Introduction
A macroscopic theory of diffusion-controlled reactions usually relies on the diffusion equation with appropriate boundary conditions [1 , 2 , 3 , 4 , 5 , 6 ] . It is therefore common to employ the Laplacian eigenfunctions to get spectral expansions for most quantities of interest, such as the diffusion propagator, the survival probability, the moments and the probability density of the first-reaction time on a target region, the concentration of diffusing particles, or the diffusive flux [7 , 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 ] . Despite its impressive progress, the conventional theory is limited to basic surface reactions on perfectly or partially reactive regions of the boundary that correspond respectively to Dirichlet and Robin boundary conditions. In turn, recent developments of the encounter-based approach [16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 ] employ the Steklov eigenfunctions to describe repetitive returns of a diffusing particle to the reactive regions and thus to incorporate more sophisticated reactions, such as non-Markovian binding, activation/passivation of a catalytic germ, targets with encounter-dependent reactivity, resetting mechanisms, permeation across membranes, etc [25 , 26 , 27 , 28 , 29 , 30 , 31 , 32 , 33 ] (see also Sec. 5).
For a bounded Euclidean domain Ω ⊂ R d \Omega\subset\mathbb{R}^{d} Ω ⊂ R d with a smooth boundary ∂ Ω \partial\Omega ∂ Ω , the conventional Steklov problem consists in finding the eigenpairs { μ , V } \{\mu,V\} { μ , V } satisfying
Δ V = 0 in Ω , ∂ n V = μ V on ∂ Ω , (1) \Delta V = 0 \quad\text{in } \Omega,\qquad\partial_n V = \mu V \quad\text{on } \partial\Omega, \tag*{(1)} Δ V = 0 in Ω , ∂ n V = μ V on ∂ Ω , (1)
where Δ \Delta Δ is the Laplace operator and ∂ n \partial_n ∂ n is the normal derivative on the boundary ∂ Ω \partial\Omega ∂ Ω oriented outwards the domain Ω \Omega Ω [34 , 35 ] . The reminiscent feature of the Steklov problem is that the spectral parameter μ \mu μ stands in the boundary condition. For a broad class of domains, the spectrum is known to be discrete, i.e., there is a countable sequence of eigenpairs solving the Steklov problem [36 , 37 ] . This spectral problem plays an important role in spectral geometry [38 ] and finds numerous applications in applied mathematics and physics such as approximation of harmonic functions [39 , 40 , 41 , 42 , 43 ] , domain decomposition [44 , 45 , 46 , 47 ] , electric impedance tomography [48 , 49 , 50 , 51 , 52 ] , and the aforementioned encounter-based approach to diffusion-controlled reactions.
When surface reactions occur on an open subset Γ \Gamma Γ of the otherwise impenetrable reflecting boundary, one needs to treat separately the reactive subset and the remaining passive surface of the confinement, denoted as ∂ Ω N = ∂ Ω ∖ Γ ‾ \partial\Omega_N = \partial\Omega\setminus\overline{\Gamma} ∂ Ω N = ∂ Ω ∖ Γ (Fig. 1a). The subset Γ \Gamma Γ can represent a specific target, a catalytic germ, an ion channel, an enzyme, an escape window, etc., depending on the physical, chemical or biological context. Accordingly, the above Steklov problem needs to be extended by imposing the Steklov condition on Γ \Gamma Γ and the Neumann condition on ∂ Ω N \partial\Omega_N ∂ Ω N . In this setting, one deals with the mixed Steklov-Neumann problem and searches for the eigenpairs { μ k ( p , Γ ) , V k ( p , Γ ) } \{\mu_k^{(p,\Gamma)}, V_k^{(p,\Gamma)}\} { μ k ( p , Γ ) , V k ( p , Γ ) } , enumerated by the index k = 0 , 1 , 2 , … k = 0, 1, 2, \ldots k = 0 , 1 , 2 , … and satisfying
( p − D Δ ) V k ( p , Γ ) = 0 in Ω , (2) (p-D\Delta)V_k^{(p,\Gamma)}=0 \quad\text{in }\Omega, \tag*{(2)} ( p − D Δ ) V k ( p , Γ ) = 0 in Ω , (2)
∂ n V k ( p , Γ ) = μ k ( p , Γ ) V k ( p , Γ ) on Γ , (3) \partial_n V_k^{(p,\Gamma)}=\mu_k^{(p,\Gamma)}V_k^{(p,\Gamma)} \quad\text{on }\Gamma, \tag*{(3)} ∂ n V k ( p , Γ ) = μ k ( p , Γ ) V k ( p , Γ ) on Γ , (3)
∂ n V k ( p , Γ ) = 0 on ∂ Ω N , (4) \partial_n V_k^{(p,\Gamma)}=0 \quad\text{on }\partial\Omega_N, \tag*{(4)} ∂ n V k ( p , Γ ) = 0 on ∂ Ω N , (4)
where D > 0 D>0 D > 0 is a diffusion constant and p ≥ 0 p\geq0 p ≥ 0 is a nonnegative parameter, which is introduced to give a broader perspective of the Steklov problem, as explained below.
The mixed Steklov-Neumann spectral problem for the Laplace equation (i.e., at p = 0 p=0 p = 0 ), also known as the sloshing problem (or ice-fishing problem) in hydrodynamics [53 , 54 , 55 , 56 , 57 ] , was discussed in the mathematical literature (see [58 ] and references therein). For instance, Bañuelos et al. derived several inequalities for the eigenvalues [59 ] , whereas Mayrand et al. studied the asymptotic behavior of the eigenvalues for the three-dimensional sloshing problem on a triangular prism [60 ] . Ammari et al. analyzed the behavior of the Green’s function in planar domains and proposed a method for optimizing this function at a given source-receiver pair [61 ] . In particular, they presented asymptotic expressions for the change in the Steklov-Neumann eigenpairs when a small portion of the boundary is changed from Steklov to Neumann condition. Despite this progress, some practically relevant questions remain unattended, e.g., the asymptotic behavior of eigenvalues and eigenfunctions as the subset Γ \Gamma Γ shrinks. In the context of first-passage phenomena [7 , 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 ] , this limit is known as the narrow escape problem; in particular, the asymptotic behavior of the mean first-escape time and Laplacian eigenvalues under singular perturbations of the boundary was thoroughly investigated [62 , 63 , 64 , 65 , 66 , 67 , 68 , 69 , 70 , 71 , 72 , 73 , 74 , 75 , 76 , 77 , 78 , 79 , 80 , 81 , 82 , 83 , 84 , 85 , 86 ] . We aim to inspect the behavior of the Steklov-Neumann eigenvalues and eigenfunctions in the small-target limit Γ → ∅ \Gamma\to\varnothing Γ → ∅ and then to relate the derived spectral results to the narrow escape problem.
The paper is organized as follows. In Sec. 2, we start by recalling the basic properties of the Steklov-Neumann problem and its spectral expansions. In particular, we argue that the restriction of the Green’s function G ~ 0 ( x , p ∣ x 0 ) \widetilde{G}_0(\boldsymbol{x},p|\boldsymbol{x}_0) G 0 ( x , p ∣ x 0 ) of the modified Helmholtz equation (1.2a) to Γ × Γ \Gamma\times\Gamma Γ × Γ is the kernel of an integral operator that determines the eigenpairs { μ k ( p , Γ ) , V k ( p , Γ ) } \{\mu_k^{(p,\Gamma)},V_k^{(p,\Gamma)}\} { μ k ( p , Γ ) , V k ( p , Γ ) } of the spectral problem (1.2). This property allows one to construct the eigenpairs numerically for any subset Γ \Gamma Γ and to highlight the impact of its shape. We also discuss the specific features of this analysis in the limit p = 0 p=0 p = 0 and the central role of the pseudo-Green’s function. Section 3 presents the main asymptotic results for a bounded domain Ω ⊂ R d \Omega\subset\mathbb{R}^d Ω ⊂ R d with a smooth boundary ∂ Ω \partial\Omega ∂ Ω . We consider either a connected subset Γ ⊂ ∂ Ω \Gamma\subset\partial\Omega Γ ⊂ ∂ Ω of perimeter 2 ϵ 2\epsilon 2 ϵ for the case d = 2 d=2 d = 2 , or a (curved) disk-like subset Γ \Gamma Γ of radius ϵ \epsilon ϵ for the case d = 3 d=3 d = 3 (e.g., a cap on a sphere). In the limit ϵ → 0 \epsilon\to0 ϵ → 0 , we obtain the following leading-order asymptotic behavior of the eigenvalues μ k ( p , Γ ) \mu_k^{(p,\Gamma)} μ k ( p , Γ ) and eigenfunctions V k ( p , Γ ) V_k^{(p,\Gamma)} V k ( p , Γ ) restricted to Γ \Gamma Γ :
μ k ( p , Γ ) ≈ μ ^ k ϵ , V k ( p , Γ ) ( x ) ≈ V ^ k ( x / ϵ ) ϵ ( d − 1 ) / 2 ( x ∈ Γ ) , (5) \mu_k^{(p,\Gamma)} \approx\frac{\widehat{\mu}_k}{\epsilon}, \qquad V_k^{(p,\Gamma)}(\boldsymbol{x}) \approx\frac{\widehat{V}_k(\boldsymbol{x}/\epsilon)}{\epsilon^{(d-1)/2}} \quad(\boldsymbol{x} \in\Gamma), \tag*{(5)} μ k ( p , Γ ) ≈ ϵ μ k , V k ( p , Γ ) ( x ) ≈ ϵ ( d − 1 ) /2 V k ( x / ϵ ) ( x ∈ Γ ) , (5)
for any 0 ≤ p ≪ 1 / ( D ϵ 2 ) 0 \le p \ll1/(D\epsilon^2) 0 ≤ p ≪ 1/ ( D ϵ 2 ) . Here { μ ^ k , V ^ k } \{\widehat{\mu}_k,\widehat{V}_k\} { μ k , V k } are the eigenpairs of an auxiliary Steklov-Neumann problem for the exterior of an interval in the half-plane (d = 2 d=2 d = 2 ), or for the exterior of a disk in the half-space (d = 3 d=3 d = 3 ). We discuss different ways to construct these eigenpairs, in particular, by reformulating the Steklov-Neumann problem as a spectral problem for an integral operator whose kernel is explicitly derived. Section 4 provides numerical illustrations for two relevant examples: an arc-shaped subset Γ \Gamma Γ on the boundary of a disk (Figure 1 b) and a spherical cap Γ \Gamma Γ on the boundary of a ball (Figure 1 c). In these settings, the Green’s function and the pseudo-Green’s function are known explicitly, which allows us to construct efficiently the eigenpairs { μ k ( p , Γ ) , V k ( p , Γ ) } \{\mu_k^{(p,\Gamma)},V_k^{(p,\Gamma)}\} { μ k ( p , Γ ) , V k ( p , Γ ) } for any size of the subset Γ \Gamma Γ and any positive p p p . In this way, we illustrate the accuracy and the validity range of the asymptotic relations (1.3). Section 5 presents a straightforward application of these spectral results to first-passage problems. In particular, we express the mean first-reaction time (MFRT) in terms of the eigenpairs { μ k ( 0 , Γ ) , V k ( 0 , Γ ) } \{\mu_k^{(0,\Gamma)},V_k^{(0,\Gamma)}\} { μ k ( 0 , Γ ) , V k ( 0 , Γ ) } , discuss its asymptotic behavior and re-analyze the validity of the constant-flux approximation for the MFRT [83 ] . Most importantly, we reveal the effect of partial reactivity over the whole range of reactivities and extend the asymptotic analysis to more sophisticated surface reactions that go beyond the conventional narrow escape problem. Section 6 summarizes the main results and concludes the paper.
Figure 1. Schematic illustration of the mixed Steklov-Neumann problem (1.2) in a generic Euclidean domain (a), in a disk (b), and in a ball (c). The solid blue line indicates the reflecting boundary ∂ Ω N \partial\Omega_N ∂ Ω N with Neumann boundary condition, while the dashed red line shows the subset Γ \Gamma Γ with the Steklov condition.
We emphasize that the presented analysis remains at a formal level and lacks rigorous proofs. For instance, the convergence of spectral expansions, integrations by parts, the use of Dirac distributions, and probabilistic interpretations are not yet properly demonstrated. For this reason, the related statements are not formulated as theorems. The omission of proofs is partly compensated by an extensive numerical analysis that validates the asymptotic results and illustrates their accuracy. We expect that the presented results can stimulate a broader interest among mathematicians to this topic to ensure its more rigorous treatment in the future.
General spectral properties
This section prepares the theoretical ground to deal with the mixed Steklov-Neumann problem (1.2). In Sec. 2.1, we recall the basic properties of the eigenvalues and eigenfunctions and their relation to the Green’s function of the modified Helmholtz equation and to the Neumann Laplacian. Section 2.2 focuses on the asymptotic behavior of the Steklov eigenpairs { μ k ( p , Γ ) , V k ( p , Γ ) } \{\mu_k^{(p,\Gamma)},V_k^{(p,\Gamma)}\} { μ k ( p , Γ ) , V k ( p , Γ ) } in the limit p → 0 p \to0 p → 0 . In particular, we show how μ k ( 0 , Γ ) \mu_k^{(0,\Gamma)} μ k ( 0 , Γ ) and V k ( 0 , Γ ) V_k^{(0,\Gamma)} V k ( 0 , Γ ) can be obtained as eigenpairs of the integral equation (43) , whose kernel in Eq. (41) is related to the pseudo-Green’s function. Section 2.3 presents useful interpretations of the first-order corrections A Γ A_{\Gamma} A Γ and W 0 ( Γ ) W_{0}^{(\Gamma)} W 0 ( Γ ) to the principal eigenvalue μ 0 ( p , Γ ) \mu_{0}^{(p,\Gamma)} μ 0 ( p , Γ ) and eigenfunction V 0 ( p , Γ ) V_{0}^{(p,\Gamma)} V 0 ( p , Γ ) in the limit p → 0 p \to0 p → 0 . For instance, we show how A Γ A_{\Gamma} A Γ determines the variance of the boundary local time on Γ \Gamma Γ and the leading-order term of the mean first-passage time (MFPT) in the narrow escape problem.
Mixed Steklov-Neumann problem
We consider a bounded Euclidean domain Ω ⊂ R d \Omega\subset\mathbb{R}^{d} Ω ⊂ R d with a smooth connected boundary ∂ Ω \partial\Omega ∂ Ω . An open connected set Γ ⊂ ∂ Ω \Gamma\subset\partial\Omega Γ ⊂ ∂ Ω is called a “reactive subset” of the boundary ∂ Ω \partial\Omega ∂ Ω , whereas its reflecting part is denoted as ∂ Ω N = ∂ Ω \ Γ ‾ \partial\Omega_{N} = \partial\Omega\backslash\overline{\Gamma} ∂ Ω N = ∂ Ω\ Γ (Figure 1 a). To avoid possible technical issues on the “junction” between Γ \Gamma Γ and ∂ Ω N \partial\Omega_{N} ∂ Ω N , we assume that ∂ Γ \partial\Gamma ∂ Γ is smooth. In this setting, the mixed Steklov-Neumann problem (1.2) is known to have a discrete spectrum for any p ≥ 0 p \ge0 p ≥ 0 , i.e., there is a countable sequence of eigenpairs { μ k ( p , Γ ) , V k ( p , Γ ) } \{\mu_{k}^{(p,\Gamma)}, V_{k}^{(p,\Gamma)}\} { μ k ( p , Γ ) , V k ( p , Γ ) } satisfying (1.2); in particular, the eigenvalues can be enumerated in increasing order:
0 ≤ μ 0 ( p , Γ ) < μ 1 ( p , Γ ) ≤ μ 2 ( p , Γ ) ≤ … ↗ + ∞ , (6) 0 \le\mu_{0}^{(p,\Gamma)} < \mu_{1}^{(p,\Gamma)} \le\mu_{2}^{(p,\Gamma)} \le\ldots\nearrow+\infty, \tag*{(6)} 0 ≤ μ 0 ( p , Γ ) < μ 1 ( p , Γ ) ≤ μ 2 ( p , Γ ) ≤ … ↗ + ∞ , (6)
and the first eigenvalue μ 0 ( p , Γ ) \mu_{0}^{(p,\Gamma)} μ 0 ( p , Γ ) is simple (see, e.g., [38 , 59 ] for mathematical details in the case p = 0 p = 0 p = 0 ; an extension to p > 0 p > 0 p > 0 is straightforward). The eigenfunctions V k ( p , Γ ) V_{k}^{(p,\Gamma)} V k ( p , Γ ) belong to the Sobolev space H 1 ( Ω ) H^{1}(\Omega) H 1 ( Ω ) ; moreover, since p ≥ 0 p \ge0 p ≥ 0 , the eigenfunctions V k ( p , Γ ) V_{k}^{(p,\Gamma)} V k ( p , Γ ) can be chosen to be real. Their restrictions to Γ \Gamma Γ , denoted as v k ( p , Γ ) = V k ( p , Γ ) ∣ Γ v_{k}^{(p,\Gamma)} = V_{k}^{(p,\Gamma)}\rvert_{\Gamma} v k ( p , Γ ) = V k ( p , Γ ) ∣ Γ , form a complete orthonormal basis of the space L 2 ( Γ ) L^{2}(\Gamma) L 2 ( Γ ) of square-integrable functions on Γ \Gamma Γ , i.e.,
∫ Γ v k ( p , Γ ) v k ′ ( p , Γ ) = δ k , k ′ , (7) \int_{\Gamma} v_{k}^{(p,\Gamma)}v_{k'}^{(p,\Gamma)} = \delta_{k,k'}, \tag*{(7)} ∫ Γ v k ( p , Γ ) v k ′ ( p , Γ ) = δ k , k ′ , (7)
where δ k , k ′ = 1 \delta_{k,k'} = 1 δ k , k ′ = 1 for k = k ′ k = k' k = k ′ and 0 otherwise. This relation fixes the normalization of v k ( p , Γ ) v_{k}^{(p,\Gamma)} v k ( p , Γ ) and thus of V k ( p , Γ ) V_{k}^{(p,\Gamma)} V k ( p , Γ ) . In turn, the L 2 ( Ω ) L^{2}(\Omega) L 2 ( Ω ) norm of V k ( p , Γ ) V_{k}^{(p,\Gamma)} V k ( p , Γ ) can be found via the identity [16 , 87 , 88 ] :
∫ Ω ∣ V k ( p , Γ ) ∣ 2 = D ∂ p μ k ( p , Γ ) (8) \int_{\Omega} \left|V_{k}^{(p,\Gamma)}\right|^{2} = D\partial_{p}\mu_{k}^{(p,\Gamma)} \tag*{(8)} ∫ Ω V k ( p , Γ ) 2 = D ∂ p μ k ( p , Γ ) (8)
(see Appendix A.§ 1 for a formal derivation and discussion). Note that μ k ( p , Γ ) \mu_{k}^{(p,\Gamma)} μ k ( p , Γ ) and v k ( p , Γ ) v_{k}^{(p,\Gamma)} v k ( p , Γ ) turn out to be the eigenvalues and eigenfunctions of the Dirichlet-to-Neumann operator M p ( Γ ) \mathcal{M}_{p}^{(\Gamma)} M p ( Γ ) :
M p ( Γ ) v k ( p , Γ ) = μ k ( p , Γ ) v k ( p , Γ ) . (9) \mathcal{M}_{p}^{(\Gamma)}v_{k}^{(p,\Gamma)} = \mu_{k}^{(p,\Gamma)}v_{k}^{(p,\Gamma)}. \tag*{(9)} M p ( Γ ) v k ( p , Γ ) = μ k ( p , Γ ) v k ( p , Γ ) . (9)
This operator acts from H 1 2 ( Γ ) H^{\frac{1}{2}}(\Gamma) H 2 1 ( Γ ) to H − 1 2 ( Γ ) H^{-\frac{1}{2}}(\Gamma) H − 2 1 ( Γ ) and associates to a given function f f f on Γ \Gamma Γ another function g g g on Γ \Gamma Γ such that M p ( Γ ) f = g = ( ∂ n u ) ∣ Γ \mathcal{M}_{p}^{(\Gamma)}f = g = (\partial_{n}u)\rvert_{\Gamma} M p ( Γ ) f = g = ( ∂ n u ) ∣ Γ , where u u u is the unique solution of the boundary value problem:
( p − D Δ ) u = 0 in Ω , u ∣ Γ = f , ( ∂ n u ) ∣ ∂ Ω N = 0 (10) (p - D\Delta)u = 0 \quad\text{in } \Omega,\qquad u\rvert_{\Gamma} = f,\qquad(\partial_{n}u)\rvert_{\partial\Omega_{N}} = 0 \tag*{(10)} ( p − D Δ ) u = 0 in Ω , u ∣ Γ = f , ( ∂ n u ) ∣ ∂ Ω N = 0 (10)
(see [38 , 89 ] for mathematical details and the definition of the Sobolev spaces H ± 1 2 ( Γ ) H^{\pm\frac{1}{2}}(\Gamma) H ± 2 1 ( Γ ) ).
We also introduce the Green’s function G ~ q ( x , p ∣ x 0 ) \widetilde{G}_q(x,p\mid x_0) G q ( x , p ∣ x 0 ) of the modified Helmholtz equation, which satisfies for any x 0 ∈ Ω x_0 \in\Omega x 0 ∈ Ω :
( p − D Δ ) G ~ q ( x , p ∣ x 0 ) = δ ( x − x 0 ) ( x ∈ Ω ) , (11) (p-D\Delta)\widetilde{G}_q(x,p\mid x_0)=\delta(x-x_0) \qquad(x\in\Omega), \tag*{(11)} ( p − D Δ ) G q ( x , p ∣ x 0 ) = δ ( x − x 0 ) ( x ∈ Ω ) , (11)
∂ n G ~ q ( x , p ∣ x 0 ) + q G ~ q ( x , p ∣ x 0 ) = 0 ( x ∈ Γ ) , (12) \partial_n\widetilde{G}_q(x,p\mid x_0)+q\widetilde{G}_q(x,p\mid x_0)=0 \qquad(x\in\Gamma), \tag*{(12)} ∂ n G q ( x , p ∣ x 0 ) + q G q ( x , p ∣ x 0 ) = 0 ( x ∈ Γ ) , (12)
∂ n G ~ q ( x , p ∣ x 0 ) = 0 ( x ∈ ∂ Ω N ) , (13) \partial_n\widetilde{G}_q(x,p\mid x_0)=0 \qquad(x\in\partial\Omega_N), \tag*{(13)} ∂ n G q ( x , p ∣ x 0 ) = 0 ( x ∈ ∂ Ω N ) , (13)
where δ ( x − x 0 ) \delta(x-x_0) δ ( x − x 0 ) is the Dirac distribution, and 0 ≤ q ≤ + ∞ 0\le q\le+\infty 0 ≤ q ≤ + ∞ is a fixed parameter; in the following, we mainly use two limits: q = 0 q=0 q = 0 (Neumann condition on Γ \Gamma Γ ) and q = ∞ q=\infty q = ∞ (Dirichlet condition on Γ \Gamma Γ ).
Following [18 ] , we get the spectral expansion of the Green’s function in terms of the Steklov-Neumann eigenpairs:
G ~ q ( x , p ∣ x 0 ) = G ~ ∞ ( x , p ∣ x 0 ) + 1 D ∑ k = 0 ∞ V k ( p , Γ ) ( x ) V k ( p , Γ ) ( x 0 ) q + μ k ( p , Γ ) . (14) \widetilde{G}_q(x,p\mid x_0)=\widetilde{G}_\infty(x,p\mid x_0)+\frac{1}{D}\sum_{k=0}^{\infty}\frac{V_k^{(p,\Gamma)}(x)V_k^{(p,\Gamma)}(x_0)}{q+\mu_k^{(p,\Gamma)}}. \tag*{(14)} G q ( x , p ∣ x 0 ) = G ∞ ( x , p ∣ x 0 ) + D 1 k = 0 ∑ ∞ q + μ k ( p , Γ ) V k ( p , Γ ) ( x ) V k ( p , Γ ) ( x 0 ) . (14)
(see Appendix A.§ 2 for a formal derivation and discussion on its convergence). Setting q = 0 q=0 q = 0 and restricting both x x x and x 0 x_0 x 0 to Γ \Gamma Γ , we have
D G ~ 0 ( x , p ∣ x 0 ) = ∑ k = 0 ∞ v k ( p , Γ ) ( x ) v k ( p , Γ ) ( x 0 ) μ k ( p , Γ ) ( x , x 0 ∈ Γ ) . (15) D\widetilde{G}_0(x,p\mid x_0)=\sum_{k=0}^{\infty}\frac{v_k^{(p,\Gamma)}(x)v_k^{(p,\Gamma)}(x_0)}{\mu_k^{(p,\Gamma)}} \qquad(x,x_0\in\Gamma). \tag*{(15)} D G 0 ( x , p ∣ x 0 ) = k = 0 ∑ ∞ μ k ( p , Γ ) v k ( p , Γ ) ( x ) v k ( p , Γ ) ( x 0 ) ( x , x 0 ∈ Γ ) . (15)
One sees that the restriction of D G ~ 0 ( x , p ∣ x 0 ) D\widetilde{G}_0(x,p\mid x_0) D G 0 ( x , p ∣ x 0 ) to Γ × Γ \Gamma\times\Gamma Γ × Γ is the kernel of an integral operator, which is the inverse of the Dirichlet-to-Neumann operator M p ( Γ ) \mathcal{M}_p^{(\Gamma)} M p ( Γ ) on Γ \Gamma Γ . As a consequence, one can search for its eigenvalues μ k ( p , Γ ) \mu_k^{(p,\Gamma)} μ k ( p , Γ ) and eigenfunctions v k ( p , Γ ) v_k^{(p,\Gamma)} v k ( p , Γ ) as the eigenpairs of this integral operator:
∫ Γ d x D G ~ 0 ( x , p ∣ x 0 ) v k ( p , Γ ) ( x ) = v k ( p , Γ ) ( x 0 ) μ k ( p , Γ ) ( x 0 ∈ Γ , k ≥ 0 ) . (16) \int_{\Gamma} dx\,D\widetilde{G}_0(x,p\mid x_0)v_k^{(p,\Gamma)}(x)=\frac{v_k^{(p,\Gamma)}(x_0)}{\mu_k^{(p,\Gamma)}} \qquad(x_0\in\Gamma,\ k\ge0). \tag*{(16)} ∫ Γ d x D G 0 ( x , p ∣ x 0 ) v k ( p , Γ ) ( x ) = μ k ( p , Γ ) v k ( p , Γ ) ( x 0 ) ( x 0 ∈ Γ , k ≥ 0 ) . (16)
Once v k ( p , Γ ) v_k^{(p,\Gamma)} v k ( p , Γ ) is obtained on Γ \Gamma Γ , its extension V k ( p , Γ ) ( x 0 ) V_k^{(p,\Gamma)}(x_0) V k ( p , Γ ) ( x 0 ) to the whole domain x 0 ∈ Ω x_0\in\Omega x 0 ∈ Ω follows as
V k ( p , Γ ) ( x 0 ) = μ k ( p , Γ ) ∫ Γ d x D G ~ 0 ( x , p ∣ x 0 ) v k ( p , Γ ) ( x ) ( x 0 ∈ Ω ‾ ) . (17) V_k^{(p,\Gamma)}(x_0)=\mu_k^{(p,\Gamma)}\int_{\Gamma} dx\,D\widetilde{G}_0(x,p\mid x_0)v_k^{(p,\Gamma)}(x) \qquad(x_0\in\overline{\Omega}). \tag*{(17)} V k ( p , Γ ) ( x 0 ) = μ k ( p , Γ ) ∫ Γ d x D G 0 ( x , p ∣ x 0 ) v k ( p , Γ ) ( x ) ( x 0 ∈ Ω ) . (17)
This relation is obtained by multiplying Eq. (2.6a) with q = 0 q=0 q = 0 by V k ( p , Γ ) ( x ) V_k^{(p,\Gamma)}(x) V k ( p , Γ ) ( x ) , multiplying Eq. (1.2a) by G ~ 0 ( x , p ∣ x 0 ) \widetilde{G}_0(x,p\mid x_0) G 0 ( x , p ∣ x 0 ) , subtracting these equations, integrating over x ∈ Ω x\in\Omega x ∈ Ω , and using the Green’s formula and the boundary conditions (1.2b, 2.6b).
Importantly, the Green’s function G ~ 0 ( x , p ∣ x 0 ) \widetilde{G}_0(x,p\mid x_0) G 0 ( x , p ∣ x 0 ) does not depend on Γ \Gamma Γ . Indeed, at q = 0 q=0 q = 0 , the Robin boundary condition (2.6b) is reduced to the Neumann one so that one deals with the Neumann boundary condition on the whole boundary ∂ Ω \partial\Omega ∂ Ω . In other words, if the Green’s function G ~ 0 ( x , p ∣ x 0 ) \widetilde{G}_0(x,p\mid x_0) G 0 ( x , p ∣ x 0 ) is known for a given domain Ω \Omega Ω , its restriction to Γ × Γ \Gamma\times\Gamma Γ × Γ determines the integral operator, whose eigenmodes solve the mixed Steklov-Neumann problem. For some simple domains, the Green’s function is known explicitly (e.g., see Appendices B and C for a disk and a ball). In general, it admits the standard spectral expansion
G ~ 0 ( x , p ∣ x 0 ) = ∑ k = 0 ∞ u k N ( x ) u k N ( x 0 ) λ k N + p / D ( x , x 0 ∈ Ω ‾ ) (18) \widetilde{G}_0(x,p\mid x_0)=\sum_{k=0}^{\infty}\frac{u_k^N(x)u_k^N(x_0)}{\lambda_k^N+p/D} \qquad(x,x_0\in\overline{\Omega}) \tag*{(18)} G 0 ( x , p ∣ x 0 ) = k = 0 ∑ ∞ λ k N + p / D u k N ( x ) u k N ( x 0 ) ( x , x 0 ∈ Ω ) (18)
over the L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) -normalized eigenfunctions u k N u_k^N u k N and eigenvalues λ k N \lambda_k^N λ k N of the Laplace operator with Neumann boundary condition:
− Δ u k N = λ k N u k N in Ω , ∂ n u k N = 0 on ∂ Ω . (19) -\Delta u_k^N = \lambda_k^N u_k^N \quad\text{in } \Omega,\qquad\partial_n u_k^N = 0 \quad\text{on } \partial\Omega. \tag*{(19)} − Δ u k N = λ k N u k N in Ω , ∂ n u k N = 0 on ∂ Ω. (19)
The expansion (18) can be verified by applying the operator ( p − D Δ ) (p-D\Delta) ( p − D Δ ) to the right-hand side and using the completeness relation
∑ k = 0 ∞ u k N ( x ) u k N ( x 0 ) = δ ( x − x 0 ) ( x , x 0 ∈ Ω ) , (20) \sum_{k=0}^{\infty} u_k^N(x)u_k^N(x_0)=\delta(x-x_0)\qquad(x,x_0\in\Omega), \tag*{(20)} k = 0 ∑ ∞ u k N ( x ) u k N ( x 0 ) = δ ( x − x 0 ) ( x , x 0 ∈ Ω ) , (20)
which reflects the completeness of the basis of the Neumann Laplacian eigenfunctions { u k N } \{u_k^N\} { u k N } in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) (this relation can also be formally understood as the expansion of the Dirac distribution onto the basis { u k N } \{u_k^N\} { u k N } ). Alternatively, one can use the expansion (15) with Γ = ∂ Ω \Gamma=\partial\Omega Γ = ∂ Ω :
G ~ 0 ( x , p ∣ x 0 ) = ∑ k = 0 ∞ v k ( p , ∂ Ω ) ( x ) v k ( p , ∂ Ω ) ( x 0 ) D μ k ( p , ∂ Ω ) ( x , x 0 ∈ ∂ Ω ) , (21) \widetilde{G}_0(x,p\mid x_0)=\sum_{k=0}^{\infty}\frac{v_k^{(p,\partial\Omega)}(x)v_k^{(p,\partial\Omega)}(x_0)}{D\mu_k^{(p,\partial\Omega)}}\qquad(x,x_0\in\partial\Omega), \tag*{(21)} G 0 ( x , p ∣ x 0 ) = k = 0 ∑ ∞ D μ k ( p , ∂ Ω ) v k ( p , ∂ Ω ) ( x ) v k ( p , ∂ Ω ) ( x 0 ) ( x , x 0 ∈ ∂ Ω ) , (21)
which is based on the Steklov eigenfunctions v k ( p , ∂ Ω ) v_k^{(p,\partial\Omega)} v k ( p , ∂ Ω ) on the whole boundary ∂ Ω \partial\Omega ∂ Ω and is thus independent of the subset Γ \Gamma Γ .
Limit p = 0 p=0 p = 0
At p = 0 p=0 p = 0 , the smallest Steklov eigenvalue is zero, while the corresponding eigenfunction is constant:
μ 0 ( 0 , Γ ) = 0 , V 0 ( 0 , Γ ) = 1 ∣ Γ ∣ , (22) \mu_0^{(0,\Gamma)}=0,\qquad V_0^{(0,\Gamma)}=\frac{1}{\sqrt{|\Gamma|}}, \tag*{(22)} μ 0 ( 0 , Γ ) = 0 , V 0 ( 0 , Γ ) = ∣Γ∣ 1 , (22)
where ∣ Γ ∣ |\Gamma| ∣Γ∣ denotes the area of Γ \Gamma Γ , and we used the normalization (7) of the Steklov eigenfunctions. As a consequence, its contribution to the spectral expansion (15) diverges, which reflects the well-known fact that the conventional Green’s function for the Laplace equation does not exist for the interior Neumann problem. This divergence can be easily amended by subtracting the divergent term and thus dealing with pseudo-Green’s function (also known as Neumann Green’s function [65 ] ):
G 0 ( x ∣ x 0 ) = lim p → 0 ( D ~ G 0 ( x , p ∣ x 0 ) − 1 ∣ Ω ∣ ( p / D ) ) , (23) G_0(x\mid x_0)=\lim_{p\to0}\left(\widetilde{D}G_0(x,p\mid x_0)-\frac{1}{|\Omega|(p/D)}\right), \tag*{(23)} G 0 ( x ∣ x 0 ) = p → 0 lim ( D G 0 ( x , p ∣ x 0 ) − ∣Ω∣ ( p / D ) 1 ) , (23)
where ∣ Ω ∣ |\Omega| ∣Ω∣ denotes the volume of Ω \Omega Ω . Since λ 0 N = 0 \lambda_0^N=0 λ 0 N = 0 and u 0 N = 1 / ∣ Ω ∣ u_0^N=1/\sqrt{|\Omega|} u 0 N = 1/ ∣Ω∣ , the subtracted term is precisely the contribution of the principal eigenmode (k = 0 k=0 k = 0 ) to the spectral expansion (18) . As a consequence, the pseudo-Green’s function admits the same spectral expansion without the principal eigenmode:
G 0 ( x ∣ x 0 ) = ∑ k = 1 ∞ u k N ( x ) u k N ( x 0 ) λ k N ( x , x 0 ∈ Ω ‾ ) . (24) G_0(x\mid x_0)=\sum_{k=1}^{\infty}\frac{u_k^N(x)u_k^N(x_0)}{\lambda_k^N}\qquad(x,x_0\in\overline{\Omega}). \tag*{(24)} G 0 ( x ∣ x 0 ) = k = 1 ∑ ∞ λ k N u k N ( x ) u k N ( x 0 ) ( x , x 0 ∈ Ω ) . (24)
Applying the Laplace operator to the spectral expansion (24) , one can check that the pseudo-Green’s function satisfies for any x 0 ∈ Ω x_0\in\Omega x 0 ∈ Ω :
Δ G 0 ( x ∣ x 0 ) = 1 ∣ Ω ∣ − δ ( x − x 0 ) ( x ∈ Ω ) , (25) \Delta G_0(x\mid x_0)=\frac{1}{|\Omega|}-\delta(x-x_0)\qquad(x\in\Omega), \tag*{(25)} Δ G 0 ( x ∣ x 0 ) = ∣Ω∣ 1 − δ ( x − x 0 ) ( x ∈ Ω ) , (25)
∂ n G 0 ( x ∣ x 0 ) = 0 ( x ∈ ∂ Ω ) , (26) \partial_nG_0(x\mid x_0)=0\qquad(x\in\partial\Omega), \tag*{(26)} ∂ n G 0 ( x ∣ x 0 ) = 0 ( x ∈ ∂ Ω ) , (26)
where we used the completeness relation (20) . While the limit in Eq. (23) and the spectral expansion (24) determine G 0 ( x ∣ x 0 ) \mathcal{G}_0(x|x_0) G 0 ( x ∣ x 0 ) uniquely, the boundary value problem (2.18) defines the pseudo-Green’s function up to an additive constant, which can be fixed by the additional condition:
∫ Ω d x G 0 ( x ∣ x 0 ) = 0. (27) \int_{\Omega} dx\,\mathcal{G}_0(x|x_0)=0. \tag*{(27)} ∫ Ω d x G 0 ( x ∣ x 0 ) = 0. (27)
In fact, this condition is consistent with the fact that the integral of the spectral expansion (24) over Ω \Omega Ω is zero because all eigenfunctions u k N u_k^N u k N with k ≥ 1 k \ge1 k ≥ 1 are orthogonal to the constant function u N u^N u N . Note that the symmetry of the pseudo-Green’s function, G 0 ( x ∣ x 0 ) = G 0 ( x 0 ∣ x ) \mathcal{G}_0(x|x_0)=\mathcal{G}_0(x_0|x) G 0 ( x ∣ x 0 ) = G 0 ( x 0 ∣ x ) , implies that
∫ Ω d x 0 G 0 ( x ∣ x 0 ) = 0. (28) \int_{\Omega} dx_0\,\mathcal{G}_0(x|x_0)=0. \tag*{(28)} ∫ Ω d x 0 G 0 ( x ∣ x 0 ) = 0. (28)
We stress that the unusual constant 1 / ∣ Ω ∣ 1/|\Omega| 1/∣Ω∣ in Eq. (2.18a) is necessary to satisfy the convergence theorem; indeed, the integral of Eq. (2.18a) over x ∈ Ω x \in\Omega x ∈ Ω reads
0 = ∫ Ω d x ( 1 ∣ Ω ∣ − δ ( x − x 0 ) ) = ∫ Ω d x Δ G 0 ( x ∣ x 0 ) = ∫ ∂ Ω d x ∂ n G 0 ( x ∣ x 0 ) = 0 , 0=\int_{\Omega} dx\left(\frac{1}{|\Omega|}-\delta(x-x_0)\right)=\int_{\Omega} dx\,\Delta\mathcal{G}_0(x|x_0)=\int_{\partial\Omega}dx\,\partial_n\mathcal{G}_0(x|x_0)=0, 0 = ∫ Ω d x ( ∣Ω∣ 1 − δ ( x − x 0 ) ) = ∫ Ω d x Δ G 0 ( x ∣ x 0 ) = ∫ ∂ Ω d x ∂ n G 0 ( x ∣ x 0 ) = 0 ,
and this identity could not be satisfied without the constant 1 / ∣ Ω ∣ 1/|\Omega| 1/∣Ω∣ . As no such constant is present in Eq. (2.6a) for other Green’s functions, G 0 ( x ∣ x 0 ) \mathcal{G}_0(x|x_0) G 0 ( x ∣ x 0 ) is called pseudo -Green’s function.
The last step consists in expressing G 0 ( x ∣ x 0 ) \mathcal{G}_0(x|x_0) G 0 ( x ∣ x 0 ) in terms of the Steklov eigenfunctions V k ( 0 , Γ ) V_k^{(0,\Gamma)} V k ( 0 , Γ ) , in analogy to Eq. (24) . For this purpose, we need to compute the limit in Eq. (23) . This computation relies on the analyticity of Dirichlet-to-Neumann maps, which was established in [90 ] (Corollary 4.7). When p ≥ 0 p \ge0 p ≥ 0 , the Dirichlet-to-Neumann operators M p ( Γ ) \mathcal{M}_p^{(\Gamma)} M p ( Γ ) form a self-adjoint holomorphic family of type (A), in the terminology used in Kato’s book [91 ] . A general analytic perturbation theory for linear elliptic operators [91 ] (Chapter 7) allows therefore to establish the analyticity of eigenvalues and eigenprojectors (under certain conditions). Alternatively, the analyticity of the eigenvalues of the Steklov problem (i.e., with Γ = ∂ Ω \Gamma=\partial\Omega Γ = ∂ Ω ) follows from a corresponding result for Robin Laplacian eigenvalues [92 ] (p. 82) and the duality between Robin and Steklov spectral problems [38 ] . An extension of this result to the Steklov-Neumann problem is straightforward. Indeed, the eigenvalues of the Laplace operator with mixed Robin-Neumann boundary conditions remain analytic in p p p near 0 0 0 because the domain H 1 ( Ω ) H^1(\Omega) H 1 ( Ω ) of the Laplace operator does not change, while the associated form is still analytic in p p p .
The analyticity near 0 0 0 justifies a Taylor expansion of the principal eigenmode of the Steklov-Neumann problem in powers of p p p as p → 0 p \to0 p → 0 :
μ 0 ( p , Γ ) = p ∣ Ω ∣ D ∣ Γ ∣ ( 1 − A Γ p D ∣ Ω ∣ + O ( p 2 ) ) , (29) \mu_0^{(p,\Gamma)}=\frac{p|\Omega|}{D|\Gamma|}\left(1-A_\Gamma\frac{p}{D}|\Omega|+O(p^2)\right), \tag*{(29)} μ 0 ( p , Γ ) = D ∣Γ∣ p ∣Ω∣ ( 1 − A Γ D p ∣Ω∣ + O ( p 2 ) ) , (29)
V 0 ( p , Γ ) ( x ) = 1 ∣ Γ ∣ ( 1 + W 0 ( Γ ) ( x ) p D ∣ Ω ∣ + O ( p 2 ) ) , (30) V_0^{(p,\Gamma)}(x)=\frac{1}{\sqrt{|\Gamma|}}\left(1+W_0^{(\Gamma)}(x)\frac{p}{D}|\Omega|+O(p^2)\right), \tag*{(30)} V 0 ( p , Γ ) ( x ) = ∣Γ∣ 1 ( 1 + W 0 ( Γ ) ( x ) D p ∣Ω∣ + O ( p 2 ) ) , (30)
where A Γ A_\Gamma A Γ and W 0 ( Γ ) ( x ) W_0^{(\Gamma)}(x) W 0 ( Γ ) ( x ) are the first-order corrections to the principal eigenvalue and eigenfunction, respectively:
A Γ = − D 2 ∣ Γ ∣ 2 ∣ Ω ∣ 2 lim p → 0 ∂ p 2 μ 0 ( p , Γ ) , (31) \mathcal{A}_{\Gamma}=-\frac{D^{2}|\Gamma|}{2|\Omega|^{2}}\lim_{p\to0}\partial_{p}^{2}\mu_{0}^{(p,\Gamma)}, \tag*{(31)} A Γ = − 2∣Ω ∣ 2 D 2 ∣Γ∣ p → 0 lim ∂ p 2 μ 0 ( p , Γ ) , (31)
W 0 ( Γ ) ( x ) = D ∣ Γ ∣ ∣ Ω ∣ lim p → 0 ∂ p V 0 ( p , Γ ) ( x ) . (32) W_{0}^{(\Gamma)}(x)=\frac{D\sqrt{|\Gamma|}}{|\Omega|}\lim_{p\to0}\partial_{p}V_{0}^{(p,\Gamma)}(x). \tag*{(32)} W 0 ( Γ ) ( x ) = ∣Ω∣ D ∣Γ∣ p → 0 lim ∂ p V 0 ( p , Γ ) ( x ) . (32)
According to Eq. (8) , the leading term in Eq. (2.21a) is determined by the first derivative of μ 0 ( p , Γ ) \mu_{0}^{(p,\Gamma)} μ 0 ( p , Γ ) , evaluated at p = 0 p=0 p = 0 :
lim p → 0 ∂ p μ 0 ( p , Γ ) = 1 D ∫ Ω d x [ V 0 ( 0 ) ] 2 = ∣ Ω ∣ D ∣ Γ ∣ . (33) \lim_{p\to0}\partial_{p}\mu_{0}^{(p,\Gamma)}
=\frac{1}{D}\int_{\Omega}dx\left[V_{0}^{(0)}\right]^{2}
=\frac{|\Omega|}{D|\Gamma|}. \tag*{(33)} p → 0 lim ∂ p μ 0 ( p , Γ ) = D 1 ∫ Ω d x [ V 0 ( 0 ) ] 2 = D ∣Γ∣ ∣Ω∣ . (33)
Since the first eigenvalue is simple, there is no ambiguity in choosing the analytic branch. We stress that the above arguments on the analyticity of eigenvalues and eigenfunctions of the mixed Steklov-Neumann problem (1.2) lack technical details and rigorous proofs. For this reason, the Taylor expansion (2.21b) is treated at a formal level, and its rigorous proof remains an open problem.
