Introduction
The ratio of eigenvalues characterizes the separation of the low part of the spectrum and is independent of the size of the domain. For a vibrating membrane, the sharp inequality of Payne–Pólya–Weinberger shows that the ball maximizes the ratio of the first two eigenvalues [7 ] . The fourth order buckling problem admits a parallel conclusion.
It connects geometric eigenvalue inequalities with elastic stability and spectral shape optimization, and reveals how the extremal geometry depends on the differential operator and on the clamped boundary conditions. The present paper investigates this question after interior constraints, smooth perforation, and a change of topology.
Let Ω ⊂ R d \Omega\subset\mathbb{R}^{d} Ω ⊂ R d be a bounded connected open set with smooth boundary, where R d \mathbb{R}^{d} R d is d d d -dimensional Euclidean space. Write Δ = ∑ i = 1 d ∂ x i 2 \Delta= \sum_{i=1}^{d} \partial_{x_i}^{2} Δ = ∑ i = 1 d ∂ x i 2 for the Euclidean Laplacian, and let λ k D ( Ω ) \lambda_k^{D}(\Omega) λ k D ( Ω ) be the eigenvalues of the Dirichlet problem
− Δ v = λ v in Ω , v = 0 on ∂ Ω , -\Delta v = \lambda v \quad\text{in } \Omega, \qquad v = 0 \quad\text{on } \partial\Omega, − Δ v = λ v in Ω , v = 0 on ∂ Ω ,
arranged in nondecreasing order and counted with multiplicity. In the planar case, the Payne–Pólya–Weinberger conjecture [19 ] asserts that the disk maximizes the ratio of the first two eigenvalues. Its higher dimensional analogue is
λ 2 D ( Ω ) λ 1 D ( Ω ) ≤ λ 2 D ( B ) λ 1 D ( B ) = j d / 2 , 1 2 j d / 2 − 1 , 1 2 . \frac{\lambda_2^{D}(\Omega)}{\lambda_1^{D}(\Omega)} \leq\frac{\lambda_2^{D}(B)}{\lambda_1^{D}(B)} = \frac{j_{d/2,1}^{2}}{j_{d/2-1,1}^{2}}. λ 1 D ( Ω ) λ 2 D ( Ω ) ≤ λ 1 D ( B ) λ 2 D ( B ) = j d /2 − 1 , 1 2 j d /2 , 1 2 .
Here B = B 1 ( 0 ) B = B_1(0) B = B 1 ( 0 ) is the open unit ball, J α J_{\alpha} J α is the Bessel function of the first kind of order α \alpha α , and j α , k j_{\alpha,k} j α , k is its k k k -th positive zero. This planar conjecture was also included by Yau in his problem collection [24 ] as Problem 77. Ashbaugh and Benguria [7 ] proved this sharp inequality in every dimension, with equality only for the ball. This Euclidean Dirichlet theorem provides the prototype for the buckling ratio conjecture.
For a membrane with fixed boundary, the vibrational frequencies are proportional to the square roots of the Dirichlet eigenvalues, so the ratio of eigenvalues is the square of the ratio of the first two frequencies. Kac’s question “Can one hear the shape of a drum?” [16 ] concerns the geometric information contained in the entire Dirichlet spectrum, whereas the PPW problem gives a sharp constraint on two quantities in the spectrum.
For a clamped elastic plate, the corresponding stability problem is
Δ 2 u = − Λ Δ u in Ω , u = ∂ ν u = 0 on ∂ Ω , \Delta^{2}u = -\Lambda\Delta u \quad\text{in } \Omega, \qquad u = \partial_{\nu}u = 0 \quad\text{on } \partial\Omega, Δ 2 u = − ΛΔ u in Ω , u = ∂ ν u = 0 on ∂ Ω ,
where Δ 2 = Δ ∘ Δ \Delta^{2} = \Delta\circ\Delta Δ 2 = Δ ∘ Δ is the biharmonic operator, ν \nu ν is the outward unit normal, and ∂ ν u \partial_{\nu}u ∂ ν u is the normal derivative. Denote its positive eigenvalues by Λ 1 ( Ω ) ≤ Λ 2 ( Ω ) ≤ ⋯ \Lambda_1(\Omega) \leq\Lambda_2(\Omega) \leq\cdots Λ 1 ( Ω ) ≤ Λ 2 ( Ω ) ≤ ⋯ , counted with multiplicity. In the planar thin plate model, u u u is the transverse displacement and Λ 1 \Lambda_1 Λ 1 is proportional to the critical uniform compressive load at which the plate loses stability, see [11 , 20 ] . Higher eigenvalues correspond to higher buckling modes of the linearized problem, and the separation of these modal thresholds is measured by their ratios, which are independent of the size of the plate. This ratio is not the same as the ratio of vibrational frequencies. The vibrations of a clamped plate satisfy Δ 2 u = Γ u \Delta^{2}u = \Gamma u Δ 2 u = Γ u , where Γ \Gamma Γ is the vibrational eigenvalue, with the same boundary conditions.
In the original planar problem the domain used for comparison is the disk. Rayleigh’s conjecture [21 p. 382] states that among planar domains of given area the disk minimizes the first clamped vibrational eigenvalue. The citation refers to the second edition of Volume I, see also [5 Section 2.4] . Pólya and Szegő [20 Note F, p. 231] proposed the corresponding fixed-area minimization problem for the first buckling load. Both original statements concern only d = 2 d=2 d = 2 . Comparison with the ball in higher dimensions is an extension rather than the original formulation, see [11 p. 383] . Talenti’s rearrangement method [22 ] , together with the subsequent work of Nadirashvili [18 ] and of Ashbaugh and Benguria [8 ] , eventually settled Rayleigh’s conjecture and its three-dimensional extension. For the buckling problem, Ashbaugh and Laugesen [11 ] obtained dimension-dependent lower bounds, in which the comparison factor tends to 1 as the dimension tends to infinity. However, the comparison of the first eigenvalue and the ratio problem require different arguments, see also the survey [9 ] .
The buckling problem on the ball can be solved by separation of variables, which yields
Λ 1 ( B ) = j d / 2 , 1 2 and Λ 2 ( B ) = j d / 2 + 1 , 1 2 , \Lambda_1(B)=j_{d/2,1}^{2}\quad\text{and}\quad\Lambda_2(B)=j_{d/2+1,1}^{2}, Λ 1 ( B ) = j d /2 , 1 2 and Λ 2 ( B ) = j d /2 + 1 , 1 2 ,
see [10 Section 11.3] . Since Λ k ( t Ω ) = t − 2 Λ k ( Ω ) \Lambda_k(t\Omega)=t^{-2}\Lambda_k(\Omega) Λ k ( t Ω ) = t − 2 Λ k ( Ω ) holds for t > 0 t>0 t > 0 , this ratio is invariant under scaling. In 1999, Ashbaugh [5 Section 4, equation (4.8)] formulated the following buckling analogue of the PPW conjecture:
Λ 2 ( Ω ) Λ 1 ( Ω ) ≤ C d , C d : = j d / 2 + 1 , 1 2 j d / 2 , 1 2 . (1) \frac{\Lambda_2(\Omega)}{\Lambda_1(\Omega)}\leq C_d,\qquad C_d:=\frac{j_{d/2+1,1}^{2}}{j_{d/2,1}^{2}}. \tag*{(1)} Λ 1 ( Ω ) Λ 2 ( Ω ) ≤ C d , C d := j d /2 , 1 2 j d /2 + 1 , 1 2 . (1)
This also appears in [10 Remark 8.10] . Payne, Pólya and Weinberger earlier established the planar bound Λ 2 / Λ 1 < 3 \Lambda_2/\Lambda_1<3 Λ 2 / Λ 1 < 3 [19 ] . Hile and Yeh [14 ] improved and extended this bound to
Λ 2 ( Ω ) Λ 1 ( Ω ) ≤ d 2 + 8 d + 20 ( d + 2 ) 2 . \frac{\Lambda_2(\Omega)}{\Lambda_1(\Omega)}\leq\frac{d^2+8d+20}{(d+2)^2}. Λ 1 ( Ω ) Λ 2 ( Ω ) ≤ ( d + 2 ) 2 d 2 + 8 d + 20 .
Later work has focused on stronger conclusions for several buckling eigenvalues. Ashbaugh [5 , 6 ] proved that
∑ i = 1 d Λ i + 1 ≤ ( d + 4 ) Λ 1 , \sum_{i=1}^{d}\Lambda_{i+1}\leq(d+4)\Lambda_1, i = 1 ∑ d Λ i + 1 ≤ ( d + 4 ) Λ 1 ,
and Jost et al. [15 Theorem 4.2] strengthened this to
∑ i = 1 d Λ i + 1 + 4 ( Λ 2 − Λ 1 ) d + 4 ≤ ( d + 4 ) Λ 1 . \sum_{i=1}^{d}\Lambda_{i+1}+\frac{4(\Lambda_2-\Lambda_1)}{d+4}\leq(d+4)\Lambda_1. i = 1 ∑ d Λ i + 1 + d + 4 4 ( Λ 2 − Λ 1 ) ≤ ( d + 4 ) Λ 1 .
Looking only at the first ratio, the preceding inequality gives
d Λ 2 + 4 ( Λ 2 − Λ 1 ) d + 4 ≤ ∑ i = 1 d Λ i + 1 + 4 ( Λ 2 − Λ 1 ) d + 4 ≤ ( d + 4 ) Λ 1 , d\Lambda_2+\frac{4(\Lambda_2-\Lambda_1)}{d+4}\leq\sum_{i=1}^{d}\Lambda_{i+1}+\frac{4(\Lambda_2-\Lambda_1)}{d+4}\leq(d+4)\Lambda_1, d Λ 2 + d + 4 4 ( Λ 2 − Λ 1 ) ≤ i = 1 ∑ d Λ i + 1 + d + 4 4 ( Λ 2 − Λ 1 ) ≤ ( d + 4 ) Λ 1 ,
Λ 2 Λ 1 ≤ ( d + 4 ) 2 + 4 d ( d + 4 ) + 4 = d 2 + 8 d + 20 ( d + 2 ) 2 . \frac{\Lambda_2}{\Lambda_1}\leq\frac{(d+4)^2+4}{d(d+4)+4}=\frac{d^2+8d+20}{(d+2)^2}. Λ 1 Λ 2 ≤ d ( d + 4 ) + 4 ( d + 4 ) 2 + 4 = ( d + 2 ) 2 d 2 + 8 d + 20 .
Thus the improved inequality recovers the constant of Hile and Yeh and controls several further low eigenvalues without weakening the upper bound for Λ 2 / Λ 1 \Lambda_2/\Lambda_1 Λ 2 / Λ 1 .
For any k ≥ 1 k \ge1 k ≥ 1 , Cheng and Yang [12 ] established a general quadratic inequality with coefficient 4 ( d + 2 ) / d 2 4(d+2)/d^2 4 ( d + 2 ) / d 2 , and later improved it in [13 Theorem 1.1] to
∑ i = 1 k ( Λ k + 1 − Λ i ) 2 ≤ 4 ( d + 4 / 3 ) d 2 ∑ i = 1 k ( Λ k + 1 − Λ i ) Λ i . \sum_{i=1}^{k}(\Lambda_{k+1}-\Lambda_i)^2 \le\frac{4(d+4/3)}{d^2}\sum_{i=1}^{k}(\Lambda_{k+1}-\Lambda_i)\Lambda_i. i = 1 ∑ k ( Λ k + 1 − Λ i ) 2 ≤ d 2 4 ( d + 4/3 ) i = 1 ∑ k ( Λ k + 1 − Λ i ) Λ i .
Taking k = 1 k=1 k = 1 and Λ 2 > Λ 1 \Lambda_2>\Lambda_1 Λ 2 > Λ 1 gives
( Λ 2 − Λ 1 ) 2 ≤ 4 ( d + 4 / 3 ) d 2 ( Λ 2 − Λ 1 ) Λ 1 ⟹ Λ 2 Λ 1 ≤ 1 + 4 ( d + 4 / 3 ) d 2 . (\Lambda_2-\Lambda_1)^2 \le\frac{4(d+4/3)}{d^2}(\Lambda_2-\Lambda_1)\Lambda_1 \quad\Longrightarrow\quad\frac{\Lambda_2}{\Lambda_1}\le1+\frac{4(d+4/3)}{d^2}. ( Λ 2 − Λ 1 ) 2 ≤ d 2 4 ( d + 4/3 ) ( Λ 2 − Λ 1 ) Λ 1 ⟹ Λ 1 Λ 2 ≤ 1 + d 2 4 ( d + 4/3 ) .
The ratio bound also holds when Λ 2 = Λ 1 \Lambda_2=\Lambda_1 Λ 2 = Λ 1 . However, it is weaker than the Hile and Yeh bound:
1 + 4 ( d + 4 / 3 ) d 2 − d 2 + 8 d + 20 ( d + 2 ) 2 = 16 ( d 2 + 7 d + 4 ) 3 d 2 ( d + 2 ) 2 > 0. 1+\frac{4(d+4/3)}{d^2}-\frac{d^2+8d+20}{(d+2)^2}=\frac{16(d^2+7d+4)}{3d^2(d+2)^2}>0. 1 + d 2 4 ( d + 4/3 ) − ( d + 2 ) 2 d 2 + 8 d + 20 = 3 d 2 ( d + 2 ) 2 16 ( d 2 + 7 d + 4 ) > 0.
Thus these later general inequalities are meaningful progress for the whole spectrum, but the conclusions about the first ratio derived from them cannot replace the Hile–Yeh estimate used here.
The theorem for the membrane problem, together with a few simple symmetric examples, makes this conjecture appear quite plausible. For the square ( 0 , 1 ) 2 (0,1)^2 ( 0 , 1 ) 2 , separation of variables for the membrane problem gives
λ 2 D ( ( 0 , 1 ) 2 ) λ 1 D ( ( 0 , 1 ) 2 ) = 5 π 2 2 π 2 = 5 2 < j 1 , 1 2 j 0 , 1 2 ≈ 2.5387. \frac{\lambda^{D}_2((0,1)^2)}{\lambda^{D}_1((0,1)^2)}=\frac{5\pi^2}{2\pi^2}=\frac{5}{2}<\frac{j_{1,1}^2}{j_{0,1}^2}\approx2.5387. λ 1 D (( 0 , 1 ) 2 ) λ 2 D (( 0 , 1 ) 2 ) = 2 π 2 5 π 2 = 2 5 < j 0 , 1 2 j 1 , 1 2 ≈ 2.5387.
Numerical computations in the buckling case likewise place the square below the disk, the ratio being about 1.76, while C 2 ≈ 1.7964 C_2\approx1.7964 C 2 ≈ 1.7964 , see [4 Section 3.2.2 and Figure 7] . We note here that these buckling values are only numerical evidence, since the spectrum of the clamped square has no explicit formula. Antunes computed triangular, quadrilateral and octagonal domains and found that within each class the regular polygon maximizes the ratio, with the disk as the limiting case. Thus the membrane theorem and the computations on polygons both point toward the ball, whereas the later general inequalities do not address this specific comparison.
A direct transfer of the membrane proof encounters structural difficulties. The argument of Ashbaugh and Benguria uses product-type trial functions together with Chiti’s comparison theorem. The latter compares the rearranged positive first eigenfunction with the first eigenfunction on a ball having the same first eigenvalue, see [5 Section 2.3, equations (2.20)–(2.28)] . Let v > 0 v>0 v > 0 be the first Dirichlet eigenfunction and g ∈ C ∞ ( Ω ‾ ) g\in C^\infty(\overline{\Omega}) g ∈ C ∞ ( Ω ) a real-valued function, here d x dx d x denotes the Lebesgue volume measure. Taking g 2 v g^2v g 2 v as a test function in the equation for v v v yields
∫ Ω ∣ ∇ ( g v ) ∣ 2 d x − λ 1 D ∫ Ω g 2 v 2 d x = ∫ Ω v 2 ∣ ∇ g ∣ 2 d x . (2) \int_{\Omega}|\nabla(gv)|^2\,dx-\lambda^D_1\int_{\Omega}g^2v^2\,dx=\int_{\Omega}v^2|\nabla g|^2\,dx. \tag*{(2)} ∫ Ω ∣∇ ( g v ) ∣ 2 d x − λ 1 D ∫ Ω g 2 v 2 d x = ∫ Ω v 2 ∣∇ g ∣ 2 d x . (2)
If g v ≢ 0 gv \not\equiv0 g v ≡ 0 and ∫ Ω g v 2 d x = 0 \int_{\Omega} gv^{2}\,\mathrm{d}x=0 ∫ Ω g v 2 d x = 0 , then
( λ 2 D − λ 1 D ) ∫ Ω g 2 v 2 d x ≤ ∫ Ω v 2 ∣ ∇ g ∣ 2 d x . \left(\lambda_{2}^{\mathrm{D}}-\lambda_{1}^{\mathrm{D}}\right)\int_{\Omega}g^{2}v^{2}\,\mathrm{d}x\leq\int_{\Omega}v^{2}|\nabla g|^{2}\,\mathrm{d}x. ( λ 2 D − λ 1 D ) ∫ Ω g 2 v 2 d x ≤ ∫ Ω v 2 ∣∇ g ∣ 2 d x .
Both integrals have the positive weight v 2 v^{2} v 2 . Once a suitable radial multiplier is chosen and an eigenfunction comparison is made, one obtains the sharp membrane estimate.
For the first buckling eigenfunction u u u , the situation is different. The trial function g u gu g u must satisfy the gradient orthogonality condition
∫ Ω ∇ ( g u ) ⋅ ∇ u d x = − ∫ Ω g u Δ u d x = 0. (3) \int_{\Omega}\nabla(gu)\cdot\nabla u\,\mathrm{d}x=-\int_{\Omega}gu\Delta u\,\mathrm{d}x=0. \tag*{(3)} ∫ Ω ∇ ( g u ) ⋅ ∇ u d x = − ∫ Ω g u Δ u d x = 0. (3)
The clamped conditions imply g u , g 2 u ∈ H 0 2 ( Ω ) gu,g^{2}u\in H_{0}^{2}(\Omega) g u , g 2 u ∈ H 0 2 ( Ω ) , and hence
∫ Ω Δ u Δ ( g 2 u ) d x = Λ 1 ∫ Ω ∇ u ⋅ ∇ ( g 2 u ) d x . \int_{\Omega}\Delta u\Delta(g^{2}u)\,\mathrm{d}x=\Lambda_{1}\int_{\Omega}\nabla u\cdot\nabla(g^{2}u)\,\mathrm{d}x. ∫ Ω Δ u Δ ( g 2 u ) d x = Λ 1 ∫ Ω ∇ u ⋅ ∇ ( g 2 u ) d x .
By the product rule we have
Δ ( g u ) = g Δ u + 2 ∇ g ⋅ ∇ u + u Δ g , Δ ( g 2 u ) = g 2 Δ u + 4 g ∇ g ⋅ ∇ u + 2 u ∣ ∇ g ∣ 2 + 2 g u Δ g , ∣ ∇ ( g u ) ∣ 2 − ∇ u ⋅ ∇ ( g 2 u ) = u 2 ∣ ∇ g ∣ 2 . \begin{aligned}
\Delta(gu)&=g\Delta u+2\nabla g\cdot\nabla u+u\Delta g,\\
\Delta(g^{2}u)&=g^{2}\Delta u+4g\nabla g\cdot\nabla u+2u|\nabla g|^{2}+2gu\Delta g,\\
|\nabla(gu)|^{2}-\nabla u\cdot\nabla(g^{2}u)&=u^{2}|\nabla g|^{2}.
\end{aligned} Δ ( g u ) Δ ( g 2 u ) ∣∇ ( g u ) ∣ 2 − ∇ u ⋅ ∇ ( g 2 u ) = g Δ u + 2∇ g ⋅ ∇ u + u Δ g , = g 2 Δ u + 4 g ∇ g ⋅ ∇ u + 2 u ∣∇ g ∣ 2 + 2 g u Δ g , = u 2 ∣∇ g ∣ 2 .
Substituting these expressions and then subtracting the tested equation gives
∫ Ω [ Δ ( g u ) ] 2 d x − Λ 1 ∫ Ω ∣ ∇ ( g u ) ∣ 2 d x = ∫ Ω ( 2 ∇ g ⋅ ∇ u + u Δ g ) 2 d x − ∫ Ω ( 2 u Δ u + Λ 1 u 2 ) ∣ ∇ g ∣ 2 d x . (4) \begin{aligned}
\int_{\Omega}[\Delta(gu)]^{2}\,\mathrm{d}x-\Lambda_{1}\int_{\Omega}|\nabla(gu)|^{2}\,\mathrm{d}x
&=\int_{\Omega}(2\nabla g\cdot\nabla u+u\Delta g)^{2}\,\mathrm{d}x\\
&\quad-\int_{\Omega}(2u\Delta u+\Lambda_{1}u^{2})|\nabla g|^{2}\,\mathrm{d}x. \tag*{(4)}
\end{aligned} ∫ Ω [ Δ ( g u ) ] 2 d x − Λ 1 ∫ Ω ∣∇ ( g u ) ∣ 2 d x = ∫ Ω ( 2∇ g ⋅ ∇ u + u Δ g ) 2 d x − ∫ Ω ( 2 u Δ u + Λ 1 u 2 ) ∣∇ g ∣ 2 d x . (4)
Under (3) , provided g u ≢ 0 gu\not\equiv0 g u ≡ 0 , the left-hand side is bounded below by ( Λ 2 − Λ 1 ) ∫ Ω ∣ ∇ ( g u ) ∣ 2 d x (\Lambda_{2}-\Lambda_{1})\int_{\Omega}|\nabla(gu)|^{2}\,\mathrm{d}x ( Λ 2 − Λ 1 ) ∫ Ω ∣∇ ( g u ) ∣ 2 d x . In other words, both the orthogonality condition and the gap estimate involve derivatives of u u u , unlike the membrane case where only the positive weight u 2 u^{2} u 2 appears. Even on the unit disk, this orthogonality weight changes sign. The first radial mode on the disk satisfies
u ( r ) = J 0 ( j 1 , 1 r ) − J 0 ( j 1 , 1 ) , u ( 1 ) = 0 , u(r)=J_{0}(j_{1,1}r)-J_{0}(j_{1,1}),\qquad u(1)=0, u ( r ) = J 0 ( j 1 , 1 r ) − J 0 ( j 1 , 1 ) , u ( 1 ) = 0 ,
u ′ ( r ) = − j 1 , 1 J 1 ( j 1 , 1 r ) < 0 , 0 < r < 1 , u'(r)=-j_{1,1}J_{1}(j_{1,1}r)<0,\qquad0<r<1, u ′ ( r ) = − j 1 , 1 J 1 ( j 1 , 1 r ) < 0 , 0 < r < 1 ,
− u ( r ) Δ u ( r ) = j 1 , 1 2 u ( r ) J 0 ( j 1 , 1 r ) . -u(r)\Delta u(r)=j_{1,1}^{2}u(r)J_{0}(j_{1,1}r). − u ( r ) Δ u ( r ) = j 1 , 1 2 u ( r ) J 0 ( j 1 , 1 r ) .
Hence u ( r ) > 0 u(r)>0 u ( r ) > 0 for 0 ≤ r < 1 0\leq r<1 0 ≤ r < 1 , while − u Δ u -u\Delta u − u Δ u changes sign at r = j 0 , 1 / j 1 , 1 r=j_{0,1}/j_{1,1} r = j 0 , 1 / j 1 , 1 . Such derivative identities and the interlacing property of zeros are standard [23 p. 45 and pp. 477–487] . On general domains, the first buckling eigenfunction itself may change sign [9 , 10 ] , and taking its absolute value does not necessarily preserve H 2 H^{2} H 2 regularity. The membrane comparison provides neither a positive weight suitable for the orthogonality condition (3) nor the information about derivative terms required in (4) . This is the essential obstruction that makes the proof strategy difficult to carry over directly.
The membrane theorem and the computations on polygons both support an affirmative expectation, yet the corresponding proof cannot be transferred. More than twenty-five years after the conjecture was formulated, we reverse this expected comparison within the class of smooth connected domains. The construction relies on interior holes that are absent from those polygonal examples.
Our main result is as follows.
Theorem 1.1. For every integer d ≥ 2 d \ge2 d ≥ 2 , there is a bounded connected domain Ω d ⊂ R d \Omega_d \subset\mathbb{R}^d Ω d ⊂ R d with C ∞ C^\infty C ∞ boundary such that
Λ 2 ( Ω d ) Λ 1 ( Ω d ) > C d , \frac{\Lambda_2(\Omega_d)}{\Lambda_1(\Omega_d)} > C_d, Λ 1 ( Ω d ) Λ 2 ( Ω d ) > C d , where C d C_d C d is defined by (1) . The domain is the unit ball with finitely many disjoint closed tubular neighborhoods of round ( d − 2 ) (d-2) ( d − 2 ) -spheres removed. All their locations and radii are specified explicitly in terms of d d d in (2.3)–(2.4). For d = 2 d=2 d = 2 , the removed connected components are ordinary closed disks.
Theorem 1.1 shows that over all smooth connected domains, the value given by the ball cannot be a sharp universal constant. The constructed domains have interior holes. This result does not address the comparison with the ball among convex or simply connected domains, nor does it determine the optimal universal ratio. It is not in conflict with the comparison of the first eigenvalue at fixed volume mentioned above. Thus the ratio problem remains worth studying under additional geometric restrictions, or in cases where the sharp constant differs from C d C_d C d . The corresponding counterexamples to the clamped-plate eigenvalue ratio conjecture will be presented in our subsequent paper.
The idea is to use interior constraints so that the relative increment of the second eigenvalue is larger than that of the first. Domain monotonicity alone is not enough. Removing a subset raises both eigenvalues and reveals nothing about their ratio. A removal of small volume is also insufficient to estimate the cost of imposing the two clamped boundary conditions. We therefore compute the relative shifts directly, rule out competing spectral branches, and then realize this constraint through perforations whose energy cost we can control.
The proof of Theorem 1.1 starts from a rotationally invariant auxiliary problem that first isolates the effect of the perforation. For 0 < a < 1 0<a<1 0 < a < 1 , let S a = { x : ∣ x ∣ = a } S_a=\{x:|x|=a\} S a = { x : ∣ x ∣ = a } . In the space of clamped functions on B B B we impose only the zero value condition u ∣ S a = 0 u|_{S_a}=0 u ∣ S a = 0 , with no restriction on the normal derivative, so that the traces from the two sides of S a S_a S a coincide. The constraint is imposed in H 0 2 ( B ) H^2_0(B) H 0 2 ( B ) , defined in § 2 . This is not the buckling problem on the disjoint ball and annulus obtained by clamping the entire inner sphere. Let μ 1 ( a ) \mu_1(a) μ 1 ( a ) and μ 2 ( a ) \mu_2(a) μ 2 ( a ) be the first two variational eigenvalues of this problem. Let
h = 1 − a h = 1 - a h = 1 − a . For fixed d d d , the calculation in § 3 gives
μ 1 ( 1 − h ) j d / 2 , 1 2 = 1 + 3 2 h + d + 26 16 h 2 + O d ( h 3 ) , μ 2 ( 1 − h ) j d / 2 + 1 , 1 2 = 1 + 3 2 h + d + 28 16 h 2 + O d ( h 3 ) . (5) \begin{aligned}
\frac{\mu_1(1-h)}{j_{d/2,1}^2} &= 1 + \frac{3}{2}h + \frac{d+26}{16}h^2 + O_d(h^3),\\
\frac{\mu_2(1-h)}{j_{d/2+1,1}^2} &= 1 + \frac{3}{2}h + \frac{d+28}{16}h^2 + O_d(h^3). \tag*{(5)}
\end{aligned} j d /2 , 1 2 μ 1 ( 1 − h ) j d /2 + 1 , 1 2 μ 2 ( 1 − h ) = 1 + 2 3 h + 16 d + 26 h 2 + O d ( h 3 ) , = 1 + 2 3 h + 16 d + 28 h 2 + O d ( h 3 ) . (5)
Here O d ( h 3 ) O_d(h^3) O d ( h 3 ) denotes a remainder controlled as h → 0 h \to0 h → 0 by a constant depending on d d d times h 3 h^3 h 3 . In particular,
μ 2 ( 1 − h ) μ 1 ( 1 − h ) = C d [ 1 + 1 8 h 2 + O d ( h 3 ) 1 + 3 2 h + d + 26 16 h 2 + O d ( h 3 ) ] = C d ( 1 + 1 8 h 2 + O d ( h 3 ) ) . \begin{aligned}
\frac{\mu_2(1-h)}{\mu_1(1-h)}
= C_d\left[1+\frac{\frac{1}{8}h^2+O_d(h^3)}{1+\frac{3}{2}h+\frac{d+26}{16}h^2+O_d(h^3)}\right] \\
&= C_d\left(1+\frac{1}{8}h^2+O_d(h^3)\right).
\end{aligned} μ 1 ( 1 − h ) μ 2 ( 1 − h ) = C d [ 1 + 1 + 2 3 h + 16 d + 26 h 2 + O d ( h 3 ) 8 1 h 2 + O d ( h 3 ) ] = C d ( 1 + 8 1 h 2 + O d ( h 3 ) ) .
Thus the first order calculation alone does not reveal that (1.1) already fails. It is the positive second order difference in (1.5) that gives rise to the counterexample.
Before using this expansion, we first prove a uniform analytic bound on the remainder and rule out all competing angular and radial branches. Figure 1 shows the cancellation that makes the second order calculation necessary.
Figure 1. Taylor polynomials of the relevant quantities for d = 2 d = 2 d = 2 . The left panel has the common term 1 + 3 h / 2 1 + 3h/2 1 + 3 h /2 subtracted, and the right panel shows the ratio gain h 2 / 8 h^2/8 h 2 /8 . The figure only illustrates the coefficients and does not show actual computed eigenvalues. In the proof the range of h h h is smaller and is specified in § 2 .
The next step is to realize this constraint on genuine smooth holes. On S a S_a S a we choose finitely many round codimension-two latitude sets. When these latitudes spacing is sufficiently fine, the vanishing of a function on them forces its trace on S a S_a S a to be small. We proceed by adding a surface penalty term κ ∫ S a u 2 d S \kappa\int_{S_a}u^2\,dS κ ∫ S a u 2 d S to the bending energy ∫ B ( Δ u ) 2 d x \int_B(\Delta u)^2\,dx ∫ B ( Δ u ) 2 d x , which yields a lower bound for the second eigenvalue. Here d S dS d S denotes the surface measure and κ > 0 \kappa> 0 κ > 0 is an explicit coefficient. The first eigenfunction under the spherical constraint already vanishes on every latitude. Multiplying it by a logarithmic cutoff near these sets therefore gives a clamped trial function on the perforated domain whose energy error can be made arbitrarily small. The codimension being two is crucial here. The core has dimension d − 2 d-2 d − 2 , and the normal variables span a plane. The logarithmic transition makes the additional energy required to clamp the first derivatives tend to zero, which places the construction exactly at the critical case for the cost of derivative cutoff. At the same time, the latitude geometry allows us to perform only one-dimensional sampling estimates on the sphere in every dimension d ≥ 2 d \ge2 d ≥ 2 . This is the classical critical cutoff mechanism in Sobolev capacity theory [2], where capacity describes precisely the energy required to impose a local constraint. All estimates needed in this paper are proved directly with explicit constants.
Figure 2 summarizes the transition from the spherical value constraint to the final domain. To the best of our knowledge, this construction is new. It combines the second order ratio gain produced by the interior value constraint with a quantitative realization through smooth codimension-two perforations. The variational methods, Bessel functions and capacity tools used here are themselves classical. The same idea may apply to other spectral ratio problems provided that the first order relative shifts coincide exactly and the interior constraint can be realized at a sufficiently small energy cost. However, such extensions require separate spectral estimates and cutoff estimates.
Figure 2. Planar construction: the value constraint on S a S_a S a , the paired core points, and the clamped holes. The shaded region is the final domain, and the dashed circle only serves to locate the cores. Seven representative latitudes are shown, with the spacing and hole radii exaggerated.
The rest of this paper is organized as follows. § 2 fixes the variational framework and the domain parameters. § 3 establishes the strict spectral gap and identifies exactly which two eigenvalues it concerns. § 9 maintains the lower bound for the second eigenvalue using only finitely many cores. § 11 controls the derivatives of the first constrained mode so that the cost of the cutoff can be estimated explicitly. § 13 combines these two bounds and verifies that the resulting domain is smooth and connected. None of the intermediate constraint steps can be omitted. They are precisely what allows us to pass from the solvable radial problem to genuine clamped holes while preserving the strict ratio gap.
Variational framework and the explicit domains
This section first introduces a unified variational framework on the unit ball in which the auxiliary constrained problem and the perforated domain problem can be compared. We then specify the geometric parameters of the domain in Theorem 1.1 .
Unless complex parameters are explicitly introduced, the function spaces below are taken to be real. For an open set G G G , H k ( G ) H^k(G) H k ( G ) is the Sobolev space consisting of functions that are square integrable together with their weak derivatives up to order k k k . H 0 k ( G ) H^k_0(G) H 0 k ( G ) is the closure of C c ∞ ( G ) C^\infty_c(G) C c ∞ ( G ) in the corresponding norm, where C c ∞ ( G ) C^\infty_c(G) C c ∞ ( G ) denotes the smooth functions whose support is compactly contained in G G G . Fractional Sobolev spaces on smooth surfaces are defined in local coordinates, and we use their standard trace and compact embedding theorems, see [1 ] .
Write D u = ( ∂ x 1 u , … , ∂ x d u ) Du=(\partial_{x_1}u,\ldots,\partial_{x_d}u) D u = ( ∂ x 1 u , … , ∂ x d u ) for the weak gradient. On H 0 2 ( B ) H^2_0(B) H 0 2 ( B ) , define the bending form and the Dirichlet form
Q ( u , v ) = ∫ B Δ u Δ v d x , E ( u , v ) = ∫ B D u ⋅ D v d x . Q(u,v)=\int_B \Delta u\Delta v\,\mathrm{d}x,\qquad\mathcal{E}(u,v)=\int_B Du\cdot Dv\,\mathrm{d}x. Q ( u , v ) = ∫ B Δ u Δ v d x , E ( u , v ) = ∫ B D u ⋅ D v d x .
We abbreviate Q ( u , u ) Q(u,u) Q ( u , u ) and E ( u , u ) \mathcal{E}(u,u) E ( u , u ) as Q ( u ) Q(u) Q ( u ) and E ( u ) \mathcal{E}(u) E ( u ) , respectively. Let D 2 u D^2u D 2 u denote the Hessian matrix formed by the second weak derivatives. By standard integration by parts and the Poincaré estimate [1 ] , we have
∥ D 2 u ∥ L 2 ( B ) 2 = Q ( u ) , ∥ u ∥ L 2 ( B ) ≤ 2 π ∥ D u ∥ L 2 ( B ) , ∥ D u ∥ L 2 ( B ) ≤ 2 π ∥ Δ u ∥ L 2 ( B ) . (6) \lVert D^2u\rVert_{L^2(B)}^2=Q(u),\qquad\lVert u\rVert_{L^2(B)}\le\frac{2}{\pi}\lVert Du\rVert_{L^2(B)},\qquad\lVert Du\rVert_{L^2(B)}\le\frac{2}{\pi}\lVert\Delta u\rVert_{L^2(B)}. \tag*{(6)} ∥ D 2 u ∥ L 2 ( B ) 2 = Q ( u ) , ∥ u ∥ L 2 ( B ) ≤ π 2 ∥ D u ∥ L 2 ( B ) , ∥ D u ∥ L 2 ( B ) ≤ π 2 ∥ Δ u ∥ L 2 ( B ) . (6)
Here the factor 2 / π 2/\pi 2/ π comes from the reciprocal square root of the first Dirichlet eigenvalue π 2 / 4 \pi^2/4 π 2 /4 on the interval ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) . Extending u u u by zero to ( − 1 , 1 ) d (-1,1)^d ( − 1 , 1 ) d and then applying this one-dimensional inequality along each parallel line yields the estimates above. For u ∈ C c ∞ ( B ) u\in C^\infty_c(B) u ∈ C c ∞ ( B ) , integration by parts gives
∫ B ( Δ u ) 2 d x = ∑ i , j = 1 d ∫ B ( ∂ x i 2 u ) ( ∂ x j 2 u ) d x = − ∑ i , j = 1 d ∫ B ( ∂ x i u ) ( ∂ x i ∂ x j 2 u ) d x = ∑ i , j = 1 d ∫ B ( ∂ x i ∂ x j u ) 2 d x = ∫ B ∣ D 2 u ∣ 2 d x . \begin{aligned}
\int_B(\Delta u)^2\,\mathrm{d}x
&=\sum_{i,j=1}^{d}\int_B(\partial_{x_i}^2u)(\partial_{x_j}^2u)\,\mathrm{d}x\\
&=-\sum_{i,j=1}^{d}\int_B(\partial_{x_i}u)(\partial_{x_i}\partial_{x_j}^2u)\,\mathrm{d}x\\
&=\sum_{i,j=1}^{d}\int_B(\partial_{x_i}\partial_{x_j}u)^2\,\mathrm{d}x
=\int_B\lvert D^2u\rvert^2\,\mathrm{d}x.
