , s 2 + t 2 < τ 2 . (19) F_j(\omega,s,t)=((R_j+s)\omega,a_j+t),\qquad\omega\in\mathbb{S}^{d-2},\qquad s^2+t^2<\tau^2. \tag*{(19)} F j ( ω , s , t ) = (( R j + s ) ω , a j + t ) , ω ∈ S d − 2 , s 2 + t 2 < τ 2 . (19)
These maps parametrize disjoint tubular neighborhoods contained in B B B , and dist ( F j ( ω , s , t ) , M j ) = s 2 + t 2 \operatorname{dist}(F_j(\omega,s,t),M_j)=\sqrt{s^2+t^2} dist ( F j ( ω , s , t ) , M j ) = s 2 + t 2 ; see [14 Section 6, Steps 3 and 6] .
For 0 < η < τ 0<\eta<\tau 0 < η < τ define
K j , η = { x : dist ( x , M j ) ≤ η } , Ω N , η = B ∖ ⋃ j = 1 N K j , η . K_{j,\eta}=\{x:\operatorname{dist}(x,M_j)\leq\eta\},\qquad\Omega_{N,\eta}=B\setminus\bigcup_{j=1}^{N}K_{j,\eta}. K j , η = { x : dist ( x , M j ) ≤ η } , Ω N , η = B ∖ j = 1 ⋃ N K j , η .
The normal-coordinate description identifies each hole with S d − 2 × { ( s , t ) : s 2 + t 2 ≤ η 2 } \mathbb{S}^{d-2}\times\{(s,t):s^2+t^2\leq\eta^2\} S d − 2 × {( s , t ) : s 2 + t 2 ≤ η 2 } . The smoothness argument in [14 Section 6, Step 6] therefore applies: the inner boundaries are the regular level sets s 2 + t 2 = η 2 s^2+t^2=\eta^2 s 2 + t 2 = η 2 , and all boundary components are disjoint. For d = 2 d=2 d = 2 , each latitude consists of two points and its two disk neighborhoods are disjoint because η < τ < R j / 4 \eta<\tau<R_j/4 η < τ < R j /4 .
The following lemma proves spectral convergence for the vibrating plate and records the connectedness supplied by the same geometric argument.
Lemma 5.3. For every fixed N N N and k k k , the ordered eigenvalues satisfy
Γ k ( Ω N , η ) ↓ λ k ( V N ) as η ↓ 0. (20) \Gamma_k(\Omega_{N,\eta})\downarrow\lambda_k(V_N)\quad\text{as }\eta\downarrow0. \tag*{(20)} Γ k ( Ω N , η ) ↓ λ k ( V N ) as η ↓ 0. (20) Every Ω N , η \Omega_{N,\eta} Ω N , η with 0 < η < τ 0<\eta<\tau 0 < η < τ is connected.
Proof. Step 1. Approximation of the trial space.
Throughout this argument, N N N and the supports are fixed. By Lemma 5.1 , applied to their finite disjoint union,
H 0 2 ( B ∖ Σ ) = V N . H_0^2(B\setminus\Sigma)=V_N. H 0 2 ( B ∖ Σ ) = V N .
For 0 < η 1 < η 2 < τ 0<\eta_1<\eta_2<\tau 0 < η 1 < η 2 < τ , zero extension and domain inclusion therefore give
H 0 2 ( Ω N , η 2 ) ⊂ H 0 2 ( Ω N , η 1 ) ⊂ V N . H_0^2(\Omega_{N,\eta_2})\subset H_0^2(\Omega_{N,\eta_1})\subset V_N. H 0 2 ( Ω N , η 2 ) ⊂ H 0 2 ( Ω N , η 1 ) ⊂ V N .
The corresponding min–max inequalities are
λ k ( V N ) ≤ Γ k ( Ω N , η 1 ) ≤ Γ k ( Ω N , η 2 ) . \lambda_k(V_N)\le\Gamma_k(\Omega_{N,\eta_1})\le\Gamma_k(\Omega_{N,\eta_2}). λ k ( V N ) ≤ Γ k ( Ω N , η 1 ) ≤ Γ k ( Ω N , η 2 ) .