Substituting Eqs. (2.21) into Eq. (23) , we evaluate the pseudo-Green’s function as
G 0 ( x ∣ x 0 ) = D G ~ ∞ ( x , 0 ∣ x 0 ) + A Γ + W 0 ( Γ ) ( x ) + W 0 ( Γ ) ( x 0 ) + ∑ k = 1 ∞ V k ( 0 , Γ ) ( x ) V k ( 0 , Γ ) ( x 0 ) μ k ( 0 , Γ ) ( x , x 0 ∈ Ω ) . (34) G_{0}(x|x_{0})=D\widetilde{G}_{\infty}(x,0|x_{0})+\mathcal{A}_{\Gamma}+W_{0}^{(\Gamma)}(x)+W_{0}^{(\Gamma)}(x_{0})+\sum_{k=1}^{\infty}\frac{V_{k}^{(0,\Gamma)}(x)V_{k}^{(0,\Gamma)}(x_{0})}{\mu_{k}^{(0,\Gamma)}}\qquad(x,x_{0}\in\Omega). \tag*{(34)} G 0 ( x ∣ x 0 ) = D G ∞ ( x , 0∣ x 0 ) + A Γ + W 0 ( Γ ) ( x ) + W 0 ( Γ ) ( x 0 ) + k = 1 ∑ ∞ μ k ( 0 , Γ ) V k ( 0 , Γ ) ( x ) V k ( 0 , Γ ) ( x 0 ) ( x , x 0 ∈ Ω ) . (34)
In turn, its restriction to Γ × Γ \Gamma\times\Gamma Γ × Γ eliminates the first term and gives
G 0 ( x ∣ x 0 ) = A Γ + w 0 ( Γ ) ( x ) + w 0 ( Γ ) ( x 0 ) + ∑ k = 1 ∞ V k ( 0 , Γ ) ( x ) V k ( 0 , Γ ) ( x 0 ) μ k ( 0 , Γ ) ( x , x 0 ∈ Γ ) , (35) G_{0}(x|x_{0})=\mathcal{A}_{\Gamma}+w_{0}^{(\Gamma)}(x)+w_{0}^{(\Gamma)}(x_{0})+\sum_{k=1}^{\infty}\frac{V_{k}^{(0,\Gamma)}(x)V_{k}^{(0,\Gamma)}(x_{0})}{\mu_{k}^{(0,\Gamma)}}\qquad(x,x_{0}\in\Gamma), \tag*{(35)} G 0 ( x ∣ x 0 ) = A Γ + w 0 ( Γ ) ( x ) + w 0 ( Γ ) ( x 0 ) + k = 1 ∑ ∞ μ k ( 0 , Γ ) V k ( 0 , Γ ) ( x ) V k ( 0 , Γ ) ( x 0 ) ( x , x 0 ∈ Γ ) , (35)
where w 0 ( Γ ) w_{0}^{(\Gamma)} w 0 ( Γ ) is the restriction of W 0 ( Γ ) W_{0}^{(\Gamma)} W 0 ( Γ ) to Γ \Gamma Γ . Integrating this expression over x ∈ Γ x\in\Gamma x ∈ Γ yields
∫ Γ d x G 0 ( x ∣ x 0 ) = ∣ Γ ∣ ( w 0 ( Γ ) ( x 0 ) + A Γ ) ( x 0 ∈ Γ ) , (36) \int_{\Gamma}dx\,G_{0}(x|x_{0})=|\Gamma|\left(w_{0}^{(\Gamma)}(x_{0})+\mathcal{A}_{\Gamma}\right)\qquad(x_{0}\in\Gamma), \tag*{(36)} ∫ Γ d x G 0 ( x ∣ x 0 ) = ∣Γ∣ ( w 0 ( Γ ) ( x 0 ) + A Γ ) ( x 0 ∈ Γ ) , (36)
where we used the orthogonality of w 0 ( Γ ) w_{0}^{(\Gamma)} w 0 ( Γ ) to a constant,
∫ Γ w 0 ( Γ ) = 0 , (37) \int_{\Gamma}w_{0}^{(\Gamma)}=0, \tag*{(37)} ∫ Γ w 0 ( Γ ) = 0 , (37)
which follows from the normalization of the eigenfunction v 0 ( p , Γ ) v_{0}^{(p,\Gamma)} v 0 ( p , Γ ) :
1 = ∫ Γ ∣ v 0 ( p , Γ ) ∣ 2 = ∫ Γ ∣ v 0 ( 0 , Γ ) ∣ 2 + p 2 ∣ Ω ∣ D ∣ Γ ∣ ∫ Γ w 0 ( Γ ) + O ( p 2 ) . (38) 1=\int_{\Gamma}\left|v_{0}^{(p,\Gamma)}\right|^{2}=\int_{\Gamma}\left|v_{0}^{(0,\Gamma)}\right|^{2}+p\frac{2|\Omega|}{D|\Gamma|}\int_{\Gamma}w_{0}^{(\Gamma)}+O(p^{2}). \tag*{(38)} 1 = ∫ Γ v 0 ( p , Γ ) 2 = ∫ Γ v 0 ( 0 , Γ ) 2 + p D ∣Γ∣ 2∣Ω∣ ∫ Γ w 0 ( Γ ) + O ( p 2 ) . (38)
The integral of Eq. (36) over x 0 ∈ Γ x_0 \in\Gamma x 0 ∈ Γ determines the constant A Γ A_\Gamma A Γ ,
A Γ = 1 ∣ Γ ∣ 2 ∫ Γ d x 0 ∫ Γ d x G 0 ( x ∣ x 0 ) , (39) A_\Gamma= \frac{1}{|\Gamma|^2} \int_\Gamma dx_0 \int_\Gamma dx\,G_0(x|x_0), \tag*{(39)} A Γ = ∣Γ ∣ 2 1 ∫ Γ d x 0 ∫ Γ d x G 0 ( x ∣ x 0 ) , (39)
from which the function w 0 ( Γ ) ( x 0 ) w_0^{(\Gamma)}(x_0) w 0 ( Γ ) ( x 0 ) follows as
w 0 ( Γ ) ( x 0 ) = − A Γ + 1 ∣ Γ ∣ ∫ Γ d x G 0 ( x ∣ x 0 ) ( x 0 ∈ Γ ) . (40) w_0^{(\Gamma)}(x_0) = -A_\Gamma+ \frac{1}{|\Gamma|} \int_\Gamma dx\,G_0(x|x_0) \qquad(x_0 \in\Gamma). \tag*{(40)} w 0 ( Γ ) ( x 0 ) = − A Γ + ∣Γ∣ 1 ∫ Γ d x G 0 ( x ∣ x 0 ) ( x 0 ∈ Γ ) . (40)
Combining these results, we define the integral kernel:
G ( x ∣ x 0 ) = G 0 ( x ∣ x 0 ) − ( w 0 ( Γ ) ( x ) + w 0 ( Γ ) ( x 0 ) + A Γ ) , (41) G(x|x_0) = G_0(x|x_0) - \left(w_0^{(\Gamma)}(x) + w_0^{(\Gamma)}(x_0) + A_\Gamma\right), \tag*{(41)} G ( x ∣ x 0 ) = G 0 ( x ∣ x 0 ) − ( w 0 ( Γ ) ( x ) + w 0 ( Γ ) ( x 0 ) + A Γ ) , (41)
which admits the following expansion according to Eq. (35) :
G ( x ∣ x 0 ) = ∑ k = 1 ∞ v k ( 0 , Γ ) ( x ) v k ( 0 , Γ ) ( x 0 ) μ k ( 0 , Γ ) ( x , x 0 ∈ Γ ) . (42) G(x|x_0) = \sum_{k=1}^{\infty} \frac{v_k^{(0,\Gamma)}(x)v_k^{(0,\Gamma)}(x_0)}{\mu_k^{(0,\Gamma)}} \qquad(x,x_0 \in\Gamma). \tag*{(42)} G ( x ∣ x 0 ) = k = 1 ∑ ∞ μ k ( 0 , Γ ) v k ( 0 , Γ ) ( x ) v k ( 0 , Γ ) ( x 0 ) ( x , x 0 ∈ Γ ) . (42)
Once the kernel G ( x ∣ x 0 ) G(x|x_0) G ( x ∣ x 0 ) is found, one can search for the eigenvalues and eigenfunctions of the associated integral operator:
∫ Γ d x G ( x ∣ x 0 ) v k ( 0 , Γ ) ( x ) = 1 μ k ( 0 , Γ ) v k ( 0 , Γ ) ( x 0 ) ( x 0 ∈ Γ , k ≥ 1 ) , (43) \int_\Gamma dx\,G(x|x_0)v_k^{(0,\Gamma)}(x) = \frac{1}{\mu_k^{(0,\Gamma)}}v_k^{(0,\Gamma)}(x_0) \qquad(x_0 \in\Gamma,\ k \ge1), \tag*{(43)} ∫ Γ d x G ( x ∣ x 0 ) v k ( 0 , Γ ) ( x ) = μ k ( 0 , Γ ) 1 v k ( 0 , Γ ) ( x 0 ) ( x 0 ∈ Γ , k ≥ 1 ) , (43)
and thus determine the spectral properties of the Dirichlet-to-Neumann operator M 0 ( Γ ) \mathcal{M}_0^{(\Gamma)} M 0 ( Γ ) , cf. Eq. (9) . As the constant eigenfunction v 0 ( 0 , Γ ) v_0^{(0,\Gamma)} v 0 ( 0 , Γ ) is excluded from the sum in Eq. (42) , one has
∫ Γ d x G ( x ∣ x 0 ) = 0 ( x 0 ∈ Ω ) . (44) \int_\Gamma dx\,G(x|x_0) = 0 \qquad(x_0 \in\Omega). \tag*{(44)} ∫ Γ d x G ( x ∣ x 0 ) = 0 ( x 0 ∈ Ω ) . (44)
This property ensures that the eigenfunctions of the above integral operator are orthogonal to a constant, as it should be.
As previously, once the eigenfunction v k ( 0 , Γ ) v_k^{(0,\Gamma)} v k ( 0 , Γ ) is found on Γ \Gamma Γ , it can be extended into Ω \Omega Ω . As V 0 ( 0 , Γ ) = v 0 ( 0 , Γ ) = 1 / ∣ Γ ∣ V_0^{(0,\Gamma)} = v_0^{(0,\Gamma)} = 1/\sqrt{|\Gamma|} V 0 ( 0 , Γ ) = v 0 ( 0 , Γ ) = 1/ ∣Γ∣ is known, we focus on k > 0 k > 0 k > 0 . For this extension, one can multiply Eq. (1.2a) with p = 0 p=0 p = 0 by G 0 ( x ∣ x 0 ) G_0(x|x_0) G 0 ( x ∣ x 0 ) , multiply Eq. (2.18a) by V k ( 0 , Γ ) ( x ) V_k^{(0,\Gamma)}(x) V k ( 0 , Γ ) ( x ) , subtract them, integrate over x ∈ Ω x \in\Omega x ∈ Ω , and use the Green’s formula and the boundary conditions to get
V k ( 0 , Γ ) ( x 0 ) = b k ( Γ ) + μ k ( 0 , Γ ) ∫ Γ d x G 0 ( x ∣ x 0 ) v k ( 0 , Γ ) ( x ) ( x 0 ∈ Ω ) , (45) V_k^{(0,\Gamma)}(x_0) = b_k^{(\Gamma)} + \mu_k^{(0,\Gamma)} \int_\Gamma dx\,G_0(x|x_0)v_k^{(0,\Gamma)}(x) \qquad(x_0 \in\Omega), \tag*{(45)} V k ( 0 , Γ ) ( x 0 ) = b k ( Γ ) + μ k ( 0 , Γ ) ∫ Γ d x G 0 ( x ∣ x 0 ) v k ( 0 , Γ ) ( x ) ( x 0 ∈ Ω ) , (45)
where
b k ( Γ ) = 1 ∣ Ω ∣ ∫ Ω d x V k ( 0 , Γ ) ( x ) ( k = 1 , 2 , … ) . (46) b_k^{(\Gamma)} = \frac{1}{|\Omega|} \int_\Omega dx\,V_k^{(0,\Gamma)}(x) \qquad(k = 1,2,\ldots). \tag*{(46)} b k ( Γ ) = ∣Ω∣ 1 ∫ Ω d x V k ( 0 , Γ ) ( x ) ( k = 1 , 2 , … ) . (46)
Expectedly, the pseudo-Green’s function G 0 ( x ∣ x 0 ) G_0(x|x_0) G 0 ( x ∣ x 0 ) determines V k ( 0 , Γ ) V_k^{(0,\Gamma)} V k ( 0 , Γ ) up to an additive constant b k ( Γ ) b_k^{(\Gamma)} b k ( Γ ) , which can be fixed from the values of v k ( 0 , Γ ) v_k^{(0,\Gamma)} v k ( 0 , Γ ) on Γ \Gamma Γ . For this purpose, we first substitute the expansion (2.21b) into the boundary value problem (1.2) to show that the first-order correction W 0 ( Γ ) ( x ) W_0^{(\Gamma)}(x) W 0 ( Γ ) ( x ) of the principal eigenfunction satisfies
Δ W 0 ( Γ ) ( x ) = 1 ∣ Ω ∣ , ( x ∈ Ω ) , (47) \Delta W_0^{(\Gamma)}(x)=\frac{1}{|\Omega|}, \qquad(x\in\Omega), \tag*{(47)} Δ W 0 ( Γ ) ( x ) = ∣Ω∣ 1 , ( x ∈ Ω ) , (47)
∂ n W 0 ( Γ ) ( x ) = I Γ ( x ) ∣ Γ ∣ , ( x ∈ ∂ Ω ) , (48) \partial_n W_0^{(\Gamma)}(x)=\frac{\mathbb{I}_{\Gamma}(x)}{|\Gamma|}, \qquad(x\in\partial\Omega), \tag*{(48)} ∂ n W 0 ( Γ ) ( x ) = ∣Γ∣ I Γ ( x ) , ( x ∈ ∂ Ω ) , (48)
where I Γ ( x ) \mathbb{I}_{\Gamma}(x) I Γ ( x ) is the indicator function of Γ \Gamma Γ , which is equal to 1 for x ∈ Γ x\in\Gamma x ∈ Γ and 0 otherwise. The solution of this problem is determined up to an additive constant, which can be fixed by using (37) . Equivalently, substituting eq:2.21 into both sides of the identity (2.3) at k = 0 k=0 k = 0 , we get
∫ Ω ∣ V 0 ( p , Γ ) ∣ 2 = ∣ Ω ∣ ∣ Γ ∣ + p 2 ∣ Ω ∣ D ∣ Γ ∣ ∫ Ω W 0 ( Γ ) + O ( p 2 ) = ∣ Ω ∣ ∣ Γ ∣ − 2 A Γ ∣ Ω ∣ 2 D ∣ Γ ∣ p + O ( p 2 ) = D ∂ p μ 0 ( p , Γ ) , (49) \begin{aligned}
\int_{\Omega}\left|V_0^{(p,\Gamma)}\right|^2
&=\frac{|\Omega|}{|\Gamma|}+p\frac{2|\Omega|}{D|\Gamma|}\int_{\Omega}W_0^{(\Gamma)}+O(p^2)\\
&=\frac{|\Omega|}{|\Gamma|}-2A_{\Gamma}\frac{|\Omega|^2}{D|\Gamma|}p+O(p^2)=D\partial_p\mu_0^{(p,\Gamma)}, \tag*{(49)}
\end{aligned} ∫ Ω V 0 ( p , Γ ) 2 = ∣Γ∣ ∣Ω∣ + p D ∣Γ∣ 2∣Ω∣ ∫ Ω W 0 ( Γ ) + O ( p 2 ) = ∣Γ∣ ∣Ω∣ − 2 A Γ D ∣Γ∣ ∣Ω ∣ 2 p + O ( p 2 ) = D ∂ p μ 0 ( p , Γ ) , (49)
from which
1 ∣ Ω ∣ ∫ Ω W 0 ( Γ ) = − A Γ , (50) \frac{1}{|\Omega|}\int_{\Omega}W_0^{(\Gamma)}=-A_{\Gamma}, \tag*{(50)} ∣Ω∣ 1 ∫ Ω W 0 ( Γ ) = − A Γ , (50)
that fixes the additive constant.
In order to fix b k ( Γ ) b_k^{(\Gamma)} b k ( Γ ) , one can multiply (2) with p = 0 p=0 p = 0 by W 0 ( Γ ) ( x ) W_0^{(\Gamma)}(x) W 0 ( Γ ) ( x ) , multiply (47) by V k ( 0 , Γ ) ( x ) V_k^{(0,\Gamma)}(x) V k ( 0 , Γ ) ( x ) , subtract them, integrate over x ∈ Ω x\in\Omega x ∈ Ω , and use the Green’s formula and the boundary conditions:
b k ( Γ ) = − 1 μ k ( 0 , Γ ) ∫ Γ d x w 0 ( Γ ) ( x ) v k ( 0 , Γ ) ( x ) ( k = 1 , 2 , … ) . (51) b_k^{(\Gamma)}=-\frac{1}{\mu_k^{(0,\Gamma)}}\int_{\Gamma}dx\,w_0^{(\Gamma)}(x)v_k^{(0,\Gamma)}(x)\qquad(k=1,2,\ldots). \tag*{(51)} b k ( Γ ) = − μ k ( 0 , Γ ) 1 ∫ Γ d x w 0 ( Γ ) ( x ) v k ( 0 , Γ ) ( x ) ( k = 1 , 2 , … ) . (51)
Note also that the integral of (2) with p = 0 p=0 p = 0 over x ∈ Ω x\in\Omega x ∈ Ω yields the identity
1 ∣ Ω ∣ ∫ Ω d x V k ( p , Γ ) = D μ k ( p , Γ ) p ∣ Ω ∣ ∫ Γ d x v k ( p , Γ ) ( p > 0 , k = 0 , 1 , … ) . (52) \frac{1}{|\Omega|}\int_{\Omega}dx\,V_k^{(p,\Gamma)}
=\frac{D\mu_k^{(p,\Gamma)}}{p|\Omega|}\int_{\Gamma}dx\,v_k^{(p,\Gamma)}
\qquad(p>0,\ k=0,1,\ldots). \tag*{(52)} ∣Ω∣ 1 ∫ Ω d x V k ( p , Γ ) = p ∣Ω∣ D μ k ( p , Γ ) ∫ Γ d x v k ( p , Γ ) ( p > 0 , k = 0 , 1 , … ) . (52)
In the limit p → 0 p\to0 p → 0 , the left-hand side approaches b k ( Γ ) b_k^{(\Gamma)} b k ( Γ ) , yielding an alternative interpretation of b k ( Γ ) b_k^{(\Gamma)} b k ( Γ ) as
b k ( Γ ) = D ∣ Ω ∣ lim p → 0 μ k ( p , Γ ) p ∫ Γ v k ( p , Γ ) ( k = 1 , 2 , … ) (53) b_k^{(\Gamma)}=\frac{D}{|\Omega|}\lim_{p\to0}\frac{\mu_k^{(p,\Gamma)}}{p}\int_{\Gamma}v_k^{(p,\Gamma)}
\qquad(k=1,2,\ldots) \tag*{(53)} b k ( Γ ) = ∣Ω∣ D p → 0 lim p μ k ( p , Γ ) ∫ Γ v k ( p , Γ ) ( k = 1 , 2 , … ) (53)
(when the associated eigenvalue μ k ( 0 , Γ ) \mu_k^{(0,\Gamma)} μ k ( 0 , Γ ) is degenerate, one needs to keep following the k k k -th analytic branch).
Role of the first-order corrections
According to Eq. (2.31), the kernel G ( x ∣ x 0 ) G(x|x_0) G ( x ∣ x 0 ) is determined by the pseudo-Green’s function G 0 ( x ∣ x 0 ) G_0(x|x_0) G 0 ( x ∣ x 0 ) , which was thoroughly investigated in the mathematical and physical literature (see [65 , 75 , 93 , 94 ] and references therein). However, Eq. (2.31) includes two other contributions, namely, the first-order corrections A Γ A_{\Gamma} A Γ and w 0 ( Γ ) w_{0}^{(\Gamma)} w 0 ( Γ ) to the principal eigenvalue and eigenfunction of M p ( Γ ) \mathcal{M}_{p}^{(\Gamma)} M p ( Γ ) as p → 0 p \to0 p → 0 . In other words, our construction of the eigenmodes of the operator M 0 ( Γ ) \mathcal{M}_{0}^{(\Gamma)} M 0 ( Γ ) is not limited to the case p = 0 p=0 p = 0 but also requires the knowledge of spectral properties of M p ( Γ ) \mathcal{M}_{p}^{(\Gamma)} M p ( Γ ) in the vicinity of p ≈ 0 p \approx0 p ≈ 0 . As most former works dealt directly with the conventional problem at p = 0 p=0 p = 0 , the role of the first-order corrections A Γ A_{\Gamma} A Γ and w 0 ( Γ ) w_{0}^{(\Gamma)} w 0 ( Γ ) seems to be overlooked. In this section, we briefly discuss their properties.
The correction w 0 ( Γ ) w_{0}^{(\Gamma)} w 0 ( Γ ) was formally introduced in Eq. (2.22b) via the derivative ∂ p v 0 ( p , Γ ) \partial_{p}v_{0}^{(p,\Gamma)} ∂ p v 0 ( p , Γ ) at p = 0 p=0 p = 0 but it could also be obtained by solving the boundary value problem (2.37), with the condition (2.27). Curiously, the function W 0 ( Γ ) W_{0}^{(\Gamma)} W 0 ( Γ ) turns out to be related to the constant-flux approximation T q ( a p p ) ( x 0 ) T_{q}^{(\mathrm{app})}(x_{0}) T q ( app ) ( x 0 ) of the MFRT T q ( x 0 ) T_{q}(x_{0}) T q ( x 0 ) [83 , 84 , 95 , 96 , 97 ] , which can be written as
T q ( a p p ) ( x 0 ) = ∣ Ω ∣ D ( 1 q ∣ Γ ∣ − W 0 ( Γ ) ( x 0 ) ) ( q > 0 , x 0 ∈ Ω ‾ ) . (54) T_{q}^{(\mathrm{app})}(x_{0})=\frac{|\Omega|}{D}\left(\frac{1}{q|\Gamma|}-W_{0}^{(\Gamma)}(x_{0})\right)\qquad(q>0,\ x_{0}\in\overline{\Omega}). \tag*{(54)} T q ( app ) ( x 0 ) = D ∣Ω∣ ( q ∣Γ∣ 1 − W 0 ( Γ ) ( x 0 ) ) ( q > 0 , x 0 ∈ Ω ) . (54)
One can easily check that this function satisfies for any q > 0 q>0 q > 0 :
D Δ T q ( a p p ) ( x 0 ) = − 1 ( x 0 ∈ Ω ) , (55) D\Delta T_{q}^{(\mathrm{app})}(x_{0})=-1\qquad(x_{0}\in\Omega), \tag*{(55)} D Δ T q ( app ) ( x 0 ) = − 1 ( x 0 ∈ Ω ) , (55)
− ∂ n T q ( a p p ) = ∣ Ω ∣ D ∣ Γ ∣ I Γ ( x 0 ) ( x 0 ∈ ∂ Ω ) , (56) -\partial_{n}T_{q}^{(\mathrm{app})}=\frac{|\Omega|}{D|\Gamma|}I_{\Gamma}(x_{0})\qquad(x_{0}\in\partial\Omega), \tag*{(56)} − ∂ n T q ( app ) = D ∣Γ∣ ∣Ω∣ I Γ ( x 0 ) ( x 0 ∈ ∂ Ω ) , (56)
q D ∫ Γ d x 0 T q ( a p p ) = ∣ Ω ∣ . (57) qD\int_{\Gamma}\mathrm{d}x_{0}\,T_{q}^{(\mathrm{app})}=|\Omega|. \tag*{(57)} q D ∫ Γ d x 0 T q ( app ) = ∣Ω∣. (57)
In turn, the MFRT T q ( x 0 ) T_{q}(x_{0}) T q ( x 0 ) satisfies (see § 5 for more details)
D Δ T q ( x 0 ) = − 1 ( x 0 ∈ Ω ) , (58) D\Delta T_{q}(x_{0})=-1\qquad(x_{0}\in\Omega), \tag*{(58)} D Δ T q ( x 0 ) = − 1 ( x 0 ∈ Ω ) , (58)
− ∂ n T q = q T q ( x 0 ) I Γ ( x 0 ) ( x 0 ∈ ∂ Ω ) . (59) -\partial_{n}T_{q}=qT_{q}(x_{0})I_{\Gamma}(x_{0})\qquad(x_{0}\in\partial\Omega). \tag*{(59)} − ∂ n T q = q T q ( x 0 ) I Γ ( x 0 ) ( x 0 ∈ ∂ Ω ) . (59)
One sees that the constant-flux approximation consists in replacing the Robin boundary condition on the subset Γ \Gamma Γ , − ∂ n T q = q T q -\partial_{n}T_{q}=qT_{q} − ∂ n T q = q T q , by the constant-flux condition − ∂ n T q ( a p p ) = ∣ Ω ∣ / ( D ∣ Γ ∣ ) -\partial_{n}T_{q}^{(\mathrm{app})}=|\Omega|/(D|\Gamma|) − ∂ n T q ( app ) = ∣Ω∣/ ( D ∣Γ∣ ) . In addition, the self-consistent closure relation (2.44c) is imposed to fix the constant and thus to get the unique solution. We return to the approximation (2.43) in § 5 .
We also highlight the versatile roles of the constant A Γ A_{\Gamma} A Γ emerging in different contexts. This constant was introduced in Eq. (2.22a) via the second derivative of the principal eigenvalue μ 0 ( p , Γ ) \mu_{0}^{(p,\Gamma)} μ 0 ( p , Γ ) , evaluated at p = 0 p=0 p = 0 . In turn, Eq. (2.29) gives its representation as the double integral of the pseudo-Green’s function G 0 ( x ∣ x 0 ) G_{0}(x|x_{0}) G 0 ( x ∣ x 0 ) over the subset Γ \Gamma Γ . For the conventional Steklov problem (with Γ = ∂ Ω \Gamma=\partial\Omega Γ = ∂ Ω ), this relation was earlier recognized in [17 ] . In this reference, the constant A Γ A_{\Gamma} A Γ was shown to determine the long-time behavior of the variance of the boundary local time, which is the proxy of the number of encounters between a particle and the boundary. In Appendix D, we extend these former results to our setting of a given subset Γ \Gamma Γ and get the long-time asymptotic behavior of the variance of the boundary local time ℓ t \ell_{t} ℓ t on Γ \Gamma Γ :
Var { ℓ t } ≈ A Γ 2 ∣ Γ ∣ 2 ∣ Ω ∣ D t + O ( 1 ) ( t → ∞ ) . (60) \operatorname{Var}\{\ell_{t}\}\approx A_{\Gamma}\frac{2|\Gamma|^{2}}{|\Omega|}Dt+O(1)\qquad(t\to\infty). \tag*{(60)} Var { ℓ t } ≈ A Γ ∣Ω∣ 2∣Γ ∣ 2 D t + O ( 1 ) ( t → ∞ ) . (60)
Finally, we will show in § 5 that A Γ A_{\Gamma} A Γ also controls the asymptotic behavior of the MFPT to Γ \Gamma Γ in the small-target limit. In this limit, the correction A Γ A_{\Gamma} A Γ admits a universal asymptotic behavior (see § 3 ).
We conclude this section by noting that Dittmar used the Neumann function N ( x ∣ x 0 ) \mathcal{N}(x|x_{0}) N ( x ∣ x 0 ) to analyze the Steklov eigenfunctions for the conventional case p = 0 p = 0 p = 0 and Γ = ∂ Ω \Gamma= \partial\Omega Γ = ∂ Ω in planar domains [98 ] . For any x 0 ∈ Ω \mathbf{x}_{0} \in\Omega x 0 ∈ Ω , this function is defined as the unique solution of the boundary value problem:
− Δ N ( x ∣ x 0 ) = δ ( x − x 0 ) ( x ∈ Ω ) , (61) -\Delta\mathcal{N}(\mathbf{x}|\mathbf{x}_{0}) = \delta(\mathbf{x}-\mathbf{x}_{0}) \qquad(\mathbf{x} \in\Omega), \tag*{(61)} − Δ N ( x ∣ x 0 ) = δ ( x − x 0 ) ( x ∈ Ω ) , (61)
∂ n N = − 1 ∣ ∂ Ω ∣ ( x ∈ ∂ Ω ) , (62) \partial_{n}\mathcal{N} = -\frac{1}{|\partial\Omega|} \qquad(\mathbf{x} \in\partial\Omega), \tag*{(62)} ∂ n N = − ∣ ∂ Ω∣ 1 ( x ∈ ∂ Ω ) , (62)
∫ ∂ Ω d x N ( x ∣ x 0 ) = 0 , (63) \int_{\partial\Omega} \mathrm{d}x\,\mathcal{N}(\mathbf{x}|\mathbf{x}_{0}) = 0, \tag*{(63)} ∫ ∂ Ω d x N ( x ∣ x 0 ) = 0 , (63)
and thus differs from the pseudo-Green’s function G 0 ( x ∣ x 0 ) \mathcal{G}_{0}(\mathbf{x}|\mathbf{x}_{0}) G 0 ( x ∣ x 0 ) satisfying Eqs. (2.18, (27) . Dittmar showed that the restriction of N ( x ∣ x 0 ) \mathcal{N}(\mathbf{x}|\mathbf{x}_{0}) N ( x ∣ x 0 ) to ∂ Ω × ∂ Ω \partial\Omega\times\partial\Omega ∂ Ω × ∂ Ω reads
N ( x ∣ x 0 ) = ∑ k = 1 ∞ ν k ( 0 , ∂ Ω ) ( x ) ν k ( 0 , ∂ Ω ) ( x 0 ) μ k ( 0 , ∂ Ω ) ( x , x 0 ∈ ∂ Ω ) , (64) \mathcal{N}(\mathbf{x}|\mathbf{x}_{0}) = \sum_{k=1}^{\infty}\frac{\nu_{k}^{(0,\partial\Omega)}(\mathbf{x})\nu_{k}^{(0,\partial\Omega)}(\mathbf{x}_{0})}{\mu_{k}^{(0,\partial\Omega)}} \qquad(\mathbf{x},\mathbf{x}_{0} \in\partial\Omega), \tag*{(64)} N ( x ∣ x 0 ) = k = 1 ∑ ∞ μ k ( 0 , ∂ Ω ) ν k ( 0 , ∂ Ω ) ( x ) ν k ( 0 , ∂ Ω ) ( x 0 ) ( x , x 0 ∈ ∂ Ω ) , (64)
and therefore allows one to compute the eigenvalues and eigenfunctions of the Dirichlet-to-Neumann operator M ( ∂ Ω ) \mathcal{M}^{(\partial\Omega)} M ( ∂ Ω ) on the boundary ∂ Ω \partial\Omega ∂ Ω . In this setting, the restriction of N ( x ∣ x 0 ) \mathcal{N}(\mathbf{x}|\mathbf{x}_{0}) N ( x ∣ x 0 ) to the boundary ∂ Ω \partial\Omega ∂ Ω is precisely our kernel G ( x ∣ x 0 ) \mathcal{G}(\mathbf{x}|\mathbf{x}_{0}) G ( x ∣ x 0 ) defined in Eq. (42) . An explicit form of N ( x ∣ x 0 ) \mathcal{N}(\mathbf{x}|\mathbf{x}_{0}) N ( x ∣ x 0 ) for the interior and exterior of a sphere is well known [99 ] (see also [100 ] for rectangular domains), whereas some general properties of the Neumann function and its applications in electrical impedance imaging were discussed in [101 ] .
The Dittmar’s approach can be extended to deal with the mixed Steklov-Neumann problem by modifying the definition of the Neumann function, namely, by replacing 1 / ∣ ∂ Ω ∣ 1/|\partial\Omega| 1/∣ ∂ Ω∣ by I Γ ( x ) / ∣ Γ ∣ \mathbb{I}_{\Gamma}(\mathbf{x})/|\Gamma| I Γ ( x ) /∣Γ∣ in the Neumann boundary condition (2.47b). One can easily get the spectral expansion of the modified function, denoted as N ( Γ ) ( x ∣ x 0 ) \mathcal{N}^{(\Gamma)}(\mathbf{x}|\mathbf{x}_{0}) N ( Γ ) ( x ∣ x 0 ) :
N ( Γ ) ( x ∣ x 0 ) = D G ~ ∞ ( x , 0 ∣ x 0 ) + ∑ k = 1 ∞ ν k ( 0 , Γ ) ( x ) ν k ( 0 , Γ ) ( x 0 ) μ k ( 0 , Γ ) ( x , x 0 ∈ Ω ) , (65) \mathcal{N}^{(\Gamma)}(\mathbf{x}|\mathbf{x}_{0}) = D\widetilde{\mathcal{G}}_{\infty}(\mathbf{x},0|\mathbf{x}_{0}) + \sum_{k=1}^{\infty}\frac{\nu_{k}^{(0,\Gamma)}(\mathbf{x})\nu_{k}^{(0,\Gamma)}(\mathbf{x}_{0})}{\mu_{k}^{(0,\Gamma)}} \qquad(\mathbf{x},\mathbf{x}_{0} \in\Omega), \tag*{(65)} N ( Γ ) ( x ∣ x 0 ) = D G ∞ ( x , 0∣ x 0 ) + k = 1 ∑ ∞ μ k ( 0 , Γ ) ν k ( 0 , Γ ) ( x ) ν k ( 0 , Γ ) ( x 0 ) ( x , x 0 ∈ Ω ) , (65)
so that its restriction to Γ × Γ \Gamma\times\Gamma Γ × Γ becomes again identical to the kernel G ( x ∣ x 0 ) \mathcal{G}(\mathbf{x}|\mathbf{x}_{0}) G ( x ∣ x 0 ) and thus allows the construction of the eigenmodes of the Dirichlet-to-Neumann operator M 0 ( Γ ) \mathcal{M}_{0}^{(\Gamma)} M 0 ( Γ ) . However, the major drawback of this construction is that the modified Neumann function N ( Γ ) ( x ∣ x 0 ) \mathcal{N}^{(\Gamma)}(\mathbf{x}|\mathbf{x}_{0}) N ( Γ ) ( x ∣ x 0 ) depends on Γ \Gamma Γ , whereas the pseudo-Green’s function G 0 ( x ∣ x 0 ) \mathcal{G}_{0}(\mathbf{x}|\mathbf{x}_{0}) G 0 ( x ∣ x 0 ) that we used, does not. In summary, both the modified Neumann function and the pseudo-Green’s function can be employed but the latter is independent of Γ \Gamma Γ and thus simpler.
Asymptotic results
In this section, we present our main asymptotic results in the small-target limit Γ → ∅ \Gamma\to\varnothing Γ → ∅ . In Sec. 3.1, we employ a dilation argument to reduce the original Steklov-Neumann problem (1.2) for a small subset Γ \Gamma Γ to an auxiliary Steklov-Neumann problem, either for an interval ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) in the half-plane (d = 2 d = 2 d = 2 ), or for the unit disk in the half-space (d = 3 d = 3 d = 3 ). We recall the spectral properties of these two auxiliary problems. In Sec. 3.2, we reformulate these two problems in terms of integral operators and compute explicitly their kernels. We also describe the asymptotic behavior of the corrections A Γ \mathcal{A}_{\Gamma} A Γ and w 0 ( Γ ) ( x ) w_{0}^{(\Gamma)}(\mathbf{x}) w 0 ( Γ ) ( x ) and of the coefficients b k ( Γ ) b_{k}^{(\Gamma)} b k ( Γ ) . The role of boundary connectivity is discussed in Sec. 3.3.
Dilation argument and two auxiliary Steklov-Neumann problems
In two dimensions, we consider Γ ⊂ ∂ Ω \Gamma\subset\partial\Omega Γ ⊂ ∂ Ω to be a connected subset of perimeter 2 ϵ 2\epsilon 2 ϵ (i.e., a smooth curve of arbitrary shape). In three dimensions, we focus on the particular case when the subset Γ \Gamma Γ is a (curved) disk of radius ϵ \epsilon ϵ (an extension to other shapes of Γ \Gamma Γ will be mentioned below). We are interested in the limit ϵ → 0 \epsilon\to0 ϵ → 0 ; in practice, the derived results are applicable when the size ϵ \epsilon ϵ of the subset Γ \Gamma Γ is the smallest length of the problem. In particular, we require that
ϵ ≪ 1 max x ∈ ∂ Ω { H ( x ) } , (66) \epsilon\ll\frac{1}{\max_{x \in\partial\Omega}\{H(x)\}}, \tag*{(66)} ϵ ≪ max x ∈ ∂ Ω { H ( x )} 1 , (66)
where H ( x ) H(x) H ( x ) is the mean curvature of the boundary at the point x ∈ ∂ Ω x \in\partial\Omega x ∈ ∂ Ω . Moreover, the reflecting part of the boundary, ∂ Ω N \partial\Omega_N ∂ Ω N , should not appear close to the center of the subset Γ \Gamma Γ (more formally, we require that the radius of the largest ball that is inscribed into Ω \Omega Ω and touches any point of Γ \Gamma Γ , is much larger than ϵ \epsilon ϵ ).
Denoting by x Γ ∈ ∂ Ω x_{\Gamma} \in\partial\Omega x Γ ∈ ∂ Ω the center of the subset Γ \Gamma Γ , we rescale the coordinates as x → x ^ = ( x − x Γ ) / ϵ \mathbf{x} \to\hat{\mathbf{x}} = (\mathbf{x} - \mathbf{x}_{\Gamma})/\epsilon x → x ^ = ( x − x Γ ) / ϵ that dilates Ω \Omega Ω , ∂ Ω \partial\Omega ∂ Ω , and Γ \Gamma Γ into Ω ^ \hat{\Omega} Ω ^ , ∂ Ω ^ \hat{\partial\Omega} ∂ Ω ^ , and Γ ^ \hat{\Gamma} Γ ^ , and transforms the original Steklov-Neumann problem (1.2) to
( ϵ 2 p / D − Δ ) V k ( p , Γ ) ( x ^ ) = 0 ( x ^ ∈ Ω ^ ) , (67) (\epsilon^2p/D-\Delta)V_k^{(p,\Gamma)}(\hat{\mathbf{x}})=0 \quad(\hat{\mathbf{x}}\in\hat{\Omega}), \tag*{(67)} ( ϵ 2 p / D − Δ ) V k ( p , Γ ) ( x ^ ) = 0 ( x ^ ∈ Ω ^ ) , (67)
∂ n V k ( p , Γ ) ( x ^ ) = ϵ μ k ( p , Γ ) V k ( p , Γ ) ( x ^ ) ( x ^ ∈ Γ ^ ) , (68) \partial_n V_k^{(p,\Gamma)}(\hat{\mathbf{x}})=\epsilon\mu_k^{(p,\Gamma)}V_k^{(p,\Gamma)}(\hat{\mathbf{x}}) \quad(\hat{\mathbf{x}}\in\hat{\Gamma}), \tag*{(68)} ∂ n V k ( p , Γ ) ( x ^ ) = ϵ μ k ( p , Γ ) V k ( p , Γ ) ( x ^ ) ( x ^ ∈ Γ ^ ) , (68)
∂ n V k ( p , Γ ) ( x ^ ) = 0 ( x ^ ∈ ∂ Ω ^ \ Γ ^ ) . (69) \partial_n V_k^{(p,\Gamma)}(\hat{\mathbf{x}})=0 \quad(\hat{\mathbf{x}}\in\partial\hat{\Omega}\backslash\hat{\Gamma}). \tag*{(69)} ∂ n V k ( p , Γ ) ( x ^ ) = 0 ( x ^ ∈ ∂ Ω ^ \ Γ ^ ) . (69)
In the limit ϵ → 0 \epsilon\to0 ϵ → 0 , the domain Ω ^ \hat{\Omega} Ω ^ blows up and can be substituted by the half-plane or the half-space, whereas ∂ Ω ^ \hat{\partial\Omega} ∂ Ω ^ is substituted by the horizontal axis or the horizontal plane (an additional rotation may be needed to choose a suitable orientation of the coordinates). In turn, the size of the rescaled subset Γ ^ \hat{\Gamma} Γ ^ remains unchanged, but it becomes flatter as ϵ → 0 \epsilon\to0 ϵ → 0 , and can thus be substituted either by the interval ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) lying on the horizontal axis (Fig. 2b), or by the unit disk lying on the horizontal plane. The condition (3.1) ensures that the curvature of Γ \Gamma Γ can… leading order. Moreover, for a fixed p ≥ 0 p \ge0 p ≥ 0 , the term ϵ 2 p / D \epsilon^{2p/D} ϵ 2 p / D vanishes in the limit ϵ → 0 \epsilon\to0 ϵ → 0 . As a consequence, the rescaled problem (3.2) can be formally substituted by an auxiliary Steklov-Neumann problem:
(i) in two dimensions, one searches for the eigenpairs { μ ^ k , V ^ k ( x , y ) } \{\widehat{\mu}_k, \widehat{V}_k(x,y)\} { μ k , V k ( x , y )} , satisfying in the half-plane H 2 = R × R + H_2 = \mathbb{R} \times\mathbb{R}_+ H 2 = R × R + :
Δ V ^ k = 0 ( y > 0 ) , (70) \Delta\widehat{V}_k = 0 \qquad(y > 0), \tag*{(70)} Δ V k = 0 ( y > 0 ) , (70)
− ( ∂ y V ^ k ) y = 0 = { 0 ( ∣ x ∣ ≥ 1 ) , μ ^ k V ^ k ( x , 0 ) ( ∣ x ∣ < 1 ) , (71) -\left(\partial_y \widehat{V}_k\right)_{y=0} =
\begin{cases}
0 & (|x| \ge1),\\
\widehat{\mu}_k \widehat{V}_k(x,0) & (|x| < 1),
\end{cases} \tag*{(71)} − ( ∂ y V k ) y = 0 = { 0 μ k V k ( x , 0 ) ( ∣ x ∣ ≥ 1 ) , ( ∣ x ∣ < 1 ) , (71)
∣ x ^ ∣ ∣ ∇ V ^ k ∣ → 0 ( ∣ x ^ ∣ → ∞ ) , (72) |\widehat{x}|\,|\nabla\widehat{V}_k| \to0 \qquad(|\widehat{x}| \to\infty), \tag*{(72)} ∣ x ∣ ∣∇ V k ∣ → 0 ( ∣ x ∣ → ∞ ) , (72)
with x ^ = ( x , y ) \widehat{x} = (x,y) x = ( x , y ) , and the normalization
∫ − 1 1 d x ∣ v ^ k ( x ) ∣ 2 = 1 , (73) \int_{-1}^{1} dx\,|\widehat{v}_k(x)|^2 = 1, \tag*{(73)} ∫ − 1 1 d x ∣ v k ( x ) ∣ 2 = 1 , (73)
where v ^ k ( x ) = V ^ k ( x , 0 ) \widehat{v}_k(x) = \widehat{V}_k(x,0) v k ( x ) = V k ( x , 0 ) is the restriction of V ^ k ( x ^ ) \widehat{V}_k(\widehat{x}) V k ( x ) to the interval ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) .
(ii) in three dimensions, one searches for the eigenpairs { μ ^ k , V ^ k ( x , y , z ) } \{\widehat{\mu}_k, \widehat{V}_k(x,y,z)\} { μ k , V k ( x , y , z )} , satisfying in the half-space H 3 = R 2 × R + H_3 = \mathbb{R}^2 \times\mathbb{R}_+ H 3 = R 2 × R + :
Δ V ^ k = 0 ( z > 0 ) , (74) \Delta\widehat{V}_k = 0 \qquad(z > 0), \tag*{(74)} Δ V k = 0 ( z > 0 ) , (74)
− ( ∂ z V ^ k ) z = 0 = { 0 ( x 2 + y 2 ≥ 1 ) , μ ^ k V ^ k ( x , y , 0 ) ( x 2 + y 2 < 1 ) , (75) -\left(\partial_z \widehat{V}_k\right)_{z=0} =
\begin{cases}
0 & (x^2 + y^2 \ge1),\\
\widehat{\mu}_k \widehat{V}_k(x,y,0) & (x^2 + y^2 < 1),
\end{cases} \tag*{(75)} − ( ∂ z V k ) z = 0 = { 0 μ k V k ( x , y , 0 ) ( x 2 + y 2 ≥ 1 ) , ( x 2 + y 2 < 1 ) , (75)
∣ x ^ ∣ 2 ∣ ∇ V ^ k ∣ → 0 ( ∣ x ^ ∣ → ∞ ) , (76) |\widehat{x}|^2 |\nabla\widehat{V}_k| \to0 \qquad(|\widehat{x}| \to\infty), \tag*{(76)} ∣ x ∣ 2 ∣∇ V k ∣ → 0 ( ∣ x ∣ → ∞ ) , (76)
with x ^ = ( x , y , z ) \widehat{x} = (x,y,z) x = ( x , y , z ) , and the normalization
∫ Γ 1 d x ^ ∣ v ^ k ( x ^ ) ∣ 2 = 1 , (77) \int_{\Gamma_1} d\widehat{x}\,|\widehat{v}_k(\widehat{x})|^2 = 1, \tag*{(77)} ∫ Γ 1 d x ∣ v k ( x ) ∣ 2 = 1 , (77)
where v ^ k ( x ^ ) \widehat{v}_k(\widehat{x}) v k ( x ) is the restriction of V ^ k ( x ^ ) \widehat{V}_k(\widehat{x}) V k ( x ) to the unit disk Γ 1 \Gamma_1 Γ 1 on the horizontal plane z = 0 z = 0 z = 0 .
Even though the domains H 2 H_2 H 2 and H 3 H_3 H 3 are unbounded, the subset with the Steklov condition (interval or disk) is bounded, so that the spectrum of each auxiliary Steklov-Neumann problem is discrete (see [53 , 54 , 55 , 56 , 57 ] for details). In particular, the eigenvalues μ ^ k \widehat{\mu}_k μ k can be ordered as:
0 = μ ^ 0 < μ ^ 1 ≤ μ ^ 2 ≤ ⋯ ↗ + ∞ , (78) 0 = \widehat{\mu}_0 < \widehat{\mu}_1 \le\widehat{\mu}_2 \le\cdots\nearrow+\infty, \tag*{(78)} 0 = μ 0 < μ 1 ≤ μ 2 ≤ ⋯ ↗ + ∞ , (78)
whereas the eigenfunctions { v ^ k } \{\widehat{v}_k\} { v k } form an orthonormal complete basis of the L 2 L^2 L 2 space on the interval ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) (d = 2 d = 2 d = 2 ) or the unit disk (d = 3 d = 3 d = 3 ). Note that the condition (76) was formulated in [53 ] to ensure that the eigenfunctions { v ^ k } \{\widehat{v}_k\} { v k } are orthogonal to a constant v ^ 0 ∝ 1 \widehat{v}_0 \propto1 v 0 ∝ 1 . We emphasize that the conditions (3.3c, 3.5c) allow for V ^ k ( x ^ ) \widehat{V}_k(\widehat{x}) V k ( x ) to have a constant nonzero limit V ^ k ( ∞ ) \widehat{V}_k(\infty) V k ( ∞ ) as l x ^ l → ∞ l\widehat{x}l \to\infty l x l → ∞ . These limiting values will play an important role in the following asymptotic analysis.
Comparing the rescaled problem (3.2) to the auxiliary problems (3.3, 3.5), one gets immediately the asymptotic relations (5) . Note that the factor ϵ − ( d − 1 ) / 2 \epsilon^{-(d-1)/2} ϵ − ( d − 1 ) /2 provides the correct rescaling of the normalization conditions (3.4, 3.6) to ensure the L 2 ( Γ ) L^2(\Gamma) L 2 ( Γ ) normalization (2.2) of the eigenfunctions v k ( p , Γ ) v_k^{(p,\Gamma)} v k ( p , Γ ) . We stress that the above dilation argument provides an asymptotic approximation of the eigenvalues μ k ( p , Γ ) \mu_k^{(p,\Gamma)} μ k ( p , Γ ) and of the eigenfunctions V k ( p , Γ ) ( x ) V_k^{(p,\Gamma)}(\mathbf{x}) V k ( p , Γ ) ( x ) on or near the subset Γ \Gamma Γ . In turn, the behavior of V k ( p , Γ ) ( x ) V_k^{(p,\Gamma)}(\mathbf{x}) V k ( p , Γ ) ( x ) far from Γ \Gamma Γ (i.e., when the Euclidean distance ∣ x − Γ ∣ |\mathbf{x}-\Gamma| ∣ x − Γ∣ is large) is not well captured by V ^ k ( x ^ ) \widehat{V}_k(\widehat{\mathbf{x}}) V k ( x ) . We also emphasize that the dilation argument is not a rigorous proof; in particular, we do not discuss the convergence of eigenfunctions, nor a more subtle situation of degenerate eigenvalues, in which case eigenprojectors are needed. A rigorous reformulation and validation of this formal analysis presents an interesting perspective of this work.
Quite remarkably, the asymptotic relations (1.3) imply that the eigenpairs { μ k ( p , Γ ) , v k ( p , Γ ) } \{\mu_k^{(p,\Gamma)},v_k^{(p,\Gamma)}\} { μ k ( p , Γ ) , v k ( p , Γ ) } do not depend on a fixed parameter p p p in the leading order as ϵ → 0 \epsilon\to0 ϵ → 0 . In practice, if ϵ \epsilon ϵ is fixed, the above dilation argument is valid when ϵ 2 p / D ≪ 1 \epsilon^2p/D \ll1 ϵ 2 p / D ≪ 1 . In other words, the asymptotic relations (1.3) provide an approximation when 0 ≤ p ≪ D / ϵ 2 0 \le p \ll D/\epsilon^2 0 ≤ p ≪ D / ϵ 2 . Since D / p \sqrt{D/p} D / p is a lengthscale, this condition is consistent with our earlier claim that ϵ \epsilon ϵ should be the smallest length of the problem.
In three dimensions, the axial symmetry of the Steklov-Neumann problem (3.5) implies that its eigenfunctions V ^ k \widehat{V}_k V k can be searched in the cylindrical coordinates ( r , z , φ ) (r,z,\varphi) ( r , z , φ ) as [ c m , n 1 cos ( m φ ) + c m , n 2 sin ( m φ ) ] V ^ m , n ( r , z ) [c^1_{m,n}\cos(m\varphi)+c^2_{m,n}\sin(m\varphi)]\widehat{V}_{m,n}(r,z) [ c m , n 1 cos ( m φ ) + c m , n 2 sin ( m φ )] V m , n ( r , z ) , for any m = 0 , 1 , 2 , … m=0,1,2,\ldots m = 0 , 1 , 2 , … , with arbitrary constants c m , n 1 c^1_{m,n} c m , n 1 and c m , n 2 c^2_{m,n} c m , n 2 . It is therefore convenient to employ the double index ( m , n ) (m,n) ( m , n ) to account for the 2 π 2\pi 2 π -periodicity of an eigenfunction on the angle φ \varphi φ , and the index n = 0 , 1 , 2 , … n=0,1,2,\ldots n = 0 , 1 , 2 , … to distinguish different eigenfunctions within the m m m -th family. Accordingly, the associated eigenvalues are denoted as μ ^ m , n \widehat{\mu}_{m,n} μ m , n . An extensive numerical analysis from [102 ] for spheroids suggested that the eigenvalues μ ^ 0 , n \widehat{\mu}_{0,n} μ 0 , n are simple, whereas μ ^ m , n \widehat{\mu}_{m,n} μ m , n are twice degenerate for m > 0 m>0 m > 0 , in agreement with the above linear combination of sine and cosine functions. Re-ordering the ensemble of all eigenvalues { μ ^ m , n } \{\widehat{\mu}_{m,n}\} { μ m , n } into an increasing sequence, one can recover the former enumeration by a single index k k k ; in other words, for each k = 0 , 1 , 2 , … k=0,1,2,\ldots k = 0 , 1 , 2 , … , there exist indices m k m_k m k and n k n_k n k such that μ ^ k = μ ^ m k , n k \widehat{\mu}_k=\widehat{\mu}_{m_k,n_k} μ k = μ m k , n k and
V ^ k ∝ [ c m k , n k 1 cos ( m k φ ) + c m k , n k 2 sin ( m k φ ) ] V ^ m k , n k ( r , z ) . (79) \widehat{V}_k \propto\left[c^1_{m_k,n_k}\cos(m_k\varphi)+c^2_{m_k,n_k}\sin(m_k\varphi)\right]\widehat{V}_{m_k,n_k}(r,z). \tag*{(79)} V k ∝ [ c m k , n k 1 cos ( m k φ ) + c m k , n k 2 sin ( m k φ ) ] V m k , n k ( r , z ) . (79)
In the following, we will use interchangeably both the double index ( m , n ) (m,n) ( m , n ) and the single index k k k .