\end{aligned} ∫ B ( Δ u ) 2 d x = i , j = 1 ∑ d ∫ B ( ∂ x i 2 u ) ( ∂ x j 2 u ) d x = − i , j = 1 ∑ d ∫ B ( ∂ x i u ) ( ∂ x i ∂ x j 2 u ) d x = i , j = 1 ∑ d ∫ B ( ∂ x i ∂ x j u ) 2 d x = ∫ B ∣ D 2 u ∣ 2 d x .
Extending u u u by zero to ( − 1 , 1 ) d (-1,1)^d ( − 1 , 1 ) d , the one-dimensional Dirichlet inequality and Fubini’s theorem give
∫ B u 2 d x ≤ 4 π 2 ∫ B ( ∂ x 1 u ) 2 d x ≤ 4 π 2 ∫ B ∣ D u ∣ 2 d x . \int_Bu^2\,\mathrm{d}x\le\frac{4}{\pi^2}\int_B(\partial_{x_1}u)^2\,\mathrm{d}x\le\frac{4}{\pi^2}\int_B\lvert Du\rvert^2\,\mathrm{d}x. ∫ B u 2 d x ≤ π 2 4 ∫ B ( ∂ x 1 u ) 2 d x ≤ π 2 4 ∫ B ∣ D u ∣ 2 d x .
On the other hand,
∥ D u ∥ 2 2 = − ∫ B u Δ u d x ≤ ∥ u ∥ 2 ∥ Δ u ∥ 2 ≤ 2 π ∥ D u ∥ 2 ∥ Δ u ∥ 2 . \lVert Du\rVert_2^2=-\int_Bu\Delta u\,\mathrm{d}x\le\lVert u\rVert_2\lVert\Delta u\rVert_2\le\frac{2}{\pi}\lVert Du\rVert_2\lVert\Delta u\rVert_2. ∥ D u ∥ 2 2 = − ∫ B u Δ u d x ≤ ∥ u ∥ 2 ∥ Δ u ∥ 2 ≤ π 2 ∥ D u ∥ 2 ∥ Δ u ∥ 2 .
Thus,
∥ D u ∥ 2 ≤ 2 π ∥ Δ u ∥ 2 , Q ( u ) ≤ ∥ u ∥ 2 2 + ∥ D u ∥ 2 2 + ∥ D 2 u ∥ 2 2 ≤ ( 1 + 4 π 2 + 16 π 4 ) Q ( u ) . \lVert Du\rVert_2 \le\frac{2}{\pi}\lVert\Delta u\rVert_2,\qquad\mathcal{Q}(u) \le\lVert u\rVert_2^2+\lVert Du\rVert_2^2+\lVert D^2u\rVert_2^2 \le\left(1+\frac{4}{\pi^2}+\frac{16}{\pi^4}\right)\mathcal{Q}(u). ∥ D u ∥ 2 ≤ π 2 ∥ Δ u ∥ 2 , Q ( u ) ≤ ∥ u ∥ 2 2 + ∥ D u ∥ 2 2 + ∥ D 2 u ∥ 2 2 ≤ ( 1 + π 2 4 + π 4 16 ) Q ( u ) .
By density, these formulas also hold on H 0 2 ( B ) H_0^2(B) H 0 2 ( B ) , and hence Q 1 / 2 \mathcal{Q}^{1/2} Q 1/2 is an equivalent norm on this space. From here on, every L 2 L^2 L 2 norm without a subscript is taken over the domain under consideration at that point. The symbol ∥ ⋅ ∥ ∞ \lVert\cdot\rVert_\infty ∥ ⋅ ∥ ∞ denotes the essential supremum norm.
For any closed infinite-dimensional subspace W ⊂ H 0 2 ( B ) W\subset H_0^2(B) W ⊂ H 0 2 ( B ) , define
λ k ( W ) = min F ⊂ W dim F = k max 0 ≠ u ∈ F Q ( u ) E ( u ) . (7) \lambda_k(W)=\min_{\substack{F\subset W\\ \dim F=k}}\ \max_{0\ne u\in F}\frac{\mathcal{Q}(u)}{\mathcal{E}(u)}. \tag*{(7)} λ k ( W ) = F ⊂ W d i m F = k min 0 = u ∈ F max E ( u ) Q ( u ) . (7)
The compact embedding H 0 2 ( B ) ↪ H 0 1 ( B ) H_0^2(B)\hookrightarrow H_0^1(B) H 0 2 ( B ) ↪ H 0 1 ( B ) ensures that these min-max values are attained and that the spectrum is discrete. This is the variational formulation commonly used for the buckling problem, see [10 Sections 7–8] . If G ⊂ B G\subset B G ⊂ B , we identify each function in H 0 2 ( G ) H_0^2(G) H 0 2 ( G ) with its zero extension, so that Λ k ( G ) = λ k ( H 0 2 ( G ) ) \Lambda_k(G)=\lambda_k(H_0^2(G)) Λ k ( G ) = λ k ( H 0 2 ( G )) . In particular,
W 1 ⊂ W 2 ⟹ λ k ( W 1 ) ≥ λ k ( W 2 ) . W_1\subset W_2\quad\Longrightarrow\quad\lambda_k(W_1)\ge\lambda_k(W_2). W 1 ⊂ W 2 ⟹ λ k ( W 1 ) ≥ λ k ( W 2 ) .
Next consider scaling. Let u t ( x ) = u ( x / t ) u_t(x)=u(x/t) u t ( x ) = u ( x / t ) . A direct calculation gives
∫ t G ( Δ u t ) 2 d x ∫ t G ∣ D u t ∣ 2 d x = t d − 4 ∫ G ( Δ u ) 2 d x t d − 2 ∫ G ∣ D u ∣ 2 d x = t − 2 ∫ G ( Δ u ) 2 d x ∫ G ∣ D u ∣ 2 d x . \frac{\int_{tG}(\Delta u_t)^2\,\mathrm{d}x}{\int_{tG}|Du_t|^2\,\mathrm{d}x}
=\frac{t^{d-4}\int_G(\Delta u)^2\,\mathrm{d}x}{t^{d-2}\int_G|Du|^2\,\mathrm{d}x}
=t^{-2}\frac{\int_G(\Delta u)^2\,\mathrm{d}x}{\int_G|Du|^2\,\mathrm{d}x}. ∫ tG ∣ D u t ∣ 2 d x ∫ tG ( Δ u t ) 2 d x = t d − 2 ∫ G ∣ D u ∣ 2 d x t d − 4 ∫ G ( Δ u ) 2 d x = t − 2 ∫ G ∣ D u ∣ 2 d x ∫ G ( Δ u ) 2 d x .
The map H 0 2 ( G ) ∋ u ↦ u t ∈ H 0 2 ( t G ) H_0^2(G)\ni u\mapsto u_t\in H_0^2(tG) H 0 2 ( G ) ∋ u ↦ u t ∈ H 0 2 ( tG ) is a bijection. Substituting it into the same min-max formula yields Λ k ( t G ) = t − 2 Λ k ( G ) \Lambda_k(tG)=t^{-2}\Lambda_k(G) Λ k ( tG ) = t − 2 Λ k ( G ) . For smooth domains, membership in H 0 2 ( G ) H_0^2(G) H 0 2 ( G ) is equivalent to the vanishing of both the trace and the trace of the normal derivative [1 ] . The subsequent spectral comparisons and the construction of piecewise smooth trial functions both rely on these facts.
The construction involves three successively smaller geometric scales: the distance from the constrained sphere to the outer boundary, the spacing between the latitudes, and the radius of the tubes. The following choice of parameters ensures that all spectral comparisons are strict. The exponents are all taken in dyadic form, and the numerical values are only required to lie near the sufficient thresholds given later. We make no claim about the optimality of the underlying analytic bounds and derivative bounds.
We explain why the three scales are chosen in this order. The spectral calculation below yields a gain of order h 2 h^2 h 2 , while the remainder is a constant depending on d d d times h 3 h^3 h 3 . We first choose h h h sufficiently small so that the remainder accounts for only a fixed fraction of the gain. Next we approximate the sphere by latitudes. The surface penalty term is of order h − 4 h^{-4} h − 4 , and the trace estimate contributes an additional factor h − 1 h^{-1} h − 1 . To make the combined error smaller than order h 2 h^2 h 2 , it suffices to require that the square of the latitude spacing be less than a constant depending on d d d times h 7 h^7 h 7 . The specific choice satisfies
h − 4 h − 12 − 2 N h 2 = 2 7 H − 2 N ≤ 2 − d − 23 , \frac{h^{-4}h^{-12-2N}}{h^2}=2^{7H-2N}\le2^{-d-23}, h 2 h − 4 h − 12 − 2 N = 2 7 H − 2 N ≤ 2 − d − 23 ,
which is exactly why the factor 7 H 7H 7 H appears in N N N .
Only after the sphere and the finite set of latitudes have been fixed do we turn to the hole radii. For a cutoff function supported in small tubes, the cost consists of two parts. The first is the volume of the support, and the second is the cost of modifying the derivatives. The former is reduced by shrinking the outer cutoff radius 2 − K 2^{-K} 2 − K , while the latter is reduced by increasing the logarithmic width 2 Z 2^Z 2 Z . Together they give
ε = 2 − K exp ( − 2 Z ) . \varepsilon=2^{-K}\exp(-2^Z). ε = 2 − K exp ( − 2 Z ) .
The fact that the hole radii can be chosen extremely small is precisely a consequence of the logarithmic cutoff estimate.
The letters H H H , N N N , K K K , Z Z Z below record the choices made above. Powers of two are used so that the prescriptions are explicit and the subsequent comparisons are easy to check. The value h h h is the largest dyadic value allowed by the remainder estimate, and N N N is the smallest integer satisfying the coarse grid condition given later. The cutoff exponents are determined by explicit sufficient energy bounds, and the holes could perhaps be larger if these bounds are improved in the future. The relevant comparisons are verified in § 3 –§ 13 .
For an integer d ≥ 2 d\ge2 d ≥ 2 , let
m = d + 4 , s = ⌈ log 2 m ⌉ , H = 32 + ⌈ 3 log 2 m ⌉ , N = ⌈ d + 23 + 7 H 2 ⌉ . m=d+4,\qquad s=\lceil\log_2 m\rceil,\qquad H=32+\lceil3\log_2 m\rceil,\qquad N=\left\lceil\frac{d+23+7H}{2}\right\rceil. m = d + 4 , s = ⌈ log 2 m ⌉ , H = 32 + ⌈ 3 log 2 m ⌉ , N = ⌈ 2 d + 23 + 7 H ⌉ .
K = 13 m 2 + ⌈ N / 2 ⌉ + d + 4 H + 2 s + 23 , K=13m^2+\lceil N/2\rceil+d+4H+2s+23, K = 13 m 2 + ⌈ N /2 ⌉ + d + 4 H + 2 s + 23 ,
Z = 26 m 2 + N + 2 d + 8 H + 4 s + 50. Z=26m^2+N+2d+8H+4s+50. Z = 26 m 2 + N + 2 d + 8 H + 4 s + 50.
Here ⌈ t ⌉ \lceil t\rceil ⌈ t ⌉ denotes the least integer not smaller than t t t . Define the geometric scales by
h = 2 − H , a = 1 − h , ε = 2 − K exp ( − 2 Z ) . (8) h=2^{-H},\qquad a=1-h,\qquad\varepsilon=2^{-K}\exp(-2^Z). \tag*{(8)} h = 2 − H , a = 1 − h , ε = 2 − K exp ( − 2 Z ) . (8)
Set θ j = j π / 2 N \theta_j=j\pi/2^N θ j = j π / 2 N and put
M j = { x ∈ R d : x d = a cos θ j , ∣ ( x 1 , … , x d − 1 ) ∣ = a sin θ j } , 1 ≤ j < 2 N , M_j=\left\{x\in\mathbb{R}^d:x_d=a\cos\theta_j,\ |(x_1,\ldots,x_{d-1})|=a\sin\theta_j\right\},\qquad1\le j<2^N, M j = { x ∈ R d : x d = a cos θ j , ∣ ( x 1 , … , x d − 1 ) ∣ = a sin θ j } , 1 ≤ j < 2 N ,
and K N = ⋃ j M j K_N=\bigcup_j M_j K N = ⋃ j M j . Each M j M_j M j is a round sphere of dimension d − 2 d-2 d − 2 , which degenerates to two points when d = 2 d=2 d = 2 . The union K N K_N K N is called the core set. The two lowest-dimensional cases are
d = 2 : M j = { ( ± a sin θ j , a cos θ j ) } , d=2:\qquad M_j=\{(\pm a\sin\theta_j,a\cos\theta_j)\}, d = 2 : M j = {( ± a sin θ j , a cos θ j )} ,
d = 3 : M j = { ( a sin θ j cos φ , a sin θ j sin φ , a cos θ j ) : 0 ≤ φ < 2 π } . d=3:\qquad M_j=\{(a\sin\theta_j\cos\varphi,a\sin\theta_j\sin\varphi,a\cos\theta_j):0\le\varphi<2\pi\}. d = 3 : M j = {( a sin θ j cos φ , a sin θ j sin φ , a cos θ j ) : 0 ≤ φ < 2 π } .
Thus when d = 2 d = 2 d = 2 a latitude gives a pair of points, and when d = 3 d = 3 d = 3 it gives a horizontal circle, see Figure 3 . Their closed ε \varepsilon ε -neighborhoods are two disks and a solid torus, respectively. Note that the cores themselves are not part of the boundary.
Figure 3. A core M j M_j M j on S a S_a S a in the two- and three-dimensional cases. The poles are omitted in the figure. The right panel is only a schematic projection, and the dashed arc represents the rear part of the circle.
The required domain is finally defined as
Ω d = B 1 ( 0 ) ∖ { x : dist ( x , K N ) ≤ ε } . (9) \Omega_d = B_1(0) \setminus\{x : \operatorname{dist}(x,K_N) \le\varepsilon\}. \tag*{(9)} Ω d = B 1 ( 0 ) ∖ { x : dist ( x , K N ) ≤ ε } . (9)
Here dist \operatorname{dist} dist denotes the Euclidean distance. We write S q S^q S q for the unit q q q -sphere and ∣ S q ∣ |S^q| ∣ S q ∣ for its surface measure, with the convention ∣ S 0 ∣ = 2 |S^0| = 2 ∣ S 0 ∣ = 2 . The above parameters distinguish the three scales clearly. The quantity h h h is the distance from S a S_a S a to the outer boundary, 2 − N 2^{-N} 2 − N controls the latitude spacing, and ε \varepsilon ε is the thickness of each hole. In what follows we first establish the spectral gap for the constrained problem on the whole sphere, and then show that neither the latitude spacing nor the hole thickness destroys this gap.
The spherical constraint and its spectral gap
First we impose the zero value condition on the sphere that contains the core set. By rotational symmetry, this auxiliary problem reduces to individual radial sectors. A uniform expansion together with a quantitative separation between sectors then yields the strict ratio gap needed for the perforation.
Radial reduction and quantitative spectral separation
To study the dependence on the distance to the boundary, we temporarily allow 0 < a < 1 0 < a < 1 0 < a < 1 to vary and write h = 1 − a h = 1 - a h = 1 − a . For S a = { x : ∣ x ∣ = a } S_a = \{x : |x| = a\} S a = { x : ∣ x ∣ = a } , define
V a = { u ∈ H 0 2 ( B ) : Tr S a u = 0 } , μ k ( a ) : = λ k ( V a ) . V_a = \{u \in H_0^2(B) : \operatorname{Tr}_{S_a} u = 0\}, \qquad\mu_k(a) := \lambda_k(V_a). V a = { u ∈ H 0 2 ( B ) : Tr S a u = 0 } , μ k ( a ) := λ k ( V a ) .
Here Tr S a \operatorname{Tr}_{S_a} Tr S a is the Sobolev trace operator. Its kernel is closed and contains C ∞ ( B a ) C^\infty(B_a) C ∞ ( B a ) , so V a V_a V a is infinite-dimensional and (2.2) applies directly. No restriction is imposed on the slope at the interface. The following proposition combines the radial reduction, the transmission conditions and the comparison estimates.
Proposition 3.1. Let d ≥ 2 d \ge2 d ≥ 2 , 0 < a < 1 0 < a < 1 0 < a < 1 , and let Y l Y_l Y l be a spherical harmonic of degree l l l satisfying ∫ S d − 1 Y l 2 d S = 1 \int_{S^{d-1}} Y_l^2 \,\mathrm{d}S = 1 ∫ S d − 1 Y l 2 d S = 1 . Write D = d + 2 l D = d + 2l D = d + 2 l .
(i) For u = r l v ( r ) Y l u = r^l v(r)Y_l u = r l v ( r ) Y l , the two sector forms read
Q ( u ) = ∫ 0 1 r D − 1 ( L D v ) 2 d r , L D = ∂ r 2 + ( D − 1 ) r − 1 ∂ r , Q(u) = \int_0^1 r^{D-1}(L_Dv)^2\,\mathrm{d}r, \qquad L_D = \partial_r^2 + (D-1)r^{-1}\partial_r, Q ( u ) = ∫ 0 1 r D − 1 ( L D v ) 2 d r , L D = ∂ r 2 + ( D − 1 ) r − 1 ∂ r , E ( u ) = ∫ 0 1 r D − 1 ( v ′ ) 2 d r . (10) \mathcal{E}(u) = \int_0^1 r^{D-1}(v')^2\,\mathrm{d}r. \tag*{(10)} E ( u ) = ∫ 0 1 r D − 1 ( v ′ ) 2 d r . (10) The regular radial form domain is the closure in the bending norm of smooth profiles that vanish near 1 1 1 and are of order r 2 r^2 r 2 near 0 0 0 . The outer clamped conditions are written as v ( 1 ) = v ′ ( 1 ) = 0 v(1) = v'(1) = 0 v ( 1 ) = v ′ ( 1 ) = 0 . With respect to an orthogonal basis of spherical harmonics, distinct basis elements are orthogonal for both forms, so the full spectrum is the union of the spectra of the sectors, with angular multiplicities counted.
(ii) In effective dimension D D D , the spectrum of the unconstrained ball is
j D / 2 , n 2 , n = 1 , 2 , … (11) j_{D/2,n}^2, \qquad n = 1,2,\ldots \tag*{(11)} j D /2 , n 2 , n = 1 , 2 , … (11) (iii) Let ρ D ( a ) \rho_D(a) ρ D ( a ) be the first radial eigenvalue under the constraint v ( a ) = 0 v(a) = 0 v ( a ) = 0 . A nonzero profile is a constrained eigenfunction corresponding to the eigenvalue k 2 > 0 k^2 > 0 k 2 > 0 if and only if it is regular at 0 0 0 , clamped at 1 1 1 , satisfies
( L D 2 + k 2 L D ) v = 0 on ( 0 , a ) ∪ ( a , 1 ) , (12) (L_D^2 + k^2L_D)v = 0 \qquad\text{on } (0,a) \cup(a,1), \tag*{(12)} ( L D 2 + k 2 L D ) v = 0 on ( 0 , a ) ∪ ( a , 1 ) , (12) and in addition satisfies
v ( a − ) = v ( a + ) = 0 , v(a-) = v(a+) = 0, v ( a − ) = v ( a + ) = 0 , v ′ ( a − ) = v ′ ( a + ) , ( L D v ) ( a − ) = ( L D v ) ( a + ) . (13) v'(a-) = v'(a+), \qquad(L_Dv)(a-) = (L_Dv)(a+). \tag*{(13)} v ′ ( a − ) = v ′ ( a + ) , ( L D v ) ( a − ) = ( L D v ) ( a + ) . (13) No condition is imposed on the jump of ( L D v ) ′ (L_Dv)' ( L D v ) ′ .
(iv) Write α D = j D / 2 , 1 \alpha_D = j_{D/2,1} α D = j D /2 , 1 . Then
D / 2 < α D ≤ D + 2 , D ≥ 2. (14) D/2 < \alpha_D \le D+2, \qquad D \ge2. \tag*{(14)} D /2 < α D ≤ D + 2 , D ≥ 2. (14) The gaps between zeros satisfy
j ν + 1 , n 2 − j ν , n 2 ≥ 2 ν + 1 , ν > 0 , n ≥ 1 , (15) j_{\nu+1,n}^2 - j_{\nu,n}^2 \ge2\nu+1, \qquad\nu> 0, \qquad n \ge1, \tag*{(15)} j ν + 1 , n 2 − j ν , n 2 ≥ 2 ν + 1 , ν > 0 , n ≥ 1 , (15) and moreover
j ν , 2 2 − j ν + 1 , 1 2 > 2 ν + 2 , ν ≥ 1. (16) j_{\nu,2}^2 - j_{\nu+1,1}^2 > 2\nu+2, \qquad\nu\ge1. \tag*{(16)} j ν , 2 2 − j ν + 1 , 1 2 > 2 ν + 2 , ν ≥ 1. (16)
Proof. Step 1. Radial forms and their domains.
In each space of spherical harmonics we choose an orthonormal basis. For one basis element Y l Y_l Y l ,
− Δ S d − 1 Y l = l ( l + d − 2 ) Y l , ∫ S d − 1 Y l 2 d S = 1. -\Delta_{S^{d-1}}Y_l=l(l+d-2)Y_l,\qquad\int_{S^{d-1}}Y_l^2\,\mathrm{d}S=1. − Δ S d − 1 Y l = l ( l + d − 2 ) Y l , ∫ S d − 1 Y l 2 d S = 1.
Integration by parts on the sphere gives
∫ S d − 1 ∣ ∇ S Y l ∣ 2 d S = l ( l + d − 2 ) . \int_{S^{d-1}}|\nabla_S Y_l|^2\,\mathrm{d}S=l(l+d-2). ∫ S d − 1 ∣ ∇ S Y l ∣ 2 d S = l ( l + d − 2 ) .
In this angular sector we write
u ( r , ω ) = r l v ( r ) Y l ( ω ) , D = d + 2 l , r = ∣ x ∣ , ω = x / ∣ x ∣ . u(r,\omega)=r^l v(r)Y_l(\omega),\qquad D=d+2l,\qquad r=|x|,\qquad\omega=x/|x|. u ( r , ω ) = r l v ( r ) Y l ( ω ) , D = d + 2 l , r = ∣ x ∣ , ω = x /∣ x ∣.
The Laplacian in polar coordinates is
Δ = ∂ r 2 + d − 1 r ∂ r + 1 r 2 Δ S d − 1 . \Delta=\partial_r^2+\frac{d-1}{r}\partial_r+\frac{1}{r^2}\Delta_{S^{d-1}}. Δ = ∂ r 2 + r d − 1 ∂ r + r 2 1 Δ S d − 1 .
Substituting the angular eigenvalue gives
Δ u = r l Y l [ v ′ ′ + 2 l + d − 1 r v ′ + l ( l − 1 ) + ( d − 1 ) l − l ( l + d − 2 ) r 2 v ] = r l Y l [ v ′ ′ + D − 1 r v ′ ] . \begin{aligned}
\Delta u=r^lY_l\left[v^{\prime\prime}+\frac{2l+d-1}{r}v^\prime+\frac{l(l-1)+(d-1)l-l(l+d-2)}{r^2}v\right] \\
&=r^lY_l\left[v^{\prime\prime}+\frac{D-1}{r}v^\prime\right].
\end{aligned} Δ u = r l Y l [ v ′′ + r 2 l + d − 1 v ′ + r 2 l ( l − 1 ) + ( d − 1 ) l − l ( l + d − 2 ) v ] = r l Y l [ v ′′ + r D − 1 v ′ ] .
The zeroth order term cancels exactly. Since the volume element is r d − 1 d r d S r^{d-1}\,\mathrm{d}r\,\mathrm{d}S r d − 1 d r d S , squaring the identity above shows that the weight in the bending energy is r 2 l + d − 1 = r D − 1 r^{2l+d-1}=r^{D-1} r 2 l + d − 1 = r D − 1 .
For l = 0 l=0 l = 0 , the gradient is radial, and the second identity in (3.1) follows immediately. For l ≥ 1 l\geq1 l ≥ 1 , separating the radial and angular parts gives
E ( u ) = ∫ 0 1 r d − 1 [ ( l r l − 1 v + r l v ′ ) 2 + l ( l + d − 2 ) r 2 l − 2 v 2 ] d r = ∫ 0 1 r D − 1 ( v ′ ) 2 d r + 2 l ∫ 0 1 r D − 2 v v ′ d r + l ( D − 2 ) ∫ 0 1 r D − 3 v 2 d r . \begin{aligned}
\mathcal{E}(u)=\int_0^1 r^{d-1}\left[(lr^{l-1}v+r^lv^\prime)^2+l(l+d-2)r^{2l-2}v^2\right]\,\mathrm{d}r \\
&=\int_0^1 r^{D-1}(v^\prime)^2\,\mathrm{d}r+2l\int_0^1 r^{D-2}vv^\prime\,\mathrm{d}r+l(D-2)\int_0^1 r^{D-3}v^2\,\mathrm{d}r.
\end{aligned} E ( u ) = ∫ 0 1 r d − 1 [ ( l r l − 1 v + r l v ′ ) 2 + l ( l + d − 2 ) r 2 l − 2 v 2 ] d r = ∫ 0 1 r D − 1 ( v ′ ) 2 d r + 2 l ∫ 0 1 r D − 2 v v ′ d r + l ( D − 2 ) ∫ 0 1 r D − 3 v 2 d r .
The last two terms can be combined into a boundary term, since
2 l ∫ 0 1 r D − 2 v v ′ d r = l [ r D − 2 v 2 ] 0 1 − l ( D − 2 ) ∫ 0 1 r D − 3 v 2 d r . 2l\int_0^1 r^{D-2}vv^\prime\,\mathrm{d}r=l\left[r^{D-2}v^2\right]_0^1-l(D-2)\int_0^1 r^{D-3}v^2\,\mathrm{d}r. 2 l ∫ 0 1 r D − 2 v v ′ d r = l [ r D − 2 v 2 ] 0 1 − l ( D − 2 ) ∫ 0 1 r D − 3 v 2 d r .
For the smooth profiles in the statement, v ( 1 ) = 0 v(1)=0 v ( 1 ) = 0 . When l ≥ 1 l\geq1 l ≥ 1 , we also have r D − 2 v ( r ) 2 → 0 r^{D-2}v(r)^2\to0 r D − 2 v ( r ) 2 → 0 as r ↓ 0 r\downarrow0 r ↓ 0 . This proves (3.1), the case l = 0 l=0 l = 0 having been treated separately above. Both forms have the radial weight in dimension D D D . The preceding calculation is first carried out for smooth profiles that vanish near r = 1 r=1 r = 1 and are smooth functions of order r 2 r^2 r 2 near r = 0 r=0 r = 0 . The regular radial form domain is the closure of these profiles in the bending norm of (3.1). Spherical harmonic projections of smooth approximating functions are dense in the physical angular sector, so these identities extend to the completed space. Equivalently, after removing the constant spherical area factor, the resulting radial space is exactly the radial subspace of H 0 2 ( B 1 ⊂ R D ) H^2_0(B_1 \subset\mathbb{R}^D) H 0 2 ( B 1 ⊂ R D ) . In particular, the origin is an interior point of the problem in effective dimension, rather than an additional boundary on which singular solutions can be chosen arbitrarily.
The outer clamped conditions become v ( 1 ) = v ′ ( 1 ) = 0 v(1)=v'(1)=0 v ( 1 ) = v ′ ( 1 ) = 0 . Indeed, u ( 1 , ω ) = v ( 1 ) Y l ( ω ) u(1,\omega)=v(1)Y_l(\omega) u ( 1 , ω ) = v ( 1 ) Y l ( ω ) and ∂ r u ( 1 , ω ) = ( l v ( 1 ) + v ′ ( 1 ) ) Y l ( ω ) \partial_r u(1,\omega)=(lv(1)+v'(1))Y_l(\omega) ∂ r u ( 1 , ω ) = ( l v ( 1 ) + v ′ ( 1 )) Y l ( ω ) . Similarly, for a > 0 a>0 a > 0 , the condition u ∣ S a = 0 u|_{S_a}=0 u ∣ S a = 0 is equivalent to v ( a ) = 0 v(a)=0 v ( a ) = 0 . Distinct basis elements are orthogonal with respect to both forms. For the bending form, this orthogonality follows directly from L 2 L^2 L 2 orthogonality. For the gradient form, we also use
∫ S d − 1 ∇ S Y l ⋅ ∇ S Y l ~ d S = l ~ ( l ~ + d − 2 ) ∫ S d − 1 Y l Y l ~ d S . \int_{S^{d-1}} \nabla_S Y_l \cdot\nabla_S Y_{\tilde l}\,\mathrm{d}S
= \tilde l(\tilde l+d-2)\int_{S^{d-1}}Y_lY_{\tilde l}\,\mathrm{d}S. ∫ S d − 1 ∇ S Y l ⋅ ∇ S Y l ~ d S = l ~ ( l ~ + d − 2 ) ∫ S d − 1 Y l Y l ~ d S .
By the completeness of the spherical harmonics, we obtain the spectral decomposition in (i).
Ball spectrum
Set ν = D / 2 \nu=D/2 ν = D /2 , and denote the spectral parameter by k 2 k^2 k 2 , k > 0 k>0 k > 0 . On a region without interior constraints, the radial weak equation is
( L D 2 + k 2 L D ) v = 0. (L_D^2+k^2L_D)v=0. ( L D 2 + k 2 L D ) v = 0.
The weak equation in effective dimension holds for every test function. For a radial solution, rotationally averaging a test function does not change its inner product with the equation. Hence interior elliptic regularity is available near the origin. For regular solutions there, ( L D + k 2 ) v (L_D+k^2)v ( L D + k 2 ) v is a regular radial harmonic function and hence must be constant. More precisely, its radial derivative is a constant multiple of r 1 − D r^{1-D} r 1 − D , and regularity at the origin forces this multiple to vanish. Solving the remaining inhomogeneous Helmholtz equation gives
v ( r ) = c 0 + c 1 r 1 − ν J ν − 1 ( k r ) . v(r)=c_0+c_1r^{1-\nu}J_{\nu-1}(kr). v ( r ) = c 0 + c 1 r 1 − ν J ν − 1 ( k r ) .
The regular Bessel solutions here are standard [23 pp. 38–45] . Interior regularity rules out the other local solutions, even if they have finite energy on the punctured ball.
The condition v ( 1 ) = 0 v(1)=0 v ( 1 ) = 0 yields
v ( r ) = c 1 [ r 1 − ν J ν − 1 ( k r ) − J ν − 1 ( k ) ] . v(r)=c_1\left[r^{1-\nu}J_{\nu-1}(kr)-J_{\nu-1}(k)\right]. v ( r ) = c 1 [ r 1 − ν J ν − 1 ( k r ) − J ν − 1 ( k ) ] .
The derivative recurrence formula [23 p. 45] gives
d d r [ r 1 − ν J ν − 1 ( k r ) ] = − k r 1 − ν J ν ( k r ) . (17) \frac{\mathrm{d}}{\mathrm{d}r}\left[r^{1-\nu}J_{\nu-1}(kr)\right]
=-kr^{1-\nu}J_\nu(kr). \tag*{(17)} d r d [ r 1 − ν J ν − 1 ( k r ) ] = − k r 1 − ν J ν ( k r ) . (17)
Thus for a nonzero mode, v ′ ( 1 ) = 0 v'(1)=0 v ′ ( 1 ) = 0 is equivalent to J ν ( k ) = 0 J_\nu(k)=0 J ν ( k ) = 0 . This establishes (3.2), which also appears in [10 Section 11.3] .
Transmission conditions
Away from a a a , the weak equation gives (12) . Since the weight near a a a is bounded above and below by positive constants, standard one-dimensional H 2 H^2 H 2 regularity implies that the traces of v v v and v ′ v' v ′ agree. The remaining conditions come from the weak equation itself rather than from any additional clamped boundary conditions.
On a subinterval I I I on which the solution is smooth, the identity r D − 1 L D w = ( r D − 1 w ′ ) ′ r^{D-1}L_Dw=(r^{D-1}w')' r D − 1 L D w = ( r D − 1 w ′ ) ′ can be integrated by parts twice. In the displays below, the boundary brackets always denote the value at the right endpoint minus the value at the left endpoint:
∫ I r D − 1 ( L D v ) ( L D w ) d r = [ r D − 1 ( L D v ) w ′ ] ∂ I − ∫ I r D − 1 ( L D v ) ′ w ′ d r = [ r D − 1 ( ( L D v ) w ′ − ( L D v ) ′ w ) ] ∂ I + ∫ I r D − 1 L D 2 v w d r , \begin{aligned}
\int_I r^{D-1}(L_Dv)(L_Dw)\,\mathrm{d}r
= [r^{D-1}(L_Dv)w']_{\partial I}
-\int_I r^{D-1}(L_Dv)'w'\,\mathrm{d}r \\
&= [r^{D-1}((L_Dv)w'-(L_Dv)'w)]_{\partial I}
+\int_I r^{D-1}L_D^2v\,w\,\mathrm{d}r,
\end{aligned} ∫ I r D − 1 ( L D v ) ( L D w ) d r = [ r D − 1 ( L D v ) w ′ ] ∂ I − ∫ I r D − 1 ( L D v ) ′ w ′ d r = [ r D − 1 (( L D v ) w ′ − ( L D v ) ′ w ) ] ∂ I + ∫ I r D − 1 L D 2 v w d r ,
− k 2 ∫ I r D − 1 v ′ w ′ d r = [ − k 2 r D − 1 v ′ w ] ∂ I + k 2 ∫ I r D − 1 L D v w d r . -k^2\int_I r^{D-1}v'w'\,\mathrm{d}r
=[-k^2r^{D-1}v'w]_{\partial I}
+k^2\int_I r^{D-1}L_Dv\,w\,\mathrm{d}r. − k 2 ∫ I r D − 1 v ′ w ′ d r = [ − k 2 r D − 1 v ′ w ] ∂ I + k 2 ∫ I r D − 1 L D v w d r .
Adding them yields the weighted Green formula
∫ I r D − 1 [ ( L D v ) ( L D w ) − k 2 v ′ w ′ ] d r = ∫ I r D − 1 ( L D 2 v + k 2 L D v ) w d r + [ r D − 1 ( ( L D v ) w ′ − ( L D v ) ′ w − k 2 v ′ w ) ] ∂ I . (18) \int_I r^{D-1}\left[(L_Dv)(L_Dw)-k^2v'w'\right]\,\mathrm{d}r
=\int_I r^{D-1}\left(L_D^2v+k^2L_Dv\right)w\,\mathrm{d}r
+[r^{D-1}((L_Dv)w'-(L_Dv)'w-k^2v'w)]_{\partial I}. \tag*{(18)} ∫ I r D − 1 [ ( L D v ) ( L D w ) − k 2 v ′ w ′ ] d r = ∫ I r D − 1 ( L D 2 v + k 2 L D v ) w d r + [ r D − 1 (( L D v ) w ′ − ( L D v ) ′ w − k 2 v ′ w ) ] ∂ I . (18)
At r = 1 r=1 r = 1 , both the test function and its derivative vanish. At the origin, the regular solution and the smooth radial test function satisfy
v ′ , w ′ , ( L D v ) ′ = O ( r ) , w , L D v = O ( 1 ) , v',w',(L_Dv)'=O(r), \qquad w,L_Dv=O(1), v ′ , w ′ , ( L D v ) ′ = O ( r ) , w , L D v = O ( 1 ) ,
so
r D − 1 ( ( L D v ) w ′ − ( L D v ) ′ w − k 2 v ′ w ) = O ( r D ) ⟶ 0. r^{D-1}((L_Dv)w'-(L_Dv)'w-k^2v'w)=O(r^D)\longrightarrow0. r D − 1 (( L D v ) w ′ − ( L D v ) ′ w − k 2 v ′ w ) = O ( r D ) ⟶ 0.
