Introduction and statement of the results
The inverse problem and related works
The Calderón problem, we are interested in, asks whether an interior coefficient can be recovered from the corresponding boundary measurement map. In its original formulation, the coefficient is an electrical conductivity and the measurements are encoded by the Dirichlet-to-Neumann map [12 ] . The advancement of the problem began with the construction of complex geometrical optics (CGO) solutions, by Sylvester and Uhlmann, in [42 ] , which established global uniqueness for n ≥ 3 n \ge3 n ≥ 3 , see also the original work by Ramm [43 ] . Later on, Nachman [34 ] and Novikov [36 ] developed methods to compute the interior potential from the boundary measurements. The initial CGO-based results led to a research direction to determine the minimal regularity conditions on the conductivity under which uniqueness holds. These assumptions have been relaxed from smooth coefficients down to conductivities in various fractional Sobolev and Besov spaces (see, [7 , 38 , 28 , 8 , 37 , 24 ] ). Finally, uniqueness results for Lipschitz conductivities were established in [23 ] , and further relaxations of conductivities are established in [14 ] . The two-dimensional problem requires different arguments, with global uniqueness proved using ∂ ˉ \bar{\partial} ∂ ˉ methods [35 ] and extended to L ∞ L^\infty L ∞ conductivities using quasiconformal mappings [6 ] . Bukhgeim [11 ] introduced an alternative framework using stationary phase techniques. For an account of these developments, see [44 , 39 ] .
The stability estimates in [1 ] and the instability result in [31 ] describe an important difficulty of recovery from remote measurements. One way to remedy to this instability, and obtain additional information, is to perturb the medium by injecting contrast agents whose size, shape and material parameters are known. The resulting inverse problem uses the change in the measured field. When the agents are small and highly contrasting, their resonances can produce a leading-order perturbation which contains information about the surrounding medium. This enters into the framework of imaging using contrast agents. In the recent years, we proposed a modeling and analysis of such approaches where the main idea is the use of the resonant behavior of the contrast agents in its different manifestations. In this direction, this approach has been analyzed for acoustic imaging with nearly resonating bubbles [15 ] , for photo-acoustic imaging with dielectric particles [18 ] , and for plasmonic particles in the full Maxwell system [19 , 20 ] . Optical inversion using plasmonic contrast agents is considered in [13 ] . In the time domain, injected droplets have been used for recovering a wave source [41 ] and, subsequently, both the wave speed and the source [40 ] . We refer to [17 ] for an introduction to the mathematical analysis of imaging modalities using small contrast agents, and for a review of the resonance mechanisms and their use in inverse problems. These works concern different measurement settings and different resonances. We use, in the current work, the acoustic droplet model and the boundary-map formulation of [21 ] .
In [21 ] , all the droplets are injected at once contrary to [15 , 18 , 19 , 20 , 41 , 40 ] where the agents were injected in isolation (i.e. one after another). A real incident frequency is chosen near a Newtonian resonance. The boundary map of the perturbed medium is shown to approach that of an effective medium with coefficient ω 2 n 2 − P 2 \omega^2 n^2 - P^2 ω 2 n 2 − P 2 . The parameter P 2 > 0 P^2 > 0 P 2 > 0 can be made large by adjusting the detuning. This gives a coercive reference problem and allows the effective boundary map to be linearized before applying CGO solutions. Our purpose here is to replace the periodic distribution in [21 ] by a random distribution with a prescribed, possibly nonconstant, density. We formulate the configurations on a probability space and work on events on which the geometry and the empirical measure satisfy uniform bounds. For each realization in such an event, the proof is deterministic: we derive the Foldy–Lax system and compare it with a density-weighted integral equation. We do not use stochastic two-scale convergence. The dependence on P P P is traced in each estimate. The use of geometrically constrained random configurations is also present in [9 ] , where the authors establish resonant homogenization for dielectric rods with random positions, radii and permittivities through stochastic two-scale analysis. Here, we use the point-interaction approximation and quantitative empirical-measure estimates to derive an effective equation for randomly distributed acoustic droplets. Our motivation is to apply this approximation to the Calderón problem, with boundary-map estimates that retain the dependence on the tunable resonant amplitude.
The effective perturbation is − P 2 ρ ( x ) -P^{2}\rho(x) − P 2 ρ ( x ) . Thus, for a nonconstant spatial density, the reference operator is − Δ + P 2 ρ ( x ) -\Delta+P^{2}\rho(x) − Δ + P 2 ρ ( x ) rather than − Δ + P 2 -\Delta+P^{2} − Δ + P 2 . We prove the corresponding resolvent estimates and linearization, and construct CGO solutions for this known reference operator. The boundary norms of these solutions grow with their complex phase. We therefore retain these norms when applying the linearization to measured data which quantify to which extent the reconstruction is valid. Appendix § 10 treats reconstruction from the full boundary map by a boundary integral equation, following the constructive approach of [34 , 36 ] while the perturbation estimates are related to those used in [29 ] .
Our reconstruction strategy relies on tuning the incident frequency near a common droplet resonance to generate a negative effective perturbation of controllably large magnitude. Random variations in the material parameters would shift the individual resonances and could weaken this collective response or introduce contributions of opposite signs. Preserving the required sign and amplitude would therefore require additional restrictions relating the resonance spread to the chosen detuning, which becomes smaller as the desired amplitude increases. We consequently keep the droplet size, shape and material parameters fixed and introduce randomness only in their centers and occupancies. This choice also reflects a natural idealization for contrast-agent imaging, namely the injected agents are prepared with prescribed and calibrated characteristics, whereas their individual locations after injection may not be controlled. Our analysis thus isolates spatial uncertainty while preserving the resonant mechanism used for reconstruction.
We would like to finish this introduction with the following statement. In the current work, we start from the original model − Δ − ω 2 n 2 -\Delta-\omega^{2}n^{2} − Δ − ω 2 n 2 , where n n n is the unknown coefficient and ω \omega ω is the operating incident frequency. Then after injecting the droplets, all at once, with a given distribution of their locations, and passing to the homogenized model, we have the two choices according to whether the used incident frequency is chosen close to the resonance from the left-side or the right-side. Accordingly, we have two outcome:
Using frequencies from the left-side generates an effective model with the background of the form − Δ + P 2 ρ -\Delta+P^{2}\rho − Δ + P 2 ρ . In this case, in terms of inverse problems, we can linearize the effective measurement map. The drawback is that we get limited band-frequency for the stability estimates with Fourier modes at most in the band [ 0 , ω ] [0,\omega] [ 0 , ω ] .
Using incident frequencies at the right-side of the resonant frequency generate an effective background of the form − Δ − P 2 ρ -\Delta-P^{2}\rho − Δ − P 2 ρ . In this case, a linearization of the effective boundary map, in terms of P P P , is not possible. But the inversion of the non-linear map provides stable Fourier bands of the form [ 0 , ω 2 + P 2 ] [0,\sqrt{\omega^{2}+P^{2}}] [ 0 , ω 2 + P 2 ] (in the case ρ = 1 \rho=1 ρ = 1 ), which suggest, in principle, and increase of the quantitative stability as P P P grows. Such results will be reported in a forthcoming work.
As a concluding message, using such contrast agents we can breakdown the two well known obstacles of the Calderón-problem, namely non-linearity and instability, in the following sense.
using left-side resonant frequencies, we can linearize the boundary map but we are limited in stability.
using right-side resonant frequencies, we cannot linearize the boundary map but we can improve the stability by increasing the stable Fourier-band.
Therefore, either we breakdown the non-linearity or the instability but not both. To simultaneously linearize and stabilize the inverse problem, one tentative is to choose the resonant incident frequencies, with a small damping, to be able to replace the effective potential ± P 2 \pm P^{2} ± P 2 by a convenient complex parameter having convenient (large) real and imaginary parts that could allow linearization (with large imaginary part) and stability with (large) real part. But this understanding needs to be confirmed.
Let us emphasize that we are not suggesting to take the incident frequency large to improve stability, as it is suggested by Isakov [26 ] and others [27 ] , rather, we use the resonant effect to increase the effective background potential (i.e. P 2 P^{2} P 2 ).
The acoustic model
Let Ω ⊂ R 3 \Omega\subset\mathbb{R}^{3} Ω ⊂ R 3 be a bounded domain with C 2 , α C^{2,\alpha} C 2 , α boundary, where 0 < α < 1 0 < \alpha< 1 0 < α < 1 . The physical mass density is a constant r 0 > 0 r_{0} > 0 r 0 > 0 . We write
n 2 = r 0 k b ∈ W 1 , ∞ ( Ω ) , 0 < n ∗ 2 ≤ n 2 ≤ n ∗ 2 , n^{2} = \frac{r_{0}}{k_{b}} \in W^{1,\infty}(\Omega), \qquad0 < n_{*}^{2} \le n^{2} \le n^{*2}, n 2 = k b r 0 ∈ W 1 , ∞ ( Ω ) , 0 < n ∗ 2 ≤ n 2 ≤ n ∗ 2 ,
where k b k_{b} k b is the bulk modulus of the background. The notation ρ \rho ρ will be reserved for the spatial density of the droplet centers. Set
X = H − 1 / 2 ( ∂ Ω ) , Y = H 1 / 2 ( ∂ Ω ) , q ω = ω 2 n 2 . \mathcal{X} = H^{-1/2}(\partial\Omega), \qquad\mathcal{Y} = H^{1/2}(\partial\Omega), \qquad q_{\omega} = \omega^{2}n^{2}. X = H − 1/2 ( ∂ Ω ) , Y = H 1/2 ( ∂ Ω ) , q ω = ω 2 n 2 .
Unless indicated otherwise, boundary duality pairings are bilinear, obtained by complex-linear extension of the real duality. Conjugates will be written explicitly in energy estimates.
We assume that there is a compact interval I ⋐ ( 0 , ∞ ) I \Subset(0,\infty) I ⋐ ( 0 , ∞ ) , with ω 0 \omega_{0} ω 0 in its interior, such that the Neumann realization of A 0 , ω = − Δ − q ω A_{0,\omega} = -\Delta- q_{\omega} A 0 , ω = − Δ − q ω satisfies 2
sup ω ∈ I ∥ A 0 , ω − 1 ∥ L 2 ( Ω ) → L 2 ( Ω ) ≤ C I . (1) \sup_{\omega\in I} \left\|A_{0,\omega}^{-1}\right\|_{L^{2}(\Omega)\to L^{2}(\Omega)} \le C_{I}. \tag*{(1)} ω ∈ I sup A 0 , ω − 1 L 2 ( Ω ) → L 2 ( Ω ) ≤ C I . (1)
Here ω 0 \omega_{0} ω 0 is the droplet resonance specified below. For f ∈ X f \in\mathcal{X} f ∈ X , let S f S^{f} S f solve
A 0 , ω S f = 0 in Ω , ∂ ν S f = f on ∂ Ω , Λ 0 , ω f = γ S f . (2) A_{0,\omega}S^{f} = 0 \quad\text{in } \Omega, \qquad\partial_{\nu}S^{f} = f \quad\text{on } \partial\Omega, \qquad\Lambda_{0,\omega}f = \gamma S^{f}. \tag*{(2)} A 0 , ω S f = 0 in Ω , ∂ ν S f = f on ∂ Ω , Λ 0 , ω f = γ S f . (2)
The assumption (1) gives ∥ S f ∥ H 1 ( Ω ) ≤ C ∥ f ∥ X \left\|S^{f}\right\|_{H^{1}(\Omega)} \le C\left\|f\right\|_{\mathcal{X}} S f H 1 ( Ω ) ≤ C ∥ f ∥ X , uniformly for ω ∈ I \omega\in I ω ∈ I . A justification, including the kernel estimates used below, is given in Appendix § 12 .
Let G ω G_{\omega} G ω be the real symmetric Neumann Green kernel defined by
( − Δ x − q ω ( x ) ) G ω ( x , y ) = δ y , ∂ ν x G ω ( x , y ) = 0. (3) (-\Delta_{x} - q_{\omega}(x))G_{\omega}(x,y) = \delta_{y}, \qquad\partial_{\nu_{x}}G_{\omega}(x,y) = 0. \tag*{(3)} ( − Δ x − q ω ( x )) G ω ( x , y ) = δ y , ∂ ν x G ω ( x , y ) = 0. (3)
We use the decomposition
G ω ( x , y ) = Φ 0 ( x , y ) + H ω ( x , y ) , Φ 0 ( x , y ) = 1 4 π ∣ x − y ∣ . G_{\omega}(x,y) = \Phi_{0}(x,y) + H_{\omega}(x,y), \qquad\Phi_{0}(x,y) = \frac{1}{4\pi\lvert x-y\rvert}. G ω ( x , y ) = Φ 0 ( x , y ) + H ω ( x , y ) , Φ 0 ( x , y ) = 4 π ∣ x − y ∣ 1 .
The estimates needed in the sequel are, see Lemma B.1,
∣ G ω ( x , y ) ∣ ≤ C ∣ x − y ∣ − 1 , ∣ ∇ x G ω ( x , y ) ∣ + ∣ ∇ y G ω ( x , y ) ∣ ≤ C ∣ x − y ∣ − 2 , (4) \lvert G_{\omega}(x,y)\rvert\le C\lvert x-y\rvert^{-1}, \qquad\lvert\nabla_{x}G_{\omega}(x,y)\rvert+ \lvert\nabla_{y}G_{\omega}(x,y)\rvert\le C\lvert x-y\rvert^{-2}, \tag*{(4)} ∣ G ω ( x , y )∣ ≤ C ∣ x − y ∣ − 1 , ∣ ∇ x G ω ( x , y )∣ + ∣ ∇ y G ω ( x , y )∣ ≤ C ∣ x − y ∣ − 2 , (4)
∣ H ω ( x , y ) ∣ ≤ C b − 1 , ∣ ∇ x H ω ( x , y ) ∣ + ∣ ∇ y H ω ( x , y ) ∣ ≤ C b − 2 . (5) \lvert H_{\omega}(x,y)\rvert\le Cb^{-1}, \qquad\lvert\nabla_{x}H_{\omega}(x,y)\rvert+ \lvert\nabla_{y}H_{\omega}(x,y)\rvert\le Cb^{-2}. \tag*{(5)} ∣ H ω ( x , y )∣ ≤ C b − 1 , ∣ ∇ x H ω ( x , y )∣ + ∣ ∇ y H ω ( x , y )∣ ≤ C b − 2 . (5)
In (5) , 0 < b ≤ b 1 0 < b \le b_{1} 0 < b ≤ b 1 , and x , y x,y x , y belong to a ball of radius comparable to b b b , whose distance from ∂ Ω \partial\Omega ∂ Ω is at least a fixed multiple of b b b . The value estimate holds on the diagonal; the derivative estimate is understood off the diagonal, or as a Lipschitz estimate. Constants are uniform for ω ∈ I \omega\in I ω ∈ I .
With the background framework and its associated resolvent estimates established, we now introduce the physical perturbation. The continuous background Ω \Omega Ω is subsequently perturbed by injecting a large ensemble of M M M randomly distributed, subwavelength acoustic droplets. The geometric properties and resonant scaling of these individual inclusions are formalised as follows.
Assumption 1.1 (The droplets and the resonance). Each droplet is of the form D j = z j + a B D_j = z_j + aB D j = z j + a B , where B ⊂ R 3 B \subset\mathbb{R}^{3} B ⊂ R 3 is a fixed bounded C 1 , 1 C^{1,1} C 1 , 1 domain containing the origin. Its mass density is ρ 0 \rho_0 ρ 0 and its bulk modulus is k 0 a 2 k_0a^2 k 0 a 2 , with k 0 > 0 k_0 > 0 k 0 > 0 . Let
( N B v ) ( ξ ) = ∫ B v ( η ) 4 π ∣ ξ − η ∣ d η . (N_Bv)(\xi) = \int_B \frac{v(\eta)}{4\pi|\xi-\eta|}\,\mathrm{d}\eta. ( N B v ) ( ξ ) = ∫ B 4 π ∣ ξ − η ∣ v ( η ) d η . Choose an isolated eigenvalue λ ∗ > 0 \lambda_* > 0 λ ∗ > 0 of N B N_B N B and denote its orthogonal spectral projection by Π ∗ \Pi_* Π ∗ . We assume
b ∗ : = ∥ Π ∗ 1 ∥ L 2 ( B ) 2 > 0 , ω 0 2 = k 0 ρ 0 λ ∗ . (6) b_* := \|\Pi_*1\|_{L^2(B)}^2 > 0,\qquad\omega_0^2 = \frac{k_0}{\rho_0\lambda_*}. \tag*{(6)} b ∗ := ∥ Π ∗ 1 ∥ L 2 ( B ) 2 > 0 , ω 0 2 = ρ 0 λ ∗ k 0 . (6) Only the centers, and possibly their total number, are random.