There are different ways to construct numerically the eigenpairs { μ ^ k , V ^ k } \{\widehat{\mu}_k,\widehat{V}_k\} { μ k , V k } . For instance, the interval ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) in the plane R 2 \mathbb{R}^2 R 2 can be seen as the limit b → 0 b\to0 b → 0 of ellipses with semi-axes 1 1 1 and b b b (Figure 2 e), so that solutions of the spectral problem (3.3) can be obtained by inspecting the Steklov problem for the exterior of an ellipse (see Appendix § 11 ). Similarly, the unit disk in R 3 \mathbb{R}^3 R 3 can be seen as the limit of oblate spheroids, allowing one to construct solutions of the spectral problem (3.5) by looking at the Steklov problem for the exterior of spheroids (see Appendix § 12 and [102 ] ). In turn, the next section presents an alternative way to compute the eigenpairs { μ ^ k , V ^ k } \{\widehat{\mu}_k,\widehat{V}_k\} { μ k , V k } , which is based on the asymptotic behavior of the kernel G ( x x 0 ) \mathcal{G}(\mathbf{x}\mathbf{x}_0) G ( x x 0 ) and the limit of the integral equation (2.33) (a similar approach was used in [53 ] ). This alternative computation re-enforces the dilation argument described above.
Figure 2. In the small-target limit, zooming in near the subset Γ \Gamma Γ transforms the original Steklov-Neumann problem in a bounded domain (a) into the auxiliary spectral problem (3.3) in the upper half-plane H 2 = R × R + \mathbb{H}_2 = \mathbb{R} \times\mathbb{R}_+ H 2 = R × R + (b), with Steklov condition on the interval ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) (dashed red line) and Neumann condition on the remaining horizontal axis (solid blue line). The latter is equivalent to the Steklov problem in the exterior of the interval ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) in the plane R 2 \mathbb{R}^2 R 2 (c). The last two problems can be seen as the limiting cases (b = 0 b=0 b = 0 ) of two equivalent problems for the exterior of an ellipse with semi-axes a = 1 a=1 a = 1 and b b b in the plane (e) and for the exterior of a half-ellipse in the upper half-plane (d).
We conclude that the asymptotic relations (1.3) do not depend on the shape of the domain Ω \Omega Ω (see also [103 ] for a proper justification of this statement in the related context of the narrow-escape problem). In contrast, the shape of the subset Γ \Gamma Γ is relevant, as it defines the eigenpairs of the Steklov-Neumann problem (3.5) in the half-space. For instance, if the subset Γ \Gamma Γ had an elliptic shape, the dilation argument would still work but the asymptotic behavior would be determined by the eigenpairs of an auxiliary Steklov-Neumann problem for an ellipse in the half-space.
Asymptotic behavior of the pseudo-Green's function and related quantities
As in Sec. 2, the auxiliary Steklov-Neumann problems (3.3, 3.5) can be reformulated as spectral problems for suitable integral operators, like Eq. (43) . For this purpose, one needs to determine the asymptotic form of the kernel G ( x ∣ x 0 ) G(x|x_0) G ( x ∣ x 0 ) from Eq. (41) . Given that both x x x and x 0 x_0 x 0 belong to a small subset Γ \Gamma Γ and thus are close to each other, one can use the well-known asymptotic behavior of the pseudo-Green’s function G 0 ( x ∣ x 0 ) G_0(x|x_0) G 0 ( x ∣ x 0 ) as x → x 0 x \to x_0 x → x 0 . Re-delegating technical computations to Appendix G, we summarize the main results.
Two dimensions
When x 0 ∈ ∂ Ω x_0 \in\partial\Omega x 0 ∈ ∂ Ω , the pseudo-Green’s function in the limit x → x 0 x \to x_0 x → x 0 behaves as (see [65 ] and references therein)
G 0 ( x ∣ x 0 ) = − 1 π ln ∣ x − x 0 ∣ + R 0 ( x 0 ) + o ( 1 ) ( x → x 0 ∈ ∂ Ω ) , (80) G_0(x|x_0)=-\frac{1}{\pi}\ln|x-x_0|+R_0(x_0)+o(1) \qquad(x\to x_0\in\partial\Omega), \tag*{(80)} G 0 ( x ∣ x 0 ) = − π 1 ln ∣ x − x 0 ∣ + R 0 ( x 0 ) + o ( 1 ) ( x → x 0 ∈ ∂ Ω ) , (80)
where the regular part R 0 ( x 0 ) R_0(x_0) R 0 ( x 0 ) accounts for the contribution of order O ( 1 ) O(1) O ( 1 ) , while o ( 1 ) o(1) o ( 1 ) denotes higher-order corrections vanishing as x → x 0 x \to x_0 x → x 0 . We outline the minor abuse of notation when the dimensional distance ∣ x − x 0 ∣ |x-x_0| ∣ x − x 0 ∣ stands in the argument of the logarithmic function; a more accurate expression would be G 0 ( x ∣ x 0 ) = − 1 π ln ( ∣ x − x 0 ∣ / ℓ ) + R ˉ 0 ( x 0 ) + o ( 1 ) G_0(x|x_0)=-\frac{1}{\pi}\ln(|x-x_0|/\ell)+\bar{R}_0(x_0)+o(1) G 0 ( x ∣ x 0 ) = − π 1 ln ( ∣ x − x 0 ∣/ ℓ ) + R ˉ 0 ( x 0 ) + o ( 1 ) with a suitable length ℓ \ell ℓ , but we keep using (3.9) in the following.
Substituting this expression into Eqs. (39) , (40) yields (see details in Appendix G):
A Γ ≈ − ln ( 2 ϵ ) π + 3 2 π + R 0 ( x Γ ) + o ( 1 ) , (81) \mathcal{A}_{\Gamma} \approx-\frac{\ln(2\epsilon)}{\pi}+\frac{3}{2\pi}+R_{0}(x_{\Gamma})+o(1), \tag*{(81)} A Γ ≈ − π ln ( 2 ϵ ) + 2 π 3 + R 0 ( x Γ ) + o ( 1 ) , (81)
w 0 ( Γ ) ( x 0 ) ≈ w ^ 0 ( s 0 / ϵ ) ( x 0 ∈ Γ ) , (82) w_{0}^{(\Gamma)}(x_{0}) \approx\widehat{w}_{0}(s_{0}/\epsilon) \qquad(x_{0} \in\Gamma), \tag*{(82)} w 0 ( Γ ) ( x 0 ) ≈ w 0 ( s 0 / ϵ ) ( x 0 ∈ Γ ) , (82)
where x Γ ∈ ∂ Ω x_{\Gamma} \in\partial\Omega x Γ ∈ ∂ Ω is the center of the subset Γ \Gamma Γ , s 0 s_{0} s 0 is the curvilinear coordinate of x 0 ∈ Γ x_{0} \in\Gamma x 0 ∈ Γ (ranging from − ϵ -\epsilon − ϵ to ϵ \epsilon ϵ ), and
w ^ 0 ( x ) = − ( 1 + x ) ln ( 1 + x ) + ( 1 − x ) ln ( 1 − x ) + 1 − 2 ln 2 2 π . (83) \widehat{w}_{0}(x)=-\frac{(1+x)\ln(1+x)+(1-x)\ln(1-x)+1-2\ln2}{2\pi}. \tag*{(83)} w 0 ( x ) = − 2 π ( 1 + x ) ln ( 1 + x ) + ( 1 − x ) ln ( 1 − x ) + 1 − 2 ln 2 . (83)
Moreover, the coefficients b k ( Γ ) b_{k}^{(\Gamma)} b k ( Γ ) from Eq. (46) behave in the leading order as
b k ( Γ ) ≈ ϵ − 1 / 2 V ^ k ( ∞ ) , (84) b_{k}^{(\Gamma)} \approx\epsilon^{-1/2}\widehat{V}_{k}(\infty), \tag*{(84)} b k ( Γ ) ≈ ϵ − 1/2 V k ( ∞ ) , (84)
where the factor ϵ − 1 / 2 \epsilon^{-1/2} ϵ − 1/2 agrees with the rescaling of eigenfunctions in Eq. (5) , and V ^ k ( ∞ ) \widehat{V}_{k}(\infty) V k ( ∞ ) is the limit of V ^ k ( x ^ ) \widehat{V}_{k}(\widehat{x}) V k ( x ) as ∣ x ^ ∣ → ∞ |\widehat{x}| \to\infty ∣ x ∣ → ∞ . The first ten eigenvalues μ ^ k \widehat{\mu}_{k} μ k and the associated limits V ^ k ( ∞ ) \widehat{V}_{k}(\infty) V k ( ∞ ) are reported in Table 1.
Substituting Eqs. (81) , (82) into Eq. (41) , we get
G ( x ∣ x 0 ) ≈ G ^ ( s / ϵ ∣ s 0 / ϵ ) + o ( 1 ) , (85) G(x|x_{0}) \approx\widehat{G}(s/\epsilon|s_{0}/\epsilon)+o(1), \tag*{(85)} G ( x ∣ x 0 ) ≈ G ( s / ϵ ∣ s 0 / ϵ ) + o ( 1 ) , (85)
where s s s and s 0 s_{0} s 0 are the curvilinear coordinates of x x x and x 0 x_{0} x 0 on the subset Γ \Gamma Γ , and
G ^ ( x ∣ x 0 ) = − 1 π ln ∣ x − x 0 ∣ + 1 2 π G ^ s y m ( x ∣ x 0 ) , (86) \widehat{G}(x|x_{0})=-\frac{1}{\pi}\ln|x-x_{0}|+\frac{1}{2\pi}\widehat{G}_{\mathrm{sym}}(x|x_{0}), \tag*{(86)} G ( x ∣ x 0 ) = − π 1 ln ∣ x − x 0 ∣ + 2 π 1 G sym ( x ∣ x 0 ) , (86)
with
G ^ s y m ( x ∣ x 0 ) = ( 1 + x 0 ) ln ( 1 + x 0 ) + ( 1 − x 0 ) ln ( 1 − x 0 ) + ( 1 + x ) ln ( 1 + x ) + ( 1 − x ) ln ( 1 − x ) − 1 − 2 ln ( 2 ) . (87) \widehat{G}_{\mathrm{sym}}(x|x_{0})=(1+x_{0})\ln(1+x_{0})+(1-x_{0})\ln(1-x_{0})+(1+x)\ln(1+x)+(1-x)\ln(1-x)-1-2\ln(2). \tag*{(87)} G sym ( x ∣ x 0 ) = ( 1 + x 0 ) ln ( 1 + x 0 ) + ( 1 − x 0 ) ln ( 1 − x 0 ) + ( 1 + x ) ln ( 1 + x ) + ( 1 − x ) ln ( 1 − x ) − 1 − 2 ln ( 2 ) . (87)
This explicit kernel defines an integral operator on L 2 ( ( − 1 , 1 ) ) L^{2}((-1,1)) L 2 (( − 1 , 1 )) that determines the eigenpairs of the Steklov-Neumann problem (3.3):
∫ − 1 1 d x v ^ k ( x ) G ^ ( x ∣ x 0 ) = 1 μ ^ k v ^ k ( x 0 ) ( − 1 < x 0 < 1 , k = 1 , 2 , … ) (88) \int_{-1}^{1} dx\,\widehat{v}_{k}(x)\widehat{G}(x|x_{0})=\frac{1}{\widehat{\mu}_{k}}\widehat{v}_{k}(x_{0}) \qquad(-1<x_{0}<1,\ k=1,2,\ldots) \tag*{(88)} ∫ − 1 1 d x v k ( x ) G ( x ∣ x 0 ) = μ k 1 v k ( x 0 ) ( − 1 < x 0 < 1 , k = 1 , 2 , … ) (88)
(note that the trivial eigenpair μ ^ 0 = 0 \widehat{\mu}_{0}=0 μ 0 = 0 and v ^ 0 = 1 / 2 \widehat{v}_{0}=1/\sqrt{2} v 0 = 1/ 2 is not captured by this equation and should be added separately). If the eigenvalue μ ^ k \widehat{\mu}_{k} μ k is simple, the symmetry of the kernel, G ^ ( − x ∣ x 0 ) = G ^ ( x ∣ − x 0 ) \widehat{G}(-x|x_{0})=\widehat{G}(x|-x_{0}) G ( − x ∣ x 0 ) = G ( x ∣ − x 0 ) , implies that the associated eigenfunction v ^ k ( x ) \widehat{v}_{k}(x) v k ( x ) is either symmetric, or antisymmetric. If an eigenfunction v ^ k ( x ) \widehat{v}_{k}(x) v k ( x ) is antisymmetric, its integral with the symmetric part G ^ s y m ( x ∣ x 0 ) \widehat{G}_{\mathrm{sym}}(x|x_{0}) G sym ( x ∣ x 0 ) of the kernel G ^ ( x ∣ x 0 ) \widehat{G}(x|x_{0}) G ( x ∣ x 0 ) vanishes. In other words, the antisymmetric eigenfunctions can be obtained by solving the simpler spectral problem for the integral operator with the logarithmic kernel:
∫ − 1 1 d x v ^ 2 k − 1 ( x ) 1 π ln ( 1 ∣ x − x 0 ∣ ) = v ^ 2 k − 1 ( x 0 ) μ ^ 2 k − 1 ( − 1 < x 0 < 1 ) , (89) \int_{-1}^{1} dx\,\widehat{v}_{2k-1}(x)\frac{1}{\pi}\ln\left(\frac{1}{|x-x_{0}|}\right)=\frac{\widehat{v}_{2k-1}(x_{0})}{\widehat{\mu}_{2k-1}} \qquad(-1<x_{0}<1), \tag*{(89)} ∫ − 1 1 d x v 2 k − 1 ( x ) π 1 ln ( ∣ x − x 0 ∣ 1 ) = μ 2 k − 1 v 2 k − 1 ( x 0 ) ( − 1 < x 0 < 1 ) , (89)
where we used odd indices 2 k − 1 2k-1 2 k − 1 to enumerate antisymmetric eigenfunctions. Such spectral problems were thoroughly studied in the past (see [104 , 105 , 106 , 107 ] [108] and references therein). In particular, the large-k k k asymptotic behavior of the eigenvalues μ ^ 2 k − 1 \hat{\mu}_{2k-1} μ ^ 2 k − 1 of the antisymmetric eigenfunctions of Eq. (89) was given in [108]. In the leading order, one has μ ^ 2 k − 1 ≈ π k + O ( 1 ) \hat{\mu}_{2k-1} \approx\pi k + O(1) μ ^ 2 k − 1 ≈ π k + O ( 1 ) . While all eigenvalues that we computed numerically were simple (see Table 1 ), a rigorous proof of this conjecture for all eigenvalues remains an open question. We also checked numerically that the eigenvalues μ ^ 2 k \hat{\mu}_{2k} μ ^ 2 k corresponding to the symmetric eigenfunctions of the kernel G ^ ( x ∣ x 0 ) \hat{\mathcal{G}}(x|x_0) G ^ ( x ∣ x 0 ) obey the same asymptotic relation in the leading order. We conclude that
Table 1. Table 1 (Top six rows) First ten eigenvalues μ ^ k \hat{\mu}_k μ ^ k of the Steklov-Neumann problem (3.3), apart from μ ^ 0 = 0 \hat{\mu}_0=0 μ ^ 0 = 0 , as well as the squared coefficients [ V ^ k ( ∞ ) ] 2 [\hat{V}_k(\infty)]^2 [ V ^ k ( ∞ ) ] 2 determining the asymptotic behavior of b k ( 1 ) b_k^{(1)} b k ( 1 ) and the function Ψ 2 d ( z ) \Psi_{2\mathrm{d}}(z) Ψ 2 d ( z ) from Eq. (124) . Note that 1 / μ ^ k 1/\hat{\mu}_k 1/ μ ^ k are the eigenvalues of the kernel G ^ ( x ∣ x 0 ) \hat{G}(x|x_0) G ^ ( x ∣ x 0 ) from Eq. (86) . They were obtained by a numerical diagonalization of the matrix M ( 0 ) M(0) M ( 0 ) from Eq. (E.12) of size 200 × 200 200 \times200 200 × 200 , see Appendix § 11 . (Bottom six rows) First ten eigenvalues μ ^ 0 , k \hat{\mu}_{0,k} μ ^ 0 , k corresponding to axially symmetric eigenfunctions of the Steklov-Neumann problem (3.5), apart from μ ^ 0 , 0 = 0 \hat{\mu}_{0,0}=0 μ ^ 0 , 0 = 0 , as well as the squared coefficients [ V ^ 0 , k ( ∞ ) ] 2 [\hat{V}_{0,k}(\infty)]^2 [ V ^ 0 , k ( ∞ ) ] 2 determining the asymptotic behavior of b 0 , k ( 1 ) b_{0,k}^{(1)} b 0 , k ( 1 ) and the function Ψ 3 d ( z ) \Psi_{3\mathrm{d}}(z) Ψ 3 d ( z ) from Eq. (131) . They were obtained by a numerical diagonalization of the matrix G G G from Eq. (F.8) of size 100 × 100 100 \times100 100 × 100 ; note that only even eigenmodes were kept, see Appendix § 12 . For all data, the shown digits do not change as the truncation size increases. These eigenvalues are in excellent agreement with those reported in [53 ] , see Tables 6.1 and 6.2.
μ ^ k ≈ π 2 k + O ( 1 ) ( k → ∞ ) , (90) \hat{\mu}_k \approx\frac{\pi}{2} k + O(1) \qquad(k \to\infty), \tag*{(90)} μ ^ k ≈ 2 π k + O ( 1 ) ( k → ∞ ) , (90)
which agrees with lower and upper bounds discussed in Appendix § 14 .
Three dimensions
When x 0 ∈ ∂ Ω x_0 \in\partial\Omega x 0 ∈ ∂ Ω , the pseudo-Green’s function in the limit x → x 0 x \to x_0 x → x 0 behaves as
G 0 ( x ∣ x 0 ) = 1 2 π ∣ x − x 0 ∣ − H ( x 0 ) 4 π ln ∣ x − x 0 ∣ + R 0 ( x 0 ) + o ( 1 ) ( x → x 0 ∈ ∂ Ω ) , (91) \mathcal{G}_0(x|x_0) = \frac{1}{2\pi|x-x_0|} - \frac{H(x_0)}{4\pi}\ln|x-x_0| + R_0(x_0) + o(1) \qquad(x \to x_0 \in\partial\Omega), \tag*{(91)} G 0 ( x ∣ x 0 ) = 2 π ∣ x − x 0 ∣ 1 − 4 π H ( x 0 ) ln ∣ x − x 0 ∣ + R 0 ( x 0 ) + o ( 1 ) ( x → x 0 ∈ ∂ Ω ) , (91)
where R 0 ( x 0 ) R_0(x_0) R 0 ( x 0 ) is the regular part, and H ( x 0 ) H(x_0) H ( x 0 ) is the mean curvature of the boundary at the point x 0 ∈ ∂ Ω x_0 \in\partial\Omega x 0 ∈ ∂ Ω [71 , 109 , 110 ] . Substituting this expression into Eqs. (39) , (40) yields (see details in Appendix G):
A Γ ≈ 8 3 π 2 ϵ − H ( x Γ ) 4 π ( ln ϵ − 1 4 ) + R 0 ( x Γ ) + o ( 1 ) , (92) \mathcal{A}_{\Gamma} \approx\frac{8}{3\pi^2\epsilon} - \frac{H(x_{\Gamma})}{4\pi}\left(\ln\epsilon- \frac{1}{4}\right) + R_0(x_{\Gamma}) + o(1), \tag*{(92)} A Γ ≈ 3 π 2 ϵ 8 − 4 π H ( x Γ ) ( ln ϵ − 4 1 ) + R 0 ( x Γ ) + o ( 1 ) , (92)
w 0 ( Γ ) ( x 0 ) ≈ 1 2 π ϵ w ^ 0 ( r 0 / ϵ ) + O ( 1 ) ( x 0 ∈ Γ ) , (93) w_0^{(\Gamma)}(x_0) \approx\frac{1}{2\pi\epsilon}\hat{w}_0(r_0/\epsilon) + O(1) \qquad(x_0 \in\Gamma), \tag*{(93)} w 0 ( Γ ) ( x 0 ) ≈ 2 π ϵ 1 w ^ 0 ( r 0 / ϵ ) + O ( 1 ) ( x 0 ∈ Γ ) , (93)
where r 0 = ∣ x 0 − x Γ ∣ r_0 = |x_0-x_{\Gamma}| r 0 = ∣ x 0 − x Γ ∣ ,
w ^ 0 ( r ^ ) = 4 π ( E ( r ^ ) − 4 3 ) , (94) \hat{w}_0(\hat{r}) = \frac{4}{\pi}\left(E(\hat{r})-\frac{4}{3}\right), \tag*{(94)} w ^ 0 ( r ^ ) = π 4 ( E ( r ^ ) − 3 4 ) , (94)
and E ( k ) E(k) E ( k ) is the complete elliptic integral of the second kind:
E ( k ) = ∫ 0 π / 2 d z 1 − k 2 sin 2 z . (95) E(k) = \int_0^{\pi/2} dz \sqrt{1-k^2\sin^2 z}. \tag*{(95)} E ( k ) = ∫ 0 π /2 d z 1 − k 2 sin 2 z . (95)
As a consequence, the coefficients b k ( Γ ) b_k^{(\Gamma)} b k ( Γ ) from Eq. (46) behave as
b k ( Γ ) = ϵ − 1 V ^ k ( ∞ ) + O ( 1 ) ( ϵ → 0 ) , (96) b_k^{(\Gamma)} = \epsilon^{-1}\hat{V}_k(\infty) + O(1) \qquad(\epsilon\to0), \tag*{(96)} b k ( Γ ) = ϵ − 1 V ^ k ( ∞ ) + O ( 1 ) ( ϵ → 0 ) , (96)
where the factor ϵ − 1 \epsilon^{-1} ϵ − 1 agrees with the rescaling of eigenfunctions in Eq. (5) . If V ^ k \hat{V}_k V ^ k is a periodically oscillating (along the angle φ \varphi φ ), non-axially-symmetric eigenfunction, it vanishes at infinity so that the corresponding b k ( Γ ) b_k^{(\Gamma)} b k ( Γ ) behaves as O ( 1 ) O(1) O ( 1 ) . The first ten eigenvalues μ ^ 0 , n \hat{\mu}_{0,n} μ ^ 0 , n and the limiting values V ^ 0 , n ( ∞ ) \hat{V}_{0,n}(\infty) V ^ 0 , n ( ∞ ) of the associated axially symmetric eigenfunctions are reported in Table 1 .
Let us now deduce an alternative spectral problem for determining the eigenpairs { μ ^ k , v ^ k } \{\hat{\mu}_k,\hat{v}_k\} { μ ^ k , v ^ k } . Substituting Eqs. (92) , (93) into Eq. (41) , we determine the asymptotic behavior of the kernel G ( x ∣ x 0 ) \mathcal{G}(x|x_0) G ( x ∣ x 0 ) :
G ( x ∣ x 0 ) = 1 2 π ∣ x − x 0 ∣ − H ( x Γ ) 4 π ln ∣ x − x 0 ∣ − w ^ 0 ( r 0 / ϵ ) 2 π ϵ − w ^ 0 ( r / ϵ ) 2 π ϵ − 8 3 π 2 ϵ + H ( x Γ ) 4 π ln ϵ + O ( 1 ) , (97) \mathcal{G}(x|x_0) = \frac{1}{2\pi|x-x_0|} - \frac{H(x_{\Gamma})}{4\pi}\ln|x-x_0| - \frac{\hat{w}_0(r_0/\epsilon)}{2\pi\epsilon} - \frac{\hat{w}_0(r/\epsilon)}{2\pi\epsilon} - \frac{8}{3\pi^2\epsilon} + \frac{H(x_{\Gamma})}{4\pi}\ln\epsilon+ O(1), \tag*{(97)} G ( x ∣ x 0 ) = 2 π ∣ x − x 0 ∣ 1 − 4 π H ( x Γ ) ln ∣ x − x 0 ∣ − 2 π ϵ w ^ 0 ( r 0 / ϵ ) − 2 π ϵ w ^ 0 ( r / ϵ ) − 3 π 2 ϵ 8 + 4 π H ( x Γ ) ln ϵ + O ( 1 ) , (97)
where we replaced H ( x 0 ) H(x_0) H ( x 0 ) by H ( x Γ ) H(x_\Gamma) H ( x Γ ) in Eq. (91) , up to a negligible error O ( ϵ ) O(\epsilon) O ( ϵ ) . One sees that the leading-order terms of the kernel G ( x ∣ x 0 ) \mathcal{G}(x|x_0) G ( x ∣ x 0 ) respect the rotational symmetry of the disk and can thus be decomposed in the cylindrical coordinates x = ( r , 0 , φ ) x=(r,0,\varphi) x = ( r , 0 , φ ) and x 0 = ( r 0 , 0 , φ 0 ) x_0=(r_0,0,\varphi_0) x 0 = ( r 0 , 0 , φ 0 ) as
G ( x ∣ x 0 ) = 1 2 π ϵ ∑ m = − ∞ ∞ e − i m ( φ − φ 0 ) G ^ ( m ) ( r / ϵ ∣ r 0 / ϵ ) + O ( 1 ) , (98) \mathcal{G}(x|x_0)=\frac{1}{2\pi\epsilon}\sum_{m=-\infty}^{\infty}e^{-im(\varphi-\varphi_0)}\widehat{\mathcal{G}}^{(m)}(r/\epsilon|r_0/\epsilon)+O(1), \tag*{(98)} G ( x ∣ x 0 ) = 2 π ϵ 1 m = − ∞ ∑ ∞ e − im ( φ − φ 0 ) G ( m ) ( r / ϵ ∣ r 0 / ϵ ) + O ( 1 ) , (98)
where the Fourier coefficients determine the kernels
G ^ ( m ) ( r ^ ∣ r ^ 0 ) = lim ϵ → 0 ϵ ∫ 0 2 π d φ e i m ( φ − φ 0 ) G ( ϵ x ^ ∣ ϵ x ^ 0 ) , (99) \widehat{\mathcal{G}}^{(m)}(\hat{r}|\hat{r}_0)=\lim_{\epsilon\to0}\epsilon\int_{0}^{2\pi}d\varphi e^{im(\varphi-\varphi_0)}\mathcal{G}(\epsilon\hat{x}|\epsilon\hat{x}_0), \tag*{(99)} G ( m ) ( r ^ ∣ r ^ 0 ) = ϵ → 0 lim ϵ ∫ 0 2 π d φ e im ( φ − φ 0 ) G ( ϵ x ^ ∣ ϵ x ^ 0 ) , (99)
with rescaled positions x ^ = x / ϵ = ( r ^ , 0 , φ ) \hat{x}=x/\epsilon=(\hat{r},0,\varphi) x ^ = x / ϵ = ( r ^ , 0 , φ ) and x ^ 0 = x 0 / ϵ = ( r ^ 0 , 0 , φ 0 ) \hat{x}_0=x_0/\epsilon=(\hat{r}_0,0,\varphi_0) x ^ 0 = x 0 / ϵ = ( r ^ 0 , 0 , φ 0 ) in the unit disk. For any integer m ≠ 0 m\ne0 m = 0 , we get
G ^ ( m ) ( r ^ ∣ r ^ 0 ) = ∫ 0 2 π d φ cos ( m φ ) 2 π r ^ 2 + r ^ 0 2 − 2 r ^ r ^ 0 cos φ , (100) \widehat{\mathcal{G}}^{(m)}(\hat{r}|\hat{r}_0)=\int_{0}^{2\pi}d\varphi\frac{\cos(m\varphi)}{2\pi\sqrt{\hat{r}^{2}+\hat{r}_0^{2}-2\hat{r}\hat{r}_0\cos\varphi}}, \tag*{(100)} G ( m ) ( r ^ ∣ r ^ 0 ) = ∫ 0 2 π d φ 2 π r ^ 2 + r ^ 0 2 − 2 r ^ r ^ 0 cos φ cos ( m φ ) , (100)
which comes from the first term in Eq. (97) ; note the contribution of the second term was O ( ϵ ) O(\epsilon) O ( ϵ ) and thus vanished. For m = 0 m=0 m = 0 , we use Eqs. (G.16, G.18) to get
∫ 0 2 π d φ G ( x ∣ x 0 ) = 2 K ( r < / r > ) π r > − H ( x Γ ) 4 ln ( r 2 + r 0 2 + ∣ r 2 − r 0 2 ∣ 2 ) − w ^ 0 ( r 0 / ϵ ) + w ^ 0 ( r / ϵ ) + 16 π / 3 ϵ + H ( x Γ ) 2 ln ϵ + O ( 1 ) , \begin{aligned}
\int_{0}^{2\pi}d\varphi\mathcal{G}(x|x_0)
&=\frac{2K(r_{<}/r_{>})}{\pi r_{>}}-\frac{H(x_\Gamma)}{4}\ln\left(\frac{r^{2}+r_0^{2}+|r^{2}-r_0^{2}|}{2}\right)\\
&\quad-\frac{\widehat{w}_0(r_0/\epsilon)+\widehat{w}_0(r/\epsilon)+16\pi/3}{\epsilon}+\frac{H(x_\Gamma)}{2}\ln\epsilon+O(1),
\end{aligned} ∫ 0 2 π d φ G ( x ∣ x 0 ) = π r > 2 K ( r < / r > ) − 4 H ( x Γ ) ln ( 2 r 2 + r 0 2 + ∣ r 2 − r 0 2 ∣ ) − ϵ w 0 ( r 0 / ϵ ) + w 0 ( r / ϵ ) + 16 π /3 + 2 H ( x Γ ) ln ϵ + O ( 1 ) ,
where r < = min { r , r 0 } r_{<}=\min\{r,r_0\} r < = min { r , r 0 } , r > = max { r , r 0 } r_{>}=\max\{r,r_0\} r > = max { r , r 0 } , and K ( k ) K(k) K ( k ) is the complete elliptic integral of the first kind:
K ( k ) = ∫ 0 π / 2 d z 1 − k 2 sin 2 z . (101) K(k)=\int_{0}^{\pi/2}\frac{dz}{\sqrt{1-k^{2}\sin^{2}z}}. \tag*{(101)} K ( k ) = ∫ 0 π /2 1 − k 2 sin 2 z d z . (101)
Substituting r = ϵ r ^ r=\epsilon\hat{r} r = ϵ r ^ and r 0 = ϵ r ^ 0 r_0=\epsilon\hat{r}_0 r 0 = ϵ r ^ 0 , we see that the second term is subleading and thus vanishes in the limit ϵ → 0 \epsilon\to0 ϵ → 0 , yielding
G ^ ( 0 ) ( r ^ ∣ r ^ 0 ) = 4 π [ K ( r ^ < / r ^ > ) 2 r ^ > − E ( r ^ ) − E ( r ^ 0 ) + 4 3 ] . (102) \widehat{\mathcal{G}}^{(0)}(\hat{r}|\hat{r}_0)=\frac{4}{\pi}\left[\frac{K(\hat{r}_{<}/\hat{r}_{>})}{2\hat{r}_{>}}-E(\hat{r})-E(\hat{r}_0)+\frac{4}{3}\right]. \tag*{(102)} G ( 0 ) ( r ^ ∣ r ^ 0 ) = π 4 [ 2 r ^ > K ( r ^ < / r ^ > ) − E ( r ^ ) − E ( r ^ 0 ) + 3 4 ] . (102)
The Fourier expansion (98) of the kernel G ( x ∣ x 0 ) \mathcal{G}(x|x_0) G ( x ∣ x 0 ) suggests to search the leading-order term of an eigenfunction v k ( 0 , Γ ) v_k^{(0,\Gamma)} v k ( 0 , Γ ) of the Dirichlet-to-Neumann operator M 0 ( Γ ) \mathcal{M}^{(\Gamma)}_0 M 0 ( Γ ) in the form e i m φ v ^ m , n ( r / ϵ ) e^{im\varphi}\hat{v}_{m,n}(r/\epsilon) e im φ v ^ m , n ( r / ϵ ) , with suitable indices m m m and n n n . Substitution of this form and Eq. (98) into Eq. (43) yields
e i m φ 0 v ^ m , n ( r 0 / ϵ ) μ m , n = ∫ 0 ϵ d r r ∫ 0 2 π d φ e i m φ v ^ m , n ( r / ϵ ) 1 2 π ϵ ∑ m ′ = − ∞ ∞ e − i m ′ ( φ − φ 0 ) G ^ ( m ′ ) ( r / ϵ ∣ r 0 / ϵ ) = e i m φ 0 ϵ ∫ 0 1 d r ^ r ^ v ^ m , n ( r ^ ) G ^ ( m ) ( r ^ ∣ r ^ 0 ) , \begin{aligned}
\frac{e^{im\varphi_0}\hat{v}_{m,n}(r_0/\epsilon)}{\mu_{m,n}}
&=\int_{0}^{\epsilon}dr\,r\int_{0}^{2\pi}d\varphi\,e^{im\varphi}\hat{v}_{m,n}(r/\epsilon)\frac{1}{2\pi\epsilon}\sum_{m'=-\infty}^{\infty}e^{-im'(\varphi-\varphi_0)}\widehat{\mathcal{G}}^{(m')}(r/\epsilon|r_0/\epsilon)\\
&=e^{im\varphi_0}\epsilon\int_{0}^{1}d\hat{r}\,\hat{r}\,\hat{v}_{m,n}(\hat{r})\widehat{\mathcal{G}}^{(m)}(\hat{r}|\hat{r}_0),
\end{aligned} μ m , n e im φ 0 v ^ m , n ( r 0 / ϵ ) = ∫ 0 ϵ d r r ∫ 0 2 π d φ e im φ v ^ m , n ( r / ϵ ) 2 π ϵ 1 m ′ = − ∞ ∑ ∞ e − i m ′ ( φ − φ 0 ) G ( m ′ ) ( r / ϵ ∣ r 0 / ϵ ) = e im φ 0 ϵ ∫ 0 1 d r ^ r ^ v ^ m , n ( r ^ ) G ( m ) ( r ^ ∣ r ^ 0 ) ,
i.e.,
∫ 0 1 d r ^ r ^ v ^ m , n ( r ^ ) G ^ ( m ) ( r ^ ∣ r ^ 0 ) = v ^ m , n ( r ^ 0 ) μ ^ m , n ( 0 < r ^ 0 < 1 , n = 1 , 2 , … ) . (103) \int_{0}^{1} d\hat{r}\,\hat{r}\,\hat{v}_{m,n}(\hat{r})\,\hat{G}^{(m)}(\hat{r}|\hat{r}_0)=\frac{\hat{v}_{m,n}(\hat{r}_0)}{\hat{\mu}_{m,n}}
\qquad(0<\hat{r}_0<1,\ n=1,2,\ldots). \tag*{(103)} ∫ 0 1 d r ^ r ^ v ^ m , n ( r ^ ) G ^ ( m ) ( r ^ ∣ r ^ 0 ) = μ ^ m , n v ^ m , n ( r ^ 0 ) ( 0 < r ^ 0 < 1 , n = 1 , 2 , … ) . (103)
One sees that for each integer m m m , the kernel G ^ ( m ) ( r ^ ∣ r ^ 0 ) \hat{G}^{(m)}(\hat{r}|\hat{r}_0) G ^ ( m ) ( r ^ ∣ r ^ 0 ) defines an integral operator in L 2 ( ( 0 , 1 ) ) L^2((0,1)) L 2 (( 0 , 1 )) (with the scalar product including the weight r ^ \hat{r} r ^ ) that determines the eigenvalues 1 / μ ^ m , n 1/\hat{\mu}_{m,n} 1/ μ ^ m , n and eigenfunctions v ^ m , n ( r ^ ) \hat{v}_{m,n}(\hat{r}) v ^ m , n ( r ^ ) . As previously, we impose the normalization
2 π ∫ 0 1 d r ^ r ^ ∣ v ^ m , n ( r ^ ) ∣ 2 = 1. (104) 2\pi\int_{0}^{1}d\hat{r}\,\hat{r}\,|\hat{v}_{m,n}(\hat{r})|^2=1. \tag*{(104)} 2 π ∫ 0 1 d r ^ r ^ ∣ v ^ m , n ( r ^ ) ∣ 2 = 1. (104)
As in the planar case, the principal eigenpair μ ^ 0 = 0 \hat{\mu}_0=0 μ ^ 0 = 0 and v ^ 0 = 1 / π \hat{v}_0=1/\sqrt{\pi} v ^ 0 = 1/ π should be added separately. Note that we used here the complex-valued form e i m φ v ^ m , n ( r ^ ) e^{im\varphi}\hat{v}_{m,n}(\hat{r}) e im φ v ^ m , n ( r ^ ) instead of the real-valued form in Eq. (79) . Given that G ^ ( − m ) ( r ^ ∣ r ^ 0 ) = G ^ ( m ) ( r ^ ∣ r ^ 0 ) \hat{G}^{(-m)}(\hat{r}|\hat{r}_0)=\hat{G}^{(m)}(\hat{r}|\hat{r}_0) G ^ ( − m ) ( r ^ ∣ r ^ 0 ) = G ^ ( m ) ( r ^ ∣ r ^ 0 ) , one can easily translate one form to the other.
Since the kernels G ^ ( m ) ( r ^ ∣ r ^ 0 ) \hat{G}^{(m)}(\hat{r}|\hat{r}_0) G ^ ( m ) ( r ^ ∣ r ^ 0 ) were obtained as the leading-order contributions to the kernel G ( x ∣ x 0 ) G(\boldsymbol{x}|\boldsymbol{x}_0) G ( x ∣ x 0 ) in the limit ϵ → 0 \epsilon\to0 ϵ → 0 , the eigenvalues μ ^ m , n \hat{\mu}_{m,n} μ ^ m , n and eigenfunctions e i m φ v ^ m , n ( r ^ ) e^{im\varphi}\hat{v}_{m,n}(\hat{r}) e im φ v ^ m , n ( r ^ ) are suitable candidates to represent the leading-order behavior of the eigenvalues and eigenfunctions of the Dirichlet-to-Neumann operator M 0 ( Γ ) \mathcal{M}_0^{(\Gamma)} M 0 ( Γ ) . In other words, the spectral problem (103) provides an alternative formulation of the Steklov-Neumann problem (3.5). We emphasize that the equivalence between these two formulations requires a mathematical proof (see also [53 ] ). Note that the curvature of the boundary, H ( x Γ ) H(\boldsymbol{x}_{\Gamma}) H ( x Γ ) , that appeared in the kernel G ( x ∣ x 0 ) G(\boldsymbol{x}|\boldsymbol{x}_0) G ( x ∣ x 0 ) in the subleading term in Eq. (97) , does not affect the leading-order contributions to the eigenvalues and eigenfunctions. This analysis justifies the possibility of replacing a curved subset Γ \Gamma Γ by a disk in the limit ϵ → 0 \epsilon\to0 ϵ → 0 .
Role of the boundary connectivity
It is also instructive to outline the role of the assumption that the smooth boundary ∂ Ω \partial\Omega ∂ Ω is connected, i.e., the distance between the subset Γ \Gamma Γ and the reflecting part ∂ Ω N \partial\Omega_N ∂ Ω N is zero. This is a typical setting for diffusive applications when a particle has to react on a reactive patch Γ \Gamma Γ on an inert boundary or to escape from a domain through a “hole” Γ \Gamma Γ on the confining wall. In other applications, however, reactive patches Γ \Gamma Γ are dispersed in the volume, which is enclosed by a reflecting boundary ∂ Ω N \partial\Omega_N ∂ Ω N so that the boundary ∂ Ω \partial\Omega ∂ Ω (still formed by Γ \Gamma Γ and ∂ Ω N \partial\Omega_N ∂ Ω N ) is not connected. This setting is often referred to as the narrow capture problem to distinguish it from the narrow escape problem. In both cases, the subset Γ \Gamma Γ is referred to as a target to be reached or searched. If the target size were to be much smaller than the distance between Γ \Gamma Γ and ∂ Ω N \partial\Omega_N ∂ Ω N , the dilation of the domain would push the boundary ∂ Ω N \partial\Omega_N ∂ Ω N to infinity, and the asymptotic behavior would be determined from the analysis of the exterior Steklov problem for the rescaled target Γ ^ = Γ / ϵ \hat{\Gamma}=\Gamma/\epsilon Γ ^ = Γ/ ϵ . In particular, even though the 1 / ϵ 1/\epsilon 1/ ϵ scaling of the eigenvalues μ k ( 0 , Γ ) \mu_k^{(0,\Gamma)} μ k ( 0 , Γ ) would remain valid, the coefficients μ ^ k \hat{\mu}_k μ ^ k would in general be different.
To illustrate this point, let us consider the annular domain Ω = { x ∈ R d : ϵ R < ∣ x ∣ < R } \Omega=\{\boldsymbol{x}\in\mathbb{R}^d:\epsilon R<|\boldsymbol{x}|<R\} Ω = { x ∈ R d : ϵ R < ∣ x ∣ < R } between two concentric spheres of radii ϵ = ϵ R \epsilon=\epsilon R ϵ = ϵ R and R R R , with the Steklov condition on the inner sphere Γ \Gamma Γ of radius ϵ \epsilon ϵ and the Neumann condition on the outer sphere ∂ Ω N \partial\Omega_N ∂ Ω N of radius R R R . The rotational invariance of this problem allows one to get the eigenvalues μ k ( 0 , Γ ) \mu_k^{(0,\Gamma)} μ k ( 0 , Γ ) explicitly (see, e.g., [20 , 111 ] ):
μ k ( 0 , Γ ) = k ε R ( k + d − 2 ) ( 1 − ε 2 k + d − 2 ) k + ( k + d − 2 ) ε 2 k + d − 2 . (105) \mu_k^{(0,\Gamma)}=\frac{k}{\varepsilon R}\frac{(k+d-2)(1-\varepsilon^{2k+d-2})}{k+(k+d-2)\varepsilon^{2k+d-2}} . \tag*{(105)} μ k ( 0 , Γ ) = εR k k + ( k + d − 2 ) ε 2 k + d − 2 ( k + d − 2 ) ( 1 − ε 2 k + d − 2 ) . (105)
When ε ≪ 1 \varepsilon\ll1 ε ≪ 1 and d = 2 d=2 d = 2 , one has μ k ( 0 , Γ ) ≈ k / ϵ \mu_k^{(0,\Gamma)}\approx k/\epsilon μ k ( 0 , Γ ) ≈ k / ϵ . In turn, the asymptotic relation (5) for the planar case with a connected boundary yields μ k ( 0 , Γ ) ≈ π k / ( 2 ϵ ) \mu_k^{(0,\Gamma)}\approx\pi k/(2\epsilon) μ k ( 0 , Γ ) ≈ π k / ( 2 ϵ ) at large k k k , where we used Eq. (90) . One sees that the scaling of μ k ( 0 , Γ ) \mu_k^{(0,\Gamma)} μ k ( 0 , Γ ) as 1 / ϵ 1/\epsilon 1/ ϵ is the same for both settings, but the prefactor differs by π / 2 \pi/2 π /2 .
Two examples
In this section, we illustrate the above general results with two examples: an arc-shaped subset Γ \Gamma Γ on the boundary of a disk (§ 4.1 ) and a spherical cap on the boundary of a ball (§ 4.2 ). In both cases, the explicit formulas for the Green’s function G ~ 0 ( x , p ∣ x 0 ) \widetilde{G}_0(x,p\mathbin{|}x_0) G 0 ( x , p ∣ x 0 ) and the pseudo-Green’s function G 0 ( x ∣ x 0 ) G_0(x\mathbin{|}x_0) G 0 ( x ∣ x 0 ) allow us to compute accurately the eigenvalues and eigenfunctions of the Steklov-Neumann problem for Γ \Gamma Γ of any size and for any p ≥ 0 p\geq0 p ≥ 0 . In this way, we can access the accuracy and the validity range of the asymptotic results deduced in § 3 .
Disk
As the first example, we consider the domain Ω = { x ∈ R 2 : ∣ x ∣ < R } \Omega=\{x\in\mathbb{R}^2: |x|<R\} Ω = { x ∈ R 2 : ∣ x ∣ < R } to be a disk of radius R R R , and set Γ = { ( R , θ ) ∈ ∂ Ω : ∣ θ ∣ < ε } \Gamma=\{(R,\theta)\in\partial\Omega: |\theta|<\varepsilon\} Γ = {( R , θ ) ∈ ∂ Ω : ∣ θ ∣ < ε } to be an arc of angle 2 ε 2\varepsilon 2 ε on the circular boundary ∂ Ω \partial\Omega ∂ Ω (Figure 1 b), where we used the polar coordinates x = ( r , θ ) x=(r,\theta) x = ( r , θ ) , and ϵ = ε R \epsilon=\varepsilon R ϵ = εR . In Appendix B.1, we recall the exact explicit expressions for the Green’s function G ~ 0 ( x , p ∣ x 0 ) \widetilde{G}_0(x,p\mathbin{|}x_0) G 0 ( x , p ∣ x 0 ) and the pseudo-Green’s function G 0 ( x ∣ x 0 ) G_0(x\mathbin{|}x_0) G 0 ( x ∣ x 0 ) , and derive the corrections w 0 ( Γ ) ( x ) w_0^{(\Gamma)}(x) w 0 ( Γ ) ( x ) and A Γ A_\Gamma A Γ for any 0 < ε ≤ π 0<\varepsilon\leq\pi 0 < ε ≤ π . These expressions allow us to determine the eigenpairs { μ k ( p , Γ ) , V k ( p , Γ ) } \{\mu_k^{(p,\Gamma)},V_k^{(p,\Gamma)}\} { μ k ( p , Γ ) , V k ( p , Γ ) } numerically for any arc Γ \Gamma Γ . A numerical solution of the spectral problems (2.33, 3.17) is described in Appendix B.2.
Figure 3 illustrates the dependence of the first four eigenvalues μ k ( 0 , Γ ) \mu_k^{(0,\Gamma)} μ k ( 0 , Γ ) on the arc half-angle ε \varepsilon ε . At ε = π \varepsilon=\pi ε = π , one retrieves the conventional Steklov problem for the disk (i.e., Γ = ∂ Ω \Gamma=\partial\Omega Γ = ∂ Ω ), for which μ k ( 0 , ∂ Ω ) = k / R \mu_k^{(0,\partial\Omega)}=k/R μ k ( 0 , ∂ Ω ) = k / R . The eigenvalue μ 0 ( 0 , Γ ) = 0 \mu_0^{(0,\Gamma)}=0 μ 0 ( 0 , Γ ) = 0 (not shown) is simple, whereas the other eigenvalues for k > 0 k>0 k > 0 are twice degenerate. In turn, as soon as ε < π \varepsilon<\pi ε < π , the degeneracy is removed, and the considered first eigenvalues for the Steklov-Neumann problem turn out to be simple. Moreover, they rapidly become close to their asymptotic form (5) , with μ ^ k \hat{\mu}_k μ ^ k reported in Table 1 . The bottom panel of Figure 3 presents the ratio between the eigenvalue and its asymptotic form (5) to highlight its approach to 1 as ε \varepsilon ε decreases. The asymptotic form gets more and more accurate as k k k increases, even for large ε \varepsilon ε . We conclude that the asymptotic relation (1.3) provides an accurate approximation for μ k ( 0 , Γ ) \mu_k^{(0,\Gamma)} μ k ( 0 , Γ ) even for moderately large ε \varepsilon ε .