By density, this formula extends to admissible Sobolev test functions. At a a a , the contribution from the inner interval enters with a plus sign, and that from the outer interval enters with a minus sign. Since every admissible test function satisfies w ( a ) = 0 w(a)=0 w ( a ) = 0 , adding the two contributions leaves only
a D − 1 ( ( L D v ) ( a − ) − ( L D v ) ( a + ) ) w ′ ( a ) . a^{D-1}\big((L_Dv)(a-)-(L_Dv)(a+)\big)w'(a). a D − 1 ( ( L D v ) ( a − ) − ( L D v ) ( a + ) ) w ′ ( a ) .
Since w ′ ( a ) w'(a) w ′ ( a ) is arbitrary, this is equivalent to (13) . The term containing ( L D v ) ′ (L_Dv)' ( L D v ) ′ carries the zero factor w ( a ) w(a) w ( a ) . Conversely, (13) makes the boundary term in (18) vanish, which proves the weak equation. Figure 4 illustrates these conditions.
Figure 4. Schematic of the radial transmission: the slope is free at the common interface, and only the function value vanishes at a a a .
Put ν = D / 2 \nu=D/2 ν = D /2 . To prove the quantitative estimates in (iv), we consider the second order Bessel Dirichlet problem [23 Chapter III] .
Take ψ ( r ) = J ν ( λ r ) \psi(r)=J_{\nu}(\sqrt{\lambda}r) ψ ( r ) = J ν ( λ r ) . Its equation and boundary condition are
− ( r ψ ′ ) ′ + ν 2 r ψ = λ r ψ , ψ ( 1 ) = 0. -(r\psi')'+\frac{\nu^2}{r}\psi=\lambda r\psi,\qquad\psi(1)=0. − ( r ψ ′ ) ′ + r ν 2 ψ = λ r ψ , ψ ( 1 ) = 0.
Multiplying both sides by ψ \psi ψ and integrating yields the Rayleigh quotient
R ν ( ψ ) = ∫ 0 1 [ r ( ψ ′ ) 2 + ν 2 ψ 2 / r ] d r ∫ 0 1 r ψ 2 d r . \mathcal{R}_{\nu}(\psi)=\frac{\int_0^1\left[r(\psi')^2+\nu^2\psi^2/r\right]\,dr}{\int_0^1r\psi^2\,dr}. R ν ( ψ ) = ∫ 0 1 r ψ 2 d r ∫ 0 1 [ r ( ψ ′ ) 2 + ν 2 ψ 2 / r ] d r .
Near the origin the regular eigenfunction equals a constant times r ν r^\nu r ν , and there the boundary term in the integral vanishes. More generally, the form domain is the closure of C c ∞ ( 0 , 1 ) C^\infty_c(0,1) C c ∞ ( 0 , 1 ) with respect to the norms determined by the numerator and the denominator. For all positive orders this is the same space, since changing the positive coefficient in ∫ ψ 2 / r \int\psi^2/r ∫ ψ 2 / r yields equivalent norms. The Bessel comparison below applies for real order ν > 0 \nu>0 ν > 0 , while the upper bound in (3.5) uses ν ≥ 1 \nu\ge1 ν ≥ 1 . The standard local behaviors are r ν r^\nu r ν and r − ν r^{-\nu} r − ν [23 pp. 38–45, 57–64] , the latter being excluded because
∫ 0 t ( r − ν ) 2 r d r = ∫ 0 t r − 2 ν − 1 d r = + ∞ , ν > 0. \int_0^t\frac{(r^{-\nu})^2}{r}\,dr=\int_0^t r^{-2\nu-1}\,dr=+\infty,\qquad\nu>0. ∫ 0 t r ( r − ν ) 2 d r = ∫ 0 t r − 2 ν − 1 d r = + ∞ , ν > 0.
The embedding into L 2 ( ( 0 , 1 ) , r d r ) L^2((0,1),r\,dr) L 2 (( 0 , 1 ) , r d r ) is compact: ordinary compactness is used away from the origin, and
∫ 0 t r ψ 2 d r ≤ t 2 ∫ 0 t ψ 2 r d r ≤ t 2 ν 2 ∫ 0 1 [ r ( ψ ′ ) 2 + ν 2 r ψ 2 ] d r . \int_0^t r\psi^2\,dr\le t^2\int_0^t\frac{\psi^2}{r}\,dr\le\frac{t^2}{\nu^2}\int_0^1\left[r(\psi')^2+\frac{\nu^2}{r}\psi^2\right]\,dr. ∫ 0 t r ψ 2 d r ≤ t 2 ∫ 0 t r ψ 2 d r ≤ ν 2 t 2 ∫ 0 1 [ r ( ψ ′ ) 2 + r ν 2 ψ 2 ] d r .
Hence the discrete eigenvalues are exactly j ν , n 2 j_{\nu,n}^2 j ν , n 2 . This follows from regularity at the origin and from J ν ( λ ) = 0 J_\nu(\sqrt{\lambda})=0 J ν ( λ ) = 0 at 1.
Bounds for the first zero
For the upper bound in (3.5), take ψ ( r ) = r ν ( 1 − r 2 ) \psi(r)=r^\nu(1-r^2) ψ ( r ) = r ν ( 1 − r 2 ) , whose derivative is ψ ′ = ν r ν − 1 − ( ν + 2 ) r ν + 1 \psi'=\nu r^{\nu-1}-(\nu+2)r^{\nu+1} ψ ′ = ν r ν − 1 − ( ν + 2 ) r ν + 1 . This function is admissible. It vanishes at 1, and near the origin both energy densities are O ( r 2 ν − 1 ) O(r^{2\nu-1}) O ( r 2 ν − 1 ) , where 2 ν − 1 > − 1 2\nu-1>-1 2 ν − 1 > − 1 . Truncating it on [ 0 , t ] [0,t] [ 0 , t ] increases the energy by O ( t 2 ν ) → 0 O(t^{2\nu})\to0 O ( t 2 ν ) → 0 . The denominator is
∫ 0 1 r ψ 2 d r = 1 2 ν + 2 − 2 2 ν + 4 + 1 2 ν + 6 = 1 ( ν + 1 ) ( ν + 2 ) ( ν + 3 ) . \int_0^1 r\psi^2\,\mathrm{d}r=\frac{1}{2\nu+2}-\frac{2}{2\nu+4}+\frac{1}{2\nu+6}
=\frac{1}{(\nu+1)(\nu+2)(\nu+3)}. ∫ 0 1 r ψ 2 d r = 2 ν + 2 1 − 2 ν + 4 2 + 2 ν + 6 1 = ( ν + 1 ) ( ν + 2 ) ( ν + 3 ) 1 .
Expanding the two terms in the numerator gives
∫ 0 1 [ r ( ψ ′ ) 2 + ν 2 r ψ 2 ] d r = ∫ 0 1 [ 2 ν 2 r 2 ν − 1 − 4 ν ( ν + 1 ) r 2 ν + 1 + ( 2 ν 2 + 4 ν + 4 ) r 2 ν + 3 ] d r = ν − 2 ν + ν 2 + 2 ν + 2 ν + 2 = 2 ν + 2 . \int_0^1\left[r(\psi')^2+\frac{\nu^2}{r}\psi^2\right]\,\mathrm{d}r
=\int_0^1\left[2\nu^2r^{2\nu-1}-4\nu(\nu+1)r^{2\nu+1}+(2\nu^2+4\nu+4)r^{2\nu+3}\right]\,\mathrm{d}r
=\nu-2\nu+\frac{\nu^2+2\nu+2}{\nu+2}=\frac{2}{\nu+2}. ∫ 0 1 [ r ( ψ ′ ) 2 + r ν 2 ψ 2 ] d r = ∫ 0 1 [ 2 ν 2 r 2 ν − 1 − 4 ν ( ν + 1 ) r 2 ν + 1 + ( 2 ν 2 + 4 ν + 4 ) r 2 ν + 3 ] d r = ν − 2 ν + ν + 2 ν 2 + 2 ν + 2 = ν + 2 2 .
By the Rayleigh principle,
α D 2 ≤ 2 ( ν + 1 ) ( ν + 3 ) ≤ 4 ( ν + 1 ) 2 = ( D + 2 ) 2 , \alpha_D^2\leq2(\nu+1)(\nu+3)\leq4(\nu+1)^2=(D+2)^2, α D 2 ≤ 2 ( ν + 1 ) ( ν + 3 ) ≤ 4 ( ν + 1 ) 2 = ( D + 2 ) 2 ,
where the second inequality follows from ν ≥ 1 \nu\geq1 ν ≥ 1 . For the lower bound, any nonzero admissible function ψ \psi ψ satisfies
∫ 0 1 [ r ( ψ ′ ) 2 + ν 2 r ψ 2 ] d r − ν 2 ∫ 0 1 r ψ 2 d r = ∫ 0 1 r ( ψ ′ ) 2 d r + ν 2 ∫ 0 1 ( 1 r − r ) ψ 2 d r > 0. \int_0^1\left[r(\psi')^2+\frac{\nu^2}{r}\psi^2\right]\,\mathrm{d}r-\nu^2\int_0^1r\psi^2\,\mathrm{d}r
=\int_0^1r(\psi')^2\,\mathrm{d}r+\nu^2\int_0^1\left(\frac{1}{r}-r\right)\psi^2\,\mathrm{d}r>0. ∫ 0 1 [ r ( ψ ′ ) 2 + r ν 2 ψ 2 ] d r − ν 2 ∫ 0 1 r ψ 2 d r = ∫ 0 1 r ( ψ ′ ) 2 d r + ν 2 ∫ 0 1 ( r 1 − r ) ψ 2 d r > 0.
Applying this strict inequality to the first eigenfunction gives α D 2 > ν 2 \alpha_D^2>\nu^2 α D 2 > ν 2 , which is the lower bound in (3.5).
Order gap and radial gap
On the common form domain for positive orders,
R ν + 1 ( ψ ) − R ν ( ψ ) = ( 2 ν + 1 ) ∫ 0 1 ψ 2 / r d r ∫ 0 1 r ψ 2 d r ≥ 2 ν + 1. \mathcal{R}_{\nu+1}(\psi)-\mathcal{R}_{\nu}(\psi)=(2\nu+1)\frac{\int_0^1\psi^2/r\,\mathrm{d}r}{\int_0^1r\psi^2\,\mathrm{d}r}\geq2\nu+1. R ν + 1 ( ψ ) − R ν ( ψ ) = ( 2 ν + 1 ) ∫ 0 1 r ψ 2 d r ∫ 0 1 ψ 2 / r d r ≥ 2 ν + 1.
Maximizing first over each n n n -dimensional subspace and then minimizing proves (3.6).
We next consider (3.7). Let β = j ν + 1 , 1 \beta=j_{\nu+1,1} β = j ν + 1 , 1 and γ = j ν , 2 \gamma=j_{\nu,2} γ = j ν , 2 . The standard interlacing theorem [23 pp. 477–487] gives j ν , 1 < β < γ j_{\nu,1}<\beta<\gamma j ν , 1 < β < γ . To quantify the separation, set y ( x ) = x J ν ( x ) y(x)=\sqrt{x}J_\nu(x) y ( x ) = x J ν ( x ) . Substituting directly into the Bessel equation yields
y ′ ′ + [ 1 − ν 2 − 1 / 4 x 2 ] y = 0. y''+\left[1-\frac{\nu^2-1/4}{x^2}\right]y=0. y ′′ + [ 1 − x 2 ν 2 − 1/4 ] y = 0.
At x = β x=\beta x = β , by the recurrence formula J ν ′ = ν J ν / x − J ν + 1 J_\nu'=\nu J_\nu/x-J_{\nu+1} J ν ′ = ν J ν / x − J ν + 1 , one has
y ′ ( β ) y ( β ) = ν + 1 / 2 β > 0. \frac{y'(\beta)}{y(\beta)}=\frac{\nu+1/2}{\beta}>0. y ( β ) y ′ ( β ) = β ν + 1/2 > 0.
Interlacing implies y ( β ) ≠ 0 y(\beta) \ne0 y ( β ) = 0 . Since y ( γ ) = 0 y(\gamma)=0 y ( γ ) = 0 , multiplying the equation by y y y and integrating by parts gives
∫ β γ ( y ′ ) 2 d x = ∫ β γ [ 1 − ν 2 − 1 / 4 x 2 ] y 2 d x − y ( β ) y ′ ( β ) < ∫ β γ y 2 d x . \int_{\beta}^{\gamma} (y')^2\,\mathrm{d}x = \int_{\beta}^{\gamma} \left[1-\frac{\nu^2-1/4}{x^2}\right]y^2\,\mathrm{d}x-y(\beta)y'(\beta)
<\int_{\beta}^{\gamma} y^2\,\mathrm{d}x. ∫ β γ ( y ′ ) 2 d x = ∫ β γ [ 1 − x 2 ν 2 − 1/4 ] y 2 d x − y ( β ) y ′ ( β ) < ∫ β γ y 2 d x .
The last inequality follows from ν 2 − 1 / 4 > 0 \nu^2-1/4>0 ν 2 − 1/4 > 0 and y ( β ) y ′ ( β ) > 0 y(\beta)y'(\beta)>0 y ( β ) y ′ ( β ) > 0 . The one-sided Dirichlet Poincaré inequality is given by
∫ β γ y 2 d x ≤ 4 ( γ − β ) 2 π 2 ∫ β γ ( y ′ ) 2 d x . \int_{\beta}^{\gamma} y^2\,\mathrm{d}x \le\frac{4(\gamma-\beta)^2}{\pi^2}\int_{\beta}^{\gamma}(y')^2\,\mathrm{d}x. ∫ β γ y 2 d x ≤ π 2 4 ( γ − β ) 2 ∫ β γ ( y ′ ) 2 d x .
This constant can be obtained by extending y y y evenly about β \beta β . The extended function belongs to H 0 1 H_0^1 H 0 1 on an interval of length 2 ( γ − β ) 2(\gamma-\beta) 2 ( γ − β ) , and both integrals are doubled. Combining the two inequalities above forces
γ − β > π / 2 > 1. \gamma-\beta>\pi/2>1. γ − β > π /2 > 1.
Finally, applying (14) at order ν + 1 \nu+1 ν + 1 gives β > ν + 1 \beta>\nu+1 β > ν + 1 , and hence
γ 2 − β 2 = ( γ − β ) ( γ + β ) > 2 β > 2 ν + 2. \gamma^2-\beta^2=(\gamma-\beta)(\gamma+\beta)>2\beta>2\nu+2. γ 2 − β 2 = ( γ − β ) ( γ + β ) > 2 β > 2 ν + 2.
This proves (16) and completes the proof of the proposition.
A uniform expansion near the outer boundary
The ratio comparison requires a second order expansion with a controlled remainder. We obtain both from a scalar matching equation and a uniform complex-analytic estimate.
Lemma 3.2. For every integer D ≥ 2 D\ge2 D ≥ 2 , the normalized first radial branch r D ( h ) = ρ D ( 1 − h ) / α D 2 r_D(h)=\rho_D(1-h)/\alpha_D^2 r D ( h ) = ρ D ( 1 − h ) / α D 2 admits a holomorphic (complex analytic) continuation to a neighborhood of the closed disk
∣ h ∣ ≤ R D : = 1 256 ( D + 2 ) (19) |h|\le R_D:=\frac{1}{256(D+2)} \tag*{(19)} ∣ h ∣ ≤ R D := 256 ( D + 2 ) 1 (19) and satisfies ∣ r D ( h ) ∣ ≤ 2 |r_D(h)|\le2 ∣ r D ( h ) ∣ ≤ 2 on this disk. Moreover,
r D ( h ) = 1 + 3 2 h + D + 26 16 h 2 + R D ( h ) , r_D(h)=1+\frac{3}{2}h+\frac{D+26}{16}h^2+R_D(h), r D ( h ) = 1 + 2 3 h + 16 D + 26 h 2 + R D ( h ) , ∣ R D ( h ) ∣ ≤ 4 ∣ h ∣ 3 R D 3 ( ∣ h ∣ ≤ R D / 2 ) . (20) |R_D(h)|\le\frac{4|h|^3}{R_D^3}\qquad(|h|\le R_D/2). \tag*{(20)} ∣ R D ( h ) ∣ ≤ R D 3 4∣ h ∣ 3 ( ∣ h ∣ ≤ R D /2 ) . (20)
Proof. Set ν = D / 2 \nu=D/2 ν = D /2 , α = α D = j ν , 1 \alpha=\alpha_D=j_{\nu,1} α = α D = j ν , 1 , b = D − 1 b=D-1 b = D − 1 , and a = 1 − h a=1-h a = 1 − h . We first construct the eigenvalue branch before verifying that it is the first radial eigenvalue. For complex h h h , the construction involves only analytic formulas on fixed parameter intervals and does not require a domain in the complex radius variable.
Normalize the slope and choose the matching quantity.
By Proposition 3.1 (iii), it suffices to match (13) . We normalize the inner and outer solutions by requiring that
v i n ( a ) = v o u t ( a ) = 0 , v i n ′ ( a ) = v o u t ′ ( a ) = 1. v_{\mathrm{in}}(a)=v_{\mathrm{out}}(a)=0,\qquad v_{\mathrm{in}}'(a)=v_{\mathrm{out}}'(a)=1. v in ( a ) = v out ( a ) = 0 , v in ′ ( a ) = v out ′ ( a ) = 1.
Thus the first two transmission conditions hold automatically, and only the equality of L D v L_Dv L D v remains to be imposed. More generally, the quotient L D v / v ′ L_Dv/v' L D v / v ′ is unchanged when a solution is multiplied by a nonzero constant. It represents the bending quantity L D v L_Dv L D v corresponding to an interface slope normalized to one and is a convenient scalar matching quantity. Write the eigenvalue as k 2 k^2 k 2 . On the inner interval, regularity at the origin leaves only the Bessel solution and a constant. Subtracting the value at a a a and normalizing the slope gives
v i n ( r ) = C [ r 1 − ν J ν − 1 ( k r ) − a 1 − ν J ν − 1 ( k a ) ] , C = − [ k a 1 − ν J ν ( k a ) ] − 1 . (21) v_{\mathrm{in}}(r)=C\left[r^{1-\nu}J_{\nu-1}(kr)-a^{1-\nu}J_{\nu-1}(ka)\right],
\qquad
C=-\left[ka^{1-\nu}J_\nu(ka)\right]^{-1}. \tag*{(21)} v in ( r ) = C [ r 1 − ν J ν − 1 ( k r ) − a 1 − ν J ν − 1 ( k a ) ] , C = − [ k a 1 − ν J ν ( k a ) ] − 1 . (21)
From (17) and the Bessel equation we obtain
L D v i n ( a ) v i n ′ ( a ) = k J ν − 1 ( k a ) J ν ( k a ) = 1 a H ( k a ) , H ( t ) : = t J ν ′ ( t ) J ν ( t ) + ν . \frac{L_Dv_{\mathrm{in}}(a)}{v_{\mathrm{in}}'(a)}
=k\frac{J_{\nu-1}(ka)}{J_\nu(ka)}
=\frac{1}{a}\mathcal{H}(ka),
\qquad
\mathcal{H}(t):=t\frac{J_\nu'(t)}{J_\nu(t)}+\nu. v in ′ ( a ) L D v in ( a ) = k J ν ( k a ) J ν − 1 ( k a ) = a 1 H ( k a ) , H ( t ) := t J ν ( t ) J ν ′ ( t ) + ν .
Here J ν − 1 ( t ) = J ν ′ ( t ) + ν J ν ( t ) / t J_{\nu-1}(t)=J_\nu'(t)+\nu J_\nu(t)/t J ν − 1 ( t ) = J ν ′ ( t ) + ν J ν ( t ) / t [23 p. 45] . Hence H \mathcal{H} H is a a a times the interface quotient. To motivate this choice of the spectral parameter, let e = k a − α e=ka-\alpha e = k a − α . Since the zero α \alpha α of J ν J_\nu J ν is simple, the leading term of H ( α + e ) \mathcal{H}(\alpha+e) H ( α + e ) is α / e \alpha/e α / e . On the outer interval of length h h h , a function with unit slope at the interface has second derivatives of order h − 1 h^{-1} h − 1 . Matching these quantities implies e = O ( h ) e=O(h) e = O ( h ) , and hence we write
e = α h z , k = α 1 + h z 1 − h . e=\alpha h z,\qquad k=\alpha\frac{1+hz}{1-h}. e = α h z , k = α 1 − h 1 + h z .
The leading-order outer profile computed below has scaled second derivative − 4 -4 − 4 at the interface, so the leading-order matching equation is 1 / z = − 4 1/z=-4 1/ z = − 4 . This is why we work in the fixed region
∣ z + 1 4 ∣ ≤ 1 8 , 1 8 ≤ ∣ z ∣ ≤ 3 8 . \left|z+\frac{1}{4}\right|\leq\frac{1}{8},\qquad\frac{1}{8}\leq|z|\leq\frac{3}{8}. z + 4 1 ≤ 8 1 , 8 1 ≤ ∣ z ∣ ≤ 8 3 .
The radius 1 / 8 1/8 1/8 is half the distance from the center of the disk to the pole z = 0 z=0 z = 0 . The whole disk therefore remains sufficiently far from this pole while containing a fixed neighborhood of the expected root. We show later that this disk indeed contains a root. The center of the disk itself is not an assumption about the eigenvalue.
The small radius R D = 1 / [ 256 ( D + 2 ) ] R_D=1/[256(D+2)] R D = 1/ [ 256 ( D + 2 )] has a similarly concrete role. The outer equation is controlled by ( D + 2 ) ∣ h ∣ (D+2)|h| ( D + 2 ) ∣ h ∣ , and choosing a radius proportional to ( D + 2 ) − 1 (D+2)^{-1} ( D + 2 ) − 1 makes the perturbation uniformly small in D D D . The factor 256 ensures that the perturbation of the inverse operator does not exceed 1 / 32 1/32 1/32 and that the matching error does not exceed 1 / 2 1/2 1/2 . These are convenient strict margins rather than sharp thresholds.
Step 2. Bound the inner logarithmic derivative uniformly.
We need a bound near α \alpha α that does not involve the possibly small quantity J ν ′ ( α ) J_\nu'(\alpha) J ν ′ ( α ) . We normalize this factor away by setting
y ( e ) = J ν ( α + e ) J ν ′ ( α ) . y(e) = \frac{J_\nu(\alpha+e)}{J_\nu'(\alpha)}. y ( e ) = J ν ′ ( α ) J ν ( α + e ) .
The positive zeros are simple [23 pp. 477–487] , and the denominator is nonzero. The equation and initial conditions are
y ′ ′ + 1 α + e y ′ + ( 1 − ν 2 ( α + e ) 2 ) y = 0 , y ( 0 ) = 0 , y ′ ( 0 ) = 1. y''+\frac{1}{\alpha+e}y'+\left(1-\frac{\nu^2}{(\alpha+e)^2}\right)y=0,\qquad y(0)=0,\qquad y'(0)=1. y ′′ + α + e 1 y ′ + ( 1 − ( α + e ) 2 ν 2 ) y = 0 , y ( 0 ) = 0 , y ′ ( 0 ) = 1.
The analytic local solution provides the required expression. It also applies for noninteger ν \nu ν , with the branch obtained by continuation from the positive point α \alpha α . For ∣ e ∣ ≤ 1 / 16 |e|\leq1/16 ∣ e ∣ ≤ 1/16 , the lower bound α > ν ≥ 1 \alpha>\nu\geq1 α > ν ≥ 1 yields
∣ 1 α + e ∣ ≤ 16 15 < 2 , ∣ 1 − ν 2 ( α + e ) 2 ∣ ≤ 1 + ( 16 15 ) 2 < 3. \left|\frac{1}{\alpha+e}\right|\leq\frac{16}{15}<2,\qquad\left|1-\frac{\nu^2}{(\alpha+e)^2}\right|\leq1+\left(\frac{16}{15}\right)^2<3. α + e 1 ≤ 15 16 < 2 , 1 − ( α + e ) 2 ν 2 ≤ 1 + ( 15 16 ) 2 < 3.
In the first order system for ( y , y ′ ) (y,y') ( y , y ′ ) , the maximum row sum norm of the coefficient matrix is at most 5 5 5 . Integrating along the line segment from 0 0 0 to e e e and then applying elementary integral inequalities for the system gives
max { ∣ y ( e ) ∣ , ∣ y ′ ( e ) ∣ } ≤ exp ( 5 ∣ e ∣ ) < 2 , ∣ y ′ ′ ( e ) ∣ ≤ 2 ∣ y ′ ( e ) ∣ + 3 ∣ y ( e ) ∣ ≤ 10. \max\{|y(e)|,|y'(e)|\}\leq\exp(5|e|)<2,\qquad|y''(e)|\leq2|y'(e)|+3|y(e)|\leq10. max { ∣ y ( e ) ∣ , ∣ y ′ ( e ) ∣ } ≤ exp ( 5∣ e ∣ ) < 2 , ∣ y ′′ ( e ) ∣ ≤ 2∣ y ′ ( e ) ∣ + 3∣ y ( e ) ∣ ≤ 10.
Here the first bound follows by parametrizing the segment as t e te t e (0 ≤ t ≤ 1 0\leq t\leq1 0 ≤ t ≤ 1 ) and bounding the integral kernel by 5 ∣ e ∣ 5|e| 5∣ e ∣ . Write the simple zero as y ( e ) = e A ( e ) y(e)=eA(e) y ( e ) = e A ( e ) , where A ( 0 ) = 1 A(0)=1 A ( 0 ) = 1 . The Taylor integral formula together with the identity A ( e ) = ∫ 0 1 y ′ ( t e ) d t A(e)=\int_0^1 y'(te)\,dt A ( e ) = ∫ 0 1 y ′ ( t e ) d t gives
A ( e ) − 1 = e ∫ 0 1 ( 1 − t ) y ′ ′ ( t e ) d t , A ′ ( e ) = ∫ 0 1 t y ′ ′ ( t e ) d t . A(e)-1=e\int_0^1(1-t)y''(te)\,dt,\qquad A'(e)=\int_0^1ty''(te)\,dt. A ( e ) − 1 = e ∫ 0 1 ( 1 − t ) y ′′ ( t e ) d t , A ′ ( e ) = ∫ 0 1 t y ′′ ( t e ) d t .
Therefore,
∣ A − 1 ∣ ≤ 5 ∣ e ∣ , ∣ A ′ ∣ ≤ 5 , ∣ A ′ A ∣ < 8. (22) |A-1|\leq5|e|,\qquad|A'|\leq5,\qquad\left|\frac{A'}{A}\right|<8. \tag*{(22)} ∣ A − 1∣ ≤ 5∣ e ∣ , ∣ A ′ ∣ ≤ 5 , A A ′ < 8. (22)
For the last inequality,
∣ A ∣ ≥ 1 − 5 16 = 11 16 , ∣ A ′ A ∣ ≤ 5 11 / 16 = 80 11 < 8. |A|\geq1-\frac{5}{16}=\frac{11}{16},\qquad\left|\frac{A'}{A}\right|\leq\frac{5}{11/16}=\frac{80}{11}<8. ∣ A ∣ ≥ 1 − 16 5 = 16 11 , A A ′ ≤ 11/16 5 = 11 80 < 8.
In particular, A A A has no zero in the disk under consideration.
If ∣ h ∣ ≤ R D |h|\leq R_D ∣ h ∣ ≤ R D and ∣ z + 1 / 4 ∣ ≤ 1 / 8 |z+1/4|\leq1/8 ∣ z + 1/4∣ ≤ 1/8 , then
∣ e ∣ = ∣ α h z ∣ ≤ ( D + 2 ) R D 3 8 = 3 2048 < 1 16 . |e|=|\alpha hz|\leq(D+2)R_D\frac{3}{8}=\frac{3}{2048}<\frac{1}{16}. ∣ e ∣ = ∣ α h z ∣ ≤ ( D + 2 ) R D 8 3 = 2048 3 < 16 1 .
Hence the estimates above apply on the entire parameter set. For h ≠ 0 h \ne0 h = 0 , since z ≠ 0 z \ne0 z = 0 and A ( e ) ≠ 0 \mathcal{A}(e) \ne0 A ( e ) = 0 , we have J ν ( k a ) ≠ 0 J_{\nu}(ka) \ne0 J ν ( k a ) = 0 , which justifies the inner normalization in Step 1. Moreover,
H ( α + e ) = ( α + e ) ( 1 e + A ′ ( e ) A ( e ) ) + ν , \mathcal{H}(\alpha+e)=(\alpha+e)\left(\frac{1}{e}+\frac{\mathcal{A}'(e)}{\mathcal{A}(e)}\right)+\nu, H ( α + e ) = ( α + e ) ( e 1 + A ( e ) A ′ ( e ) ) + ν ,
h H ( α + α h z ) − 1 z = h ( 1 + ν ) + h ( α + e ) A ′ ( e ) A ( e ) . h\mathcal{H}(\alpha+\alpha hz)-\frac{1}{z}=h(1+\nu)+h(\alpha+e)\frac{\mathcal{A}'(e)}{\mathcal{A}(e)}. h H ( α + α h z ) − z 1 = h ( 1 + ν ) + h ( α + e ) A ( e ) A ′ ( e ) .
From α ≤ D + 2 \alpha\le D+2 α ≤ D + 2 and ∣ e ∣ ≤ 1 / 16 |e|\le1/16 ∣ e ∣ ≤ 1/16 , we obtain
∣ h H ( α + α h z ) − 1 z ∣ ≤ [ 1 + D 2 + 8 ( D + 2 + 1 16 ) ] ∣ h ∣ ≤ 10 ( D + 2 ) ∣ h ∣ . \left|h\mathcal{H}(\alpha+\alpha hz)-\frac{1}{z}\right|\le\left[1+\frac{D}{2}+8\left(D+2+\frac{1}{16}\right)\right]|h|\le10(D+2)|h|. h H ( α + α h z ) − z 1 ≤ [ 1 + 2 D + 8 ( D + 2 + 16 1 ) ] ∣ h ∣ ≤ 10 ( D + 2 ) ∣ h ∣.
This exact identity also shows that after multiplication by h h h , the singularity at h = 0 h=0 h = 0 disappears. It is precisely this regularization that makes a uniformly analytic matching equation possible.
Rescale the outer interval to a fixed interval.
Set
x = r − a h , v o u t ( r ) = h f ( x ) , r = 1 − h + h x . x=\frac{r-a}{h},\qquad v_{\mathrm{out}}(r)=hf(x),\qquad r=1-h+hx. x = h r − a , v out ( r ) = h f ( x ) , r = 1 − h + h x .
The factor h h h is chosen so that v o u t ′ ( r ) = f ′ ( x ) v_{\mathrm{out}}'(r)=f'(x) v out ′ ( r ) = f ′ ( x ) , and the normalization of the interface slope is therefore independent of h h h . The four endpoint conditions become
f ( 0 ) = 0 , f ′ ( 0 ) = 1 , f ( 1 ) = 0 , f ′ ( 1 ) = 0. f(0)=0,\qquad f'(0)=1,\qquad f(1)=0,\qquad f'(1)=0. f ( 0 ) = 0 , f ′ ( 0 ) = 1 , f ( 1 ) = 0 , f ′ ( 1 ) = 0.
To derive the equation, we first expand the radial operator before scaling:
L D 2 v = v ′ ′ ′ ′ + 2 b r v ′ ′ ′ + b ( b − 2 ) r 2 v ′ ′ − b ( b − 2 ) r 3 v ′ . L_D^2v=v^{\prime\prime\prime\prime}+\frac{2b}{r}v^{\prime\prime\prime}+\frac{b(b-2)}{r^2}v^{\prime\prime}-\frac{b(b-2)}{r^3}v'. L D 2 v = v ′′′′ + r 2 b v ′′′ + r 2 b ( b − 2 ) v ′′ − r 3 b ( b − 2 ) v ′ .
Using the relations v ′ = f ′ v'=f' v ′ = f ′ , v ′ ′ = h − 1 f ′ ′ v''=h^{-1}f'' v ′′ = h − 1 f ′′ , v ′ ′ ′ = h − 2 f ′ ′ ′ v'''=h^{-2}f''' v ′′′ = h − 2 f ′′′ , and v ′ ′ ′ ′ = h − 3 f ′ ′ ′ ′ v''''=h^{-3}f'''' v ′′′′ = h − 3 f ′′′′ , we multiply L D 2 v + k 2 L D v = 0 L_D^2v+k^2L_Dv=0 L D 2 v + k 2 L D v = 0 by h 3 h^3 h 3 and obtain
0 = f ′ ′ ′ ′ + 2 b h r f ′ ′ ′ + h 2 ( b ( b − 2 ) r 2 + k 2 ) f ′ ′ + b h 3 ( k 2 r − b − 2 r 3 ) f ′ , r = 1 − h + h x . (23) \begin{aligned}
0={}&f^{\prime\prime\prime\prime}+\frac{2bh}{r}f^{\prime\prime\prime}+h^2\left(\frac{b(b-2)}{r^2}+k^2\right)f''\\
&+bh^3\left(\frac{k^2}{r}-\frac{b-2}{r^3}\right)f',\qquad r=1-h+hx. \tag*{(23)}
\end{aligned} 0 = f ′′′′ + r 2 bh f ′′′ + h 2 ( r 2 b ( b − 2 ) + k 2 ) f ′′ + b h 3 ( r k 2 − r 3 b − 2 ) f ′ , r = 1 − h + h x . (23)
At h = 0 h=0 h = 0 the equation reduces to f ′ ′ ′ ′ = 0 f^{\prime\prime\prime\prime}=0 f ′′′′ = 0 , and together with the endpoint conditions it determines the cubic polynomial
f 0 ( x ) = x ( 1 − x ) 2 = x − 2 x 2 + x 3 , f 0 ′ ′ ( 0 ) = − 4 , f 0 ′ ′ ′ = 6. (24) f_0(x)=x(1-x)^2=x-2x^2+x^3,\qquad f_0''(0)=-4,\qquad f_0'''=6. \tag*{(24)} f 0 ( x ) = x ( 1 − x ) 2 = x − 2 x 2 + x 3 , f 0 ′′ ( 0 ) = − 4 , f 0 ′′′ = 6. (24)
Thus the outer problem is a small perturbation of an explicitly invertible fourth order equation on [ 0 , 1 ] [0,1] [ 0 , 1 ] .
Construct and estimate the outer inverse
We first collect the constants needed for the perturbation argument. If w ′ ′ ′ = p w'''=p w ′′′ = p and w = w ′ = 0 w=w'=0 w = w ′ = 0 at both endpoints, integrating from zero gives
w ( x ) = w ′ ′ ( 0 ) x 2 2 + w ′ ′ ′ ( 0 ) x 3 6 + 1 6 ∫ 0 x ( x − t ) 3 p ( t ) d t . w(x)=\frac{w''(0)x^2}{2}+\frac{w'''(0)x^3}{6}+\frac{1}{6}\int_0^x(x-t)^3p(t)\,\mathrm{d}t. w ( x ) = 2 w ′′ ( 0 ) x 2 + 6 w ′′′ ( 0 ) x 3 + 6 1 ∫ 0 x ( x − t ) 3 p ( t ) d t .
and the remaining two initial derivatives are determined by the conditions at x = 1 x=1 x = 1 :
3 w ′ ′ ( 0 ) + w ′ ′ ′ ( 0 ) = − ∫ 0 1 ( 1 − t ) 3 p ( t ) d t , 3w''(0)+w'''(0)=-\int_0^1(1-t)^3p(t)\,\mathrm{d}t, 3 w ′′ ( 0 ) + w ′′′ ( 0 ) = − ∫ 0 1 ( 1 − t ) 3 p ( t ) d t ,
2 w ′ ′ ( 0 ) + w ′ ′ ′ ( 0 ) = − ∫ 0 1 ( 1 − t ) 2 p ( t ) d t . 2w''(0)+w'''(0)=-\int_0^1(1-t)^2p(t)\,\mathrm{d}t. 2 w ′′ ( 0 ) + w ′′′ ( 0 ) = − ∫ 0 1 ( 1 − t ) 2 p ( t ) d t .