Fix 1 / 3 < h < 1 1/3 < h < 1 1/3 < h < 1 and introduce
w = a 1 − h , d = w 1 / 3 , t = P 2 ≥ t 0 ≥ 1. (7) w = a^{1-h},\qquad d = w^{1/3},\qquad t = P^2 \ge t_0 \ge1. \tag*{(7)} w = a 1 − h , d = w 1/3 , t = P 2 ≥ t 0 ≥ 1. (7) The scale d d d is deterministic. It provides a lower bound, up to a constant, for the actual distance between distinct centers. We describe the configurations on a probability space ( Ξ , F , P ) (\Xi,\mathcal{F},\mathbb{P}) ( Ξ , F , P ) and set ϖ ∈ Ξ \varpi\in\Xi ϖ ∈ Ξ to denote a realization. For each a a a , M ( a , ϖ ) M(a,\varpi) M ( a , ϖ ) denotes the total number of droplets in realization ϖ \varpi ϖ . The map M ( a , ⋅ ) M(a,\cdot) M ( a , ⋅ ) is integer-valued and measurable, and z j z_j z j is measurable on { M ( a , ⋅ ) ≥ j } \{M(a,\cdot) \ge j\} { M ( a , ⋅ ) ≥ j } .
Assumption 1.2 (Separation and boundary clearance). There are measurable events E a ⊂ Ξ E_a \subset\Xi E a ⊂ Ξ and deterministic constants c 0 , c b , C M > 0 c_0,c_b,C_M > 0 c 0 , c b , C M > 0 such that, for ϖ ∈ E a \varpi\in E_a ϖ ∈ E a ,
min i ≠ j ∣ z i − z j ∣ ≥ c 0 d , min j dist ( z j , ∂ Ω ) ≥ c b d , w M ≤ C M . \min_{i\ne j}|z_i-z_j| \ge c_0d,\qquad\min_j \operatorname{dist}(z_j,\partial\Omega) \ge c_bd,\qquad wM \le C_M. i = j min ∣ z i − z j ∣ ≥ c 0 d , j min dist ( z j , ∂ Ω ) ≥ c b d , w M ≤ C M . The events and constants do not depend on P P P or on the boundary input.
Since a / d → 0 a/d \to0 a / d → 0 , this assumption implies, for sufficiently small a a a , that the droplets are disjoint and contained in Ω \Omega Ω . Their union will be denoted by D D D .
Assumption 1.3 (The distribution of the centers). Let ρ ∈ C 1 ( Ω ‾ ) \rho\in C^1(\overline{\Omega}) ρ ∈ C 1 ( Ω ) be known and satisfy 0 < ρ ∗ ≤ ρ ≤ ρ ∗ 0 < \rho_* \le\rho\le\rho^* 0 < ρ ∗ ≤ ρ ≤ ρ ∗ . Define
μ a ( ϖ ) : = w ∑ j = 1 M ( a , ϖ ) δ z j ( ϖ ) , D a ( ϖ ) : = sup ∥ φ ∥ C 1 ( Ω ‾ ) ≤ 1 ∣ ∫ φ d μ a − ∫ Ω ρ φ d x ∣ . \mu_a(\varpi) := w\sum_{j=1}^{M(a,\varpi)}\delta_{z_j(\varpi)},\qquad\mathcal{D}_a(\varpi) := \sup_{\|\varphi\|_{C^1(\overline{\Omega})}\le1}\left|\int\varphi\,\mathrm{d}\mu_a-\int_\Omega\rho\varphi\,\mathrm{d}x\right|. μ a ( ϖ ) := w j = 1 ∑ M ( a , ϖ ) δ z j ( ϖ ) , D a ( ϖ ) := ∥ φ ∥ C 1 ( Ω ) ≤ 1 sup ∫ φ d μ a − ∫ Ω ρφ d x . where ∥ φ ∥ C 1 = ∥ φ ∥ ∞ + ∥ ∇ φ ∥ ∞ \|\varphi\|_{C^1} = \|\varphi\|_\infty+ \|\nabla\varphi\|_\infty ∥ φ ∥ C 1 = ∥ φ ∥ ∞ + ∥∇ φ ∥ ∞ . There are deterministic quantities D a → 0 D_a \to0 D a → 0 and p a → 0 p_a \to0 p a → 0 such that
P ( E a ) ≥ 1 − p a , D a ( ϖ ) ≤ D a for ϖ ∈ E a . (8) \mathbb{P}(E_a) \ge1-p_a,\qquad\mathcal{D}_a(\varpi) \le D_a\quad\text{for }\varpi\in E_a. \tag*{(8)} P ( E a ) ≥ 1 − p a , D a ( ϖ ) ≤ D a for ϖ ∈ E a . (8) The supremum in this assumption can be taken over a countable dense subset of the C 1 C^1 C 1 unit ball, so that D a \mathcal{D}_a D a is measurable. This is a common-event estimate, not a law of large numbers for one fixed test function. We shall apply it to regularized Green kernels and to
products of functions depending on P P P . It also determines the normalization of the total number of droplets:
∣ w M − ∫ Ω ρ d x ∣ ≤ D a . (9) \left|wM-\int_{\Omega}\rho\,\mathrm{d}x\right|\leq\mathcal{D}_a. \tag*{(9)} w M − ∫ Ω ρ d x ≤ D a . (9)
In particular, no normalization ∫ Ω ρ = 1 \int_{\Omega}\rho=1 ∫ Ω ρ = 1 is imposed. Section § 8 constructs random configurations satisfying Assumptions 1.2 and 1.3 for every fixed density in this class.
The near-resonant frequency and the effective equation
We choose the real frequency
ω a , P 2 = ω 0 2 ( 1 + b ∗ λ ∗ P 2 a h ) , c ∗ ( P ) = − k 0 b ∗ λ ∗ P 2 < 0. (10) \omega_{a,P}^{2}=\omega_{0}^{2}\left(1+\frac{b_*}{\lambda_*P^{2}}a^{h}\right),\qquad c_*(P)=-\frac{k_0b_*}{\lambda_*P^{2}}<0. \tag*{(10)} ω a , P 2 = ω 0 2 ( 1 + λ ∗ P 2 b ∗ a h ) , c ∗ ( P ) = − λ ∗ P 2 k 0 b ∗ < 0. (10)
For fixed P P P , this is the detuning in [21 ] , equation (1.6). The choice of its sign will give a negative scattering coefficient. For a a a sufficiently small, the frequencies in (1.10) belong to I I I , uniformly for P ≥ P 0 P\geq P_0 P ≥ P 0 . In every comparison below, all the boundary maps are evaluated at this same frequency. We often write ω = ω a , P \omega=\omega_{a,P} ω = ω a , P and suppress it in G ω G_\omega G ω .
The exact acoustic field is the solution of
{ [ − Δ − q ω ( 1 − 1 D ) − ω 0 2 r 0 2 k 0 a 2 1 D ] u a f = 0 in Ω , ∂ ν u a f = f on ∂ Ω . (11) \left\{
\begin{aligned}
\left[-\Delta-q_\omega(1-\mathbb{1}_D)-\frac{\omega_0^2r_0^2}{k_0a^2}\mathbb{1}_D\right]u_a^f&=0 &&\text{in }\Omega,\\
\partial_\nu u_a^f&=f &&\text{on }\partial\Omega.
\end{aligned}
\right. \tag*{(11)} ⎩ ⎨ ⎧ [ − Δ − q ω ( 1 − 1 D ) − k 0 a 2 ω 0 2 r 0 2 1 D ] u a f ∂ ν u a f = 0 = f in Ω , on ∂ Ω. (11)
We set Λ a , P ( ϖ ) f = γ u a f \Lambda_{a,P}(\varpi)f=\gamma u_a^f Λ a , P ( ϖ ) f = γ u a f . The constant mass density makes (11) a scalar equation with a zeroth-order contrast (i.e. no contrast occurs in its principal part). The effective field and its boundary map are defined by
{ ( − Δ + P 2 ρ − q ω ) u ρ , P f = 0 in Ω , ∂ ν u ρ , P f = f on ∂ Ω , Λ ρ , P e f f f = γ u ρ , P f . (12) \left\{
\begin{aligned}
(-\Delta+P^2\rho-q_\omega)u_{\rho,P}^f&=0 &&\text{in }\Omega,\\
\partial_\nu u_{\rho,P}^f&=f &&\text{on }\partial\Omega,
\end{aligned}
\right.
\qquad
\Lambda_{\rho,P}^{\mathrm{eff}}f=\gamma u_{\rho,P}^f. \tag*{(12)} { ( − Δ + P 2 ρ − q ω ) u ρ , P f ∂ ν u ρ , P f = 0 = f in Ω , on ∂ Ω , Λ ρ , P eff f = γ u ρ , P f . (12)
Theorem 1.4. Suppose that Assumptions 1.1–1.3 hold. Put
ϵ a = ( D a + d ) 1 / 5 , E ( a , P ) = P 2 a h + P 4 d + P ϵ a . (13) \epsilon_a=(\mathcal{D}_a+d)^{1/5},\qquad E(a,P)=P^2a^h+P^4d+P\epsilon_a. \tag*{(13)} ϵ a = ( D a + d ) 1/5 , E ( a , P ) = P 2 a h + P 4 d + P ϵ a . (13) There are constants a 0 , c a d m , C > 0 a_0,c_{\mathrm{adm}},C>0 a 0 , c adm , C > 0 and P 0 ≥ 1 P_0\geq1 P 0 ≥ 1 such that, if
0 < a < a 0 , P ≥ P 0 , P ϵ a ≤ c a d m , (14) 0<a<a_0,\qquad P\geq P_0,\qquad P\epsilon_a\leq c_{\mathrm{adm}}, \tag*{(14)} 0 < a < a 0 , P ≥ P 0 , P ϵ a ≤ c adm , (14) then (11) is uniquely solvable for every ϖ ∈ E a \varpi\in E_a ϖ ∈ E a , and
∥ Λ a , P ( ϖ ) − Λ ρ , P e f f ∥ L ( X , Y ) ≤ C ( P 2 a h + P 4 d + P ϵ a ) . (15) \left\|\Lambda_{a,P}(\varpi)-\Lambda_{\rho,P}^{\mathrm{eff}}\right\|_{\mathcal{L}(X,Y)}
\leq C\left(P^2a^h+P^4d+P\epsilon_a\right). \tag*{(15)} Λ a , P ( ϖ ) − Λ ρ , P eff L ( X , Y ) ≤ C ( P 2 a h + P 4 d + P ϵ a ) . (15) Moreover, the Foldy–Lax matrix
( B a , P Y ) i = Y i + P 2 ω ∑ j ≠ i G ω ( z i , z j ) Y j (16) (B_{a,P}Y)_i=Y_i+P^2\omega\sum_{j\ne i}G_\omega(z_i,z_j)Y_j \tag*{(16)} ( B a , P Y ) i = Y i + P 2 ω j = i ∑ G ω ( z i , z j ) Y j (16) is invertible and satisfies
∥ B a , P − 1 ∥ ℓ ω 2 → ℓ ω 2 ≤ C , ∥ Y ∥ ℓ ω 2 2 = ω ∑ i ∣ Y i ∣ 2 . (17) \left\|B_{a,P}^{-1}\right\|_{\ell_\omega^2\to\ell_\omega^2}\leq C,\qquad
\|Y\|_{\ell_\omega^2}^2=\omega\sum_i|Y_i|^2. \tag*{(17)} B a , P − 1 ℓ ω 2 → ℓ ω 2 ≤ C , ∥ Y ∥ ℓ ω 2 2 = ω i ∑ ∣ Y i ∣ 2 . (17) The constants are independent of a a a , P P P , M M M , ϖ \varpi ϖ and of the boundary inputs. The assertions hold on the same event for all the parameters in (14) .
Here and below, the constants may depend on Ω \Omega Ω , B B B , h h h , k 0 k_0 k 0 , r 0 r_0 r 0 , the spectral gap of the chosen eigenvalue, the bounds on n 2 n^2 n 2 , the background resolvent bound, the geometric constants in Assumption 1.2 , and ρ ∗ − 1 \rho_*^{-1} ρ ∗ − 1 , ρ ∗ \rho_* ρ ∗ , ∥ ρ ∥ C 1 \|\rho\|_{C^1} ∥ ρ ∥ C 1 . Uniformity over a class of densities means that these bounds are common to the class. The strict positivity of ρ \rho ρ is used in the large-P P P coercivity and linearization estimates.
The proof is given in § 3 –§ 7 . We first compute the scattering coefficient, then establish the continuous and discrete inverse estimates, and finally derive the point-interaction system and compare the boundary maps. The weighted norm in (17) makes these comparisons independent of the growing number of droplets.
Let us emphasize the difference between the present argument and the one in [21 ] . In [21 Assumption 1] , the centers are deterministic: there is one droplet in each translated cube of volume a 1 − h = w a^{1-h}=w a 1 − h = w , the array is periodic in the interior of the domain, and the boundary layer is handled deterministically. In that setting the discrete measure
μ a P e r i o d i c = w ∑ j δ z j \mu_a^{\mathrm{Periodic}}=w\sum_j\delta_{z_j} μ a Periodic = w j ∑ δ z j
is a deterministic Riemann-sum measure; after the normalization used in [21 ] , its limit is Lebesgue measure. Consequently, the passage from the Foldy–Lax system to the continuous equation is based on cellwise quadrature for a fixed configuration.
Here neither the number of centers in a cell nor their positions are fixed. The basic object is the random empirical measure
μ a ( ϖ ) = w ∑ j = 1 M ( a , ϖ ) δ z j ( ϖ ) , \mu_a(\varpi)=w\sum_{j=1}^{M(a,\varpi)}\delta_{z_j(\varpi)}, μ a ( ϖ ) = w j = 1 ∑ M ( a , ϖ ) δ z j ( ϖ ) ,
and the deterministic periodic quadrature condition of [21 ] is replaced by the common-event estimate
P { sup ∥ ϕ ∥ C 1 ( Ω ‾ ) ≤ 1 ∣ ∫ ϕ d μ a − ∫ Ω ρ ϕ d x ∣ ≤ D a } ≥ 1 − p a . (18) \mathbb{P}\left\{\sup_{\|\phi\|_{C^1(\overline{\Omega})}\leq1}\left|\int\phi\,\mathrm{d}\mu_a-\int_{\Omega}\rho\phi\,\mathrm{d}x\right|\leq D_a\right\}\geq1-p_a. \tag*{(18)} P { ∥ ϕ ∥ C 1 ( Ω ) ≤ 1 sup ∫ ϕ d μ a − ∫ Ω ρϕ d x ≤ D a } ≥ 1 − p a . (18)
combined with the hard-core separation and boundary-clearance conditions in Assumption 1.2 . Thus the randomness is isolated in the construction of an event E a \mathcal{E}_a E a of large probability. Once ϖ ∈ E a \varpi\in\mathcal{E}_a ϖ ∈ E a is fixed, every estimate in § 4 –§ 7 is a deterministic estimate, with constants independent of the realization. In particular, no expectation of the field, no stochastic corrector and no stochastic two-scale limit enters the Foldy–Lax analysis. This is also different in purpose from the stochastic homogenization framework of [9 ] , namely we use a probability space to quantify admissible random configurations, but the effective-medium comparison is proved realization by realization on E a \mathcal{E}_a E a .
There are three further consequences of this change. First, the event in (18) controls the whole C 1 C^1 C 1 unit ball on one and the same realization. It is stronger than convergence for a single fixed test function and can therefore be applied to the regularized Green kernels and to P P P -dependent products used later. Second, the limiting coefficient is P 2 ρ ( x ) P^2\rho(x) P 2 ρ ( x ) , rather than the spatially constant effective coefficient produced by the periodic unit-density array in [21 ] . Third, since E a \mathcal{E}_a E a is independent of P P P and of the boundary inputs, the deterministic estimates obtained on E a \mathcal{E}_a E a hold simultaneously for all admissible P P P and all boundary data. This produces both the high-probability statement below and, along summable sequences of scales, an almost-sure statement by Borel–Cantelli.
Corollary 1.5. For a deterministic choice P = P ( a ) P=P(a) P = P ( a ) satisfying (14) ,
P ( ∥ Λ a , P − Λ P , ρ e f f ∥ ≤ C E ( a , P ) ) ≥ 1 − p a . \mathbb{P}\left(\left\|\Lambda_{a,P}-\Lambda^{\mathrm{eff}}_{P,\rho}\right\|\leq CE(a,P)\right)\geq1-p_a. P ( Λ a , P − Λ P , ρ eff ≤ C E ( a , P ) ) ≥ 1 − p a .
If E ( a , P ( a ) ) → 0 E(a,P(a)) \to0 E ( a , P ( a )) → 0 , the difference of these maps converges to zero in probability. If a k ↓ 0 a_k \downarrow0 a k ↓ 0 and ∑ k p a k < ∞ \sum_k p_{a_k} < \infty ∑ k p a k < ∞ , there is a measurable set Ξ t y p ⊂ Ξ \Xi_{\mathrm{typ}} \subset\Xi Ξ typ ⊂ Ξ of probability one such that (15) holds for every ϖ ∈ Ξ t y p \varpi\in\Xi_{\mathrm{typ}} ϖ ∈ Ξ typ and all sufficiently large k k k , simultaneously for all admissible P P P .
Proof. The boundary-map error is measurable. To see this, fix a a a , P P P , M M M and regard the centers as parameters. On the fixed space L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) , the Lippmann–Schwinger operator is
T D = A 0 , ω − 1 [ ( ω 2 r 0 k 0 a 2 − q ω ) 1 D ⋅ ] . T_D = A_{0,\omega}^{-1}\left[\left(\frac{\omega^2 r_0}{k_0a^2} - q_\omega\right)1_D \cdot\right]. T D = A 0 , ω − 1 [ ( k 0 a 2 ω 2 r 0 − q ω ) 1 D ⋅ ] .