Figure 3. (a) First four eigenvalues μ k ( 0 , Γ ) \mu^{(0,\Gamma)}_k μ k ( 0 , Γ ) of the Dirichlet-to-Neumann operator M ( Γ ) \mathcal{M}^{(\Gamma)} M ( Γ ) for the arc-shaped subset Γ \Gamma Γ on the boundary of the unit disk (R = 1 R=1 R = 1 ) as functions of the arc half-angle ε = ϵ / R \varepsilon=\epsilon/R ε = ϵ / R . Symbols present the results of a numerical diagonalization of the matrix G ( ϵ ) G(\epsilon) G ( ϵ ) from Eq. (B.11), with N = 30 N=30 N = 30 and k max = 20000 k_{\max}=20000 k m a x = 20000 . Lines show the asymptotic relation (5) , in which μ ^ k \hat{\mu}_k μ ^ k are given in Table 1. (b) Ratio μ k ( 0 , Γ ) / ( μ ^ k / ϵ ) \mu^{(0,\Gamma)}_k/(\hat{\mu}_k/\epsilon) μ k ( 0 , Γ ) / ( μ ^ k / ϵ ) between the eigenvalue μ k ( 0 , Γ ) \mu^{(0,\Gamma)}_k μ k ( 0 , Γ ) and its asymptotic form.
Figure 4 illustrates the behavior of the four eigenfunctions v k ( 0 , Γ ) v_k^{(0,\Gamma)} v k ( 0 , Γ ) for several values of ε \varepsilon ε . When ε = π \varepsilon=\pi ε = π (i.e., Γ = ∂ Ω \Gamma=\partial\Omega Γ = ∂ Ω ), the eigenfunctions are known explicitly as cos ( k θ ) \cos(k\theta) cos ( k θ ) and sin ( k θ ) \sin(k\theta) sin ( k θ ) . These eigenfunctions are shown by dashed line. As ε \varepsilon ε decreases, the eigenfunctions v k ( 0 , Γ ) v_k^{(0,\Gamma)} v k ( 0 , Γ ) preserve their oscillating character but deviate from sine and cosine functions. Note that even for ε = π / 2 \varepsilon=\pi/2 ε = π /2 , the eigenfunctions are very close to the limiting ones, v ^ k \hat{v}_k v ^ k , obtained by diagonalizing the matrix G \mathbf{G} G from Eq. (B.18) as the eigenfunctions of the kernel G ^ ( x ∣ x 0 ) \hat{G}(x\mathbin{|}x_0) G ^ ( x ∣ x 0 ) from Eq. (86) . As k k k increases, the eigenfunctions v ^ k \hat{v}_k v ^ k become closer to sine and cosine functions, i.e., to the case ε = π \varepsilon=\pi ε = π . In other words, higher-order eigenfunctions do not almost depend on ϵ \epsilon ϵ (up to rescaling).
Ball
We replicate the above illustration for a ball of radius R R R , Ω = { x ∈ R 3 : ∣ x ∣ < R } \Omega=\{x\in\mathbb{R}^3: |x|<R\} Ω = { x ∈ R 3 : ∣ x ∣ < R } , with Γ \Gamma Γ being the spherical cap of angle ε \varepsilon ε around the North pole: Γ = { x ∈ ∂ Ω : 0 ≤ θ < ε } \Gamma=\{x\in\partial\Omega: 0\leq\theta<\varepsilon\} Γ = { x ∈ ∂ Ω : 0 ≤ θ < ε } (Figure 1 c), where we used the spherical coordinates ( r , θ , φ ) (r,\theta,\varphi) ( r , θ , φ ) , and ϵ = ε R \epsilon=\varepsilon R ϵ = εR . The axial symmetry of the domain allows us to search the eigenfunctions in the form e i m φ V m , n ( p , Γ ) ( r , θ ) e^{im\varphi}V^{(p,\Gamma)}_{m,n}(r,\theta) e im φ V m , n ( p , Γ ) ( r , θ ) , with an integer m m m , enumerated by n = 1 , 2 , … n=1,2,\ldots n = 1 , 2 , … . In Appendix C.1, we recall the exact explicit expressions for the Green’s function G ~ 0 ( x , p ∣ x 0 ) \widetilde{G}_0(\boldsymbol{x},p|\boldsymbol{x}_0) G 0 ( x , p ∣ x 0 ) and the pseudo-Green’s function G 0 ( x ∣ x 0 ) G_0(\boldsymbol{x}|\boldsymbol{x}_0) G 0 ( x ∣ x 0 ) , and derive the corrections w ( Γ ) ( x ) w^{(\Gamma)}(\boldsymbol{x}) w ( Γ ) ( x ) and A Γ A_\Gamma A Γ . These expressions allow one to determine the eigenpairs { μ m , n ( p , Γ ) , V m , n ( p , Γ ) } \{\mu^{(p,\Gamma)}_{m,n},V^{(p,\Gamma)}_{m,n}\} { μ m , n ( p , Γ ) , V m , n ( p , Γ ) } numerically for any 0 < ϵ ≤ π 0<\epsilon\le\pi 0 < ϵ ≤ π . Throughout this section, we focus on axially symmetric eigenpairs with m = 0 m=0 m = 0 , for which the asymptotic relations (1.3) read as
μ 0 , n ( 0 , Γ ) ≈ μ ^ 0 , n ϵ , V 0 , n ( 0 , Γ ) ( R , θ , φ ) ≈ v ^ 0 , n ( θ / ϵ ) ϵ . (106) \mu^{(0,\Gamma)}_{0,n}\approx\frac{\hat{\mu}_{0,n}}{\epsilon},\qquad V^{(0,\Gamma)}_{0,n}(R,\theta,\varphi)\approx\frac{\hat{v}_{0,n}(\theta/\epsilon)}{\epsilon}. \tag*{(106)} μ 0 , n ( 0 , Γ ) ≈ ϵ μ ^ 0 , n , V 0 , n ( 0 , Γ ) ( R , θ , φ ) ≈ ϵ v ^ 0 , n ( θ / ϵ ) . (106)
A numerical solution of the spectral problems (2.33, 3.32) is described in Appendix C.2.
Figure 5(a) shows the first four eigenvalues μ 0 , n ( 0 , Γ ) \mu^{(0,\Gamma)}_{0,n} μ 0 , n ( 0 , Γ ) as functions of the angle ϵ \epsilon ϵ . Expectedly, one retrieves the well-known eigenvalues n / R n/R n / R in the classical case ϵ = π \epsilon=\pi ϵ = π when the subset Γ \Gamma Γ covers the whole boundary (i.e., Γ = ∂ Ω \Gamma=\partial\Omega Γ = ∂ Ω ). As ϵ \epsilon ϵ decreases, the asymptotic relation (4.1) rapidly becomes an accurate approximation of μ 0 , n ( 0 , Γ ) \mu^{(0,\Gamma)}_{0,n} μ 0 , n ( 0 , Γ ) . The accuracy of this approximation is illustrated on Figure 5 (b). Note that the maximal relative deviation between μ 0 , n ( 0 , Γ ) \mu^{(0,\Gamma)}_{0,n} μ 0 , n ( 0 , Γ ) and its asymptotic form μ ^ 0 , n / ϵ \hat{\mu}_{0,n}/\epsilon μ ^ 0 , n / ϵ is below 27% (see the minimum of the blue curve corresponding to n = 1 n=1 n = 1 ). The first ten eigenvalues μ ^ 0 , n \hat{\mu}_{0,n} μ ^ 0 , n are listed in Table 1.
Figure 5. (a) First four eigenvalues μ 0 , n ( 0 , Γ ) \mu_{0,n}^{(0,\Gamma)} μ 0 , n ( 0 , Γ ) of the Dirichlet-to-Neumann operator M 0 ( Γ ) \mathcal{M}_{0}^{(\Gamma)} M 0 ( Γ ) for the spherical cap Γ \Gamma Γ of angle ε \varepsilon ε on the unit sphere (R = 1 R=1 R = 1 ) as functions of the angle ε \varepsilon ε . Symbols present the results of a numerical diagonalization of the matrix G ( ε ) G^{(\varepsilon)} G ( ε ) from Eq. (C.17), with N = 100 N=100 N = 100 and k max = 1000 k_{\max}=1000 k m a x = 1000 . Lines show the asymptotic relations (106) , in which μ ^ 0 , n \widehat{\mu}_{0,n} μ 0 , n are given in Table 1 . (b) Ratio μ 0 , n ( 0 , Γ ) / ( μ ^ 0 , n / ε ) \mu_{0,n}^{(0,\Gamma)}/(\widehat{\mu}_{0,n}/\varepsilon) μ 0 , n ( 0 , Γ ) / ( μ 0 , n / ε ) between the eigenvalue μ 0 , n ( 0 , Γ ) \mu_{0,n}^{(0,\Gamma)} μ 0 , n ( 0 , Γ ) and its asymptotic form.
Figure 6 presents the related axially symmetric eigenfunctions V 0 , n ( 0 , Γ ) ( R , θ , φ ) V^{(0,\Gamma)}_{0,n}(R,\theta,\varphi) V 0 , n ( 0 , Γ ) ( R , θ , φ ) , restricted to Γ \Gamma Γ , for three values of ϵ \epsilon ϵ . When ϵ = π \epsilon=\pi ϵ = π , these eigenfunctions are the (rescaled) Legendre polynomials, n + 1 / 2 P n ( cos θ ) \sqrt{n+1/2}P_n(\cos\theta) n + 1/2 P n ( cos θ ) , as shown by blue dashed line for n = 1 , 2 , 3 , 4 n=1,2,3,4 n = 1 , 2 , 3 , 4 on four panels. As ϵ \epsilon ϵ decreases, the rescaled eigenfunctions rapidly approach v ^ 0 , n \hat{v}_{0,n} v ^ 0 , n , highlighting the accuracy of the asymptotic relation (4.1), even for ϵ = π / 2 \epsilon=\pi/2 ϵ = π /2 when the subset Γ \Gamma Γ covers half of the sphere.
Figure 6. First four axially symmetric eigenfunctions 1 − cos ε V 0 , n ( 0 , Γ ) ( R , θ , ϕ ) \sqrt{1-\cos\varepsilon}\,V_{0,n}^{(0,\Gamma)}(R,\theta,\phi) 1 − cos ε V 0 , n ( 0 , Γ ) ( R , θ , ϕ ) of the Dirichlet-to-Neumann operator M ( Γ ) M^{(\Gamma)} M ( Γ ) on the spherical cap Γ \Gamma Γ of angle ε \varepsilon ε on the unit sphere ( R = 1 ) (R=1) ( R = 1 ) , for three values of the angle ε \varepsilon ε as indicated in the legend, with y = ( 1 − cos θ ) / ( 1 − cos ε ) y=(1-\cos\theta)/(1-\cos\varepsilon) y = ( 1 − cos θ ) / ( 1 − cos ε ) . They were obtained by diagonalizing the matrix G ( ε ) G^{(\varepsilon)} G ( ε ) from Eq. (C.17), with N = 100 N=100 N = 100 and k max = 1000 k_{\max}=1000 k m a x = 1000 . Note that the factor 1 − cos ε \sqrt{1-\cos\varepsilon} 1 − cos ε ensures the correct normalization and comparable amplitudes of eigenfunctions for different ε > 0 \varepsilon>0 ε > 0 . For ε = 0 \varepsilon=0 ε = 0 , we plotted the eigenfunctions v ^ 0 , n ( y ) \hat{v}_{0,n}(\sqrt{y}) v ^ 0 , n ( y ) , that were obtained by diagonalizing the matrix G G G from Eq. (F.8) of size 30 × 30 30\times30 30 × 30 .
Mean first-reaction time
Mean first-reaction time. The eigenmodes of the Dirichlet-to-Neumann operator determine many quantities characterizing first-passage processes and related diffusion-controlled reactions [18 ] . In this section, we illustrate a straightforward application of the asymptotic results by considering restricted diffusion in a bounded domain Ω \Omega Ω and focusing on the first-reaction time τ \tau τ on a reactive target Γ ⊂ ∂ Ω \Gamma\subset\partial\Omega Γ ⊂ ∂ Ω [7 , 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 ] . The moments and the distribution of the random variable τ \tau τ were thoroughly investigated in the past, especially in the narrow escape limit [62 , 63 , 64 , 65 , 66 , 67 , 68 , 69 , 70 , 71 , 72 , 73 , 74 , 75 , 76 , 77 , 78 , 79 , 80 , 81 , 82 , 83 , 84 , 85 , 86 ] . Some of these former results will be used for comparison with our predictions. Most importantly, the knowledge of the eigenpairs of the Dirichlet-to-Neumann operator allows us to go beyond the classical setting and to investigate first-reaction times for more sophisticated surface reaction mechanisms on Γ \Gamma Γ , including activation or passivation of the target, encounter-dependent reactivity, etc. [18 ] We start with the classical setting and then briefly describe its extensions.
Figure 4. First four eigenfunctions ε v 1 ( 0 , Γ ) \sqrt{\varepsilon}\,v_1^{(0,\Gamma)} ε v 1 ( 0 , Γ ) of the Dirichlet-to-Neumann operator M 0 ( Γ ) \mathcal{M}_0^{(\Gamma)} M 0 ( Γ ) for the arc-shaped subset Γ \Gamma Γ on the boundary of the unit disk (R = 1 R = 1 R = 1 ), for three values of the arc half-angle ε \varepsilon ε as indicated in the legend. They were obtained by diagonalizing the matrix G ( ε ) \mathbf{G}^{(\varepsilon)} G ( ε ) from Eq. (B.11), with N = 30 N = 30 N = 30 and k max = 20000 k_{\max} = 20000 k m a x = 20000 . Note that the factor ε \sqrt{\varepsilon} ε ensures the correct normalization and comparable amplitudes of eigenfunctions for different ε > 0 \varepsilon> 0 ε > 0 . For ε = 0 \varepsilon= 0 ε = 0 , we plotted the eigenfunctions v ^ k ( x ) \hat{v}_k(x) v ^ k ( x ) that were obtained by diagonalizing the matrix G \mathbf{G} G from Eq. (B.18) of size 30 × 30 30 \times30 30 × 30 .
Constant reactivity
For a diffusing particle started from a point x 0 ∈ Ω x_0 \in\Omega x 0 ∈ Ω at time 0, we are interested in the first-reaction time τ \tau τ on a target Γ ⊂ ∂ Ω \Gamma\subset\partial\Omega Γ ⊂ ∂ Ω with a constant reactivity parameter q > 0 q > 0 q > 0 . The diffusing particle executes a partially reflected Brownian motion inside Ω \Omega Ω , with normal reflections on ∂ Ω N \partial\Omega_N ∂ Ω N and eventual reactions on Γ \Gamma Γ [18 , 114 ] . The distribution of the random variable τ \tau τ is characterized by the survival probability S q ( t ∣ x 0 ) = P x 0 { τ > t } S_q(t|x_0) = \mathbb{P}_{x_0}\{\tau> t\} S q ( t ∣ x 0 ) = P x 0 { τ > t } , i.e., the probability of no reaction on Γ \Gamma Γ up to time t t t . The survival probability S q ( t ∣ x 0 ) S_q(t|x_0) S q ( t ∣ x 0 ) satisfies the diffusion equation with mixed Robin-Neumann boundary conditions:
∂ t S q ( t ∣ x 0 ) = D Δ S q ( t ∣ x 0 ) ( x 0 ∈ Ω ) , (107) \partial_t S_q(t|x_0) = D\Delta S_q(t|x_0) \qquad(x_0 \in\Omega), \tag*{(107)} ∂ t S q ( t ∣ x 0 ) = D Δ S q ( t ∣ x 0 ) ( x 0 ∈ Ω ) , (107)
− ∂ n S q ( t ∣ x 0 ) = q S q ( t ∣ x 0 ) I Γ ( x 0 ) ( x 0 ∈ ∂ Ω ) , (108) -\partial_n S_q(t|x_0) = q S_q(t|x_0)\mathbb{I}_{\Gamma}(x_0) \qquad(x_0 \in\partial\Omega), \tag*{(108)} − ∂ n S q ( t ∣ x 0 ) = q S q ( t ∣ x 0 ) I Γ ( x 0 ) ( x 0 ∈ ∂ Ω ) , (108)
subject to the initial condition S q ( 0 ∣ x 0 ) = 1 S_q(0|x_0) = 1 S q ( 0∣ x 0 ) = 1 (we recall that I Γ ( x 0 ) \mathbb{I}_{\Gamma}(x_0) I Γ ( x 0 ) is the indicator function of $\Gamma). The notion of partial reactivity was introduced by Collins and Kimball [112 ] and later investigated in various contexts (see [6 , 113 , 114 , 115 , 116 , 117 ] and references therein). When the particle arrives onto a partially reactive target, it can either react or be reflected to resume its diffusion, until the next arrival on Γ \Gamma Γ , and so on. Depending on the considered application, such reflections can represent the arrival at a passive boundary point (due to either microscopic heterogeneity of reactive sites on Γ \Gamma Γ or their temporal activation/passivation dynamics), a failure to overcome an energy activation barrier in a chemical reaction, a failure to squeeze through a narrow channel to escape, etc. In this way, the parameter q q q is related to the probability of the reaction event and thus characterizes the reactivity of the target Γ \Gamma Γ , by ranging from 0 0 0 for an inert, passive target (without any reaction), to + ∞ +\infty + ∞ for a perfectly reactive target, on which the particle reacts instantly upon the first arrival [6 ] .
The Laplace transform of the survival probability,
S ~ q ( p ∣ x 0 ) = ∫ 0 ∞ d t e − p t S q ( t ∣ x 0 ) ( p ≥ 0 ) , (109) \widetilde{S}_q(p|x_0)=\int_{0}^{\infty}dt\,e^{-pt}S_q(t|x_0)\qquad(p\geq0), \tag*{(109)} S q ( p ∣ x 0 ) = ∫ 0 ∞ d t e − pt S q ( t ∣ x 0 ) ( p ≥ 0 ) , (109)
satisfies the modified Helmholtz equation with mixed Robin-Neumann boundary conditions,
( p − D Δ ) S ~ q ( p ∣ x 0 ) = 1 ( x 0 ∈ Ω ) , (110) (p-D\Delta)\widetilde{S}_q(p|x_0)=1\qquad(x_0\in\Omega), \tag*{(110)} ( p − D Δ ) S q ( p ∣ x 0 ) = 1 ( x 0 ∈ Ω ) , (110)
∂ n S ~ q ( p ∣ x 0 ) + q S ~ q ( p ∣ x 0 ) I Γ ( x 0 ) = 0 ( x 0 ∈ ∂ Ω ) . (111) \partial_n\widetilde{S}_q(p|x_0)+q\widetilde{S}_q(p|x_0)I_{\Gamma}(x_0)=0\qquad(x_0\in\partial\Omega). \tag*{(111)} ∂ n S q ( p ∣ x 0 ) + q S q ( p ∣ x 0 ) I Γ ( x 0 ) = 0 ( x 0 ∈ ∂ Ω ) . (111)
As this problem can be obtained by integrating Eq. (2.6) over x ∈ Ω x\in\Omega x ∈ Ω , one has
S ~ q ( p ∣ x 0 ) = ∫ Ω d x G ~ q ( x , p ∣ x 0 ) , (112) \widetilde{S}_q(p|x_0)=\int_{\Omega}dx\,\widetilde{G}_q(x,p|x_0), \tag*{(112)} S q ( p ∣ x 0 ) = ∫ Ω d x G q ( x , p ∣ x 0 ) , (112)
so that the spectral expansion (14) for the Green’s function yields after some simplifications (see [17 , 18 ] for more details):
S ~ q ( p ∣ x 0 ) = 1 p − 1 p ∑ k = 0 ∞ V k ( p , Γ ) ( x 0 ) 1 + μ k ( p , Γ ) / q ∫ Γ v k ( p , Γ ) ( x 0 ∈ Ω ‾ , q > 0 ) . (113) \widetilde{S}_q(p|x_0)=\frac{1}{p}-\frac{1}{p}\sum_{k=0}^{\infty}\frac{V_k^{(p,\Gamma)}(x_0)}{1+\mu_k^{(p,\Gamma)}/q}\int_{\Gamma}v_k^{(p,\Gamma)}\qquad(x_0\in\overline{\Omega},\ q>0). \tag*{(113)} S q ( p ∣ x 0 ) = p 1 − p 1 k = 0 ∑ ∞ 1 + μ k ( p , Γ ) / q V k ( p , Γ ) ( x 0 ) ∫ Γ v k ( p , Γ ) ( x 0 ∈ Ω , q > 0 ) . (113)
Since the negative time derivative of the survival probability determines the probability density of the first-reaction time, H q ( t ∣ x 0 ) = − ∂ t S q ( t ∣ x 0 ) H_q(t|x_0)=-\partial_tS_q(t|x_0) H q ( t ∣ x 0 ) = − ∂ t S q ( t ∣ x 0 ) , its Laplace transform reads
H ~ q ( p ∣ x 0 ) = 1 − p S ~ q ( p ∣ x 0 ) = ∑ k = 0 ∞ V k ( p , Γ ) ( x 0 ) 1 + μ k ( p , Γ ) / q ∫ Γ v k ( p , Γ ) (114) \widetilde{H}_q(p|x_0)=1-p\widetilde{S}_q(p|x_0)=\sum_{k=0}^{\infty}\frac{V_k^{(p,\Gamma)}(x_0)}{1+\mu_k^{(p,\Gamma)}/q}\int_{\Gamma}v_k^{(p,\Gamma)} \tag*{(114)} H q ( p ∣ x 0 ) = 1 − p S q ( p ∣ x 0 ) = k = 0 ∑ ∞ 1 + μ k ( p , Γ ) / q V k ( p , Γ ) ( x 0 ) ∫ Γ v k ( p , Γ ) (114)
( x 0 ∈ Ω ‾ , q > 0 ) (x_0\in\overline{\Omega},\ q>0) ( x 0 ∈ Ω , q > 0 ) .
By definition, this is the moment-generating function of the random time τ \tau τ :
H ~ q ( p ∣ x 0 ) = ∫ 0 ∞ d t e − p t H q ( t ∣ x 0 ) = E x 0 { e − p τ } . (115) \widetilde{H}_q(p|x_0)=\int_0^{\infty}dt\,e^{-pt}H_q(t|x_0)=\mathbb{E}_{x_0}\{e^{-p\tau}\}. \tag*{(115)} H q ( p ∣ x 0 ) = ∫ 0 ∞ d t e − pt H q ( t ∣ x 0 ) = E x 0 { e − p τ } . (115)
In the following, we focus on the MFRT:
T q ( x 0 ) = E x 0 { τ } = ∫ 0 ∞ d t t H q ( t ∣ x 0 ) = = − ∂ t S q ( t ∣ x 0 ) ∫ 0 ∞ d t S q ( t ∣ x 0 ) = S ~ q ( 0 ∣ x 0 ) , (116) T_q(x_0)=\mathbb{E}_{x_0}\{\tau\}=\int_0^{\infty}dt\,tH_q(t|x_0)\underset{=-\partial_tS_q(t|x_0)}{=}\int_0^{\infty}dt\,S_q(t|x_0)=\widetilde{S}_q(0|x_0), \tag*{(116)} T q ( x 0 ) = E x 0 { τ } = ∫ 0 ∞ d t t H q ( t ∣ x 0 ) = − ∂ t S q ( t ∣ x 0 ) = ∫ 0 ∞ d t S q ( t ∣ x 0 ) = S q ( 0∣ x 0 ) , (116)
which satisfies the boundary value problem (2.45). In the limit q → ∞ q\to\infty q → ∞ , the Robin boundary condition on Γ \Gamma Γ turns into the Dirichlet one, and T ∞ ( x 0 ) T_{\infty}(x_0) T ∞ ( x 0 ) becomes the mean first-passage time to Γ \Gamma Γ .
To compute the MFRT, we evaluate a series expansion of H ~ q ( p ∣ x 0 ) \widetilde{H}_q(p|x_0) H q ( p ∣ x 0 ) in powers of p p p up to the first order (or, equivalently, the series expansion of S ~ q ( p ∣ x 0 ) \widetilde{S}_q(p|x_0) S q ( p ∣ x 0 ) up to the zeroth order). Since v k ( 0 , Γ ) v_k^{(0,\Gamma)} v k ( 0 , Γ ) for any k > 0 k>0 k > 0 is orthogonal to the constant function v 0 ( 0 , Γ ) = 1 / ∣ Γ ∣ v_0^{(0,\Gamma)}=1/\sqrt{|\Gamma|} v 0 ( 0 , Γ ) = 1/ ∣Γ∣ , the integral of v k ( p , Γ ) v_k^{(p,\Gamma)} v k ( p , Γ ) over Γ \Gamma Γ vanishes as p → 0 p\to0 p → 0 . In turn, we use the series expansions (2.21) for the eigenmode with k = 0 k=0 k = 0 to get as p → 0 p\to0 p → 0 :
H ~ q ( p ∣ x 0 ) = 1 + W 0 ( Γ ) ( x 0 ) p ∣ Ω ∣ D 1 + p ∣ Ω ∣ q ∣ Γ ∣ ( 1 + p ∣ Ω ∣ D ∣ Γ ∣ ∫ Γ d x w 0 ( Γ ) ( x ) ) + ∑ k = 1 ∞ V k ( 0 , Γ ) ( x 0 ) 1 + μ k ( 0 , Γ ) / q ∫ Γ d x v k ( p , Γ ) ( x ) + O ( p 2 ) , \widetilde{H}_q(p|x_0)=\frac{1+W_0^{(\Gamma)}(x_0)\frac{p|\Omega|}{D}}{1+\frac{p|\Omega|}{q|\Gamma|}}\left(1+\frac{p|\Omega|}{D|\Gamma|}\int_{\Gamma}dx\,w_0^{(\Gamma)}(x)\right)+\sum_{k=1}^{\infty}\frac{V_k^{(0,\Gamma)}(x_0)}{1+\mu_k^{(0,\Gamma)}/q}\int_{\Gamma}dx\,v_k^{(p,\Gamma)}(x)+O(p^2), H q ( p ∣ x 0 ) = 1 + q ∣Γ∣ p ∣Ω∣ 1 + W 0 ( Γ ) ( x 0 ) D p ∣Ω∣ ( 1 + D ∣Γ∣ p ∣Ω∣ ∫ Γ d x w 0 ( Γ ) ( x ) ) + k = 1 ∑ ∞ 1 + μ k ( 0 , Γ ) / q V k ( 0 , Γ ) ( x 0 ) ∫ Γ d x v k ( p , Γ ) ( x ) + O ( p 2 ) ,
∫ Γ d x v k ( p , Γ ) ( x ) ⏟ = p b k ( Γ ) ∣ Ω ∣ / ( D μ k ( 0 , Γ ) ) + O ( p 2 ) \underbrace{\int_{\Gamma}dx\,v_k^{(p,\Gamma)}(x)}_{=p b_k^{(\Gamma)}|\Omega|/(D\mu_k^{(0,\Gamma)})+O(p^2)} = p b k ( Γ ) ∣Ω∣/ ( D μ k ( 0 , Γ ) ) + O ( p 2 ) ∫ Γ d x v k ( p , Γ ) ( x )
from which
T q ( x 0 ) = ∣ Ω ∣ D ( 1 q ∣ Γ ∣ − W 0 ( Γ ) ( x 0 ) − ∑ k = 1 ∞ b k ( Γ ) V k ( 0 , Γ ) ( x 0 ) μ k ( 0 , Γ ) ( 1 + μ k ( 0 , Γ ) / q ) ) , ( x 0 ∈ Ω ‾ ) , (117) T_q(x_0)=\frac{|\Omega|}{D}\left(\frac{1}{q|\Gamma|}-W_0^{(\Gamma)}(x_0)-\sum_{k=1}^{\infty}\frac{b_k^{(\Gamma)}V_k^{(0,\Gamma)}(x_0)}{\mu_k^{(0,\Gamma)}(1+\mu_k^{(0,\Gamma)}/q)}\right),\qquad(x_0\in\overline{\Omega}), \tag*{(117)} T q ( x 0 ) = D ∣Ω∣ ( q ∣Γ∣ 1 − W 0 ( Γ ) ( x 0 ) − k = 1 ∑ ∞ μ k ( 0 , Γ ) ( 1 + μ k ( 0 , Γ ) / q ) b k ( Γ ) V k ( 0 , Γ ) ( x 0 ) ) , ( x 0 ∈ Ω ) , (117)
where we used (53) , whereas the integral of w 0 ( Γ ) w_0^{(\Gamma)} w 0 ( Γ ) vanished according to (37) . Solving the boundary value problem (2.37), one can find W 0 ( Γ ) ( x 0 ) W_0^{(\Gamma)}(x_0) W 0 ( Γ ) ( x 0 ) that determines the coefficients b k ( Γ ) b_k^{(\Gamma)} b k ( Γ ) via (51) , while V k ( 0 , Γ ) ( x 0 ) V_k^{(0,\Gamma)}(x_0) V k ( 0 , Γ ) ( x 0 ) are given by (45) . By recalling the relation (2.43), one can also represent the above expression as
T q ( x 0 ) = T q ( a p p ) ( x 0 ) − ∣ Ω ∣ D ∑ k = 1 ∞ b k ( Γ ) V k ( 0 , Γ ) ( x 0 ) μ k ( 0 , Γ ) ( 1 + μ k ( 0 , Γ ) / q ) , (118) T_q(x_0)=T_q^{(\mathrm{app})}(x_0)-\frac{|\Omega|}{D}\sum_{k=1}^{\infty}\frac{b_k^{(\Gamma)}V_k^{(0,\Gamma)}(x_0)}{\mu_k^{(0,\Gamma)}(1+\mu_k^{(0,\Gamma)}/q)}, \tag*{(118)} T q ( x 0 ) = T q ( app ) ( x 0 ) − D ∣Ω∣ k = 1 ∑ ∞ μ k ( 0 , Γ ) ( 1 + μ k ( 0 , Γ ) / q ) b k ( Γ ) V k ( 0 , Γ ) ( x 0 ) , (118)
where T q ( a p p ) ( x 0 ) T_q^{(\mathrm{app})}(x_0) T q ( app ) ( x 0 ) is the constant-flux approximation to the MFRT [83 ] . In other words, the spectral expansion in the second term provides the correction to this approximation. The exact expression (5.10) determines the dependence of the MFRT on the reactivity parameter q q q over the entire range of 0 < q ≤ + ∞ 0<q\le+\infty 0 < q ≤ + ∞ .
It is also instructive to consider the volume-averaged MFRT,
T ‾ q = 1 ∣ Ω ∣ ∫ Ω d x 0 T q ( x 0 ) , (119) \overline{T}_q=\frac{1}{|\Omega|}\int_{\Omega}dx_0\,T_q(x_0), \tag*{(119)} T q = ∣Ω∣ 1 ∫ Ω d x 0 T q ( x 0 ) , (119)
as if the starting point x 0 x_0 x 0 was uniformly distributed in Ω \Omega Ω . Using Eqs. (2.36, 2.39), we get then
T ‾ q = ∣ Ω ∣ D ( 1 q ∣ Γ ∣ + A Γ − ∑ k = 1 ∞ [ b k ( Γ ) ] 2 μ k ( 0 , Γ ) ( 1 + μ k ( 0 , Γ ) / q ) ) . (120) \overline{T}_q=\frac{|\Omega|}{D}\left(\frac{1}{q|\Gamma|}+A_{\Gamma}-\sum_{k=1}^{\infty}\frac{[b_k^{(\Gamma)}]^2}{\mu_k^{(0,\Gamma)}(1+\mu_k^{(0,\Gamma)}/q)}\right). \tag*{(120)} T q = D ∣Ω∣ ( q ∣Γ∣ 1 + A Γ − k = 1 ∑ ∞ μ k ( 0 , Γ ) ( 1 + μ k ( 0 , Γ ) / q ) [ b k ( Γ ) ] 2 ) . (120)
In the limit q → ∞ q\to\infty q → ∞ , we find the volume-averaged MFPT to the target Γ \Gamma Γ :
T ‾ ∞ = ∣ Ω ∣ D ( A Γ − ∑ k = 1 ∞ [ b k ( Γ ) ] 2 μ k ( 0 , Γ ) ) . (121) \overline{T}_{\infty}=\frac{|\Omega|}{D}\left(A_{\Gamma}-\sum_{k=1}^{\infty}\frac{[b_k^{(\Gamma)}]^2}{\mu_k^{(0,\Gamma)}}\right). \tag*{(121)} T ∞ = D ∣Ω∣ ( A Γ − k = 1 ∑ ∞ μ k ( 0 , Γ ) [ b k ( Γ ) ] 2 ) . (121)
While these exact representations are valid for arbitrary bounded domains with smooth boundaries, they require the knowledge of the Steklov eigenfunctions and eigenvalues. In the next section, we discuss their consequences in the small-target limit.
Small-target limit
As stated earlier, the small-target limit of the MFPT was thoroughly investigated [62 , 63 , 64 , 65 , 66 , 67 , 68 , 69 , 70 , 71 , 72 , 73 , 74 , 75 , 76 , 77 , 78 , 79 , 80 , 81 , 82 , 83 , 84 , 85 , 86 ] . As the detailed discussion of this topic can be found elsewhere (see, e.g., the review [11 ] ), we just outline some novel insights onto the volume-averaged quantities T ‾ ∞ \overline{T}_{\infty} T ∞ and T ‾ q \overline{T}_{q} T q from our asymptotic results.
Two dimensions. For planar bounded domains, we substitute the derived asymptotic relations (1.3, 3.10, 3.13) into Eq. (120) to get in the small-target limit ϵ → 0 \epsilon\to0 ϵ → 0 :
T ‾ q ≈ ∣ Ω ∣ D ( 1 2 q ϵ − ln ( 2 ϵ ) π + 3 2 π + R 0 ( x Γ ) − ∑ k = 1 ∞ [ V ^ k ( ∞ ) ] 2 1 + μ ^ k / ( q ϵ ) ) . (122) \overline{T}_{q} \approx\frac{|\Omega|}{D}\left(\frac{1}{2q\epsilon}-\frac{\ln(2\epsilon)}{\pi}+\frac{3}{2\pi}+R_{0}(x_{\Gamma})-\sum_{k=1}^{\infty}\frac{[\widehat{V}_{k}(\infty)]^{2}}{1+\widehat{\mu}_{k}/(q\epsilon)}\right). \tag*{(122)} T q ≈ D ∣Ω∣ ( 2 q ϵ 1 − π ln ( 2 ϵ ) + 2 π 3 + R 0 ( x Γ ) − k = 1 ∑ ∞ 1 + μ k / ( q ϵ ) [ V k ( ∞ ) ] 2 ) . (122)
Let us briefly comment on this asymptotic relation.
(i) for a perfectly reactive target (q = ∞ q=\infty q = ∞ ), this relation is reduced to
T ‾ ∞ ≈ ∣ Ω ∣ π D ( ln ( 2 / ϵ ) + π R 0 ( x Γ ) + o ( 1 ) ) , (123) \overline{T}_{\infty} \approx\frac{|\Omega|}{\pi D}\left(\ln(2/\epsilon)+\pi R_{0}(x_{\Gamma})+o(1)\right), \tag*{(123)} T ∞ ≈ π D ∣Ω∣ ( ln ( 2/ ϵ ) + π R 0 ( x Γ ) + o ( 1 ) ) , (123)
where we used Eq. (G.11) to evaluate the sum. The logarithmic leading-order term, which was derived by Singer et al. [66 ] , is universal (see further discussions on a more general form of the small parameter in [81 ] ). In turn, the next-order term O ( 1 ) O(1) O ( 1 ) depends on the confining domain and the location x Γ x_{\Gamma} x Γ of the target through the regular part R 0 ( x Γ ) R_{0}(x_{\Gamma}) R 0 ( x Γ ) of the pseudo-Green’s function. For instance, for the disk of radius R R R with the arc-shaped target Γ \Gamma Γ of angle 2 ϵ 2\epsilon 2 ϵ , one has R 0 ( x Γ ) = 1 / ( 8 π ) R_{0}(x_{\Gamma})=1/(8\pi) R 0 ( x Γ ) = 1/ ( 8 π ) and thus retrieves the asymptotic result from [67 ] .
(ii) For a partially reactive target (0 < q < ∞ 0<q<\infty 0 < q < ∞ ), the situation is drastically different. In fact, the dominant contribution to Eq. (122) comes from the first term ∣ Ω ∣ / ( 2 q D ϵ ) |\Omega|/(2qD\epsilon) ∣Ω∣/ ( 2 q D ϵ ) , which diverges as 1 / ϵ 1/\epsilon 1/ ϵ . The crucial role of partial reactivity in the narrow escape problem was thoroughly discussed in [83 ] . When q q q is fixed, the sum in Eq. (122) vanishes as O ( ϵ ) O(\epsilon) O ( ϵ ) and can thus be neglected. However, if q ϵ q\epsilon q ϵ is fixed (i.e., if q q q grows as ϵ → 0 \epsilon\to0 ϵ → 0 ), one gets a different scaling. Moreover, one can fix ϵ \epsilon ϵ small enough and study the dependence of the MFRT on the reactivity parameter q q q , in which case Eq. (122) provides a universal dependence of ( T ‾ q − T ‾ ∞ ) / ∣ Ω ∣ (\overline{T}_{q}-\overline{T}_{\infty})/|\Omega| ( T q − T ∞ ) /∣Ω∣ on q ϵ q\epsilon q ϵ , regardless of the domain:
T ‾ q − T ‾ ∞ ≈ ∣ Ω ∣ D Ψ 2 ( q ϵ ) , Ψ 2 ( z ) = 1 2 z + ∑ k = 1 ∞ [ V ^ k ( ∞ ) ] 2 μ ^ k + z . (124) \overline{T}_{q}-\overline{T}_{\infty}\approx\frac{|\Omega|}{D}\Psi_{2}(q\epsilon),\qquad\Psi_{2}(z)=\frac{1}{2z}+\sum_{k=1}^{\infty}\frac{[\widehat{V}_{k}(\infty)]^{2}}{\widehat{\mu}_{k}+z}. \tag*{(124)} T q − T ∞ ≈ D ∣Ω∣ Ψ 2 ( q ϵ ) , Ψ 2 ( z ) = 2 z 1 + k = 1 ∑ ∞ μ k + z [ V k ( ∞ ) ] 2 . (124)
This unique function monotonously decreases from + ∞ +\infty + ∞ at z = 0 z=0 z = 0 to 0 0 0 as z → ∞ z\to\infty z → ∞ , and thus characterizes the impact of partial reactivity onto the MFRT on a small target. At small z z z , the dominant contribution comes from the first term, 1 / ( 2 z ) 1/(2z) 1/ ( 2 z ) , whereas the sum gives higher-order corrections in z = q ϵ z=q\epsilon z = q ϵ :
z Ψ 2 ( z ) ≈ 1 2 + ∑ n = 1 ∞ C n z n , C n = ∑ k = 1 ∞ [ V ^ k ( ∞ ) ] 2 [ μ ^ k ] n . (125) z\Psi_{2}(z)\approx\frac{1}{2}+\sum_{n=1}^{\infty}C_{n}z^{n},\qquad C_{n}=\sum_{k=1}^{\infty}\frac{[\widehat{V}_{k}(\infty)]^{2}}{[\widehat{\mu}_{k}]^{n}}. \tag*{(125)} z Ψ 2 ( z ) ≈ 2 1 + n = 1 ∑ ∞ C n z n , C n = k = 1 ∑ ∞ [ μ k ] n [ V k ( ∞ ) ] 2 . (125)
The first two coefficients C 1 C_{1} C 1 and C 2 C_{2} C 2 are given by Eqs. (G.11) and (G.9), respectively. Comparing Eq. (124) with Eqs. [2,3] from [86 ] , we get the asymptotic behavior of the function Ψ 2 ( z ) \Psi_{2}(z) Ψ 2 ( z ) at large z z z :
z Ψ 2 ( z ) ≈ 1 π 2 ( ln ( 8 z ) + γ + 1 ) ( z → ∞ ) , (126) z\Psi_{2}(z)\approx\frac{1}{\pi^{2}}\left(\ln(8z)+\gamma+1\right)\qquad(z\to\infty), \tag*{(126)} z Ψ 2 ( z ) ≈ π 2 1 ( ln ( 8 z ) + γ + 1 ) ( z → ∞ ) , (126)
where γ ≈ 0.5772 \gamma\approx0.5772 γ ≈ 0.5772 is the Euler constant. Figure 7 (a) presents the function Ψ 2 ( z ) \Psi_{2}(z) Ψ 2 ( z ) and its asymptotic behavior (126) , which are in excellent agreement.
Figure 7. (a) The function z Ψ 2 ( z ) z\Psi_2(z) z Ψ 2 ( z ) from Eq. (124) , shown by solid line, and its large-z z z asymptotic behavior (5.18), shown by dashed line. (b) The function z Ψ 3 ( z ) z\Psi_3(z) z Ψ 3 ( z ) from Eq. (131) , shown by solid line, and its large-z z z asymptotic behavior (5.25), shown by dashed line.
Three dimensions. In three dimensions, we substitute the asymptotic relations (1.3, 3.21, 3.25) into Eq. (120) to get in the small-target limit:
T ‾ q ≈ ∣ Ω ∣ π D ( 1 q ϵ 2 + 8 3 π ϵ + H ( x Γ ) 4 ln ( 1 / ϵ ) − π ϵ ∑ k = 1 ∞ [ V ^ k ( ∞ ) ] 2 μ ^ k ( 1 + μ ^ k / ( q ϵ ) ) + O ( 1 ) ) . (127) \overline{T}_q \approx\frac{|\Omega|}{\pi D}\left(\frac{1}{q\epsilon^2}+\frac{8}{3\pi\epsilon}+\frac{H(x_\Gamma)}{4}\ln(1/\epsilon)-\frac{\pi}{\epsilon}\sum_{k=1}^{\infty}\frac{[\widehat{V}_k(\infty)]^2}{\widehat{\mu}_k\left(1+\widehat{\mu}_k/(q\epsilon)\right)}+O(1)\right). \tag*{(127)} T q ≈ π D ∣Ω∣ ( q ϵ 2 1 + 3 π ϵ 8 + 4 H ( x Γ ) ln ( 1/ ϵ ) − ϵ π k = 1 ∑ ∞ μ k ( 1 + μ k / ( q ϵ ) ) [ V k ( ∞ ) ] 2 + O ( 1 ) ) . (127)
As previously, we distinguish two cases:
for a perfectly reactive target (q = ∞ q=\infty q = ∞ ), one gets
T ‾ ∞ ≈ ∣ Ω ∣ π D ( 8 3 π ϵ + H ( x Γ ) 4 ln ( 1 / ϵ ) − π ϵ ∑ k = 1 ∞ [ V ^ k ( ∞ ) ] 2 μ ^ k + O ( 1 ) ) . (128) \overline{T}_{\infty} \approx\frac{|\Omega|}{\pi D}\left(\frac{8}{3\pi\epsilon}+\frac{H(x_\Gamma)}{4}\ln(1/\epsilon)-\frac{\pi}{\epsilon}\sum_{k=1}^{\infty}\frac{[\widehat{V}_k(\infty)]^2}{\widehat{\mu}_k}+O(1)\right). \tag*{(128)} T ∞ ≈ π D ∣Ω∣ ( 3 π ϵ 8 + 4 H ( x Γ ) ln ( 1/ ϵ ) − ϵ π k = 1 ∑ ∞ μ k [ V k ( ∞ ) ] 2 + O ( 1 ) ) . (128)
While the term A Γ \mathcal{A}_\Gamma A Γ provided the dominant contribution to the volume-averaged MFPT in the planar case, this is not true anymore in three dimensions; indeed, the leading-order scaling 1 / ϵ 1/\epsilon 1/ ϵ comes from both A Γ \mathcal{A}_\Gamma A Γ and the spectral expansion. This distinction explains why the constant-flux approximation from [83 ] that ignores the spectral expansion, was more accurate in two dimensions than in three dimensions. Using Eq. (G.33) to evaluate the sum in Eq. (128) , we get
T ‾ ∞ ≈ ∣ Ω ∣ 4 D ( 1 ϵ + H ( x Γ ) π ln ( 1 / ϵ ) + O ( 1 ) ) . (129) \overline{T}_{\infty} \approx\frac{|\Omega|}{4D}\left(\frac{1}{\epsilon}+\frac{H(x_\Gamma)}{\pi}\ln(1/\epsilon)+O(1)\right). \tag*{(129)} T ∞ ≈ 4 D ∣Ω∣ ( ϵ 1 + π H ( x Γ ) ln ( 1/ ϵ ) + O ( 1 ) ) . (129)
The leading term of this expression goes back to Lord Rayleigh [118 ] (see [66 ] for more discussions). As expected, the curvature of the boundary appears in the subleading, logarithmic term.
for a partially reactive target (0 < q < ∞ 0<q<\infty 0 < q < ∞ ), Eq. (127) can be rewritten as
T ‾ q ≈ ∣ Ω ∣ π D ( 1 q ϵ 2 + 8 3 π ϵ + H ( x Γ ) 4 ln ( 1 / ϵ ) − q π ∑ k = 1 ∞ [ V ^ k ( ∞ ) ] 2 μ ^ k 2 + q 2 ϵ π ∑ k = 1 ∞ [ V ^ k ( ∞ ) ] 2 / μ ^ k q ϵ + μ ^ k + O ( 1 ) ) . (130) \begin{aligned}
\overline{T}_q \approx{}& \frac{|\Omega|}{\pi D}\left(\frac{1}{q\epsilon^2}+\frac{8}{3\pi\epsilon}+\frac{H(x_\Gamma)}{4}\ln(1/\epsilon)\right.\\
&\left.{}-q\pi\sum_{k=1}^{\infty}\frac{[\widehat{V}_k(\infty)]^2}{\widehat{\mu}_k^2}+q^2\epsilon\pi\sum_{k=1}^{\infty}\frac{[\widehat{V}_k(\infty)]^2/\widehat{\mu}_k}{q\epsilon+\widehat{\mu}_k}+O(1)\right). \tag*{(130)}
\end{aligned} T q ≈ π D ∣Ω∣ ( q ϵ 2 1 + 3 π ϵ 8 + 4 H ( x Γ ) ln ( 1/ ϵ ) − q π k = 1 ∑ ∞ μ k 2 [ V k ( ∞ ) ] 2 + q 2 ϵ π k = 1 ∑ ∞ q ϵ + μ k [ V k ( ∞ ) ] 2 / μ k + O ( 1 ) ) . (130)
When q q q is fixed, the last two terms are of the order O ( 1 ) O(1) O ( 1 ) and O ( ϵ ) O(\epsilon) O ( ϵ ) , respectively. In turn, the first three terms provide the leading and subleading contributions to T ‾ q \overline{T}_q T q . Moreover, as the second and the third terms came from A Γ \mathcal{A}_\Gamma A Γ , they are accessible within the constant-flux approximation, which turns out to be much more accurate for a partially reactive target. As previously, we obtain the universal dependence of ( T ‾ q − T ‾ ∞ ) / ∣ Ω ∣ (\overline{T}_q-\overline{T}_\infty)/|\Omega| ( T q − T ∞ ) /∣Ω∣ on q ϵ q\epsilon q ϵ , regardless of the domain:
T ‾ q − T ‾ ∞ ∣ Ω ∣ ≈ 1 ϵ D Ψ 3 ( q ϵ ) , Ψ 3 ( z ) = 1 z π + ∑ k = 1 ∞ [ V ^ k ( ∞ ) ] 2 z + μ ^ k . (131) \frac{\overline{T}_q-\overline{T}_\infty}{|\Omega|}\approx\frac{1}{\epsilon D}\Psi_3(q\epsilon),\qquad\Psi_3(z)=\frac{1}{z\pi}+\sum_{k=1}^{\infty}\frac{[\widehat{V}_k(\infty)]^2}{z+\widehat{\mu}_k}. \tag*{(131)} ∣Ω∣ T q − T ∞ ≈ ϵD 1 Ψ 3 ( q ϵ ) , Ψ 3 ( z ) = z π 1 + k = 1 ∑ ∞ z + μ k [ V k ( ∞ ) ] 2 . (131)
At small z z z , the dominant contribution comes from the first term, 1 / ( z π ) 1/(z\pi) 1/ ( z π ) , whereas the sum gives higher-order corrections in z = q ϵ z=q\epsilon z = q ϵ :
z Ψ 3 ( z ) ≈ 1 π + ∑ n = 1 ∞ C n z n , C n = ∑ k = 1 ∞ [ V ^ k ( ∞ ) ] 2 [ μ ^ k ] n . (132) z\Psi_3(z)\approx\frac{1}{\pi}+\sum_{n=1}^{\infty}C_nz^n,\qquad C_n=\sum_{k=1}^{\infty}\frac{[\widehat{V}_k(\infty)]^2}{[\widehat{\mu}_k]^n}. \tag*{(132)} z Ψ 3 ( z ) ≈ π 1 + n = 1 ∑ ∞ C n z n , C n = k = 1 ∑ ∞ [ μ k ] n [ V k ( ∞ ) ] 2 . (132)
The first two coefficients C 1 C_1 C 1 and C 2 C_2 C 2 are given by Eqs. (G.33) and (G.30), respectively. Comparing Eq. (131) with Eqs. [2,3] from [86 ] , we get the asymptotic behavior of the function Ψ 3 ( z ) \Psi_3(z) Ψ 3 ( z ) at large z z z :
z Ψ 3 ( z ) ≈ 1 4 π ( ln ( 2 z ) + γ + 1 ) ( z → ∞ ) . (133) z\Psi_3(z)\approx\frac{1}{4\pi}\left(\ln(2z)+\gamma+1\right)\qquad(z\to\infty). \tag*{(133)} z Ψ 3 ( z ) ≈ 4 π 1 ( ln ( 2 z ) + γ + 1 ) ( z → ∞ ) . (133)
Figure 7 (b) illustrates excellent accuracy of this expression.