Solving this 2 × 2 2\times2 2 × 2 system gives
w ′ ′ ( 0 ) = ∫ 0 1 t ( 1 − t ) 2 p ( t ) d t , w''(0)=\int_0^1t(1-t)^2p(t)\,\mathrm{d}t, w ′′ ( 0 ) = ∫ 0 1 t ( 1 − t ) 2 p ( t ) d t ,
w ′ ′ ′ ( 0 ) = − ∫ 0 1 ( 1 − t ) 2 ( 1 + 2 t ) p ( t ) d t . w'''(0)=-\int_0^1(1-t)^2(1+2t)p(t)\,\mathrm{d}t. w ′′′ ( 0 ) = − ∫ 0 1 ( 1 − t ) 2 ( 1 + 2 t ) p ( t ) d t .
The kernel integrals are
∫ 0 1 t ( 1 − t ) 2 d t = 1 2 − 2 3 + 1 4 = 1 12 , \int_0^1t(1-t)^2\,\mathrm{d}t=\frac{1}{2}-\frac{2}{3}+\frac{1}{4}=\frac{1}{12}, ∫ 0 1 t ( 1 − t ) 2 d t = 2 1 − 3 2 + 4 1 = 12 1 ,
∫ 0 1 ( 1 − t ) 2 ( 1 + 2 t ) d t = ∫ 0 1 ( 1 − 3 t 2 + 2 t 3 ) d t = 1 − 1 + 1 2 = 1 2 . \int_0^1(1-t)^2(1+2t)\,\mathrm{d}t=\int_0^1(1-3t^2+2t^3)\,\mathrm{d}t=1-1+\frac{1}{2}=\frac{1}{2}. ∫ 0 1 ( 1 − t ) 2 ( 1 + 2 t ) d t = ∫ 0 1 ( 1 − 3 t 2 + 2 t 3 ) d t = 1 − 1 + 2 1 = 2 1 .
Since these kernels are nonnegative,
∣ w ′ ′ ( 0 ) ∣ ≤ 1 12 ∥ p ∥ ∞ , ∣ w ′ ′ ′ ( 0 ) ∣ ≤ 1 2 ∥ p ∥ ∞ . \lvert w''(0)\rvert\leq\frac{1}{12}\lVert p\rVert_\infty,\qquad\lvert w'''(0)\rvert\leq\frac{1}{2}\lVert p\rVert_\infty. ∣ w ′′ ( 0 )∣ ≤ 12 1 ∥ p ∥ ∞ , ∣ w ′′′ ( 0 )∣ ≤ 2 1 ∥ p ∥ ∞ .
Differentiating the integral formula yields
w ′ ( x ) = w ′ ′ ( 0 ) x + w ′ ′ ′ ( 0 ) x 2 / 2 + 1 2 ∫ 0 x ( x − t ) 2 p ( t ) d t , w'(x)=w''(0)x+w'''(0)x^2/2+\frac{1}{2}\int_0^x(x-t)^2p(t)\,\mathrm{d}t, w ′ ( x ) = w ′′ ( 0 ) x + w ′′′ ( 0 ) x 2 /2 + 2 1 ∫ 0 x ( x − t ) 2 p ( t ) d t ,
w ′ ′ ( x ) = w ′ ′ ( 0 ) + w ′ ′ ′ ( 0 ) x + ∫ 0 x ( x − t ) p ( t ) d t , w''(x)=w''(0)+w'''(0)x+\int_0^x(x-t)p(t)\,\mathrm{d}t, w ′′ ( x ) = w ′′ ( 0 ) + w ′′′ ( 0 ) x + ∫ 0 x ( x − t ) p ( t ) d t ,
w ′ ′ ′ ( x ) = w ′ ′ ′ ( 0 ) + ∫ 0 x p ( t ) d t . w'''(x)=w'''(0)+\int_0^xp(t)\,\mathrm{d}t. w ′′′ ( x ) = w ′′′ ( 0 ) + ∫ 0 x p ( t ) d t .
On 0 ≤ x ≤ 1 0\leq x\leq1 0 ≤ x ≤ 1 we estimate term by term:
∥ w ∥ ∞ ≤ ( 1 24 + 1 12 + 1 24 ) ∥ p ∥ ∞ = 1 6 ∥ p ∥ ∞ , \lVert w\rVert_\infty\leq\left(\frac{1}{24}+\frac{1}{12}+\frac{1}{24}\right)\lVert p\rVert_\infty=\frac{1}{6}\lVert p\rVert_\infty, ∥ w ∥ ∞ ≤ ( 24 1 + 12 1 + 24 1 ) ∥ p ∥ ∞ = 6 1 ∥ p ∥ ∞ ,
∥ w ′ ∥ ∞ ≤ ( 1 12 + 1 4 + 1 6 ) ∥ p ∥ ∞ = 1 2 ∥ p ∥ ∞ , \lVert w'\rVert_\infty\leq\left(\frac{1}{12}+\frac{1}{4}+\frac{1}{6}\right)\lVert p\rVert_\infty=\frac{1}{2}\lVert p\rVert_\infty, ∥ w ′ ∥ ∞ ≤ ( 12 1 + 4 1 + 6 1 ) ∥ p ∥ ∞ = 2 1 ∥ p ∥ ∞ ,
∥ w ′ ′ ∥ ∞ ≤ ( 1 12 + 1 2 + 1 2 ) ∥ p ∥ ∞ = 13 12 ∥ p ∥ ∞ , \lVert w''\rVert_\infty\leq\left(\frac{1}{12}+\frac{1}{2}+\frac{1}{2}\right)\lVert p\rVert_\infty=\frac{13}{12}\lVert p\rVert_\infty, ∥ w ′′ ∥ ∞ ≤ ( 12 1 + 2 1 + 2 1 ) ∥ p ∥ ∞ = 12 13 ∥ p ∥ ∞ ,
∥ w ′ ′ ′ ∥ ∞ ≤ ( 1 2 + 1 ) ∥ p ∥ ∞ = 3 2 ∥ p ∥ ∞ . \lVert w'''\rVert_\infty\leq\left(\frac{1}{2}+1\right)\lVert p\rVert_\infty=\frac{3}{2}\lVert p\rVert_\infty. ∥ w ′′′ ∥ ∞ ≤ ( 2 1 + 1 ) ∥ p ∥ ∞ = 2 3 ∥ p ∥ ∞ .
Let G G G denote this inverse operator with homogeneous endpoint data. The norm on C j ( [ 0 , 1 ] ) C^j([0,1]) C j ([ 0 , 1 ]) is taken to be ∥ w ∥ C j = max 0 ≤ i ≤ j ∥ w ( i ) ∥ ∞ \lVert w\rVert_{C^j}=\max_{0\le i\le j}\lVert w^{(i)}\rVert_\infty ∥ w ∥ C j = max 0 ≤ i ≤ j ∥ w ( i ) ∥ ∞ . From the bounds above and w ′ ′ ′ ′ = p w^{\prime\prime\prime\prime}=p w ′′′′ = p , we have
∥ G ∥ C 0 → C 3 ≤ 3 2 , ∥ G ∥ C 0 → C 4 ≤ max { 3 2 , 1 } = 3 2 . \lVert G\rVert_{C^0\to C^3}\le\frac{3}{2},\qquad\lVert G\rVert_{C^0\to C^4}\le\max\left\{\frac{3}{2},1\right\}=\frac{3}{2}. ∥ G ∥ C 0 → C 3 ≤ 2 3 , ∥ G ∥ C 0 → C 4 ≤ max { 2 3 , 1 } = 2 3 .
These estimates also apply to complex-valued functions.
Put u = ( D + 2 ) ∣ h ∣ ≤ 1 / 256 u=(D+2)|h|\le1/256 u = ( D + 2 ) ∣ h ∣ ≤ 1/256 , and let c 3 , c 2 , c 1 c_3,c_2,c_1 c 3 , c 2 , c 1 denote the coefficients of f ′ ′ ′ f^{\prime\prime\prime} f ′′′ , f ′ ′ f^{\prime\prime} f ′′ , f ′ f' f ′ in (23) , respectively. Uniformly for x ∈ [ 0 , 1 ] x\in[0,1] x ∈ [ 0 , 1 ] and ∣ z + 1 / 4 ∣ ≤ 1 / 8 |z+1/4|\le1/8 ∣ z + 1/4∣ ≤ 1/8 ,
∣ r ∣ ≥ 1 − ∣ h ∣ ≥ 1 2 , ∣ k ∣ ≤ α 1 + 3 ∣ h ∣ / 8 1 − ∣ h ∣ ≤ 2 ( D + 2 ) . |r|\ge1-|h|\ge\frac{1}{2},\qquad|k|\le\alpha\frac{1+3|h|/8}{1-|h|}\le2(D+2). ∣ r ∣ ≥ 1 − ∣ h ∣ ≥ 2 1 , ∣ k ∣ ≤ α 1 − ∣ h ∣ 1 + 3∣ h ∣/8 ≤ 2 ( D + 2 ) .
Since b ≤ D + 2 b\le D+2 b ≤ D + 2 and ∣ b − 2 ∣ ≤ D + 2 |b-2|\le D+2 ∣ b − 2∣ ≤ D + 2 ,
∣ c 3 ∣ ≤ 4 u , |c_3|\le4u, ∣ c 3 ∣ ≤ 4 u ,
∣ c 2 ∣ ≤ ∣ h ∣ 2 [ 4 ∣ b ( b − 2 ) ∣ + 4 ( D + 2 ) 2 ] ≤ 8 u 2 , |c_2|\le|h|^2\left[4|b(b-2)|+4(D+2)^2\right]\le8u^2, ∣ c 2 ∣ ≤ ∣ h ∣ 2 [ 4∣ b ( b − 2 ) ∣ + 4 ( D + 2 ) 2 ] ≤ 8 u 2 ,
∣ c 1 ∣ ≤ b ∣ h ∣ 3 [ 8 ( D + 2 ) 2 + 8 ∣ b − 2 ∣ ] ≤ 16 u 3 . |c_1|\le b|h|^3\left[8(D+2)^2+8|b-2|\right]\le16u^3. ∣ c 1 ∣ ≤ b ∣ h ∣ 3 [ 8 ( D + 2 ) 2 + 8∣ b − 2∣ ] ≤ 16 u 3 .
Their sum satisfies
∣ c 3 ∣ + ∣ c 2 ∣ + ∣ c 1 ∣ ≤ u ( 4 + 8 u + 16 u 2 ) ≤ u ( 4 + 1 32 + 1 4096 ) ≤ 5 u . |c_3|+|c_2|+|c_1|\le u(4+8u+16u^2)\le u\left(4+\frac{1}{32}+\frac{1}{4096}\right)\le5u. ∣ c 3 ∣ + ∣ c 2 ∣ + ∣ c 1 ∣ ≤ u ( 4 + 8 u + 16 u 2 ) ≤ u ( 4 + 32 1 + 4096 1 ) ≤ 5 u .
Substituting f = f 0 + w f=f_0+w f = f 0 + w into the equation and the endpoint conditions gives
f = f 0 − G ( c 3 f ′ ′ ′ + c 2 f ′ ′ + c 1 f ′ ) . f=f_0-G\left(c_3f^{\prime\prime\prime}+c_2f^{\prime\prime}+c_1f'\right). f = f 0 − G ( c 3 f ′′′ + c 2 f ′′ + c 1 f ′ ) .
The operator bound is
∥ G ( c 3 f ′ ′ ′ + c 2 f ′ ′ + c 1 f ′ ) ∥ C 3 ≤ 3 2 ( 4 u + 8 u 2 + 16 u 3 ) ∥ f ∥ C 3 ≤ 15 2 u ∥ f ∥ C 3 ≤ 8 u ∥ f ∥ C 3 . \begin{aligned}
\left\lVert G\left(c_3f^{\prime\prime\prime}+c_2f^{\prime\prime}+c_1f'\right)\right\rVert_{C^3}
&\le\frac{3}{2}(4u+8u^2+16u^3)\lVert f\rVert_{C^3}\\
&\le\frac{15}{2}u\lVert f\rVert_{C^3}\le8u\lVert f\rVert_{C^3}.
\end{aligned} ∥ G ( c 3 f ′′′ + c 2 f ′′ + c 1 f ′ ) ∥ C 3 ≤ 2 3 ( 4 u + 8 u 2 + 16 u 3 ) ∥ f ∥ C 3 ≤ 2 15 u ∥ f ∥ C 3 ≤ 8 u ∥ f ∥ C 3 .
Since 8 u ≤ 1 / 32 < 1 8u\le1/32<1 8 u ≤ 1/32 < 1 , successive substitution yields a Neumann series that converges in C 3 C^3 C 3 , and the equation further implies f ∈ C 4 f\in C^4 f ∈ C 4 . In addition,
f 0 ′ = 1 − 4 x + 3 x 2 , f 0 ′ ′ = − 4 + 6 x , f 0 ′ ′ ′ = 6 , f'_0=1-4x+3x^2,\qquad f^{\prime\prime}_0=-4+6x,\qquad f^{\prime\prime\prime}_0=6, f 0 ′ = 1 − 4 x + 3 x 2 , f 0 ′′ = − 4 + 6 x , f 0 ′′′ = 6 ,
∥ f 0 ∥ C 3 = max { 4 27 , 1 , 4 , 6 } = 6. \lVert f_0\rVert_{C^3}=\max\left\{\frac{4}{27},1,4,6\right\}=6. ∥ f 0 ∥ C 3 = max { 27 4 , 1 , 4 , 6 } = 6.
Hence
∥ f ∥ C 3 ≤ 6 1 − 8 u ≤ 12 , ∥ f − f 0 ∥ C 3 ≤ 48 u 1 − 8 u ≤ 96 u . \lVert f\rVert_{C^3}\le\frac{6}{1-8u}\le12,\qquad\lVert f-f_0\rVert_{C^3}\le\frac{48u}{1-8u}\le96u. ∥ f ∥ C 3 ≤ 1 − 8 u 6 ≤ 12 , ∥ f − f 0 ∥ C 3 ≤ 1 − 8 u 48 u ≤ 96 u .
We retain these bounds in the following form
∥ f ∥ C 3 ≤ 12 , ∥ f − f 0 ∥ C 3 ≤ 96 u . (25) \lVert f\rVert_{C^3}\le12,\qquad\lVert f-f_0\rVert_{C^3}\le96u. \tag*{(25)} ∥ f ∥ C 3 ≤ 12 , ∥ f − f 0 ∥ C 3 ≤ 96 u . (25)
The coefficients are holomorphic in ( h , z ) (h,z) ( h , z ) , and the Neumann series converges locally uniformly in operator norm, so the holomorphic dependence of f f f , together with that of f ′ ′ ( 0 ) f^{\prime\prime}(0) f ′′ ( 0 ) , on these parameters is proved. The fixed point formula and the bound G : C 0 → C 4 G:C^0\to C^4 G : C 0 → C 4 promote this dependence to C 4 C^4 C 4 , which justifies differentiating the fourth order equation in Step 7 below.
For the outer solution, the slope is one at a a a and
L D v o u t ( a ) v o u t ′ ( a ) = f ′ ′ ( 0 ) h + b a . \frac{L_Dv_{\mathrm{out}}(a)}{v_{\mathrm{out}}^{\prime}(a)}=\frac{f^{\prime\prime}(0)}{h}+\frac{b}{a}. v out ′ ( a ) L D v out ( a ) = h f ′′ ( 0 ) + a b .
Setting this equal to H ( k a ) / a \mathcal{H}(ka)/a H ( k a ) / a and multiplying both sides by a h ah ah gives the regularized matching equation
F D ( h , z ) : = h H ( α + α h z ) − ( 1 − h ) f ′ ′ ( 0 ) − b h = 0. (26) F_D(h,z):=h\mathcal{H}(\alpha+\alpha hz)-(1-h)f^{\prime\prime}(0)-bh=0. \tag*{(26)} F D ( h , z ) := h H ( α + α h z ) − ( 1 − h ) f ′′ ( 0 ) − bh = 0. (26)
Here f = f ( h , z , ⋅ ) f=f(h,z,\cdot) f = f ( h , z , ⋅ ) is the unique outer solution from Step 4. Step 2 has removed the apparent singularity at h = 0 h=0 h = 0 , so F D F_D F D is holomorphic there and
F D ( 0 , z ) = 1 z + 4. F_D(0,z)=\frac{1}{z}+4. F D ( 0 , z ) = z 1 + 4.
To find a root for each ∣ h ∣ ≤ R D |h|\le R_D ∣ h ∣ ≤ R D , we estimate its difference from the simple limiting equation
∣ F D ( h , z ) − ( 1 / z + 4 ) ∣ ≤ 10 u + ∣ f ′ ′ ( 0 ) + 4 ∣ + ∣ h ∣ ∣ f ′ ′ ( 0 ) ∣ + b ∣ h ∣ \left|F_D(h,z)-(1/z+4)\right|\le10u+\left|f^{\prime\prime}(0)+4\right|+|h|\left|f^{\prime\prime}(0)\right|+b|h| ∣ F D ( h , z ) − ( 1/ z + 4 ) ∣ ≤ 10 u + ∣ f ′′ ( 0 ) + 4 ∣ + ∣ h ∣ ∣ f ′′ ( 0 ) ∣ + b ∣ h ∣
≤ 10 u + 96 u + 12 ∣ h ∣ + b ∣ h ∣ ≤ ( 10 + 96 + 12 + 1 ) u = 119 u ≤ 128 u ≤ 1 2 . \begin{aligned}
\le10u+96u+12|h|+b|h| \\
&\le(10+96+12+1)u=119u\le128u\le\frac{1}{2}.
\end{aligned} ≤ 10 u + 96 u + 12∣ h ∣ + b ∣ h ∣ ≤ ( 10 + 96 + 12 + 1 ) u = 119 u ≤ 128 u ≤ 2 1 .
The penultimate bound uses 12 ∣ h ∣ ≤ 12 u 12|h|\le12u 12∣ h ∣ ≤ 12 u and b ∣ h ∣ ≤ u b|h|\le u b ∣ h ∣ ≤ u . On the circle ∣ z + 1 / 4 ∣ = 1 / 8 |z+1/4|=1/8 ∣ z + 1/4∣ = 1/8 ,
∣ 1 z + 4 ∣ = 4 ∣ z + 1 / 4 ∣ ∣ z ∣ ≥ 4 ( 1 / 8 ) 3 / 8 = 4 3 . \left|\frac{1}{z}+4\right|=\frac{4|z+1/4|}{|z|}\ge\frac{4(1/8)}{3/8}=\frac{4}{3}. z 1 + 4 = ∣ z ∣ 4∣ z + 1/4∣ ≥ 3/8 4 ( 1/8 ) = 3 4 .
Hence the perturbation is strictly smaller than the limiting function on the whole boundary, and this comparison is uniform in h h h .
The disk ∣ z + 1 / 4 ∣ ≤ 1 / 8 |z+1/4|\le1/8 ∣ z + 1/4∣ ≤ 1/8 does not contain z = 0 z=0 z = 0 . So the function 1 / z + 4 1/z+4 1/ z + 4 is holomorphic in a neighborhood of the disk and has exactly one zero there, namely − 1 / 4 -1/4 − 1/4 . By Rouché’s theorem [3 ] , F D ( h , ⋅ ) F_D(h,\cdot) F D ( h , ⋅ ) also has exactly one zero in this disk, counted with multiplicity, which we denote by z D ( h ) z_D(h) z D ( h ) . Counting multiplicity shows that this zero is simple. The holomorphic implicit function theorem gives its local holomorphic dependence on h h h , and uniqueness makes these local functions agree on overlaps. Thus z D z_D z D is holomorphic on the whole parameter disk, and
∣ z D ( h ) + 1 4 ∣ < 1 8 , ∣ z D ( h ) ∣ < 3 8 . \left|z_D(h)+\frac{1}{4}\right|<\frac{1}{8},\qquad|z_D(h)|<\frac{3}{8}. z D ( h ) + 4 1 < 8 1 , ∣ z D ( h ) ∣ < 8 3 .
To extend the continuation beyond the closed disk, we repeat the estimates with ∣ h ∣ ≤ 2 R D |h| \le2R_D ∣ h ∣ ≤ 2 R D :
∣ e ∣ ≤ 3 1024 < 1 16 , u ≤ 1 128 , |e| \le\frac{3}{1024} < \frac{1}{16}, \qquad u \le\frac{1}{128}, ∣ e ∣ ≤ 1024 3 < 16 1 , u ≤ 128 1 ,
4 + 8 u + 16 u 2 ≤ 4 + 1 16 + 1 1024 < 5 , 8 u ≤ 1 16 < 1 , 4 + 8u + 16u^2 \le4 + \frac{1}{16} + \frac{1}{1024} < 5, \qquad8u \le\frac{1}{16} < 1, 4 + 8 u + 16 u 2 ≤ 4 + 16 1 + 1024 1 < 5 , 8 u ≤ 16 1 < 1 ,
∥ f ∥ C 3 ≤ 6 1 − 8 u ≤ 96 15 < 12 , \|f\|_{C^3} \le\frac{6}{1-8u} \le\frac{96}{15} < 12, ∥ f ∥ C 3 ≤ 1 − 8 u 6 ≤ 15 96 < 12 ,
∥ f − f 0 ∥ C 3 ≤ 48 u 1 − 8 u ≤ 768 15 u ≤ 96 u , \|f-f_0\|_{C^3} \le\frac{48u}{1-8u} \le\frac{768}{15}u \le96u, ∥ f − f 0 ∥ C 3 ≤ 1 − 8 u 48 u ≤ 15 768 u ≤ 96 u ,
∣ F D − ( 1 / z + 4 ) ∣ ≤ 128 u ≤ 1 < 4 3 . |F_D-(1/z+4)| \le128u \le1 < \frac{4}{3}. ∣ F D − ( 1/ z + 4 ) ∣ ≤ 128 u ≤ 1 < 3 4 .
The same Rouché and uniqueness argument yields a holomorphic z D z_D z D on ∣ h ∣ < 2 R D |h|<2R_D ∣ h ∣ < 2 R D , which is a neighborhood of the original closed disk.
For the moment denote the candidate normalized eigenvalue by
r ~ D ( h ) = ( 1 + h z D ( h ) 1 − h ) 2 . \widetilde r_D(h)=\left(\frac{1+hz_D(h)}{1-h}\right)^2. r D ( h ) = ( 1 − h 1 + h z D ( h ) ) 2 .
It is holomorphic, and since ∣ h ∣ ≤ R D < 1 / 16 |h|\le R_D<1/16 ∣ h ∣ ≤ R D < 1/16 ,
∣ r ~ D ( h ) ∣ ≤ ( 1 + ( 3 / 8 ) ( 1 / 16 ) 1 − 1 / 16 ) 2 = ( 131 120 ) 2 < 2. |\widetilde r_D(h)|\le\left(\frac{1+(3/8)(1/16)}{1-1/16}\right)^2=\left(\frac{131}{120}\right)^2<2. ∣ r D ( h ) ∣ ≤ ( 1 − 1/16 1 + ( 3/8 ) ( 1/16 ) ) 2 = ( 120 131 ) 2 < 2.
Step 6. Identify the root with the first radial branch.
For real 0 < h ≤ R D 0<h\le R_D 0 < h ≤ R D , the coefficients have the usual conjugation symmetry, so the conjugate of a root is another root in the same disk. Uniqueness forces z D ( h ) z_D(h) z D ( h ) to be real and gives
− 3 8 < z D ( h ) < − 1 8 , k α − 1 = h ( 1 + z D ( h ) ) 1 − h > 0. -\frac{3}{8}<z_D(h)<-\frac{1}{8}, \qquad\frac{k}{\alpha}-1=\frac{h(1+z_D(h))}{1-h}>0. − 8 3 < z D ( h ) < − 8 1 , α k − 1 = 1 − h h ( 1 + z D ( h )) > 0.
The inner and outer pieces have the same value, slope, and L D v L_Dv L D v at a a a . The regularized Bessel expression on the inner ball is smooth in r 2 r^2 r 2 near the origin. Since the pieces are smooth on either side of a a a , their matching values and first derivatives give an H 2 H^2 H 2 function across the interface. Its weak second derivatives contain no measure supported on that interface. Together with the clamped trace conditions at 1 1 1 , this places the function in the regular radial form domain. It is nonzero because v ′ ( a ) = 1 v'(a)=1 v ′ ( a ) = 1 . Moreover,
v ′ ′ ( a ± ) = ( L D v ) ( a ± ) − D − 1 a v ′ ( a ± ) , v''(a\pm)=(L_Dv)(a\pm)-\frac{D-1}{a}v'(a\pm), v ′′ ( a ± ) = ( L D v ) ( a ± ) − a D − 1 v ′ ( a ± ) ,
so the traces of the second radial derivative also agree. The piecewise function is C 2 C^2 C 2 across the sphere, and the derivative estimates below rely on this fact. Since the test function in (18) vanishes at a a a , k 2 k^2 k 2 is a constrained radial eigenvalue. To determine which eigenvalue it is, use
r ~ D ( h ) − 1 = 2 h ( 1 + z D ( h ) ) + h 2 ( z D ( h ) 2 − 1 ) ( 1 − h ) 2 . \widetilde r_D(h)-1=\frac{2h(1+z_D(h))+h^2(z_D(h)^2-1)}{(1-h)^2}. r D ( h ) − 1 = ( 1 − h ) 2 2 h ( 1 + z D ( h )) + h 2 ( z D ( h ) 2 − 1 ) .
For complex h h h with 0 < ∣ h ∣ ≤ R D 0 < |h| \le R_D 0 < ∣ h ∣ ≤ R D , we have
∣ r ~ D ( h ) − 1 ∣ ≤ ( 11 / 4 ) ∣ h ∣ + ( 73 / 64 ) ∣ h ∣ 2 ( 1 − ∣ h ∣ ) 2 < 4 ∣ h ∣ . |\widetilde r_D(h)-1| \le\frac{(11/4)|h|+(73/64)|h|^2}{(1-|h|)^2}<4|h|. ∣ r D ( h ) − 1∣ ≤ ( 1 − ∣ h ∣ ) 2 ( 11/4 ) ∣ h ∣ + ( 73/64 ) ∣ h ∣ 2 < 4∣ h ∣.
Indeed,
11 / 4 + ( 73 / 64 ) ∣ h ∣ ( 1 − ∣ h ∣ ) 2 ≤ ( 11 4 + 73 1024 ) ( 16 15 ) 2 = 321 100 < 4. \frac{11/4+(73/64)|h|}{(1-|h|)^2}\le\left(\frac{11}{4}+\frac{73}{1024}\right)\left(\frac{16}{15}\right)^2=\frac{321}{100}<4. ( 1 − ∣ h ∣ ) 2 11/4 + ( 73/64 ) ∣ h ∣ ≤ ( 4 11 + 1024 73 ) ( 15 16 ) 2 = 100 321 < 4.
Hence for real 0 < h ≤ R D 0<h\le R_D 0 < h ≤ R D ,
0 < k 2 − α 2 < 4 α 2 h ≤ 4 ( D + 2 ) 2 R D = D + 2 64 < D + 1. (27) 0<k^2-\alpha^2<4\alpha^2h\le4(D+2)^2R_D=\frac{D+2}{64}<D+1. \tag*{(27)} 0 < k 2 − α 2 < 4 α 2 h ≤ 4 ( D + 2 ) 2 R D = 64 D + 2 < D + 1. (27)
By (16) and (15) , the second unconstrained radial eigenvalue satisfies
j ~ D / 2 , 2 2 > j D / 2 + 1 , 1 2 + D + 2 ≥ α 2 + ( D + 1 ) + ( D + 2 ) > α 2 + D + 1. \tilde j_{D/2,2}^{\,2}>j_{D/2+1,1}^{\,2}+D+2\ge\alpha^2+(D+1)+(D+2)>\alpha^2+D+1. j ~ D /2 , 2 2 > j D /2 + 1 , 1 2 + D + 2 ≥ α 2 + ( D + 1 ) + ( D + 2 ) > α 2 + D + 1.
Restricting the variational space cannot lower the second radial eigenvalue, and the eigenvalue we constructed is strictly below this threshold, so it must be the first constrained radial eigenvalue. The same comparison with the second eigenvalue, counted with multiplicity, also proves that it is simple. Thus for real positive h h h in this interval, r ~ D ( h ) = ρ D ( 1 − h ) / α 2 = r D ( h ) \widetilde r_D(h)=\rho_D(1-h)/\alpha^2=r_D(h) r D ( h ) = ρ D ( 1 − h ) / α 2 = r D ( h ) . The candidate constructed in Step 5 gives exactly the asserted holomorphic continuation of r D r_D r D .
Step 7. Compute the first two Taylor coefficients.
The analytic construction ensures that expanding in h h h along z = z D ( h ) z=z_D(h) z = z D ( h ) is legitimate. In the outer equation the coefficient of f ′ ′ ′ f''' f ′′′ is 2 b h + O ( h 2 ) 2bh+O(h^2) 2 bh + O ( h 2 ) , while the other two perturbation coefficients begin at orders h 2 h^2 h 2 and h 3 h^3 h 3 , respectively. We therefore write in C 4 C^4 C 4
f = f 0 + h f 1 + O ( h 2 ) , f=f_0+hf_1+O(h^2), f = f 0 + h f 1 + O ( h 2 ) ,
and the coefficient of h h h satisfies
f 1 ′ ′ ′ ′ = − 2 b f 0 ′ ′ ′ = − 12 b , f 1 ( 0 ) = f 1 ′ ( 0 ) = f 1 ( 1 ) = f 1 ′ ( 1 ) = 0. f_1''''=-2bf_0'''=-12b,\qquad f_1(0)=f_1'(0)=f_1(1)=f_1'(1)=0. f 1 ′′′′ = − 2 b f 0 ′′′ = − 12 b , f 1 ( 0 ) = f 1 ′ ( 0 ) = f 1 ( 1 ) = f 1 ′ ( 1 ) = 0.
These conditions determine
f 1 ( x ) = − b 2 x 2 ( 1 − x ) 2 , f 1 ′ ′ ( 0 ) = − b . f_1(x)=-\frac{b}{2}x^2(1-x)^2,\qquad f_1''(0)=-b. f 1 ( x ) = − 2 b x 2 ( 1 − x ) 2 , f 1 ′′ ( 0 ) = − b .
A direct verification gives
f 1 ′ ( x ) = − b x ( 1 − x ) ( 1 − 2 x ) , f 1 ′ ′ ′ ′ ( x ) = − b 2 ⋅ 24 = − 12 b , f_1'(x)=-bx(1-x)(1-2x),\qquad f_1''''(x)=-\frac{b}{2}\cdot24=-12b, f 1 ′ ( x ) = − b x ( 1 − x ) ( 1 − 2 x ) , f 1 ′′′′ ( x ) = − 2 b ⋅ 24 = − 12 b ,
f 1 ( 0 ) = f 1 ( 1 ) = f 1 ′ ( 0 ) = f 1 ′ ( 1 ) = 0. f_1(0)=f_1(1)=f_1'(0)=f_1'(1)=0. f 1 ( 0 ) = f 1 ( 1 ) = f 1 ′ ( 0 ) = f 1 ′ ( 1 ) = 0.
Thus the outer side of the matching equation has the expansion
( 1 − h ) f ′ ′ ( 0 ) + b h = ( 1 − h ) ( − 4 − b h + O ( h 2 ) ) + b h = − 4 + 4 h + O ( h 2 ) . (1-h)f''(0)+bh=(1-h)(-4-bh+O(h^2))+bh
=-4+4h+O(h^2). ( 1 − h ) f ′′ ( 0 ) + bh = ( 1 − h ) ( − 4 − bh + O ( h 2 )) + bh = − 4 + 4 h + O ( h 2 ) .
The terms containing b b b cancel here. Writing this cancellation explicitly is useful, since the dimensional dependence of the next coefficient is then confined entirely to the inner expansion.
At the Bessel zero α \alpha α , the equation gives J ν ′ ′ ( α ) = − J ν ′ ( α ) / α J_{\nu}''(\alpha)=-J_{\nu}'(\alpha)/\alpha J ν ′′ ( α ) = − J ν ′ ( α ) / α . Accordingly,
J ν ( α + e ) = J ν ′ ( α ) e ( 1 − e 2 α + O ( e 2 ) ) , J_{\nu}(\alpha+e)=J_{\nu}'(\alpha)e\left(1-\frac{e}{2\alpha}+O(e^2)\right), J ν ( α + e ) = J ν ′ ( α ) e ( 1 − 2 α e + O ( e 2 ) ) ,
J ν ′ ( α + e ) = J ν ′ ( α ) ( 1 − e α + O ( e 2 ) ) , J_{\nu}'(\alpha+e)=J_{\nu}'(\alpha)\left(1-\frac{e}{\alpha}+O(e^2)\right), J ν ′ ( α + e ) = J ν ′ ( α ) ( 1 − α e + O ( e 2 ) ) ,
J ν ′ ( α + e ) J ν ( α + e ) = 1 e − 1 2 α + O ( e ) . \frac{J_{\nu}'(\alpha+e)}{J_{\nu}(\alpha+e)}=\frac{1}{e}-\frac{1}{2\alpha}+O(e). J ν ( α + e ) J ν ′ ( α + e ) = e 1 − 2 α 1 + O ( e ) .
Multiplying the last line by α + e \alpha+e α + e and adding ν \nu ν yields
H ( α + e ) = α e + ν + 1 2 + O ( e ) = α e + D + 1 2 + O ( e ) . \mathcal{H}(\alpha+e)=\frac{\alpha}{e}+\nu+\frac{1}{2}+O(e)=\frac{\alpha}{e}+\frac{D+1}{2}+O(e). H ( α + e ) = e α + ν + 2 1 + O ( e ) = e α + 2 D + 1 + O ( e ) .
Let z D ( h ) = z 0 + z 1 h + O ( h 2 ) z_D(h)=z_0+z_1h+O(h^2) z D ( h ) = z 0 + z 1 h + O ( h 2 ) . Substituting e = α h z D ( h ) e=\alpha hz_D(h) e = α h z D ( h ) into the matching equation gives
1 z D ( h ) + D + 1 2 h + O ( h 2 ) = − 4 + 4 h + O ( h 2 ) . \frac{1}{z_D(h)}+\frac{D+1}{2}h+O(h^2)=-4+4h+O(h^2). z D ( h ) 1 + 2 D + 1 h + O ( h 2 ) = − 4 + 4 h + O ( h 2 ) .
Also 1 / z D ( h ) = 1 / z 0 − ( z 1 / z 0 2 ) h + O ( h 2 ) 1/z_D(h)=1/z_0-(z_1/z_0^2)h+O(h^2) 1/ z D ( h ) = 1/ z 0 − ( z 1 / z 0 2 ) h + O ( h 2 ) , and comparing constant and linear terms yields
1 z 0 = − 4 , − 16 z 1 + D + 1 2 = 4. \frac{1}{z_0}=-4,\qquad-16z_1+\frac{D+1}{2}=4. z 0 1 = − 4 , − 16 z 1 + 2 D + 1 = 4.
Hence
z 0 = − 1 4 , z 1 = D − 7 32 . z_0=-\frac{1}{4},\qquad z_1=\frac{D-7}{32}. z 0 = − 4 1 , z 1 = 32 D − 7 .
Finally,
( 1 + h z D ( h ) ) 2 = 1 + 2 z 0 h + ( z 0 2 + 2 z 1 ) h 2 + O ( h 3 ) , (1+hz_D(h))^2=1+2z_0h+(z_0^2+2z_1)h^2+O(h^3), ( 1 + h z D ( h ) ) 2 = 1 + 2 z 0 h + ( z 0 2 + 2 z 1 ) h 2 + O ( h 3 ) ,
( 1 − h ) − 2 = 1 + 2 h + 3 h 2 + O ( h 3 ) , (1-h)^{-2}=1+2h+3h^2+O(h^3), ( 1 − h ) − 2 = 1 + 2 h + 3 h 2 + O ( h 3 ) ,
r D ( h ) = 1 + ( 2 + 2 z 0 ) h + ( 3 + 4 z 0 + z 0 2 + 2 z 1 ) h 2 + O ( h 3 ) . r_D(h)=1+(2+2z_0)h+(3+4z_0+z_0^2+2z_1)h^2+O(h^3). r D ( h ) = 1 + ( 2 + 2 z 0 ) h + ( 3 + 4 z 0 + z 0 2 + 2 z 1 ) h 2 + O ( h 3 ) .