Thus the eigenvalues have a decreasing limit as η ↓ 0 \eta\downarrow0 η ↓ 0 , and this limit is at least λ k ( V N ) \lambda_k(V_N) λ k ( V N ) . For the opposite inequality, choose real L 2 L^2 L 2 -orthonormal eigenvectors v 1 , … , v k v_1,\ldots,v_k v 1 , … , v k corresponding to the first k k k eigenvalues on V N V_N V N . Given δ > 0 \delta>0 δ > 0 , the definition of H 0 2 ( B ∖ Σ ) H_0^2(B\setminus\Sigma) H 0 2 ( B ∖ Σ ) and the displayed space identity provide ϕ i ∈ C c ∞ ( B ∖ Σ ) \phi_i\in C_c^\infty(B\setminus\Sigma) ϕ i ∈ C c ∞ ( B ∖ Σ ) such that
∥ ϕ i − v i ∥ 2 2 + ∥ Δ ( ϕ i − v i ) ∥ 2 2 < δ 2 , 1 ≤ i ≤ k . \|\phi_i-v_i\|_2^2+\|\Delta(\phi_i-v_i)\|_2^2<\delta^2,\qquad1\le i\le k. ∥ ϕ i − v i ∥ 2 2 + ∥Δ ( ϕ i − v i ) ∥ 2 2 < δ 2 , 1 ≤ i ≤ k .
We need an estimate on every linear combination of these functions. Put e i = ϕ i − v i e_i=\phi_i-v_i e i = ϕ i − v i . For c ∈ R k c\in\mathbb{R}^k c ∈ R k with Euclidean norm ∣ c ∣ = 1 |c|=1 ∣ c ∣ = 1 , Cauchy–Schwarz in the product Hilbert space L 2 ( B ) × L 2 ( B ) L^2(B)\times L^2(B) L 2 ( B ) × L 2 ( B ) gives
∥ ∑ i c i e i ∥ 2 2 + ∥ Δ ∑ i c i e i ∥ 2 2 ≤ ( ∑ i ∣ c i ∣ 2 ) ∑ i { ∥ e i ∥ 2 2 + ∥ Δ e i ∥ 2 2 } < k δ 2 . \left\|\sum_i c_i e_i\right\|_2^2+\left\|\Delta\sum_i c_i e_i\right\|_2^2\le\left(\sum_i|c_i|^2\right)\sum_i\{\|e_i\|_2^2+\|\Delta e_i\|_2^2\}<k\delta^2. i ∑ c i e i 2 2 + Δ i ∑ c i e i 2 2 ≤ ( i ∑ ∣ c i ∣ 2 ) i ∑ { ∥ e i ∥ 2 2 + ∥Δ e i ∥ 2 2 } < k δ 2 .
On the exact spectral space, orthonormality and the eigenvalue equations give
∥ ∑ i c i v i ∥ 2 2 = 1 , ∥ Δ ∑ i c i v i ∥ 2 2 = ∑ i c i 2 λ i ( V N ) ≤ λ k ( V N ) . \left\|\sum_i c_i v_i\right\|_2^2=1,\qquad\left\|\Delta\sum_i c_i v_i\right\|_2^2=\sum_i c_i^2\lambda_i(V_N)\le\lambda_k(V_N). i ∑ c i v i 2 2 = 1 , Δ i ∑ c i v i 2 2 = i ∑ c i 2 λ i ( V N ) ≤ λ k ( V N ) .
Apply the reverse triangle inequality to the mass norm and the triangle inequality to the energy norm. This gives, for every ∣ c ∣ = 1 |c|=1 ∣ c ∣ = 1 ,
∥ ∑ i c i ϕ i ∥ 2 ≥ 1 − k δ , ∥ Δ ∑ i c i ϕ i ∥ 2 ≤ λ k ( V N ) + k δ . \left\|\sum_i c_i\phi_i\right\|_2 \ge1-\sqrt{k}\delta,\qquad\left\|\Delta\sum_i c_i\phi_i\right\|_2 \le\sqrt{\lambda_k(V_N)}+\sqrt{k}\delta. i ∑ c i ϕ i 2 ≥ 1 − k δ , Δ i ∑ c i ϕ i 2 ≤ λ k ( V N ) + k δ .