A common way to account for the effect of partial reactivity consists in splitting the first-reaction time into two contributions: the first-passage time to the target (the first arrival) and the first-reaction time after restarting from the target. If the first arrival point onto the target was uniformly distributed, one would simply get
T ‾ q ≃ T ‾ ∞ + 1 ∣ Γ ∣ ∫ Γ d x T q ( x ) = T ‾ ∞ + ∣ Ω ∣ q D ∣ Γ ∣ (134) \overline{T}_q \mathrel{\mathop{\simeq}} \overline{T}_\infty+\frac{1}{|\Gamma|}\int_{\Gamma} dx\,T_q(x)=\overline{T}_\infty+\frac{|\Omega|}{qD|\Gamma|} \tag*{(134)} T q ≃ T ∞ + ∣Γ∣ 1 ∫ Γ d x T q ( x ) = T ∞ + q D ∣Γ∣ ∣Ω∣ (134)
where the integral over Γ \Gamma Γ was evaluated exactly by integrating Eq. (2.45a) over x 0 ∈ Ω x_0\in\Omega x 0 ∈ Ω , using the Green’s formula and the boundary condition (59) . One sees that the hypothetical relation (5.26) corresponds to an approximation of the asymptotic relations (5.16, 5.23) by replacing Ψ d ( z ) \Psi_d(z) Ψ d ( z ) by 1 / ( 2 z ) 1/(2z) 1/ ( 2 z ) (d = 2 d=2 d = 2 ) or 1 / ( z π ) 1/(z\pi) 1/ ( z π ) (d = 3 d=3 d = 3 ). In other words, Eq. (134) captures the first term of Ψ d ( z ) \Psi_d(z) Ψ d ( z ) and ignores the remaining sum. As a consequence, this sum originates from the fact that the first arrival point on the target is distributed according to the harmonic measure density, which is highly non-uniform and even exhibits singularity at the edges of the target, see Eqs. (G.10, G.32).
On one hand, the smallness of the coefficients [ V ^ k ( ∞ ) ] 2 [\widehat{V}_k(\infty)]^2 [ V k ( ∞ ) ] 2 (see Table 1 ) suggests that the approximation (5.26) is accurate at small or even moderate values of the parameter z = q ϵ z=q\epsilon z = q ϵ , i.e., it justifies the approximation (5.26). On the other hand, the sum in Eqs. (124) , (131) is responsible for the peculiar large-z z z asymptotic behavior in Eqs. (126) , (133) that was discovered by Guérin et al. [86 ] . While their work does not mention the Steklov-Neumann problem, many employed tools are similar. For instance, the integral equation [24 ] from [86 ] involves the Green’s function G ~ 0 ( x , 0 ∣ x 0 ) \widetilde{G}_0(x,0\mid x_0) G 0 ( x , 0 ∣ x 0 ) in the half space and the related kernels. The asymptotic analysis from [86 ] may thus be reformulated in terms of the Dirichlet-to-Neumann operator and the Steklov-Neumann eigenfunctions. To illustrate this point, we mention that the first-order correction f 1 ( x ) f_1(x) f 1 ( x ) for the three-dimensional setting, defined via Eq. [88] from [86 ] , can be found explicitly by using the Landen transformation:
f 1 ( x ) = 2 π 2 ( 4 3 − E ( x ) ) , c 1 = 8 3 π 2 ≈ 0.2702 , (135) f_1(x)=\frac{2}{\pi^2}\left(\frac{4}{3}-E(x)\right),\qquad c_1=\frac{8}{3\pi^2}\approx0.2702, \tag*{(135)} f 1 ( x ) = π 2 2 ( 3 4 − E ( x ) ) , c 1 = 3 π 2 8 ≈ 0.2702 , (135)
where the constant c 1 c_1 c 1 (with unknown value) is defined by Eq. [85] from [86 ] , and E ( x ) E(x) E ( x ) is given by Eq. (95) . The function f 1 ( x ) f_1(x) f 1 ( x ) , which determines the first-order correction to the MFRT in the small reactivity regime, is proportional to our function w ^ 0 ( r ^ ) \widehat{w}_0(\widehat{r}) w 0 ( r ) from Eq. (94) , as it should.
To give a broader picture of the small-target limit, we outline a complementary approach [27 ] that was developed to study the narrow capture problem when the target is a small trap inside the confining domain, which is separated from the reflecting wall ∂ Ω N \partial\Omega_N ∂ Ω N by a large distance (i.e., the boundary ∂ Ω \partial\Omega ∂ Ω is disconnected, see § 3.3 ). Bressloff employed the matched asymptotic analysis to the encounter-based formulation of diffusion-controlled reactions and derived an asymptotic expansion of the joint probability density for particle position and the boundary local time, without employing spectral expansions like Eq. (14) on the Steklov-Neumann eigenfunctions. In this light, his asymptotic analysis brings complementary insights to our spectral results.
Beyond constant reactivity
A general framework for dealing with surface reactions on a target, known as the encounter-based approach, was introduced in [18 ] . At each encounter with the target Γ \Gamma Γ , the particle attempts to react but may fail and thus resume its diffusion. The surface reaction occurs after a sufficient number of encounters, which is described by the boundary local time ℓ t \ell_t ℓ t on Γ \Gamma Γ [119 , 120 , 121 , 122 ] . The first-reaction time τ \tau τ is thus defined as the first time instance when ℓ t \ell_t ℓ t exceeds some threshold ℓ ^ \widehat{\ell} ℓ :
τ = inf { t > 0 : ℓ t > ℓ ^ } . (136) \tau= \inf\{t > 0 : \ell_t > \widehat{\ell}\}. \tag*{(136)} τ = inf { t > 0 : ℓ t > ℓ } . (136)
The conventional setting of a target with a constant reactivity q > 0 q > 0 q > 0 , standing in the Robin boundary condition like Eq. (2.6b) or (108) , corresponds to the exponentially distributed threshold: P { ℓ ^ > ℓ } = e − q ℓ \mathbb{P}\{\widehat{\ell} > \ell\} = e^{-q\ell} P { ℓ > ℓ } = e − q ℓ [18 ] . In turn, other distributions of the random threshold ℓ ^ \widehat{\ell} ℓ yield more sophisticated surface reactions. In analogy to Eq. (114) , the spectral expansion of the moment-generating function of the associated first-reaction time τ \tau τ reads [18 ]
H ~ ( p ∣ x 0 ) = ∑ k = 0 ∞ V k ( p , Γ ) ( x 0 ) Υ ( μ k ( p , Γ ) ) ∫ Γ v k ( p , Γ ) . (137) \widetilde{H}(p|x_0)=\sum_{k=0}^{\infty}V_k^{(p,\Gamma)}(x_0)\Upsilon(\mu_k^{(p,\Gamma)})\int_{\Gamma}v_k^{(p,\Gamma)} \, . \tag*{(137)} H ( p ∣ x 0 ) = k = 0 ∑ ∞ V k ( p , Γ ) ( x 0 ) Υ ( μ k ( p , Γ ) ) ∫ Γ v k ( p , Γ ) . (137)
where Υ ( μ ) = E { e − μ ℓ ^ } \Upsilon(\mu)=\mathbb{E}\{e^{-\mu\widehat{\ell}}\} Υ ( μ ) = E { e − μ ℓ } is the moment-generating function of the threshold ℓ ^ \widehat{\ell} ℓ . When ℓ ^ \widehat{\ell} ℓ obeys the exponential law, its moment-generating function is Υ ( μ ) = 1 / ( 1 + μ / q ) \Upsilon(\mu)=1/(1+\mu/q) Υ ( μ ) = 1/ ( 1 + μ / q ) , and Eq. (137) is reduced to Eq. (114) . In general, if the mean value of ℓ ^ \widehat{\ell} ℓ is finite, its moment-generating function behaves as Υ ( μ ) = 1 − μ E { ℓ ^ } + o ( μ ) \Upsilon(\mu)=1-\mu\mathbb{E}\{\widehat{\ell}\}+o(\mu) Υ ( μ ) = 1 − μ E { ℓ } + o ( μ ) as μ → 0 \mu\to0 μ → 0 . Repeating the small-p p p asymptotic analysis from § 5.1 , we deduce the MFRT:
E x 0 { τ } = ∣ Ω ∣ D ( E { ℓ ^ } ∣ Γ ∣ − w 0 ( Γ ) ( x 0 ) − ∑ k = 1 ∞ V k ( 0 , Γ ) ( x 0 ) Υ ( μ k ( 0 , Γ ) ) b k ( Γ ) μ k ( 0 , Γ ) ) . \mathbb{E}_{x_0}\{\tau\}=\frac{|\Omega|}{D}\left(\frac{\mathbb{E}\{\widehat{\ell}\}}{|\Gamma|}-w_0^{(\Gamma)}(x_0)-\sum_{k=1}^{\infty}V_k^{(0,\Gamma)}(x_0)\Upsilon(\mu_k^{(0,\Gamma)})\frac{b_k^{(\Gamma)}}{\mu_k^{(0,\Gamma)}}\right). E x 0 { τ } = D ∣Ω∣ ( ∣Γ∣ E { ℓ } − w 0 ( Γ ) ( x 0 ) − k = 1 ∑ ∞ V k ( 0 , Γ ) ( x 0 ) Υ ( μ k ( 0 , Γ ) ) μ k ( 0 , Γ ) b k ( Γ ) ) .
Similarly, the volume-averaged MFRT reads
T ‾ = 1 ∣ Ω ∣ ∫ Ω d x 0 E x 0 { τ } = ∣ Ω ∣ D ( E { ℓ ^ } ∣ Γ ∣ + A Γ − ∑ k = 1 ∞ Υ ( μ k ( 0 , Γ ) ) [ b k ( Γ ) μ k ( 0 , Γ ) ] 2 ) . \overline{T}=\frac{1}{|\Omega|}\int_{\Omega}dx_0\,\mathbb{E}_{x_0}\{\tau\}=\frac{|\Omega|}{D}\left(\frac{\mathbb{E}\{\widehat{\ell}\}}{|\Gamma|}+A_{\Gamma}-\sum_{k=1}^{\infty}\Upsilon(\mu_k^{(0,\Gamma)})\left[\frac{b_k^{(\Gamma)}}{\mu_k^{(0,\Gamma)}}\right]^2\right). T = ∣Ω∣ 1 ∫ Ω d x 0 E x 0 { τ } = D ∣Ω∣ ∣Γ∣ E { ℓ } + A Γ − k = 1 ∑ ∞ Υ ( μ k ( 0 , Γ ) ) [ μ k ( 0 , Γ ) b k ( Γ ) ] 2 .
These expressions extend the spectral expansions (5.9, 5.12) to more general surface reactions (note that E { ℓ ^ } = 1 / q \mathbb{E}\{\widehat{\ell}\}=1/q E { ℓ } = 1/ q for the exponentially distributed threshold ℓ ^ \widehat{\ell} ℓ ). While the dominant contribution (the first two terms) does not change (except that 1 / q 1/q 1/ q is replaced by E { ℓ ^ } \mathbb{E}\{\hat{\ell}\} E { ℓ ^ } ), the correction term is affected by the chosen surface reaction via Υ ( μ ) \Upsilon(\mu) Υ ( μ ) . Our asymptotic results for μ k ( 0 , Γ ) \mu_k^{(0,\Gamma)} μ k ( 0 , Γ ) and b k ( Γ ) b_k^{(\Gamma)} b k ( Γ ) can thus help understanding sophisticated surface reactions on small targets (see also [27 ] ). Moreover, if the mean threshold is infinite, E { ℓ ^ } = + ∞ \mathbb{E}\{\hat{\ell}\}=+\infty E { ℓ ^ } = + ∞ , the MFRT is also infinite, despite the fact that diffusion occurs in a bounded domain.
Conclusion
In this paper, we revisited the mixed Steklov-Neumann spectral problem in a bounded Euclidean domain Ω \Omega Ω , with the Steklov condition on a subset Γ \Gamma Γ of a smooth boundary ∂ Ω \partial\Omega ∂ Ω . In the first step, we discussed a general scheme for constructing the eigenvalues μ k ( p , Γ ) \mu_k^{(p,\Gamma)} μ k ( p , Γ ) and eigenfunctions V k ( p , Γ ) V_k^{(p,\Gamma)} V k ( p , Γ ) with p > 0 p>0 p > 0 from the restriction of the Green’s function G ~ 0 ( x , p ∣ x 0 ) \widetilde{G}_0(x,p\mid x_0) G 0 ( x , p ∣ x 0 ) to Γ × Γ \Gamma\times\Gamma Γ × Γ . In the limit p → 0 p\to0 p → 0 , this construction involved the pseudo-Green’s function G 0 ( x ∣ x 0 ) G_0(x\mid x_0) G 0 ( x ∣ x 0 ) and the first-order corrections A Γ \mathcal{A}_{\Gamma} A Γ and W 0 ( Γ ) W_0^{(\Gamma)} W 0 ( Γ ) to the principal eigenvalue μ 0 ( p , Γ ) \mu_0^{(p,\Gamma)} μ 0 ( p , Γ ) and eigenfunction V 0 ( p , Γ ) V_0^{(p,\Gamma)} V 0 ( p , Γ ) as p → 0 p\to0 p → 0 . These corrections were shown to emerge in different contexts such as the long-time asymptotic behavior of the variance of the boundary local time on Γ \Gamma Γ , the asymptotic behavior of the MFRT on Γ \Gamma Γ , its constant-flux approximation, etc.
In the second step, we obtained the asymptotic relations (5) for the eigenvalues and eigenfunctions in the small-target limit when the subset Γ \Gamma Γ shrinks. For this purpose, we identified two auxiliary Steklov-Neumann problems and also constructed the explicit kernels of integral operators that determine the eigenpairs μ ^ k \hat{\mu}_k μ ^ k and V ^ k \hat{V}_k V ^ k appearing in Eq. (5) . We also described efficient matrix representations to compute these eigenpairs numerically with high precision. The asymptotic behavior of the corrections A Γ \mathcal{A}_{\Gamma} A Γ and w 0 ( Γ ) w_0^{(\Gamma)} w 0 ( Γ ) and of the coefficients b k ( Γ ) b_k^{(\Gamma)} b k ( Γ ) was derived. The general results were illustrated for two basic examples: an arc-shaped subset Γ \Gamma Γ on a circle and a spherical cap on a sphere. By solving the original Steklov-Neumann problem numerically, we revealed high accuracy of the asymptotic relations (5) , even for moderately large subsets Γ \Gamma Γ .
In the third step, we presented a direct application of the derived asymptotic results to the analysis of first-reaction times and related diffusion-controlled reactions. Since the moment-generating function of the first-reaction time on a subset Γ \Gamma Γ admits a spectral expansion over the Steklov-Neumann eigenpairs, its asymptotic behavior in the small-target limit can be directly accessed. In particular, we retrieved and generalized the asymptotic behavior of the MFRT for both perfect and partially reactive targets. We also revealed why the constant-flux approximation provided the correct leading terms in two dimensions but yielded wrong prefactors in three dimensions. We derived a universal, geometry-independent formula for the difference between volume-averaged MFRT (for a finite reactivity) and MFPT (for infinite reactivity) in the small-target limit. Moreover, we also discussed an extension to more sophisticated surface reactions beyond the conventional setting of constant reactivity. Further investigations of the narrow escape limit in the context of the mixed Steklov-Neumann problem can significantly improve our understanding of diffusion-controlled reactions in complex environments such as biological tissues or porous media.
Acknowledgments. The author thanks professors A. F. M. ter Elst, P. Freitas, M. Levitin, and I. Polterovich for fruitful discussions, and A. Chaigneau for preliminary numerical results. The author acknowledges the Simons Foundation for supporting his sabbatical sojourn in 2024 at the CRM (CNRS – University of Montréal, Canada), as well as the Alexander von Humboldt Foundation for support within a Bessel Prize award.
Some derivations
Identity (2.3)
For the Steklov problem, the identity (2.3) was rigorously established in [87 ] (see also [16 , 88 ] ). Its extension to the Steklov-Neumann problem seems to rather straightforward but we are not aware of its proof. In the following, we provide a formal derivation of this identity.
Multiplying Eq. (1.2a) by V k ( p + δ , Γ ) V_k^{(p+\delta,\Gamma)} V k ( p + δ , Γ ) with some δ > 0 \delta> 0 δ > 0 , multiplying Eq. (1.2a) with p ′ = p + δ p' = p + \delta p ′ = p + δ by V k ( p , Γ ) V_k^{(p,\Gamma)} V k ( p , Γ ) , subtracting them, integrating over Ω \Omega Ω and using the Green’s formula, one gets
δ ∫ Ω V k ( p , Γ ) V k ( p + δ , Γ ) = D ∫ Γ ( V k ( p , Γ ) ∂ n V k ( p + δ , Γ ) − V k ( p + δ , Γ ) ∂ n V k ( p , Γ ) ) = D ( μ k ( p + δ , Γ ) − μ k ( p , Γ ) ) ∫ Γ v k ( p , Γ ) v k ( p + δ , Γ ) . \begin{aligned}
\delta\int_{\Omega} V_k^{(p,\Gamma)} V_k^{(p+\delta,\Gamma)} = D \int_{\Gamma} \left(V_k^{(p,\Gamma)} \partial_n V_k^{(p+\delta,\Gamma)} - V_k^{(p+\delta,\Gamma)} \partial_n V_k^{(p,\Gamma)}\right) \\
&= D\left(\mu_k^{(p+\delta,\Gamma)}-\mu_k^{(p,\Gamma)}\right)\int_{\Gamma}v_k^{(p,\Gamma)}v_k^{(p+\delta,\Gamma)}.
\end{aligned} δ ∫ Ω V k ( p , Γ ) V k ( p + δ , Γ ) = D ∫ Γ ( V k ( p , Γ ) ∂ n V k ( p + δ , Γ ) − V k ( p + δ , Γ ) ∂ n V k ( p , Γ ) ) = D ( μ k ( p + δ , Γ ) − μ k ( p , Γ ) ) ∫ Γ v k ( p , Γ ) v k ( p + δ , Γ ) .
Dividing this identity by δ \delta δ and taking the limit δ → 0 \delta\to0 δ → 0 yield
∂ p μ k ( p , Γ ) = lim δ → 0 μ k ( p + δ , Γ ) − μ k ( p , Γ ) δ = 1 D ∫ Ω ∣ V k ( p , Γ ) ∣ 2 . \partial_p\mu_k^{(p,\Gamma)} = \lim_{\delta\to0}\frac{\mu_k^{(p+\delta,\Gamma)}-\mu_k^{(p,\Gamma)}}{\delta} = \frac{1}{D}\int_{\Omega}\left|V_k^{(p,\Gamma)}\right|^2. ∂ p μ k ( p , Γ ) = δ → 0 lim δ μ k ( p + δ , Γ ) − μ k ( p , Γ ) = D 1 ∫ Ω V k ( p , Γ ) 2 .
where we used the normalization (2.2) of eigenfunctions v k ( p , Γ ) v_k^{(p,\Gamma)} v k ( p , Γ ) . The existence of the limit follows from the analyticity of the Dirichlet-to-Neumann map, see [90 ] and the discussion before Eq. (2.21). Strictly speaking, one has to follow the same (k k k -th) analytical branch for p p p and p + δ p+\delta p + δ that may perturb the increasing order of eigenvalues.
Spectral expansion
Even though the spectral expansion (2.7) resembles that reported in [18 ] , we sketch here its formal derivation for mixed Robin-Neumann conditions. The expression (2.7) can be deduced by searching G ~ q ( x , p ∣ x 0 ) \widetilde{G}_q(x,p|x_0) G q ( x , p ∣ x 0 ) as G ~ ∞ ( x , p ∣ x 0 ) + g ~ q ( x , p ∣ x 0 ) \widetilde{G}_{\infty}(x,p|x_0) + \widetilde{g}_q(x,p|x_0) G ∞ ( x , p ∣ x 0 ) + g q ( x , p ∣ x 0 ) , with an unknown function g ~ q ( x , p ∣ x 0 ) \widetilde{g}_q(x,p|x_0) g q ( x , p ∣ x 0 ) that satisfies the homogeneous modified Helmholtz equation and can thus be expressed by using the completeness of the eigenfunctions v k ( p , Γ ) v_k^{(p,\Gamma)} v k ( p , Γ ) in L 2 ( Γ ) L^2(\Gamma) L 2 ( Γ ) . Here we follow an alternative way by verifying directly that Eq. (14) satisfies the boundary value problem (2.6).
As Eqs. (2.6a, 2.6c) are satisfied by construction, it remains to check the Robin boundary condition (2.6b). Substituting Eq. (14) into Eq. (2.6b), one gets for any x 0 ∈ Ω x_0 \in\Omega x 0 ∈ Ω and x ∈ Γ x \in\Gamma x ∈ Γ :
D ( ∂ n + q ) G ~ q ( x , p ∣ x 0 ) = − j ~ ∞ ( x , p ∣ x 0 ) + ∑ k = 0 ∞ V k ( p , Γ ) ( x 0 ) v k ( p , Γ ) ( x ) , (138) D(\partial_n+q)\widetilde{G}_q(x,p|x_0)=-\widetilde{j}_{\infty}(x,p|x_0)+\sum_{k=0}^{\infty}V_k^{(p,\Gamma)}(x_0)v_k^{(p,\Gamma)}(x), \tag*{(138)} D ( ∂ n + q ) G q ( x , p ∣ x 0 ) = − j ∞ ( x , p ∣ x 0 ) + k = 0 ∑ ∞ V k ( p , Γ ) ( x 0 ) v k ( p , Γ ) ( x ) , (138)
where j ~ ∞ ( x , p ∣ x 0 ) = − D ∂ n G ~ ∞ ( x , p ∣ x 0 ) \widetilde{j}_{\infty}(x,p|x_0)=-D\partial_n\widetilde{G}_{\infty}(x,p|x_0) j ∞ ( x , p ∣ x 0 ) = − D ∂ n G ∞ ( x , p ∣ x 0 ) . To check that the right-hand side is zero, we first note that the Steklov eigenfunction V k ( p , Γ ) ( x 0 ) V_k^{(p,\Gamma)}(x_0) V k ( p , Γ ) ( x 0 ) for x 0 ∈ Ω x_0 \in\Omega x 0 ∈ Ω can be obtained from its restriction v k ( p , Γ ) v_k^{(p,\Gamma)} v k ( p , Γ ) to Γ \Gamma Γ as
V k ( p , Γ ) ( x 0 ) = ∫ Γ d x ′ j ~ ∞ ( x ′ , p ∣ x 0 ) v k ( p , Γ ) ( x ′ ) ( x 0 ∈ Ω ) . (139) V_k^{(p,\Gamma)}(x_0)=\int_{\Gamma}dx'\,\widetilde{j}_{\infty}(x',p|x_0)v_k^{(p,\Gamma)}(x') \qquad(x_0\in\Omega). \tag*{(139)} V k ( p , Γ ) ( x 0 ) = ∫ Γ d x ′ j ∞ ( x ′ , p ∣ x 0 ) v k ( p , Γ ) ( x ′ ) ( x 0 ∈ Ω ) . (139)
This relation is deduced in a standard way by multiplying Eq. (1.2a) by G ~ ∞ ( x , p ∣ x 0 ) \widetilde{G}_{\infty}(x,p|x_0) G ∞ ( x , p ∣ x 0 ) , multiplying Eq. (2.6a) by V k ( p , Γ ) ( x ) V_k^{(p,\Gamma)}(x) V k ( p , Γ ) ( x ) , subtracting these equations, integrating over x ∈ Ω x \in\Omega x ∈ Ω , applying the Green’s formula, and using the boundary conditions. Substituting the representation (A.2) into Eq. (A.1), we have
D ( ∂ n + q ) G ~ q ( x , p ∣ x 0 ) = − j ~ ∞ ( x , p ∣ x 0 ) + ∫ Γ d x ′ j ~ ∞ ( x ′ , p ∣ x 0 ) ∑ k = 0 ∞ v k ( p , Γ ) ( x ′ ) v k ( p , Γ ) ( x ) D(\partial_n+q)\widetilde{G}_q(x,p|x_0)=-\widetilde{j}_\infty(x,p|x_0)+\int_{\Gamma}dx'\,\widetilde{j}_\infty(x',p|x_0)\sum_{k=0}^{\infty}v_k^{(p,\Gamma)}(x')v_k^{(p,\Gamma)}(x) D ( ∂ n + q ) G q ( x , p ∣ x 0 ) = − j ∞ ( x , p ∣ x 0 ) + ∫ Γ d x ′ j ∞ ( x ′ , p ∣ x 0 ) k = 0 ∑ ∞ v k ( p , Γ ) ( x ′ ) v k ( p , Γ ) ( x )
where the order of sum and integral was exchanged. Second, we employ the completeness relation
∑ k = 0 ∞ v k ( p , Γ ) ( x ′ ) v k ( p , Γ ) ( x ) = δ ( x − x ′ ) (140) \sum_{k=0}^{\infty}v_k^{(p,\Gamma)}(x')v_k^{(p,\Gamma)}(x)=\delta(x-x') \tag*{(140)} k = 0 ∑ ∞ v k ( p , Γ ) ( x ′ ) v k ( p , Γ ) ( x ) = δ ( x − x ′ ) (140)
that formally reflects that the basis of the eigenfunctions { v k ( p , Γ ) } \{v_k^{(p,\Gamma)}\} { v k ( p , Γ ) } is complete in L 2 ( Γ ) L^2(\Gamma) L 2 ( Γ ) (this relation can also be formally understood as the expansion of the Dirac distribution on the basis { v k ( p , Γ ) } \{v_k^{(p,\Gamma)}\} { v k ( p , Γ ) } ). As a consequence, one sees that the right-hand side is zero, so that the spectral expansion (2.7) satisfies the Robin boundary condition.
We emphasize that the above derivation is not a mathematical proof; in particular, we do not discuss the convergence of the above expansions, as well as the possibility of exchanging the order sum and normal derivative, integration by parts, etc. Spectral expansions of Green’s functions over Laplacian eigenfunctions (such as Eq. (18) ) are fairly standard. In turn, we are not aware of convergence results for Eq. (14) and related expressions used in the manuscript. For instance, a variant of Eq. (14) for q = 0 q=0 q = 0 and Steklov problem (i.e., Γ = ∂ Ω \Gamma=\partial\Omega Γ = ∂ Ω ) was rigorously established in [90 ] (Theorem 4.9) in a much more general setting. We note that similar expansions involving Steklov eigenfunctions and single/double layer potentials were rigorously established in [43 ] . Moreover, if (i) Γ = ∂ Ω \Gamma=\partial\Omega Γ = ∂ Ω , (ii) the boundary ∂ Ω \partial\Omega ∂ Ω is real-analytic, and (iii) at least one point x x x or x 0 x_0 x 0 does not belong to ∂ Ω \partial\Omega ∂ Ω , then the Steklov eigenfunctions V k ( p , ∂ Ω ) V_k^{(p,\partial\Omega)} V k ( p , ∂ Ω ) are known to decay exponentially with k k k [123 ] , thus ensuring the fast convergence of the series. To our knowledge, similar results are not yet established for mixed Steklov-Neumann problem. A rigorous justification of the spectral expansion (2.7) presents an open mathematical problem.
Appendix B. Disk
In this Appendix, we recall the exact explicit formulas for the Green’s function G ~ 0 ( x , p ∣ x 0 ) \widetilde{G}_0(x,p|x_0) G 0 ( x , p ∣ x 0 ) and the pseudo-Green’s function G 0 ( x ∣ x 0 ) G_0(x|x_0) G 0 ( x ∣ x 0 ) for the disk of radius R R R . We also deduce the exact expressions for the corrections A Γ A_\Gamma A Γ and w 0 ( Γ ) w_0^{(\Gamma)} w 0 ( Γ ) . We finally present an implementation for computing numerically the eigenpairs { μ k ( p , Γ ) , V k ( p , Γ ) } \{\mu_k^{(p,\Gamma)},V_k^{(p,\Gamma)}\} { μ k ( p , Γ ) , V k ( p , Γ ) } .
General solution
The eigenmodes of the conventional Steklov problem, as well as the Green’s function, are fairly well-known for the disk [20 , 38 ] . For instance, the restriction of the Green’s function to the boundary ∂ Ω \partial\Omega ∂ Ω reads in polar coordinates as (see, e.g., [20 ] )
G ~ 0 ( θ , p ∣ θ 0 ) = 1 π D R ( 1 2 μ 0 ( p , ∂ Ω ) + ∑ k = 1 ∞ cos ( k ( θ − θ 0 ) ) μ k ( p , ∂ Ω ) ) (141) \widetilde{G}_0(\theta,p|\theta_0)=\frac{1}{\pi D R}\left(\frac{1}{2\mu_0^{(p,\partial\Omega)}}+\sum_{k=1}^{\infty}\frac{\cos(k(\theta-\theta_0))}{\mu_k^{(p,\partial\Omega)}}\right) \tag*{(141)} G 0 ( θ , p ∣ θ 0 ) = π D R 1 ( 2 μ 0 ( p , ∂ Ω ) 1 + k = 1 ∑ ∞ μ k ( p , ∂ Ω ) cos ( k ( θ − θ 0 )) ) (141)
where
μ k ( p , ∂ Ω ) = p D I k ′ ( R p D ) I k ( R p D ) (142) \mu_k^{(p,\partial\Omega)}=\sqrt{\frac{p}{D}}\frac{I_k'\left(R\sqrt{\frac{p}{D}}\right)}{I_k\left(R\sqrt{\frac{p}{D}}\right)} \tag*{(142)} μ k ( p , ∂ Ω ) = D p I k ( R D p ) I k ′ ( R D p ) (142)
with I ν ( z ) I_{\nu}(z) I ν ( z ) being the modified Bessel function of the first kind, and prime denoting the derivative with respect to the argument. Writing cos ( k ( θ − θ 0 ) ) = cos ( k θ ) cos ( k θ 0 ) + sin ( k θ ) sin ( k θ 0 ) \cos(k(\theta-\theta_0))=\cos(k\theta)\cos(k\theta_0)+\sin(k\theta)\sin(k\theta_0) cos ( k ( θ − θ 0 )) = cos ( k θ ) cos ( k θ 0 ) + sin ( k θ ) sin ( k θ 0 ) , one can recognize the spectral expansion (24) over the Steklov eigenvalues μ k ( p , ∂ Ω ) \mu_k^{(p,\partial\Omega)} μ k ( p , ∂ Ω ) and eigenfunctions v k ( p , ∂ Ω ) v_k^{(p,\partial\Omega)} v k ( p , ∂ Ω ) , which are given by cosine and sine functions [20 ] . The principal eigenvalue μ 0 ( p , ∂ Ω ) \mu_0^{(p,\partial\Omega)} μ 0 ( p , ∂ Ω ) is simple, whereas the other eigenvalues are twice degenerate (throughout this Appendix, we ignore the degeneracy and enumerate by k k k distinct eigenvalues).
We consider the subset Γ \Gamma Γ to be an arc of angle 2 ε 2\varepsilon 2 ε on the boundary of the disk: Γ = { ( R , θ ) ∈ ∂ Ω : ∣ θ ∣ < ε } \Gamma=\{(R,\theta)\in\partial\Omega:\ |\theta|<\varepsilon\} Γ = {( R , θ ) ∈ ∂ Ω : ∣ θ ∣ < ε } . According to Eq. (16) , the restriction of the Green’s function G ~ 0 ( θ , p ∣ θ 0 ) \widetilde{G}_0(\theta,p\mid\theta_0) G 0 ( θ , p ∣ θ 0 ) to Γ × Γ \Gamma\times\Gamma Γ × Γ defines an integral operator that determines the eigenvalues and eigenfunctions of the Dirichlet-to-Neumann operator M p ( Γ ) \mathcal{M}_p^{(\Gamma)} M p ( Γ ) for any ε \varepsilon ε and any p > 0 p>0 p > 0 :
∫ − ε ε d θ D G ~ 0 ( θ , p ∣ θ 0 ) v k ( p , Γ ) ( θ ) = v k ( p , Γ ) ( θ 0 ) R μ k ( p , Γ ) ( ∣ θ 0 ∣ < ε , k ≥ 0 ) . (143) \int_{-\varepsilon}^{\varepsilon}d\theta\,D\widetilde{G}_0(\theta,p\mid\theta_0)v_k^{(p,\Gamma)}(\theta)=\frac{v_k^{(p,\Gamma)}(\theta_0)}{R\mu_k^{(p,\Gamma)}}\qquad(|\theta_0|<\varepsilon,\ k\ge0). \tag*{(143)} ∫ − ε ε d θ D G 0 ( θ , p ∣ θ 0 ) v k ( p , Γ ) ( θ ) = R μ k ( p , Γ ) v k ( p , Γ ) ( θ 0 ) ( ∣ θ 0 ∣ < ε , k ≥ 0 ) . (143)
Replacing this integral by a Riemann sum yields an approximate matrix representation of the integral operator, while its numerical diagonalization approximates eigenvalues and eigenfunctions (see Appendix B.2).
The analysis is a little subtler at p = 0 p=0 p = 0 . Since μ k ( 0 , ∂ Ω ) = k / R \mu_k^{(0,\partial\Omega)}=k/R μ k ( 0 , ∂ Ω ) = k / R , one needs to remove the diverging term 1 / μ 0 ( p , ∂ Ω ) 1/\mu_0^{(p,\partial\Omega)} 1/ μ 0 ( p , ∂ Ω ) in Eq. (B.1) by subtracting D / ( p ∣ Ω ∣ ) D/(p|\Omega|) D / ( p ∣Ω∣ ) that yields the pseudo-Green’s function
G 0 ( θ ∣ θ 0 ) = 1 8 π + 1 π ∑ k = 1 ∞ cos ( k ( θ − θ 0 ) ) k = 1 8 π − 1 2 π ln ( 2 − 2 cos ( θ − θ 0 ) ) (144) G_0(\theta\mid\theta_0)=\frac{1}{8\pi}+\frac{1}{\pi}\sum_{k=1}^{\infty}\frac{\cos(k(\theta-\theta_0))}{k}=\frac{1}{8\pi}-\frac{1}{2\pi}\ln(2-2\cos(\theta-\theta_0)) \tag*{(144)} G 0 ( θ ∣ θ 0 ) = 8 π 1 + π 1 k = 1 ∑ ∞ k cos ( k ( θ − θ 0 )) = 8 π 1 − 2 π 1 ln ( 2 − 2 cos ( θ − θ 0 )) (144)
(note that a more general explicit form of the pseudo-Green’s function inside the unit disk was given in [65 ] , see Eq. [4.3a]). In particular, in the limit θ → θ 0 \theta\to\theta_0 θ → θ 0 , one retrieves the asymptotic behavior (80) , with
R 0 = 1 8 π , (145) R_0=\frac{1}{8\pi}, \tag*{(145)} R 0 = 8 π 1 , (145)
independently of θ 0 \theta_0 θ 0 . Substituting this expression into Eqs. (39) , (40) , we find
w 0 ( Γ ) ( θ 0 ) = − A Γ + 1 2 ε ∫ − ε ε d θ G 0 ( θ ∣ θ 0 ) = − A Γ + 1 8 π + 1 π ε ∑ k = 1 ∞ sin ( k ε ) cos ( k θ 0 ) k 2 = − A Γ + 1 8 π + Li 2 ( e i ( ε + θ 0 ) ) − Li 2 ( e − i ( ε + θ 0 ) ) + Li 2 ( e i ( ε − θ 0 ) ) − Li 2 ( e − i ( ε − θ 0 ) ) 4 i π ε , (146) \begin{aligned}
w_0^{(\Gamma)}(\theta_0)=-A_\Gamma+\frac{1}{2\varepsilon}\int_{-\varepsilon}^{\varepsilon}d\theta\,G_0(\theta\mid\theta_0)=-A_\Gamma+\frac{1}{8\pi}+\frac{1}{\pi\varepsilon}\sum_{k=1}^{\infty}\frac{\sin(k\varepsilon)\cos(k\theta_0)}{k^2} \\
&=-A_\Gamma+\frac{1}{8\pi}+\frac{\operatorname{Li}_2(e^{i(\varepsilon+\theta_0)})-\operatorname{Li}_2(e^{-i(\varepsilon+\theta_0)})+\operatorname{Li}_2(e^{i(\varepsilon-\theta_0)})-\operatorname{Li}_2(e^{-i(\varepsilon-\theta_0)})}{4i\pi\varepsilon}, \tag*{(146)}
\end{aligned} w 0 ( Γ ) ( θ 0 ) = − A Γ + 2 ε 1 ∫ − ε ε d θ G 0 ( θ ∣ θ 0 ) = − A Γ + 8 π 1 + π ε 1 k = 1 ∑ ∞ k 2 sin ( k ε ) cos ( k θ 0 ) = − A Γ + 8 π 1 + 4 iπ ε Li 2 ( e i ( ε + θ 0 ) ) − Li 2 ( e − i ( ε + θ 0 ) ) + Li 2 ( e i ( ε − θ 0 ) ) − Li 2 ( e − i ( ε − θ 0 ) ) , (146)
and
A Γ = 1 8 π + 1 π ε 2 ∑ k = 1 ∞ sin 2 ( k ε ) k 3 = 1 8 π + 2 Li 3 ( 1 ) − Li 3 ( e 2 i ε ) − Li 3 ( e − 2 i ε ) 4 π ε 2 , (147) A_\Gamma=\frac{1}{8\pi}+\frac{1}{\pi\varepsilon^2}\sum_{k=1}^{\infty}\frac{\sin^2(k\varepsilon)}{k^3}=\frac{1}{8\pi}+\frac{2\operatorname{Li}_3(1)-\operatorname{Li}_3(e^{2i\varepsilon})-\operatorname{Li}_3(e^{-2i\varepsilon})}{4\pi\varepsilon^2}, \tag*{(147)} A Γ = 8 π 1 + π ε 2 1 k = 1 ∑ ∞ k 3 sin 2 ( k ε ) = 8 π 1 + 4 π ε 2 2 Li 3 ( 1 ) − Li 3 ( e 2 i ε ) − Li 3 ( e − 2 i ε ) , (147)
where
Li n ( z ) = ∑ k = 1 ∞ z k k n (148) \operatorname{Li}_n(z)=\sum_{k=1}^{\infty}\frac{z^k}{k^n} \tag*{(148)} Li n ( z ) = k = 1 ∑ ∞ k n z k (148)
is the polylogarithm. Substitution of these relations in Eq. (41) yields the kernel
G ( θ ∣ θ 0 ) = − 1 2 π ln ( 2 − 2 cos ( θ − θ 0 ) ) − 1 π ∑ k = 1 ∞ ( sin ( k ε ) [ cos ( k θ 0 ) + cos ( k θ ) ] ε k 2 − sin 2 ( k ε ) ε 2 k 3 ) . (149) \mathcal{G}(\theta\mid\theta_0)=-\frac{1}{2\pi}\ln(2-2\cos(\theta-\theta_0))-\frac{1}{\pi}\sum_{k=1}^{\infty}\left(\frac{\sin(k\varepsilon)[\cos(k\theta_0)+\cos(k\theta)]}{\varepsilon k^2}-\frac{\sin^2(k\varepsilon)}{\varepsilon^2 k^3}\right). \tag*{(149)} G ( θ ∣ θ 0 ) = − 2 π 1 ln ( 2 − 2 cos ( θ − θ 0 )) − π 1 k = 1 ∑ ∞ ( ε k 2 sin ( k ε ) [ cos ( k θ 0 ) + cos ( k θ )] − ε 2 k 3 sin 2 ( k ε ) ) . (149)
which can also be expressed in terms of polylogarithms. In summary, the eigenvalue problem (2.33) reads here as
∫ − ε ε d θ G ( θ ∣ θ 0 ) v k ( 0 , Γ ) ( θ ) = 1 R μ k ( 0 , Γ ) v k ( 0 , Γ ) ( θ 0 ) ( ∣ θ 0 ∣ < ε , k ≥ 1 ) . (150) \int_{-\varepsilon}^{\varepsilon}d\theta\,\mathcal{G}(\theta\mid\theta_0)v_k^{(0,\Gamma)}(\theta)=\frac{1}{R\mu_k^{(0,\Gamma)}}v_k^{(0,\Gamma)}(\theta_0)\quad(|\theta_0|<\varepsilon,\ k\geq1). \tag*{(150)} ∫ − ε ε d θ G ( θ ∣ θ 0 ) v k ( 0 , Γ ) ( θ ) = R μ k ( 0 , Γ ) 1 v k ( 0 , Γ ) ( θ 0 ) ( ∣ θ 0 ∣ < ε , k ≥ 1 ) . (150)
This equation is valid for any 0 < ε ≤ π 0<\varepsilon\leq\pi 0 < ε ≤ π but requires a numerical treatment (see Appendix B.2). If the eigenvalue μ k ( 0 , Γ ) \mu_k^{(0,\Gamma)} μ k ( 0 , Γ ) is simple, the symmetry of the considered Steklov-Neumann problem implies that the associated eigenfunction v k ( 0 , Γ ) ( θ ) v_k^{(0,\Gamma)}(\theta) v k ( 0 , Γ ) ( θ ) is either symmetric or antisymmetric: v k ( 0 , Γ ) ( − θ ) = ± v k ( 0 , Γ ) ( θ ) v_k^{(0,\Gamma)}(-\theta)=\pm v_k^{(0,\Gamma)}(\theta) v k ( 0 , Γ ) ( − θ ) = ± v k ( 0 , Γ ) ( θ ) . Lower and upper bounds on the eigenvalues are obtained in Appendix H.
In the limit ε → 0 \varepsilon\to0 ε → 0 , one can use the behavior of the polylogarithm Li n ( z ) \operatorname{Li}_n(z) Li n ( z ) as z → 1 z\to1 z → 1 [124 ] to investigate the asymptotic behavior of Eqs. (B.6, B.7, B.9. In this way, one can retrieve the asymptotic relations (3.10, 3.11, 3.15); moreover, this alternative derivation illustrates our statement that the curvature of the boundary does not appear in the leading order.
Numerical implementation
Let us briefly discuss the numerical implementation of the considered spectral problems that was realized in Matlab.
Arbitrary arc. For a numerical implementation of the spectral problem (B.10), the interval ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) is discretized into 2 N 2N 2 N segments, centered at y n = ( n − N − 1 / 2 ) / N y_n=(n-N-1/2)/N y n = ( n − N − 1/2 ) / N , with n = 1 , 2 , … , 2 N n=1,2,\ldots,2N n = 1 , 2 , … , 2 N . Setting η = 1 / ( 2 N ) \eta=1/(2N) η = 1/ ( 2 N ) , we compute the matrix of size ( 2 N ) × ( 2 N ) (2N)\times(2N) ( 2 N ) × ( 2 N ) :
G n , n ′ ( ε ) = ∫ ε ( y n − η ) ε ( y n + η ) d θ G ( θ ∣ ε y n ′ ) = ε ∫ y n − η y n + η d y G ( ε y ∣ ε y n ′ ) ≈ 2 R π ∑ k = 1 k max [ cos ( k ε ( y n − y n ′ ) ) sin ( k ε η ) k 2 + η sin 2 ( k ε ) ε k 3 − sin ( k ε ) ε k 2 ( η ε cos ( k ε y n ′ ) + cos ( k ε y n ) sin ( k ε η ) k ) ] . (151) \begin{aligned}
G_{n,n'}^{(\varepsilon)}&=\int_{\varepsilon(y_n-\eta)}^{\varepsilon(y_n+\eta)}d\theta\,\mathcal{G}(\theta\mid\varepsilon y_{n'})=\varepsilon\int_{y_n-\eta}^{y_n+\eta}dy\,\mathcal{G}(\varepsilon y\mid\varepsilon y_{n'})\\
&\approx\frac{2R}{\pi}\sum_{k=1}^{k_{\max}}\left[\frac{\cos(k\varepsilon(y_n-y_{n'}))\sin(k\varepsilon\eta)}{k^2}+\eta\frac{\sin^2(k\varepsilon)}{\varepsilon k^3}-\frac{\sin(k\varepsilon)}{\varepsilon k^2}\left(\eta\varepsilon\cos(k\varepsilon y_{n'})+\cos(k\varepsilon y_n)\frac{\sin(k\varepsilon\eta)}{k}\right)\right]. \tag*{(151)}
\end{aligned} G n , n ′ ( ε ) = ∫ ε ( y n − η ) ε ( y n + η ) d θ G ( θ ∣ ε y n ′ ) = ε ∫ y n − η y n + η d y G ( ε y ∣ ε y n ′ ) ≈ π 2 R k = 1 ∑ k m a x [ k 2 cos ( k ε ( y n − y n ′ )) sin ( k ε η ) + η ε k 3 sin 2 ( k ε ) − ε k 2 sin ( k ε ) ( η ε cos ( k ε y n ′ ) + cos ( k ε y n ) k sin ( k ε η ) ) ] . (151)
where the kernel G ( θ ∣ θ 0 ) \mathcal{G}(\theta\mid\theta_0) G ( θ ∣ θ 0 ) was given by Eq. (B.9), and k max k_{\max} k m a x is the truncation order that may need to be large if ε \varepsilon ε is small. The integral eigenvalue problem (B.10) is then approximated as
∑ n = 1 2 N G n , n ′ ( ε ) v k ( 0 , Γ ) ( ε y n ) ≈ 1 R μ k ( 0 , Γ ) v k ( 0 , Γ ) ( ε y n ′ ) ( n = 1 , 2 , … , 2 N ) . (152) \sum_{n=1}^{2N}G_{n,n'}^{(\varepsilon)}v_k^{(0,\Gamma)}(\varepsilon y_n)\approx\frac{1}{R\mu_k^{(0,\Gamma)}}v_k^{(0,\Gamma)}(\varepsilon y_{n'})\quad(n=1,2,\ldots,2N). \tag*{(152)} n = 1 ∑ 2 N G n , n ′ ( ε ) v k ( 0 , Γ ) ( ε y n ) ≈ R μ k ( 0 , Γ ) 1 v k ( 0 , Γ ) ( ε y n ′ ) ( n = 1 , 2 , … , 2 N ) . (152)
In other words, one needs to diagonalize the matrix G ( ε ) G^{(\varepsilon)} G ( ε ) in order to approximate the (inverse) eigenvalues and eigenfunctions of the Dirichlet-to-Neumann operator M 0 ( Γ ) \mathcal{M}_0^{(\Gamma)} M 0 ( Γ ) .
We checked numerically the convergence of several first eigenvalues and eigenfunctions as N N N increases. Moreover, the eigenvalues and eigenfunctions were compared to those obtained independently by a finite-element method [88 ] .
A very similar computation is also applicable to the eigenvalue problem (88) for the kernel G ^ ( x ∣ x 0 ) \widehat{G}(x|x_0) G ( x ∣ x 0 ) from Eq. (86) . However, its simpler form allows for a more efficient computation described in the next section.