The two coefficients are
2 + 2 z 0 = 3 2 , 3 + 4 z 0 + z 0 2 + 2 z 1 = 3 − 1 + 1 16 + D − 7 16 = D + 26 16 . 2+2z_0=\frac{3}{2},\qquad3+4z_0+z_0^2+2z_1=3-1+\frac{1}{16}+\frac{D-7}{16}=\frac{D+26}{16}. 2 + 2 z 0 = 2 3 , 3 + 4 z 0 + z 0 2 + 2 z 1 = 3 − 1 + 16 1 + 16 D − 7 = 16 D + 26 .
Bound all remaining Taylor terms
Write r D ( h ) = ∑ n = 0 + ∞ a n h n r_D(h)=\sum_{n=0}^{+\infty}a_nh^n r D ( h ) = ∑ n = 0 + ∞ a n h n . The continuation is holomorphic on a neighborhood of the closed disk, so Cauchy’s formula can be applied directly on the boundary. Since ∣ r D ∣ ≤ 2 |r_D|\le2 ∣ r D ∣ ≤ 2 ,
a n = 1 2 π i ∫ ∣ ζ ∣ = R D r D ( ζ ) ζ n + 1 d ζ , ∣ a n ∣ ≤ ( 2 π R D ) 2 2 π R D n + 1 = 2 R D − n . a_n=\frac{1}{2\pi i}\int_{|\zeta|=R_D}\frac{r_D(\zeta)}{\zeta^{n+1}}\,d\zeta,\qquad|a_n|\le\frac{(2\pi R_D)\,2}{2\pi R_D^{n+1}}=2R_D^{-n}. a n = 2 π i 1 ∫ ∣ ζ ∣ = R D ζ n + 1 r D ( ζ ) d ζ , ∣ a n ∣ ≤ 2 π R D n + 1 ( 2 π R D ) 2 = 2 R D − n .
After removing the three coefficients computed in Step 7, the remaining series satisfies, for ∣ h ∣ ≤ R D / 2 |h| \le R_D/2 ∣ h ∣ ≤ R D /2 ,
∣ R D ( h ) ∣ ≤ ∑ n = 3 + ∞ 2 ( ∣ h ∣ R D ) n = 2 ( ∣ h ∣ / R D ) 3 1 − ∣ h ∣ / R D ≤ 4 ∣ h ∣ 3 R D 3 . |\mathcal{R}_D(h)| \le\sum_{n=3}^{+\infty} 2\left(\frac{|h|}{R_D}\right)^n
= \frac{2(|h|/R_D)^3}{1-|h|/R_D} \le\frac{4|h|^3}{R_D^3}. ∣ R D ( h ) ∣ ≤ n = 3 ∑ + ∞ 2 ( R D ∣ h ∣ ) n = 1 − ∣ h ∣/ R D 2 ( ∣ h ∣/ R D ) 3 ≤ R D 3 4∣ h ∣ 3 .
This proves the expansion and the remainder with explicit dependence on the dimension.
□ \square □
Figure 5 shows the fixed disk used in solving (3.17). Figure 6 shows the leading-order profile (3.15) and its nonzero slope at the interface.
Figure 5. The matching disk excludes the pole 0 and encloses the limiting root − 1 / 4 -1/4 − 1/4 . The bound shown on the boundary guarantees that a simple zero persists under the perturbation.
Figure 6. The leading-order outer profile: f 0 ( 0 ) = f 0 ( 1 ) = f 0 ′ ( 1 ) = 0 f_0(0)=f_0(1)=f_0'(1)=0 f 0 ( 0 ) = f 0 ( 1 ) = f 0 ′ ( 1 ) = 0 and f 0 ′ ( 0 ) = 1 f_0'(0)=1 f 0 ′ ( 0 ) = 1 .
Identification of the first two eigenvalues
The preceding lemma estimates a radial branch in each effective dimension. The next proposition identifies the first two eigenvalues of the fully constrained problem and provides the strict gap that subsequent approximations must preserve.
Proposition 3.3. Let d ≥ 2 d \ge2 d ≥ 2 , and let h h h , a a a be as in (8) . Then
μ 1 ( a ) = ρ d ( a ) , μ 2 ( a ) = ⋯ = μ d + 1 ( a ) = ρ d + 2 ( a ) . \mu_1(a)=\rho_d(a),\qquad\mu_2(a)=\cdots=\mu_{d+1}(a)=\rho_{d+2}(a). μ 1 ( a ) = ρ d ( a ) , μ 2 ( a ) = ⋯ = μ d + 1 ( a ) = ρ d + 2 ( a ) . The first eigenvalue is simple, and the second has exactly multiplicity d d d . Moreover,
μ 2 ( a ) − C d μ 1 ( a ) ≥ 3 32 α d + 2 2 h 2 > h 2 4 , μ 2 ( a ) μ 1 ( a ) > C d . \mu_2(a)-C_d\mu_1(a)\ge\frac{3}{32}\alpha_{d+2}^2h^2>\frac{h^2}{4},\qquad\frac{\mu_2(a)}{\mu_1(a)}>C_d. μ 2 ( a ) − C d μ 1 ( a ) ≥ 32 3 α d + 2 2 h 2 > 4 h 2 , μ 1 ( a ) μ 2 ( a ) > C d .
Proof. The proof consists of four steps. We first control the remainder, then rule out competing modes, next compute the gap, and finally reserve part of the gap for the surface penalty term and the latitude constraints in § 9 .
Step 1. Check the chosen scale against both analytic radii.
The difference between the second order coefficients of the two branches is 2 / 16 = 1 / 8 2/16=1/8 2/16 = 1/8 . Each branch allows an error of only h 2 / 64 h^2/64 h 2 /64 , so the difference used is only 2 / 64 2/64 2/64 , leaving a positive margin of 3 / 32 3/32 3/32 . This explains why h h h must be chosen so small. At the common analytic radius below, the condition 4 h 3 / R 0 3 ≤ h 2 / 64 4h^3/R_0^3\le h^2/64 4 h 3 / R 0 3 ≤ h 2 /64 is equivalent to h ≤ 2 − 32 m − 3 h\le2^{-32}m^{-3} h ≤ 2 − 32 m − 3 . The dyadic parameters in § 2 realize exactly this convenient and conservative choice.
Returning to m = d + 4 m=d+4 m = d + 4 , H = 32 + ⌈ 3 log 2 m ⌉ H=32+\lceil3\log_2 m\rceil H = 32 + ⌈ 3 log 2 m ⌉ , and h = 2 − H h=2^{-H} h = 2 − H from § 2 , we have
3 log 2 m ≤ ⌈ 3 log 2 m ⌉ < 3 log 2 m + 1. 3\log_2m\le\lceil3\log_2m\rceil<3\log_2m+1. 3 log 2 m ≤ ⌈ 3 log 2 m ⌉ < 3 log 2 m + 1.
and hence
2 − 33 m − 3 < h = 2 − 32 − ⌈ 3 log 2 m ⌉ ≤ 2 − 32 m − 3 . 2^{-33}m^{-3}<h=2^{-32-\lceil3\log_2m\rceil}\le2^{-32}m^{-3}. 2 − 33 m − 3 < h = 2 − 32 − ⌈ 3 l o g 2 m ⌉ ≤ 2 − 32 m − 3 .
This is the largest dyadic value satisfying the sufficient remainder condition h ≤ 2 − 32 m − 3 h\le2^{-32}m^{-3} h ≤ 2 − 32 m − 3 . Doubling it violates the condition. The cost of the dyadic restriction is therefore less than a factor of two. The two relevant effective dimensions are d d d and d + 2 d+2 d + 2 , and their common analytic radius is
R 0 : = 1 256 m = 1 256 ( d + 4 ) = R d + 2 ≤ R d . R_0:=\frac{1}{256m}=\frac{1}{256(d+4)}=R_{d+2}\le R_d. R 0 := 256 m 1 = 256 ( d + 4 ) 1 = R d + 2 ≤ R d .
In particular,
h R 0 = 256 m h ≤ 2 − 24 m − 2 < 1 2 . \frac{h}{R_0}=256mh\le2^{-24}m^{-2}<\frac{1}{2}. R 0 h = 256 mh ≤ 2 − 24 m − 2 < 2 1 .
Thus the remainder bound in (20) applies to both branches. Its comparison with the desired second order margin is an exact power of two calculation:
4 h 3 R 0 3 = 4 ⋅ 256 3 m 3 h 3 = 2 26 m 3 h 3 ≤ 2 26 m 3 ( 2 − 32 m − 3 ) h 2 = h 2 64 . \frac{4h^3}{R_0^3}=4\cdot256^3m^3h^3=2^{26}m^3h^3\le2^{26}m^3(2^{-32}m^{-3})h^2=\frac{h^2}{64}. R 0 3 4 h 3 = 4 ⋅ 25 6 3 m 3 h 3 = 2 26 m 3 h 3 ≤ 2 26 m 3 ( 2 − 32 m − 3 ) h 2 = 64 h 2 .
The uniform analytic estimate is therefore small enough to resolve the difference between the two second order coefficients.
Set
A = α d 2 , B 0 = α d + 2 2 , C d = B 0 A . A=\alpha_d^2,\qquad B_0=\alpha_{d+2}^2,\qquad C_d=\frac{B_0}{A}. A = α d 2 , B 0 = α d + 2 2 , C d = A B 0 .
and define the explicit upper and lower comparison
U = A ( 1 + 3 2 h + d + 26 16 h 2 + h 2 64 ) , V = B 0 ( 1 + 3 2 h + d + 28 16 h 2 − h 2 64 ) . (28) \begin{aligned}
U &= A\left(1+\frac{3}{2}h+\frac{d+26}{16}h^2+\frac{h^2}{64}\right),\\
V &= B_0\left(1+\frac{3}{2}h+\frac{d+28}{16}h^2-\frac{h^2}{64}\right). \tag*{(28)}
\end{aligned} U V = A ( 1 + 2 3 h + 16 d + 26 h 2 + 64 h 2 ) , = B 0 ( 1 + 2 3 h + 16 d + 28 h 2 − 64 h 2 ) . (28)
By Lemma 3.2, we have
ρ d ( a ) ≤ U , ρ d + 2 ( a ) ≥ V . \rho_d(a) \le U,\qquad\rho_{d+2}(a) \ge V. ρ d ( a ) ≤ U , ρ d + 2 ( a ) ≥ V .
The added h 2 / 64 h^2/64 h 2 /64 in U U U and the subtracted h 2 / 64 h^2/64 h 2 /64 in V V V exactly cover the possible signs of the two analytic remainders, respectively.
Step 2. Locate the two branches before excluding all competitors.
For D = d , d + 2 D=d,d+2 D = d , d + 2 , (14) gives α D 2 ≤ m 2 \alpha_D^2 \le m^2 α D 2 ≤ m 2 . The displacement bound (27) yields
0 < ρ D ( a ) − α D 2 < 4 α D 2 h ≤ 4 m 2 h ≤ 2 − 30 m − 1 < 1. 0 < \rho_D(a)-\alpha_D^2 < 4\alpha_D^2h \le4m^2h \le2^{-30}m^{-1} < 1. 0 < ρ D ( a ) − α D 2 < 4 α D 2 h ≤ 4 m 2 h ≤ 2 − 30 m − 1 < 1.
Taking ν = d / 2 \nu=d/2 ν = d /2 in (15) gives
B 0 − A ≥ d + 1 > 1. B_0-A \ge d+1>1. B 0 − A ≥ d + 1 > 1.
Hence the candidate first and second branches lie in disjoint intervals:
A < ρ d ( a ) < A + 1 < B 0 < ρ d + 2 ( a ) < B 0 + 1. A<\rho_d(a)<A+1<B_0<\rho_{d+2}(a)<B_0+1. A < ρ d ( a ) < A + 1 < B 0 < ρ d + 2 ( a ) < B 0 + 1.
This alone does not yet rule out higher radial branches or other angular degrees. Using the monotonicity of the sector eigenvalues under the value constraint together with the estimates on the ball, we obtain the following complete list:
Table 1.
Here the first row is (16) with ν = d / 2 \nu=d/2 ν = d /2 . For the second row, raising the degree does not decrease the eigenvalue when the radial index is fixed. For the third row, we first reduce to the smallest degree and radial index in that row and then apply the order gap with ν = d / 2 + 1 \nu=d/2+1 ν = d /2 + 1 :
j d / 2 + 2 , 1 2 − j d / 2 + 1 , 1 2 ≥ d + 3. j_{d/2+2,1}^2-j_{d/2+1,1}^2\ge d+3. j d /2 + 2 , 1 2 − j d /2 + 1 , 1 2 ≥ d + 3.
Every bound in the table exceeds B 0 + 1 B_0+1 B 0 + 1 . Restricting the sector form domain only raises the min-max values with the same index, so the same exclusion remains valid after adding v ( a ) = 0 v(a)=0 v ( a ) = 0 . In particular, the third row handles arbitrarily large angular degrees at once, with no asymptotic expansion in these degrees.
It follows that
μ 1 ( a ) = ρ d ( a ) ≤ U , μ 2 ( a ) = ρ d + 2 ( a ) ≥ V . (29) \mu_1(a)=\rho_d(a)\le U,\qquad\mu_2(a)=\rho_{d+2}(a)\ge V. \tag*{(29)} μ 1 ( a ) = ρ d ( a ) ≤ U , μ 2 ( a ) = ρ d + 2 ( a ) ≥ V . (29)
The bounds for n ≥ 2 n \ge2 n ≥ 2 also establish simplicity in the two relevant radial sectors. If the first eigenvalue were repeated, the min-max value of the second sector would equal it, which contradicts
ρ D ( a ) < B 0 + 1 < j D / 2 , 2 2 , D = d , d + 2. \rho_D(a)<B_0+1<j^2_{D/2,2}, \qquad D=d,d+2. ρ D ( a ) < B 0 + 1 < j D /2 , 2 2 , D = d , d + 2.
The space of degree-zero spherical harmonics is one-dimensional, so μ 1 \mu_1 μ 1 is simple. The degree-one spherical harmonics are spanned by the restrictions of the d d d coordinate functions to the sphere. The first degree-one radial eigenvalue is simple in each copy of the radial sector, and the table rules out any coincidence with another sector. Hence the multiplicity of μ 2 \mu_2 μ 2 is exactly d d d , namely μ 2 = ⋯ = μ d + 1 \mu_2=\cdots=\mu_{d+1} μ 2 = ⋯ = μ d + 1 . Figure 7 shows its separation from all excluded branches.
Figure 7. Schematic of the spectral ordering: both shifts of the first two branches are less than one, all excluded branches lie above B 0 + d + 2 B_0+d+2 B 0 + d + 2 , and the second branch has multiplicity d d d .
Verify the comparison levels and subtract the common terms.
We shall also use two convenient bounds later, namely U < B 0 U<B_0 U < B 0 and V < B 0 + 1 V<B_0+1 V < B 0 + 1 . Since m = d + 4 ≥ 6 m=d+4\ge6 m = d + 4 ≥ 6 , the second order coefficients in these two comparisons satisfy
d + 26 16 + 1 64 = 4 m + 89 64 ≤ m , d + 28 16 − 1 64 = 4 m + 95 64 ≤ m . \frac{d+26}{16}+\frac{1}{64}=\frac{4m+89}{64}\le m,\qquad
\frac{d+28}{16}-\frac{1}{64}=\frac{4m+95}{64}\le m. 16 d + 26 + 64 1 = 64 4 m + 89 ≤ m , 16 d + 28 − 64 1 = 64 4 m + 95 ≤ m .
We also have m h ≤ 2 − 32 m − 2 < 1 / 2 mh\le2^{-32}m^{-2}<1/2 mh ≤ 2 − 32 m − 2 < 1/2 , so
U A ≤ 1 + 3 2 h + m h 2 < 1 + 2 h , V B 0 ≤ 1 + 3 2 h + m h 2 < 1 + 2 h . \frac{U}{A}\le1+\frac{3}{2}h+mh^2<1+2h,\qquad
\frac{V}{B_0}\le1+\frac{3}{2}h+mh^2<1+2h. A U ≤ 1 + 2 3 h + m h 2 < 1 + 2 h , B 0 V ≤ 1 + 2 3 h + m h 2 < 1 + 2 h .
Since 2 m 2 h ≤ 2 − 31 m − 1 < 1 2m^2h\le2^{-31}m^{-1}<1 2 m 2 h ≤ 2 − 31 m − 1 < 1 , these inequalities give U < A + 1 < B 0 U<A+1<B_0 U < A + 1 < B 0 and B 0 < V < B 0 + 1 B_0<V<B_0+1 B 0 < V < B 0 + 1 .
The useful comparison is V − C d U V-C_dU V − C d U . Multiplying by C d = B 0 / A C_d=B_0/A C d = B 0 / A puts the two expansions on the same scale, and their constant and linear terms cancel exactly:
V − C d U = B 0 [ ( d + 28 16 − 1 64 ) − ( d + 26 16 + 1 64 ) ] h 2 = B 0 ( 2 16 − 2 64 ) h 2 = 3 32 B 0 h 2 . \begin{aligned}
V-C_dU
&=B_0\left[\left(\frac{d+28}{16}-\frac{1}{64}\right)-\left(\frac{d+26}{16}+\frac{1}{64}\right)\right]h^2\\
&=B_0\left(\frac{2}{16}-\frac{2}{64}\right)h^2=\frac{3}{32}B_0h^2.
\end{aligned} V − C d U = B 0 [ ( 16 d + 28 − 64 1 ) − ( 16 d + 26 + 64 1 ) ] h 2 = B 0 ( 16 2 − 64 2 ) h 2 = 32 3 B 0 h 2 .
The remainder bounds therefore leave a positive gap
δ : = V − C d U = 3 32 B 0 h 2 > h 2 / 4. (30) \delta:=V-C_dU=\frac{3}{32}B_0h^2>h^2/4. \tag*{(30)} δ := V − C d U = 32 3 B 0 h 2 > h 2 /4. (30)
By (14) ,
B 0 > ( d + 2 2 ) 2 ≥ 4 , 3 32 B 0 > 3 8 > 1 4 . B_0 > \left(\frac{d+2}{2}\right)^2 \ge4,\qquad\frac{3}{32}B_0 > \frac{3}{8} > \frac{1}{4}. B 0 > ( 2 d + 2 ) 2 ≥ 4 , 32 3 B 0 > 8 3 > 4 1 .
For the actual eigenvalues,
μ 2 − C d μ 1 ≥ V − C d U = δ > 0 , \mu_2-C_d\mu_1 \ge V-C_dU=\delta>0, μ 2 − C d μ 1 ≥ V − C d U = δ > 0 ,
and hence μ 2 / μ 1 > C d \mu_2/\mu_1>C_d μ 2 / μ 1 > C d .
Step 4. Reserve separate portions of the gap for the approximations.
Define two levels T T T and P P P :
T : = C d U + δ / 4 , P : = C d U + δ / 2. (31) T:=C_dU+\delta/4,\qquad P:=C_dU+\delta/2. \tag*{(31)} T := C d U + δ /4 , P := C d U + δ /2. (31)
One has that
1 < C d < 36 , U + 1 < 3 m 2 , T < 2 m 2 , P < V . (32) 1<C_d<36,\qquad U+1<3m^2,\qquad T<2m^2,\qquad P<V. \tag*{(32)} 1 < C d < 36 , U + 1 < 3 m 2 , T < 2 m 2 , P < V . (32)
These bounds all follow from the estimates already proved. First, the order comparison gives C d > 1 C_d>1 C d > 1 , while
C d = B 0 A < ( d + 4 ) 2 d 2 / 4 = 4 ( 1 + 4 d ) 2 ≤ 36. C_d=\frac{B_0}{A}<\frac{(d+4)^2}{d^2/4}=4\left(1+\frac{4}{d}\right)^2\le36. C d = A B 0 < d 2 /4 ( d + 4 ) 2 = 4 ( 1 + d 4 ) 2 ≤ 36.
Second, from A ≤ m 2 A\le m^2 A ≤ m 2 and h < 1 / 2 h<1/2 h < 1/2 ,
U + 1 < A ( 1 + 2 h ) + 1 < 2 m 2 + 1 < 3 m 2 . U+1<A(1+2h)+1<2m^2+1<3m^2. U + 1 < A ( 1 + 2 h ) + 1 < 2 m 2 + 1 < 3 m 2 .
Finally, using δ = ( 3 / 32 ) B 0 h 2 \delta=(3/32)B_0h^2 δ = ( 3/32 ) B 0 h 2 and h < 1 / 4 h<1/4 h < 1/4 ,
T B 0 = U A + 3 128 h 2 < 1 + 2 h + 3 128 h 2 < 1 + 1 2 + 3 2048 < 2. \frac{T}{B_0}=\frac{U}{A}+\frac{3}{128}h^2<1+2h+\frac{3}{128}h^2<1+\frac{1}{2}+\frac{3}{2048}<2. B 0 T = A U + 128 3 h 2 < 1 + 2 h + 128 3 h 2 < 1 + 2 1 + 2048 3 < 2.
Since B 0 ≤ m 2 B_0\le m^2 B 0 ≤ m 2 , this gives T < 2 m 2 T<2m^2 T < 2 m 2 . The relation V = C d U + δ V=C_dU+\delta V = C d U + δ directly yields
C d U < T < P < V , P − T = δ 4 , V − P = δ 2 . C_dU<T<P<V,\qquad P-T=\frac{\delta}{4},\qquad V-P=\frac{\delta}{2}. C d U < T < P < V , P − T = 4 δ , V − P = 2 δ .
Also, C d U = B 0 ( U / A ) > B 0 C_dU=B_0(U/A)>B_0 C d U = B 0 ( U / A ) > B 0 , so both T T T and P P P lie strictly above B 0 B_0 B 0 and below V < B 0 + 1 V<B_0+1 V < B 0 + 1 . The surface penalty below can therefore be applied in the interval ( B 0 , P ) (B_0,P) ( B 0 , P ) without reaching the ball eigenvalues. These levels play different roles: the spherical problem provides the lower bound V V V , the finite surface penalty preserves P P P , and the latitude constraints preserve T T T .
For fixed d d d , the same branch identification holds for every 0 < h ≤ 2 − H 0<h\le2^{-H} 0 < h ≤ 2 − H , since the estimate of the remainder relative to h 2 h^2 h 2 and the branch-shift estimates only improve as h h h decreases. Hence the two individual expansions in (5) both follow from Lemma 3.2.
To check their quotient explicitly, we first subtract the denominator from the numerator before dividing:
μ 2 / μ 1 C d = 1 + 3 2 h + d + 28 16 h 2 + O d ( h 3 ) 1 + 3 2 h + d + 26 16 h 2 + O d ( h 3 ) = ( d + 28 16 − d + 26 16 ) h 2 + O d ( h 3 ) 1 + 3 2 h + d + 26 16 h 2 + O d ( h 3 ) + 1 = 1 + 1 8 h 2 + O d ( h 3 ) . \frac{\mu_2/\mu_1}{C_d}=\frac{1+\frac{3}{2}h+\frac{d+28}{16}h^2+O_d(h^3)}{1+\frac{3}{2}h+\frac{d+26}{16}h^2+O_d(h^3)}
=\frac{\left(\frac{d+28}{16}-\frac{d+26}{16}\right)h^2+O_d(h^3)}{1+\frac{3}{2}h+\frac{d+26}{16}h^2+O_d(h^3)}+1
=1+\frac{1}{8}h^2+O_d(h^3). C d μ 2 / μ 1 = 1 + 2 3 h + 16 d + 26 h 2 + O d ( h 3 ) 1 + 2 3 h + 16 d + 28 h 2 + O d ( h 3 ) = 1 + 2 3 h + 16 d + 26 h 2 + O d ( h 3 ) ( 16 d + 28 − 16 d + 26 ) h 2 + O d ( h 3 ) + 1 = 1 + 8 1 h 2 + O d ( h 3 ) .
where the last equality uses
[ 1 + O d ( h ) ] − 1 = 1 + O d ( h ) , h 2 O d ( h ) = O d ( h 3 ) . [1+O_d(h)]^{-1}=1+O_d(h),\qquad h^2O_d(h)=O_d(h^3). [ 1 + O d ( h ) ] − 1 = 1 + O d ( h ) , h 2 O d ( h ) = O d ( h 3 ) .
We now keep h = 2 − H h=2^{-H} h = 2 − H fixed. The positive margin δ \delta δ is sufficient to absorb the errors incurred when the spherical constraint is replaced first by finitely many latitude constraints and then by smooth holes.
From the sphere to finitely many latitudes
We now transfer the spherical comparison in Proposition 3.3 to functions constrained to vanish only on the finite core set K N K_N K N . We first replace the constraint on the whole sphere by a finite surface penalty term so that the second eigenvalue exceeds P P P , and then estimate the penalty term between adjacent latitudes to obtain a lower bound above T T T for the core-constrained problem. The powers in the parameter choices can be traced back to the available gap δ = 3 B 0 h 2 / 32 \delta=3B_0h^2/32 δ = 3 B 0 h 2 /32 . The lower bound derived below for the negative resolvent value − g D ( P ; a ) -g_D(P;a) − g D ( P ; a ) is proportional to h 4 h^4 h 4 , which explains the choice κ = 2 13 h − 4 \kappa=2^{13}h^{-4} κ = 2 13 h − 4 . The sampling error that follows has leading size κ Δ θ 2 / h \kappa\Delta_\theta^2/h κ Δ θ 2 / h . To make it smaller than a gap of order h 2 h^2 h 2 , we need Δ θ 2 \Delta_\theta^2 Δ θ 2 of size at most h 7 h^7 h 7 , up to dimension-dependent factors. This is the reason for the term 7 H 7H 7 H in the integer N N N . These observations explain the design. All comparisons used in the proof are the explicit inequalities below. For κ > 0 \kappa>0 κ > 0 , define the penalized bending form
Q κ ( u ) = Q ( u ) + κ ∫ S a ∣ u ∣ 2 d S , u ∈ H 0 2 ( B ) . Q_\kappa(u)=Q(u)+\kappa\int_{S_a}|u|^2\,\mathrm{d}S,\qquad u\in H_0^2(B). Q κ ( u ) = Q ( u ) + κ ∫ S a ∣ u ∣ 2 d S , u ∈ H 0 2 ( B ) .
Restricted to the inner ball B a ( 0 ) B_a(0) B a ( 0 ) , whose boundary is S a S_a S a , the trace theorem for smooth boundaries [17 Theorem 3.37, p. 102] shows that for fixed κ > 0 \kappa>0 κ > 0 , Q κ 1 / 2 Q_\kappa^{1/2} Q κ 1/2 is an equivalent Hilbert norm on H 0 2 ( B ) H_0^2(B) H 0 2 ( B ) . Let β k ( κ ) \beta_k(\kappa) β k ( κ ) denote the min-max eigenvalues defined by (7) with E \mathcal{E} E as the denominator.
A quantitative surface penalty
We first quantify the size of the surface penalty needed to preserve the lower threshold P P P in the spherical comparison.
Proposition 4.1. Let d ≥ 2 d \ge2 d ≥ 2 be an integer, and take h h h , a a a as in (8) and P P P as in (31) . For κ = 2 13 h − 4 \kappa= 2^{13}h^{-4} κ = 2 13 h − 4 , the penalized second eigenvalue satisfies
β 2 ( κ ) > P . \beta_2(\kappa) > P. β 2 ( κ ) > P .
Proof. Step 1. Normalize the radial ball modes.
The ball eigenvalue bounds in § 3 already keep all higher angular and radial branches above P P P . Thus the task here is to raise the first degree-one branch above P P P . A physical degree-one mode has the form u = r v ( r ) Y 1 u = rv(r)Y_1 u = r v ( r ) Y 1 , and its effective radial dimension is D = d + 2 D = d + 2 D = d + 2 . In this proof, Q D Q_D Q D and E D E_D E D are the radial forms in (10) , without the spherical area factor. Put ν = D / 2 \nu= D/2 ν = D /2 . In what follows, the letter z z z denotes the Bessel zero j ν , k j_{\nu,k} j ν , k of the mode under consideration. The ball modes normalized with respect to E D E_D E D are
ϕ D , k ( r ) = 2 z J D / 2 − 1 ( z ) [ r 1 − D / 2 J D / 2 − 1 ( z r ) − J D / 2 − 1 ( z ) ] , z = j D / 2 , k . \phi_{D,k}(r) = \frac{\sqrt{2}}{zJ_{D/2-1}(z)}\left[r^{1-D/2}J_{D/2-1}(zr)-J_{D/2-1}(z)\right], \qquad z=j_{D/2,k}. ϕ D , k ( r ) = z J D /2 − 1 ( z ) 2 [ r 1 − D /2 J D /2 − 1 ( z r ) − J D /2 − 1 ( z ) ] , z = j D /2 , k .
To verify the normalization explicitly, we differentiate the expression in brackets using the recurrence formula from § 3 and obtain
ϕ D , k ′ ( r ) = − 2 J ν − 1 ( z ) r 1 − ν J ν ( z r ) . \phi'_{D,k}(r)=-\frac{\sqrt{2}}{J_{\nu-1}(z)}r^{1-\nu}J_\nu(zr). ϕ D , k ′ ( r ) = − J ν − 1 ( z ) 2 r 1 − ν J ν ( z r ) .
At a zero of J ν J_\nu J ν , by the recurrence formula, one has
J ν − 1 ( z ) = J ν ′ ( z ) + ν z J ν ( z ) = J ν ′ ( z ) . J_{\nu-1}(z)=J'_\nu(z)+\frac{\nu}{z}J_\nu(z)=J'_\nu(z). J ν − 1 ( z ) = J ν ′ ( z ) + z ν J ν ( z ) = J ν ′ ( z ) .
The standard Bessel norm identity at a zero [23 Chapter V] reads
∫ 0 1 r J ν ( z r ) 2 d r = 1 2 J ν ′ ( z ) 2 , J ν ( z ) = 0. (33) \int_0^1 rJ_\nu(zr)^2\,\mathrm{d}r=\frac{1}{2}J'_\nu(z)^2,\qquad J_\nu(z)=0. \tag*{(33)} ∫ 0 1 r J ν ( z r ) 2 d r = 2 1 J ν ′ ( z ) 2 , J ν ( z ) = 0. (33)
Since D − 1 + 2 ( 1 − ν ) = 1 D-1+2(1-\nu)=1 D − 1 + 2 ( 1 − ν ) = 1 , substitution yields
E D ( ϕ D , k ) = 2 J ν − 1 ( z ) 2 ∫ 0 1 r J ν ( z r ) 2 d r = 1. E_D(\phi_{D,k})=\frac{2}{J_{\nu-1}(z)^2}\int_0^1 rJ_\nu(zr)^2\,\mathrm{d}r=1. E D ( ϕ D , k ) = J ν − 1 ( z ) 2 2 ∫ 0 1 r J ν ( z r ) 2 d r = 1.
The weak eigenvalue equation gives Q D ( ϕ D , k , v ) = j ν , k 2 E D ( ϕ D , k , v ) Q_D(\phi_{D,k},v)=j_{\nu,k}^2E_D(\phi_{D,k},v) Q D ( ϕ D , k , v ) = j ν , k 2 E D ( ϕ D , k , v ) . By symmetry, distinct modes are E D E_D E D -orthogonal. Subtracting the two weak equations, we have that
( j ν , i 2 − j ν , j 2 ) E D ( ϕ D , i , ϕ D , j ) = 0. \left(j_{\nu,i}^2-j_{\nu,j}^2\right)E_D(\phi_{D,i},\phi_{D,j})=0. ( j ν , i 2 − j ν , j 2 ) E D ( ϕ D , i , ϕ D , j ) = 0.
Thus they form the energy-orthonormal radial eigenbasis supplied by the compact variational problem.
Express the value constraint by a scalar resolvent.
The evaluation functional ℓ ( v ) = v ( a ) \ell(v)=v(a) ℓ ( v ) = v ( a ) is bounded on this radial H 2 H^2 H 2 form space. Indeed, a > 0 a>0 a > 0 , and ordinary one-dimensional H 2 H^2 H 2 regularity holds in a neighborhood of a a a . Taking a smooth radial bump function equal to one at a a a shows that ℓ \ell ℓ is nonzero. For real x x x outside the radial ball spectrum, define
g D ( x ; a ) = ∑ k ≥ 1 ϕ D , k ( a ) 2 j ν , k 2 − x . g_D(x;a)=\sum_{k\geq1}\frac{\phi_{D,k}(a)^2}{j_{\nu,k}^2-x}. g D ( x ; a ) = k ≥ 1 ∑ j ν , k 2 − x ϕ D , k ( a ) 2 .
The evaluation functional is bounded with respect to the Q D 1 / 2 Q_D^{1/2} Q D 1/2 norm, so
∑ k ≥ 1 ϕ D , k ( a ) 2 j ν , k 2 < + ∞ . \sum_{k\geq1}\frac{\phi_{D,k}(a)^2}{j_{\nu,k}^2}<+\infty. k ≥ 1 ∑ j ν , k 2 ϕ D , k ( a ) 2 < + ∞.
For completeness, the Q D Q_D Q D -orthonormal basis is ϕ D , k / j ν , k \phi_{D,k}/j_{\nu,k} ϕ D , k / j ν , k , and the sum above is the square of the Q D Q_D Q D norm of the Riesz representer of ℓ \ell ℓ . For x x x outside the radial ball spectrum, set
w x = ∑ k ≥ 1 ϕ D , k ( a ) j ν , k 2 − x ϕ D , k . w_x=\sum_{k\geq1}\frac{\phi_{D,k}(a)}{j_{\nu,k}^2-x}\phi_{D,k}. w x = k ≥ 1 ∑ j ν , k 2 − x ϕ D , k ( a ) ϕ D , k .
This series converges in the radial form space since
Q D ( w x ) = ∑ k ≥ 1 j ν , k 2 ϕ D , k ( a ) 2 ( j ν , k 2 − x ) 2 < + ∞ . Q_D(w_x)=\sum_{k\geq1}\frac{j_{\nu,k}^2\phi_{D,k}(a)^2}{(j_{\nu,k}^2-x)^2}<+\infty. Q D ( w x ) = k ≥ 1 ∑ ( j ν , k 2 − x ) 2 j ν , k 2 ϕ D , k ( a ) 2 < + ∞.
Indeed, for all sufficiently large k k k , the summand is bounded by 4 ϕ D , k ( a ) 2 / j ν , k 2 4\phi_{D,k}(a)^2/j_{\nu,k}^2 4 ϕ D , k ( a ) 2 / j ν , k 2 , and only finitely many terms remain. Checking term by term on finite modes and then passing to the form norm limit gives
( Q D − x E D ) ( w x , v ) = ℓ ( v ) , ℓ ( w x ) = g D ( x ; a ) . (Q_D-xE_D)(w_x,v)=\ell(v),\qquad\ell(w_x)=g_D(x;a). ( Q D − x E D ) ( w x , v ) = ℓ ( v ) , ℓ ( w x ) = g D ( x ; a ) .
On the form-orthonormal basis ϕ D , k / j ν , k \phi_{D,k}/j_{\nu,k} ϕ D , k / j ν , k , the diagonal entries of the operator corresponding to Q D − x E D Q_D-xE_D Q D − x E D are 1 − x / j ν , k 2 1-x/j_{\nu,k}^2 1 − x / j ν , k 2 . Also,
inf k ≥ 1 ∣ 1 − x j ν , k 2 ∣ > 0 , \inf_{k\geq1}\left|1-\frac{x}{j_{\nu,k}^2}\right|>0, k ≥ 1 inf 1 − j ν , k 2 x > 0 ,
so it has a bounded inverse on the form space. For x x x in a compact set avoiding the ball spectrum and for all sufficiently large k k k , we have uniformly
ϕ D , k ( a ) 2 ∣ j ν , k 2 − x ∣ ≤ 2 ϕ D , k ( a ) 2 j ν , k 2 , ϕ D , k ( a ) 2 ( j ν , k 2 − x ) 2 ≤ 4 ϕ D , k ( a ) 2 j ν , k 4 . \frac{\phi_{D,k}(a)^2}{\left|j_{\nu,k}^2-x\right|}\leq\frac{2\phi_{D,k}(a)^2}{j_{\nu,k}^2},\qquad
\frac{\phi_{D,k}(a)^2}{\left(j_{\nu,k}^2-x\right)^2}\leq\frac{4\phi_{D,k}(a)^2}{j_{\nu,k}^4}. j ν , k 2 − x ϕ D , k ( a ) 2 ≤ j ν , k 2 2 ϕ D , k ( a ) 2 , ( j ν , k 2 − x ) 2 ϕ D , k ( a ) 2 ≤ j ν , k 4 4 ϕ D , k ( a ) 2 .