Choose δ \delta δ so that k δ < 1 \sqrt{k}\delta<1 k δ < 1 . By homogeneity, the first estimate for arbitrary c c c has right side ( 1 − k δ ) ∣ c ∣ \left(1-\sqrt{k}\delta\right)|c| ( 1 − k δ ) ∣ c ∣ . Consequently no nonzero coefficient vector can represent the zero function: the span E δ E_\delta E δ of the ϕ i \phi_i ϕ i has dimension k k k . The two bounds yield the uniform quotient estimate
max 0 ≠ ϕ ∈ E δ ∥ Δ ϕ ∥ 2 2 ∥ ϕ ∥ 2 2 ≤ ( λ k ( V N ) + k δ ) 2 ( 1 − k δ ) 2 . \max_{0\ne\phi\in E_\delta}\frac{\|\Delta\phi\|_2^2}{\|\phi\|_2^2}\le\frac{\left(\sqrt{\lambda_k(V_N)}+\sqrt{k}\delta\right)^2}{\left(1-\sqrt{k}\delta\right)^2}. 0 = ϕ ∈ E δ max ∥ ϕ ∥ 2 2 ∥Δ ϕ ∥ 2 2 ≤ ( 1 − k δ ) 2 ( λ k ( V N ) + k δ ) 2 .
For this fixed choice of δ \delta δ and of the approximating functions, let K = ⋃ i = 1 k supp ϕ i K=\bigcup_{i=1}^k\operatorname{supp}\phi_i K = ⋃ i = 1 k supp ϕ i . It is a compact subset of B ∖ Σ B\setminus\Sigma B ∖ Σ . Therefore both dist ( K , Σ ) \operatorname{dist}(K,\Sigma) dist ( K , Σ ) and dist ( K , ∂ B ) \operatorname{dist}(K,\partial B) dist ( K , ∂ B ) are positive. If
0 < η < min { τ , dist ( K , Σ ) } , 0<\eta<\min\{\tau,\operatorname{dist}(K,\Sigma)\}, 0 < η < min { τ , dist ( K , Σ )} ,
then K K K is compactly contained in Ω N , η \Omega_{N,\eta} Ω N , η , and every vector of E δ E_\delta E δ is an admissible trial function there. Min–max consequently gives, for all such η \eta η ,
Γ k ( Ω N , η ) ≤ ( λ k ( V N ) + k δ ) 2 ( 1 − k δ ) 2 . \Gamma_k(\Omega_{N,\eta})\le\frac{\left(\sqrt{\lambda_k(V_N)}+\sqrt{k}\delta\right)^2}{\left(1-\sqrt{k}\delta\right)^2}. Γ k ( Ω N , η ) ≤ ( 1 − k δ ) 2 ( λ k ( V N ) + k δ ) 2 .
First let η ↓ 0 \eta\downarrow0 η ↓ 0 with δ \delta δ fixed, and then let δ ↓ 0 \delta\downarrow0 δ ↓ 0 . The right side tends to λ k ( V N ) \lambda_k(V_N) λ k ( V N ) . Combined with the lower bound, this proves (20) . The functions used in the trial space may depend on δ \delta δ ; their common admissible range of η \eta η is chosen only after they have been fixed.
Step 2. Connectedness.