Translations of the droplets are continuous in measure, and sup y ∥ G ω ( ⋅ , y ) ∥ 2 ≤ C \sup_y \lVert G_\omega(\cdot,y)\rVert_2 \le C sup y ∥ G ω ( ⋅ , y ) ∥ 2 ≤ C . Hence T D T_D T D is continuous in Hilbert–Schmidt norm. On the open set where I − T D I-T_D I − T D is invertible, the field is continuous in L 2 L^2 L 2 for each fixed boundary input. The volume representation and A 0 , ω − 1 : L 2 → H 2 A_{0,\omega}^{-1}:L^2 \to H^2 A 0 , ω − 1 : L 2 → H 2 then give continuity of its boundary trace. Countable dense subsets of the unit balls of X \mathcal{X} X and Y ∗ \mathcal{Y}^* Y ∗ determine the operator norm, proving its measurability. Taking the countable union over M M M gives the same conclusion for the random configuration.
Put
A a ( P ) = { ϖ : ∥ Λ a , P ( ϖ ) − Λ P , P e f f ∥ ≤ C E ( a , P ) } . A_a(P) = \left\{\varpi: \lVert\Lambda_{a,P}(\varpi)-\Lambda_{P,P}^{\mathrm{eff}}\rVert\le CE(a,P)\right\}. A a ( P ) = { ϖ : ∥ Λ a , P ( ϖ ) − Λ P , P eff ∥ ≤ C E ( a , P ) } .
Theorem 1.4 states precisely that E a ⊂ A a ( P ) \mathcal{E}_a \subset A_a(P) E a ⊂ A a ( P ) for every admissible P P P , with the same constant C C C . Therefore
P ( A a ( P ) ) ≥ P ( E a ) ≥ 1 − p a , \mathbb{P}(A_a(P)) \ge\mathbb{P}(\mathcal{E}_a) \ge1-p_a, P ( A a ( P )) ≥ P ( E a ) ≥ 1 − p a ,
which is the first assertion.
Now let P = P ( a ) P=P(a) P = P ( a ) be deterministic and assume E ( a , P ( a ) ) → 0 E(a,P(a)) \to0 E ( a , P ( a )) → 0 . Fix η > 0 \eta>0 η > 0 . There exists a η > 0 a_\eta>0 a η > 0 such that C E ( a , P ( a ) ) ≤ η CE(a,P(a)) \le\eta C E ( a , P ( a )) ≤ η whenever a < a η a<a_\eta a < a η . For these values of a a a ,
{ ∥ Λ a , P ( a ) − Λ P ( a ) , P ( a ) e f f ∥ > η } ⊂ E a c . \left\{\lVert\Lambda_{a,P(a)}-\Lambda_{P(a),P(a)}^{\mathrm{eff}}\rVert> \eta\right\} \subset\mathcal{E}_a^c. { ∥ Λ a , P ( a ) − Λ P ( a ) , P ( a ) eff ∥ > η } ⊂ E a c .
Hence
P { ∥ Λ a , P ( a ) − Λ P ( a ) , P ( a ) e f f ∥ > η } ≤ p a ⟶ 0 , \mathbb{P}\left\{\lVert\Lambda_{a,P(a)}-\Lambda_{P(a),P(a)}^{\mathrm{eff}}\rVert> \eta\right\} \le p_a \longrightarrow0, P { ∥ Λ a , P ( a ) − Λ P ( a ) , P ( a ) eff ∥ > η } ≤ p a ⟶ 0 ,
which is exactly convergence in probability.
For the almost-sure statement, set F k = E a k c F_k=\mathcal{E}_{a_k}^c F k = E a k c . By (8) ,
∑ k = 1 ∞ P ( F k ) ≤ ∑ k = 1 ∞ p a k < ∞ . \sum_{k=1}^{\infty}\mathbb{P}(F_k) \le\sum_{k=1}^{\infty}p_{a_k} < \infty. k = 1 ∑ ∞ P ( F k ) ≤ k = 1 ∑ ∞ p a k < ∞.
The first Borel–Cantelli lemma, which does not require independence, gives P ( lim sup k F k ) = 0 \mathbb{P}(\limsup_k F_k)=0 P ( lim sup k F k ) = 0 . Equivalently,
Ξ t y p : = ( lim sup k F k ) c = ⋃ K = 1 ∞ ⋂ k ≥ K E a k , P ( Ξ t y p ) = 1. (19) \Xi_{\mathrm{typ}} := \left(\limsup_k F_k\right)^c = \bigcup_{K=1}^{\infty}\bigcap_{k\ge K}\mathcal{E}_{a_k}, \qquad\mathbb{P}(\Xi_{\mathrm{typ}})=1. \tag*{(19)} Ξ typ := ( k lim sup F k ) c = K = 1 ⋃ ∞ k ≥ K ⋂ E a k , P ( Ξ typ ) = 1. (19)
Thus, for every ϖ ∈ Ξ t y p \varpi\in\Xi_{\mathrm{typ}} ϖ ∈ Ξ typ , there exists a finite K ( ϖ ) K(\varpi) K ( ϖ ) such that ϖ ∈ E a k \varpi\in\mathcal{E}_{a_k} ϖ ∈ E a k for all k ≥ K ( ϖ ) k \ge K(\varpi) k ≥ K ( ϖ ) . Applying Theorem 1.4 at each of those scales gives (15) . Since membership in E a k \mathcal{E}_{a_k} E a k does not involve P P P , and the theorem is uniform in P P P , the same K ( ϖ ) K(\varpi) K ( ϖ ) works simultaneously for every P P P satisfying (14) at the corresponding scale, and also for every boundary input. □ \square □
The scattering coefficient
We begin with the single-droplet computation. It is the spectral calculation of [21 Appendix A.5] , with the detuning parameter written in terms of P P P . Set
θ = ω a , P 2 r 0 k 0 = 1 λ ∗ ( 1 + b ∗ a h λ ∗ t ) , κ a = θ a − 2 , t = P 2 . (20) \theta= \frac{\omega_{a,P}^2 r_0}{k_0} = \frac{1}{\lambda_*}\left(1+\frac{b_*a^h}{\lambda_*t}\right), \qquad\kappa_a=\theta a^{-2}, \qquad t=P^2. \tag*{(20)} θ = k 0 ω a , P 2 r 0 = λ ∗ 1 ( 1 + λ ∗ t b ∗ a h ) , κ a = θ a − 2 , t = P 2 . (20)
On L 2 ( B ) L^2(B) L 2 ( B ) , define
R a = ( I − θ N B ) − 1 , ψ a = R a 1 , m a = ∫ B ψ a , α a = a θ m a . (21) R_a=(I-\theta N_B)^{-1},\qquad\psi_a=R_a1,\qquad m_a=\int_B\psi_a,\qquad\alpha_a=a\theta m_a. \tag*{(21)} R a = ( I − θ N B ) − 1 , ψ a = R a 1 , m a = ∫ B ψ a , α a = a θ m a . (21)
The operators and quantities in (21) also depend on P P P ; this dependence is suppressed in the notation.
Lemma 2.1. For a sufficiently small a a a and t ≥ t 0 ≥ 1 t\geq t_0\geq1 t ≥ t 0 ≥ 1 , the inverse R a R_a R a exists. There are constants independent of a a a and t t t such that
m a = − λ ∗ t a − h + O ( 1 ) , α a = − t a 1 − h + a r a , ∣ r a ∣ ≤ C , (22) m_a=-\lambda_*ta^{-h}+O(1),\qquad\alpha_a=-ta^{1-h}+ar_a,\qquad|r_a|\leq C, \tag*{(22)} m a = − λ ∗ t a − h + O ( 1 ) , α a = − t a 1 − h + a r a , ∣ r a ∣ ≤ C , (22) ∥ R a ∥ L 2 ( B ) → L 2 ( B ) + ∥ ψ a ∥ L 2 ( B ) ≤ C t a − h , c t a − h ≤ ∣ m a ∣ ≤ C t a − h . (23) \|R_a\|_{L^2(B)\to L^2(B)}+\|\psi_a\|_{L^2(B)}\leq Cta^{-h},\qquad cta^{-h}\leq|m_a|\leq Cta^{-h}. \tag*{(23)} ∥ R a ∥ L 2 ( B ) → L 2 ( B ) + ∥ ψ a ∥ L 2 ( B ) ≤ C t a − h , c t a − h ≤ ∣ m a ∣ ≤ C t a − h . (23) In particular, α a < 0 \alpha_a<0 α a < 0 . The functional
ℓ a ( v ) = m a − 1 ∫ B R a v = ∫ B ψ a m a v (24) \ell_a(v)=m_a^{-1}\int_B R_av=\int_B\frac{\psi_a}{m_a}v \tag*{(24)} ℓ a ( v ) = m a − 1 ∫ B R a v = ∫ B m a ψ a v (24) is bounded on L 2 ( B ) L^2(B) L 2 ( B ) with norm at most C C C , and ℓ a ( 1 ) = 1 \ell_a(1)=1 ℓ a ( 1 ) = 1 .
Proof. The Newtonian operator is compact and self-adjoint on L 2 ( B ) L^2(B) L 2 ( B ) . Since λ ∗ \lambda_* λ ∗ is isolated, there is δ ∗ > 0 \delta_*>0 δ ∗ > 0 such that
∣ 1 − λ λ ∗ ∣ ≥ δ ∗ for λ ∈ σ ( N B ) ∖ { λ ∗ } . \left|1-\frac{\lambda}{\lambda_*}\right|\geq\delta_*\qquad\text{for }\lambda\in\sigma(N_B)\setminus\{\lambda_*\}. 1 − λ ∗ λ ≥ δ ∗ for λ ∈ σ ( N B ) ∖ { λ ∗ } .
The possible accumulation point 0 0 0 causes no difficulty in this inequality. By (20) , θ → λ ∗ − 1 \theta\to\lambda_*^{-1} θ → λ ∗ − 1 uniformly for t ≥ t 0 t\geq t_0 t ≥ t 0 . After reducing a 0 a_0 a 0 , we therefore have ∣ 1 − θ λ ∣ ≥ δ ∗ / 2 |1-\theta\lambda|\geq\delta_*/2 ∣1 − θ λ ∣ ≥ δ ∗ /2 on the complementary spectrum. The spectral decomposition gives
R a = Π ∗ 1 − θ λ ∗ + R a ⊥ , R a ⊥ = ( I − θ N B ) − 1 ( I − Π ∗ ) , ∥ R a ⊥ ∥ ≤ 2 δ ∗ . (25) R_a=\frac{\Pi_*}{1-\theta\lambda_*}+R_a^\perp,\qquad R_a^\perp=(I-\theta N_B)^{-1}(I-\Pi_*),\qquad\|R_a^\perp\|\leq\frac{2}{\delta_*}. \tag*{(25)} R a = 1 − θ λ ∗ Π ∗ + R a ⊥ , R a ⊥ = ( I − θ N B ) − 1 ( I − Π ∗ ) , ∥ R a ⊥ ∥ ≤ δ ∗ 2 . (25)
From (20) , we have,
1 − θ λ ∗ = − b ∗ a h λ ∗ t . 1-\theta\lambda_*=-\frac{b_*a^h}{\lambda_*t}. 1 − θ λ ∗ = − λ ∗ t b ∗ a h .
thus the denominator is nonzero. Taking the inner product of (25) with 1 1 1 yields
m a = b ∗ 1 − θ λ ∗ + ⟨ R a ⊥ 1 , 1 ⟩ L 2 ( B ) = − λ ∗ t a − h + β a , ∣ β a ∣ ≤ C . (26) m_a=\frac{b_*}{1-\theta\lambda_*}+\langle R_a^\perp1,1\rangle_{L^2(B)}
=-\lambda_*ta^{-h}+\beta_a,\qquad|\beta_a|\leq C. \tag*{(26)} m a = 1 − θ λ ∗ b ∗ + ⟨ R a ⊥ 1 , 1 ⟩ L 2 ( B ) = − λ ∗ t a − h + β a , ∣ β a ∣ ≤ C . (26)
Multiplication by a θ a\theta a θ now gives
α a = a 1 λ ∗ ( 1 + b ∗ a h λ ∗ t ) ( − λ ∗ t a − h + β a ) = − t a 1 − h + a ( − b ∗ λ ∗ + θ β a ) . \begin{aligned}
\alpha_a&=a\frac{1}{\lambda_*}\left(1+\frac{b_*a^h}{\lambda_*t}\right)(-\lambda_*ta^{-h}+\beta_a)\\
&=-ta^{1-h}+a\left(-\frac{b_*}{\lambda_*}+\theta\beta_a\right).
\end{aligned} α a = a λ ∗ 1 ( 1 + λ ∗ t b ∗ a h ) ( − λ ∗ t a − h + β a ) = − t a 1 − h + a ( − λ ∗ b ∗ + θ β a ) .
The term in parentheses is uniformly bounded, because θ \theta θ is uniformly bounded. This proves (22) without placing any t t t -dependence in its remainder.
The operator estimate in (23) follows directly from (25) . Since t a − h → ∞ ta^{-h}\to\infty t a − h → ∞ uniformly for t ≥ t 0 t\geq t_0 t ≥ t 0 , the bounded term β a \beta_a β a can be absorbed in (26) , proving the two-sided estimate for m a m_a m a . The same observation shows that the main term in α a \alpha_a α a dominates a r a ar_a a r a , so that α a < 0 \alpha_a<0 α a < 0 .
Finally, R a R_a R a is real and self-adjoint. Consequently ∫ B R a v = ∫ B ψ a v \int_B R_av=\int_B\psi_av ∫ B R a v = ∫ B ψ a v , first for real v v v and then by complex linearity. By Cauchy–Schwarz and (23) ,
∣ ℓ a ( v ) ∣ ≤ ∥ ψ a ∥ L 2 ( B ) ∣ m a ∣ ∥ v ∥ L 2 ( B ) ≤ C ∥ v ∥ L 2 ( B ) . |\ell_a(v)|\leq\frac{\|\psi_a\|_{L^2(B)}}{|m_a|}\|v\|_{L^2(B)}\leq C\|v\|_{L^2(B)}. ∣ ℓ a ( v ) ∣ ≤ ∣ m a ∣ ∥ ψ a ∥ L 2 ( B ) ∥ v ∥ L 2 ( B ) ≤ C ∥ v ∥ L 2 ( B ) .
The equality ℓ a ( 1 ) = 1 \ell_a(1)=1 ℓ a ( 1 ) = 1 follows from the definition of m a m_a m a .
For completeness, the scattering coefficient in physical variables is obtained as follows. The unitary rescaling
( U j v ) ( ξ ) = a 3 / 2 v ( z j + a ξ ) (\mathcal{U}_jv)(\xi)=a^{3/2}v(z_j+a\xi) ( U j v ) ( ξ ) = a 3/2 v ( z j + a ξ )
satisfies U j N D j U j − 1 = a 2 N B \mathcal{U}_jN_{D_j}\mathcal{U}_j^{-1}=a^2N_B U j N D j U j − 1 = a 2 N B . Let
W j = ( κ a − 1 I − N D j ) − 1 1. (27) W_j=(\kappa_a^{-1}I-N_{D_j})^{-1}\mathbf{1}. \tag*{(27)} W j = ( κ a − 1 I − N D j ) − 1 1 . (27)
Since ( κ a − 1 I − N D j ) − 1 = κ a ( I − κ a N D j ) − 1 (\kappa_a^{-1}I-N_{D_j})^{-1}=\kappa_a(I-\kappa_aN_{D_j})^{-1} ( κ a − 1 I − N D j ) − 1 = κ a ( I − κ a N D j ) − 1 , the rescaled function is
W j ( z j + a ξ ) = θ a − 2 ψ a ( ξ ) . W_j(z_j+a\xi)=\theta a^{-2}\psi_a(\xi). W j ( z j + a ξ ) = θ a − 2 ψ a ( ξ ) .
We get then the properties of the scattering coefficient in the original variables:
∫ D j W j = α a , ∥ W j ∥ L 2 ( D j ) = θ a − 1 / 2 ∥ ψ a ∥ L 2 ( B ) ≤ C t a − 1 / 2 − h . (28) \int_{D_j}W_j=\alpha_a,\qquad\|W_j\|_{L^2(D_j)}=\theta a^{-1/2}\|\psi_a\|_{L^2(B)}\leq Cta^{-1/2-h}. \tag*{(28)} ∫ D j W j = α a , ∥ W j ∥ L 2 ( D j ) = θ a − 1/2 ∥ ψ a ∥ L 2 ( B ) ≤ C t a − 1/2 − h . (28)
The projection Π ∗ \Pi_* Π ∗ in the proof includes the whole eigenspace. Therefore no simplicity assumption on λ ∗ \lambda_* λ ∗ is needed. Only its coupling to the constant function, expressed by b ∗ > 0 b_*>0 b ∗ > 0 , is used.
Estimates for the distribution of the centers
Throughout § 4 –§ 7 , we fix a a a and a realization ϖ ∈ E a \varpi\in\mathcal{E}_a ϖ ∈ E a . All constants are independent of this choice. We first derive consequences of separation and of Assumption 1.3.