These summable bounds justify the series for g D g_D g D and its differentiated series. Between B 0 = j ν , 1 2 B_0=j_{\nu,1}^2 B 0 = j ν , 1 2 and ρ D ( a ) \rho_D(a) ρ D ( a ) ,
g D ( ρ D ; a ) = 0 , g D ′ ( x ; a ) = ∑ k ≥ 1 ϕ D , k ( a ) 2 ( j D / 2 , k 2 − x ) 2 > 0. g_D(\rho_D;a)=0,\qquad g_D'(x;a)=\sum_{k\geq1}\frac{\phi_{D,k}(a)^2}{\left(j_{D/2,k}^2-x\right)^2}>0. g D ( ρ D ; a ) = 0 , g D ′ ( x ; a ) = k ≥ 1 ∑ ( j D /2 , k 2 − x ) 2 ϕ D , k ( a ) 2 > 0.
We next verify the zeros of g D g_D g D . Recall from § 3 that ρ D \rho_D ρ D lies strictly between j ν , 1 2 j_{\nu,1}^{2} j ν , 1 2 and j ν , 2 2 j_{\nu,2}^{2} j ν , 2 2 . Let v ρ v_\rho v ρ be a nonzero constrained eigenfunction. Consider the continuous functional
v ⟼ ( Q D − ρ D E D ) ( v ρ , v ) . v \longmapsto(Q_D-\rho_D E_D)(v_\rho,v). v ⟼ ( Q D − ρ D E D ) ( v ρ , v ) .
It vanishes on the codimension-one subspace ker l \ker l ker l , and hence must be a scalar multiple of l ( v ) l(v) l ( v ) . This multiple cannot vanish, otherwise v ρ v_\rho v ρ would be a ball eigenfunction corresponding to the non-ball eigenvalue ρ D \rho_D ρ D . By the invertibility of Q D − ρ D E D Q_D-\rho_D E_D Q D − ρ D E D , v ρ v_\rho v ρ is a nonzero scalar multiple of w ρ D w_{\rho_D} w ρ D . Evaluating at a a a gives g D ( ρ D ; a ) = 0 g_D(\rho_D;a)=0 g D ( ρ D ; a ) = 0 , as claimed. The first coefficient ϕ D , 1 ( a ) \phi_{D,1}(a) ϕ D , 1 ( a ) is nonzero, as will be checked quantitatively in the next step, so the derivative series is strictly positive. Since B 0 < P < V ≤ ρ D B_0<P<V\leq\rho_D B 0 < P < V ≤ ρ D , we obtain g D ( P ; a ) < 0 g_D(P;a)<0 g D ( P ; a ) < 0 .
Step 3. Quantify the negative value at P P P .
For the first normalized ball mode, one has
ϕ D , 1 ′ ( r ) = − 2 r 1 − D / 2 J D / 2 ( α D r ) J D / 2 ′ ( α D ) . \phi'_{D,1}(r)=-\sqrt{2}r^{1-D/2}\frac{J_{D/2}(\alpha_Dr)}{J'_{D/2}(\alpha_D)}. ϕ D , 1 ′ ( r ) = − 2 r 1 − D /2 J D /2 ′ ( α D ) J D /2 ( α D r ) .
For r ∈ [ a , 1 ] r\in[a,1] r ∈ [ a , 1 ] , the local estimate applies since
∣ α D ( r − 1 ) ∣ ≤ α D h ≤ m h ≤ 2 − 32 m − 2 < 1 16 . \left|\alpha_D(r-1)\right|\leq\alpha_Dh\leq mh\leq2^{-32}m^{-2}<\frac{1}{16}. ∣ α D ( r − 1 ) ∣ ≤ α D h ≤ mh ≤ 2 − 32 m − 2 < 16 1 .
Using (22) , we see that
J D / 2 ( α D r ) J D / 2 ′ ( α D ) = − α D ( 1 − r ) A ( − α D ( 1 − r ) ) , \frac{J_{D/2}(\alpha_Dr)}{J'_{D/2}(\alpha_D)}
=-\alpha_D(1-r)\mathcal{A}(-\alpha_D(1-r)), J D /2 ′ ( α D ) J D /2 ( α D r ) = − α D ( 1 − r ) A ( − α D ( 1 − r )) ,
where A ≥ 1 − 5 m h > 1 / 2 \mathcal{A}\geq1-5mh>1/2 A ≥ 1 − 5 mh > 1/2 . Hence
ϕ D , 1 ′ ( r ) = 2 α D r 1 − D / 2 ( 1 − r ) A ( − α D ( 1 − r ) ) ≥ α D ( 1 − r ) 2 . \phi'_{D,1}(r)=\sqrt{2}\alpha_Dr^{1-D/2}(1-r)\mathcal{A}(-\alpha_D(1-r))
\geq\frac{\alpha_D(1-r)}{\sqrt{2}}. ϕ D , 1 ′ ( r ) = 2 α D r 1 − D /2 ( 1 − r ) A ( − α D ( 1 − r )) ≥ 2 α D ( 1 − r ) .
Since ϕ D , 1 ( 1 ) = 0 \phi_{D,1}(1)=0 ϕ D , 1 ( 1 ) = 0 ,
− ϕ D , 1 ( a ) = ∫ a 1 ϕ D , 1 ′ ( r ) d r ≥ α D 2 ∫ a 1 ( 1 − r ) d r = α D h 2 2 2 . -\phi_{D,1}(a)=\int_a^1\phi'_{D,1}(r)\,dr\geq\frac{\alpha_D}{\sqrt{2}}\int_a^1(1-r)\,dr=\frac{\alpha_Dh^2}{2\sqrt{2}}. − ϕ D , 1 ( a ) = ∫ a 1 ϕ D , 1 ′ ( r ) d r ≥ 2 α D ∫ a 1 ( 1 − r ) d r = 2 2 α D h 2 .
Therefore,
ϕ D , 1 ( a ) 2 ≥ α D 2 h 4 8 = B 0 h 4 8 . \phi_{D,1}(a)^2\geq\frac{\alpha_D^2h^4}{8}=\frac{B_0h^4}{8}. ϕ D , 1 ( a ) 2 ≥ 8 α D 2 h 4 = 8 B 0 h 4 .
The expansion in § 3 gives 0 < ρ D − B 0 ≤ 4 B 0 h 0<\rho_D-B_0\leq4B_0h 0 < ρ D − B 0 ≤ 4 B 0 h . Retaining only the first nonnegative summand in the derivative series, for P ≤ x ≤ ρ D P\leq x\leq\rho_D P ≤ x ≤ ρ D we have
g D ′ ( x ; a ) ≥ ϕ D , 1 ( a ) 2 ( x − B 0 ) 2 ≥ B 0 h 4 / 8 ( 4 B 0 h ) 2 = h 2 128 B 0 . g'_D(x;a)\geq\frac{\phi_{D,1}(a)^2}{(x-B_0)^2}\geq\frac{B_0h^4/8}{(4B_0h)^2}=\frac{h^2}{128B_0}. g D ′ ( x ; a ) ≥ ( x − B 0 ) 2 ϕ D , 1 ( a ) 2 ≥ ( 4 B 0 h ) 2 B 0 h 4 /8 = 128 B 0 h 2 .
Since ρ D − P ≥ V − P = δ / 2 \rho_D-P\ge V-P=\delta/2 ρ D − P ≥ V − P = δ /2 and g D ( ρ D ; a ) = 0 g_D(\rho_D;a)=0 g D ( ρ D ; a ) = 0 ,
− g D ( P ; a ) = ∫ P ρ D g D ′ ( x ; a ) d x ≥ ( ρ D − P ) h 2 128 B 0 ≥ δ h 2 256 B 0 = 3 h 4 8192 . -g_D(P;a)=\int_P^{\rho_D}g_D'(x;a)\,\mathrm{d}x
\ge\frac{(\rho_D-P)h^2}{128B_0}\ge\frac{\delta h^2}{256B_0}=\frac{3h^4}{8192}. − g D ( P ; a ) = ∫ P ρ D g D ′ ( x ; a ) d x ≥ 128 B 0 ( ρ D − P ) h 2 ≥ 256 B 0 δ h 2 = 8192 3 h 4 .
By Bernoulli’s inequality and the prescribed value of κ \kappa κ ,
a d + 1 = ( 1 − h ) d + 1 ≥ 1 − ( d + 1 ) h ≥ 1 − m h ≥ 1 − 2 − 32 m − 2 > 1 2 , a^{d+1}=(1-h)^{d+1}\ge1-(d+1)h\ge1-mh\ge1-2^{-32}m^{-2}>\frac{1}{2}, a d + 1 = ( 1 − h ) d + 1 ≥ 1 − ( d + 1 ) h ≥ 1 − mh ≥ 1 − 2 − 32 m − 2 > 2 1 ,
− κ a d + 1 g D ( P ; a ) > 2 13 h − 4 ⋅ 1 2 ⋅ 3 h 4 8192 = 3 2 . -\kappa a^{d+1}g_D(P;a)>2^{13}h^{-4}\cdot\frac{1}{2}\cdot\frac{3h^4}{8192}=\frac{3}{2}. − κ a d + 1 g D ( P ; a ) > 2 13 h − 4 ⋅ 2 1 ⋅ 8192 3 h 4 = 2 3 .
This bound ensures − κ a d + 1 g D ( P ; a ) > 1 -\kappa a^{d+1}g_D(P;a)>1 − κ a d + 1 g D ( P ; a ) > 1 provided κ h 4 ≥ 16384 / 3 \kappa h^4\ge16384/3 κ h 4 ≥ 16384/3 . Since
2 12 < 16384 3 < 2 13 , 2^{12}<\frac{16384}{3}<2^{13}, 2 12 < 3 16384 < 2 13 ,
2 13 2^{13} 2 13 is the smallest dyadic coefficient permitted by this sufficient condition.
Here rank one means that the added form depends only on the scalar l ( v ) l(v) l ( v ) . In the effective radial space, write b P = Q D − P E D b_P=Q_D-PE_D b P = Q D − P E D and set w = w P w=w_P w = w P and g = l ( w ) = g D ( P ; a ) < 0 g=l(w)=g_D(P;a)<0 g = l ( w ) = g D ( P ; a ) < 0 . For any v v v , put t = l ( v ) / g t=l(v)/g t = l ( v ) / g and v 0 = v − t w v_0=v-tw v 0 = v − tw . Then v 0 ∈ ker l v_0\in\ker l v 0 ∈ ker l , and this decomposition is unique. From the defining equation for w w w ,
b P ( w , v 0 ) = l ( v 0 ) = 0 , b P ( w , w ) = g . b_P(w,v_0)=l(v_0)=0,\qquad b_P(w,w)=g. b P ( w , v 0 ) = l ( v 0 ) = 0 , b P ( w , w ) = g .
Thus for every τ > 0 \tau>0 τ > 0 ,
( b P + τ l 2 ) ( v , v ) = b P ( v 0 , v 0 ) + t 2 g ( 1 + τ g ) . (b_P+\tau l^2)(v,v)=b_P(v_0,v_0)+t^2g(1+\tau g). ( b P + τ l 2 ) ( v , v ) = b P ( v 0 , v 0 ) + t 2 g ( 1 + τ g ) .
By the constrained Rayleigh principle,
b P ( v 0 , v 0 ) ≥ ( ρ D − P ) E D ( v 0 ) > 0 ( v 0 ≠ 0 ) . b_P(v_0,v_0)\ge(\rho_D-P)E_D(v_0)>0\qquad(v_0\ne0). b P ( v 0 , v 0 ) ≥ ( ρ D − P ) E D ( v 0 ) > 0 ( v 0 = 0 ) .
Since g < 0 g<0 g < 0 , for t ≠ 0 t\ne0 t = 0 the second term is positive if and only if 1 + τ g < 0 1+\tau g<0 1 + τ g < 0 . This sign condition is exactly what is needed to remove the only possible negative direction of b P b_P b P . Take a physical degree-one harmonic normalized on S d − 1 S^{d-1} S d − 1 , so that
∫ S a ∣ r v ( r ) Y 1 ∣ 2 d S = a d + 1 v ( a ) 2 . \int_{S_a}|rv(r)Y_1|^2\,\mathrm{d}S=a^{d+1}v(a)^2. ∫ S a ∣ r v ( r ) Y 1 ∣ 2 d S = a d + 1 v ( a ) 2 .
Hence τ = κ a d + 1 \tau=\kappa a^{d+1} τ = κ a d + 1 , and Step 3 gives 1 + τ g < − 1 / 2 1+\tau g<-1/2 1 + τ g < − 1/2 . For v ≠ 0 v\ne0 v = 0 there are two cases:
t ≠ 0 ⟹ t 2 g ( 1 + τ g ) > 0 , b P ( v 0 , v 0 ) ≥ 0 , t = 0 ⟹ v 0 = v ≠ 0 , b P ( v 0 , v 0 ) > 0. \begin{aligned}
t\ne0&\quad\Longrightarrow\quad t^2g(1+\tau g)>0,\qquad b_P(v_0,v_0)\ge0,\\
t=0&\quad\Longrightarrow\quad v_0=v\ne0,\qquad b_P(v_0,v_0)>0.
\end{aligned} t = 0 t = 0 ⟹ t 2 g ( 1 + τ g ) > 0 , b P ( v 0 , v 0 ) ≥ 0 , ⟹ v 0 = v = 0 , b P ( v 0 , v 0 ) > 0.
In both cases ( b P + τ ℓ 2 ) ( v , v ) > 0 (b_P+\tau\ell^2)(v,v)>0 ( b P + τ ℓ 2 ) ( v , v ) > 0 . The first Rayleigh minimum is attained by compactness, so the first eigenvalue is strictly larger than P P P .
The scalar picture in Figure 8 also uses the same resolvent representation. For x x x outside the ball spectrum, a penalized eigenfunction satisfies
Figure 8. The increasing resolvent on the first radial gap (schematic). The zero is ρ D \rho_D ρ D , and its intersection with − 1 / τ -1/\tau − 1/ τ is the penalized eigenvalue, which lies above P P P .
( Q D − x E D ) ( v , ⋅ ) = − τ ℓ ( v ) ℓ ( ⋅ ) , v = − τ ℓ ( v ) w x . (Q_D-xE_D)(v,\cdot)=-\tau\ell(v)\ell(\cdot), \qquad v=-\tau\ell(v)w_x. ( Q D − x E D ) ( v , ⋅ ) = − τ ℓ ( v ) ℓ ( ⋅ ) , v = − τ ℓ ( v ) w x .
Here ℓ ( v ) ≠ 0 \ell(v)\ne0 ℓ ( v ) = 0 , since otherwise invertibility would force v = 0 v=0 v = 0 . Evaluating at a a a gives
1 + τ g D ( x ; a ) = 0. 1+\tau g_D(x;a)=0. 1 + τ g D ( x ; a ) = 0.
Conversely, this scalar equation makes the nonzero function w x w_x w x a penalized eigenfunction. By strict monotonicity of g D g_D g D and
g D ( P ; a ) < − 1 / τ < g D ( ρ D ; a ) = 0 , g_D(P;a)<-1/\tau<g_D(\rho_D;a)=0, g D ( P ; a ) < − 1/ τ < g D ( ρ D ; a ) = 0 ,
the corresponding eigenvalue lies in ( P , ρ D ) (P,\rho_D) ( P , ρ D ) . It is the first radial eigenvalue, since the second penalized eigenvalue is no smaller than the second ball eigenvalue and the latter lies above ρ D \rho_D ρ D .
Step 5. Count the low eigenvalues in the full space.
The penalty is rotationally invariant, so the angular sectors remain decoupled. The first eigenvalue in degree zero is still at most U < P U<P U < P . Indeed, the surface penalty of a constrained first eigenfunction vanishes, and its Rayleigh quotient is at most U U U . The second eigenvalue and all eigenvalues in degrees l ≥ 2 l\ge2 l ≥ 2 remain above P P P . Since P < V < B 0 + 1 P<V<B_0+1 P < V < B 0 + 1 , and their unconstrained lower bounds have already been listed in § 3 , these bounds persist after adding a nonnegative penalty. Hence the degree-one sector contributes no eigenvalue at or below P P P , and the full space has exactly one eigenvalue at or below this level. Therefore β 2 ( κ ) > P \beta_2(\kappa)>P β 2 ( κ ) > P .
Sampling the trace on latitude sets
We now transfer the penalized lower bound to functions that vanish on the finite core set. The required comparison follows by estimating the traces between adjacent latitudes. Define the corresponding variational space and eigenvalues by
H N = C c ∞ ( B ∖ K N ) ‾ H 2 ( B ) , λ k , N = λ k ( H N ) . H_N=\overline{C_c^\infty(B\setminus K_N)}^{H^2(B)}, \qquad\lambda_{k,N}=\lambda_k(H_N). H N = C c ∞ ( B ∖ K N ) H 2 ( B ) , λ k , N = λ k ( H N ) .
This is a closed subspace of H 0 2 ( B ) H^2_0(B) H 0 2 ( B ) and is infinite-dimensional. Indeed, it contains the C c ∞ C^\infty_c C c ∞ functions of any small ball whose closure lies in B ∖ K N B \setminus K_N B ∖ K N . Write Δ θ = π 2 − N \Delta_\theta= \pi2^{-N} Δ θ = π 2 − N for the latitude spacing. The sampling estimate below shows that this finite constraint set still preserves the lower threshold T T T for the second eigenvalue. Figure 9 shows the interior band and the two polar caps used in the estimate.
Figure 9. Latitude sampling: bounded weight ratios on interior bands and weighted integration on polar caps. No value is imposed at a pole; for d = 2 d=2 d = 2 the weight is one.
Proposition 4.2. For the explicit parameters h h h , a a a , N N N in § 2 and the threshold T T T in (31) , the second eigenvalue on H N H_N H N satisfies
λ 2 , N > T . \lambda_{2,N} > T. λ 2 , N > T .
Proof. Step 1. Identify the vanishing values.
As in Proposition 4.1, we fix κ = 2 13 h − 4 \kappa= 2^{13}h^{-4} κ = 2 13 h − 4 . By the trace theorem on smooth submanifolds, restriction from H 2 ( B ) H^2(B) H 2 ( B ) to the codimension-two cores maps continuously into H 1 ( M j ) H^1(M_j) H 1 ( M j ) [1 ] [2]. In the zero-dimensional case this is point evaluation. Hence every element of H N H_N H N has zero trace on each M j M_j M j . Next consider the trace on S a S_a S a . We first restrict to the inner ball and use its smooth boundary:
H 2 ( B ) ⟶ H 2 ( B a ( 0 ) ) → Tr S a H 3 / 2 ( S a ) . H^2(B) \longrightarrow H^2(B_a(0)) \xrightarrow{\operatorname{Tr}_{S_a}} H^{3/2}(S_a). H 2 ( B ) ⟶ H 2 ( B a ( 0 )) Tr S a H 3/2 ( S a ) .
Both arrows are continuous. The second follows from [17 Theorem 3.37, p. 102] with k = 3 k=3 k = 3 and s = 2 s=2 s = 2 . Restricting this trace further to M j M_j M j agrees with direct restriction. The two continuous maps coincide on smooth functions. On the latitude band, this also identifies the endpoint values of the Sobolev sections with values in L 2 ( S d − 2 ) L^2(S^{d-2}) L 2 ( S d − 2 ) . The estimates below do not use the constants in these qualitative trace theorems.
Put s N = 2 d − 2 a 2 Δ θ 2 s_N = 2^{d-2}a^2\Delta_\theta^2 s N = 2 d − 2 a 2 Δ θ 2 . Writing ∇ T \nabla_T ∇ T for the tangential gradient on S a S_a S a , we shall prove the weighted one-dimensional sampling estimate
∥ u ∣ S a ∥ 2 2 ≤ s N ∥ ∇ T u ∣ S a ∥ 2 2 . (34) \left\|u|_{S_a}\right\|_2^2 \le s_N \left\|\nabla_Tu|_{S_a}\right\|_2^2. \tag*{(34)} ∥ u ∣ S a ∥ 2 2 ≤ s N ∥ ∇ T u ∣ S a ∥ 2 2 . (34)
Step 2. Estimate each interior latitude band.
Fix ω ∈ S d − 2 \omega\in S^{d-2} ω ∈ S d − 2 and write
f ( θ ) = u ( a sin θ ω , a cos θ ) , 0 < θ < π . f(\theta)=u(a\sin\theta\,\omega,a\cos\theta), \qquad0<\theta<\pi. f ( θ ) = u ( a sin θ ω , a cos θ ) , 0 < θ < π .
The surface measure is a d − 1 sin d − 2 θ d θ d S ( ω ) a^{d-1}\sin^{d-2}\theta\,d\theta\,dS(\omega) a d − 1 sin d − 2 θ d θ d S ( ω ) , and ∣ f ′ ( θ ) ∣ ≤ a ∣ ∇ T u ∣ |f'(\theta)| \le a|\nabla_T u| ∣ f ′ ( θ ) ∣ ≤ a ∣ ∇ T u ∣ . On an interior grid interval [ θ j , θ j + 1 ] [\theta_j,\theta_{j+1}] [ θ j , θ j + 1 ] , the function f f f vanishes at both endpoints. Starting from the left endpoint and applying Cauchy–Schwarz gives
∣ f ( θ ) ∣ 2 ≤ ( θ − θ j ) ∫ θ j θ j + 1 ∣ f ′ ( t ) ∣ 2 d t . |f(\theta)|^2 \le(\theta-\theta_j)\int_{\theta_j}^{\theta_{j+1}} |f'(t)|^2\,dt. ∣ f ( θ ) ∣ 2 ≤ ( θ − θ j ) ∫ θ j θ j + 1 ∣ f ′ ( t ) ∣ 2 d t .
Integrating in θ \theta θ yields an unweighted bound with constant Δ θ 2 / 2 \Delta_\theta^2/2 Δ θ 2 /2 , and in particular with the larger constant Δ θ 2 \Delta_\theta^2 Δ θ 2 . On each interior interval, the ratio of the largest to the smallest value of the sine factor is at most 2. On the half of the sphere where the sine factor is increasing, this is immediate from the decrease of sin t / t \sin t/t sin t / t on ( 0 , π / 2 ] (0,\pi/2] ( 0 , π /2 ] :
sin ( ( j + 1 ) Δ θ ) sin ( j Δ θ ) ≤ j + 1 j ≤ 2. \frac{\sin((j+1)\Delta_\theta)}{\sin(j\Delta_\theta)} \le\frac{j+1}{j} \le2. sin ( j Δ θ ) sin (( j + 1 ) Δ θ ) ≤ j j + 1 ≤ 2.
The decreasing half is handled by reflection about π / 2 \pi/2 π /2 , and π / 2 \pi/2 π /2 itself is a grid point. Hence the ratio of the maximum to the minimum of the weight sin d − 2 θ \sin^{d-2}\theta sin d − 2 θ is at most 2 d − 2 2^{d-2} 2 d − 2 . Multiplying the unweighted estimate by the maximum weight and then inserting the minimum weight on the right, we obtain
∫ θ j θ j + 1 sin d − 2 θ ∣ f ∣ 2 d θ ≤ 2 d − 2 Δ θ 2 ∫ θ j θ j + 1 sin d − 2 θ ∣ f ′ ∣ 2 d θ . \int_{\theta_j}^{\theta_{j+1}} \sin^{d-2}\theta\,|f|^2\,d\theta\le2^{d-2}\Delta_\theta^2\int_{\theta_j}^{\theta_{j+1}} \sin^{d-2}\theta\,|f'|^2\,d\theta. ∫ θ j θ j + 1 sin d − 2 θ ∣ f ∣ 2 d θ ≤ 2 d − 2 Δ θ 2 ∫ θ j θ j + 1 sin d − 2 θ ∣ f ′ ∣ 2 d θ .
Step 3. Estimate the polar caps without a value at the pole.
For d > 2 d>2 d > 2 the weight vanishes at the north pole, so comparison with its minimum is unavailable. Put q = d − 2 q=d-2 q = d − 2 and instead use the following elementary estimate:
∫ 0 Δ θ t q f 2 d t ≤ Δ θ 2 2 ( q + 1 ) ∫ 0 Δ θ ( f ′ ) 2 d t , f ( Δ θ ) = 0. \int_0^{\Delta_\theta} t^q f^2\,dt \le\frac{\Delta_\theta^2}{2(q+1)}\int_0^{\Delta_\theta} (f')^2\,dt,\qquad f(\Delta_\theta)=0. ∫ 0 Δ θ t q f 2 d t ≤ 2 ( q + 1 ) Δ θ 2 ∫ 0 Δ θ ( f ′ ) 2 d t , f ( Δ θ ) = 0.
To verify this, write f ( t ) = − ∫ t Δ θ f ′ ( r ) d r f(t)=-\int_t^{\Delta_\theta} f'(r)\,dr f ( t ) = − ∫ t Δ θ f ′ ( r ) d r and use
∣ f ( t ) ∣ 2 ≤ ( ∫ t Δ θ r − q d r ) ( ∫ t Δ θ r q ∣ f ′ ( r ) ∣ 2 d r ) . |f(t)|^2 \le\left(\int_t^{\Delta_\theta} r^{-q}\,dr\right)\left(\int_t^{\Delta_\theta} r^q|f'(r)|^2\,dr\right). ∣ f ( t ) ∣ 2 ≤ ( ∫ t Δ θ r − q d r ) ( ∫ t Δ θ r q ∣ f ′ ( r ) ∣ 2 d r ) .
The second factor is controlled by its integral over the whole interval, while the first factor is integrated against t q d t t^q\,dt t q d t . By Fubini’s theorem,
∫ 0 Δ θ t q ∫ t Δ θ r − q d r d t = 1 q + 1 ∫ 0 Δ θ r d r = Δ θ 2 2 ( q + 1 ) . \int_0^{\Delta_\theta} t^q\int_t^{\Delta_\theta} r^{-q}\,dr\,dt
=\frac{1}{q+1}\int_0^{\Delta_\theta} r\,dr
=\frac{\Delta_\theta^2}{2(q+1)}. ∫ 0 Δ θ t q ∫ t Δ θ r − q d r d t = q + 1 1 ∫ 0 Δ θ r d r = 2 ( q + 1 ) Δ θ 2 .
The sine weight is recovered from two independent inequalities:
∫ 0 Δ θ sin q t ∣ f ∣ 2 d t ≤ ∫ 0 Δ θ t q ∣ f ∣ 2 d t ≤ Δ θ 2 2 ( q + 1 ) ( π 2 ) q ∫ 0 Δ θ sin q t ∣ f ′ ∣ 2 d t ≤ 2 q Δ θ 2 ∫ 0 Δ θ sin q t ∣ f ′ ∣ 2 d t . \begin{aligned}
\int_0^{\Delta_\theta}\sin^q t\,|f|^2\,dt
&\le\int_0^{\Delta_\theta}t^q|f|^2\,dt\\
&\le\frac{\Delta_\theta^2}{2(q+1)}\left(\frac{\pi}{2}\right)^q\int_0^{\Delta_\theta}\sin^q t\,|f'|^2\,dt\\
&\le2^q\Delta_\theta^2\int_0^{\Delta_\theta}\sin^q t\,|f'|^2\,dt.
\end{aligned} ∫ 0 Δ θ sin q t ∣ f ∣ 2 d t ≤ ∫ 0 Δ θ t q ∣ f ∣ 2 d t ≤ 2 ( q + 1 ) Δ θ 2 ( 2 π ) q ∫ 0 Δ θ sin q t ∣ f ′ ∣ 2 d t ≤ 2 q Δ θ 2 ∫ 0 Δ θ sin q t ∣ f ′ ∣ 2 d t .
Here 2 t / π ≤ sin t ≤ t 2t/\pi\le\sin t \le t 2 t / π ≤ sin t ≤ t , and ( π / 2 ) q 2 ( q + 1 ) ≤ 2 q \frac{(\pi/2)q}{2(q+1)} \le2^q 2 ( q + 1 ) ( π /2 ) q ≤ 2 q for q ≥ 0 q \ge0 q ≥ 0 . The same estimate at the south pole follows by reflection. Adding the contributions of the bands and the polar caps, multiplying by a d − 1 a^{d-1} a d − 1 , integrating in ω \omega ω , and using ∣ f ′ ∣ ≤ a ∣ ∇ T u ∣ |f'| \le a|\nabla_T u| ∣ f ′ ∣ ≤ a ∣ ∇ T u ∣ gives exactly (4.2). The factor a 2 a^2 a 2 in s N s_N s N comes from passing from the angular derivative to the tangential derivative. For nonsmooth traces the same proof applies directly to the one-dimensional Sobolev sections on each interior interval. On a cap we first use the weighted argument away from the pole and then let that endpoint tend to zero. This requires no zero value at the pole and introduces no boundary term there. Integrating the section inequalities over S d − 2 S^{d-2} S d − 2 proves (4.2). The same conclusion holds when d = 2 d=2 d = 2 and S 0 S^0 S 0 consists of two points.
To turn (4.2) into a comparison of quadratic forms, we estimate the gradient trace in a shell inside S a S_a S a . For a scalar or vector function g g g on [ a − h , a ] [a-h,a] [ a − h , a ] and a fixed t t t in this interval, the fundamental theorem of calculus gives
∣ g ( a ) ∣ 2 = ∣ g ( t ) ∣ 2 + 2 ∫ t a g ⋅ g ′ d r ≤ ∣ g ( t ) ∣ 2 + 1 h ∫ a − h a ∣ g ∣ 2 d r + h ∫ a − h a ∣ g ′ ∣ 2 d r . |g(a)|^2 = |g(t)|^2 + 2\int_t^a g\cdot g'\,\mathrm{d}r \le|g(t)|^2 + \frac{1}{h}\int_{a-h}^a |g|^2\,\mathrm{d}r + h\int_{a-h}^a |g'|^2\,\mathrm{d}r. ∣ g ( a ) ∣ 2 = ∣ g ( t ) ∣ 2 + 2 ∫ t a g ⋅ g ′ d r ≤ ∣ g ( t ) ∣ 2 + h 1 ∫ a − h a ∣ g ∣ 2 d r + h ∫ a − h a ∣ g ′ ∣ 2 d r .
Averaging this inequality over t ∈ [ a − h , a ] t\in[a-h,a] t ∈ [ a − h , a ] , the averaged first term contributes an additional h − 1 ∫ a − h a ∣ g ∣ 2 d r h^{-1}\int_{a-h}^a |g|^2\,\mathrm{d}r h − 1 ∫ a − h a ∣ g ∣ 2 d r , and hence
∣ g ( a ) ∣ 2 ≤ 2 h ∫ a − h a ∣ g ∣ 2 d r + h ∫ a − h a ∣ g ′ ∣ 2 d r . |g(a)|^2 \le\frac{2}{h}\int_{a-h}^a |g|^2\,\mathrm{d}r + h\int_{a-h}^a |g'|^2\,\mathrm{d}r. ∣ g ( a ) ∣ 2 ≤ h 2 ∫ a − h a ∣ g ∣ 2 d r + h ∫ a − h a ∣ g ′ ∣ 2 d r .
Here a − h = 1 − 2 h > 0 a-h=1-2h>0 a − h = 1 − 2 h > 0 , and
log a a − h ≤ h 1 − 2 h , ( a a − h ) d − 1 < 2. \log\frac{a}{a-h}\le\frac{h}{1-2h},\qquad\left(\frac{a}{a-h}\right)^{d-1}<2. log a − h a ≤ 1 − 2 h h , ( a − h a ) d − 1 < 2.
Indeed, a / ( a − h ) = 1 + h / ( 1 − 2 h ) a/(a-h)=1+h/(1-2h) a / ( a − h ) = 1 + h / ( 1 − 2 h ) , and log ( 1 + t ) ≤ t \log(1+t)\le t log ( 1 + t ) ≤ t . Also h < 1 / 4 h<1/4 h < 1/4 and ( d − 1 ) h ≤ m h ≤ 2 − 32 m − 2 (d-1)h\le mh\le2^{-32}m^{-2} ( d − 1 ) h ≤ mh ≤ 2 − 32 m − 2 , so the exponent satisfies
( d − 1 ) h 1 − 2 h < 2 m h ≤ 2 − 31 m − 2 < 1 2 < log 2. \frac{(d-1)h}{1-2h}<2mh\le2^{-31}m^{-2}<\frac{1}{2}<\log2. 1 − 2 h ( d − 1 ) h < 2 mh ≤ 2 − 31 m − 2 < 2 1 < log 2.
This proves the second inequality. Hence every r r r in the shell satisfies a d − 1 ≤ 2 r d − 1 a^{d-1}\le2r^{d-1} a d − 1 ≤ 2 r d − 1 . First for smooth u u u , apply the trace estimate to g ( r ) = D u ( r ω ) g(r)=Du(r\omega) g ( r ) = D u ( r ω ) with ω ∈ S d − 1 \omega\in S^{d-1} ω ∈ S d − 1 , multiply by a d − 1 a^{d-1} a d − 1 , and integrate in ω \omega ω . Since ∣ ∂ r D u ∣ ≤ ∣ D 2 u ∣ |\partial_rDu|\le|D^2u| ∣ ∂ r D u ∣ ≤ ∣ D 2 u ∣ , together with the weight comparison and (2.1), we obtain
∥ D u ∣ S a ∥ 2 ≤ 4 h E ( u ) + 2 h Q ( u ) . \|Du|_{S_a}\|^2 \le\frac{4}{h}\mathcal{E}(u)+2h\mathcal{Q}(u). ∥ D u ∣ S a ∥ 2 ≤ h 4 E ( u ) + 2 h Q ( u ) .
By density and continuity of the gradient trace, this estimate extends to H 0 2 ( B ) H^2_0(B) H 0 2 ( B ) . Hence on H N H_N H N ,
Q κ ( u ) ≤ ( 1 + 2 κ h s N ) Q ( u ) + 4 κ s N h E ( u ) . \mathcal{Q}_\kappa(u)\le(1+2\kappa h s_N)\mathcal{Q}(u)+\frac{4\kappa s_N}{h}\mathcal{E}(u). Q κ ( u ) ≤ ( 1 + 2 κh s N ) Q ( u ) + h 4 κ s N E ( u ) .
The two error coefficients have different origins. The term involving Q \mathcal{Q} Q comes from the radial derivative of D u Du D u , while the term involving E \mathcal{E} E comes from D u Du D u itself.
Step 5. Choose the mesh to fit inside P − T P-T P − T .
By the min-max principle, the second penalized eigenvalue on H N H_N H N is no smaller than the second penalized eigenvalue on H 0 2 ( B ) H_0^2(B) H 0 2 ( B ) . Hence Proposition 4.1 shows that it is larger than P P P . The final form inequality also gives the upper bound ( 1 + 2 κ h s N ) λ 2 , N + 4 κ s N / h (1+2\kappa h s_N)\lambda_{2,N}+4\kappa s_N/h ( 1 + 2 κh s N ) λ 2 , N + 4 κ s N / h for it. More explicitly,
λ 2 , N > P − 4 κ s N / h 1 + 2 κ h s N . \lambda_{2,N}>\frac{P-4\kappa s_N/h}{1+2\kappa h s_N}. λ 2 , N > 1 + 2 κh s N P − 4 κ s N / h .
The denominator is positive, and the right-hand side exceeds T T T if and only if P − T > κ s N ( 4 / h + 2 h T ) P-T>\kappa s_N(4/h+2hT) P − T > κ s N ( 4/ h + 2 h T ) . Substituting s N = 2 d − 2 a 2 π 2 / 4 N s_N=2^{d-2}a^2\pi^2/4^N s N = 2 d − 2 a 2 π 2 / 4 N gives the sufficient condition
4 N > 2 d − 2 a 2 π 2 κ 4 / h + 2 h T P − T . (35) 4^N>2^{d-2}a^2\pi^2\kappa\frac{4/h+2hT}{P-T}. \tag*{(35)} 4 N > 2 d − 2 a 2 π 2 κ P − T 4/ h + 2 h T . (35)
We now check the factors in this condition one by one. First, P − T = δ / 4 > h 2 / 16 P-T=\delta/4>h^2/16 P − T = δ /4 > h 2 /16 . Second, since T < 2 m 2 T<2m^2 T < 2 m 2 ,
4 / h + 2 h T < 4 / h + 4 m 2 h < 8 / h , 4/h+2hT<4/h+4m^2h<8/h, 4/ h + 2 h T < 4/ h + 4 m 2 h < 8/ h ,
where the last step uses m 2 h 2 < 1 m^2h^2<1 m 2 h 2 < 1 . Together with a 2 < 1 a^2<1 a 2 < 1 , π 2 < 16 \pi^2<16 π 2 < 16 , and κ = 2 13 h − 4 \kappa=2^{13}h^{-4} κ = 2 13 h − 4 , the right-hand side of (35) is bounded by
2 d − 2 2 4 2 13 h − 4 2 3 h − 1 2 4 h − 2 = 2 d + 22 h − 7 . 2^{d-2}2^4 2^{13}h^{-4}2^3h^{-1}2^4h^{-2}=2^{d+22}h^{-7}. 2 d − 2 2 4 2 13 h − 4 2 3 h − 1 2 4 h − 2 = 2 d + 22 h − 7 .