Apply the path construction in [14 Section 6, Step 6] to the finite codimension-two union Σ \Sigma Σ . Its hypotheses hold because the τ \tau τ -neighborhoods are disjoint and contained in B B B . For any 0 < η < τ 0<\eta<\tau 0 < η < τ , choose η < η ˉ < τ \eta<\bar{\eta}<\tau η < η ˉ < τ and use η ˉ \bar{\eta} η ˉ as the target normal radius in that construction (in place of 2 ε 2\varepsilon 2 ε ). The argument gives a path in Ω N , η \Omega_{N,\eta} Ω N , η between any two of its points. Thus Ω N , η \Omega_{N,\eta} Ω N , η is connected for every 0 < η < τ 0<\eta<\tau 0 < η < τ . □
Proof of Theorem 1.1
The spherical constraint gives a strict spectral gap. The two approximation results allow that gap to be retained with first finitely many supports and then positive hole thickness. We now make these choices explicitly in terms of the gaps they must preserve.
Proof. Step 1. Fix the spherical gap.
Fix d ≥ 2 d \ge2 d ≥ 2 and choose r r r by Proposition 3.2 . Since μ 1 ( r ) > 0 \mu_1(r)>0 μ 1 ( r ) > 0 , the strict ratio inequality is equivalent to
g = μ 2 ( r ) − ρ d μ 1 ( r ) = μ 1 ( r ) { μ 2 ( r ) μ 1 ( r ) − ρ d } > 0. g=\mu_2(r)-\rho_d\mu_1(r)=\mu_1(r)\left\{\frac{\mu_2(r)}{\mu_1(r)}-\rho_d\right\}>0. g = μ 2 ( r ) − ρ d μ 1 ( r ) = μ 1 ( r ) { μ 1 ( r ) μ 2 ( r ) − ρ d } > 0.
Step 2. Choose finitely many latitudes.
Keep r r r and this number g g g fixed. By (15) , there is a finite N N N such that
0 ≤ μ 2 ( r ) − λ 2 ( V N ) < g 2 . 0\le\mu_2(r)-\lambda_2(V_N)<\frac{g}{2}. 0 ≤ μ 2 ( r ) − λ 2 ( V N ) < 2 g .
The same space inclusion gives λ 1 ( V N ) ≤ μ 1 ( r ) \lambda_1(V_N)\le\mu_1(r) λ 1 ( V N ) ≤ μ 1 ( r ) without any further choice of N N N . Therefore the finite supports retain the strictly positive gap
g N = λ 2 ( V N ) − ρ d λ 1 ( V N ) > μ 2 ( r ) − g 2 − ρ d μ 1 ( r ) = g 2 > 0. \begin{aligned}
g_N=\lambda_2(V_N)-\rho_d\lambda_1(V_N) \\
&>\mu_2(r)-\frac{g}{2}-\rho_d\mu_1(r)=\frac{g}{2}>0.
\end{aligned} g N = λ 2 ( V N ) − ρ d λ 1 ( V N ) > μ 2 ( r ) − 2 g − ρ d μ 1 ( r ) = 2 g > 0.
In particular, it is enough to approximate the second eigenvalue at this stage: the first eigenvalue is already bounded in the direction needed for this lower estimate.
Step 3. Choose the hole radius.
Fix this N N N , choose τ \tau τ according to (18) , and retain the resulting family of tubular neighborhoods. Since ρ d > 0 \rho_d>0 ρ d > 0 , the positive quantity g N / ( 2 ρ d ) g_N/(2\rho_d) g N / ( 2 ρ d ) is an admissible error tolerance. Applying (20) with k = 1 k=1 k = 1 , choose 0 < η < τ 0<\eta<\tau 0 < η < τ so small that
0 ≤ Γ 1 ( Ω N , η ) − λ 1 ( V N ) < g N 2 ρ d . 0\le\Gamma_1(\Omega_{N,\eta})-\lambda_1(V_N)<\frac{g_N}{2\rho_d}. 0 ≤ Γ 1 ( Ω N , η ) − λ 1 ( V N ) < 2 ρ d g N .