Matrix representation of the kernel G ^ ( x ∣ x 0 ) \widehat{G}(x|x_0) G ( x ∣ x 0 ) . The eigenvalue problem (88) admits a simple matrix representation that relies on the following expansion in terms of Chebyshev polynomials of the first kind for x , x 0 ∈ [ − 1 , 1 ] x,x_0 \in[-1,1] x , x 0 ∈ [ − 1 , 1 ] and x ≠ x 0 x \ne x_0 x = x 0 (see, e.g., [128 ] ):
ln ( 2 ∣ x − x 0 ∣ ) = − ∑ n = 1 ∞ 2 n T n ( x ) T n ( x 0 ) . (153) \ln(2|x-x_0|)=-\sum_{n=1}^{\infty}\frac{2}{n}T_n(x)T_n(x_0). \tag*{(153)} ln ( 2∣ x − x 0 ∣ ) = − n = 1 ∑ ∞ n 2 T n ( x ) T n ( x 0 ) . (153)
The kernel G ^ ( x ∣ x 0 ) \widehat{G}(x|x_0) G ( x ∣ x 0 ) can then be represented as
G ^ ( x ∣ x 0 ) = − 1 + 2 ln 2 2 π + 1 2 π ( ( 1 + x 0 ) ln ( 1 + x 0 ) + ( 1 − x 0 ) ln ( 1 − x 0 ) ) + 1 π ∑ n = 1 ∞ 1 n { 2 T n ( x ) T n ( x 0 ) − T n ( x ) ( 1 + ( − 1 ) n − x ( 1 − ( − 1 ) n ) ) } . (154) \begin{aligned}
\widehat{G}(x|x_0)=-\frac{1+2\ln2}{2\pi}+\frac{1}{2\pi}\left((1+x_0)\ln(1+x_0)+(1-x_0)\ln(1-x_0)\right) \\
&+\frac{1}{\pi}\sum_{n=1}^{\infty}\frac{1}{n}\left\{2T_n(x)T_n(x_0)-T_n(x)\left(1+(-1)^n-x\left(1-(-1)^n\right)\right)\right\}. \tag*{(154)}
\end{aligned} G ( x ∣ x 0 ) = − 2 π 1 + 2 ln 2 + 2 π 1 ( ( 1 + x 0 ) ln ( 1 + x 0 ) + ( 1 − x 0 ) ln ( 1 − x 0 ) ) + π 1 n = 1 ∑ ∞ n 1 { 2 T n ( x ) T n ( x 0 ) − T n ( x ) ( 1 + ( − 1 ) n − x ( 1 − ( − 1 ) n ) ) } . (154)
Let us search for an eigenfunction v ^ k ( x ) \widehat{v}_k(x) v k ( x ) in the form
v ^ k ( x ) = 1 1 − x 2 ∑ n ′ = 1 ∞ C k , n ′ T n ′ ( x ) , (155) \widehat{v}_k(x)=\frac{1}{\sqrt{1-x^2}}\sum_{n'=1}^{\infty}C_{k,n'}T_{n'}(x), \tag*{(155)} v k ( x ) = 1 − x 2 1 n ′ = 1 ∑ ∞ C k , n ′ T n ′ ( x ) , (155)
with unknown coefficients C k , n ′ C_{k,n'} C k , n ′ , where the term n ′ = 0 n'=0 n ′ = 0 was excluded from the sum to ensure that v ^ k ( x ) \widehat{v}_k(x) v k ( x ) is orthogonal to a constant T 0 ( x ) = 1 T_0(x)=1 T 0 ( x ) = 1 . Substituting this expression into Eq. (88) , multiplying by T m ( x 0 ) T_m(x_0) T m ( x 0 ) and integrating over x 0 x_0 x 0 from − 1 -1 − 1 to 1 1 1 , we reduce the original integral problem to the equivalent matrix problem
∑ n = 1 ∞ C k , n G n , m = 1 μ ^ k C k , m ( m = 1 , 2 , … ) , (156) \sum_{n=1}^{\infty}C_{k,n}G_{n,m}=\frac{1}{\widehat{\mu}_k}C_{k,m}\qquad(m=1,2,\ldots), \tag*{(156)} n = 1 ∑ ∞ C k , n G n , m = μ k 1 C k , m ( m = 1 , 2 , … ) , (156)
where
G n , m = 2 π ∫ − 1 1 d x 1 − x 2 ∫ − 1 1 d x 0 T n ( x ) G ^ ( x ∣ x 0 ) T m ( x 0 ) , (157) G_{n,m}=\frac{2}{\pi}\int_{-1}^{1}\frac{dx}{\sqrt{1-x^2}}\int_{-1}^{1}dx_0\,T_n(x)\widehat{G}(x|x_0)T_m(x_0), \tag*{(157)} G n , m = π 2 ∫ − 1 1 1 − x 2 d x ∫ − 1 1 d x 0 T n ( x ) G ( x ∣ x 0 ) T m ( x 0 ) , (157)
and we used the orthogonality and normalization of Chebyshev polynomials. Substituting Eq. (B.14), we can evaluate the matrix elements explicitly:
G n , m = 1 + ( − 1 ) m + n π n ( 1 1 − ( m − n ) 2 + 1 1 − ( m + n ) 2 ) − ( 1 + ( − 1 ) n ) ( 1 + ( − 1 ) m ) π n ( 1 − n 2 ) ( 1 − m 2 ) ( m , n = 1 , 2 , … ) (158) G_{n,m}=\frac{1+(-1)^{m+n}}{\pi n}\left(\frac{1}{1-(m-n)^2}+\frac{1}{1-(m+n)^2}\right)-\frac{(1+(-1)^n)(1+(-1)^m)}{\pi n(1-n^2)(1-m^2)}\qquad(m,n=1,2,\ldots) \tag*{(158)} G n , m = π n 1 + ( − 1 ) m + n ( 1 − ( m − n ) 2 1 + 1 − ( m + n ) 2 1 ) − π n ( 1 − n 2 ) ( 1 − m 2 ) ( 1 + ( − 1 ) n ) ( 1 + ( − 1 ) m ) ( m , n = 1 , 2 , … ) (158)
(note that the “seemingly divergent” terms with m = 1 m=1 m = 1 , n = 1 n=1 n = 1 , ( m − n ) 2 = 1 (m-n)^2=1 ( m − n ) 2 = 1 or ( n + m ) 2 = 1 (n+m)^2=1 ( n + m ) 2 = 1 are set to zero). The eigenvalues of the truncated matrix G \mathbf{G} G approximate 1 / μ ^ k 1/\widehat{\mu}_k 1/ μ k , whereas its left eigenvectors yield the coefficients C k , n C_{k,n} C k , n determining the eigenfunctions v ^ k ( x ) \hat{v}_k(x) v ^ k ( x ) via Eq. (B.15). We checked a rapid convergence of the numerical values of 1 / μ ^ k 1/\hat{\mu}_k 1/ μ ^ k as the truncation size increases. In fact, this method is much faster and more accurate that a direct discretization of the integral equation.
Appendix C. Ball
In this Appendix, we recall the exact explicit formulas for the Green’s function G ~ 0 ( x , p ∣ x 0 ) \tilde{G}_0(x,p|x_0) G ~ 0 ( x , p ∣ x 0 ) and the pseudo-Green’s function G 0 ( x ∣ x 0 ) G_0(x|x_0) G 0 ( x ∣ x 0 ) for the ball of radius R R R . We also deduce the exact expressions for the corrections A Γ A_{\Gamma} A Γ and w 0 ( Γ ) w_0^{(\Gamma)} w 0 ( Γ ) . We finally present an implementation for computing numerically the eigenpairs { μ k ( p , Γ ) , V k ( p , Γ ) } \{ \mu_k^{(p,\Gamma)}, V_k^{(p,\Gamma)} \} { μ k ( p , Γ ) , V k ( p , Γ ) } .
General solution
The restriction of the Green’s function to the boundary ∂ Ω \partial\Omega ∂ Ω reads in the spherical coordinates as (see, e.g., [20 ] )
G ~ 0 ( x , p ∣ x 0 ) = 1 R 2 D ∑ n = 0 ∞ ∑ m = − n n Y m n ( θ , ϕ ) Y m n ∗ ( θ 0 , ϕ 0 ) μ n ( p , ∂ Ω ) , (159) \tilde{G}_0(x,p|x_0)=\frac{1}{R^2D}\sum_{n=0}^{\infty}\sum_{m=-n}^{n}\frac{Y_{mn}(\theta,\phi)Y_{mn}^{*}(\theta_0,\phi_0)}{\mu_n^{(p,\partial\Omega)}}, \tag*{(159)} G ~ 0 ( x , p ∣ x 0 ) = R 2 D 1 n = 0 ∑ ∞ m = − n ∑ n μ n ( p , ∂ Ω ) Y mn ( θ , ϕ ) Y mn ∗ ( θ 0 , ϕ 0 ) , (159)
where
μ n ( p , ∂ Ω ) = p D i n ′ ( R p D ) i n ( R p D ) , (160) \mu_n^{(p,\partial\Omega)}=\sqrt{\frac{p}{D}}\frac{i_n'\left(R\sqrt{\frac{p}{D}}\right)}{i_n\left(R\sqrt{\frac{p}{D}}\right)}, \tag*{(160)} μ n ( p , ∂ Ω ) = D p i n ( R D p ) i n ′ ( R D p ) , (160)
i n ( z ) i_n(z) i n ( z ) is the modified spherical Bessel function of the first kind,
Y m n ( θ , ϕ ) = ( 2 n + 1 ) ( n − m ) ! 4 π ( n + m ) ! P n m ( cos θ ) e i m ϕ (161) Y_{mn}(\theta,\phi)=\sqrt{\frac{(2n+1)(n-m)!}{4\pi(n+m)!}}P_n^m(\cos\theta)e^{im\phi} \tag*{(161)} Y mn ( θ , ϕ ) = 4 π ( n + m )! ( 2 n + 1 ) ( n − m )! P n m ( cos θ ) e im ϕ (161)
are the normalized spherical harmonics, and P n m ( z ) P_n^m(z) P n m ( z ) are the associated Legendre polynomials. Note that the eigenvalue μ n ( p , ∂ Ω ) \mu_n^{(p,\partial\Omega)} μ n ( p , ∂ Ω ) is ( 2 n + 1 ) (2n+1) ( 2 n + 1 ) times degenerate, whereas the eigenfunctions of the Dirichlet-to-Neumann operator M p ( ∂ Ω ) \mathcal{M}_p^{(\partial\Omega)} M p ( ∂ Ω ) , given by spherical harmonics, do not depend on p p p . These properties are specific to the ball and follow from the rotational invariance of the problem. According to Eq. (16) , the restriction of D ~ G ~ 0 ( x , p ∣ x 0 ) \tilde{D}\tilde{G}_0(x,p|x_0) D ~ G ~ 0 ( x , p ∣ x 0 ) to Γ × Γ \Gamma\times\Gamma Γ × Γ defines an integral operator, which determines the eigenvalues and eigenfunctions of the Dirichlet-to-Neumann operator M p ( Γ ) \mathcal{M}_p^{(\Gamma)} M p ( Γ ) . This construction is valid for any 0 < ϵ ≤ π 0<\epsilon\le\pi 0 < ϵ ≤ π and p > 0 p>0 p > 0 , though it requires a numerical treatment. In the following, we focus on the limit p = 0 p=0 p = 0 .
Substituting the expansion
1 μ 0 ( p , ∂ Ω ) ≈ 3 R p / D + R 5 + O ( p ) ( p → 0 ) \frac{1}{\mu_0^{(p,\partial\Omega)}}\approx\frac{3}{Rp/D}+\frac{R}{5}+O(p)\qquad(p\to0) μ 0 ( p , ∂ Ω ) 1 ≈ R p / D 3 + 5 R + O ( p ) ( p → 0 )
into Eq. (23) , we retrieve the restriction of the pseudo-Green’s function to the boundary ∂ Ω \partial\Omega ∂ Ω ,
G 0 ( x ∣ x 0 ) = 1 R ( 1 20 π + ∑ n = 1 ∞ ∑ m = − n n Y m n ( θ , ϕ ) Y m n ∗ ( θ 0 , ϕ 0 ) n ) , (162) G_0(x|x_0)=\frac{1}{R}\left(\frac{1}{20\pi}+\sum_{n=1}^{\infty}\sum_{m=-n}^{n}\frac{Y_{mn}(\theta,\phi)Y_{mn}^{*}(\theta_0,\phi_0)}{n}\right), \tag*{(162)} G 0 ( x ∣ x 0 ) = R 1 ( 20 π 1 + n = 1 ∑ ∞ m = − n ∑ n n Y mn ( θ , ϕ ) Y mn ∗ ( θ 0 , ϕ 0 ) ) , (162)
where we used μ n ( 0 , ∂ Ω ) = n / R \mu_n^{(0,\partial\Omega)}=n/R μ n ( 0 , ∂ Ω ) = n / R . A more explicit form of the pseudo-Green’s function inside a ball is given by Eq. [(36) ] in Ref. [75 ] . Its restriction to the boundary ∂ Ω \partial\Omega ∂ Ω reads:
G 0 ( x ∣ x 0 ) = 1 4 π R { 2 1 − cos γ − 9 5 + ln ( 2 1 − cos γ + 2 ( 1 − cos γ ) ) } , (163) G_0(x|x_0)=\frac{1}{4\pi R}\left\{\frac{\sqrt{2}}{\sqrt{1-\cos\gamma}}-\frac{9}{5}+\ln\left(\frac{2}{1-\cos\gamma+\sqrt{2(1-\cos\gamma)}}\right)\right\}, \tag*{(163)} G 0 ( x ∣ x 0 ) = 4 π R 1 { 1 − cos γ 2 − 5 9 + ln ( 1 − cos γ + 2 ( 1 − cos γ ) 2 ) } , (163)
where γ \gamma γ is the angle between two vectors pointing from the origin at x \boldsymbol{x} x and x 0 \boldsymbol{x}_0 x 0 on the unit sphere:
cos γ = cos θ cos θ 0 + sin θ sin θ 0 cos ( φ − φ 0 ) (164) \cos\gamma= \cos\theta\cos\theta_0 + \sin\theta\sin\theta_0 \cos(\varphi-\varphi_0) \tag*{(164)} cos γ = cos θ cos θ 0 + sin θ sin θ 0 cos ( φ − φ 0 ) (164)
(note that the expression (C.5) can be found in the classical textbook [129 ] , p. 247). In particular, in the limit x → x 0 \boldsymbol{x} \to\boldsymbol{x}_0 x → x 0 , one retrieves the asymptotic behavior (91) , with
R 0 ( x 0 ) = ln ( 2 ) − 9 / 5 + ln R 4 π R , H ( x 0 ) = 1 R , (165) R_0(\boldsymbol{x}_0) = \frac{\ln(2)-9/5+\ln R}{4\pi R}, \qquad H(\boldsymbol{x}_0)=\frac{1}{R}, \tag*{(165)} R 0 ( x 0 ) = 4 π R ln ( 2 ) − 9/5 + ln R , H ( x 0 ) = R 1 , (165)
independently of x 0 \boldsymbol{x}_0 x 0 .
We consider the subset Γ \Gamma Γ to be a spherical cap of angle ε \varepsilon ε around the North pole, defined in the spherical coordinates as Γ = { ( R , θ , φ ) ∈ ∂ Ω : 0 ≤ θ < ε } \Gamma=\{(R,\theta,\varphi)\in\partial\Omega:0\leq\theta<\varepsilon\} Γ = {( R , θ , φ ) ∈ ∂ Ω : 0 ≤ θ < ε } . Using Eq. (C.4), we find then
1 ∣ Γ ∣ ∫ Γ G 0 ( x ∣ x 0 ) d x = 1 4 π R ( 1 5 + ∑ n = 1 ∞ ϕ n n P n ( cos θ 0 ) ) , \frac{1}{|\Gamma|}\int_{\Gamma} \mathcal{G}_0(\boldsymbol{x}|\boldsymbol{x}_0)\,d\boldsymbol{x}=\frac{1}{4\pi R}\left(\frac{1}{5}+\sum_{n=1}^{\infty}\frac{\phi_n}{n}P_n(\cos\theta_0)\right), ∣Γ∣ 1 ∫ Γ G 0 ( x ∣ x 0 ) d x = 4 π R 1 ( 5 1 + n = 1 ∑ ∞ n ϕ n P n ( cos θ 0 ) ) ,
where ∣ Γ ∣ = 2 π R 2 ( 1 − cos ε ) |\Gamma|=2\pi R^2(1-\cos\varepsilon) ∣Γ∣ = 2 π R 2 ( 1 − cos ε ) is the area of the spherical cap Γ \Gamma Γ , P n ( z ) P_n(z) P n ( z ) are Legendre polynomials, and
ϕ n = P n − 1 ( cos ε ) − P n + 1 ( cos ε ) 1 − cos ε . (166) \phi_n=\frac{P_{n-1}(\cos\varepsilon)-P_{n+1}(\cos\varepsilon)}{1-\cos\varepsilon}. \tag*{(166)} ϕ n = 1 − cos ε P n − 1 ( cos ε ) − P n + 1 ( cos ε ) . (166)
As a consequence, we find
ω 0 ( Γ ) ( x 0 ) = − A Γ + 1 4 π R ( 1 5 + ∑ n = 1 ∞ ϕ n n P n ( cos θ 0 ) ) , (167) \omega_0^{(\Gamma)}(\boldsymbol{x}_0)=-A_\Gamma+\frac{1}{4\pi R}\left(\frac{1}{5}+\sum_{n=1}^{\infty}\frac{\phi_n}{n}P_n(\cos\theta_0)\right), \tag*{(167)} ω 0 ( Γ ) ( x 0 ) = − A Γ + 4 π R 1 ( 5 1 + n = 1 ∑ ∞ n ϕ n P n ( cos θ 0 ) ) , (167)
A Γ = 1 4 π R ( 1 5 + ∑ n = 1 ∞ ϕ n 2 n ( 2 n + 1 ) ) . (168) A_\Gamma=\frac{1}{4\pi R}\left(\frac{1}{5}+\sum_{n=1}^{\infty}\frac{\phi_n^2}{n(2n+1)}\right). \tag*{(168)} A Γ = 4 π R 1 ( 5 1 + n = 1 ∑ ∞ n ( 2 n + 1 ) ϕ n 2 ) . (168)
Substituting Eq. (C.7) into Eq. (92) gives the asymptotic behavior of A Γ A_\Gamma A Γ as ε → 0 \varepsilon\to0 ε → 0 ,
A Γ ≈ 1 4 π R ( 32 3 π ε + ln ( 1 / ε ) − 31 20 + ln 2 + O ( ε ) ) , (169) A_\Gamma\approx\frac{1}{4\pi R}\left(\frac{32}{3\pi\varepsilon}+\ln(1/\varepsilon)-\frac{31}{20}+\ln2+O(\varepsilon)\right), \tag*{(169)} A Γ ≈ 4 π R 1 ( 3 π ε 32 + ln ( 1/ ε ) − 20 31 + ln 2 + O ( ε ) ) , (169)
that was earlier derived by a different method in [83 ] . Combining the above results, we get the kernel
G ( x ∣ x 0 ) = 1 4 π R { 4 π ∑ n = 1 ∞ ∑ m = − n n Y m n ( θ , φ ) Y m n ∗ ( θ 0 , φ 0 ) n − ∑ n = 1 ∞ ( ϕ n n [ P n ( cos θ 0 ) + P n ( cos θ ) ] − ϕ n 2 n ( 2 n + 1 ) ) } , (170) \mathcal{G}(\boldsymbol{x}|\boldsymbol{x}_0)=\frac{1}{4\pi R}\left\{4\pi\sum_{n=1}^{\infty}\sum_{m=-n}^{n}\frac{Y_{mn}(\theta,\varphi)Y_{mn}^{*}(\theta_0,\varphi_0)}{n}-\sum_{n=1}^{\infty}\left(\frac{\phi_n}{n}\left[P_n(\cos\theta_0)+P_n(\cos\theta)\right]-\frac{\phi_n^2}{n(2n+1)}\right)\right\}, \tag*{(170)} G ( x ∣ x 0 ) = 4 π R 1 { 4 π n = 1 ∑ ∞ m = − n ∑ n n Y mn ( θ , φ ) Y mn ∗ ( θ 0 , φ 0 ) − n = 1 ∑ ∞ ( n ϕ n [ P n ( cos θ 0 ) + P n ( cos θ ) ] − n ( 2 n + 1 ) ϕ n 2 ) } , (170)
restricted to Γ × Γ \Gamma\times\Gamma Γ × Γ , that determines the eigenvalues and eigenfunctions of the Dirichlet-to-Neumann operator M 0 ( Γ ) \mathcal{M}_0^{(\Gamma)} M 0 ( Γ ) for the mixed Steklov-Neumann problem. The axial symmetry of the problem allows one to decompose the kernel as
G ( x ∣ x 0 ) = 1 2 π R ∑ m = − ∞ ∞ e i m ( φ − φ 0 ) G ( m ) ( θ ∣ θ 0 ) , (171) \mathcal{G}(\boldsymbol{x}|\boldsymbol{x}_0)=\frac{1}{2\pi R}\sum_{m=-\infty}^{\infty}e^{im(\varphi-\varphi_0)}\mathcal{G}^{(m)}(\theta|\theta_0), \tag*{(171)} G ( x ∣ x 0 ) = 2 π R 1 m = − ∞ ∑ ∞ e im ( φ − φ 0 ) G ( m ) ( θ ∣ θ 0 ) , (171)
with the one-dimensional kernels
G ( 0 ) ( θ ∣ θ 0 ) = ∑ n = 1 ∞ { n + 1 / 2 n P n ( cos θ ) P n ( cos θ 0 ) − ϕ n n [ P n ( cos θ 0 ) + P n ( cos θ ) ] + ϕ n 2 n ( 2 n + 1 ) } (172) \mathcal{G}^{(0)}(\theta|\theta_0)=\sum_{n=1}^{\infty}\left\{\frac{n+1/2}{n}P_n(\cos\theta)P_n(\cos\theta_0)-\frac{\phi_n}{n}\left[P_n(\cos\theta_0)+P_n(\cos\theta)\right]+\frac{\phi_n^2}{n(2n+1)}\right\} \tag*{(172)} G ( 0 ) ( θ ∣ θ 0 ) = n = 1 ∑ ∞ { n n + 1/2 P n ( cos θ ) P n ( cos θ 0 ) − n ϕ n [ P n ( cos θ 0 ) + P n ( cos θ ) ] + n ( 2 n + 1 ) ϕ n 2 } (172)
G ( m ) ( θ ∣ θ 0 ) = ∑ n = m ∞ ( n − m ) ! ( n + m ) ! n + 1 / 2 n P n m ( cos θ ) P n m ( cos θ 0 ) ( m ≥ 1 ) , (173) \mathcal{G}^{(m)}(\theta|\theta_0)=\sum_{n=m}^{\infty}\frac{(n-m)!}{(n+m)!}\frac{n+1/2}{n}P_n^m(\cos\theta)P_n^m(\cos\theta_0)\qquad(m\ge1), \tag*{(173)} G ( m ) ( θ ∣ θ 0 ) = n = m ∑ ∞ ( n + m )! ( n − m )! n n + 1/2 P n m ( cos θ ) P n m ( cos θ 0 ) ( m ≥ 1 ) , (173)
G ( m ) ( θ ∣ θ 0 ) = G ( − m ) ( θ ∣ θ 0 ) ( m ≤ − 1 ) . (174) \mathcal{G}^{(m)}(\theta|\theta_0)=\mathcal{G}^{(-m)}(\theta|\theta_0)\qquad(m\le-1). \tag*{(174)} G ( m ) ( θ ∣ θ 0 ) = G ( − m ) ( θ ∣ θ 0 ) ( m ≤ − 1 ) . (174)
As a consequence, the eigenfunctions of the kernel G ( x ∣ x 0 ) \mathcal{G}(x|x_0) G ( x ∣ x 0 ) are of the form e i m φ v m , n ( 0 , Γ ) ( θ ) e^{im\varphi}v_{m,n}^{(0,\Gamma)}(\theta) e im φ v m , n ( 0 , Γ ) ( θ ) , where the second factor is the n n n -th eigenfunction of the kernel G ( m ) ( θ ∣ θ 0 ) \mathcal{G}^{(m)}(\theta|\theta_0) G ( m ) ( θ ∣ θ 0 ) :
∫ 0 ε d θ sin θ G ( m ) ( θ ∣ θ 0 ) v m , n ( 0 , Γ ) ( θ ) = 1 R μ m , n ( 0 , Γ ) v m , n ( 0 , Γ ) ( θ 0 ) ( 0 < θ 0 < ε ) (175) \int_{0}^{\varepsilon}d\theta\sin\theta\,\mathcal{G}^{(m)}(\theta|\theta_0)v_{m,n}^{(0,\Gamma)}(\theta)=\frac{1}{R\mu_{m,n}^{(0,\Gamma)}}v_{m,n}^{(0,\Gamma)}(\theta_0)\qquad(0<\theta_0<\varepsilon) \tag*{(175)} ∫ 0 ε d θ sin θ G ( m ) ( θ ∣ θ 0 ) v m , n ( 0 , Γ ) ( θ ) = R μ m , n ( 0 , Γ ) 1 v m , n ( 0 , Γ ) ( θ 0 ) ( 0 < θ 0 < ε ) (175)
(the additional index m m m is used to distinguish different kernels G ( m ) ( θ ∣ θ 0 ) \mathcal{G}^{(m)}(\theta|\theta_0) G ( m ) ( θ ∣ θ 0 ) and their eigenmodes). For convenience, we used here a complex-valued form; however, the factors e ± i m φ e^{\pm im\varphi} e ± im φ can be replaced by a linear combination of sine and cosine functions that we employed in § 3.1 .
One sees that the original problem on the two-dimensional spherical cap Γ \Gamma Γ is decomposed into separate spectral problems for the kernels G ( m ) ( θ ∣ θ 0 ) \mathcal{G}^{(m)}(\theta|\theta_0) G ( m ) ( θ ∣ θ 0 ) . When the target covers the whole sphere (i.e., Γ = ∂ Ω \Gamma=\partial\Omega Γ = ∂ Ω ), it is easy to check that P n m ( cos θ ) P_n^m(\cos\theta) P n m ( cos θ ) is an eigenfunction of the kernel G ( m ) ( θ ∣ θ 0 ) \mathcal{G}^{(m)}(\theta|\theta_0) G ( m ) ( θ ∣ θ 0 ) , whereas the associated eigenvalue is R / n R/n R / n , as it should be (given that n / R n/R n / R is an eigenvalue of the Dirichlet-to-Neumann operator M ( ∂ Ω ) \mathcal{M}^{(\partial\Omega)} M ( ∂ Ω ) on the sphere).
In the limit ε → 0 \varepsilon\to0 ε → 0 , one can investigate the asymptotic behavior of Eqs. (C.9b, C.9a, C.13a). For instance, one can check that the leading term of each kernel G ( m ) ( θ ∣ θ 0 ) \mathcal{G}^{(m)}(\theta|\theta_0) G ( m ) ( θ ∣ θ 0 ) is ε − 1 G ^ ( m ) ( θ / ε ∣ θ 0 / ε ) \varepsilon^{-1}\widehat{\mathcal{G}}^{(m)}(\theta/\varepsilon|\theta_0/\varepsilon) ε − 1 G ( m ) ( θ / ε ∣ θ 0 / ε ) , given by Eqs. (100) , (102) , as expected. As these cumbersome computations do not bring new insights, they are not presented.
Numerical implementation
By substituting x = cos θ x=\cos\theta x = cos θ , x 0 = cos θ 0 x_0=\cos\theta_0 x 0 = cos θ 0 and a = cos ε a=\cos\varepsilon a = cos ε , it is convenient to rewrite the spectral problem (C.14) as
∫ a 1 d x G ( m ) ( x ∣ x 0 ) v m , n ( 0 , Γ ) ( x ) = 1 μ m , n ( 0 , Γ ) v m , n ( 0 , Γ ) ( x 0 ) ( a < x 0 < 1 ) , (176) \int_{a}^{1}dx\,\mathcal{G}^{(m)}(x|x_0)v_{m,n}^{(0,\Gamma)}(x)=\frac{1}{\mu_{m,n}^{(0,\Gamma)}}v_{m,n}^{(0,\Gamma)}(x_0)\qquad(a<x_0<1), \tag*{(176)} ∫ a 1 d x G ( m ) ( x ∣ x 0 ) v m , n ( 0 , Γ ) ( x ) = μ m , n ( 0 , Γ ) 1 v m , n ( 0 , Γ ) ( x 0 ) ( a < x 0 < 1 ) , (176)
where G ( m ) ( cos θ ∣ cos θ 0 ) = G ( m ) ( θ ∣ θ 0 ) \mathcal{G}^{(m)}(\cos\theta|\cos\theta_0)=\mathcal{G}^{(m)}(\theta|\theta_0) G ( m ) ( cos θ ∣ cos θ 0 ) = G ( m ) ( θ ∣ θ 0 ) . The interval ( a , 1 ) (a,1) ( a , 1 ) is discretized into N N N segments, centered at x n = a + ( 1 − a ) ( n − 1 / 2 ) / N x_n=a+(1-a)(n-1/2)/N x n = a + ( 1 − a ) ( n − 1/2 ) / N , with n = 1 , 2 , … , N n=1,2,\ldots,N n = 1 , 2 , … , N . Setting η = ( 1 − a ) / ( 2 N ) \eta=(1-a)/(2N) η = ( 1 − a ) / ( 2 N ) and using the identity
( 2 k + 1 ) P k ( x ) = d d x ( P k + 1 ( x ) − P k − 1 ( x ) ) , (177) (2k+1)P_k(x)=\frac{d}{dx}\left(P_{k+1}(x)-P_{k-1}(x)\right), \tag*{(177)} ( 2 k + 1 ) P k ( x ) = d x d ( P k + 1 ( x ) − P k − 1 ( x ) ) , (177)
we evaluate the matrix elements for m = 0 m=0 m = 0 as
G n , n ′ ( ε ) = ∫ x n − η x n + η d x G ( 0 ) ′ ( x ∣ x n ′ ) ≈ R 2 ∑ k = 1 k max 1 k { ( P k ( x n ′ ) − ϕ k 2 k + 1 ) × ( [ P k + 1 ( x n + η ) − P k − 1 ( x n + η ) ] − [ P k + 1 ( x n − η ) − P k − 1 ( x n − η ) ] ) − 2 η ( ϕ k P k ( x n ′ ) − ϕ k 2 2 k + 1 ) } . (178) \mathbf{G}^{(\varepsilon)}_{n,n'}=\int_{x_n-\eta}^{x_n+\eta}dx\,G^{(0)\prime}(x|x_{n'})\approx\frac{R}{2}\sum_{k=1}^{k_{\max}}\frac{1}{k}\left\{\left(P_k(x_{n'})-\frac{\phi_k}{2k+1}\right)\times\left([P_{k+1}(x_n+\eta)-P_{k-1}(x_n+\eta)]-[P_{k+1}(x_n-\eta)-P_{k-1}(x_n-\eta)]\right)-2\eta\left(\phi_kP_k(x_{n'})-\frac{\phi_k^2}{2k+1}\right)\right\}. \tag*{(178)} G n , n ′ ( ε ) = ∫ x n − η x n + η d x G ( 0 ) ′ ( x ∣ x n ′ ) ≈ 2 R k = 1 ∑ k m a x k 1 { ( P k ( x n ′ ) − 2 k + 1 ϕ k ) × ( [ P k + 1 ( x n + η ) − P k − 1 ( x n + η )] − [ P k + 1 ( x n − η ) − P k − 1 ( x n − η )] ) − 2 η ( ϕ k P k ( x n ′ ) − 2 k + 1 ϕ k 2 ) } . (178)
where the infinite sum was truncated up to some k max k_{\max} k m a x . A numerical diagonalization of the matrix G ( ε ) \mathbf{G}^{(\varepsilon)} G ( ε ) of size N × N N\times N N × N allows one to approximate the axially symmetric eigenfunctions v 0 , n ( 0 , Γ ) ( x ) v_{0,n}^{(0,\Gamma)}(x) v 0 , n ( 0 , Γ ) ( x ) and the associated eigenvalues μ 0 , n ( 0 , Γ ) \mu_{0,n}^{(0,\Gamma)} μ 0 , n ( 0 , Γ ) . A similar computation can be done for kernels with m = 1 , 2 , … m=1,2,\ldots m = 1 , 2 , … to approximate v m , n ( 0 , Γ ) ( x ) v_{m,n}^{(0,\Gamma)}(x) v m , n ( 0 , Γ ) ( x ) and μ m , n ( 0 , Γ ) \mu_{m,n}^{(0,\Gamma)} μ m , n ( 0 , Γ ) .
In the same vein, one can discretize the integral in Eq. (103) and then diagonalize the resulting matrix to approximate the eigenvalues μ ^ m , n \hat{\mu}_{m,n} μ ^ m , n and eigenfunctions v ^ m , n \hat{v}_{m,n} v ^ m , n . However, as the semi-analytical method described in Appendix F is much faster and more accurate, we do not present the details of the direct diagonalization.
Appendix D. Variance of the boundary local time
In this Appendix, we sketch the main steps to get the variance of the boundary local time by following Ref. [17 ] (see also [126 ] ).
Let X t X_t X t denote the reflected Brownian motion in a bounded domain Ω ⊂ R d \Omega\subset\mathbb{R}^{d} Ω ⊂ R d with a smooth boundary ∂ Ω \partial\Omega ∂ Ω (see [119 , 120 , 121 , 122 ] for a mathematical formulation and [18 ] for physical insights). The boundary local time ℓ t \ell_t ℓ t on a subset Γ \Gamma Γ of the boundary ∂ Ω \partial\Omega ∂ Ω can be introduced as the rescaled residence time in a thin boundary layer Γ a = { x ∈ Ω : ∣ x − Γ ∣ < a } \Gamma_a=\{x\in\Omega:|x-\Gamma|<a\} Γ a = { x ∈ Ω : ∣ x − Γ∣ < a } of thickness a a a near Γ \Gamma Γ :
ℓ t = lim a → 0 D a ∫ 0 t d t ′ Θ ( a − ∣ X t ′ − Γ ∣ ) , (179) \ell_t=\lim_{a\to0}\frac{D}{a}\int_0^t dt'\,\Theta(a-|X_{t'}-\Gamma|), \tag*{(179)} ℓ t = a → 0 lim a D ∫ 0 t d t ′ Θ ( a − ∣ X t ′ − Γ∣ ) , (179)
where ∣ X t ′ − Γ ∣ |X_{t'}-\Gamma| ∣ X t ′ − Γ∣ is the Euclidean distance between the position X t ′ X_{t'} X t ′ at time t ′ t' t ′ and the subset Γ \Gamma Γ , and Θ ( z ) \Theta(z) Θ ( z ) is the Heaviside step function: Θ ( z ) = 1 \Theta(z)=1 Θ ( z ) = 1 for z > 0 z>0 z > 0 and 0 otherwise. In other words, the above integral determines how long the diffusing particle stayed near the subset Γ \Gamma Γ up to time t t t . As the layer thickness a a a goes to 0, the residence time also vanishes but its rescaling by a a a yields a nontrivial limit – the boundary local time ℓ t \ell_t ℓ t (despite its name, ℓ t \ell_t ℓ t has units of length, given that the diffusivity D D D has units of squared length over time).
In [17 ] , the distribution of the boundary local time on the boundary ∂ Ω \partial\Omega ∂ Ω was obtained in terms of the Steklov eigenpairs. A straightforward extension of its derivation to the case of a subset Γ \Gamma Γ consists in using the eigenpairs of the Steklov-Neumann problem. In particular, Eq. [37] from [17 ] for the n n n -th order moment of ℓ t \ell_t ℓ t , L n ( t ∣ x 0 ) = E x 0 { ℓ t n } L_n(t|x_0)=\mathbb{E}_{x_0}\{\ell_t^n\} L n ( t ∣ x 0 ) = E x 0 { ℓ t n } , is generalized in our setting as (see also [18 ] )
L ~ n ( p ∣ x 0 ) = ∫ 0 ∞ d t e − p t L n ( t ∣ x 0 ) = ∑ k = 0 ∞ V k ( p , Γ ) ( x 0 ) [ μ k ( p , Γ ) ] n ∫ Γ v k ( p , Γ ) (180) \widetilde{L}_n(p|x_0)=\int_0^\infty dt\,e^{-pt}L_n(t|x_0)=\sum_{k=0}^\infty\frac{V_k^{(p,\Gamma)}(x_0)}{[\mu_k^{(p,\Gamma)}]^n}\int_\Gamma v_k^{(p,\Gamma)} \tag*{(180)} L n ( p ∣ x 0 ) = ∫ 0 ∞ d t e − pt L n ( t ∣ x 0 ) = k = 0 ∑ ∞ [ μ k ( p , Γ ) ] n V k ( p , Γ ) ( x 0 ) ∫ Γ v k ( p , Γ ) (180)
where x 0 ∈ Ω ‾ x_0\in\overline{\Omega} x 0 ∈ Ω is the starting point of the particle. According to the properties of the Laplace transform [127 ] , the growth of the moments L n ( t ∣ x 0 ) L_n(t|x_0) L n ( t ∣ x 0 ) at long times can be determined from the asymptotic behavior of L ~ n ( p ∣ x 0 ) \widetilde{L}_n(p|x_0) L n ( p ∣ x 0 ) as p → 0 p \to0 p → 0 . Using Eqs. (2.21), we get
L ~ 1 ( p ∣ x 0 ) ≈ D ∣ Γ ∣ ∣ Ω ∣ p 2 ( 1 + p ∣ Ω ∣ D ( A Γ + W 0 ( Γ ) ( x 0 ) ) + O ( p 2 ) ) (181) \widetilde{L}_1(p|x_0) \approx\frac{D|\Gamma|}{|\Omega|p^2}\left(1+p\frac{|\Omega|}{D}\left(A_\Gamma+W_0^{(\Gamma)}(x_0)\right)+O(p^2)\right) \tag*{(181)} L 1 ( p ∣ x 0 ) ≈ ∣Ω∣ p 2 D ∣Γ∣ ( 1 + p D ∣Ω∣ ( A Γ + W 0 ( Γ ) ( x 0 ) ) + O ( p 2 ) ) (181)
L ~ 2 ( p ∣ x 0 ) ≈ 2 D 2 ∣ Γ ∣ 2 ∣ Ω ∣ 2 p 3 ( 1 + p ∣ Ω ∣ D ( 2 A Γ + W 0 ( Γ ) ( x 0 ) ) + O ( p 2 ) ) (182) \widetilde{L}_2(p|x_0) \approx\frac{2D^2|\Gamma|^2}{|\Omega|^2p^3}\left(1+p\frac{|\Omega|}{D}\left(2A_\Gamma+W_0^{(\Gamma)}(x_0)\right)+O(p^2)\right) \tag*{(182)} L 2 ( p ∣ x 0 ) ≈ ∣Ω ∣ 2 p 3 2 D 2 ∣Γ ∣ 2 ( 1 + p D ∣Ω∣ ( 2 A Γ + W 0 ( Γ ) ( x 0 ) ) + O ( p 2 ) ) (182)
from which a formal Laplace transform inversion yields
L 1 ( t ∣ x 0 ) ≈ D ∣ Γ ∣ ∣ Ω ∣ ( t + ∣ Ω ∣ D ( A Γ + W 0 ( Γ ) ( x 0 ) ) + … ) (183) L_1(t|x_0) \approx\frac{D|\Gamma|}{|\Omega|}\left(t+\frac{|\Omega|}{D}\left(A_\Gamma+W_0^{(\Gamma)}(x_0)\right)+\ldots\right) \tag*{(183)} L 1 ( t ∣ x 0 ) ≈ ∣Ω∣ D ∣Γ∣ ( t + D ∣Ω∣ ( A Γ + W 0 ( Γ ) ( x 0 ) ) + … ) (183)
L 2 ( t ∣ x 0 ) ≈ D 2 ∣ Γ ∣ 2 ∣ Ω ∣ 2 ( t 2 + 2 t ∣ Ω ∣ D ( 2 A Γ + W 0 ( Γ ) ( x 0 ) ) + … ) (184) L_2(t|x_0) \approx\frac{D^2|\Gamma|^2}{|\Omega|^2}\left(t^2+2t\frac{|\Omega|}{D}\left(2A_\Gamma+W_0^{(\Gamma)}(x_0)\right)+\ldots\right) \tag*{(184)} L 2 ( t ∣ x 0 ) ≈ ∣Ω ∣ 2 D 2 ∣Γ ∣ 2 ( t 2 + 2 t D ∣Ω∣ ( 2 A Γ + W 0 ( Γ ) ( x 0 ) ) + … ) (184)
These relations imply the long-time asymptotic behavior (2.46) of the variance:
Var x 0 { ℓ t } = L 2 ( t ∣ x 0 ) − L 1 2 ( t ∣ x 0 ) . (185) \operatorname{Var}_{x_0}\{\ell_t\}=L_2(t|x_0)-L_1^2(t|x_0). \tag*{(185)} Var x 0 { ℓ t } = L 2 ( t ∣ x 0 ) − L 1 2 ( t ∣ x 0 ) . (185)
Expectedly, the leading-order term does not depend on the starting point x 0 x_0 x 0 , which appears in the subleading correction (not shown).
Exterior Steklov problem for an ellipse.
The auxiliary spectral problem (3.3) in the upper half-plane H 2 = R × R + \mathbb{H}_2=\mathbb{R}\times\mathbb{R}_+ H 2 = R × R + is actually the specific case of the Steklov problem in the exterior of an ellipse, Ω = { ( x , y ) ∈ R 2 : ( x / a ) 2 + ( y / b ) 2 > 1 } \Omega=\{(x,y)\in\mathbb{R}^2:(x/a)^2+(y/b)^2>1\} Ω = {( x , y ) ∈ R 2 : ( x / a ) 2 + ( y / b ) 2 > 1 } , with semi-axes a > b a>b a > b (Figure 2 e):
Δ V k = 0 in Ω , (186) \Delta V_k=0 \quad\text{in }\Omega, \tag*{(186)} Δ V k = 0 in Ω , (186)
∂ n V k = μ k V k on ∂ Ω , (187) \partial_n V_k=\mu_k V_k \quad\text{on }\partial\Omega, \tag*{(187)} ∂ n V k = μ k V k on ∂ Ω , (187)
∣ x ∣ ∣ ∇ V k ∣ → 0 ( ∣ x ∣ → ∞ ) (188) |x|\,|\nabla V_k|\to0 \quad(|x|\to\infty) \tag*{(188)} ∣ x ∣ ∣∇ V k ∣ → 0 ( ∣ x ∣ → ∞ ) (188)
(see [131 , 132 ] for more details). In this Appendix, we briefly discuss this more general problem by reproducing the analysis of the exterior Steklov problem for oblate spheroids [102 ] .
To solve the Laplace equation (E.1a), it is convenient to use the elliptic coordinates ( α , θ ) (\alpha,\theta) ( α , θ ) ,
x = a E cosh α cos θ , y = a E sinh α sin θ , (189) x=a_E\cosh\alpha\cos\theta,\qquad y=a_E\sinh\alpha\sin\theta, \tag*{(189)} x = a E cosh α cos θ , y = a E sinh α sin θ , (189)
with 0 ≤ α < + ∞ 0\leq\alpha<+\infty 0 ≤ α < + ∞ , − π < θ ≤ π -\pi<\theta\leq\pi − π < θ ≤ π , and a E = a 2 − b 2 a_E=\sqrt{a^2-b^2} a E = a 2 − b 2 being one-half of the focal distance. In these coordinates, the exterior of an ellipse reads Ω = { α > α 0 , − π < θ ≤ π } \Omega=\{\alpha>\alpha_0,\ -\pi<\theta\leq\pi\} Ω = { α > α 0 , − π < θ ≤ π } , with tanh α 0 = b / a \tanh\alpha_0=b/a tanh α 0 = b / a . The Steklov boundary condition is
∂ n V k ∣ ∂ Ω = − 1 h α ∂ α V k ∣ α = α 0 = μ k V k ∣ α = α 0 , (190) \left.\partial_nV_k\right|_{\partial\Omega}=-\left.\frac{1}{h_\alpha}\partial_\alpha V_k\right|_{\alpha=\alpha_0}=\left.\mu_kV_k\right|_{\alpha=\alpha_0}, \tag*{(190)} ∂ n V k ∣ ∂ Ω = − h α 1 ∂ α V k α = α 0 = μ k V k ∣ α = α 0 , (190)
where h α = a E cosh 2 α − cos 2 θ h_\alpha=a_E\sqrt{\cosh^2\alpha-\cos^2\theta} h α = a E cosh 2 α − cos 2 θ is the scale factor accounting for the curvature. Since the Laplace operator reads
Δ = 1 a E 2 ( cosh 2 α − cos 2 θ ) ( ∂ α 2 + ∂ θ 2 ) . (191) \Delta=\frac{1}{a_E^2(\cosh^2\alpha-\cos^2\theta)}\left(\partial_\alpha^2+\partial_\theta^2\right). \tag*{(191)} Δ = a E 2 ( cosh 2 α − cos 2 θ ) 1 ( ∂ α 2 + ∂ θ 2 ) . (191)
a Steklov eigenfunction can be searched in either of two forms:
V 2 k ( α , θ ) = ∑ n = 0 ∞ c 2 k , n cos ( n θ ) e − n ( α − α 0 ) , (192) V_{2k}(\alpha,\theta)=\sum_{n=0}^{\infty}c_{2k,n}\cos(n\theta)e^{-n(\alpha-\alpha_0)}, \tag*{(192)} V 2 k ( α , θ ) = n = 0 ∑ ∞ c 2 k , n cos ( n θ ) e − n ( α − α 0 ) , (192)
V 2 k + 1 ( α , θ ) = ∑ n = 1 ∞ c 2 k + 1 , n sin ( n θ ) e − n ( α − α 0 ) , (193) V_{2k+1}(\alpha,\theta)=\sum_{n=1}^{\infty}c_{2k+1,n}\sin(n\theta)e^{-n(\alpha-\alpha_0)}, \tag*{(193)} V 2 k + 1 ( α , θ ) = n = 1 ∑ ∞ c 2 k + 1 , n sin ( n θ ) e − n ( α − α 0 ) , (193)
with unknown coefficients c k , n c_{k,n} c k , n . Even and odd indices are used to explicitly distinguish eigenfunctions that are symmetric and antisymmetric with respect to the horizontal axis:
V 2 k ( α , − θ ) = V 2 k ( α , θ ) ⇔ V 2 k ( x , − y ) = V 2 k ( x , y ) , V_{2k}(\alpha,-\theta)=V_{2k}(\alpha,\theta)\quad\Leftrightarrow\quad V_{2k}(x,-y)=V_{2k}(x,y), V 2 k ( α , − θ ) = V 2 k ( α , θ ) ⇔ V 2 k ( x , − y ) = V 2 k ( x , y ) ,
V 2 k + 1 ( α , − θ ) = − V 2 k + 1 ( α , θ ) ⇔ V 2 k + 1 ( x , − y ) = − V 2 k + 1 ( x , y ) V_{2k+1}(\alpha,-\theta)=-V_{2k+1}(\alpha,\theta)\quad\Leftrightarrow\quad V_{2k+1}(x,-y)=-V_{2k+1}(x,y) V 2 k + 1 ( α , − θ ) = − V 2 k + 1 ( α , θ ) ⇔ V 2 k + 1 ( x , − y ) = − V 2 k + 1 ( x , y )
(with some abuse of notations, we wrote the symmetries in both elliptic and Cartesian coordinates). These symmetries imply that any eigenfunction V 2 k ( x , y ) V_{2k}(x,y) V 2 k ( x , y ) is actually the solution of the mixed Steklov-Neumann problem in the exterior of the half-ellipse in the upper half-plane (Figure 2 d):
Δ V 2 k = 0 ( in Ω ∩ H 2 ) , (194) \Delta V_{2k}=0\qquad(\text{in }\Omega\cap H_2), \tag*{(194)} Δ V 2 k = 0 ( in Ω ∩ H 2 ) , (194)
∂ n V 2 k = μ 2 k V 2 k ( on Γ ∩ H 2 ) , (195) \partial_n V_{2k}=\mu_{2k}V_{2k}\qquad(\text{on }\Gamma\cap H_2), \tag*{(195)} ∂ n V 2 k = μ 2 k V 2 k ( on Γ ∩ H 2 ) , (195)
− ∂ y V 2 k = 0 ( y = 0 , ∣ x ∣ ≥ a ) (196) -\partial_y V_{2k}=0\qquad(y=0,\ |x|\ge a) \tag*{(196)} − ∂ y V 2 k = 0 ( y = 0 , ∣ x ∣ ≥ a ) (196)
(in turn, antisymmetric eigenfunctions V 2 k + 1 ( x , y ) V_{2k+1}(x,y) V 2 k + 1 ( x , y ) solve the mixed Steklov-Dirichlet problem). In particular, one retrieves the auxiliary spectral problem (3.3) by setting a = 1 a=1 a = 1 and b = 0 b=0 b = 0 . In the following, we focus on the symmetric eigenfunctions V 2 k V_{2k} V 2 k .