Packing and quadrature estimates
Lemma 3.1. For r > 0 r>0 r > 0 and x ∈ R 3 x\in\mathbb{R}^3 x ∈ R 3 ,
# { j : ∣ z j − x ∣ < r } ≤ C ( 1 + r / d ) 3 . (29) \#\{j:|z_j-x|<r\}\leq C(1+r/d)^3. \tag*{(29)} # { j : ∣ z j − x ∣ < r } ≤ C ( 1 + r / d ) 3 . (29) For every fixed s > 0 s>0 s > 0 ,
w ∑ j ≠ i ∣ z i − z j ∣ − s ≤ { C s , 0 < s < 3 , C ( 1 + ∣ log d ∣ ) , s = 3 , C s d 3 − s , s > 3. (30) w\sum_{j\ne i}|z_i-z_j|^{-s}\leq
\begin{cases}
C_s, & 0<s<3,\\
C(1+|\log d|), & s=3,\\
C_sd^{3-s}, & s>3.
\end{cases} \tag*{(30)} w j = i ∑ ∣ z i − z j ∣ − s ≤ ⎩ ⎨ ⎧ C s , C ( 1 + ∣ log d ∣ ) , C s d 3 − s , 0 < s < 3 , s = 3 , s > 3. (30) The same bounds hold for column sums. In addition, for r ≥ c d r\geq cd r ≥ c d ,
w ∑ 0 < ∣ z j − z i ∣ < r ∣ z i − z j ∣ − 1 ≤ C r 2 . (31) w\sum_{0<|z_j-z_i|<r}|z_i-z_j|^{-1}\leq Cr^2. \tag*{(31)} w 0 < ∣ z j − z i ∣ < r ∑ ∣ z i − z j ∣ − 1 ≤ C r 2 . (31) Proof. The balls B ( z j , c 0 d / 3 ) B(z_j,c_0d/3) B ( z j , c 0 d /3 ) are disjoint. If ∣ z j − x ∣ < r |z_j-x|<r ∣ z j − x ∣ < r , each of these balls is contained in B ( x , r + c 0 d / 3 ) B(x,r+c_0d/3) B ( x , r + c 0 d /3 ) . Comparison of their volumes gives (29) . Choose c 1 > 0 c_1>0 c 1 > 0 smaller than c 0 c_0 c 0 and divide the distances from z i z_i z i into dyadic shells
2 k c 1 d ≤ ∣ z j − z i ∣ < 2 k + 1 c 1 d , k = 0 , … , K , 2^kc_1d\leq|z_j-z_i|<2^{k+1}c_1d,\qquad k=0,\ldots,K, 2 k c 1 d ≤ ∣ z j − z i ∣ < 2 k + 1 c 1 d , k = 0 , … , K , where 2 K d 2^Kd 2 K d is comparable to diam Ω \operatorname{diam}\Omega diam Ω . There are at most C 2 3 k C2^{3k} C 2 3 k centers in the k k k th shell. Since w = d 3 w=d^3 w = d 3 , its contribution to the weighted sum in (30) is at most C s d 3 − s 2 ( 3 − s ) k C_sd^{3-s}2^{(3-s)k} C s d 3 − s 2 ( 3 − s ) k . Summing this geometric series gives the three alternatives. Restricting the sum to shells with 2 k d ≲ r 2^kd\lesssim r 2 k d ≲ r and taking s = 1 s=1 s = 1 gives (31) . The column estimates are identical, since the distance is symmetric. This dyadic-shell argument, and closely related distribution-uniform interaction estimates, are also used in [10 ] .
Lemma 3.2. For ψ ∈ C 1 ( Ω ‾ ) \psi\in C^1(\overline{\Omega}) ψ ∈ C 1 ( Ω ) ,
max i ∣ w ∑ j ≠ i G ( z i , z j ) ψ ( z j ) − ∫ Ω G ( z i , y ) ρ ( y ) ψ ( y ) d y ∣ ≤ C ( D a 1 / 2 + d 2 ) ∥ ψ ∥ C 1 ( Ω ‾ ) . (32) \max_i \left| w \sum_{j\ne i} G(z_i,z_j)\psi(z_j)-\int_{\Omega}G(z_i,y)\rho(y)\psi(y)\,dy \right| \le C(D_a^{1/2}+d^2)\|\psi\|_{C^1(\overline{\Omega})}. \tag*{(32)} i max w j = i ∑ G ( z i , z j ) ψ ( z j ) − ∫ Ω G ( z i , y ) ρ ( y ) ψ ( y ) d y ≤ C ( D a 1/2 + d 2 ) ∥ ψ ∥ C 1 ( Ω ) . (32) Proof. Take a smooth cutoff χ r \chi_r χ r which vanishes on B ( z i , r ) B(z_i,r) B ( z i , r ) and equals one outside B ( z i , 2 r ) B(z_i,2r) B ( z i , 2 r ) , with ∣ ∇ χ r ∣ ≤ C / r |\nabla\chi_r|\le C/r ∣∇ χ r ∣ ≤ C / r , where r ≥ c d r\ge cd r ≥ c d . The estimates (4) imply
∥ χ r ( ⋅ ) G ( z i , ⋅ ) ψ ∥ C 1 ( Ω ‾ ) ≤ C r − 2 ∥ ψ ∥ C 1 ( Ω ‾ ) \|\chi_r(\cdot)G(z_i,\cdot)\psi\|_{C^1(\overline{\Omega})}\le Cr^{-2}\|\psi\|_{C^1(\overline{\Omega})} ∥ χ r ( ⋅ ) G ( z i , ⋅ ) ψ ∥ C 1 ( Ω ) ≤ C r − 2 ∥ ψ ∥ C 1 ( Ω ) for r r r bounded above by a fixed constant. Assumption 1.3 therefore bounds the discrepancy of the cutoff expression by C D a r − 2 ∥ ψ ∥ C 1 CD_ar^{-2}\|\psi\|_{C^1} C D a r − 2 ∥ ψ ∥ C 1 . Its omitted discrete part is bounded by C r 2 ∥ ψ ∥ ∞ Cr^2\|\psi\|_\infty C r 2 ∥ ψ ∥ ∞ , by (31) . The omitted integral has the same bound, since
∫ Ω ∩ B ( z i , 2 r ) ∣ z i − y ∣ − 1 d y ≤ C r 2 \int_{\Omega\cap B(z_i,2r)}|z_i-y|^{-1}\,dy\le Cr^2 ∫ Ω ∩ B ( z i , 2 r ) ∣ z i − y ∣ − 1 d y ≤ C r 2 and ρ \rho ρ is bounded. The complete difference is consequently bounded by C ( r 2 + D a r − 2 ) ∥ ψ ∥ C 1 C(r^2+D_ar^{-2})\|\psi\|_{C^1} C ( r 2 + D a r − 2 ) ∥ ψ ∥ C 1 . Taking r = max { c d , D a 1 / 4 } r=\max\{cd,D_a^{1/4}\} r = max { c d , D a 1/4 } proves (32) . If D a = 0 D_a=0 D a = 0 , take r = c d r=cd r = c d .
The estimate (32) gives the expected density-weighted integral equation for smooth functions. For a uniform inverse estimate, we also need a comparison on Sobolev spaces. We establish it by replacing each atom with a constant density on a small ball.
An auxiliary absolutely continuous measure
Choose b 0 > 0 b_0>0 b 0 > 0 , depending only on c 0 , c b c_0,c_b c 0 , c b , so that the balls B ( z j , 4 b 0 d ) B(z_j,4b_0d) B ( z j , 4 b 0 d ) are disjoint and contained in Ω \Omega Ω . Set
B j d = B ( z j , b 0 d ) , v d = ∣ B j d ∣ , ρ a a v = w v d ∑ j 1 B j d , ν a = ρ a a v d x . (33) B_j^d=B(z_j,b_0d),\qquad v_d=|B_j^d|,\qquad\rho_a^{\mathrm{av}}=\frac{w}{v_d}\sum_j\mathbf{1}_{B_j^d},\qquad\nu_a=\rho_a^{\mathrm{av}}\,dx. \tag*{(33)} B j d = B ( z j , b 0 d ) , v d = ∣ B j d ∣ , ρ a av = v d w j ∑ 1 B j d , ν a = ρ a av d x . (33)
Thus ν a ( B j d ) = w \nu_a(B_j^d)=w ν a ( B j d ) = w and ∥ ρ a a v ∥ ∞ ≤ C \|\rho_a^{\mathrm{av}}\|_\infty\le C ∥ ρ a av ∥ ∞ ≤ C . For every φ ∈ C 1 ( Ω ‾ ) \varphi\in C^1(\overline{\Omega}) φ ∈ C 1 ( Ω ) ,
∣ ∫ Ω φ d ν a − ∫ Ω ρ φ d x ∣ ≤ ∣ w ∑ j 1 v d ∫ B j d ( φ ( x ) − φ ( z j ) ) d x ∣ + D a ∥ φ ∥ C 1 ≤ C ( D a + d ) ∥ φ ∥ C 1 . (34) \left|\int_\Omega\varphi\,d\nu_a-\int_\Omega\rho\varphi\,dx\right|
\le\left|w\sum_j\frac{1}{v_d}\int_{B_j^d}(\varphi(x)-\varphi(z_j))\,dx\right|+D_a\|\varphi\|_{C^1}
\le C(D_a+d)\|\varphi\|_{C^1}. \tag*{(34)} ∫ Ω φ d ν a − ∫ Ω ρφ d x ≤ w j ∑ v d 1 ∫ B j d ( φ ( x ) − φ ( z j )) d x + D a ∥ φ ∥ C 1 ≤ C ( D a + d ) ∥ φ ∥ C 1 . (34)
Here we used w M ≤ C M wM\le C_M w M ≤ C M . These balls are used only to compare operators; they are not substituted for the droplets in the acoustic problem.
Lemma 3.3. With ϵ a = ( D a + d ) 1 / 5 \epsilon_a=(D_a+d)^{1/5} ϵ a = ( D a + d ) 1/5 , one has
∣ ∫ Ω ( ρ a a v − ρ ) u ‾ v d x ∣ ≤ C ϵ a ( ∥ u ∥ H 1 ∥ v ∥ L 2 + ∥ u ∥ L 2 ∥ v ∥ H 1 ) , u , v ∈ H 1 ( Ω ) . (35) \left|\int_\Omega(\rho_a^{\mathrm{av}}-\rho)\overline{u}v\,dx\right|\le C\epsilon_a\left(\|u\|_{H^1}\|v\|_{L^2}+\|u\|_{L^2}\|v\|_{H^1}\right),\qquad u,v\in H^1(\Omega). \tag*{(35)} ∫ Ω ( ρ a av − ρ ) u v d x ≤ C ϵ a ( ∥ u ∥ H 1 ∥ v ∥ L 2 + ∥ u ∥ L 2 ∥ v ∥ H 1 ) , u , v ∈ H 1 ( Ω ) . (35) Proof. Set δ a = D a + d \delta_a=D_a+d δ a = D a + d . Since Ω \Omega Ω is a bounded C 2 , α C^{2,\alpha} C 2 , α domain, there is a linear extension operator
E : H 1 ( Ω ) ⟶ H 1 ( R 3 ) \mathcal{E}:H^1(\Omega)\longrightarrow H^1(\mathbb{R}^3) E : H 1 ( Ω ) ⟶ H 1 ( R 3 ) which is also bounded from L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) to L 2 ( R 3 ) L^2(\mathbb{R}^3) L 2 ( R 3 ) . Let η ∈ C c ∞ ( B ( 0 , 1 ) ) \eta\in C^\infty_c(B(0,1)) η ∈ C c ∞ ( B ( 0 , 1 )) satisfy η ≥ 0 \eta\ge0 η ≥ 0 and ∫ η = 1 \int\eta=1 ∫ η = 1 , and write η r ( x ) = r − 3 η ( x / r ) \eta_r(x)=r^{-3}\eta(x/r) η r ( x ) = r − 3 η ( x / r ) . For 0 < r < r 1 0<r<r_1 0 < r < r 1 define
u r = η r ∗ ( E u ) , v r = η r ∗ ( E v ) . u_r=\eta_r*(\mathcal{E}u),\qquad v_r=\eta_r*(\mathcal{E}v). u r = η r ∗ ( E u ) , v r = η r ∗ ( E v ) . We use the same symbols for their restrictions to Ω \Omega Ω .
We first record the smoothing estimates with their scaling. Since ∫ η r = 1 \int\eta_r=1 ∫ η r = 1 , translation estimates in H 1 ( R 3 ) H^1(\mathbb{R}^3) H 1 ( R 3 ) give
∥ u − u r ∥ L 2 ( Ω ) ≤ C r ∥ u ∥ H 1 ( Ω ) , ∥ u r ∥ L 2 ( Ω ) ≤ C ∥ u ∥ L 2 ( Ω ) . (36) \|u-u_r\|_{L^2(\Omega)}\le Cr\|u\|_{H^1(\Omega)},\qquad\|u_r\|_{L^2(\Omega)}\le C\|u\|_{L^2(\Omega)}. \tag*{(36)} ∥ u − u r ∥ L 2 ( Ω ) ≤ C r ∥ u ∥ H 1 ( Ω ) , ∥ u r ∥ L 2 ( Ω ) ≤ C ∥ u ∥ L 2 ( Ω ) . (36)
Young’s inequality in the form ∥ f ∗ g ∥ ∞ ≤ ∥ f ∥ 2 ∥ g ∥ 2 \|f*g\|_\infty\le\|f\|_2\|g\|_2 ∥ f ∗ g ∥ ∞ ≤ ∥ f ∥ 2 ∥ g ∥ 2 gives
∥ u r ∥ ∞ ≤ ∥ η r ∥ 2 ∥ E u ∥ 2 ≤ C r − 3 / 2 ∥ u ∥ 2 , \|u_r\|_\infty\le\|\eta_r\|_2\|\mathcal{E}u\|_2\le Cr^{-3/2}\|u\|_2, ∥ u r ∥ ∞ ≤ ∥ η r ∥ 2 ∥ E u ∥ 2 ≤ C r − 3/2 ∥ u ∥ 2 ,
while ∥ ∇ η r ∥ 2 = r − 5 / 2 ∥ ∇ η ∥ 2 \|\nabla\eta_r\|_2=r^{-5/2}\|\nabla\eta\|_2 ∥∇ η r ∥ 2 = r − 5/2 ∥∇ η ∥ 2 yields
∥ ∇ u r ∥ ∞ ≤ C r − 5 / 2 ∥ u ∥ 2 . \|\nabla u_r\|_\infty\le Cr^{-5/2}\|u\|_2. ∥∇ u r ∥ ∞ ≤ C r − 5/2 ∥ u ∥ 2 .
The same estimates hold for v r v_r v r . Therefore
∥ u r v r ‾ ∥ ∞ ≤ C r − 3 ∥ u ∥ 2 ∥ v ∥ 2 , \|u_r\overline{v_r}\|_\infty\le Cr^{-3}\|u\|_2\|v\|_2, ∥ u r v r ∥ ∞ ≤ C r − 3 ∥ u ∥ 2 ∥ v ∥ 2 ,
∥ ∇ ( u r v r ‾ ) ∥ ∞ ≤ ∥ ∇ u r ∥ ∞ ∥ v r ∥ ∞ + ∥ u r ∥ ∞ ∥ ∇ v r ∥ ∞ ≤ C r − 4 ∥ u ∥ 2 ∥ v ∥ 2 . (37) \begin{aligned}
\|\nabla(u_r\overline{v_r})\|_\infty
&\le\|\nabla u_r\|_\infty\|v_r\|_\infty+\|u_r\|_\infty\|\nabla v_r\|_\infty\\
&\le Cr^{-4}\|u\|_2\|v\|_2. \tag*{(37)}
\end{aligned} ∥∇ ( u r v r ) ∥ ∞ ≤ ∥∇ u r ∥ ∞ ∥ v r ∥ ∞ + ∥ u r ∥ ∞ ∥∇ v r ∥ ∞ ≤ C r − 4 ∥ u ∥ 2 ∥ v ∥ 2 . (37)
After decreasing r 1 r_1 r 1 so that r ≤ 1 r\le1 r ≤ 1 , the r − 4 r^{-4} r − 4 term dominates, and hence
∥ u r v r ‾ ∥ C 1 ( Ω ‾ ) ≤ C r − 4 ∥ u ∥ L 2 ( Ω ) ∥ v ∥ L 2 ( Ω ) . (38) \|u_r\overline{v_r}\|_{C^1(\overline{\Omega})}\le Cr^{-4}\|u\|_{L^2(\Omega)}\|v\|_{L^2(\Omega)}. \tag*{(38)} ∥ u r v r ∥ C 1 ( Ω ) ≤ C r − 4 ∥ u ∥ L 2 ( Ω ) ∥ v ∥ L 2 ( Ω ) . (38)
We now decompose the desired form. Put q a = ρ a a v − ρ q_a=\rho_a^{\mathrm{av}}-\rho q a = ρ a av − ρ . From (33) and the bounds on ρ \rho ρ , ∥ q a ∥ ∞ ≤ C \|q_a\|_\infty\le C ∥ q a ∥ ∞ ≤ C uniformly in a a a . Thus
∣ ∫ Ω q a ( u v ‾ − u r v r ‾ ) d x ∣ ≤ C ∥ u − u r ∥ 2 ∥ v ∥ 2 + C ∥ u r ∥ 2 ∥ v − v r ∥ 2 ≤ C r ( ∥ u ∥ H 1 ∥ v ∥ 2 + ∥ u ∥ 2 ∥ v ∥ H 1 ) , \left|\int_\Omega q_a(u\overline{v}-u_r\overline{v_r})\,\mathrm{d}x\right|
\le C\|u-u_r\|_2\|v\|_2+C\|u_r\|_2\|v-v_r\|_2
\le Cr\left(\|u\|_{H^1}\|v\|_2+\|u\|_2\|v\|_{H^1}\right), ∫ Ω q a ( u v − u r v r ) d x ≤ C ∥ u − u r ∥ 2 ∥ v ∥ 2 + C ∥ u r ∥ 2 ∥ v − v r ∥ 2 ≤ C r ( ∥ u ∥ H 1 ∥ v ∥ 2 + ∥ u ∥ 2 ∥ v ∥ H 1 ) ,
where (36) was used in the last line.