For the second eigenvalue, inclusion already gives Γ 2 ( Ω N , η ) ≥ λ 2 ( V N ) \Gamma_2(\Omega_{N,\eta})\ge\lambda_2(V_N) Γ 2 ( Ω N , η ) ≥ λ 2 ( V N ) . Combining these two one-sided estimates yields
Γ 2 ( Ω N , η ) − ρ d Γ 1 ( Ω N , η ) > λ 2 ( V N ) − ρ d { λ 1 ( V N ) + g N 2 ρ d } = g N 2 > 0. \begin{aligned}
\Gamma_2(\Omega_{N,\eta})-\rho_d\Gamma_1(\Omega_{N,\eta}) \\
&>\lambda_2(V_N)-\rho_d\left\{\lambda_1(V_N)+\frac{g_N}{2\rho_d}\right\}=\frac{g_N}{2}>0.
\end{aligned} Γ 2 ( Ω N , η ) − ρ d Γ 1 ( Ω N , η ) > λ 2 ( V N ) − ρ d { λ 1 ( V N ) + 2 ρ d g N } = 2 g N > 0.
The first clamped eigenvalue of Ω N , η \Omega_{N,\eta} Ω N , η is positive, so division gives
R ( Ω N , η ) − ρ d > g N 2 Γ 1 ( Ω N , η ) > 0. R(\Omega_{N,\eta})-\rho_d>\frac{g_N}{2\Gamma_1(\Omega_{N,\eta})}>0. R ( Ω N , η ) − ρ d > 2 Γ 1 ( Ω N , η ) g N > 0.
Step 4. Verify the geometric conclusions.
The sets removed from B B B are precisely the finitely many closed tubular neighborhoods K j , η K_{j,\eta} K j , η of the embedded S d − 2 \mathbb{S}^{d-2} S d − 2 supports. Their boundary smoothness and separation were verified using (5.4), and Lemma 5.3 proves connectedness of the complement. For d = 2 d=2 d = 2 these neighborhoods are pairs of disks; for d = 3 d=3 d = 3 they are solid tori. Thus Ω = Ω N , η \Omega=\Omega_{N,\eta} Ω = Ω N , η has every asserted geometric and spectral property. The choices were made in the order d d d , r r r , finite N N N , and finally η \eta η ; the second approximation was taken with the supports fixed. This completes the proof of Theorem 1.1 .
The spectral continuity recalled in § 2 also shows that the strict comparison persists under sufficiently small smooth changes of each constructed domain. Indeed, write g = Γ 2 ( Ω ) − ρ d Γ 1 ( Ω ) > 0 g = \Gamma_2(\Omega) - \rho_d\Gamma_1(\Omega) > 0 g = Γ 2 ( Ω ) − ρ d Γ 1 ( Ω ) > 0 . For a C 2 C^2 C 2 diffeomorphism T T T sufficiently close to the identity,
∣ Γ 2 ( T Ω ) − Γ 2 ( Ω ) ∣ + ρ d ∣ Γ 1 ( T Ω ) − Γ 1 ( Ω ) ∣ < g 2 . \left|\Gamma_2(T\Omega)-\Gamma_2(\Omega)\right|+\rho_d\left|\Gamma_1(T\Omega)-\Gamma_1(\Omega)\right|<\frac{g}{2}. ∣ Γ 2 ( T Ω ) − Γ 2 ( Ω ) ∣ + ρ d ∣ Γ 1 ( T Ω ) − Γ 1 ( Ω ) ∣ < 2 g .
Subtracting these possible errors from g g g gives Γ 2 ( T Ω ) − ρ d Γ 1 ( T Ω ) > g / 2 \Gamma_2(T\Omega)-\rho_d\Gamma_1(T\Omega)>g/2 Γ 2 ( T Ω ) − ρ d Γ 1 ( T Ω ) > g /2 , so the ratio remains strictly greater than ρ d \rho_d ρ d .
Data availability. All quantities used in the proofs are specified in this article.
Conflict of interest. The authors declare that they have no conflicts of interest.
Use of AI tools. The authors used the 6.1 Sol models in ChatGPT Pro (OpenAI) as an auxiliary tool to generate the figures, to help identify illustrative and computational examples, and to assist with some calculations. The overall research ideas, mathematical methods, proofs, and interpretation are the authors’ own. All AI-assisted outputs, including calculations and examples, were independently verified by the authors, who take full responsibility for the content of the manuscript.