Since the Steklov condition (E.3) should be satisfied for any θ \theta θ , it is convenient to multiply it by h α 0 ( θ ) cos ( m θ ) h_{\alpha_0}(\theta)\cos(m\theta) h α 0 ( θ ) cos ( m θ ) and integrate over θ \theta θ from 0 0 0 to π \pi π to get
π 2 m c 2 k , m = μ 2 k ∑ n = 0 ∞ c 2 k , n ∫ 0 π d θ h α 0 ( θ ) cos ( n θ ) cos ( m θ ) ( m = 1 , 2 , … ) . (197) \frac{\pi}{2}mc_{2k,m}=\mu_{2k}\sum_{n=0}^{\infty}c_{2k,n}\int_0^\pi d\theta\,h_{\alpha_0}(\theta)\cos(n\theta)\cos(m\theta)\qquad(m=1,2,\ldots). \tag*{(197)} 2 π m c 2 k , m = μ 2 k n = 0 ∑ ∞ c 2 k , n ∫ 0 π d θ h α 0 ( θ ) cos ( n θ ) cos ( m θ ) ( m = 1 , 2 , … ) . (197)
Denoting
A n , m ( α 0 ) = 2 π ∫ 0 π d θ cos ( m θ ) cos ( n θ ) cosh 2 α 0 − cos 2 θ , (198) A_{n,m}(\alpha_0)=\frac{2}{\pi}\int_0^\pi d\theta\,\cos(m\theta)\cos(n\theta)\sqrt{\cosh^2\alpha_0-\cos^2\theta}, \tag*{(198)} A n , m ( α 0 ) = π 2 ∫ 0 π d θ cos ( m θ ) cos ( n θ ) cosh 2 α 0 − cos 2 θ , (198)
we rewrite the above equation as
m a E μ 2 k c 2 k , m = c 2 k , 0 A m , 0 ( α 0 ) + ∑ n = 1 ∞ c 2 k , n A n , m ( α 0 ) . (199) \frac{m}{a_E\mu_{2k}}c_{2k,m}=c_{2k,0}A_{m,0}(\alpha_0)+\sum_{n=1}^{\infty}c_{2k,n}A_{n,m}(\alpha_0). \tag*{(199)} a E μ 2 k m c 2 k , m = c 2 k , 0 A m , 0 ( α 0 ) + n = 1 ∑ ∞ c 2 k , n A n , m ( α 0 ) . (199)
In turn, multiplying Eq. (E.3) by h α 0 ( θ ) h_{\alpha_0}(\theta) h α 0 ( θ ) and integrating over θ \theta θ from 0 0 0 to π \pi π yield
c 2 k , 0 = − 1 A 0 , 0 ( α 0 ) ∑ n = 1 ∞ c 2 k , n A n , 0 ( α 0 ) . (200) c_{2k,0}=-\frac{1}{A_{0,0}(\alpha_0)}\sum_{n=1}^{\infty}c_{2k,n}A_{n,0}(\alpha_0). \tag*{(200)} c 2 k , 0 = − A 0 , 0 ( α 0 ) 1 n = 1 ∑ ∞ c 2 k , n A n , 0 ( α 0 ) . (200)
Substituting this expression into Eq. (E.9), we finally get the matrix spectral problem
∑ n = 1 ∞ c 2 k , n M n , m ( α 0 ) = 1 a E μ 2 k c 2 k , m ( m = 1 , 2 , … ) , (201) \sum_{n=1}^{\infty}c_{2k,n}M_{n,m}(\alpha_0)=\frac{1}{a_E\mu_{2k}}c_{2k,m}\qquad(m=1,2,\ldots), \tag*{(201)} n = 1 ∑ ∞ c 2 k , n M n , m ( α 0 ) = a E μ 2 k 1 c 2 k , m ( m = 1 , 2 , … ) , (201)
where
M n , m ( α 0 ) = 1 m [ A n , m ( α 0 ) − A n , 0 ( α 0 ) A 0 , m ( α 0 ) A 0 , 0 ( α 0 ) ] . (202) \mathbf{M}_{n,m}(\alpha_0)=\frac{1}{m}\left[\mathbf{A}_{n,m}(\alpha_0)-\frac{\mathbf{A}_{n,0}(\alpha_0)\mathbf{A}_{0,m}(\alpha_0)}{\mathbf{A}_{0,0}(\alpha_0)}\right]. \tag*{(202)} M n , m ( α 0 ) = m 1 [ A n , m ( α 0 ) − A 0 , 0 ( α 0 ) A n , 0 ( α 0 ) A 0 , m ( α 0 ) ] . (202)
A numerical diagonalization of the truncated matrix M ( α 0 ) \mathbf{M}(\alpha_0) M ( α 0 ) allows one to approximate the eigenvalues μ 2 k \mu_{2k} μ 2 k and to construct the symmetric eigenfunctions V 2 k V_{2k} V 2 k of the exterior Steklov problem.
Note that the antisymmetric eigenmodes satisfy a similar eigenvalue problem,
∑ n = 1 ∞ c 2 k + 1 , n M n , m ′ ( α 0 ) = 1 a E μ 2 k + 1 c 2 k + 1 , m ( m = 1 , 2 , … ) , (203) \sum_{n=1}^{\infty}c_{2k+1,n}\mathbf{M}'_{n,m}(\alpha_0)=\frac{1}{a_E\mu_{2k+1}}c_{2k+1,m}\qquad(m=1,2,\ldots), \tag*{(203)} n = 1 ∑ ∞ c 2 k + 1 , n M n , m ′ ( α 0 ) = a E μ 2 k + 1 1 c 2 k + 1 , m ( m = 1 , 2 , … ) , (203)
and are obtained by diagonalizing the matrix
M n , m ′ ( α 0 ) = 2 π m ∫ 0 π d θ sin ( m θ ) sin ( n θ ) cosh 2 α 0 − cos 2 θ . (204) \mathbf{M}'_{n,m}(\alpha_0)=\frac{2}{\pi m}\int_{0}^{\pi}d\theta\,\sin(m\theta)\sin(n\theta)\sqrt{\cosh^{2}\alpha_0-\cos^{2}\theta}. \tag*{(204)} M n , m ′ ( α 0 ) = π m 2 ∫ 0 π d θ sin ( m θ ) sin ( n θ ) cosh 2 α 0 − cos 2 θ . (204)
Finally, we compute the normalization of the Steklov eigenfunctions:
∫ ∂ Ω d x ∣ V 2 k ∣ 2 = ∫ − π π d θ h α 0 ( θ ) ∣ V 2 k ( α 0 , θ ) ∣ 2 = a E π ∑ m = 0 ∞ ∑ n = 0 ∞ c 2 k , n c 2 k , m A n , m ( α 0 ) , (205) \int_{\partial\Omega}dx\,|V_{2k}|^{2}=\int_{-\pi}^{\pi}d\theta\,h_{\alpha_0}(\theta)|V_{2k}(\alpha_0,\theta)|^{2}=a_E\pi\sum_{m=0}^{\infty}\sum_{n=0}^{\infty}c_{2k,n}c_{2k,m}A_{n,m}(\alpha_0), \tag*{(205)} ∫ ∂ Ω d x ∣ V 2 k ∣ 2 = ∫ − π π d θ h α 0 ( θ ) ∣ V 2 k ( α 0 , θ ) ∣ 2 = a E π m = 0 ∑ ∞ n = 0 ∑ ∞ c 2 k , n c 2 k , m A n , m ( α 0 ) , (205)
where we substituted Eq. (E.5a) and used Eq. (E.8). Using Eqs. (E.9, E.10) to evaluate the sums over n n n , we get
∫ ∂ Ω d x ∣ V 2 k ∣ 2 = a E π [ c 2 k , 0 ( c 2 k , 0 A 0 , 0 ( α 0 ) + ∑ n = 1 ∞ c 2 k , n A n , 0 ( α 0 ) ) + ∑ m = 1 ∞ c 2 k , m ∑ n = 0 ∞ c 2 k , n A n , m ( α 0 ) ] = π μ 2 k ∑ m = 1 ∞ m ∣ c 2 k , m ∣ 2 . (206) \int_{\partial\Omega}dx\,|V_{2k}|^{2}=a_E\pi\left[c_{2k,0}\left(c_{2k,0}A_{0,0}(\alpha_0)+\sum_{n=1}^{\infty}c_{2k,n}A_{n,0}(\alpha_0)\right)+\sum_{m=1}^{\infty}c_{2k,m}\sum_{n=0}^{\infty}c_{2k,n}A_{n,m}(\alpha_0)\right]=\frac{\pi}{\mu_{2k}}\sum_{m=1}^{\infty}m|c_{2k,m}|^{2}. \tag*{(206)} ∫ ∂ Ω d x ∣ V 2 k ∣ 2 = a E π [ c 2 k , 0 ( c 2 k , 0 A 0 , 0 ( α 0 ) + n = 1 ∑ ∞ c 2 k , n A n , 0 ( α 0 ) ) + m = 1 ∑ ∞ c 2 k , m n = 0 ∑ ∞ c 2 k , n A n , m ( α 0 ) ] = μ 2 k π m = 1 ∑ ∞ m ∣ c 2 k , m ∣ 2 . (206)
A similar relation holds for V 2 k + 1 V_{2k+1} V 2 k + 1 :
∫ ∂ Ω d x ∣ V 2 k + 1 ∣ 2 = π μ 2 k + 1 ∑ m = 1 ∞ m ∣ c 2 k + 1 , m ∣ 2 . (207) \int_{\partial\Omega}dx\,|V_{2k+1}|^{2}=\frac{\pi}{\mu_{2k+1}}\sum_{m=1}^{\infty}m|c_{2k+1,m}|^{2}. \tag*{(207)} ∫ ∂ Ω d x ∣ V 2 k + 1 ∣ 2 = μ 2 k + 1 π m = 1 ∑ ∞ m ∣ c 2 k + 1 , m ∣ 2 . (207)
These relations can be used to ensure the L 2 ( ∂ Ω ) L^{2}(\partial\Omega) L 2 ( ∂ Ω ) normalization of the Steklov eigenfunctions by rescaling the coefficients c k , m c_{k,m} c k , m .
Setting a = 1 a=1 a = 1 and b = 0 b=0 b = 0 (i.e., α 0 = 0 \alpha_0=0 α 0 = 0 ), we deal with the exterior of the interval ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) . In this limit, the elements of the matrices A ( 0 ) \mathbf{A}(0) A ( 0 ) , M ( 0 ) \mathbf{M}(0) M ( 0 ) , and M ′ ( 0 ) \mathbf{M}'(0) M ′ ( 0 ) can be found explicitly:
A n , m ( 0 ) = 1 + ( − 1 ) m + n π ( 1 1 − ( m − n ) 2 + 1 1 − ( m + n ) 2 ) , M n , m ′ ( 0 ) = 1 + ( − 1 ) m + n π m ( 1 1 − ( m − n ) 2 − 1 1 − ( m + n ) 2 ) . (208) \begin{aligned}
\mathbf{A}_{n,m}(0)&=\frac{1+(-1)^{m+n}}{\pi}\left(\frac{1}{1-(m-n)^{2}}+\frac{1}{1-(m+n)^{2}}\right),\\
\mathbf{M}'_{n,m}(0)&=\frac{1+(-1)^{m+n}}{\pi m}\left(\frac{1}{1-(m-n)^{2}}-\frac{1}{1-(m+n)^{2}}\right). \tag*{(208)}
\end{aligned} A n , m ( 0 ) M n , m ′ ( 0 ) = π 1 + ( − 1 ) m + n ( 1 − ( m − n ) 2 1 + 1 − ( m + n ) 2 1 ) , = π m 1 + ( − 1 ) m + n ( 1 − ( m − n ) 2 1 − 1 − ( m + n ) 2 1 ) . (208)
(note that the elements with m = ± n ± 1 m = \pm n \pm1 m = ± n ± 1 are zero). Substituting A n , m ( 0 ) \mathbf{A}_{n,m}(0) A n , m ( 0 ) into Eq. (E.12), one gets the matrix M ( 0 ) \mathbf{M}(0) M ( 0 ) that determines the symmetric eigenmodes. Comparing its expression with Eq. (B.18), we realize that M ( 0 ) = G † \mathbf{M}(0) = \mathbf{G}^{\dagger} M ( 0 ) = G † , i.e., the symmetric eigenfunctions v 2 k v_{2k} v 2 k and their eigenvalues μ 2 k \mu_{2k} μ 2 k of the Steklov-Neumann problem for the exterior of the interval ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) are therefore the eigenmodes of the kernel G ^ ( x ∣ x 0 ) \hat{\mathcal{G}}(x|x_0) G ^ ( x ∣ x 0 ) from Eq. (86) .
We also note that the interior Steklov problem for the ellipse can be solved in the same way. The Steklov eigenfunctions are searched in symmetric and antisymmetric forms
V 2 k ( α , θ ) = ∑ n = 0 ∞ c 2 k , n cos ( n θ ) cosh ( n α ) cosh ( n α 0 ) . (209) V_{2k}(\alpha,\theta)=\sum_{n=0}^{\infty}c_{2k,n}\cos(n\theta)\frac{\cosh(n\alpha)}{\cosh(n\alpha_0)}. \tag*{(209)} V 2 k ( α , θ ) = n = 0 ∑ ∞ c 2 k , n cos ( n θ ) cosh ( n α 0 ) cosh ( n α ) . (209)
V 2 k + 1 ( α , θ ) = ∑ n = 1 ∞ c 2 k + 1 , n sin ( n θ ) sinh ( n α ) sinh ( n α 0 ) . (210) V_{2k+1}(\alpha,\theta)=\sum_{n=1}^{\infty}c_{2k+1,n}\sin(n\theta)\frac{\sinh(n\alpha)}{\sinh(n\alpha_0)}. \tag*{(210)} V 2 k + 1 ( α , θ ) = n = 1 ∑ ∞ c 2 k + 1 , n sin ( n θ ) sinh ( n α 0 ) sinh ( n α ) . (210)
Repeating the above computation, one gets the matrix eigenvalue problem (E.11) for symmetric eigenmodes, in which the matrix elements M n , m ( α 0 ) M_{n,m}(\alpha_0) M n , m ( α 0 ) are multiplied by ctanh ( m α 0 ) \operatorname{ctanh}(m\alpha_0) ctanh ( m α 0 ) , and the matrix eigenvalue problem (E.13) for antisymmetric eigenmodes, in which the matrix elements M n , m ′ ( α 0 ) M'_{n,m}(\alpha_0) M n , m ′ ( α 0 ) are multiplied by tanh ( m α 0 ) \tanh(m\alpha_0) tanh ( m α 0 ) .
Exterior Steklov problem for an oblate spheroid
In analogy to Appendix E, one can get a semi-analytical solution of the auxiliary Steklov-Neumann problem (3.5). In [102 ] , we studied a more general Steklov problem for the exterior of an oblate spheroid (see also [134 ] ). However, the published results are not directly applicable because the Dirichlet boundary condition at infinity was employed so that all eigenvalues were strictly positive, whereas the principal eigenfunction was not a constant. Indeed, in three dimensions, the choice of the boundary condition at infinity (Dirichlet or Neumann) affects the resulting Dirichlet-to-Neumann operators M 0 D M_0^D M 0 D and M 0 N M_0^N M 0 N and their spectral properties (see [133 ] for a rigorous definition of both operators and their comparison). In order to get the correct asymptotic behavior in the small-target limit, one needs to impose Neumann boundary condition (3.5c), which ensures that the principal eigenvalue is zero, while the associated eigenfunction is constant, as in the original Steklov-Neumann problem. In this Appendix, we sketch the main steps and resulting formulas in this setting, whereas notations, explanations, and technical details are skipped and can be found in [102 ] .
We consider the Steklov problem in the exterior of an oblate spheroid with semi-axes b > a b>a b > a , Ω = { ( x , y , z ) ∈ R 3 : ( x / b ) 2 + ( y / b ) 2 + ( z / a ) 2 > 1 } \Omega=\{(x,y,z)\in\mathbb{R}^{3}:(x/b)^2+(y/b)^2+(z/a)^2>1\} Ω = {( x , y , z ) ∈ R 3 : ( x / b ) 2 + ( y / b ) 2 + ( z / a ) 2 > 1 } , i.e., we search for the eigenpairs { μ k , V k } \{\mu_k,V_k\} { μ k , V k } satisfying
Δ V k = 0 in Ω , (211) \Delta V_k=0 \quad\text{in }\Omega, \tag*{(211)} Δ V k = 0 in Ω , (211)
∂ n V k = μ k V k on ∂ Ω , (212) \partial_n V_k=\mu_k V_k \quad\text{on }\partial\Omega, \tag*{(212)} ∂ n V k = μ k V k on ∂ Ω , (212)
∣ x ∣ 2 ∣ ∇ V k ∣ → 0 ( ∣ x ∣ → ∞ ) . (213) |x|^2|\nabla V_k|\to0 \quad(|x|\to\infty). \tag*{(213)} ∣ x ∣ 2 ∣∇ V k ∣ → 0 ( ∣ x ∣ → ∞ ) . (213)
The starting point is the expansion of an eigenfunction v m , n ( θ , φ ) v_{m,n}(\theta,\varphi) v m , n ( θ , φ ) in the oblate spheroidal coordinates onto the basis of (modified) spherical harmonics Y ˉ m ′ , n ′ ( θ , φ ) \bar{Y}_{m',n'}(\theta,\varphi) Y ˉ m ′ , n ′ ( θ , φ ) (in which P n m ( cos θ ) P_n^m(\cos\theta) P n m ( cos θ ) is replaced by P n m ( sin θ ) P_n^m(\sin\theta) P n m ( sin θ ) , see [102 ] )
v m , n ( θ , φ ) = ∑ m ′ , n ′ [ V m , n ] m ′ , n ′ Y ˉ m ′ , n ′ ( θ , φ ) . (214) v_{m,n}(\theta,\varphi)=\sum_{m',n'}[\mathbf{V}_{m,n}]_{m',n'}\bar{Y}_{m',n'}(\theta,\varphi). \tag*{(214)} v m , n ( θ , φ ) = m ′ , n ′ ∑ [ V m , n ] m ′ , n ′ Y ˉ m ′ , n ′ ( θ , φ ) . (214)
where the unknown coefficients [ V m , n ] m ′ , n ′ [\mathbf{V}_{m,n}]_{m',n'} [ V m , n ] m ′ , n ′ are determined from the boundary condition, and we keep using the double indices. Replicating the derivation of Ref. [102 ] , one obtains the system in Eq. [B13] of linear algebraic equations:
μ m , n ∑ n ′ = 0 ∞ [ V m , n ] m , n ′ G ˉ m , n ′ ; m , n ′ ′ = [ V m , n ] m , n ′ ′ c m , n ′ ′ ( n ′ ′ = 0 , 1 , 2 , … ) . (215) \mu_{m,n}\sum_{n'=0}^{\infty}[\mathbf{V}_{m,n}]_{m,n'}\bar{G}_{m,n';m,n''}=[\mathbf{V}_{m,n}]_{m,n''}c_{m,n''}\qquad(n''=0,1,2,\ldots). \tag*{(215)} μ m , n n ′ = 0 ∑ ∞ [ V m , n ] m , n ′ G ˉ m , n ′ ; m , n ′′ = [ V m , n ] m , n ′′ c m , n ′′ ( n ′′ = 0 , 1 , 2 , … ) . (215)
Here the matrix elements G ˉ m , n ; m ′ , n ′ \bar{G}_{m,n;m',n'} G ˉ m , n ; m ′ , n ′ are given in Eq. [B14], and the coefficients c m , n ′ ′ c_{m,n''} c m , n ′′ are given by Eq. [43], except for c 0 , 0 c_{0,0} c 0 , 0 , which is now equal to 0 to represent the Neumann boundary condition (F.1c) at infinity (we use square brackets when referring to equations from Ref. [102 ] ). As a consequence, the construction remains unchanged for any m > 0 m>0 m > 0 as the associated eigenfunctions were periodically oscillating and thus already orthogonal to a constant. In turn, the analysis for axially symmetric eigenfunctions with m = 0 m=0 m = 0 requires modifications. For any n ′ ′ = 1 , 2 , … n''=1,2,\ldots n ′′ = 1 , 2 , … , the system (F.3) can be rewritten as
1 c 0 , n ′ ′ ( [ V 0 , n ] 0 , 0 G ˉ 0 , 0 ; 0 , n ′ ′ + ∑ n ′ = 1 ∞ [ V 0 , n ] 0 , n ′ G ˉ 0 , n ′ ; 0 , n ′ ′ ) = [ V 0 , n ] 0 , n ′ ′ μ 0 , n . (216) \frac{1}{c_{0,n''}}\left([\mathbf{V}_{0,n}]_{0,0}\bar{G}_{0,0;0,n''}+\sum_{n'=1}^{\infty}[\mathbf{V}_{0,n}]_{0,n'}\bar{G}_{0,n';0,n''}\right)=\frac{[\mathbf{V}_{0,n}]_{0,n''}}{\mu_{0,n}}. \tag*{(216)} c 0 , n ′′ 1 ( [ V 0 , n ] 0 , 0 G ˉ 0 , 0 ; 0 , n ′′ + n ′ = 1 ∑ ∞ [ V 0 , n ] 0 , n ′ G ˉ 0 , n ′ ; 0 , n ′′ ) = μ 0 , n [ V 0 , n ] 0 , n ′′ . (216)
In turn, as c 0 , 0 = 0 c_{0,0}=0 c 0 , 0 = 0 , the equation for n ′ ′ = 0 n''=0 n ′′ = 0 yields
[ V 0 , n ] 0 , 0 = − 1 G ˉ 0 , 0 ; 0 , 0 ∑ n ′ = 1 ∞ [ V 0 , n ] 0 , n ′ G ˉ 0 , n ′ ; 0 , 0 , (217) [\mathbf{V}_{0,n}]_{0,0}=-\frac{1}{\bar{G}_{0,0;0,0}}\sum_{n'=1}^{\infty}[\mathbf{V}_{0,n}]_{0,n'}\bar{G}_{0,n';0,0}, \tag*{(217)} [ V 0 , n ] 0 , 0 = − G ˉ 0 , 0 ; 0 , 0 1 n ′ = 1 ∑ ∞ [ V 0 , n ] 0 , n ′ G ˉ 0 , n ′ ; 0 , 0 , (217)
allowing one to exclude [ V 0 , n ] 0 , 0 [\mathbf{V}_{0,n}]_{0,0} [ V 0 , n ] 0 , 0 from the above equations and to get a closed system
∑ n ′ = 1 ∞ [ V 0 , n ] 0 , n ′ G n ′ , n ′ ′ = 1 μ 0 , n [ V 0 , n ] 0 , n ′ ′ ( n ′ ′ = 1 , 2 , … ) , (218) \sum_{n'=1}^{\infty}[\mathbf{V}_{0,n}]_{0,n'}G_{n',n''}=\frac{1}{\mu_{0,n}}[\mathbf{V}_{0,n}]_{0,n''}\qquad(n''=1,2,\ldots), \tag*{(218)} n ′ = 1 ∑ ∞ [ V 0 , n ] 0 , n ′ G n ′ , n ′′ = μ 0 , n 1 [ V 0 , n ] 0 , n ′′ ( n ′′ = 1 , 2 , … ) , (218)
where
G n ′ , n ′ ′ = [ G ˉ 0 , n ′ ; 0 , n ′ ′ − G ˉ 0 , n ′ ; 0 , 0 G ˉ 0 , 0 ; 0 , n ′ ′ G ˉ 0 , 0 ; 0 , 0 ] 1 c 0 , n ′ ′ . (219) G_{n',n''}=\left[\bar{G}_{0,n';0,n''}-\frac{\bar{G}_{0,n';0,0}\bar{G}_{0,0;0,n''}}{\bar{G}_{0,0;0,0}}\right]\frac{1}{c_{0,n''}}. \tag*{(219)} G n ′ , n ′′ = [ G ˉ 0 , n ′ ; 0 , n ′′ − G ˉ 0 , 0 ; 0 , 0 G ˉ 0 , n ′ ; 0 , 0 G ˉ 0 , 0 ; 0 , n ′′ ] c 0 , n ′′ 1 . (219)
One sees that { V 0 , n } \{\mathbf{V}_{0,n}\} { V 0 , n } , enumerated by the index n = 1 , 2 , … n=1,2,\ldots n = 1 , 2 , … , are the left eigenvectors of the matrix G \mathbf{G} G that correspond to the eigenvalues 1 / μ 0 , n 1/\mu_{0,n} 1/ μ 0 , n . Truncation and numerical diagonalization of this matrix allow one to approximate μ 0 , n \mu_{0,n} μ 0 , n and [ V ] 0 , n [\mathbf{V}]_{0,n} [ V ] 0 , n , from which the eigenfunctions v 0 , n v_{0,n} v 0 , n are obtained via Eq. (F.2). Note that the coefficient [ V 0 , n ] 0 , 0 [\mathbf{V}_{0,n}]_{0,0} [ V 0 , n ] 0 , 0 is evaluated via Eq. (F.5).
The exterior of the disk corresponds to the limit a = 0 a=0 a = 0 . Since the disk has two “faces”, it is convenient to describe any point on the top face at distance r r r from the center by the variable χ = 1 − r 2 \chi=\sqrt{1-r^2} χ = 1 − r 2 , whereas any point on the bottom face by χ = − 1 − r 2 \chi=-\sqrt{1-r^2} χ = − 1 − r 2 . This convention naturally follows from the oblate spheroidal coordinates (see [102 ] for details). The eigenvalues μ 0 , n \mu_{0,n} μ 0 , n of the axially symmetric eigenfunctions of M 0 N \mathcal{M}_0^N M 0 N are then obtained by diagonalizing the matrix
G n ′ , n ′ ′ = ( n ′ + 1 / 2 ) ( n ′ ′ + 1 / 2 ) c 0 , n ′ ′ [ G n ′ , n ′ ′ 0 ( 0 ) − G n ′ , 0 0 ( 0 ) G 0 , n ′ ′ 0 ( 0 ) G 0 , 0 0 ( 0 ) ] . (220) G_{n',n''}=\frac{\sqrt{(n'+1/2)(n''+1/2)}}{c_{0,n''}}\left[G^0_{n',n''}(0)-\frac{G^0_{n',0}(0)G^0_{0,n''}(0)}{G^0_{0,0}(0)}\right]. \tag*{(220)} G n ′ , n ′′ = c 0 , n ′′ ( n ′ + 1/2 ) ( n ′′ + 1/2 ) [ G n ′ , n ′′ 0 ( 0 ) − G 0 , 0 0 ( 0 ) G n ′ , 0 0 ( 0 ) G 0 , n ′′ 0 ( 0 ) ] . (220)
with n ′ , n ′ ′ = 1 , 2 , 3 , ⋯ n^\prime,n^{\prime\prime}=1,2,3,\cdots n ′ , n ′′ = 1 , 2 , 3 , ⋯ , where c 0 , n = 2 ( Γ ( n / 2 + 1 ) Γ ( n / 2 + 1 / 2 ) ) 2 c_{0,n}=2\left(\frac{\Gamma(n/2+1)}{\Gamma(n/2+1/2)}\right)^2 c 0 , n = 2 ( Γ ( n /2 + 1/2 ) Γ ( n /2 + 1 ) ) 2 was given in Eq. [51], and
G n ′ , n ′ ′ 0 ( 0 ) = ∫ − 1 1 d x ∣ x ∣ P n ′ ( x ) P n ′ ′ ( x ) , (221) G^0_{n^\prime,n^{\prime\prime}}(0)=\int_{-1}^{1}dx\,|x|\,P_{n^\prime}(x)P_{n^{\prime\prime}}(x), \tag*{(221)} G n ′ , n ′′ 0 ( 0 ) = ∫ − 1 1 d x ∣ x ∣ P n ′ ( x ) P n ′′ ( x ) , (221)
where P n ( x ) P_n(x) P n ( x ) are the Legendre polynomials. The elements of this matrix can be rapidly obtained via recurrence relations derived in [102 ] . In turn, the left eigenvectors of this matrix, V 0 , n \mathbf{V}_{0,n} V 0 , n , determine the axially symmetric eigenfunctions v 0 , n v_{0,n} v 0 , n , and thus V 0 , n \mathbf{V}_{0,n} V 0 , n ; for instance,
v 0 , n ( χ ) = ∑ n ′ = 0 ∞ [ V 0 , n ] 0 , n ′ n ′ + 1 / 2 P n ′ ( χ ) ( − 1 < χ < 1 ) . (222) v_{0,n}(\chi)=\sum_{n^\prime=0}^{\infty}[\mathbf{V}_{0,n}]_{0,n^\prime}\sqrt{n^\prime+1/2}\,P_{n^\prime}(\chi)\qquad(-1<\chi<1). \tag*{(222)} v 0 , n ( χ ) = n ′ = 0 ∑ ∞ [ V 0 , n ] 0 , n ′ n ′ + 1/2 P n ′ ( χ ) ( − 1 < χ < 1 ) . (222)
Note that these eigenfunctions need to be normalized, see details in [102 ] .
As discussed in [102 ] , the Steklov eigenfunctions V 0 , n \mathbf{V}_{0,n} V 0 , n are symmetric with respect to the horizontal plane for even n n n , and antisymmetric for odd n n n . As a consequence, the symmetric eigenfunctions provide solutions of the mixed Steklov-Neumann problem (3.5) in the upper half-space. In other words, the eigenvalues μ 0 , 2 n \mu_{0,2n} μ 0 , 2 n and eigenfunctions v 0 , 2 n ( χ ) v_{0,2n}(\chi) v 0 , 2 n ( χ ) , restricted to positive χ \chi χ (i.e., to the “top face” of the disk), coincide with the eigenpairs { μ ^ k , v ^ k ( r ^ ) } \{\hat{\mu}_k,\hat{v}_k(\hat{r})\} { μ ^ k , v ^ k ( r ^ )} of the kernel G ^ ( 0 ) ( r ^ ∣ r ^ 0 ) \hat{G}^{(0)}(\hat{r}|\hat{r}_0) G ^ ( 0 ) ( r ^ ∣ r ^ 0 ) from Eq. (102) , by setting χ = 1 − r ^ 2 \chi=\sqrt{1-\hat{r}^2} χ = 1 − r ^ 2 . This is confirmed by a numerical diagonalization of the matrix G \mathbf{G} G from Eq. (F.8).
First-order corrections in the small-target limit
In this Appendix, we discuss the small-target limit of the corrections A Γ A_\Gamma A Γ and w 0 ( Γ ) w_0^{(\Gamma)} w 0 ( Γ ) and the coefficients b k ( Γ ) b_k^{(\Gamma)} b k ( Γ ) for a bounded domain with a smooth boundary.
Two dimensions
For a connected subset Γ \Gamma Γ of perimeter ∣ Γ ∣ = 2 ϵ |\Gamma|=2\epsilon ∣Γ∣ = 2 ϵ , centered at a boundary point x Γ ∈ ∂ Ω x_\Gamma\in\partial\Omega x Γ ∈ ∂ Ω , substitution of the expansion (3.9) of the pseudo-Green’s function G 0 ( x ∣ x 0 ) G_0(x|x_0) G 0 ( x ∣ x 0 ) into Eq. (39) yields in the limit ϵ → 0 \epsilon\to0 ϵ → 0 :
A Γ ≈ 1 ( 2 ϵ ) 2 ∫ − ϵ ϵ d s ∫ − ϵ ϵ d s 0 [ − 1 π ln ∣ s − s 0 ∣ + R 0 ( x Γ ) ] = − ln ∣ Γ ∣ π + 3 2 π + R 0 ( x Γ ) + o ( 1 ) , (223) A_\Gamma\approx\frac{1}{(2\epsilon)^2}\int_{-\epsilon}^{\epsilon}ds\int_{-\epsilon}^{\epsilon}ds_0\left[-\frac{1}{\pi}\ln|s-s_0|+R_0(x_\Gamma)\right]=-\frac{\ln|\Gamma|}{\pi}+\frac{3}{2\pi}+R_0(x_\Gamma)+o(1), \tag*{(223)} A Γ ≈ ( 2 ϵ ) 2 1 ∫ − ϵ ϵ d s ∫ − ϵ ϵ d s 0 [ − π 1 ln ∣ s − s 0 ∣ + R 0 ( x Γ ) ] = − π ln ∣Γ∣ + 2 π 3 + R 0 ( x Γ ) + o ( 1 ) , (223)
where s s s and s 0 s_0 s 0 are the curvilinear coordinates of the points x x x and x 0 x_0 x 0 on the subset Γ \Gamma Γ . As the boundary ∂ Ω \partial\Omega ∂ Ω is smooth, the small subset Γ \Gamma Γ can be approximated as a flat interval ( − ϵ , ϵ ) (-\epsilon,\epsilon) ( − ϵ , ϵ ) (the curvature effect emerging in the next-order term in ϵ \epsilon ϵ ). Similarly, we compute
w 0 ( Γ ) ( x 0 ) ≈ − A Γ + 1 2 ϵ ∫ − ϵ ϵ d s [ − 1 π ln ∣ x − x 0 ∣ + R 0 ( x 0 ) ] = − A Γ + R 0 ( x Γ ) − ( 1 + s 0 / ϵ ) ln ( ϵ + s 0 ) + ( 1 − s 0 / ϵ ) ln ( ϵ − s 0 ) − 2 2 π , \begin{aligned}
w_0^{(\Gamma)}(x_0)\approx-A_\Gamma+\frac{1}{2\epsilon}\int_{-\epsilon}^{\epsilon}ds\left[-\frac{1}{\pi}\ln|x-x_0|+R_0(x_0)\right] \\
&=-A_\Gamma+R_0(x_\Gamma)-\frac{(1+s_0/\epsilon)\ln(\epsilon+s_0)+(1-s_0/\epsilon)\ln(\epsilon-s_0)-2}{2\pi},
\end{aligned} w 0 ( Γ ) ( x 0 ) ≈ − A Γ + 2 ϵ 1 ∫ − ϵ ϵ d s [ − π 1 ln ∣ x − x 0 ∣ + R 0 ( x 0 ) ] = − A Γ + R 0 ( x Γ ) − 2 π ( 1 + s 0 / ϵ ) ln ( ϵ + s 0 ) + ( 1 − s 0 / ϵ ) ln ( ϵ − s 0 ) − 2 ,
which yields Eq. (82) after substitution of A Γ A_\Gamma A Γ from Eq. (G.1). As a consequence, the coefficients b k ( Γ ) b_k^{(\Gamma)} b k ( Γ ) from Eq. (51) admit the following scaling in the leading order:
b k ( Γ ) ≈ − μ ^ k ϵ ∫ − ϵ ϵ d s w ^ 0 ( s / ϵ ) 1 ϵ v ^ k ( s / ϵ ) . b_k^{(\Gamma)}\approx-\frac{\hat{\mu}_k}{\epsilon}\int_{-\epsilon}^{\epsilon}ds\,\hat{w}_0(s/\epsilon)\frac{1}{\sqrt{\epsilon}}\hat{v}_k(s/\epsilon). b k ( Γ ) ≈ − ϵ μ ^ k ∫ − ϵ ϵ d s w ^ 0 ( s / ϵ ) ϵ 1 v ^ k ( s / ϵ ) .
yielding
b k ( Γ ) ≈ ϵ − 1 / 2 b ^ k , b ^ k = − μ ^ k ∫ − 1 1 d y w ^ 0 ( y ) v ^ k ( y ) . (224) b_k^{(\Gamma)} \approx\epsilon^{-1/2}\hat{b}_k,\qquad\hat{b}_k=-\hat{\mu}_k\int_{-1}^{1} \mathrm{d}y\,\hat{w}_0(y)\hat{v}_k(y). \tag*{(224)} b k ( Γ ) ≈ ϵ − 1/2 b ^ k , b ^ k = − μ ^ k ∫ − 1 1 d y w ^ 0 ( y ) v ^ k ( y ) . (224)
In order to avoid the numerical integration in Eq. (G.2), we derive another representation of the coefficients b ^ k \hat{b}_k b ^ k . We start by an intuitive “hand-waving” argument that relies on Eq. (46) , which is valid for a bounded domain Ω \Omega Ω . Let Ω L \Omega_L Ω L be the half-disk of radius L L L in the upper half-plane, with Γ = ( − 1 , 1 ) \Gamma=(-1,1) Γ = ( − 1 , 1 ) lying on the horizontal diameter. When L → ∞ L\to\infty L → ∞ , the domain Ω L \Omega_L Ω L approaches H 2 \mathbb{H}_2 H 2 , so that an eigenfunction V k ( 0 , Γ ) V_k^{(0,\Gamma)} V k ( 0 , Γ ) approaches V ^ k \hat{V}_k V ^ k , which remains bounded. As a consequence, Eq. (46) yields in the limit L → ∞ L\to\infty L → ∞ :
b ^ k = V ^ k ( ∞ ) . (225) \hat{b}_k=\hat{V}_k(\infty). \tag*{(225)} b ^ k = V ^ k ( ∞ ) . (225)
A more formal derivation of this relation employs the pseudo-Green’s function G 0 ( x ∣ x 0 ) G_0(\boldsymbol{x}|\boldsymbol{x}_0) G 0 ( x ∣ x 0 ) for the upper half-plane,
G 0 ( x ∣ x 0 ) = − 1 2 π ( ln ∣ x − x 0 ∣ + ln ∣ x − x 0 ′ ∣ ) , (226) G_0(\boldsymbol{x}|\boldsymbol{x}_0)=-\frac{1}{2\pi}\left(\ln|\boldsymbol{x}-\boldsymbol{x}_0|+\ln|\boldsymbol{x}-\boldsymbol{x}_0'|\right), \tag*{(226)} G 0 ( x ∣ x 0 ) = − 2 π 1 ( ln ∣ x − x 0 ∣ + ln ∣ x − x 0 ′ ∣ ) , (226)
where x 0 ′ = ( x 0 , − y 0 ) \boldsymbol{x}_0'=(x_0,-y_0) x 0 ′ = ( x 0 , − y 0 ) is the mirror reflection of the point x 0 = ( x 0 , y 0 ) \boldsymbol{x}_0=(x_0,y_0) x 0 = ( x 0 , y 0 ) . This function satisfies
− Δ G 0 ( x ∣ x 0 ) = δ ( x − x 0 ) ( x ∈ H 2 ) , (227) -\Delta G_0(\boldsymbol{x}|\boldsymbol{x}_0)=\delta(\boldsymbol{x}-\boldsymbol{x}_0)\qquad(\boldsymbol{x}\in\mathbb{H}_2), \tag*{(227)} − Δ G 0 ( x ∣ x 0 ) = δ ( x − x 0 ) ( x ∈ H 2 ) , (227)
∂ n G 0 ( x ∣ x 0 ) = 0 ( x ∈ ∂ H 2 ) , (228) \partial_nG_0(\boldsymbol{x}|\boldsymbol{x}_0)=0\qquad(\boldsymbol{x}\in\partial\mathbb{H}_2), \tag*{(228)} ∂ n G 0 ( x ∣ x 0 ) = 0 ( x ∈ ∂ H 2 ) , (228)
G 0 ( x ∣ x 0 ) ∼ − 1 π ln ∣ x ∣ + o ( 1 ) ( ∣ x ∣ → ∞ ) . (229) G_0(\boldsymbol{x}|\boldsymbol{x}_0)\sim-\frac{1}{\pi}\ln|\boldsymbol{x}|+o(1)\qquad(|\boldsymbol{x}|\to\infty). \tag*{(229)} G 0 ( x ∣ x 0 ) ∼ − π 1 ln ∣ x ∣ + o ( 1 ) ( ∣ x ∣ → ∞ ) . (229)
Multiplying Eq. (3.3a) by G 0 ( x ∣ x 0 ) G_0(\boldsymbol{x}|\boldsymbol{x}_0) G 0 ( x ∣ x 0 ) , multiplying Eq. (G.5a) by V ^ k ( x ) \hat{V}_k(\boldsymbol{x}) V ^ k ( x ) , subtracting them, integrating over x = ( x , y ) ∈ H 2 \boldsymbol{x}=(x,y)\in\mathbb{H}_2 x = ( x , y ) ∈ H 2 and using the Green’s formula, we get for any x 0 = ( x 0 , y 0 ) ∈ H 2 \boldsymbol{x}_0=(x_0,y_0)\in\mathbb{H}_2 x 0 = ( x 0 , y 0 ) ∈ H 2 :
V ^ k ( x 0 , y 0 ) = V ^ k ( ∞ ) + μ ^ k ∫ − 1 1 d x G 0 ( x , 0 ∣ x 0 , y 0 ) v ^ k ( x ) . (230) \hat{V}_k(x_0,y_0)=\hat{V}_k(\infty)+\hat{\mu}_k\int_{-1}^{1}\mathrm{d}x\,G_0(x,0|x_0,y_0)\hat{v}_k(x). \tag*{(230)} V ^ k ( x 0 , y 0 ) = V ^ k ( ∞ ) + μ ^ k ∫ − 1 1 d x G 0 ( x , 0∣ x 0 , y 0 ) v ^ k ( x ) . (230)
Substituting
1 2 ∫ − 1 1 d x G 0 ( x , 0 ∣ x 0 , 0 ) = w ^ 0 ( x 0 ) + a 0 ( − 1 < x 0 < 1 ) (231) \frac{1}{2}\int_{-1}^{1}\mathrm{d}x\,G_0(x,0|x_0,0)=\hat{w}_0(x_0)+a_0\qquad(-1<x_0<1) \tag*{(231)} 2 1 ∫ − 1 1 d x G 0 ( x , 0∣ x 0 , 0 ) = w ^ 0 ( x 0 ) + a 0 ( − 1 < x 0 < 1 ) (231)
(with a 0 = ( 3 − 2 ln 2 ) / ( 2 π ) a_0=(3-2\ln2)/(2\pi) a 0 = ( 3 − 2 ln 2 ) / ( 2 π ) ) into Eq. (G.2) yields
b ^ k = − μ ^ k ∫ − 1 1 d x v ^ k ( x ) [ − a 0 + 1 2 ∫ − 1 1 d x 0 G 0 ( x , 0 ∣ x 0 , 0 ) ] = − μ ^ k 2 ∫ − 1 1 d x 0 ∫ − 1 1 d x v ^ k ( x ) G 0 ( x , 0 ∣ x 0 , 0 ) = V ^ k ( ∞ ) , \begin{aligned}
\hat{b}_k=-\hat{\mu}_k\int_{-1}^{1}\mathrm{d}x\,\hat{v}_k(x)\left[-a_0+\frac{1}{2}\int_{-1}^{1}\mathrm{d}x_0\,G_0(x,0|x_0,0)\right] \\
&=-\frac{\hat{\mu}_k}{2}\int_{-1}^{1}\mathrm{d}x_0\int_{-1}^{1}\mathrm{d}x\,\hat{v}_k(x)G_0(x,0|x_0,0)=\hat{V}_k(\infty),
\end{aligned} b ^ k = − μ ^ k ∫ − 1 1 d x v ^ k ( x ) [ − a 0 + 2 1 ∫ − 1 1 d x 0 G 0 ( x , 0∣ x 0 , 0 ) ] = − 2 μ ^ k ∫ − 1 1 d x 0 ∫ − 1 1 d x v ^ k ( x ) G 0 ( x , 0∣ x 0 , 0 ) = V ^ k ( ∞ ) ,
where the underbraced expression is equal to ( v ^ k ( x 0 ) − V ^ k ( ∞ ) ) / μ ^ k (\hat{v}_k(x_0)-\hat{V}_k(\infty))/\hat{\mu}_k ( v ^ k ( x 0 ) − V ^ k ( ∞ )) / μ ^ k .
where we used the orthogonality of v ^ k \hat{v}_k v ^ k to a constant. In order to compute numerically μ ^ k \hat{\mu}_k μ ^ k and V ^ k ( ∞ ) \hat{V}_k(\infty) V ^ k ( ∞ ) , it is sufficient to diagonalize a truncated matrix M ( 0 ) \mathbf{M}(0) M ( 0 ) with explicitly known coefficients, as described in Appendix E. The first ten values of V ^ k ( ∞ ) \hat{V}_k(\infty) V ^ k ( ∞ ) are reported in Table 1 .
In addition, we compute two sums used in Sec. 5. First, we establish the following identity
∑ k = 1 ∞ [ b k ( Γ ) ] 2 [ μ k ( 0 , Γ ) ] 2 = ∫ Γ d x 1 w 0 ( Γ ) ( x 1 ) ∫ Γ d x 2 w 0 ( Γ ) ( x 2 ) ∑ k = 1 ∞ v k ( 0 , Γ ) ( x 1 ) v k ( 0 , Γ ) ( x 2 ) = ∫ Γ d x [ w 0 ( Γ ) ( x ) ] 2 = ∥ w 0 ( Γ ) ∥ L 2 ( Γ ) 2 , ∑ k = 1 ∞ v k ( 0 , Γ ) ( x 1 ) v k ( 0 , Γ ) ( x 2 ) = δ ( x 1 − x 2 ) − 1 / ∣ Γ ∣ (232) \sum_{k=1}^{\infty}\frac{[b_k^{(\Gamma)}]^2}{[\mu_k^{(0,\Gamma)}]^2}
=
\int_{\Gamma} dx_1\,w_0^{(\Gamma)}(x_1)
\int_{\Gamma} dx_2\,w_0^{(\Gamma)}(x_2)
\sum_{k=1}^{\infty}v_k^{(0,\Gamma)}(x_1)v_k^{(0,\Gamma)}(x_2)
=
\int_{\Gamma} dx\,[w_0^{(\Gamma)}(x)]^2
=
\lVert w_0^{(\Gamma)}\rVert_{L^2(\Gamma)}^2,
\qquad
\sum_{k=1}^{\infty}v_k^{(0,\Gamma)}(x_1)v_k^{(0,\Gamma)}(x_2)
=\delta(x_1-x_2)-1/|\Gamma| \tag*{(232)} k = 1 ∑ ∞ [ μ k ( 0 , Γ ) ] 2 [ b k ( Γ ) ] 2 = ∫ Γ d x 1 w 0 ( Γ ) ( x 1 ) ∫ Γ d x 2 w 0 ( Γ ) ( x 2 ) k = 1 ∑ ∞ v k ( 0 , Γ ) ( x 1 ) v k ( 0 , Γ ) ( x 2 ) = ∫ Γ d x [ w 0 ( Γ ) ( x ) ] 2 = ∥ w 0 ( Γ ) ∥ L 2 ( Γ ) 2 , k = 1 ∑ ∞ v k ( 0 , Γ ) ( x 1 ) v k ( 0 , Γ ) ( x 2 ) = δ ( x 1 − x 2 ) − 1/∣Γ∣ (232)
where we used Eq. (51) , the orthogonality (2.27) of w 0 ( Γ ) w_0^{(\Gamma)} w 0 ( Γ ) to a constant, and the completeness relation (A.3) at p = 0 p=0 p = 0 . In the limit ϵ → 0 \epsilon\to0 ϵ → 0 , Eq. (G.8) yields
C 2 = ∑ k = 1 ∞ [ V ^ k ( ∞ ) ] 2 μ ^ k 2 = ∫ − 1 1 d x [ w ^ 0 ( x ) ] 2 = 21 − 2 π 2 18 π 2 ≈ 0.007 , (233) C_2=\sum_{k=1}^{\infty}\frac{[\hat{V}_k(\infty)]^2}{\hat{\mu}_k^2}
=\int_{-1}^{1}dx\,[\hat{w}_0(x)]^2
=\frac{21-2\pi^2}{18\pi^2}\approx0.007, \tag*{(233)} C 2 = k = 1 ∑ ∞ μ ^ k 2 [ V ^ k ( ∞ ) ] 2 = ∫ − 1 1 d x [ w ^ 0 ( x ) ] 2 = 18 π 2 21 − 2 π 2 ≈ 0.007 , (233)
where we used Eq. (83) for w ^ 0 ( x ) \hat{w}_0(x) w ^ 0 ( x ) . Second, the divergence theorem implies
V ^ k ( ∞ ) = ∫ − 1 1 d x v ^ k ( x ) h ( x ∣ ∞ ) where h ( x ∣ ∞ ) = 1 π 1 − x 2 (234) \hat{V}_k(\infty)=\int_{-1}^{1}dx\,\hat{v}_k(x)h(x|\infty)
\qquad\text{where}\qquad
h(x|\infty)=\frac{1}{\pi\sqrt{1-x^2}} \tag*{(234)} V ^ k ( ∞ ) = ∫ − 1 1 d x v ^ k ( x ) h ( x ∣∞ ) where h ( x ∣∞ ) = π 1 − x 2 1 (234)
is the harmonic measure density on the interval ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) , seen from infinity. Combining this representation with Eqs. (G.2, G.3), we get
∑ k = 1 ∞ [ V ^ k ( ∞ ) ] 2 μ ^ k = − ∑ k = 1 ∞ ∫ − 1 1 d x v ^ k ( x ) h ( x ∣ ∞ ) ∫ − 1 1 d x ′ w ^ 0 ( x ′ ) v ^ k ( x ′ ) = − ∫ − 1 1 d x h ( x ∣ ∞ ) ∫ − 1 1 d x ′ w ^ 0 ( x ′ ) ∑ k = 1 ∞ v ^ k ( x ) v ^ k ( x ′ ) \sum_{k=1}^{\infty}\frac{[\hat{V}_k(\infty)]^2}{\hat{\mu}_k}
=
-\sum_{k=1}^{\infty}\int_{-1}^{1}dx\,\hat{v}_k(x)h(x|\infty)
\int_{-1}^{1}dx'\,\hat{w}_0(x')\hat{v}_k(x')
=
-\int_{-1}^{1}dx\,h(x|\infty)
\int_{-1}^{1}dx'\,\hat{w}_0(x')
\sum_{k=1}^{\infty}\hat{v}_k(x)\hat{v}_k(x') k = 1 ∑ ∞ μ ^ k [ V ^ k ( ∞ ) ] 2 = − k = 1 ∑ ∞ ∫ − 1 1 d x v ^ k ( x ) h ( x ∣∞ ) ∫ − 1 1 d x ′ w ^ 0 ( x ′ ) v ^ k ( x ′ ) = − ∫ − 1 1 d x h ( x ∣∞ ) ∫ − 1 1 d x ′ w ^ 0 ( x ′ ) k = 1 ∑ ∞ v ^ k ( x ) v ^ k ( x ′ )
∑ k = 1 ∞ v ^ k ( x ) v ^ k ( x ′ ) = δ ( x − x ′ ) − 1 2 . \sum_{k=1}^{\infty}\hat{v}_k(x)\hat{v}_k(x')
=\delta(x-x')-\frac{1}{2}. k = 1 ∑ ∞ v ^ k ( x ) v ^ k ( x ′ ) = δ ( x − x ′ ) − 2 1 .