For the smooth product, note that
∫ Ω q a u r v r ‾ d x = ∫ Ω u r v r ‾ d ν a − ∫ Ω ρ u r v r ‾ d x . \int_\Omega q_au_r\overline{v_r}\,\mathrm{d}x
=\int_\Omega u_r\overline{v_r}\,\mathrm{d}\nu_a-\int_\Omega\rho u_r\overline{v_r}\,\mathrm{d}x. ∫ Ω q a u r v r d x = ∫ Ω u r v r d ν a − ∫ Ω ρ u r v r d x .
Hence (34) and (38) imply
∣ ∫ Ω q a u r v r ‾ d x ∣ ≤ C δ a r − 4 ∥ u ∥ 2 ∥ v ∥ 2 . \left|\int_\Omega q_au_r\overline{v_r}\,\mathrm{d}x\right|
\le C\delta_a r^{-4}\|u\|_2\|v\|_2. ∫ Ω q a u r v r d x ≤ C δ a r − 4 ∥ u ∥ 2 ∥ v ∥ 2 .
Combining the last two estimates gives, for every 0 < r < r 1 0<r<r_1 0 < r < r 1 ,
∣ ∫ Ω q a u v ‾ d x ∣ ≤ C r ( ∥ u ∥ H 1 ∥ v ∥ 2 + ∥ u ∥ 2 ∥ v ∥ H 1 ) + C δ a r − 4 ∥ u ∥ 2 ∥ v ∥ 2 . (39) \left|\int_\Omega q_au\overline{v}\,\mathrm{d}x\right|
\le Cr\left(\|u\|_{H^1}\|v\|_2+\|u\|_2\|v\|_{H^1}\right)+C\delta_a r^{-4}\|u\|_2\|v\|_2. \tag*{(39)} ∫ Ω q a u v d x ≤ C r ( ∥ u ∥ H 1 ∥ v ∥ 2 + ∥ u ∥ 2 ∥ v ∥ H 1 ) + C δ a r − 4 ∥ u ∥ 2 ∥ v ∥ 2 . (39)
For small a a a , choose r = δ a 1 / 5 < r 1 r=\delta_a^{1/5}<r_1 r = δ a 1/5 < r 1 . Then δ a r − 4 = r = δ a 1 / 5 = ϵ a \delta_a r^{-4}=r=\delta_a^{1/5}=\epsilon_a δ a r − 4 = r = δ a 1/5 = ϵ a , and ∥ u ∥ 2 ∥ v ∥ 2 \|u\|_2\|v\|_2 ∥ u ∥ 2 ∥ v ∥ 2 is bounded by either mixed norm appearing on the right-hand side of (35) . Thus (39) gives (35) after decreasing the asymptotic threshold a 0 a_0 a 0 if necessary.
Coercivity and comparison of the auxiliary fields
For P ≥ 1 P \ge1 P ≥ 1 , use the norm
∥ u ∥ P 2 = ∥ ∇ u ∥ L 2 ( Ω ) 2 + P 2 ∥ u ∥ L 2 ( Ω ) 2 . (40) \lVert u\rVert_{P}^{2}=\lVert\nabla u\rVert_{L^{2}(\Omega)}^{2}+P^{2}\lVert u\rVert_{L^{2}(\Omega)}^{2}. \tag*{(40)} ∥ u ∥ P 2 = ∥ ∇ u ∥ L 2 ( Ω ) 2 + P 2 ∥ u ∥ L 2 ( Ω ) 2 . (40)
Let q ∗ = sup ω ∈ I ∥ q ω ∥ ∞ q_{*}=\sup_{\omega\in I}\lVert q_{\omega}\rVert_{\infty} q ∗ = sup ω ∈ I ∥ q ω ∥ ∞ . The two forms under consideration are
a ρ ( u , v ) = ∫ Ω ∇ u ⋅ ∇ v ‾ + ( P 2 ρ − q ω ) u v ‾ , a_{\rho}(u,v)=\int_{\Omega}\nabla u\cdot\nabla\overline{v}+\left(P^{2}\rho-q_{\omega}\right)u\overline{v}, a ρ ( u , v ) = ∫ Ω ∇ u ⋅ ∇ v + ( P 2 ρ − q ω ) u v ,
a a ( u , v ) = ∫ Ω ∇ u ⋅ ∇ v ‾ + ( P 2 ρ a a v − q ω ) u v ‾ . a_{a}(u,v)=\int_{\Omega}\nabla u\cdot\nabla\overline{v}+\left(P^{2}\rho_{a}^{\mathrm{av}}-q_{\omega}\right)u\overline{v}. a a ( u , v ) = ∫ Ω ∇ u ⋅ ∇ v + ( P 2 ρ a av − q ω ) u v .
Lemma 3.4. There are P 0 ≥ 1 P_{0}\ge1 P 0 ≥ 1 and c > 0 c>0 c > 0 such that, if P ≥ P 0 P\ge P_{0} P ≥ P 0 and P ϵ a ≤ c a d m P\epsilon_{a}\le c_{\mathrm{adm}} P ϵ a ≤ c adm with c a d m c_{\mathrm{adm}} c adm sufficiently small, then
a ρ ( u , u ) ≥ c ∥ u ∥ P 2 , a a ( u , u ) ≥ c ∥ u ∥ P 2 . (41) a_{\rho}(u,u)\ge c\lVert u\rVert_{P}^{2},\qquad a_{a}(u,u)\ge c\lVert u\rVert_{P}^{2}. \tag*{(41)} a ρ ( u , u ) ≥ c ∥ u ∥ P 2 , a a ( u , u ) ≥ c ∥ u ∥ P 2 . (41)
Let u a a v , f u_{a}^{\mathrm{av},f} u a av , f solve
( − Δ + P 2 ρ a a v − q ω ) u a a v , f = 0 , ∂ ν u a a v , f = f , Λ a a v f = γ u a a v , f . (42) \left(-\Delta+P^{2}\rho_{a}^{\mathrm{av}}-q_{\omega}\right)u_{a}^{\mathrm{av},f}=0,\qquad\partial_{\nu}u_{a}^{\mathrm{av},f}=f,\qquad\Lambda_{a}^{\mathrm{av}}f=\gamma u_{a}^{\mathrm{av},f}. \tag*{(42)} ( − Δ + P 2 ρ a av − q ω ) u a av , f = 0 , ∂ ν u a av , f = f , Λ a av f = γ u a av , f . (42)
Then both this problem and (12) are uniquely solvable, and
∥ u a a v , f ∥ P + ∥ u ρ , P f ∥ P ≤ C ∥ f ∥ X , (43) \lVert u_{a}^{\mathrm{av},f}\rVert_{P}+\lVert u_{\rho,P}^{f}\rVert_{P}\le C\lVert f\rVert_{\mathcal{X}}, \tag*{(43)} ∥ u a av , f ∥ P + ∥ u ρ , P f ∥ P ≤ C ∥ f ∥ X , (43)
∥ Λ a a v − Λ ρ , P e f f ∥ L ( X , Y ) ≤ C P ϵ a . (44) \lVert\Lambda_{a}^{\mathrm{av}}-\Lambda_{\rho,P}^{\mathrm{eff}}\rVert_{\mathcal{L}(\mathcal{X},\mathcal{Y})}\le CP\epsilon_{a}. \tag*{(44)} ∥ Λ a av − Λ ρ , P eff ∥ L ( X , Y ) ≤ C P ϵ a . (44)
Proof. Take P 0 2 ρ ∗ ≥ 2 q ∗ P_{0}^{2}\rho_{*}\ge2q_{*} P 0 2 ρ ∗ ≥ 2 q ∗ . Then
a ρ ( u , u ) ≥ ∥ ∇ u ∥ 2 + 1 2 ρ ∗ P 2 ∥ u ∥ 2 ≥ c 1 ∥ u ∥ P 2 , c 1 = min { 1 , ρ ∗ / 2 } . a_{\rho}(u,u)\ge\lVert\nabla u\rVert^{2}+\frac{1}{2}\rho_{*}P^{2}\lVert u\rVert^{2}\ge c_{1}\lVert u\rVert_{P}^{2},\qquad c_{1}=\min\{1,\rho_{*}/2\}. a ρ ( u , u ) ≥ ∥ ∇ u ∥ 2 + 2 1 ρ ∗ P 2 ∥ u ∥ 2 ≥ c 1 ∥ u ∥ P 2 , c 1 = min { 1 , ρ ∗ /2 } .
Lemma 3.3, together with ∥ u ∥ H 1 ≤ C ∥ u ∥ P \lVert u\rVert_{H^{1}}\le C\lVert u\rVert_{P} ∥ u ∥ H 1 ≤ C ∥ u ∥ P and ∥ u ∥ 2 ≤ P − 1 ∥ u ∥ P \lVert u\rVert_{2}\le P^{-1}\lVert u\rVert_{P} ∥ u ∥ 2 ≤ P − 1 ∥ u ∥ P , yields
∣ a a ( u , v ) − a ρ ( u , v ) ∣ ≤ C P ϵ a ∥ u ∥ P ∥ v ∥ P . (45) \left|a_{a}(u,v)-a_{\rho}(u,v)\right|\le CP\epsilon_{a}\lVert u\rVert_{P}\lVert v\rVert_{P}. \tag*{(45)} ∣ a a ( u , v ) − a ρ ( u , v ) ∣ ≤ C P ϵ a ∥ u ∥ P ∥ v ∥ P . (45)
Choosing C c a d m ≤ c 1 / 2 Cc_{\mathrm{adm}}\le c_{1}/2 C c adm ≤ c 1 /2 proves the second inequality in (41) .
The trace theorem gives ∣ ⟨ f , γ v ⟩ ∣ ≤ C ∥ f ∥ X ∥ v ∥ P \left|\langle f,\gamma v\rangle\right|\le C\lVert f\rVert_{\mathcal{X}}\lVert v\rVert_{P} ∣ ⟨ f , γ v ⟩ ∣ ≤ C ∥ f ∥ X ∥ v ∥ P . Lax–Milgram therefore gives both solutions and (43) . Subtract their weak equations and put e = u a a v , f − u ρ , P f e=u_{a}^{\mathrm{av},f}-u_{\rho,P}^{f} e = u a av , f − u ρ , P f . For v ∈ H 1 ( Ω ) v\in H^{1}(\Omega) v ∈ H 1 ( Ω ) ,
a a ( e , v ) = P 2 ∫ Ω ( ρ − ρ a a v ) u ρ , P f v ‾ . a_{a}(e,v)=P^{2}\int_{\Omega}\left(\rho-\rho_{a}^{\mathrm{av}}\right)u_{\rho,P}^{f}\overline{v}. a a ( e , v ) = P 2 ∫ Ω ( ρ − ρ a av ) u ρ , P f v .
By (45) , the right-hand side is at most C P ϵ a ∥ u ρ , P f ∥ P ∥ v ∥ P CP\epsilon_{a}\lVert u_{\rho,P}^{f}\rVert_{P}\lVert v\rVert_{P} C P ϵ a ∥ u ρ , P f ∥ P ∥ v ∥ P in modulus. Take v = e v=e v = e , use coercivity and (43) , and then apply the trace theorem. This gives (44) . □ \square □
Invertibility of the algebraic system
We prove (17) by comparing the matrix with an integral operator associated with (42) . This avoids a smallness assumption on the full interaction matrix.
The continuous inverse
Let
H a = L 2 ( Ω , ν a ) , ( T a F ) ( x ) = ∫ Ω G ( x , y ) F ( y ) d ν a ( y ) . (46) \mathcal{H}_{a}=L^{2}(\Omega,\nu_{a}),\qquad(T_{a}F)(x)=\int_{\Omega}G(x,y)F(y)\,d\nu_{a}(y). \tag*{(46)} H a = L 2 ( Ω , ν a ) , ( T a F ) ( x ) = ∫ Ω G ( x , y ) F ( y ) d ν a ( y ) . (46)
The space H a \mathcal{H}_{a} H a identifies functions which agree on the union of the balls B d j B_{d}^{j} B d j . Since ρ a a v \rho_{a}^{\mathrm{av}} ρ a av is bounded and sup x ∫ Ω ∣ x − y ∣ − 2 d y < ∞ \sup_{x}\int_{\Omega}|x-y|^{-2}\,dy<\infty sup x ∫ Ω ∣ x − y ∣ − 2 d y < ∞ , the kernel belongs to L 2 ( ν a ⊗ ν a ) L^{2}(\nu_{a}\otimes\nu_{a}) L 2 ( ν a ⊗ ν a ) . Hence T a T_{a} T a is bounded, compact and self-adjoint on H a \mathcal{H}_{a} H a .
Lemma 4.1. Under (14) , after fixing P 0 P_0 P 0 and c a d m c_{\mathrm{adm}} c adm as in Lemma 3.4 ,
∥ ( I + t T a ) − 1 ∥ H a → H a ≤ C , t = P 2 . (47) \left\|(I+tT_a)^{-1}\right\|_{\mathcal{H}_a\to\mathcal{H}_a}\le C,\qquad t=P^2. \tag*{(47)} ( I + t T a ) − 1 H a → H a ≤ C , t = P 2 . (47) Proof. For F ∈ H a F\in\mathcal{H}_a F ∈ H a , solve the zero-Neumann problem
( A 0 , ω + t ρ a a v ) v = ρ a a v F . (48) (A_{0,\omega}+t\rho_a^{\mathrm{av}})v=\rho_a^{\mathrm{av}}F. \tag*{(48)} ( A 0 , ω + t ρ a av ) v = ρ a av F . (48) The right-hand side is in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) , because
∥ ρ a a v F ∥ L 2 ( Ω ) 2 ≤ ∥ ρ a a v ∥ ∞ ∥ F ∥ H a 2 . \left\|\rho_a^{\mathrm{av}}F\right\|_{L^2(\Omega)}^2\le\left\|\rho_a^{\mathrm{av}}\right\|_\infty\|F\|_{\mathcal{H}_a}^2. ∥ ρ a av F ∥ L 2 ( Ω ) 2 ≤ ∥ ρ a av ∥ ∞ ∥ F ∥ H a 2 . By coercivity and Cauchy–Schwarz,
c ∥ v ∥ P 2 ≤ ∣ ∫ F v ‾ d ν a ∣ ≤ ∥ F ∥ H a ∥ v ∥ H a ≤ C P − 1 ∥ F ∥ H a ∥ v ∥ P . c\|v\|_P^2\le\left|\int F\overline{v}\,d\nu_a\right|\le\|F\|_{\mathcal{H}_a}\|v\|_{\mathcal{H}_a}\le CP^{-1}\|F\|_{\mathcal{H}_a}\|v\|_P. c ∥ v ∥ P 2 ≤ ∫ F v d ν a ≤ ∥ F ∥ H a ∥ v ∥ H a ≤ C P − 1 ∥ F ∥ H a ∥ v ∥ P . Consequently
∥ v ∥ P ≤ C P − 1 ∥ F ∥ H a , t ∥ v ∥ H a ≤ C ∥ F ∥ H a . (49) \|v\|_P\le CP^{-1}\|F\|_{\mathcal{H}_a},\qquad t\|v\|_{\mathcal{H}_a}\le C\|F\|_{\mathcal{H}_a}. \tag*{(49)} ∥ v ∥ P ≤ C P − 1 ∥ F ∥ H a , t ∥ v ∥ H a ≤ C ∥ F ∥ H a . (49) Applying A 0 , ω − 1 A_{0,\omega}^{-1} A 0 , ω − 1 to (48) gives, on the support of ν a \nu_a ν a ,
v = T a ( F − t v ) , ( I + t T a ) v = T a F . v=T_a(F-tv),\qquad(I+tT_a)v=T_aF. v = T a ( F − t v ) , ( I + t T a ) v = T a F . It follows that U = F − t v U=F-tv U = F − t v satisfies
( I + t T a ) U = F + t T a F − t ( v + t T a v ) = F . (I+tT_a)U=F+tT_aF-t(v+tT_av)=F. ( I + t T a ) U = F + t T a F − t ( v + t T a v ) = F . The estimate (49) yields ∥ U ∥ H a ≤ C ∥ F ∥ H a \|U\|_{\mathcal{H}_a}\le C\|F\|_{\mathcal{H}_a} ∥ U ∥ H a ≤ C ∥ F ∥ H a .