Finally, by the definition of N N N , 2 N ≥ d + 23 + 7 H 2N\ge d+23+7H 2 N ≥ d + 23 + 7 H , so
4 N = 2 2 N ≥ 2 d + 23 + 7 H = 2 d + 23 h − 7 . 4^N=2^{2N}\ge2^{d+23+7H}=2^{d+23}h^{-7}. 4 N = 2 2 N ≥ 2 d + 23 + 7 H = 2 d + 23 h − 7 .
The remaining factor is at least two, and (4.3) therefore holds strictly. The chosen N N N is the smallest integer satisfying this strict coarse inequality 2 N > d + 22 + 7 H 2^N>d+22+7H 2 N > d + 22 + 7 H . This proves λ 2 , N > T \lambda_{2,N}>T λ 2 , N > T .
Derivative bounds for the constrained first mode
Proposition 4.2 already provides the lower bound for the second eigenvalue. To control the first eigenvalue from above on the perforated domain, we truncate the first spherical mode near the cores. The bending error produced by this truncation is controlled by its gradient and Laplacian. The next lemma estimates both explicitly. § 13 uses this bound to choose the cutoff scale and complete the proof.
Lemma 5.1. Let u ∗ u_* u ∗ be the first degree-zero eigenfunction in V a V_a V a , normalized by E ( u ∗ ) = 1 \mathcal{E}(u_*)=1 E ( u ∗ ) = 1 . For m = d + 4 m=d+4 m = d + 4 and the present h h h ,
∥ D u ∗ ∥ ∞ , ∥ Δ u ∗ ∥ ∞ ≤ M : = 2 13 m 2 + 2 H . \lVert Du_*\rVert_\infty,\ \lVert\Delta u_*\rVert_\infty\le M:=2^{13m^2+2H}. ∥ D u ∗ ∥ ∞ , ∥ Δ u ∗ ∥ ∞ ≤ M := 2 13 m 2 + 2 H .
Proof. Step 1. Start with the solution whose slope is one. Set k = ρ d ( a ) k=\sqrt{\rho_d(a)} k = ρ d ( a ) , z = z d ( h ) z=z_d(h) z = z d ( h ) , and let f f f be the outer profile constructed in Lemma 3.2 for D = d D=d D = d . We first take an unnormalized real solution w w w with w ′ ( a ) = 1 w'(a)=1 w ′ ( a ) = 1 . For ν = d / 2 \nu=d/2 ν = d /2 and α = α d \alpha=\alpha_d α = α d , its inner expression is
C [ r 1 − ν J ν − 1 ( k r ) − a 1 − ν J ν − 1 ( k a ) ] , C = − [ k a 1 − ν J ν ( k a ) ] − 1 . C\left[r^{1-\nu}J_{\nu-1}(kr)-a^{1-\nu}J_{\nu-1}(ka)\right], \qquad C=-\left[ka^{1-\nu}J_\nu(ka)\right]^{-1}. C [ r 1 − ν J ν − 1 ( k r ) − a 1 − ν J ν − 1 ( k a ) ] , C = − [ k a 1 − ν J ν ( k a ) ] − 1 .
Since ∣ z ∣ ∈ [ 1 / 8 , 3 / 8 ] |z|\in[1/8,3/8] ∣ z ∣ ∈ [ 1/8 , 3/8 ] , (22) gives
∣ J ν ( k a ) ∣ ≥ ∣ J ν ′ ( α ) ∣ α h / 16. |J_\nu(ka)|\geq|J_\nu'(\alpha)|\alpha h/16. ∣ J ν ( k a ) ∣ ≥ ∣ J ν ′ ( α ) ∣ α h /16.
Indeed, k a = α + α h z ka=\alpha+\alpha hz k a = α + α h z , and
J ν ( k a ) = J ν ′ ( α ) α h z A ( α h z ) , ∣ A ( α h z ) ∣ ≥ 1 / 2. J_\nu(ka)=J_\nu'(\alpha)\alpha hz\,A(\alpha hz), \qquad|A(\alpha hz)|\geq1/2. J ν ( k a ) = J ν ′ ( α ) α h z A ( α h z ) , ∣ A ( α h z ) ∣ ≥ 1/2.
The factor 16 in the denominator comes from 1/8 and 1/2. It remains to find a bound depending only on d d d that keeps the normalization factor J ν ′ ( α ) J_\nu'(\alpha) J ν ′ ( α ) bounded away from zero.
Step 2. Obtain an explicit lower bound for the Bessel slope. We estimate the slope using the Bessel function on [ 1 / 2 , 1 ] [1/2,1] [ 1/2 , 1 ] , where its alternating series is easy to control. For Γ ( t ) = ∫ 0 + ∞ x t − 1 e − x d x \Gamma(t)=\int_0^{+\infty}x^{t-1}e^{-x}\,\mathrm{d}x Γ ( t ) = ∫ 0 + ∞ x t − 1 e − x d x with t > 0 t>0 t > 0 and 1 ≤ ν ≤ m 1\leq\nu\leq m 1 ≤ ν ≤ m , the integer and half-integer cases give respectively
Γ ( ν + 1 ) = ∏ j = 1 ν j ≤ m ν ( ν ∈ N ) , \Gamma(\nu+1)=\prod_{j=1}^{\nu}j\leq m^\nu\qquad(\nu\in\mathbb{N}), Γ ( ν + 1 ) = j = 1 ∏ ν j ≤ m ν ( ν ∈ N ) ,
Γ ( ν + 1 ) = π ∏ j = 0 ν − 1 / 2 ( j + 1 / 2 ) ≤ 2 m ν + 1 / 2 ( ν − 1 / 2 ∈ N ) . \Gamma(\nu+1)=\sqrt{\pi}\prod_{j=0}^{\nu-1/2}(j+1/2)\leq2m^{\nu+1/2}\qquad(\nu-1/2\in\mathbb{N}). Γ ( ν + 1 ) = π j = 0 ∏ ν − 1/2 ( j + 1/2 ) ≤ 2 m ν + 1/2 ( ν − 1/2 ∈ N ) .
Hence
log ( 2 m m + 1 ) = log 2 + ( m + 1 ) log m < 1 + ( m + 1 ) ( m − 1 ) = m 2 , \log(2m^{m+1})=\log2+(m+1)\log m<1+(m+1)(m-1)=m^2, log ( 2 m m + 1 ) = log 2 + ( m + 1 ) log m < 1 + ( m + 1 ) ( m − 1 ) = m 2 ,
Γ ( ν + 1 ) ≤ 2 m m + 1 ≤ e m 2 . \Gamma(\nu+1)\leq2m^{m+1}\leq e^{m^2}. Γ ( ν + 1 ) ≤ 2 m m + 1 ≤ e m 2 .
For 1 / 2 ≤ x ≤ 1 1/2\leq x\leq1 1/2 ≤ x ≤ 1 , the ratio of consecutive absolute terms in the Bessel series satisfies
x 2 4 ( n + 1 ) ( ν + n + 1 ) ≤ 1 4 ( ν + 1 ) ≤ 1 8 , n ≥ 0. \frac{x^2}{4(n+1)(\nu+n+1)}\leq\frac{1}{4(\nu+1)}\leq\frac{1}{8},\qquad n\geq0. 4 ( n + 1 ) ( ν + n + 1 ) x 2 ≤ 4 ( ν + 1 ) 1 ≤ 8 1 , n ≥ 0.
By the alternating series estimate,
Γ ( ν + 1 ) ( 2 / x ) ν J ν ( x ) ≥ 1 − x 2 4 ( ν + 1 ) ≥ 7 8 , \Gamma(\nu+1)(2/x)^\nu J_\nu(x)\geq1-\frac{x^2}{4(\nu+1)}\geq\frac{7}{8}, Γ ( ν + 1 ) ( 2/ x ) ν J ν ( x ) ≥ 1 − 4 ( ν + 1 ) x 2 ≥ 8 7 ,
J ν ( x ) ≥ 7 ( x / 2 ) ν 8 Γ ( ν + 1 ) ≥ 7 4 − ν 8 Γ ( ν + 1 ) ≥ 4 − ν 2 Γ ( ν + 1 ) . J_\nu(x)\geq\frac{7(x/2)^\nu}{8\Gamma(\nu+1)}\geq\frac{7\,4^{-\nu}}{8\Gamma(\nu+1)}\geq\frac{4^{-\nu}}{2\Gamma(\nu+1)}. J ν ( x ) ≥ 8Γ ( ν + 1 ) 7 ( x /2 ) ν ≥ 8Γ ( ν + 1 ) 7 4 − ν ≥ 2Γ ( ν + 1 ) 4 − ν .
Taking z = α z=\alpha z = α in (33) gives
∫ 0 α x J ν ( x ) 2 d x = α 2 ∫ 0 1 r J ν ( α r ) 2 d r = α 2 2 J ν ′ ( α ) 2 . \begin{aligned}
\int_0^\alpha xJ_\nu(x)^2\,\mathrm{d}x=\alpha^2\int_0^1rJ_\nu(\alpha r)^2\,\mathrm{d}r \\
&=\frac{\alpha^2}{2}J_\nu'(\alpha)^2.
\end{aligned} ∫ 0 α x J ν ( x ) 2 d x = α 2 ∫ 0 1 r J ν ( α r ) 2 d r = 2 α 2 J ν ′ ( α ) 2 .
Since α > ν ≥ 1 \alpha>\nu\ge1 α > ν ≥ 1 ,
J ν ′ ( α ) 2 = 2 α 2 ∫ 0 α x J ν ( x ) 2 d x ≥ 2 α 2 4 − 2 ν 4 Γ ( ν + 1 ) 2 ∫ 1 / 2 1 x d x = 3 4 − 2 ν 16 α 2 Γ ( ν + 1 ) 2 ≥ 4 − 2 ν 8 α 2 Γ ( ν + 1 ) 2 . \begin{aligned}
J_\nu'(\alpha)^2=\frac{2}{\alpha^2}\int_0^\alpha xJ_\nu(x)^2\,\mathrm{d}x \\
&\ge\frac{2}{\alpha^2}\frac{4^{-2\nu}}{4\Gamma(\nu+1)^2}\int_{1/2}^1 x\,\mathrm{d}x \\
&=\frac{3\,4^{-2\nu}}{16\alpha^2\Gamma(\nu+1)^2}\ge\frac{4^{-2\nu}}{8\alpha^2\Gamma(\nu+1)^2}.
\end{aligned} J ν ′ ( α ) 2 = α 2 2 ∫ 0 α x J ν ( x ) 2 d x ≥ α 2 2 4Γ ( ν + 1 ) 2 4 − 2 ν ∫ 1/2 1 x d x = 16 α 2 Γ ( ν + 1 ) 2 3 4 − 2 ν ≥ 8 α 2 Γ ( ν + 1 ) 2 4 − 2 ν .
The three factors in the inverse slope estimate satisfy
3 α ≤ 3 m ≤ m 2 ≤ e m 2 , 4 ν ≤ e m log 4 ≤ e m 2 , Γ ( ν + 1 ) ≤ e m 2 . 3\alpha\le3m\le m^2\le e^{m^2},\qquad4^\nu\le e^{m\log4}\le e^{m^2},\qquad\Gamma(\nu+1)\le e^{m^2}. 3 α ≤ 3 m ≤ m 2 ≤ e m 2 , 4 ν ≤ e m l o g 4 ≤ e m 2 , Γ ( ν + 1 ) ≤ e m 2 .
Thus
∣ J ν ′ ( α ) ∣ − 1 ≤ 8 α 4 ν Γ ( ν + 1 ) \left|J_\nu'(\alpha)\right|^{-1}\le\sqrt{8}\,\alpha\,4^\nu\Gamma(\nu+1) ∣ J ν ′ ( α ) ∣ − 1 ≤ 8 α 4 ν Γ ( ν + 1 )
≤ 3 α 4 ν Γ ( ν + 1 ) ≤ e 3 m 2 . \le3\alpha\,4^\nu\Gamma(\nu+1)\le e^{3m^2}. ≤ 3 α 4 ν Γ ( ν + 1 ) ≤ e 3 m 2 .
Also k a 1 − ν α ≥ 1 ka^{1-\nu}\alpha\ge1 k a 1 − ν α ≥ 1 and 16 ≤ e m 2 16\le e^{m^2} 16 ≤ e m 2 , so
∣ C ∣ ≤ 16 ∣ J ν ′ ( α ) ∣ − 1 k a 1 − ν h ≤ 16 e 3 m 2 h ≤ e 4 m 2 h . |C|\le\frac{16|J_\nu'(\alpha)|^{-1}}{ka^{1-\nu}h}\le\frac{16e^{3m^2}}{h}\le\frac{e^{4m^2}}{h}. ∣ C ∣ ≤ k a 1 − ν h 16∣ J ν ′ ( α ) ∣ − 1 ≤ h 16 e 3 m 2 ≤ h e 4 m 2 .
Step 3. Bound both derivatives before energy normalization.
We now control the regularized Bessel factors so that these bounds remain valid at the origin. For t ≥ 0 t\ge0 t ≥ 0 and x > 0 x>0 x > 0 , the power series [23 Chapter III] is
J t ( x ) = ( x / 2 ) t Γ ( t + 1 ) ∑ n ≥ 0 ( − x 2 / 4 ) n n ! ( t + 1 ) n . J_t(x)=\frac{(x/2)^t}{\Gamma(t+1)}\sum_{n\ge0}\frac{(-x^2/4)^n}{n!(t+1)_n}. J t ( x ) = Γ ( t + 1 ) ( x /2 ) t n ≥ 0 ∑ n ! ( t + 1 ) n ( − x 2 /4 ) n .
Here ( t + 1 ) n = ( t + 1 ) ⋯ ( t + n ) (t+1)_n=(t+1)\cdots(t+n) ( t + 1 ) n = ( t + 1 ) ⋯ ( t + n ) for n ≥ 1 n\ge1 n ≥ 1 , and ( t + 1 ) 0 = 1 (t+1)_0=1 ( t + 1 ) 0 = 1 . Therefore,
( t + 1 ) n ≥ 1 , (t+1)_n\ge1, ( t + 1 ) n ≥ 1 ,
∣ ∑ n ≥ 0 ( − x 2 / 4 ) n n ! ( t + 1 ) n ∣ ≤ ∑ n ≥ 0 ( ∣ x ∣ 2 / 4 ) n n ! = e ∣ x ∣ 2 / 4 , \left|\sum_{n\ge0}\frac{(-x^2/4)^n}{n!(t+1)_n}\right|
\le\sum_{n\ge0}\frac{(|x|^2/4)^n}{n!}=e^{|x|^2/4}, n ≥ 0 ∑ n ! ( t + 1 ) n ( − x 2 /4 ) n ≤ n ≥ 0 ∑ n ! ( ∣ x ∣ 2 /4 ) n = e ∣ x ∣ 2 /4 ,
Γ ( t + 1 ) ≥ ∫ 1 2 x t e − x d x ≥ ∫ 1 2 e − 2 d x = e − 2 . \Gamma(t+1)\ge\int_1^2x^t e^{-x}\,\mathrm{d}x\ge\int_1^2e^{-2}\,\mathrm{d}x=e^{-2}. Γ ( t + 1 ) ≥ ∫ 1 2 x t e − x d x ≥ ∫ 1 2 e − 2 d x = e − 2 .
For 0 < r ≤ 1 0<r\le1 0 < r ≤ 1 and k ≤ 2 m k\le2m k ≤ 2 m ,
r 1 − ν ( k r / 2 ) ν = ( k / 2 ) ν r ≤ m ν ≤ m m , r^{1-\nu}(kr/2)^\nu=(k/2)^\nu r\le m^\nu\le m^m, r 1 − ν ( k r /2 ) ν = ( k /2 ) ν r ≤ m ν ≤ m m ,
r 1 − ν ( k r / 2 ) ν − 1 = ( k / 2 ) ν − 1 ≤ m ν − 1 ≤ m m . r^{1-\nu}(kr/2)^{\nu-1}=(k/2)^{\nu-1}\le m^{\nu-1}\le m^m. r 1 − ν ( k r /2 ) ν − 1 = ( k /2 ) ν − 1 ≤ m ν − 1 ≤ m m .
The resulting bounds, with continuous values at r = 0 r=0 r = 0 , are
max { ∣ r 1 − ν J ν ( k r ) ∣ , ∣ r 1 − ν J ν − 1 ( k r ) ∣ } ≤ m m e 2 e m 2 = e m log m + m 2 + 2 ≤ e 2 m 2 + 2 ≤ e 3 m 2 . \begin{aligned}
\max\{|r^{1-\nu}J_\nu(kr)|,|r^{1-\nu}J_{\nu-1}(kr)|\}\le m^m e^{2}e^{m^2} \\
&=e^{m\log m+m^2+2}\le e^{2m^2+2}\le e^{3m^2}.
\end{aligned} max { ∣ r 1 − ν J ν ( k r ) ∣ , ∣ r 1 − ν J ν − 1 ( k r ) ∣ } ≤ m m e 2 e m 2 = e m l o g m + m 2 + 2 ≤ e 2 m 2 + 2 ≤ e 3 m 2 .
On the inner ball,
w ′ = − C k r 1 − ν J ν ( k r ) , L d w = − C k 2 r 1 − ν J ν − 1 ( k r ) , w'=-Ckr^{1-\nu}J_\nu(kr),\qquad L_d w=-Ck^2r^{1-\nu}J_{\nu-1}(kr), w ′ = − C k r 1 − ν J ν ( k r ) , L d w = − C k 2 r 1 − ν J ν − 1 ( k r ) ,
max { k , k 2 } = k 2 ≤ 4 m 2 ≤ e m 2 , \max\{k,k^2\}=k^2\le4m^2\le e^{m^2}, max { k , k 2 } = k 2 ≤ 4 m 2 ≤ e m 2 ,
and hence
max { ∣ w ′ ∣ , ∣ L d w ∣ } ≤ e 4 m 2 h e m 2 e 3 m 2 = e 8 m 2 h . \max\{|w'|,|L_d w|\}\le\frac{e^{4m^2}}{h}e^{m^2}e^{3m^2}=\frac{e^{8m^2}}{h}. max { ∣ w ′ ∣ , ∣ L d w ∣ } ≤ h e 4 m 2 e m 2 e 3 m 2 = h e 8 m 2 .
On the outer layer, w = h f w=hf w = h f , r ≥ a > 1 / 2 r\ge a>1/2 r ≥ a > 1/2 , and (3.16) give
∣ w ′ ∣ = ∣ f ′ ∣ ≤ 12 , |w'|=|f'|\le12, ∣ w ′ ∣ = ∣ f ′ ∣ ≤ 12 ,
∣ L d w ∣ = ∣ h − 1 f ′ ′ + ( d − 1 ) r − 1 f ′ ∣ ≤ 12 h + 24 d , |L_d w|=|h^{-1}f''+(d-1)r^{-1}f'|\le\frac{12}{h}+24d, ∣ L d w ∣ = ∣ h − 1 f ′′ + ( d − 1 ) r − 1 f ′ ∣ ≤ h 12 + 24 d ,
max { ∣ w ′ ∣ , ∣ L d w ∣ } ≤ 12 + 24 d h ≤ m 3 h ≤ e 3 m h ≤ e 8 m 2 h . \max\{|w'|,|L_d w|\}\le\frac{12+24d}{h}\le\frac{m^3}{h}\le\frac{e^{3m}}{h}\le\frac{e^{8m^2}}{h}. max { ∣ w ′ ∣ , ∣ L d w ∣ } ≤ h 12 + 24 d ≤ h m 3 ≤ h e 3 m ≤ h e 8 m 2 .
Estimate the energy used to normalize the mode
On the short piece of the outer layer adjacent to a a a , we obtain a lower bound for the normalized energy. For 0 ≤ x ≤ 1 / 8 0\le x\le1/8 0 ≤ x ≤ 1/8 ,
f 0 ′ ′ ( x ) = − 4 + 6 x ≤ − 13 4 < 0 , f_0''(x)=-4+6x\le-\frac{13}{4}<0, f 0 ′′ ( x ) = − 4 + 6 x ≤ − 4 13 < 0 ,
f 0 ′ ( x ) = 1 − 4 x + 3 x 2 ≥ f 0 ′ ( 1 / 8 ) = 35 64 > 1 2 , f_0'(x)=1-4x+3x^2\ge f_0'(1/8)=\frac{35}{64}>\frac{1}{2}, f 0 ′ ( x ) = 1 − 4 x + 3 x 2 ≥ f 0 ′ ( 1/8 ) = 64 35 > 2 1 ,
∥ f − f 0 ∥ C 3 ≤ 96 ( d + 2 ) h ≤ 96 m h ≤ 96 ⋅ 2 − 32 m − 2 < 1 4 , \|f-f_0\|_{C^3}\le96(d+2)h\le96mh\le96\cdot2^{-32}m^{-2}<\frac{1}{4}, ∥ f − f 0 ∥ C 3 ≤ 96 ( d + 2 ) h ≤ 96 mh ≤ 96 ⋅ 2 − 32 m − 2 < 4 1 ,
f ′ ( x ) ≥ f 0 ′ ( x ) − ∥ f − f 0 ∥ C 3 > 1 2 − 1 4 = 1 4 , f'(x)\ge f_0'(x)-\|f-f_0\|_{C^3}>\frac{1}{2}-\frac{1}{4}=\frac{1}{4}, f ′ ( x ) ≥ f 0 ′ ( x ) − ∥ f − f 0 ∥ C 3 > 2 1 − 4 1 = 4 1 ,
a d − 1 = ( 1 − h ) d − 1 ≥ 1 − ( d − 1 ) h ≥ 1 − 2 − 32 m − 2 > 1 2 . a^{d-1}=(1-h)^{d-1}\ge1-(d-1)h\ge1-2^{-32}m^{-2}>\frac{1}{2}. a d − 1 = ( 1 − h ) d − 1 ≥ 1 − ( d − 1 ) h ≥ 1 − 2 − 32 m − 2 > 2 1 .
Hence
E ( w ) = ∣ S d − 1 ∣ ∫ 0 1 r d − 1 ∣ w ′ ( r ) ∣ 2 d r ≥ ∣ S d − 1 ∣ ∫ a a + h / 8 r d − 1 ∣ f ′ ( r − a h ) ∣ 2 d r ≥ ∣ S d − 1 ∣ h 8 1 2 1 16 = ∣ S d − 1 ∣ h 256 . \begin{aligned}
\mathcal{E}(w)=|S^{d-1}|\int_0^1 r^{d-1}|w'(r)|^2\,dr \\
&\ge|S^{d-1}|\int_a^{a+h/8}r^{d-1}\left|f'\left(\frac{r-a}{h}\right)\right|^2\,dr \\
&\ge|S^{d-1}|\frac{h}{8}\frac{1}{2}\frac{1}{16}=\frac{|S^{d-1}|h}{256}.
\end{aligned} E ( w ) = ∣ S d − 1 ∣ ∫ 0 1 r d − 1 ∣ w ′ ( r ) ∣ 2 d r ≥ ∣ S d − 1 ∣ ∫ a a + h /8 r d − 1 f ′ ( h r − a ) 2 d r ≥ ∣ S d − 1 ∣ 8 h 2 1 16 1 = 256 ∣ S d − 1 ∣ h .
For the second bound, using the Gamma estimate from Step 2, one has
Γ ( d / 2 ) = Γ ( d / 2 + 1 ) d / 2 ≤ e − m 2 , ∣ S d − 1 ∣ = 2 π d / 2 Γ ( d / 2 ) ≥ 2 e − m 2 . \Gamma(d/2)=\frac{\Gamma(d/2+1)}{d/2}\le e^{-m^2},\qquad|S^{d-1}|=\frac{2\pi^{d/2}}{\Gamma(d/2)}\ge2e^{-m^2}. Γ ( d /2 ) = d /2 Γ ( d /2 + 1 ) ≤ e − m 2 , ∣ S d − 1 ∣ = Γ ( d /2 ) 2 π d /2 ≥ 2 e − m 2 .
Thus, since m ≥ 6 m\ge6 m ≥ 6 ,
∣ S d − 1 ∣ 256 ≥ e − m 2 128 ≥ e − 2 m 2 , E ( w ) ≥ h e − 2 m 2 . \frac{|S^{d-1}|}{256}\ge\frac{e^{-m^2}}{128}\ge e^{-2m^2},\qquad\mathcal{E}(w)\ge he^{-2m^2}. 256 ∣ S d − 1 ∣ ≥ 128 e − m 2 ≥ e − 2 m 2 , E ( w ) ≥ h e − 2 m 2 .
The energy is taken on the physical d d d -dimensional ball, so the spherical area factor must be retained in this normalization. The final dyadic bound uses
log 2 = 2 ∫ 0 1 / 3 d t 1 − t 2 > 2 ∫ 0 1 / 3 ( 1 + t 2 + t 4 ) d t = 842 1215 > 9 13 . \log2 = 2\int_{0}^{1/3}\frac{\mathrm{d}t}{1-t^2} > 2\int_{0}^{1/3}(1+t^2+t^4)\,\mathrm{d}t = \frac{842}{1215} > \frac{9}{13}. log 2 = 2 ∫ 0 1/3 1 − t 2 d t > 2 ∫ 0 1/3 ( 1 + t 2 + t 4 ) d t = 1215 842 > 13 9 .
For u ∗ = w / E ( w ) u_* = w/\sqrt{\mathcal{E}(w)} u ∗ = w / E ( w ) and 0 < h < 1 0<h<1 0 < h < 1 , it follows that
max { ∥ D u ∗ ∥ ∞ , ∥ Δ u ∗ ∥ ∞ } ≤ e 8 m 2 / h h 1 / 2 e − m 2 = e 9 m 2 h − 3 / 2 ≤ 2 13 m 2 h − 2 = 2 13 m 2 + 2 H = M . \max\{\lVert Du_*\rVert_\infty,\lVert\Delta u_*\rVert_\infty\}\leq\frac{e^{8m^2}/h}{h^{1/2}e^{-m^2}}=e^{9m^2}h^{-3/2}
\leq2^{13m^2}h^{-2}=2^{13m^2+2H}=M. max {∥ D u ∗ ∥ ∞ , ∥ Δ u ∗ ∥ ∞ } ≤ h 1/2 e − m 2 e 8 m 2 / h = e 9 m 2 h − 3/2 ≤ 2 13 m 2 h − 2 = 2 13 m 2 + 2 H = M .
The role of the exponent 13 m 2 13m^2 13 m 2 is to convert the preceding exponential estimates into an integer power of two. The earlier Bessel bounds and spherical area bounds are only sufficient, not optimal. The derivatives are understood in the weak sense, with bounded piecewise representative. For this radial mode, continuity of the radial derivative and of L d u ∗ L_d u_* L d u ∗ ensures continuity of the second radial derivative, so u ∗ u_* u ∗ is C 2 C^2 C 2 across S a S_a S a .
Proof of the main theorem
It remains to turn the core constraints into actual smooth holes while preserving the gap (30) . Proposition 4.2 already gives the lower bound for Λ 2 \Lambda_2 Λ 2 , while Lemma 5.1 controls a trial function used to estimate Λ 1 \Lambda_1 Λ 1 . We first fix the admissible error and the cutoff scales, then estimate the Rayleigh quotient of the trial function, and finally verify the geometric properties of the domain.
Proof of Theorem 1.1. Fix an integer d ≥ 2 d\geq2 d ≥ 2 and let Ω d \Omega_d Ω d be the domain in (9) , with the parameters chosen as in § 2 . The facts we shall use are
E ( u ∗ ) = 1 , Q ( u ∗ ) ≤ U , λ 2 , N > T = C d U + δ / 4. \mathcal{E}(u_*)=1,\qquad Q(u_*)\leq U,\qquad\lambda_{2,N}>T=C_dU+\delta/4. E ( u ∗ ) = 1 , Q ( u ∗ ) ≤ U , λ 2 , N > T = C d U + δ /4.
Step 1. Choose the error allowance from the spectral gap.
The lower bound for Λ 2 \Lambda_2 Λ 2 is already fixed. To preserve it, the upper bound for C d Λ 1 C_d\Lambda_1 C d Λ 1 may be raised by at most δ / 16 \delta/16 δ /16 . The Rayleigh quotient estimate in Step 5 will show that an error of size η \eta η in the bending norm costs at most 8 C d ( U + 1 ) η 8C_d(U+1)\eta 8 C d ( U + 1 ) η . We therefore require
η ≤ δ 128 C d ( U + 1 ) . (36) \eta\leq\frac{\delta}{128C_d(U+1)}. \tag*{(36)} η ≤ 128 C d ( U + 1 ) δ . (36)
The explicit choice below is h 2 h^2 h 2 divided by a generous dimensional factor. Powers of two are used so that this choice and the radii can be specified by integers. The aim is to leave ample margin.
Step 2. Choose the cutoff radius and logarithmic width.
Set
η = 2 − 16 − 2 s h 2 , R = 2 − K , L = 2 Z . \eta= 2^{-16-2s}h^2,\qquad R = 2^{-K},\qquad L = 2^Z. η = 2 − 16 − 2 s h 2 , R = 2 − K , L = 2 Z .
Here R R R is the outer cutoff radius and L L L is its logarithmic width. Let A N A_N A N be the total ( d − 2 ) (d-2) ( d − 2 ) -dimensional measure of the cores, with the counting measure when d = 2 d=2 d = 2 . The surface area of the unit ( d − 1 ) (d-1) ( d − 1 ) -sphere equals d − 1 d-1 d − 1 times the volume of the unit ( d − 1 ) (d-1) ( d − 1 ) -ball, and this ball lies in [ − 1 , 1 ] d − 1 [-1,1]^{d-1} [ − 1 , 1 ] d − 1 . Hence
∣ S d − 2 ∣ ≤ ( d − 1 ) 2 d − 1 ≤ 2 2 d , 0 ≤ a d − 2 sin d − 2 ( j Δ θ ) ≤ 1. |S^{d-2}| \le(d-1)2^{d-1} \le2^{2d},\qquad0\le a^{d-2}\sin^{d-2}(j\Delta_\theta)\le1. ∣ S d − 2 ∣ ≤ ( d − 1 ) 2 d − 1 ≤ 2 2 d , 0 ≤ a d − 2 sin d − 2 ( j Δ θ ) ≤ 1.
and
A N = ∣ S d − 2 ∣ a d − 2 ∑ j = 1 2 N − 1 sin d − 2 ( j Δ θ ) ≤ ∣ S d − 2 ∣ ( 2 N − 1 ) ≤ 2 2 d ( 2 N − 1 ) < 2 N + 2 d . \begin{aligned}
A_N=|S^{d-2}|a^{d-2}\sum_{j=1}^{2^N-1}\sin^{d-2}(j\Delta_\theta) \\
&\le|S^{d-2}|(2^N-1)\le2^{2d}(2^N-1)<2^{N+2d}.
\end{aligned} A N = ∣ S d − 2 ∣ a d − 2 j = 1 ∑ 2 N − 1 sin d − 2 ( j Δ θ ) ≤ ∣ S d − 2 ∣ ( 2 N − 1 ) ≤ 2 2 d ( 2 N − 1 ) < 2 N + 2 d .
The case d = 2 d=2 d = 2 is also included. Each latitude contributes two points, and A N = 2 ( 2 N − 1 ) A_N=2(2^N-1) A N = 2 ( 2 N − 1 ) . Since a ≥ 1 / 2 a\ge1/2 a ≥ 1/2 and 0 < Δ θ = π 2 − N ≤ π / 2 0<\Delta_\theta=\pi2^{-N}\le\pi/2 0 < Δ θ = π 2 − N ≤ π /2 , the minimal core radius satisfies
r min = min 1 ≤ j < 2 N a sin ( j Δ θ ) = a sin Δ θ ≥ a 2 Δ θ π ≥ 1 2 2 1 − N = 2 − N . \begin{aligned}
r_{\min}=\min_{1\le j<2^N}a\sin(j\Delta_\theta)=a\sin\Delta_\theta\\
&\ge a\frac{2\Delta_\theta}{\pi}\ge\frac{1}{2}2^{1-N}=2^{-N}.
\end{aligned} r m i n = 1 ≤ j < 2 N min a sin ( j Δ θ ) = a sin Δ θ ≥ a π 2 Δ θ ≥ 2 1 2 1 − N = 2 − N .
The first two bounds below ensure that the cutoff tubes lie inside the ball, are pairwise disjoint, and are located in the effective normal coordinates. The third bound makes the contribution of their support to the bending error less than η / 4 \eta/4 η /4 . The condition on L L L makes the derivative contribution less than η / 4 \eta/4 η /4 . Thus 16 = 4 ⋅ 4 16=4\cdot4 16 = 4 ⋅ 4 and 25600 = ( 4 ⋅ 40 ) 2 25600=(4\cdot40)^2 25600 = ( 4 ⋅ 40 ) 2 correspond exactly to the two terms in (39) . By the definition of the integers,
R ≤ h / 4 , R ≤ r min / ( 4 d ) , R ≤ η 16 ( M + 1 ) A N + 1 . (37) R\le h/4,\qquad R\le r_{\min}/(4d),\qquad R\le\frac{\eta}{16(M+1)\sqrt{A_N+1}}. \tag*{(37)} R ≤ h /4 , R ≤ r m i n / ( 4 d ) , R ≤ 16 ( M + 1 ) A N + 1 η . (37)
L > 1 + 25600 ( M + 1 ) 2 ( A N + 1 ) η 2 . (38) L>1+\frac{25600(M+1)^2(A_N+1)}{\eta^2}. \tag*{(38)} L > 1 + η 2 25600 ( M + 1 ) 2 ( A N + 1 ) . (38)
To verify the geometric bounds, use the choice of K K K in § 2 and N ≤ ( d + 24 + 7 H ) / 2 N\le(d+24+7H)/2 N ≤ ( d + 24 + 7 H ) /2 :
K − ( H + 2 ) = 13 m 2 + ⌈ N / 2 ⌉ + d + 3 H + 2 s + 21 > 0 , K-(H+2)=13m^2+\lceil N/2\rceil+d+3H+2s+21>0, K − ( H + 2 ) = 13 m 2 + ⌈ N /2 ⌉ + d + 3 H + 2 s + 21 > 0 ,
K − ( N + s + 2 ) = 13 m 2 − ⌊ N / 2 ⌋ + d + 4 H + s + 21 K-(N+s+2)=13m^2-\lfloor N/2\rfloor+d+4H+s+21 K − ( N + s + 2 ) = 13 m 2 − ⌊ N /2 ⌋ + d + 4 H + s + 21
≥ 13 m 2 + 3 d + 9 H + 60 4 + s > 0. \ge13m^2+\frac{3d+9H+60}{4}+s>0. ≥ 13 m 2 + 4 3 d + 9 H + 60 + s > 0.
Since d ≤ 2 s d\le2^s d ≤ 2 s and r min ≥ 2 − N r_{\min}\ge2^{-N} r m i n ≥ 2 − N ,
R = 2 − K ≤ 2 − H − 2 = h / 4 , R ≤ 2 − N − s − 2 ≤ r min 4 d . R=2^{-K}\le2^{-H-2}=h/4,\qquad R\le2^{-N-s-2}\le\frac{r_{\min}}{4d}. R = 2 − K ≤ 2 − H − 2 = h /4 , R ≤ 2 − N − s − 2 ≤ 4 d r m i n .
For the error bound on R R R , use M + 1 < 2 M M+1<2M M + 1 < 2 M and A N + 1 < 2 N + 2 d + 1 A_N+1<2^{N+2d+1} A N + 1 < 2 N + 2 d + 1 .