As a consequence, we obtain
C 1 = ∑ k = 1 ∞ [ V ^ k ( ∞ ) ] 2 μ ^ k = − ∫ − 1 1 d x w ^ 0 ( x ) π 1 − x 2 = 3 − 4 ln 2 2 π ≈ 0.0362. (235) C_1=\sum_{k=1}^{\infty}\frac{[\hat{V}_k(\infty)]^2}{\hat{\mu}_k}
=-\int_{-1}^{1}dx\,\frac{\hat{w}_0(x)}{\pi\sqrt{1-x^2}}
=\frac{3-4\ln2}{2\pi}\approx0.0362. \tag*{(235)} C 1 = k = 1 ∑ ∞ μ ^ k [ V ^ k ( ∞ ) ] 2 = − ∫ − 1 1 d x π 1 − x 2 w ^ 0 ( x ) = 2 π 3 − 4 ln 2 ≈ 0.0362. (235)
Three dimensions
We consider the subset Γ \Gamma Γ to be a small disk of radius ϵ \epsilon ϵ , centered at a point x Γ ∈ ∂ Ω x_\Gamma\in\partial\Omega x Γ ∈ ∂ Ω . Substitution of the expansion (3.20) of the pseudo-Green’s function G 0 ( x ∣ x 0 ) G_0(x|x_0) G 0 ( x ∣ x 0 ) into Eq. (39) yields as ϵ → 0 \epsilon\to0 ϵ → 0 :
A Γ ≈ R 0 ( x Γ ) + 1 ∣ Γ ∣ 2 ∫ Γ d x 0 ∫ Γ d x 2 π ∣ x − x 0 ∣ ⏟ = A Γ ( 1 ) − H ( x Γ ) 4 π ∣ Γ ∣ 2 ∫ Γ d x 0 ∫ Γ d x ln ∣ x − x 0 ∣ ⏟ = A Γ ( 2 ) + o ( 1 ) , (236) A_\Gamma\approx R_0(x_\Gamma)
+\underbrace{\frac{1}{|\Gamma|^2}\int_{\Gamma}dx_0\int_{\Gamma}\frac{dx}{2\pi|x-x_0|}}_{=A_\Gamma^{(1)}}
-\underbrace{\frac{H(x_\Gamma)}{4\pi|\Gamma|^2}\int_{\Gamma}dx_0\int_{\Gamma}dx\,\ln|x-x_0|}_{=A_\Gamma^{(2)}}
+o(1), \tag*{(236)} A Γ ≈ R 0 ( x Γ ) + = A Γ ( 1 ) ∣Γ ∣ 2 1 ∫ Γ d x 0 ∫ Γ 2 π ∣ x − x 0 ∣ d x − = A Γ ( 2 ) 4 π ∣Γ ∣ 2 H ( x Γ ) ∫ Γ d x 0 ∫ Γ d x ln ∣ x − x 0 ∣ + o ( 1 ) , (236)
where we used the smoothness of R 0 ( x 0 ) R_0(\mathbf{x}_0) R 0 ( x 0 ) and H ( x 0 ) H(\mathbf{x}_0) H ( x 0 ) to replace them by R 0 ( x Γ ) R_0(\mathbf{x}_{\Gamma}) R 0 ( x Γ ) and H ( x Γ ) H(\mathbf{x}_{\Gamma}) H ( x Γ ) , respectively.
To compute the first integral, denoted as A Γ ( 1 ) \mathcal{A}_{\Gamma}^{(1)} A Γ ( 1 ) , we use the spherical coordinates ( r , π / 2 , φ ) (r,\pi/2,\varphi) ( r , π /2 , φ ) , in which θ = π / 2 \theta=\pi/2 θ = π /2 corresponds to the disk Γ \Gamma Γ on the horizontal plane, and the Legendre expansion:
1 ∣ x − x 0 ∣ = ∑ n = 0 ∞ P n ( cos ( φ − φ 0 ) ) r < n r > n + 1 , (237) \frac{1}{\lvert\mathbf{x}-\mathbf{x}_0\rvert}=\sum_{n=0}^{\infty}P_n(\cos(\varphi-\varphi_0))\frac{r_<^n}{r_>^{n+1}}, \tag*{(237)} ∣ x − x 0 ∣ 1 = n = 0 ∑ ∞ P n ( cos ( φ − φ 0 )) r > n + 1 r < n , (237)
where r < = min { r , r 0 } r_< = \min\{r,r_0\} r < = min { r , r 0 } and r > = max { r , r 0 } r_> = \max\{r,r_0\} r > = max { r , r 0 } . As a consequence, we get
A Γ ( 1 ) = 1 2 π ( π ϵ 2 ) 2 ∫ 0 ϵ d r 0 r 0 ∫ 0 2 π d φ 0 ∫ 0 ϵ d r r ∫ 0 2 π d φ ∑ n = 0 ∞ P n ( cos ( φ − φ 0 ) ) r < n r > n + 1 = 2 3 π ϵ ∑ n = 0 ∞ P 2 n 2 ( 0 ) n + 1 , \begin{aligned}
\mathcal{A}_{\Gamma}^{(1)}=\frac{1}{2\pi(\pi\epsilon^2)^2}\int_0^\epsilon dr_0\,r_0\int_0^{2\pi}d\varphi_0\int_0^\epsilon dr\,r\int_0^{2\pi}d\varphi\sum_{n=0}^{\infty}P_n(\cos(\varphi-\varphi_0))\frac{r_<^n}{r_>^{n+1}} \\
&=\frac{2}{3\pi\epsilon}\sum_{n=0}^{\infty}\frac{P_{2n}^2(0)}{n+1},
\end{aligned} A Γ ( 1 ) = 2 π ( π ϵ 2 ) 2 1 ∫ 0 ϵ d r 0 r 0 ∫ 0 2 π d φ 0 ∫ 0 ϵ d r r ∫ 0 2 π d φ n = 0 ∑ ∞ P n ( cos ( φ − φ 0 )) r > n + 1 r < n = 3 π ϵ 2 n = 0 ∑ ∞ n + 1 P 2 n 2 ( 0 ) ,
where we used the addition theorem to evaluate
∫ 0 2 π d φ P 2 n ( cos φ ) = 2 π [ P 2 n ( 0 ) ] 2 = 2 π ( ( 2 n − 1 ) ! ! ( 2 n ) ! ! ) 2 . \int_0^{2\pi}d\varphi\,P_{2n}(\cos\varphi)=2\pi[P_{2n}(0)]^2=2\pi\left(\frac{(2n-1)!!}{(2n)!!}\right)^2. ∫ 0 2 π d φ P 2 n ( cos φ ) = 2 π [ P 2 n ( 0 ) ] 2 = 2 π ( ( 2 n )!! ( 2 n − 1 )!! ) 2 .
whereas odd-order contributions vanished. Finally, we employ the expansion of the complete elliptic integral of the first kind [124 ] ,
K ( z ) = π 2 ∑ n = 0 ∞ [ P 2 n ( 0 ) ] 2 z 2 n ( ∣ z ∣ < 1 ) , (238) K(z)=\frac{\pi}{2}\sum_{n=0}^{\infty}[P_{2n}(0)]^2z^{2n}\qquad(\lvert z\rvert<1), \tag*{(238)} K ( z ) = 2 π n = 0 ∑ ∞ [ P 2 n ( 0 ) ] 2 z 2 n (∣ z ∣ < 1 ) , (238)
to get
A Γ ( 1 ) = 2 3 π ϵ 2 π ∫ 0 1 d z K ( z ) = 8 3 π 2 ϵ . (239) \mathcal{A}_{\Gamma}^{(1)}=\frac{2}{3\pi\epsilon}\frac{2}{\pi}\int_0^1 dz\,K(\sqrt{z})=\frac{8}{3\pi^2\epsilon}. \tag*{(239)} A Γ ( 1 ) = 3 π ϵ 2 π 2 ∫ 0 1 d z K ( z ) = 3 π 2 ϵ 8 . (239)
In the same way, we also have
∫ 0 2 π d φ ∣ x − x 0 ∣ = ∑ n = 0 ∞ r < n r > n + 1 ∫ 0 2 π d φ P n ( cos ( φ − φ 0 ) ) = 2 π r > ∑ n = 0 ∞ [ P 2 n ( 0 ) ] 2 ( r < / r > ) 2 n = 4 r > K ( r < / r > ) . (240) \begin{aligned}
\int_0^{2\pi}\frac{d\varphi}{\lvert\mathbf{x}-\mathbf{x}_0\rvert}=\sum_{n=0}^{\infty}\frac{r_<^n}{r_>^{n+1}}\int_0^{2\pi}d\varphi P_n(\cos(\varphi-\varphi_0))=\frac{2\pi}{r_>}\sum_{n=0}^{\infty}[P_{2n}(0)]^2(r_</r_>)^{2n} \\
&=\frac{4}{r_>}K(r_</r_>). \tag*{(240)}
\end{aligned} ∫ 0 2 π ∣ x − x 0 ∣ d φ = n = 0 ∑ ∞ r > n + 1 r < n ∫ 0 2 π d φ P n ( cos ( φ − φ 0 )) = r > 2 π n = 0 ∑ ∞ [ P 2 n ( 0 ) ] 2 ( r < / r > ) 2 n = r > 4 K ( r < / r > ) . (240)
For the second integral in Eq. (G.12), we write ∣ x − x 0 ∣ = r 2 + r 0 2 − 2 r r 0 cos ( φ − φ 0 ) \lvert\mathbf{x}-\mathbf{x}_0\rvert=\sqrt{r^2+r_0^2-2rr_0\cos(\varphi-\varphi_0)} ∣ x − x 0 ∣ = r 2 + r 0 2 − 2 r r 0 cos ( φ − φ 0 ) and use the relation
∫ 0 2 π d φ ln ( A − B cos φ ) = 2 π ln ( A + A 2 − B 2 2 ) (241) \int_0^{2\pi}d\varphi\ln(A-B\cos\varphi)=2\pi\ln\left(\frac{A+\sqrt{A^2-B^2}}{2}\right) \tag*{(241)} ∫ 0 2 π d φ ln ( A − B cos φ ) = 2 π ln ( 2 A + A 2 − B 2 ) (241)
to evaluate first the integral over φ \varphi φ :
∫ 0 2 π d φ ln ∣ x − x 0 ∣ = π ln ( r 2 + r 0 2 + ∣ r 2 − r 0 2 ∣ 2 ) . (242) \int_{0}^{2\pi} d\varphi\ln|x-x_0|=\pi\ln\left(\frac{r^2+r_0^2+|r^2-r_0^2|}{2}\right). \tag*{(242)} ∫ 0 2 π d φ ln ∣ x − x 0 ∣ = π ln ( 2 r 2 + r 0 2 + ∣ r 2 − r 0 2 ∣ ) . (242)
As a consequence, the integral of this expression over r r r yields
1 ∣ Γ ∣ ∫ Γ d x ln ∣ x − x 0 ∣ = 1 ϵ 2 ∫ 0 ϵ d r r ln ( r 2 + r 0 2 + ∣ r 2 − r 0 2 ∣ 2 ) = ln ϵ − 1 − ( r 0 / ϵ ) 2 2 , (243) \frac{1}{|\Gamma|}\int_{\Gamma} dx\ln|x-x_0|=\frac{1}{\epsilon^2}\int_{0}^{\epsilon}dr\,r\ln\left(\frac{r^2+r_0^2+|r^2-r_0^2|}{2}\right)=\ln\epsilon-\frac{1-(r_0/\epsilon)^2}{2}, \tag*{(243)} ∣Γ∣ 1 ∫ Γ d x ln ∣ x − x 0 ∣ = ϵ 2 1 ∫ 0 ϵ d r r ln ( 2 r 2 + r 0 2 + ∣ r 2 − r 0 2 ∣ ) = ln ϵ − 2 1 − ( r 0 / ϵ ) 2 , (243)
from which
A Γ ( 2 ) = H ( x Γ ) 4 π ( ln ϵ − 1 4 ) . (244) \mathcal{A}_{\Gamma}^{(2)}=\frac{H(x_{\Gamma})}{4\pi}\left(\ln\epsilon-\frac{1}{4}\right). \tag*{(244)} A Γ ( 2 ) = 4 π H ( x Γ ) ( ln ϵ − 4 1 ) . (244)
Combining these results, we obtain the asymptotic relation (92) for A Γ \mathcal{A}_{\Gamma} A Γ . For instance, for a sphere of radius R R R , one can substitute H ( x Γ ) H(x_{\Gamma}) H ( x Γ ) and R 0 ( x Γ ) R_0(x_{\Gamma}) R 0 ( x Γ ) from Eq. (C.7) to get
A Γ ≈ 1 4 π R ( 32 3 π ϵ + ln ( 1 / ϵ ) + ln 2 − 31 20 ) + o ( 1 ) (245) \mathcal{A}_{\Gamma}\approx\frac{1}{4\pi R}\left(\frac{32}{3\pi\epsilon}+\ln(1/\epsilon)+\ln2-\frac{31}{20}\right)+o(1) \tag*{(245)} A Γ ≈ 4 π R 1 ( 3 π ϵ 32 + ln ( 1/ ϵ ) + ln 2 − 20 31 ) + o ( 1 ) (245)
(with ϵ = ε / R \epsilon=\varepsilon/R ϵ = ε / R ), which agrees with the asymptotic analysis of the series (C.9b) reported in [83 ] .
Similarly, we can compute the leading-order term in the correction w 0 ( Γ ) ( x 0 ) w_0^{(\Gamma)}(x_0) w 0 ( Γ ) ( x 0 ) by substituting Eq. (91) to Eq. (40) :
w 0 ( Γ ) ( x 0 ) ≈ 1 2 π 2 ϵ 2 ∫ 0 ϵ d r r ∫ 0 2 π d φ ∑ n = 0 ∞ P n ( cos ( φ − φ 0 ) ) r < n r > n + 1 − A Γ + R 0 ( x 0 ) − H ( x 0 ) 4 π ∣ Γ ∣ ∫ Γ d x ln ∣ x − x 0 ∣ = − A Γ + r 0 π ϵ 2 ∑ n = 0 ∞ [ P 2 n ( 0 ) ] 2 ( 1 2 n + 2 + 1 − ( r 0 / ϵ ) 2 n − 1 2 n − 1 ) + R 0 ( x 0 ) − H ( x 0 ) 4 π ( ln ϵ − 1 − ( r 0 / ϵ ) 2 2 ) . \begin{aligned}
w_0^{(\Gamma)}(x_0)&\approx\frac{1}{2\pi^2\epsilon^2}\int_{0}^{\epsilon}dr\,r\int_{0}^{2\pi}d\varphi\sum_{n=0}^{\infty}P_n(\cos(\varphi-\varphi_0))\frac{r_<^n}{r_>^{n+1}}\\
&\quad-\mathcal{A}_{\Gamma}+R_0(x_0)-\frac{H(x_0)}{4\pi|\Gamma|}\int_{\Gamma}dx\ln|x-x_0|\\
&=-\mathcal{A}_{\Gamma}+\frac{r_0}{\pi\epsilon^2}\sum_{n=0}^{\infty}[P_{2n}(0)]^2\left(\frac{1}{2n+2}+\frac{1-(r_0/\epsilon)^{2n-1}}{2n-1}\right)\\
&\quad+R_0(x_0)-\frac{H(x_0)}{4\pi}\left(\ln\epsilon-\frac{1-(r_0/\epsilon)^2}{2}\right).
\end{aligned} w 0 ( Γ ) ( x 0 ) ≈ 2 π 2 ϵ 2 1 ∫ 0 ϵ d r r ∫ 0 2 π d φ n = 0 ∑ ∞ P n ( cos ( φ − φ 0 )) r > n + 1 r < n − A Γ + R 0 ( x 0 ) − 4 π ∣Γ∣ H ( x 0 ) ∫ Γ d x ln ∣ x − x 0 ∣ = − A Γ + π ϵ 2 r 0 n = 0 ∑ ∞ [ P 2 n ( 0 ) ] 2 ( 2 n + 2 1 + 2 n − 1 1 − ( r 0 / ϵ ) 2 n − 1 ) + R 0 ( x 0 ) − 4 π H ( x 0 ) ( ln ϵ − 2 1 − ( r 0 / ϵ ) 2 ) .
Using the expansion [124 ]
E ( z ) = − π 2 ∑ n = 0 ∞ [ P 2 n ( 0 ) ] 2 2 n − 1 z 2 n ( ∣ z ∣ < 1 ) , (246) E(z)=-\frac{\pi}{2}\sum_{n=0}^{\infty}\frac{[P_{2n}(0)]^2}{2n-1}z^{2n}\qquad(|z|<1), \tag*{(246)} E ( z ) = − 2 π n = 0 ∑ ∞ 2 n − 1 [ P 2 n ( 0 ) ] 2 z 2 n ( ∣ z ∣ < 1 ) , (246)
we evaluate the sums and obtain as ϵ → 0 \epsilon\to0 ϵ → 0 :
w 0 ( Γ ) ( x 0 ) ≈ 2 π 2 ϵ ( E ( r 0 / ϵ ) − 4 3 ) + H ( x Γ ) ( 1 − 2 ( r 0 / ϵ ) 2 ) 16 π , (247) w_0^{(\Gamma)}(x_0)\approx\frac{2}{\pi^2\epsilon}\left(E(r_0/\epsilon)-\frac{4}{3}\right)+\frac{H(x_{\Gamma})(1-2(r_0/\epsilon)^2)}{16\pi}, \tag*{(247)} w 0 ( Γ ) ( x 0 ) ≈ π 2 ϵ 2 ( E ( r 0 / ϵ ) − 3 4 ) + 16 π H ( x Γ ) ( 1 − 2 ( r 0 / ϵ ) 2 ) , (247)
where we replaced H ( x 0 ) H(x_0) H ( x 0 ) by H ( x Γ ) H(x_{\Gamma}) H ( x Γ ) (the error being o ( 1 ) o(1) o ( 1 ) ). In the leading order, one can neglect the second term that yields Eq. (93) .
Substituting Eqs. (5) , (93) into Eq. (51) , we deduce the asymptotic behavior of the coefficients b k ( Γ ) b_k^{(\Gamma)} b k ( Γ ) :
b k ( Γ ) ≈ ϵ − 1 b ^ k , b ^ k = − μ ^ k 2 π ∫ Γ 1 d x ^ w ^ 0 ( ∣ x ^ ∣ ) v ^ k ( x ^ ) , (248) b_k^{(\Gamma)} \approx\epsilon^{-1}\hat{b}_k,\qquad\hat{b}_k=-\frac{\hat{\mu}_k}{2\pi}\int_{\Gamma_1}d\hat{x}\,\hat{w}_0(|\hat{x}|)\,\hat{v}_k(\hat{x}), \tag*{(248)} b k ( Γ ) ≈ ϵ − 1 b ^ k , b ^ k = − 2 π μ ^ k ∫ Γ 1 d x ^ w ^ 0 ( ∣ x ^ ∣ ) v ^ k ( x ^ ) , (248)
where Γ 1 \Gamma_1 Γ 1 is the unit disk on the horizontal plane. As the integral of the axially symmetric function w ^ 0 ( ∣ x ^ ∣ ) \hat{w}_0(|\hat{x}|) w ^ 0 ( ∣ x ^ ∣ ) with any periodically oscillating eigenfunction vanishes, only axially symmetric eigenfunctions v ^ 0 , n k \hat{v}_{0,nk} v ^ 0 , nk can provide contributions O ( ϵ − 1 ) O(\epsilon^{-1}) O ( ϵ − 1 ) to b k ( Γ ) b_k^{(\Gamma)} b k ( Γ ) in the leading order.
As in the planar case, one can employ Eq. (46) to deduce the following relation
b ^ k = V ^ k ( ∞ ) . (249) \hat{b}_k=\hat{V}_k(\infty). \tag*{(249)} b ^ k = V ^ k ( ∞ ) . (249)
We also provide a more formal derivation of this relation by using the pseudo-Green’s function G 0 ( x ∣ x 0 ) G_0(\boldsymbol{x}|\boldsymbol{x}_0) G 0 ( x ∣ x 0 ) for the upper half-space,
G 0 ( x ∣ x 0 ) = 1 4 π ( 1 ∣ x − x 0 ∣ + 1 ∣ x − x 0 ′ ∣ ) , (250) G_0(\boldsymbol{x}|\boldsymbol{x}_0)=\frac{1}{4\pi}\left(\frac{1}{|\boldsymbol{x}-\boldsymbol{x}_0|}+\frac{1}{|\boldsymbol{x}-\boldsymbol{x}_0'|}\right), \tag*{(250)} G 0 ( x ∣ x 0 ) = 4 π 1 ( ∣ x − x 0 ∣ 1 + ∣ x − x 0 ′ ∣ 1 ) , (250)
where x 0 ′ = ( x 0 , y 0 , − z 0 ) \boldsymbol{x}_0'=(x_0,y_0,-z_0) x 0 ′ = ( x 0 , y 0 , − z 0 ) is the mirror reflection of the point x 0 = ( x 0 , y 0 , z 0 ) \boldsymbol{x}_0=(x_0,y_0,z_0) x 0 = ( x 0 , y 0 , z 0 ) . This function satisfies
− Δ G 0 ( x ∣ x 0 ) = δ ( x − x 0 ) ( x ∈ H 3 ) , (251) -\Delta G_0(\boldsymbol{x}|\boldsymbol{x}_0)=\delta(\boldsymbol{x}-\boldsymbol{x}_0)\qquad(\boldsymbol{x}\in\mathbb{H}_3), \tag*{(251)} − Δ G 0 ( x ∣ x 0 ) = δ ( x − x 0 ) ( x ∈ H 3 ) , (251)
∂ n G 0 ( x ∣ x 0 ) = 0 ( x ∈ ∂ H 3 ) , (252) \partial_nG_0(\boldsymbol{x}|\boldsymbol{x}_0)=0\qquad(\boldsymbol{x}\in\partial\mathbb{H}_3), \tag*{(252)} ∂ n G 0 ( x ∣ x 0 ) = 0 ( x ∈ ∂ H 3 ) , (252)
G 0 ( x ∣ x 0 ) ∼ 1 2 π ∣ x ∣ + o ( 1 ) ( ∣ x ∣ → ∞ ) . (253) G_0(\boldsymbol{x}|\boldsymbol{x}_0)\sim\frac{1}{2\pi|\boldsymbol{x}|}+o(1)\qquad(|\boldsymbol{x}|\to\infty). \tag*{(253)} G 0 ( x ∣ x 0 ) ∼ 2 π ∣ x ∣ 1 + o ( 1 ) ( ∣ x ∣ → ∞ ) . (253)
Multiplying Eq. (3.5a) by G 0 ( x ∣ x 0 ) G_0(\boldsymbol{x}|\boldsymbol{x}_0) G 0 ( x ∣ x 0 ) , multiplying Eq. (G.27a) by V ^ k ( x ) \hat{V}_k(\boldsymbol{x}) V ^ k ( x ) , subtracting them, integrating over x ∈ H 3 \boldsymbol{x}\in\mathbb{H}_3 x ∈ H 3 and using the Green’s formula, we get for any x 0 ∈ H 3 \boldsymbol{x}_0\in\mathbb{H}_3 x 0 ∈ H 3
V ^ k ( x 0 ) = V ^ k ( ∞ ) + μ ^ k ∫ Γ 1 d x G 0 ( x ∣ x 0 ) v ^ k ( x ) . (254) \hat{V}_k(\boldsymbol{x}_0)=\hat{V}_k(\infty)+\hat{\mu}_k\int_{\Gamma_1}d\boldsymbol{x}\,G_0(\boldsymbol{x}|\boldsymbol{x}_0)\,\hat{v}_k(\boldsymbol{x}). \tag*{(254)} V ^ k ( x 0 ) = V ^ k ( ∞ ) + μ ^ k ∫ Γ 1 d x G 0 ( x ∣ x 0 ) v ^ k ( x ) . (254)
Substituting
∫ Γ 1 d x 0 G 0 ( x ∣ x 0 ) = w ^ 0 ( ∣ x ∣ ) 2 + 8 3 π ( x ∈ Γ 1 ) (255) \int_{\Gamma_1}d\boldsymbol{x}_0\,G_0(\boldsymbol{x}|\boldsymbol{x}_0)=\frac{\hat{w}_0(|\boldsymbol{x}|)}{2}+\frac{8}{3\pi}\qquad(\boldsymbol{x}\in\Gamma_1) \tag*{(255)} ∫ Γ 1 d x 0 G 0 ( x ∣ x 0 ) = 2 w ^ 0 ( ∣ x ∣ ) + 3 π 8 ( x ∈ Γ 1 ) (255)
into Eq. (G.24), we have
b ^ k = − μ ^ k π ∫ Γ 1 d x v ^ k ( x ) [ − 8 3 π + ∫ Γ 1 d x 0 G 0 ( x ∣ x 0 ) ] = − μ ^ k π ∫ Γ 1 d x 0 ∫ Γ 1 d x v ^ k ( x ) G 0 ( x ∣ x 0 ) = V ^ k ( ∞ ) , \begin{aligned}
\hat{b}_k&=-\frac{\hat{\mu}_k}{\pi}\int_{\Gamma_1}d\boldsymbol{x}\,\hat{v}_k(\boldsymbol{x})\left[-\frac{8}{3\pi}+\int_{\Gamma_1}d\boldsymbol{x}_0\,G_0(\boldsymbol{x}|\boldsymbol{x}_0)\right]\\
&=-\frac{\hat{\mu}_k}{\pi}\int_{\Gamma_1}d\boldsymbol{x}_0\int_{\Gamma_1}d\boldsymbol{x}\,\hat{v}_k(\boldsymbol{x})G_0(\boldsymbol{x}|\boldsymbol{x}_0)=\hat{V}_k(\infty),
\end{aligned} b ^ k = − π μ ^ k ∫ Γ 1 d x v ^ k ( x ) [ − 3 π 8 + ∫ Γ 1 d x 0 G 0 ( x ∣ x 0 ) ] = − π μ ^ k ∫ Γ 1 d x 0 ∫ Γ 1 d x v ^ k ( x ) G 0 ( x ∣ x 0 ) = V ^ k ( ∞ ) ,
where we used the orthogonality of v ^ k \hat{v}_k v ^ k to a constant.
To complete this section, we also evaluate two sums used in Sec. 5. According to Eqs. (G.8, G.24), we have
C 2 = ∑ k = 1 ∞ b ^ k 2 μ ^ k 2 = 1 2 π ∫ 0 1 d r ^ r ^ w ^ 0 2 ( r ^ ) = 8 π 3 ∫ 0 1 d r ^ r ^ ( E ( r ^ ) − 4 3 ) 2 ≈ 0.0031 , (256) C_2=\sum_{k=1}^{\infty}\frac{\hat{b}_k^2}{\hat{\mu}_k^2}=\frac{1}{2\pi}\int_0^1 d\hat{r}\,\hat{r}\,\hat{w}_0^2(\hat{r})=\frac{8}{\pi^3}\int_0^1 d\hat{r}\,\hat{r}\left(E(\hat{r})-\frac{4}{3}\right)^2\approx0.0031, \tag*{(256)} C 2 = k = 1 ∑ ∞ μ ^ k 2 b ^ k 2 = 2 π 1 ∫ 0 1 d r ^ r ^ w ^ 0 2 ( r ^ ) = π 3 8 ∫ 0 1 d r ^ r ^ ( E ( r ^ ) − 3 4 ) 2 ≈ 0.0031 , (256)
where this integral involving the explicit function w ^ 0 ( r ^ ) \hat{w}_0(\hat{r}) w ^ 0 ( r ^ ) from Eq. (94) was computed numerically.
To proceed, we recall that a harmonic function V ^ k ( x 0 ) \hat{V}_k(\boldsymbol{x}_0) V ^ k ( x 0 ) can be extended via the harmonic measure density h ( x ∣ x 0 ) h(\boldsymbol{x}|\boldsymbol{x}_0) h ( x ∣ x 0 ) on the unit disk Γ 1 \Gamma_1 Γ 1 in H 3 \mathbb{H}_3 H 3 :
V ^ k ( x 0 ) = ∫ Γ 1 d x v ^ k ( x ) h ( x ∣ x 0 ) ( x 0 ∈ H 3 ) . (257) \hat{V}_k(\boldsymbol{x}_0)=\int_{\Gamma_1}d\boldsymbol{x}\,\hat{v}_k(\boldsymbol{x})h(\boldsymbol{x}|\boldsymbol{x}_0)\qquad(\boldsymbol{x}_0\in\mathbb{H}_3). \tag*{(257)} V ^ k ( x 0 ) = ∫ Γ 1 d x v ^ k ( x ) h ( x ∣ x 0 ) ( x 0 ∈ H 3 ) . (257)
In the limit ∣ x 0 ∣ → ∞ |\boldsymbol{x}_0|\to\infty ∣ x 0 ∣ → ∞ , the density h ( x ∣ x 0 ) h(\boldsymbol{x}|\boldsymbol{x}_0) h ( x ∣ x 0 ) admits a simple form,
h ( x ∣ ∞ ) = 1 2 π 1 − ∣ x ∣ 2 ( x ∈ Γ 1 ) , (258) h(\boldsymbol{x}|\infty)=\frac{1}{2\pi\sqrt{1-|\boldsymbol{x}|^2}}\qquad(\boldsymbol{x}\in\Gamma_1), \tag*{(258)} h ( x ∣∞ ) = 2 π 1 − ∣ x ∣ 2 1 ( x ∈ Γ 1 ) , (258)
that follows from the Weber’s solution for an electrified disk [125 ] . Combining this expression with Eqs. (G.24, G.25), we obtain
∑ k = 1 ∞ [ V ^ k ( ∞ ) ] 2 μ ^ k = − ∑ k = 1 ∞ ∫ Γ 1 d x 2 π w ^ 0 ( ∣ x ∣ ) v ^ k ( x ) ∫ Γ 1 d x ′ h ( x ′ ∣ ∞ ) v ^ k ( x ′ ) = − 1 2 π ∫ Γ 1 d x w ^ 0 ( ∣ x ∣ ) ∫ Γ 1 d x ′ h ( x ′ ∣ ∞ ) ∑ k = 1 ∞ v ^ k ( x ) v ^ k ( x ′ ) ⏟ = δ ( x − x ′ ) − 1 / π = − 1 2 π ∫ Γ 1 d x w ^ 0 ( ∣ x ∣ ) h ( x ∣ ∞ ) , \begin{aligned}
\sum_{k=1}^{\infty}\frac{[\hat{V}_k(\infty)]^2}{\hat{\mu}_k}
&=-\sum_{k=1}^{\infty}\int_{\Gamma_1}\frac{d\boldsymbol{x}}{2\pi}\,\hat{w}_0(|\boldsymbol{x}|)\hat{v}_k(\boldsymbol{x})\int_{\Gamma_1}d\boldsymbol{x}'\,h(\boldsymbol{x}'|\infty)\hat{v}_k(\boldsymbol{x}')\\
&=-\frac{1}{2\pi}\int_{\Gamma_1}d\boldsymbol{x}\,\hat{w}_0(|\boldsymbol{x}|)\int_{\Gamma_1}d\boldsymbol{x}'\,h(\boldsymbol{x}'|\infty)\underbrace{\sum_{k=1}^{\infty}\hat{v}_k(\boldsymbol{x})\hat{v}_k(\boldsymbol{x}')}_{=\delta(\boldsymbol{x}-\boldsymbol{x}')-1/\pi}\\
&=-\frac{1}{2\pi}\int_{\Gamma_1}d\boldsymbol{x}\,\hat{w}_0(|\boldsymbol{x}|)h(\boldsymbol{x}|\infty),
\end{aligned} k = 1 ∑ ∞ μ ^ k [ V ^ k ( ∞ ) ] 2 = − k = 1 ∑ ∞ ∫ Γ 1 2 π d x w ^ 0 ( ∣ x ∣ ) v ^ k ( x ) ∫ Γ 1 d x ′ h ( x ′ ∣∞ ) v ^ k ( x ′ ) = − 2 π 1 ∫ Γ 1 d x w ^ 0 ( ∣ x ∣ ) ∫ Γ 1 d x ′ h ( x ′ ∣∞ ) = δ ( x − x ′ ) − 1/ π k = 1 ∑ ∞ v ^ k ( x ) v ^ k ( x ′ ) = − 2 π 1 ∫ Γ 1 d x w ^ 0 ( ∣ x ∣ ) h ( x ∣∞ ) ,
where we used the orthogonality of v ^ k \hat{v}_k v ^ k to a constant. We conclude that
C 1 = ∑ k = 1 ∞ [ V ^ k ( ∞ ) ] 2 μ ^ k = − 2 π 2 ∫ 0 1 d r ^ r ^ E ( r ^ ) − 4 / 3 1 − r ^ 2 = 8 3 π 2 − 1 4 ≈ 0.0202. (259) C_1=\sum_{k=1}^{\infty}\frac{[\hat{V}_k(\infty)]^2}{\hat{\mu}_k}=-\frac{2}{\pi^2}\int_0^1d\hat{r}\,\hat{r}\frac{E(\hat{r})-4/3}{\sqrt{1-\hat{r}^2}}=\frac{8}{3\pi^2}-\frac{1}{4}\approx0.0202. \tag*{(259)} C 1 = k = 1 ∑ ∞ μ ^ k [ V ^ k ( ∞ ) ] 2 = − π 2 2 ∫ 0 1 d r ^ r ^ 1 − r ^ 2 E ( r ^ ) − 4/3 = 3 π 2 8 − 4 1 ≈ 0.0202. (259)
Lower and upper bounds for eigenvalues
When Ω \Omega Ω is the disk of radius R R R and Γ \Gamma Γ is the arc of angle 2 ε 2\varepsilon 2 ε , one can easily get lower and upper bounds for the eigenvalues μ k ( p , Γ ) \mu_k^{(p,\Gamma)} μ k ( p , Γ ) via variational arguments:
μ k ( p , 2 ε , N ) ≤ μ k ( p , Γ ) ≤ μ k ( p , 2 ε , D ) , (260) \mu_k^{(p,2\varepsilon,N)}\leq\mu_k^{(p,\Gamma)}\leq\mu_k^{(p,2\varepsilon,D)}, \tag*{(260)} μ k ( p , 2 ε , N ) ≤ μ k ( p , Γ ) ≤ μ k ( p , 2 ε , D ) , (260)
where μ k ( p , α , N ) \mu_k^{(p,\alpha,N)} μ k ( p , α , N ) and μ k ( p , α , D ) \mu_k^{(p,\alpha,D)} μ k ( p , α , D ) are the eigenvalues of two auxiliary mixed Steklov problems in the sector of angle α \alpha α and radius R R R , Ω α = { ( r , θ ) : 0 < r < R , 0 < θ < α } \Omega_\alpha=\{(r,\theta):0<r<R,\ 0<\theta<\alpha\} Ω α = {( r , θ ) : 0 < r < R , 0 < θ < α } , when the Steklov condition is imposed on the arc ( r = R ) (r=R) ( r = R ) , while either Neumann or Dirichlet condition is imposed on two radial segments ( θ = 0 (\theta=0 ( θ = 0 and θ = α ) \theta=\alpha) θ = α ) . As this technique is fairly standard (see [38 ] for details), we just sketch the arguments for the principal eigenvalue. For the original Steklov-Neumann problem in the disk Ω \Omega Ω , one has to minimize the Rayleigh quotient:
μ 0 ( p , Γ ) = inf U ∈ H 1 ( Ω ) { ∥ ∇ U ∥ L 2 ( Ω ) 2 + p D ∥ U ∥ L 2 ( Ω ) 2 ∥ U ∣ Γ ∥ L 2 ( Γ ) 2 } . (261) \mu_0^{(p,\Gamma)}=\inf_{U\in H^1(\Omega)}\left\{\frac{\lVert\nabla U\rVert_{L^2(\Omega)}^2+\frac{p}{D}\lVert U\rVert_{L^2(\Omega)}^2}{\lVert U|_{\Gamma}\rVert_{L^2(\Gamma)}^2}\right\}. \tag*{(261)} μ 0 ( p , Γ ) = U ∈ H 1 ( Ω ) inf { ∥ U ∣ Γ ∥ L 2 ( Γ ) 2 ∥ ∇ U ∥ L 2 ( Ω ) 2 + D p ∥ U ∥ L 2 ( Ω ) 2 } . (261)
where the infimum is taken over all suitable functions (in the space H 1 ( Ω ) H^1(\Omega) H 1 ( Ω ) ). As the sector Ω 2 ε \Omega_{2\varepsilon} Ω 2 ε is a subset of the disk Ω \Omega Ω , the principal eigenvalue μ 0 ( p , 2 ε , N ) \mu_0^{(p,2\varepsilon,N)} μ 0 ( p , 2 ε , N ) in the sector is determined by the same Rayleigh quotient, in which the two norms in the numerator are evaluated over a smaller domain, implying the lower bound in Eq. (H.1). In turn, the principal eigenvalue μ 0 ( p , 2 ε , D ) \mu_0^{(p,2\varepsilon,D)} μ 0 ( p , 2 ε , D ) is determined by imposing the Dirichlet condition on two radial segments that reduces the space of functions and thus increases the infimum, implying the upper bound in Eq. (H.1).
The eigenmodes of the mixed Steklov-Neumann and Steklov-Dirichlet problems in the sector are known explicitly due to the separation of variables:
V k ( p , α , N ) ∝ I ν k N ( r p / D ) cos ( ν k N θ ) , ν k N = π k α , (262) V_k^{(p,\alpha,N)}\propto I_{\nu_k^N}\left(r\sqrt{p/D}\right)\cos\left(\nu_k^N\theta\right),\qquad\nu_k^N=\frac{\pi k}{\alpha}, \tag*{(262)} V k ( p , α , N ) ∝ I ν k N ( r p / D ) cos ( ν k N θ ) , ν k N = α π k , (262)
V k ( p , α , D ) ∝ I ν k D ( r p / D ) sin ( ν k D θ ) , ν k D = π ( k + 1 ) α , (263) V_k^{(p,\alpha,D)}\propto I_{\nu_k^D}\left(r\sqrt{p/D}\right)\sin\left(\nu_k^D\theta\right),\qquad\nu_k^D=\frac{\pi(k+1)}{\alpha}, \tag*{(263)} V k ( p , α , D ) ∝ I ν k D ( r p / D ) sin ( ν k D θ ) , ν k D = α π ( k + 1 ) , (263)
with k = 0 , 1 , 2 , … k=0,1,2,\ldots k = 0 , 1 , 2 , … , so that
μ k ( p , α , N / D ) = p / D I ν k N / D ′ ( R p / D ) I ν k N / D ( R p / D ) . (264) \mu_k^{(p,\alpha,N/D)}=\sqrt{p/D}\frac{I_{\nu_k^{N/D}}'\left(R\sqrt{p/D}\right)}{I_{\nu_k^{N/D}}\left(R\sqrt{p/D}\right)}. \tag*{(264)} μ k ( p , α , N / D ) = p / D I ν k N / D ( R p / D ) I ν k N / D ′ ( R p / D ) . (264)
One sees that these eigenvalues are of the same form, except for ν k N / D \nu_k^{N/D} ν k N / D , which are shifted between the Neumann and Dirichlet cases. We conclude that
z I π k 2 ε ′ ( z ) I π k 2 ε ( z ) ≤ R μ k ( p , Γ ) ≤ z I π ( k + 1 ) 2 ε ′ ( z ) I π ( k + 1 ) 2 ε ( z ) ( z = R p / D ) , (265) z\frac{I_{\frac{\pi k}{2\varepsilon}}'(z)}{I_{\frac{\pi k}{2\varepsilon}}(z)}\leq R\mu_k^{(p,\Gamma)}\leq z\frac{I_{\frac{\pi(k+1)}{2\varepsilon}}'(z)}{I_{\frac{\pi(k+1)}{2\varepsilon}}(z)}\qquad(z=R\sqrt{p/D}), \tag*{(265)} z I 2 ε π k ( z ) I 2 ε π k ′ ( z ) ≤ R μ k ( p , Γ ) ≤ z I 2 ε π ( k + 1 ) ( z ) I 2 ε π ( k + 1 ) ′ ( z ) ( z = R p / D ) , (265)
which are valid for any k k k , ε \varepsilon ε and p p p . In the limit p → 0 p\to0 p → 0 , one gets much simpler bounds
π k 2 ε ≤ μ k ( 0 , Γ ) ≤ π ( k + 1 ) 2 ε . (266) \frac{\pi k}{2\varepsilon}\leq\mu_k^{(0,\Gamma)}\leq\frac{\pi(k+1)}{2\varepsilon}. \tag*{(266)} 2 ε π k ≤ μ k ( 0 , Γ ) ≤ 2 ε π ( k + 1 ) . (266)
These bounds agree with the asymptotic behavior (5) as ε → 0 \varepsilon\to0 ε → 0 , together with the asymptotic behavior (90) for the eigenvalues μ ^ k \hat{\mu}_k μ ^ k as k → ∞ k\to\infty k → ∞ . Moreover, one can check that each eigenvalue μ ^ k \hat{\mu}_k μ ^ k listed in Table 1 , lies indeed between the bounds π k / 2 \pi k/2 π k /2 and π ( k + 1 ) / 2 \pi(k+1)/2 π ( k + 1 ) /2 .
In the same vein, one can derive bounds for the case of a spherical cap of angle ε \varepsilon ε on the sphere of radius R R R by considering two auxiliary mixed Steklov problems in the spherical sector of angle ε \varepsilon ε . We focus on the axially symmetric eigenfunctions in the sector, which have the form i ν k N / D ( r p / D ) P ν k N / D ( cos θ ) i_{\nu_k^{N/D}}\left(r\sqrt{p/D}\right)P_{\nu_k^{N/D}}(\cos\theta) i ν k N / D ( r p / D ) P ν k N / D ( cos θ ) , where the ν k N / D \nu_k^{N/D} ν k N / D are fixed by imposing boundary conditions on the conical boundary: P ν k D ( cos ε ) = 0 P_{\nu_k^D}(\cos\varepsilon)=0 P ν k D ( cos ε ) = 0 for the Dirichlet case, and P ν k N ′ ( cos ε ) = 0 P_{\nu_k^N}'(\cos\varepsilon)=0 P ν k N ′ ( cos ε ) = 0 for the Neumann case. The associated eigenvalues are
μ k ( p , ε , N / D ) = p / D i ν k N / D ′ ( R p / D ) i ν k N / D ( R p / D ) , (267) \mu_k^{(p,\varepsilon,N/D)}=\sqrt{p/D}\frac{i_{\nu_k^{N/D}}'\left(R\sqrt{p/D}\right)}{i_{\nu_k^{N/D}}\left(R\sqrt{p/D}\right)}, \tag*{(267)} μ k ( p , ε , N / D ) = p / D i ν k N / D ( R p / D ) i ν k N / D ′ ( R p / D ) , (267)
which are reduced to μ k ( 0 , ε , N / D ) = ν k N / D / R \mu_k^{(0,\varepsilon,N/D)}=\nu_k^{N/D}/R μ k ( 0 , ε , N / D ) = ν k N / D / R at p = 0 p=0 p = 0 . While the values of ν k N / D \nu_k^{N/D} ν k N / D need to be found numerically, their asymptotic behavior for large k k k can be obtained by using the large-ν \nu ν asymptotic relation [124 ]
P ν ( cos ε ) ≈ Γ ( ν + 1 ) Γ ( ν + 3 / 2 ) ( π sin ε 2 ) − 1 / 2 cos ( ( ν + 1 / 2 ) ε − π / 4 ) + O ( ν − 1 ) , (268) P_\nu(\cos\varepsilon)\approx\frac{\Gamma(\nu+1)}{\Gamma(\nu+3/2)}\left(\frac{\pi\sin\varepsilon}{2}\right)^{-1/2}\cos\left((\nu+1/2)\varepsilon-\pi/4\right)+O(\nu^{-1}), \tag*{(268)} P ν ( cos ε ) ≈ Γ ( ν + 3/2 ) Γ ( ν + 1 ) ( 2 π sin ε ) − 1/2 cos ( ( ν + 1/2 ) ε − π /4 ) + O ( ν − 1 ) , (268)
so that
ν k D ≈ π ( k + 3 / 4 ) ε − 1 2 ( k ≫ 1 ) . (269) \nu_k^D\approx\frac{\pi(k+3/4)}{\varepsilon}-\frac{1}{2}\qquad(k\gg1). \tag*{(269)} ν k D ≈ ε π ( k + 3/4 ) − 2 1 ( k ≫ 1 ) . (269)
Similarly, the asymptotic analysis of P ν ′ ( cos ε ) P_\nu'(\cos\varepsilon) P ν ′ ( cos ε ) yields
ν k N ≈ π ( k + 1 / 4 ) ε − 1 2 ( k ≫ 1 ) (270) \nu_k^N\approx\frac{\pi(k+1/4)}{\varepsilon}-\frac{1}{2}\qquad(k\gg1) \tag*{(270)} ν k N ≈ ε π ( k + 1/4 ) − 2 1 ( k ≫ 1 ) (270)
(see more discussions in [130 ] ). As a consequence, we get two bounds for the eigenvalues associated to axially symmetric eigenfunctions, which are valid for any 0 < ε ≤ π 0<\varepsilon\le\pi 0 < ε ≤ π :
π ( k + 1 / 4 ) ε − 1 2 ≤ R μ 0 , k ( 0 , Γ ) ≤ π ( k + 3 / 4 ) ε − 1 2 ( k ≫ 1 ) . (271) \frac{\pi(k+1/4)}{\varepsilon}-\frac{1}{2}\le R\mu_{0,k}^{(0,\Gamma)}\le\frac{\pi(k+3/4)}{\varepsilon}-\frac{1}{2}\qquad(k\gg1). \tag*{(271)} ε π ( k + 1/4 ) − 2 1 ≤ R μ 0 , k ( 0 , Γ ) ≤ ε π ( k + 3/4 ) − 2 1 ( k ≫ 1 ) . (271)
Combining these bounds with the asymptotic relation (5) as ε → 0 \varepsilon\to0 ε → 0 , one gets
μ ^ 0 , k ≃ π k ( k ≫ 1 ) . (272) \hat{\mu}_{0,k}\simeq\pi k\qquad(k\gg1). \tag*{(272)} μ ^ 0 , k ≃ π k ( k ≫ 1 ) . (272)
References