To verify uniqueness, suppose U + t T a U = 0 U+tT_aU=0 U + t T a U = 0 . Define the global function v = A 0 , ω − 1 ( ρ a a v U ) v=A_{0,\omega}^{-1}(\rho_a^{\mathrm{av}}U) v = A 0 , ω − 1 ( ρ a av U ) . On the support of ν a \nu_a ν a , v = T a U v=T_aU v = T a U and U = − t v U=-tv U = − t v . Thus ( A 0 , ω + t ρ a a v ) v = 0 (A_{0,\omega}+t\rho_a^{\mathrm{av}})v=0 ( A 0 , ω + t ρ a av ) v = 0 with zero Neumann data. Coercivity gives v = 0 v=0 v = 0 and then U = 0 U=0 U = 0 . This proves (47) .
Comparison with the point-interaction operator
Define the constant extension and averaging operators by
( J Y ) ∣ B i d = Y i , ( Q F ) i = 1 w ∫ B i d F d ν a = 1 v d ∫ B i d F d x . (50) (JY)|_{B_i^d}=Y_i,\qquad(QF)_i=\frac{1}{w}\int_{B_i^d}F\,d\nu_a=\frac{1}{v_d}\int_{B_i^d}F\,dx. \tag*{(50)} ( J Y ) ∣ B i d = Y i , ( QF ) i = w 1 ∫ B i d F d ν a = v d 1 ∫ B i d F d x . (50)
Then J : ℓ w 2 → H a J:\ell_w^2\to\mathcal{H}_a J : ℓ w 2 → H a is an isometry, Q = J ∗ Q=J^* Q = J ∗ , and Q J = I QJ=I Q J = I . In particular, J Q JQ J Q is the orthogonal projection onto the functions which are constant on each ball. Put
( K a Y ) i = w ∑ j ≠ i G ( z i , z j ) Y j , K a = J K a Q . (K_aY)_i=w\sum_{j\ne i}G(z_i,z_j)Y_j,\qquad K_a=JK_aQ. ( K a Y ) i = w j = i ∑ G ( z i , z j ) Y j , K a = J K a Q .
As an operator on H a \mathcal{H}_a H a , K a K_a K a has kernel G ( z i , z j ) G(z_i,z_j) G ( z i , z j ) on B i d × B j d B_i^d\times B_j^d B i d × B j d when i ≠ j i\ne j i = j , and zero on the diagonal blocks.
Lemma 4.2. One has
∥ T a − K a ∥ H a → H a ≤ C d . (51) \|T_a-K_a\|_{\mathcal{H}_a\to\mathcal{H}_a}\le Cd. \tag*{(51)} ∥ T a − K a ∥ H a → H a ≤ C d . (51) Consequently (17) holds.
Proof. For x ∈ B i d x \in B_i^d x ∈ B i d , the integral of the absolute value of the kernel on the diagonal block is bounded by
∫ B i d ∣ G ( x , y ) ∣ d ν a ( y ) ≤ C ∫ B ( x , 2 b 0 d ) ∣ x − y ∣ − 1 d y ≤ C d 2 . \int_{B_i^d} |G(x,y)|\,d\nu_a(y) \le C\int_{B(x,2b_0d)} |x-y|^{-1}\,dy \le Cd^2. ∫ B i d ∣ G ( x , y ) ∣ d ν a ( y ) ≤ C ∫ B ( x , 2 b 0 d ) ∣ x − y ∣ − 1 d y ≤ C d 2 .
If x ∈ B i d x \in B_i^d x ∈ B i d , y ∈ B j d y \in B_j^d y ∈ B j d and i ≠ j i \ne j i = j , our choice of b 0 b_0 b 0 ensures that every distance arising while moving x x x to z i z_i z i and y y y to z j z_j z j is comparable to ∣ z i − z j ∣ |z_i-z_j| ∣ z i − z j ∣ . The line segments lie inside the corresponding balls. The gradient estimate (4) therefore gives
∣ G ( x , y ) − G ( z i , z j ) ∣ ≤ C d ∣ z i − z j ∣ − 2 . |G(x,y)-G(z_i,z_j)| \le Cd|z_i-z_j|^{-2}. ∣ G ( x , y ) − G ( z i , z j ) ∣ ≤ C d ∣ z i − z j ∣ − 2 .
Integration in y y y and (30) , with s = 2 s=2 s = 2 , yield
sup x ∫ ∣ G ( x , y ) − K a ( x , y ) ∣ d ν a ( y ) ≤ C ( d 2 + d ) ≤ C d . \sup_x \int|G(x,y)-K_a(x,y)|\,d\nu_a(y) \le C(d^2+d) \le Cd. x sup ∫ ∣ G ( x , y ) − K a ( x , y ) ∣ d ν a ( y ) ≤ C ( d 2 + d ) ≤ C d .
The same bound holds after interchanging x x x and y y y . We spell out the Schur argument. Let L a ( x , y ) L_a(x,y) L a ( x , y ) be the kernel of T a − K a T_a-K_a T a − K a with respect to ν a \nu_a ν a . The preceding estimates say that
A a : = sup x ∫ Ω ∣ L a ( x , y ) ∣ d ν a ( y ) ≤ C d , B a : = sup y ∫ Ω ∣ L a ( x , y ) ∣ d ν a ( x ) ≤ C d . (52) A_a:=\sup_x\int_\Omega|L_a(x,y)|\,d\nu_a(y)\le Cd,\qquad B_a:=\sup_y\int_\Omega|L_a(x,y)|\,d\nu_a(x)\le Cd. \tag*{(52)} A a := x sup ∫ Ω ∣ L a ( x , y ) ∣ d ν a ( y ) ≤ C d , B a := y sup ∫ Ω ∣ L a ( x , y ) ∣ d ν a ( x ) ≤ C d . (52)
For completeness, the off-diagonal part of the first inequality is
∑ j ≠ i ∫ B j d ∣ G ( x , y ) − G ( z i , z j ) ∣ d ν a ( y ) ≤ C d w ∑ j ≠ i ∣ z i − z j ∣ − 2 ≤ C d , \sum_{j\ne i}\int_{B_j^d}|G(x,y)-G(z_i,z_j)|\,d\nu_a(y)\le Cd\,w\sum_{j\ne i}|z_i-z_j|^{-2}\le Cd, j = i ∑ ∫ B j d ∣ G ( x , y ) − G ( z i , z j ) ∣ d ν a ( y ) ≤ C d w j = i ∑ ∣ z i − z j ∣ − 2 ≤ C d ,
for x ∈ B i d x \in B_i^d x ∈ B i d , while the diagonal block contributes only C d 2 Cd^2 C d 2 . The second inequality follows in the same way (or from the symmetry of G G G and of the block kernel).
If F ∈ H a F \in\mathcal{H}_a F ∈ H a , Cauchy–Schwarz with the measure ∣ L a ( x , y ) ∣ d ν a ( y ) |L_a(x,y)|\,d\nu_a(y) ∣ L a ( x , y ) ∣ d ν a ( y ) gives, for a.e. x x x ,
∣ ( T a − K a ) F ( x ) ∣ 2 ≤ ( ∫ ∣ L a ( x , y ) ∣ d ν a ( y ) ) ( ∫ ∣ L a ( x , y ) ∣ ∣ F ( y ) ∣ 2 d ν a ( y ) ) . |(T_a-K_a)F(x)|^2\le\left(\int|L_a(x,y)|\,d\nu_a(y)\right)\left(\int|L_a(x,y)||F(y)|^2\,d\nu_a(y)\right). ∣ ( T a − K a ) F ( x ) ∣ 2 ≤ ( ∫ ∣ L a ( x , y ) ∣ d ν a ( y ) ) ( ∫ ∣ L a ( x , y ) ∣∣ F ( y ) ∣ 2 d ν a ( y ) ) .
Integrating in x x x , using Tonelli’s theorem and (52) , we obtain
∥ ( T a − K a ) F ∥ H a 2 ≤ A a ∫ Ω ∣ F ( y ) ∣ 2 ( ∫ Ω ∣ L a ( x , y ) ∣ d ν a ( x ) ) d ν a ( y ) ≤ A a B a ∥ F ∥ H a 2 ≤ C d 2 ∥ F ∥ H a 2 . \begin{aligned}
\|(T_a-K_a)F\|_{\mathcal{H}_a}^2\le A_a\int_\Omega|F(y)|^2\left(\int_\Omega|L_a(x,y)|\,d\nu_a(x)\right)d\nu_a(y) \\
&\le A_aB_a\|F\|_{\mathcal{H}_a}^2\le Cd^2\|F\|_{\mathcal{H}_a}^2.
\end{aligned} ∥ ( T a − K a ) F ∥ H a 2 ≤ A a ∫ Ω ∣ F ( y ) ∣ 2 ( ∫ Ω ∣ L a ( x , y ) ∣ d ν a ( x ) ) d ν a ( y ) ≤ A a B a ∥ F ∥ H a 2 ≤ C d 2 ∥ F ∥ H a 2 .
Thus ∥ T a − K a ∥ ≤ A a B a ≤ C d \|T_a-K_a\|\le\sqrt{A_aB_a}\le Cd ∥ T a − K a ∥ ≤ A a B a ≤ C d , which is (51) .
Now factor
I + t K a = ( I + t T a ) [ I + t ( I + t T a ) − 1 ( K a − T a ) ] . (53) I+tK_a=(I+tT_a)\left[I+t(I+tT_a)^{-1}(K_a-T_a)\right]. \tag*{(53)} I + t K a = ( I + t T a ) [ I + t ( I + t T a ) − 1 ( K a − T a ) ] . (53)
The norm of the perturbation in brackets is at most C t d Ctd C t d . For d ≤ 1 d\le1 d ≤ 1 , we have
d ≤ ϵ a 2 , t d ≤ ( P ϵ a ) 2 ≤ ϵ a d m 2 . (54) d\le\epsilon_a^2,\qquad td\le(P\epsilon_a)^2\le\epsilon_{\mathrm{adm}}^2. \tag*{(54)} d ≤ ϵ a 2 , t d ≤ ( P ϵ a ) 2 ≤ ϵ adm 2 . (54)
Indeed ϵ a 2 = ( D a + d ) 2 / 5 ≥ d 2 / 5 ≥ d \epsilon_a^2=(D_a+d)^{2/5}\ge d^{2/5}\ge d ϵ a 2 = ( D a + d ) 2/5 ≥ d 2/5 ≥ d . Choose c a d m c_{\mathrm{adm}} c adm so that C t d ≤ 1 / 2 Ctd\le1/2 C t d ≤ 1/2 . A Neumann series in (53) , together with Lemma 4.1 , gives a uniform inverse of I + t K a I+tK_a I + t K a .
It remains to pass from the inverse on H a \mathcal{H}_a H a to the inverse of the finite-dimensional Foldy–Lax matrix. Set
V a : = Ran J , Π a : = J Q . \mathcal{V}_a:=\operatorname{Ran}J,\qquad\Pi_a:=JQ. V a := Ran J , Π a := J Q .
Since Q = J ∗ Q=J^* Q = J ∗ and Q J = I QJ=I Q J = I , Π a \Pi_a Π a is the orthogonal projection of H a \mathcal{H}_a H a onto V a \mathcal{V}_a V a . Moreover,
K a = J K a Q = Π a K a Π a . K_a=JK_aQ=\Pi_aK_a\Pi_a. K a = J K a Q = Π a K a Π a .
In particular,
K a ( V a ) ⊂ V a , K a ∣ V a ⊥ = 0 , ( I − Π a ) K a = K a ( I − Π a ) = 0. (55) K_a(\mathcal{V}_a) \subset\mathcal{V}_a,\qquad K_a|_{\mathcal{V}_a^\perp}=0,\qquad(I-\Pi_a)K_a=K_a(I-\Pi_a)=0. \tag*{(55)} K a ( V a ) ⊂ V a , K a ∣ V a ⊥ = 0 , ( I − Π a ) K a = K a ( I − Π a ) = 0. (55)
Thus, with respect to the orthogonal decomposition H a = V a ⊕ V a ⊥ \mathcal{H}_a=\mathcal{V}_a\oplus\mathcal{V}_a^\perp H a = V a ⊕ V a ⊥ , the operator has the block form
I + t K a = ( ( I + t K a ) ∣ V a 0 0 I ) . I+tK_a=\begin{pmatrix}(I+tK_a)|_{\mathcal{V}_a} & 0 \\ 0 & I\end{pmatrix}. I + t K a = ( ( I + t K a ) ∣ V a 0 0 I ) .
This already shows that its inverse preserves V a \mathcal{V}_a V a . Equivalently, if ( I + t K a ) U ∈ V a (I+tK_a)U\in\mathcal{V}_a ( I + t K a ) U ∈ V a , then applying I − Π a I-\Pi_a I − Π a and using (55) gives
0 = ( I − Π a ) ( I + t K a ) U = ( I − Π a ) U , 0=(I-\Pi_a)(I+tK_a)U=(I-\Pi_a)U, 0 = ( I − Π a ) ( I + t K a ) U = ( I − Π a ) U ,
so U ∈ V a U\in\mathcal{V}_a U ∈ V a .
For Y ∈ ℓ w 2 Y\in\ell_w^2 Y ∈ ℓ w 2 , the definitions give the exact intertwining identity
( I + t K a ) J Y = J Y + t J K a Q J Y = J ( I + t K a ) Y = J B a , P Y . (56) (I+tK_a)JY=JY+tJK_aQJY=J(I+tK_a)Y=J B_{a,P}Y. \tag*{(56)} ( I + t K a ) J Y = J Y + t J K a Q J Y = J ( I + t K a ) Y = J B a , P Y . (56)
Because J J J is an isometric isomorphism from ℓ w 2 \ell_w^2 ℓ w 2 onto V a \mathcal{V}_a V a , invertibility of ( I + t K a ) ∣ V a (I+tK_a)|_{\mathcal{V}_a} ( I + t K a ) ∣ V a is equivalent to invertibility of B a , P B_{a,P} B a , P , and
J B a , P − 1 = ( I + t K a ) − 1 J . J B_{a,P}^{-1}=(I+tK_a)^{-1}J. J B a , P − 1 = ( I + t K a ) − 1 J .
Consequently, for every Y ∈ ℓ w 2 Y\in\ell_w^2 Y ∈ ℓ w 2 ,
∥ B a , P − 1 Y ∥ ℓ w 2 = ∥ ( I + t K a ) − 1 J Y ∥ H a ≤ C ∥ J Y ∥ H a = C ∥ Y ∥ ℓ w 2 . \left\|B_{a,P}^{-1}Y\right\|_{\ell_w^2}=\left\|(I+tK_a)^{-1}JY\right\|_{\mathcal{H}_a}\leq C\|JY\|_{\mathcal{H}_a}=C\|Y\|_{\ell_w^2}. B a , P − 1 Y ℓ w 2 = ( I + t K a ) − 1 J Y H a ≤ C ∥ J Y ∥ H a = C ∥ Y ∥ ℓ w 2 .
This proves (17) with exactly the weighted norm used in the statement. □
Sampling the background solutions
We shall use point values only for solutions of the background equation, not for arbitrary H 1 H^1 H 1 functions.
Lemma 4.3. Let s i f = S f ( z i ) s_i^f=S^f(z_i) s i f = S f ( z i ) and S ^ i f ( ξ ) = S f ( z i + a ξ ) \widehat{S}_i^f(\xi)=S^f(z_i+a\xi) S i f ( ξ ) = S f ( z i + a ξ ) for ξ ∈ B \xi\in B ξ ∈ B . For small a a a ,
∥ S f ∥ ℓ w 2 + ( w ∑ i sup B i d ∣ ∇ S f ∣ 2 ) 1 / 2 ≤ C ∥ f ∥ X , (57) \|S^f\|_{\ell_w^2}+\left(w\sum_i\sup_{B_i^d}|\nabla S^f|^2\right)^{1/2}\leq C\|f\|_{\mathcal{X}}, \tag*{(57)} ∥ S f ∥ ℓ w 2 + ( w i ∑ B i d sup ∣∇ S f ∣ 2 ) 1/2 ≤ C ∥ f ∥ X , (57) ∥ S f − J s f ∥ H a ≤ C d ∥ f ∥ X , (58) \|S^f-Js^f\|_{\mathcal{H}_a}\leq Cd\|f\|_{\mathcal{X}}, \tag*{(58)} ∥ S f − J s f ∥ H a ≤ C d ∥ f ∥ X , (58) ( w ∑ i ∥ S ^ i f − s i f ∥ L 2 ( B ) 2 ) 1 / 2 ≤ C a ∥ f ∥ X . (59) \left(w\sum_i\|\widehat{S}_i^f-s_i^f\|_{L^2(B)}^2\right)^{1/2}\leq Ca\|f\|_{\mathcal{X}}. \tag*{(59)} ( w i ∑ ∥ S i f − s i f ∥ L 2 ( B ) 2 ) 1/2 ≤ C a ∥ f ∥ X . (59)
Proof. Write r = b 0 d r=b_0d r = b 0 d . Interior estimates for Δ S f + q ω S f = 0 \Delta S^f+q_\omega S^f=0 Δ S f + q ω S f = 0 on B ( z i , 4 r ) B(z_i,4r) B ( z i , 4 r ) give
∣ S f ( z i ) ∣ ≤ C r − 3 / 2 ∥ S f ∥ L 2 ( B ( z i , 2 r ) ) , (60) |S^f(z_i)|\leq Cr^{-3/2}\|S^f\|_{L^2(B(z_i,2r))}, \tag*{(60)} ∣ S f ( z i ) ∣ ≤ C r − 3/2 ∥ S f ∥ L 2 ( B ( z i , 2 r )) , (60)
sup B ( z i , r ) ∣ ∇ S f ∣ ≤ C r − 3 / 2 ∥ S f ∥ H 1 ( B ( z i , 4 r ) ) . (61) \sup_{B(z_i,r)}|\nabla S^f|\leq Cr^{-3/2}\|S^f\|_{H^1(B(z_i,4r))}. \tag*{(61)} B ( z i , r ) sup ∣∇ S f ∣ ≤ C r − 3/2 ∥ S f ∥ H 1 ( B ( z i , 4 r )) . (61)
We recall why no additional inverse power of r r r is needed in the second estimate when the full H 1 H^1 H 1 norm is used. Differentiate the equation in the weak sense:
( Δ + q ω ) ∂ k S f = − ( ∂ k q ω ) S f . (\Delta+q_\omega)\partial_kS^f=-(\partial_kq_\omega)S^f. ( Δ + q ω ) ∂ k S f = − ( ∂ k q ω ) S f .