16 ( M + 1 ) A N + 1 < 32 2 M 2 N / 2 + d < 64 M 2 N / 2 + d . 16(M+1)\sqrt{A_N+1}<32\sqrt{2}\,M\,2^{N/2+d}<64M\,2^{N/2+d}. 16 ( M + 1 ) A N + 1 < 32 2 M 2 N /2 + d < 64 M 2 N /2 + d .
Substituting M = 2 13 m 2 + 2 H M=2^{13m^2+2H} M = 2 13 m 2 + 2 H , h = 2 − H h=2^{-H} h = 2 − H , and the definitions of R R R and η \eta η gives an exact cancellation,
64 M 2 N / 2 + d R η = 2 13 m 2 + N / 2 + d + 4 H + 2 s + 22 − K = 2 N / 2 − ⌈ N / 2 ⌉ − 1 ≤ 1 2 < 1. \begin{aligned}
\frac{64M2^{N/2+d}R}{\eta}=2^{13m^2+N/2+d+4H+2s+22-K} \\
&=2^{N/2-\lceil N/2\rceil-1}\le\frac{1}{2}<1.
\end{aligned} η 64 M 2 N /2 + d R = 2 13 m 2 + N /2 + d + 4 H + 2 s + 22 − K = 2 N /2 − ⌈ N /2 ⌉ − 1 ≤ 2 1 < 1.
Therefore
R < η 64 M 2 N / 2 + d < η 16 ( M + 1 ) A N + 1 . R<\frac{\eta}{64M2^{N/2+d}}<\frac{\eta}{16(M+1)\sqrt{A_N+1}}. R < 64 M 2 N /2 + d η < 16 ( M + 1 ) A N + 1 η .
This proves (6.2). The chosen integer K K K is one larger than the ceiling in the first line of the sufficient exponents. This extra binary step preserves the factor 1 / 2 1/2 1/2 above and also satisfies the two geometric bounds. On the logarithmic width side, retaining the exact coefficient 25600 ⋅ 4 ⋅ 2 = 204800 = ( 25 / 32 ) 2 18 25600\cdot4\cdot2=204800=(25/32)2^{18} 25600 ⋅ 4 ⋅ 2 = 204800 = ( 25/32 ) 2 18 gives
25600 ( M + 1 ) 2 ( A N + 1 ) η 2 < 25 32 2 18 M 2 2 N + 2 d η − 2 = 25 32 2 18 + ( 26 m 2 + 4 H ) + ( N + 2 d ) + ( 32 + 4 s + 4 H ) = 25 32 2 Z . \begin{aligned}
\frac{25600(M+1)^2(A_N+1)}{\eta^2}<\frac{25}{32}2^{18}M^2 2^{N+2d}\eta^{-2} \\
&=\frac{25}{32}2^{18+(26m^2+4H)+(N+2d)+(32+4s+4H)} \\
&=\frac{25}{32}2^Z.
\end{aligned} η 2 25600 ( M + 1 ) 2 ( A N + 1 ) < 32 25 2 18 M 2 2 N + 2 d η − 2 = 32 25 2 18 + ( 26 m 2 + 4 H ) + ( N + 2 d ) + ( 32 + 4 s + 4 H ) = 32 25 2 Z .
Since Z ≥ 3 Z\ge3 Z ≥ 3 ,
1 + 25600 ( M + 1 ) 2 ( A N + 1 ) η 2 < 1 + 25 32 L ≤ 29 32 L < L , 1+\frac{25600(M+1)^2(A_N+1)}{\eta^2}<1+\frac{25}{32}L\le\frac{29}{32}L<L, 1 + η 2 25600 ( M + 1 ) 2 ( A N + 1 ) < 1 + 32 25 L ≤ 32 29 L < L ,
which proves (6.3). The integer exponent Z Z Z leaves a margin of at least 3 L / 32 3L/32 3 L /32 in this sufficient condition. In addition,
L > 1 , 2 ε R = 2 e − L < 2 e < 1 , L>1,\qquad\frac{2\varepsilon}{R}=2e^{-L}<\frac{2}{e}<1, L > 1 , R 2 ε = 2 e − L < e 2 < 1 ,
so the normal push used in the connectivity argument always stays inside the cutoff tubes. Finally, from (30) and (32) ,
δ 128 C d ( U + 1 ) > h 2 55296 m 2 > h 2 65536 m 2 ≥ 2 − 16 − 2 s h 2 = η . \frac{\delta}{128C_d(U+1)}>\frac{h^2}{55296m^2}>\frac{h^2}{65536m^2}\ge2^{-16-2s}h^2=\eta. 128 C d ( U + 1 ) δ > 55296 m 2 h 2 > 65536 m 2 h 2 ≥ 2 − 16 − 2 s h 2 = η .
The denominator 55296 55296 55296 is 4 ⋅ 128 ⋅ 36 ⋅ 3 4\cdot128\cdot36\cdot3 4 ⋅ 128 ⋅ 36 ⋅ 3 , where the factors come respectively from δ > h 2 / 4 \delta>h^2/4 δ > h 2 /4 , the 128 128 128 in the requirement, C d < 36 C_d<36 C d < 36 , and U + 1 < 3 m 2 U+1<3m^2 U + 1 < 3 m 2 . The last inequality is m − 2 ≥ 2 − 2 s m^{-2}\ge2^{-2s} m − 2 ≥ 2 − 2 s , which follows from m ≤ 2 s m\le2^s m ≤ 2 s . Hence (6.1) holds and η < 1 / 4 \eta<1/4 η < 1/4 .
Step 3. Construct the clamped cutoff.
For the distance to the j j j -th core, let ϱ = dist ( x , M j ) \varrho=\operatorname{dist}(x,M_j) ϱ = dist ( x , M j ) and set
η j ( x ) = χ ( log ( ϱ / ε ) L ) , χ ( t ) = { 0 , t ≤ 0 , 3 t 2 − 2 t 3 , 0 < t < 1 , 1 , t ≥ 1. \eta_j(x)=\chi\left(\frac{\log(\varrho/\varepsilon)}{L}\right),\qquad
\chi(t)=
\begin{cases}
0, & t\le0,\\
3t^2-2t^3, & 0<t<1,\\
1, & t\ge1.
\end{cases} η j ( x ) = χ ( L log ( ϱ / ε ) ) , χ ( t ) = ⎩ ⎨ ⎧ 0 , 3 t 2 − 2 t 3 , 1 , t ≤ 0 , 0 < t < 1 , t ≥ 1.
We also set η j = 0 \eta_j=0 η j = 0 when ϱ = 0 \varrho=0 ϱ = 0 . Since ε = R e − L \varepsilon=Re^{-L} ε = R e − L , this cutoff equals one at distance R R R . Figure 10 shows the normal geometry and the transition as a function of the logarithmic distance. It is C 1 C^1 C 1 with bounded weak second derivatives, which suffices for H 2 H^2 H 2 . The endpoint data are
Figure 10. Normal cross-section and logarithmic cutoff profile. The zero endpoint slopes give the clamped condition at ε \varepsilon ε and an H 2 H^2 H 2 matching at R R R . A larger L L L reduces the derivative cost.
χ ( 0 ) = χ ′ ( 0 ) = 0 , χ ( 1 ) = 1 , χ ′ ( 1 ) = 0. \chi(0)=\chi'(0)=0,\qquad\chi(1)=1,\qquad\chi'(1)=0. χ ( 0 ) = χ ′ ( 0 ) = 0 , χ ( 1 ) = 1 , χ ′ ( 1 ) = 0.
The first pair of conditions gives the two clamped traces at the hole boundary. The second pair allows the function to be glued back to the unmodified function at distance R R R without a jump in the first derivative.
All R R R -tubes are pairwise disjoint. The minimal distance between cores is 2 a sin ( Δ θ / 2 ) ≥ r min 2a\sin(\Delta_\theta/2)\ge r_{\min} 2 a sin ( Δ θ /2 ) ≥ r m i n , and all tubes lie in B B B . For distinct latitudes j , k j,k j , k , minimizing in the angular direction gives the distance
( a sin θ j − a sin θ k ) 2 + ( a cos θ j − a cos θ k ) 2 = 2 a sin ∣ θ j − θ k ∣ 2 . \sqrt{(a\sin\theta_j-a\sin\theta_k)^2+(a\cos\theta_j-a\cos\theta_k)^2}=2a\sin\frac{\lvert\theta_j-\theta_k\rvert}{2}. ( a sin θ j − a sin θ k ) 2 + ( a cos θ j − a cos θ k ) 2 = 2 a sin 2 ∣ θ j − θ k ∣ .
The minimum is attained for adjacent latitudes and is at least a sin Δ θ = r min a\sin\Delta_\theta=r_{\min} a sin Δ θ = r m i n , since sin Δ θ = 2 sin ( Δ θ / 2 ) cos ( Δ θ / 2 ) \sin\Delta_\theta=2\sin(\Delta_\theta/2)\cos(\Delta_\theta/2) sin Δ θ = 2 sin ( Δ θ /2 ) cos ( Δ θ /2 ) . Also 2 R ≤ r min / ( 2 d ) < r min 2R\le r_{\min}/(2d)<r_{\min} 2 R ≤ r m i n / ( 2 d ) < r m i n . In two dimensions, the two points on the same latitude are separated by 2 r j ≥ 2 r min 2r_j\ge2r_{\min} 2 r j ≥ 2 r m i n , so their disks are also pairwise disjoint. Every point in an R R R -tube has modulus at most a + R ≤ 1 − 3 h / 4 < 1 a+R\le1-3h/4<1 a + R ≤ 1 − 3 h /4 < 1 , which shows that the tube is contained in B B B .
By Lemma 5.1 and the zero value on each core, ∣ u ∗ ∣ ≤ M ϱ \lvert u_*\rvert\le M\varrho ∣ u ∗ ∣ ≤ M ϱ , and nearby we also have ∣ D u ∗ ∣ ≤ M \lvert Du_*\rvert\le M ∣ D u ∗ ∣ ≤ M . In the R R R -tube around M j M_j M j we set v = η j u ∗ v=\eta_j u_* v = η j u ∗ , and outside all these tubes we set v = u ∗ v=u_* v = u ∗ . The disjointness proved above ensures that this definition is unambiguous. The value estimate follows by joining a point to its nearest core point along a normal segment and then integrating D u ∗ Du_* D u ∗ along this segment. The segment lies in B B B , and u ∗ u_* u ∗ vanishes at the core endpoint.
For a latitude of radius r j = a sin θ j r_j=a\sin\theta_j r j = a sin θ j , the tube coordinates are
( ( r j + z 1 ) ω , a cos θ j + z 2 ) , \bigl((r_j+z_1)\omega,a\cos\theta_j+z_2\bigr), ( ( r j + z 1 ) ω , a cos θ j + z 2 ) ,
where ω ∈ S d − 2 \omega\in S^{d-2} ω ∈ S d − 2 , z = ( z 1 , z 2 ) ∈ R 2 z=(z_1,z_2)\in\mathbb{R}^2 z = ( z 1 , z 2 ) ∈ R 2 , and ϱ = ∣ z ∣ \varrho=|z| ϱ = ∣ z ∣ . For ∣ z ∣ ≤ R |z|\le R ∣ z ∣ ≤ R , the volume factor relative to the measure on the core is ( 1 + z 1 / r j ) d − 2 < 2 (1+z_1/r_j)^{d-2}<2 ( 1 + z 1 / r j ) d − 2 < 2 , by (6.2). Indeed, since r j ≥ r min r_j\ge r_{\min} r j ≥ r m i n and R / r j ≤ 1 / ( 4 d ) R/r_j\le1/(4d) R / r j ≤ 1/ ( 4 d ) ,
( 1 + z 1 / r j ) d − 2 ≤ ( 1 + R / r j ) d − 2 ≤ exp ( ( d − 2 ) R / r j ) ≤ e 1 / 4 < 2. (1+z_1/r_j)^{d-2}\le(1+R/r_j)^{d-2}\le\exp((d-2)R/r_j)\le e^{1/4}<2. ( 1 + z 1 / r j ) d − 2 ≤ ( 1 + R / r j ) d − 2 ≤ exp (( d − 2 ) R / r j ) ≤ e 1/4 < 2.
This factor is positive because r j − R > 0 r_j-R>0 r j − R > 0 . For d = 2 d=2 d = 2 , it equals exactly one on each coordinate chart centered at a point. The Laplacian of the cutoff function is the two-dimensional normal Laplacian Δ z = ∂ z 1 2 + ∂ z 2 2 \Delta_z=\partial_{z_1}^2+\partial_{z_2}^2 Δ z = ∂ z 1 2 + ∂ z 2 2 plus ( d − 2 ) ( r j + z 1 ) − 1 ∂ z 1 η j (d-2)(r_j+z_1)^{-1}\partial_{z_1}\eta_j ( d − 2 ) ( r j + z 1 ) − 1 ∂ z 1 η j .
Step 4. Fit the bending error inside the allowance.
The cutoff in Step 3 already imposes the boundary conditions. We now use the radius and logarithmic width from Step 2 to control its cost, so that the Rayleigh principle can be applied in Step 5.
On 0 < t < 1 0<t<1 0 < t < 1 , χ ′ ( t ) = 6 t ( 1 − t ) \chi'(t)=6t(1-t) χ ′ ( t ) = 6 t ( 1 − t ) and χ ′ ′ ( t ) = 6 − 12 t \chi''(t)=6-12t χ ′′ ( t ) = 6 − 12 t , so ∣ χ ′ ∣ ≤ 3 / 2 |\chi'|\le3/2 ∣ χ ′ ∣ ≤ 3/2 and ∣ χ ′ ′ ∣ ≤ 6 |\chi''|\le6 ∣ χ ′′ ∣ ≤ 6 . We now prove
∥ Δ ( v − u ∗ ) ∥ 2 ≤ 4 M A N R + 40 M A N / L < η / 2. (39) \|\Delta(v-u_*)\|_2\le4M\sqrt{A_N}R+40M\sqrt{A_N}/L<\eta/2. \tag*{(39)} ∥Δ ( v − u ∗ ) ∥ 2 ≤ 4 M A N R + 40 M A N / L < η /2. (39)
To verify this estimate, write e = v − u ∗ e=v-u_* e = v − u ∗ on the whole ball, with v = 0 v=0 v = 0 inside the removed tubes. By the product rule, there is a support term ( η j − 1 ) Δ u ∗ (\eta_j-1)\Delta u_* ( η j − 1 ) Δ u ∗ and a derivative term 2 D η j ⋅ D u ∗ + u ∗ Δ η j 2D\eta_j\cdot Du_*+u_*\Delta\eta_j 2 D η j ⋅ D u ∗ + u ∗ Δ η j . The tubes are pairwise disjoint, so the squared norm is the sum of the contributions from the individual tubes. In each step the volume factor is bounded by 2, integration over the core gives A N A_N A N , and polar coordinates with area element 2 π ϱ d ϱ 2\pi\varrho\,d\varrho 2 π ϱ d ϱ are used in the normal plane. The norms below refer to the piecewise terms defined on the union of all tubes and set equal to zero elsewhere.
First, ∣ η j − 1 ∣ ≤ 1 |\eta_j-1|\le1 ∣ η j − 1∣ ≤ 1 and ∣ Δ u ∗ ∣ ≤ M |\Delta u_*|\le M ∣Δ u ∗ ∣ ≤ M . The support consists of both the interior holes and the transition annuli, so we use the full normal disk of radius R R R :
∥ ( η j − 1 ) Δ u ∗ ∥ 2 2 ≤ 2 A N M 2 ∫ 0 R 2 π ϱ d ϱ = 2 π A N M 2 R 2 . \|(\eta_j-1)\Delta u_*\|_2^2\le2A_NM^2\int_0^R2\pi\varrho\,d\varrho=2\pi A_NM^2R^2. ∥ ( η j − 1 ) Δ u ∗ ∥ 2 2 ≤ 2 A N M 2 ∫ 0 R 2 π ϱ d ϱ = 2 π A N M 2 R 2 .
Its norm is at most 2 π M A N R \sqrt{2\pi}M\sqrt{A_N}R 2 π M A N R , and we may also use the more convenient and more generous bound 2 π M A N R 2\sqrt{\pi}M\sqrt{A_N}R 2 π M A N R .
Second, in the transition region set t = log ( ϱ / ε ) / L t=\log(\varrho/\varepsilon)/L t = log ( ϱ / ε ) / L . By the chain rule,
∂ η j ∂ ϱ = χ ′ ( t ) L ϱ , ∣ D η j ∣ ≤ 3 2 L ϱ . \frac{\partial\eta_j}{\partial\varrho}=\frac{\chi'(t)}{L\varrho},\qquad|D\eta_j|\le\frac{3}{2L\varrho}. ∂ ϱ ∂ η j = L ϱ χ ′ ( t ) , ∣ D η j ∣ ≤ 2 L ϱ 3 .
Thus the cross term satisfies
∥ 2 D η j ⋅ D u ∗ ∥ 2 2 ≤ 2 A N ∫ ε R 9 M 2 L 2 ϱ 2 2 π ϱ d ϱ = 36 π A N M 2 L 2 log ( R / ε ) = 36 π A N M 2 L . \begin{aligned}
\|2D\eta_j\cdot Du_*\|_2^2\le2A_N\int_\varepsilon^R\frac{9M^2}{L^2\varrho^2}2\pi\varrho\,d\varrho\\
&=\frac{36\pi A_NM^2}{L^2}\log(R/\varepsilon)=\frac{36\pi A_NM^2}{L}.
\end{aligned} ∥2 D η j ⋅ D u ∗ ∥ 2 2 ≤ 2 A N ∫ ε R L 2 ϱ 2 9 M 2 2 π ϱ d ϱ = L 2 36 π A N M 2 log ( R / ε ) = L 36 π A N M 2 .
Its norm is at most 6 π M A N / L 6\sqrt{\pi}M\sqrt{A_N/L} 6 π M A N / L .
Third, the two-dimensional radial Laplacian has a useful cancellation:
∂ ϱ 2 η j = χ ′ ′ ( t ) L 2 ϱ 2 − χ ′ ( t ) L ϱ 2 , \partial_{\varrho}^{2}\eta_j=\frac{\chi^{\prime\prime}(t)}{L^{2}\varrho^{2}}-\frac{\chi^{\prime}(t)}{L\varrho^{2}}, ∂ ϱ 2 η j = L 2 ϱ 2 χ ′′ ( t ) − L ϱ 2 χ ′ ( t ) ,
Δ z η j = ∂ ϱ 2 η j + ϱ − 1 ∂ ϱ η j = χ ′ ′ ( t ) L 2 ϱ 2 . \Delta_z\eta_j=\partial_{\varrho}^{2}\eta_j+\varrho^{-1}\partial_{\varrho}\eta_j=\frac{\chi^{\prime\prime}(t)}{L^{2}\varrho^{2}}. Δ z η j = ∂ ϱ 2 η j + ϱ − 1 ∂ ϱ η j = L 2 ϱ 2 χ ′′ ( t ) .
Using ∣ u ∗ ∣ ≤ M ϱ |u_*|\le M\varrho ∣ u ∗ ∣ ≤ M ϱ , one has
∥ u ∗ Δ z η j ∥ 2 2 ≤ 2 A N ∫ 0 R 36 M 2 L 4 ϱ 2 2 π ϱ d ϱ = 144 π A N M 2 L 3 . \left\|u_*\Delta_z\eta_j\right\|_2^2\le2A_N\int_0^R\frac{36M^2}{L^4\varrho^2}2\pi\varrho\,\mathrm{d}\varrho
=\frac{144\pi A_NM^2}{L^3}. ∥ u ∗ Δ z η j ∥ 2 2 ≤ 2 A N ∫ 0 R L 4 ϱ 2 36 M 2 2 π ϱ d ϱ = L 3 144 π A N M 2 .
Its norm is at most 12 π M A N / L 3 / 2 12\sqrt{\pi}M\sqrt{A_N}/L^{3/2} 12 π M A N / L 3/2 . This cancellation is precisely where codimension two enters. The two normal variables produce the factor 1 / ϱ 1/\varrho 1/ ϱ in the radial Laplacian, and only the second derivative of the logarithmic profile remains.
Fourth, we estimate the remaining curvature term, whose factor satisfies
d − 2 r j + z 1 ∣ ∂ z 1 η j ∣ ≤ 2 d r min L ϱ . \frac{d-2}{r_j+z_1}|\partial_{z_1}\eta_j|\le\frac{2d}{r_{\min}L\varrho}. r j + z 1 d − 2 ∣ ∂ z 1 η j ∣ ≤ r m i n L ϱ 2 d .
For the numerical coefficients we use r j + z 1 ≥ r min ( 1 − 1 / ( 4 d ) ) r_j+z_1\ge r_{\min}(1-1/(4d)) r j + z 1 ≥ r m i n ( 1 − 1/ ( 4 d )) and
3 ( d − 2 ) 2 ( 1 − 1 / ( 4 d ) ) ≤ 2 d ( d ≥ 2 ) . \frac{3(d-2)}{2(1-1/(4d))}\le2d\qquad(d\ge2). 2 ( 1 − 1/ ( 4 d )) 3 ( d − 2 ) ≤ 2 d ( d ≥ 2 ) .
After multiplying by ∣ u ∗ ∣ ≤ M ϱ |u_*|\le M\varrho ∣ u ∗ ∣ ≤ M ϱ , the pointwise bound is 2 d M / ( r min L ) 2dM/(r_{\min}L) 2 d M / ( r m i n L ) . Enlarging the annulus to a disk gives
∥ u ∗ d − 2 r j + z 1 ∂ z 1 η j ∥ 2 2 ≤ 2 A N 4 d 2 M 2 r min 2 L 2 ∫ 0 R 2 π ϱ d ϱ = 8 π A N d 2 M 2 R 2 r min 2 L 2 . \left\|u_*\frac{d-2}{r_j+z_1}\partial_{z_1}\eta_j\right\|_2^2\le2A_N\frac{4d^2M^2}{r_{\min}^2L^2}\int_0^R2\pi\varrho\,\mathrm{d}\varrho=\frac{8\pi A_Nd^2M^2R^2}{r_{\min}^2L^2}. u ∗ r j + z 1 d − 2 ∂ z 1 η j 2 2 ≤ 2 A N r m i n 2 L 2 4 d 2 M 2 ∫ 0 R 2 π ϱ d ϱ = r m i n 2 L 2 8 π A N d 2 M 2 R 2 .
Its square root is at most 8 π M A N / ( 4 L ) \sqrt{8\pi}M\sqrt{A_N}/(4L) 8 π M A N / ( 4 L ) , since d R / r min ≤ 1 / 4 dR/r_{\min}\le1/4 d R / r m i n ≤ 1/4 . It may also be written as at most 2 π M A N / L 2\sqrt{\pi}M\sqrt{A_N}/L 2 π M A N / L . By L ≥ 1 L\ge1 L ≥ 1 , π < 2 \sqrt{\pi}<2 π < 2 , and 20 π < 40 20\sqrt{\pi}<40 20 π < 40 , the triangle inequality together with the four bounds above gives
∥ Δ e ∥ 2 ≤ 2 π M A N R + ( 6 + 12 + 2 ) π M A N / L \|\Delta e\|_2\le2\sqrt{\pi}M\sqrt{A_N}R+(6+12+2)\sqrt{\pi}M\sqrt{A_N/L} ∥Δ e ∥ 2 ≤ 2 π M A N R + ( 6 + 12 + 2 ) π M A N / L
≤ 4 M A N R + 40 M A N / L . \le4M\sqrt{A_NR}+40M\sqrt{A_N/L}. ≤ 4 M A N R + 40 M A N / L .
The final strict bounds follow from (6.2), which gives
4 M A N R ≤ η 4 M M + 1 A N A N + 1 < η 4 . 4M\sqrt{A_NR}\le\frac{\eta}{4}\frac{M}{M+1}\sqrt{\frac{A_N}{A_N+1}}<\frac{\eta}{4}. 4 M A N R ≤ 4 η M + 1 M A N + 1 A N < 4 η .
Using 25600 = 160 \sqrt{25600}=160 25600 = 160 in (6.3) also gives
40 M A N / L < η 4 M M + 1 A N A N + 1 < η 4 . 40M\sqrt{A_N/L}<\frac{\eta}{4}\frac{M}{M+1}\sqrt{\frac{A_N}{A_N+1}}<\frac{\eta}{4}. 40 M A N / L < 4 η M + 1 M A N + 1 A N < 4 η .
This proves (39) , including the strict inequalities. Thus the function v v v obtained from these cutoffs belongs to H 0 2 ( Ω d ) H_0^2(\Omega_d) H 0 2 ( Ω d ) . Its value and normal derivative vanish on the actual tube boundaries because χ ( 0 ) = χ ′ ( 0 ) = 0 \chi(0)=\chi'(0)=0 χ ( 0 ) = χ ′ ( 0 ) = 0 . The matching values χ ( 1 ) = 1 \chi(1)=1 χ ( 1 ) = 1 and χ ′ ( 1 ) = 0 \chi'(1)=0 χ ′ ( 1 ) = 0 also ensure that the function is H 2 H^2 H 2 across the outer edge of the cutoff support.
Compare the first two eigenvalues
For e = v − u ∗ ∈ H 0 2 ( B ) e=v-u_*\in H_0^2(B) e = v − u ∗ ∈ H 0 2 ( B ) , (6) gives ∥ D e ∥ 2 ≤ ( 2 / π ) ∥ Δ e ∥ 2 < η / 2 \lVert De\rVert_2\le(2/\pi)\lVert\Delta e\rVert_2<\eta/2 ∥ D e ∥ 2 ≤ ( 2/ π ) ∥ Δ e ∥ 2 < η /2 . Hence ∥ D v ∥ 2 ≥ 1 − η / 2 > 0 \lVert Dv\rVert_2\ge1-\eta/2>0 ∥ D v ∥ 2 ≥ 1 − η /2 > 0 , so the trial function is nonzero. From E ( u ∗ ) = 1 \mathcal{E}(u_*)=1 E ( u ∗ ) = 1 and Q ( u ∗ ) ≤ U Q(u_*)\le U Q ( u ∗ ) ≤ U , the numerator and denominator satisfy
∥ Δ v ∥ 2 ≤ ∥ Δ u ∗ ∥ 2 + ∥ Δ e ∥ 2 < U + η , \lVert\Delta v\rVert_2\le\lVert\Delta u_*\rVert_2+\lVert\Delta e\rVert_2<\sqrt{U}+\eta, ∥ Δ v ∥ 2 ≤ ∥ Δ u ∗ ∥ 2 + ∥ Δ e ∥ 2 < U + η ,
∥ D v ∥ 2 ≥ ∥ D u ∗ ∥ 2 − ∥ D e ∥ 2 > 1 − η . \lVert Dv\rVert_2\ge\lVert Du_*\rVert_2-\lVert De\rVert_2>1-\eta. ∥ D v ∥ 2 ≥ ∥ D u ∗ ∥ 2 − ∥ D e ∥ 2 > 1 − η .
The first Rayleigh minimum therefore satisfies
Λ 1 ( Ω d ) < ( U + η ) 2 ( 1 − η ) 2 ≤ U + 8 ( U + 1 ) η ≤ U + δ 16 C d . (40) \Lambda_1(\Omega_d)<\frac{(\sqrt{U}+\eta)^2}{(1-\eta)^2}\le U+8(U+1)\eta\le U+\frac{\delta}{16C_d}. \tag*{(40)} Λ 1 ( Ω d ) < ( 1 − η ) 2 ( U + η ) 2 ≤ U + 8 ( U + 1 ) η ≤ U + 16 C d δ . (40)
The middle inequality uses only U ≥ 1 U\ge1 U ≥ 1 and η ≤ 1 / 4 \eta\le1/4 η ≤ 1/4 :
( U + η ) 2 ( 1 − η ) 2 − U = 2 ( U + U ) η + ( 1 − U ) η 2 ( 1 − η ) 2 ≤ 4 U η ( 3 / 4 ) 2 = 64 9 U η ≤ 8 ( U + 1 ) η . \frac{(\sqrt{U}+\eta)^2}{(1-\eta)^2}-U
=\frac{2(\sqrt{U}+U)\eta+(1-U)\eta^2}{(1-\eta)^2}
\le\frac{4U\eta}{(3/4)^2}
=\frac{64}{9}U\eta\le8(U+1)\eta. ( 1 − η ) 2 ( U + η ) 2 − U = ( 1 − η ) 2 2 ( U + U ) η + ( 1 − U ) η 2 ≤ ( 3/4 ) 2 4 U η = 9 64 U η ≤ 8 ( U + 1 ) η .
The final bound in (6.5) follows from (6.1). For the second eigenvalue, every compactly supported smooth function on Ω d \Omega_d Ω d , after zero extension, is an element of C c ∞ ( B ∖ K N ) C^\infty_c(B\setminus K_N) C c ∞ ( B ∖ K N ) , and H 2 H^2 H 2 convergence keeps it in the closure, so
H 0 2 ( Ω d ) ⊂ H N ⟹ Λ 2 ( Ω d ) ≥ λ 2 , N > T . (41) H_0^2(\Omega_d)\subset H_N\quad\Longrightarrow\quad\Lambda_2(\Omega_d)\ge\lambda_{2,N}>T. \tag*{(41)} H 0 2 ( Ω d ) ⊂ H N ⟹ Λ 2 ( Ω d ) ≥ λ 2 , N > T . (41)
These two estimates use the two sides of the min-max principle. The trial function bounds Λ 1 \Lambda_1 Λ 1 from above, while zero extension into H N H_N H N bounds Λ 2 \Lambda_2 Λ 2 from below. Their margins are shown in Figure 11 . Combining (6.5) and (6.6) gives
Figure 11. Distribution of δ = V − C d U \delta= V - C_dU δ = V − C d U . The bounds for Λ 2 \Lambda_2 Λ 2 and C d Λ 1 C_d\Lambda_1 C d Λ 1 leave a gap larger than 3 δ / 16 3\delta/16 3 δ /16 . In the normalized coordinate, the spacings of the marks are exact.
Λ 2 ( Ω d ) − C d Λ 1 ( Ω d ) > ( C d U + δ / 4 ) − ( C d U + δ / 16 ) = 3 δ / 16 > 0. \Lambda_2(\Omega_d)-C_d\Lambda_1(\Omega_d)>(C_dU+\delta/4)-(C_dU+\delta/16)
=3\delta/16>0. Λ 2 ( Ω d ) − C d Λ 1 ( Ω d ) > ( C d U + δ /4 ) − ( C d U + δ /16 ) = 3 δ /16 > 0.
Consequently,
Λ 1 ( Ω d ) > 0 ⟹ Λ 2 ( Ω d ) Λ 1 ( Ω d ) > C d . \Lambda_1(\Omega_d)>0\quad\Longrightarrow\quad\frac{\Lambda_2(\Omega_d)}{\Lambda_1(\Omega_d)}>C_d. Λ 1 ( Ω d ) > 0 ⟹ Λ 1 ( Ω d ) Λ 2 ( Ω d ) > C d .
Verify smoothness and connectedness of the domain
The spectral comparison is complete. It remains to verify that Ω d \Omega_d Ω d indeed belongs to the geometric class in Theorem 1.1. It is bounded, and Step 3 has already produced pairwise disjoint tubular neighborhoods contained in B B B . In the normal coordinates of that step,
∂ ( removed tubes ) : ∣ z ∣ 2 = ε 2 , ∇ z ( ∣ z ∣ 2 − ε 2 ) = 2 z ≠ 0. \partial(\text{removed tubes}):\ |z|^2=\varepsilon^2,\qquad\nabla_z(|z|^2-\varepsilon^2)=2z\ne0. ∂ ( removed tubes ) : ∣ z ∣ 2 = ε 2 , ∇ z ( ∣ z ∣ 2 − ε 2 ) = 2 z = 0.
Since ε < R < r j \varepsilon< R < r_j ε < R < r j , these coordinates are smooth and injective near the level set, so the tube boundaries together with ∂ B \partial B ∂ B form a disjoint union of smooth compact boundary components.
For connectedness, take p , q ∈ Ω d p,q \in\Omega_d p , q ∈ Ω d . Figure 12 divides the construction of a path into three steps. We first choose a polygonal path that avoids the cores. Fix the endpoint p p p . If a segment starting from p p p meets some core, the corresponding intermediate points lie in
Figure 12. Connectivity. Left: a planar example with representative cores. Middle: the dashed segment is pushed radially so that it lies exactly on ∣ z ∣ ≥ 2 ε |z| \ge2\varepsilon ∣ z ∣ ≥ 2 ε . Right: the connecting arcs stay outside the removed disks, where p , p ′ p,p' p , p ′ denote normal coordinates.
⋃ j ⋃ n = 1 + ∞ { p + t ( y − p ) : y ∈ M j , 1 ≤ t ≤ n } . \bigcup_{j}\bigcup_{n=1}^{+\infty}\{p+t(y-p):y\in M_j,\ 1\le t\le n\}. j ⋃ n = 1 ⋃ + ∞ { p + t ( y − p ) : y ∈ M j , 1 ≤ t ≤ n } .
In smooth coordinate charts, each set in this union is a Lipschitz image, and
dim ( M j × [ 1 , n ] ) ≤ ( d − 2 ) + 1 = d − 1 < d . \dim(M_j\times[1,n])\le(d-2)+1=d-1<d. dim ( M j × [ 1 , n ]) ≤ ( d − 2 ) + 1 = d − 1 < d .
This union therefore has d d d -dimensional measure zero, and so does the analogous union for q q q . Choose b ∈ B b\in B b ∈ B outside both unions. By convexity,
[ p , b ] ∪ [ b , q ] ⊂ B ∖ K N . [p,b]\cup[b,q]\subset B\setminus K_N. [ p , b ] ∪ [ b , q ] ⊂ B ∖ K N .
This path may still enter the removed tubes, so a second step is needed. In each normal disk we use
z ⟼ max ( ∣ z ∣ , 2 ε ) ∣ z ∣ z , 0 < ∣ z ∣ < R , z \longmapsto\frac{\max(|z|,2\varepsilon)}{|z|}z,\qquad0<|z|<R, z ⟼ ∣ z ∣ max ( ∣ z ∣ , 2 ε ) z , 0 < ∣ z ∣ < R ,
and use the identity map elsewhere. Since 2 ε < R 2\varepsilon<R 2 ε < R , this map is continuous away from the cores, agrees with the identity map for ∣ z ∣ ≥ 2 ε |z|\ge2\varepsilon ∣ z ∣ ≥ 2 ε , and maps each perforated tube into 2 ε ≤ ∣ z ∣ < R 2\varepsilon\le|z|<R 2 ε ≤ ∣ z ∣ < R . The tubes are pairwise disjoint, so the pushed path lies in Ω d \Omega_d Ω d .
Finally, we reconnect all moved endpoints along normal rays. For either endpoint, the initial distance is ϱ > ε \varrho>\varepsilon ϱ > ε , and the pushed distance satisfies
[ ϱ , max ( ϱ , 2 ε ) ] ⊂ ( ε , R ) [\varrho,\max(\varrho,2\varepsilon)]\subset(\varepsilon,R) [ ϱ , max ( ϱ , 2 ε )] ⊂ ( ε , R )
whenever the endpoint lies in some R R R -tube.
Hence both endpoint connecting segments stay in Ω d \Omega_d Ω d . Joining them to the pushed path connects the original p p p and q q q within Ω d \Omega_d Ω d . This proves connectedness and completes the proof of Theorem 1.1.
This counterexample connects geometric eigenvalue inequalities with the analysis of localized constraints. Its two key estimates concern the relative separation of low modes and the bending energy cost of imposing additional boundary traces. Comparing these two quantities shows how interior boundary components can alter a spectral optimization problem. For elastic stability, this ratio measures the normalized separation of the first two buckling loads. For spectral geometry, this result distinguishes the sharp membrane comparison from its fourth order counterpart. The trace and cutoff estimates show that the required geometric constraints can be realized at an energy cost smaller than the spectral gain. Extending this approach to other operators would require separate estimates for the relative eigenvalue shifts and for the cost of imposing the corresponding boundary conditions.
Declaration of AI-assisted writing. During the preparation of this work, the authors used ChatGPT (OpenAI) to assist with language polishing and grammar checking of the manuscript, to perform certain elementary but lengthy algebraic and computational calculations, and to help prepare the figures. All AI-assisted output, including all calculations and figures, was independently checked and verified by the authors. The authors reviewed and edited the entire manuscript and take full responsibility for the accuracy and integrity of the content of this publication.
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.