The scaled interior L 2 L^2 L 2 -to-L ∞ L^\infty L ∞ estimate for this equation, followed by the value estimate for S f S^f S f on a larger ball, yields
sup B ( z i , r ) ∣ ∂ k S f ∣ ≤ C r − 3 / 2 ∥ ∂ k S f ∥ L 2 ( B ( z i , 2 r ) ) + C r 2 ∥ ∇ q ω ∥ ∞ sup B ( z i , 2 r ) ∣ S f ∣ ≤ C r − 3 / 2 ∥ S f ∥ H 1 ( B ( z i , 4 r ) ) . \sup_{B(z_i,r)} |\partial_k S^f| \le Cr^{-3/2}\|\partial_k S^f\|_{L^2(B(z_i,2r))}+Cr^2\|\nabla q_\omega\|_\infty\sup_{B(z_i,2r)}|S^f| \le Cr^{-3/2}\|S^f\|_{H^1(B(z_i,4r))}. B ( z i , r ) sup ∣ ∂ k S f ∣ ≤ C r − 3/2 ∥ ∂ k S f ∥ L 2 ( B ( z i , 2 r )) + C r 2 ∥∇ q ω ∥ ∞ B ( z i , 2 r ) sup ∣ S f ∣ ≤ C r − 3/2 ∥ S f ∥ H 1 ( B ( z i , 4 r )) .
The constants are uniform since r ≤ r 1 r\le r_1 r ≤ r 1 and q ω q_\omega q ω has a common W 1 , ∞ W^{1,\infty} W 1 , ∞ bound. These estimates can first be proved for smooth coefficients and then passed to the limit; alternatively, interior W 2 , p W^{2,p} W 2 , p estimates give the same conclusion.
The balls B ( z i , 4 r ) B(z_i,4r) B ( z i , 4 r ) are disjoint, and w r − 3 wr^{-3} w r − 3 is a fixed constant. Square (60) and (61) , multiply by w w w , and sum in i i i . The background bound ∥ S f ∥ H 1 ( Ω ) ≤ C ∥ f ∥ X \|S^f\|_{H^1(\Omega)}\le C\|f\|_X ∥ S f ∥ H 1 ( Ω ) ≤ C ∥ f ∥ X proves (57) . On B i d B_i^d B i d ,
∣ S f ( x ) − S f ( z i ) ∣ ≤ r sup B i d ∣ ∇ S f ∣ . |S^f(x)-S^f(z_i)|\le r\sup_{B_i^d}|\nabla S^f|. ∣ S f ( x ) − S f ( z i ) ∣ ≤ r B i d sup ∣∇ S f ∣.
Integrating against ν a \nu_a ν a and summing proves (58) . For small a a a , the set z i + a B z_i+aB z i + a B and all segments from z i z_i z i to its points lie in B i d B_i^d B i d . The same argument, with ∣ a ξ ∣ ≤ a sup B ∣ ξ ∣ |a\xi|\le a\sup_B|\xi| ∣ a ξ ∣ ≤ a sup B ∣ ξ ∣ , proves (59) . Convexity of the droplet is not required.
Derivation of the Point-Interaction approximation
We next relate the exact field to the algebraic system of the preceding section. The normalization of the unknowns is important: it removes the local resonant factor from the remainder in the algebraic equation.
The local integral equations
By (3) , the exact equation (11) is equivalent to
u a f ( x ) = S f ( x ) + ∫ D G ( x , y ) ( κ a − q ω ( y ) ) u a f ( y ) d y , x ∈ Ω . (62) u_a^f(x)=S^f(x)+\int_D G(x,y)(\kappa_a-q_\omega(y))u_a^f(y)\,dy,\qquad x\in\Omega. \tag*{(62)} u a f ( x ) = S f ( x ) + ∫ D G ( x , y ) ( κ a − q ω ( y )) u a f ( y ) d y , x ∈ Ω. (62)
For the moment, suppose that a solution exists. We shall prove an a priori estimate and then obtain existence by Fredholm theory. Write
u ^ i ( ξ ) = u a f ( z i + a ξ ) , ∥ u ^ ∥ H a 2 = w ∑ i ∥ u ^ i ∥ L 2 ( B ) 2 , U = ∥ u ^ ∥ H a . \widehat{u}_i(\xi)=u_a^f(z_i+a\xi),\qquad\|\widehat{u}\|_{\mathcal{H}_a}^2=w\sum_i\|\widehat{u}_i\|_{L^2(B)}^2,\qquad U=\|\widehat{u}\|_{\mathcal{H}_a}. u i ( ξ ) = u a f ( z i + a ξ ) , ∥ u ∥ H a 2 = w i ∑ ∥ u i ∥ L 2 ( B ) 2 , U = ∥ u ∥ H a .
The physical and rescaled norms are related by
∥ u a f ∥ L 2 ( D ) = a 3 / 2 w − 1 / 2 U . (63) \|u_a^f\|_{L^2(D)}=a^{3/2}w^{-1/2}U. \tag*{(63)} ∥ u a f ∥ L 2 ( D ) = a 3/2 w − 1/2 U . (63)
Define the matrix G H G^H G H by
G i j H = { G ( z i , z j ) , i ≠ j , H ( z i , z i ) , i = j . G^H_{ij}=
\begin{cases}
G(z_i,z_j), & i\ne j,\\
H(z_i,z_i), & i=j.
\end{cases} G ij H = { G ( z i , z j ) , H ( z i , z i ) , i = j , i = j .
Restrict (62) to D i D_i D i , use G = Φ 0 + H G=\Phi_0+H G = Φ 0 + H on its self-interaction term, and change variables to B B B . This gives the exact identity
( I − θ N B ) u ^ i = 1 [ s i f + a θ ∑ j G i j H ∫ B u ^ j + F i ] . (64) (I-\theta N_B)\widehat{u}_i=\mathbf{1}\left[s_i^f+a\theta\sum_jG^H_{ij}\int_B\widehat{u}_j+F_i\right]. \tag*{(64)} ( I − θ N B ) u i = 1 [ s i f + a θ j ∑ G ij H ∫ B u j + F i ] . (64)
where F i = F i i n c + F i o f f + F i s e l f + F i b g F_i = F_i^{\mathrm{inc}} + F_i^{\mathrm{off}} + F_i^{\mathrm{self}} + F_i^{\mathrm{bg}} F i = F i inc + F i off + F i self + F i bg , with
F i i n c ( ξ ) = S ^ i f ( ξ ) − s i f , (65) F_i^{\mathrm{inc}}(\xi) = \widehat{S}_i^f(\xi) - s_i^f, \tag*{(65)} F i inc ( ξ ) = S i f ( ξ ) − s i f , (65)
F i o f f ( ξ ) = a θ ∑ j ≠ i ∫ B [ G ( z i + a ξ , z j + a η ) − G ( z i , z j ) ] u ^ j ( η ) d η , (66) F_i^{\mathrm{off}}(\xi) = a\theta\sum_{j\ne i}\int_B [G(z_i+a\xi,z_j+a\eta)-G(z_i,z_j)]\widehat{u}_j(\eta)\,d\eta, \tag*{(66)} F i off ( ξ ) = a θ j = i ∑ ∫ B [ G ( z i + a ξ , z j + a η ) − G ( z i , z j )] u j ( η ) d η , (66)
F i s e l f ( ξ ) = a θ ∫ B [ H ( z i + a ξ , z i + a η ) − H ( z i , z i ) ] u ^ i ( η ) d η , (67) F_i^{\mathrm{self}}(\xi) = a\theta\int_B [H(z_i+a\xi,z_i+a\eta)-H(z_i,z_i)]\widehat{u}_i(\eta)\,d\eta, \tag*{(67)} F i self ( ξ ) = a θ ∫ B [ H ( z i + a ξ , z i + a η ) − H ( z i , z i )] u i ( η ) d η , (67)
F i b g ( ξ ) = − a 3 ∑ j ∫ B G ( z i + a ξ , z j + a η ) q ω ( z j + a η ) u ^ j ( η ) d η . (68) F_i^{\mathrm{bg}}(\xi) = -a^3\sum_j\int_B G(z_i+a\xi,z_j+a\eta)q_\omega(z_j+a\eta)\widehat{u}_j(\eta)\,d\eta. \tag*{(68)} F i bg ( ξ ) = − a 3 j ∑ ∫ B G ( z i + a ξ , z j + a η ) q ω ( z j + a η ) u j ( η ) d η . (68)
No interaction between distinct droplets has been discarded in (64) .
Lemma 5.1. The remainder satisfies
∥ F ∥ H a ≤ C a ℓ ∥ f ∥ X + C a 1 + h U . (69) \|F\|_{H_a} \le Ca_\ell\|f\|_X + Ca^{1+h}U. \tag*{(69)} ∥ F ∥ H a ≤ C a ℓ ∥ f ∥ X + C a 1 + h U . (69)
Proof. The incident part is bounded by (59) . Put r i j = ∣ z i − z j ∣ r_{ij}=|z_i-z_j| r ij = ∣ z i − z j ∣ . Since a / d → 0 a/d\to0 a / d → 0 , all the off-diagonal distances in (66) are comparable to r i j r_{ij} r ij . Taylor’s formula and (4) give the block-operator bound
∥ F i o f f ∥ L 2 ( B ) ≤ C a 2 ∑ j ≠ i r i j − 2 ∥ u ^ j ∥ L 2 ( B ) . \|F_i^{\mathrm{off}}\|_{L^2(B)} \le Ca^2\sum_{j\ne i}r_{ij}^{-2}\|\widehat{u}_j\|_{L^2(B)}. ∥ F i off ∥ L 2 ( B ) ≤ C a 2 j = i ∑ r ij − 2 ∥ u j ∥ L 2 ( B ) .
By (30) , the row and column sums of this scalar majorant are at most C a 2 d − 3 Ca^2d^{-3} C a 2 d − 3 . The block version of Schur’s test proved in Lemma B.2 consequently gives
∥ F o f f ∥ H a ≤ C a 2 d − 3 U . \|F^{\mathrm{off}}\|_{H_a}\le Ca^2d^{-3}U. ∥ F off ∥ H a ≤ C a 2 d − 3 U .
The Lipschitz estimate for H H H , with b b b comparable to d d d , yields
∥ F s e l f ∥ H a ≤ C a 2 d − 2 U . \|F^{\mathrm{self}}\|_{H_a}\le Ca^2d^{-2}U. ∥ F self ∥ H a ≤ C a 2 d − 2 U .
For the off-diagonal part of (68) , the scalar majorant is C a 3 r i j − 1 Ca^3r_{ij}^{-1} C a 3 r ij − 1 ; its row and column sums are at most C a 3 d − 3 Ca^3d^{-3} C a 3 d − 3 . On the diagonal, the Newtonian part is a 2 N B a^2N_B a 2 N B followed by multiplication by the bounded function q ω ( z i + a ⋅ ) q_\omega(z_i+a\cdot) q ω ( z i + a ⋅ ) . The regular part has norm at most C a 3 d − 1 Ca^3d^{-1} C a 3 d − 1 . Hence
∥ F b g ∥ H a ≤ C ( a 3 d − 3 + a 2 + a 3 d − 1 ) U . \|F^{\mathrm{bg}}\|_{H_a}\le C(a^3d^{-3}+a^2+a^3d^{-1})U. ∥ F bg ∥ H a ≤ C ( a 3 d − 3 + a 2 + a 3 d − 1 ) U .
For a , d ≤ 1 a,d\le1 a , d ≤ 1 , each displayed coefficient is bounded by C a 2 d − 3 Ca^2d^{-3} C a 2 d − 3 . Since d 3 = a 1 − h d^3=a^{1-h} d 3 = a 1 − h , we have a 2 d − 3 = a 1 + h a^2d^{-3}=a^{1+h} a 2 d − 3 = a 1 + h . Adding the four estimates proves (69) .
Normalized amplitudes and the inverse estimate
Set
Y i = 1 m a ∫ B u ^ i = κ a α a ∫ D i u a f , r i = ℓ a ( F i ) . (70) Y_i=\frac{1}{m_a}\int_B\widehat{u}_i=\frac{\kappa_a}{\alpha_a}\int_{D_i}u_a^f,\qquad r_i=\ell_a(F_i). \tag*{(70)} Y i = m a 1 ∫ B u i = α a κ a ∫ D i u a f , r i = ℓ a ( F i ) . (70)
The equality of the two expressions for Y i Y_i Y i follows from α a = a θ m a \alpha_a=a\theta m_a α a = a θ m a and κ a = θ a − 2 \kappa_a=\theta a^{-2} κ a = θ a − 2 . Apply R a R_a R a to (64) and integrate on B B B . Division by m a m_a m a gives
( I − α a G H ) Y = s f + r , ∥ r ∥ ℓ w 2 ≤ C ∥ F ∥ H a . (71) (I-\alpha_aG^H)Y=s^f+r,\qquad\|r\|_{\ell_w^2}\le C\|F\|_{H_a}. \tag*{(71)} ( I − α a G H ) Y = s f + r , ∥ r ∥ ℓ w 2 ≤ C ∥ F ∥ H a . (71)
The uniform bound on ℓ a \ell_a ℓ a in Lemma 2.1 is used in the last inequality.
Lemma 5.2. Under (14) , after reducing a 0 a_0 a 0 and c a d m c_{\mathrm{adm}} c adm if necessary,
∥ ( I − α a G H ) − 1 ∥ ℓ w 2 → ℓ w 2 ≤ C . (72) \left\lVert(I-\alpha_aG^H)^{-1}\right\rVert_{\ell^2_w\to\ell^2_w}\le C. \tag*{(72)} ( I − α a G H ) − 1 ℓ w 2 → ℓ w 2 ≤ C . (72) Moreover, with β i = 1 − α a H ( z i , z i ) \beta_i=1-\alpha_aH(z_i,z_i) β i = 1 − α a H ( z i , z i ) ,
max i ∣ β i − 1 ∣ ≤ C t d 2 . (73) \max_i|\beta_i-1|\le Ctd^2. \tag*{(73)} i max ∣ β i − 1∣ ≤ C t d 2 . (73)
Proof. The off-diagonal difference between I − α a G H I-\alpha_aG^H I − α a G H and B a , P B_{a,P} B a , P is − ( α a + t w ) G ( z i , z j ) -(\alpha_a+tw)G(z_i,z_j) − ( α a + tw ) G ( z i , z j ) . Since ∣ α a + t w ∣ ≤ C a |\alpha_a+tw|\le Ca ∣ α a + tw ∣ ≤ C a and ∥ w G o f f ∥ ℓ w 2 → ℓ w 2 ≤ C \|wG_{\mathrm{off}}\|_{\ell^2_w\to\ell^2_w}\le C ∥ w G off ∥ ℓ w 2 → ℓ w 2 ≤ C by packing and Schur’s test, this part has norm at most C a / w = C a h Ca/w=Ca^h C a / w = C a h . On the diagonal, (5) gives
∣ α a H ( z i , z i ) ∣ ≤ C t w d − 1 = C t d 2 . |\alpha_aH(z_i,z_i)|\le Ctw d^{-1}=Ctd^2. ∣ α a H ( z i , z i ) ∣ ≤ C tw d − 1 = C t d 2 .
This proves (73) and
∥ I − α a G H − B a , P ∥ ≤ C ( a h + t d 2 ) . (74) \|I-\alpha_aG^H-B_{a,P}\|\le C(a^h+td^2). \tag*{(74)} ∥ I − α a G H − B a , P ∥ ≤ C ( a h + t d 2 ) . (74)
Combine this estimate with the uniform inverse of B a , P B_{a,P} B a , P . The perturbation is smaller than 1 / 2 1/2 1/2 for small a a a and c a d m c_{\mathrm{adm}} c adm , using (54) . The resolvent identity, or a Neumann series, proves (72) .
Proposition 5.3. The exact problem (11) is uniquely solvable on E a \mathcal{E}_a E a in the regime (14) . If Z f = B a , P − 1 s f Z^f=B_{a,P}^{-1}s^f Z f = B a , P − 1 s f , then
U ≤ C t a − h ∥ f ∥ X , ∥ u a ∥ L 2 ( D ) ≤ C t a 1 − h / 2 ∥ f ∥ X , (75) U\le Cta^{-h}\|f\|_{\mathcal{X}},\qquad\|u_a\|_{L^2(D)}\le Cta^{1-h/2}\|f\|_{\mathcal{X}}, \tag*{(75)} U ≤ C t a − h ∥ f ∥ X , ∥ u a ∥ L 2 ( D ) ≤ C t a 1 − h /2 ∥ f ∥ X , (75) ∥ F ∥ H a + ∥ r ∥ ℓ w 2 ≤ C t a ∥ f ∥ X , ∥ Y ∥ ℓ w 2 + ∥ Z f ∥ ℓ w 2 ≤ C ∥ f ∥ X , (76) \|F\|_{\mathcal{H}_a}+\|r\|_{\ell^2_w}\le Cta\|f\|_{\mathcal{X}},\qquad\|Y\|_{\ell^2_w}+\|Z^f\|_{\ell^2_w}\le C\|f\|_{\mathcal{X}}, \tag*{(76)} ∥ F ∥ H a + ∥ r ∥ ℓ w 2 ≤ C t a ∥ f ∥ X , ∥ Y ∥ ℓ w 2 + ∥ Z f ∥ ℓ w 2 ≤ C ∥ f ∥ X , (76) ∥ Y − Z f ∥ ℓ w 2 ≤ C ( t a + a h + t d 2 ) ∥ f ∥ X . (77) \|Y-Z^f\|_{\ell^2_w}\le C(ta+a^h+td^2)\|f\|_{\mathcal{X}}. \tag*{(77)} ∥ Y − Z f ∥ ℓ w 2 ≤ C ( t a + a h + t d 2 ) ∥ f ∥ X . (77)
Proof. The scalar bracket in (64) equals Y i − ℓ a ( F i ) Y_i-\ell_a(F_i) Y i − ℓ a ( F i ) , by (71) . Therefore the exact reconstruction of the local field is
u ^ i = ψ a Y i + R a F i − ψ a ℓ a ( F i ) . (78) \widehat{u}_i=\psi_aY_i+R_aF_i-\psi_a\ell_a(F_i). \tag*{(78)} u i = ψ a Y i + R a F i − ψ a ℓ a ( F i ) . (78)
The last two terms have zero integral on B B B . From Lemmas 2.1 and 5.2, and (57) , we obtain
U ≤ C t a − h ( ∥ Y ∥ ℓ w 2 + ∥ F ∥ H a ) , U\le Cta^{-h}(\|Y\|_{\ell^2_w}+\|F\|_{\mathcal{H}_a}), U ≤ C t a − h ( ∥ Y ∥ ℓ w 2 + ∥ F ∥ H a ) ,
∥ Y ∥ ℓ w 2 ≤ C ( ∥ f ∥ X + ∥ F ∥ H a ) . \|Y\|_{\ell^2_w}\le C(\|f\|_{\mathcal{X}}+\|F\|_{\mathcal{H}_a}). ∥ Y ∥ ℓ w 2 ≤ C ( ∥ f ∥ X + ∥ F ∥ H a ) .
Substituting (69) into these inequalities gives
U ≤ C t a − h ∥ f ∥ X + C t a U . (79) U\le Cta^{-h}\|f\|_{\mathcal{X}}+Cta\,U. \tag*{(79)} U ≤ C t a − h ∥ f ∥ X + C t a U . (79)
Since a ≤ d a\le d a ≤ d for small a a a , the last coefficient is bounded by C t d ≤ C c a d m 2 Ctd\le Cc_{\mathrm{adm}}^2 C t d ≤ C c adm 2 . Absorbing it proves the first estimate in (75) . The second follows from (63) , because a 3 / 2 w − 1 / 2 a − h = a 1 − h / 2 a^{3/2}w^{-1/2}a^{-h}=a^{1-h/2} a 3/2 w − 1/2 a − h = a 1 − h /2 . Equation (69) now yields ∥ F ∥ H a ≤ C ( a + t a ) ∥ f ∥ X ≤ C t a ∥ f ∥ X \|F\|_{\mathcal{H}_a}\le C(a+ta)\|f\|_{\mathcal{X}}\le Cta\|f\|_{\mathcal{X}} ∥ F ∥ H a ≤ C ( a + t a ) ∥ f ∥ X ≤ C t a ∥ f ∥ X . This proves the estimate for r r r as well. Since t a ta t a is small, the estimate for Y Y Y follows from (71) ; the one for Z f Z^f Z f follows from (17) and (57) .
For the difference of the amplitudes, subtract B a , P Z f = s f B_{a,P}Z^f=s^f B a , P Z f = s f from (71) to obtain
B a , P ( Y − Z f ) = r + [ B a , P − ( I − α a G H ) ] Y . B_{a,P}(Y-Z^f)=r+[B_{a,P}-(I-\alpha_aG^H)]Y. B a , P ( Y − Z f ) = r + [ B a , P − ( I − α a G H )] Y .
Use (74) and the estimates just proved to obtain (77) .
It remains to justify existence. For a fixed a a a and a fixed configuration, the integral operator in (62) , restricted to L 2 ( D ) L^2(D) L 2 ( D ) , is compact. Indeed its kernel is a bounded multiplier, depending on a a a , times G ( x , y ) G(x,y) G ( x , y ) , and ∣ x − y ∣ − 2 |x-y|^{-2} ∣ x − y ∣ − 2 is integrable on D × D D \times D D × D . If f = 0 f=0 f = 0 , the a priori inequality (79) gives U = 0 U=0 U = 0 . Thus the homogeneous integral equation has only the zero solution. Fredholm’s alternative gives a unique solution on D D D for every S f ∣ D S f|_D S f ∣ D .
Extend this solution to Ω \Omega Ω by the right-hand side of (62) . The source ( κ a − q ω ) 1 D u f (\kappa_a-q_\omega)1_Du^f ( κ a − q ω ) 1 D u f lies in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) for each fixed a a a . The background resolvent maps it into H 2 ( Ω ) H^2(\Omega) H 2 ( Ω ) with zero Neumann trace. The extended field therefore belongs to H 1 ( Ω ) H^1(\Omega) H 1 ( Ω ) , has flux f f f , agrees with the solution on D D D , and satisfies (11) distributionally. Conversely any weak solution of that PDE satisfies (62) by the background inverse. This proves existence and uniqueness, and completes the proof.
Closure of the point-interaction approximation
We summarize explicitly how Lemmas 5.1 and 5.2 and Proposition 5.3 close the argument of this section. Lemma 5.1 reduces all errors made in replacing the exact local integral equations by their monopole parts to
∥ F ∥ H a ≤ C a ∥ f ∥ X + C a 1 + h U . \|F\|_{H_a} \leq Ca\|f\|_{\mathcal X}+Ca^{1+h}U. ∥ F ∥ H a ≤ C a ∥ f ∥ X + C a 1 + h U .
Lemma 5.2 then supplies a uniform inverse for the actual amplitude matrix I − α a G H I-\alpha_aG^H I − α a G H . These two facts are the inputs in the absorption argument of Proposition 5.3 , which gives, without assuming solvability in advance,
U ≤ C t a − h ∥ f ∥ X , ∥ r ∥ ℓ ω 2 ≤ C t a ∥ f ∥ X , ∥ Y ∥ ℓ ω 2 ≤ C ∥ f ∥ X . (80) U \leq Cta^{-h}\|f\|_{\mathcal X}, \qquad\|r\|_{\ell^2_\omega} \leq Cta\|f\|_{\mathcal X}, \qquad\|Y\|_{\ell^2_\omega} \leq C\|f\|_{\mathcal X}. \tag*{(80)} U ≤ C t a − h ∥ f ∥ X , ∥ r ∥ ℓ ω 2 ≤ C t a ∥ f ∥ X , ∥ Y ∥ ℓ ω 2 ≤ C ∥ f ∥ X . (80)
The same proposition proves existence and uniqueness of the exact physical field by Fredholm’s alternative, so no additional solvability argument is needed later.
The remaining step is to compare the physical amplitudes Y Y Y with the ideal Foldy–Lax amplitudes Z f = B a , P − 1 s f Z^f=B_{a,P}^{-1}s^f Z f = B a , P − 1 s f . Subtracting the two systems gives the exact identity
B a , P ( Y − Z f ) = r + [ B a , P − ( I − α a G H ) ] Y . (81) B_{a,P}(Y-Z^f)=r+\left[B_{a,P}-(I-\alpha_aG^H)\right]Y. \tag*{(81)} B a , P ( Y − Z f ) = r + [ B a , P − ( I − α a G H ) ] Y . (81)
Using (17) , (74) , and (80) yields
∥ Y − Z f ∥ ℓ ω 2 ≤ C ( t a + a h + t d 2 ) ∥ f ∥ X , \|Y-Z^f\|_{\ell^2_\omega}\leq C(ta+a^h+td^2)\|f\|_{\mathcal X}, ∥ Y − Z f ∥ ℓ ω 2 ≤ C ( t a + a h + t d 2 ) ∥ f ∥ X ,
which is (77) . Thus the exact many-droplet problem has been reduced, with a quantified error, to the ideal algebraic system B a , P Z f = s f B_{a,P}Z^f=s^f B a , P Z f = s f .
It is useful to record already the form in which this estimate enters the boundary-map comparison. For another boundary datum g g g , the sampling estimate gives ∥ s g ∥ ℓ ω 2 ≤ C ∥ g ∥ X \|s^g\|_{\ell^2_\omega}\leq C\|g\|_{\mathcal X} ∥ s g ∥ ℓ ω 2 ≤ C ∥ g ∥ X . Since α a = − t ω + O ( a ) \alpha_a=-t\omega+O(a) α a = − t ω + O ( a ) , weighted Cauchy–Schwarz gives
∣ α a ∑ i Y i f s i g + t ω ∑ i Z i f s i g ∣ ≤ t ∥ Y f − Z f ∥ ℓ ω 2 ∥ s g ∥ ℓ ω 2 + ∣ α a + t ω ∣ ω ∥ Y f ∥ ℓ ω 2 ∥ s g ∥ ℓ ω 2 ≤ C ( t 2 a + t a h + t 2 d 2 ) ∥ f ∥ X ∥ g ∥ X . (82) \begin{aligned}
\left|\alpha_a\sum_iY_i^fs_i^g+t\omega\sum_iZ_i^fs_i^g\right|
&\leq t\|Y^f-Z^f\|_{\ell^2_\omega}\|s^g\|_{\ell^2_\omega}
+\frac{|\alpha_a+t\omega|}{\omega}\|Y^f\|_{\ell^2_\omega}\|s^g\|_{\ell^2_\omega}\\
&\leq C(t^2a+ta^h+t^2d^2)\|f\|_{\mathcal X}\|g\|_{\mathcal X}. \tag*{(82)}
\end{aligned} α a i ∑ Y i f s i g + t ω i ∑ Z i f s i g ≤ t ∥ Y f − Z f ∥ ℓ ω 2 ∥ s g ∥ ℓ ω 2 + ω ∣ α a + t ω ∣ ∥ Y f ∥ ℓ ω 2 ∥ s g ∥ ℓ ω 2 ≤ C ( t 2 a + t a h + t 2 d 2 ) ∥ f ∥ X ∥ g ∥ X . (82)
Here ∣ α a + t ω ∣ / ω ≤ C a / ω = C a h |\alpha_a+t\omega|/\omega\leq Ca/\omega=Ca^h ∣ α a + t ω ∣/ ω ≤ C a / ω = C a h . Equation (82) is the point-interaction error needed in Section 6; that section only has to add the local variation of the test field and compare the ideal discrete system with the auxiliary continuous medium.
Proof of Theorem 1.4: Convergence of the boundary maps
We compare the exact boundary map first with a map defined by the Foldy–Lax amplitudes, and then with the auxiliary and effective maps. All pairings in this section are bilinear. They are well-defined for f , g ∈ X f,g\in\mathcal X f , g ∈ X by the trace theorem; the identities may also be obtained first for smooth data and then extended by continuity.
The exact boundary identity
The weak forms of the exact and background equations give
⟨ ( Λ a , P − Λ 0 , ω ) f , g ⟩ = ∫ D ( κ a − q ω ) u a f S g d x . (83) \left\langle(\Lambda_{a,P}-\Lambda_{0,\omega})f,g\right\rangle=\int_D(\kappa_a-q_\omega)u_a^f S^g\,\mathrm{d}x. \tag*{(83)} ⟨ ( Λ a , P − Λ 0 , ω ) f , g ⟩ = ∫ D ( κ a − q ω ) u a f S g d x . (83)
Indeed, the exact equation implies
b 0 ( u a f , S g ) = ⟨ f , γ S g ⟩ + ∫ D ( κ a − q ω ) u a f S g . b_0(u_a^f,S^g)=\langle f,\gamma S^g\rangle+\int_D(\kappa_a-q_\omega)u_a^f S^g. b 0 ( u a f , S g ) = ⟨ f , γ S g ⟩ + ∫ D ( κ a − q ω ) u a f S g .
On the other hand, the background equation for S g S^g S g gives b 0 ( u a f , S g ) = ⟨ g , γ u a f ⟩ b_0(u_a^f,S^g)=\langle g,\gamma u_a^f\rangle b 0 ( u a f , S g ) = ⟨ g , γ u a f ⟩ . The background map is symmetric, because b 0 ( S f , S g ) = b 0 ( S g , S f ) b_0(S^f,S^g)=b_0(S^g,S^f) b 0 ( S f , S g ) = b 0 ( S g , S f ) . Subtracting its boundary pairing proves (83) .
Using (70) , expand the right-hand side as
⟨ ( Λ a , P − Λ 0 , ω ) f , g ⟩ = α a ∑ i Y i f s i g + R a ( f , g ) . (84) \left\langle(\Lambda_{a,P}-\Lambda_{0,\omega})f,g\right\rangle=\alpha_a\sum_iY_i^fs_i^g+R_a(f,g). \tag*{(84)} ⟨ ( Λ a , P − Λ 0 , ω ) f , g ⟩ = α a i ∑ Y i f s i g + R a ( f , g ) . (84)
We claim that
∣ R a ( f , g ) ∣ ≤ C t a ∥ f ∥ X ∥ g ∥ X . (85) |R_a(f,g)|\leq Ct_a\|f\|_{\mathcal{X}}\|g\|_{\mathcal{X}}. \tag*{(85)} ∣ R a ( f , g ) ∣ ≤ C t a ∥ f ∥ X ∥ g ∥ X . (85)
Indeed, the part containing the variation of S g S^g S g on the droplets is bounded, after rescaling, by
κ a a 3 ∑ i ∥ u ^ i f ∥ L 2 ( B ) ∥ S ^ i g − s i g ∥ L 2 ( B ) ≤ C a ω U ( ω ∑ i ∥ S ^ i g − s i g ∥ 2 2 ) 1 / 2 ≤ C a 2 ω t a − h ∥ f ∥ X ∥ g ∥ X = C t a ∥ f ∥ X ∥ g ∥ X . \begin{aligned}
\kappa_a a^3\sum_i\|\widehat{u}_i^f\|_{L^2(B)}\|\widehat{S}_i^g-s_i^g\|_{L^2(B)}
\leq C\frac{a}{\omega}U\left(\omega\sum_i\|\widehat{S}_i^g-s_i^g\|_2^2\right)^{1/2} \\
&\leq C\frac{a^2}{\omega}t a^{-h}\|f\|_{\mathcal{X}}\|g\|_{\mathcal{X}}=Ct_a\|f\|_{\mathcal{X}}\|g\|_{\mathcal{X}}.
\end{aligned} κ a a 3 i ∑ ∥ u i f ∥ L 2 ( B ) ∥ S i g − s i g ∥ L 2 ( B ) ≤ C ω a U ( ω i ∑ ∥ S i g − s i g ∥ 2 2 ) 1/2 ≤ C ω a 2 t a − h ∥ f ∥ X ∥ g ∥ X = C t a ∥ f ∥ X ∥ g ∥ X .
Here we used (59) and (75) . The part containing q ω q_\omega q ω is at most
C a 3 ω U ( ω ∑ i ∥ S ^ i g ∥ 2 2 ) 1 / 2 ≤ C t a 2 ∥ f ∥ X ∥ g ∥ X , C\frac{a^3}{\omega}U\left(\omega\sum_i\|\widehat{S}_i^g\|_2^2\right)^{1/2}\leq Ct_a^2\|f\|_{\mathcal{X}}\|g\|_{\mathcal{X}}, C ω a 3 U ( ω i ∑ ∥ S i g ∥ 2 2 ) 1/2 ≤ C t a 2 ∥ f ∥ X ∥ g ∥ X ,
since (57) and (59) bound the last factor. This proves (85) .
The Foldy–Lax boundary map
Define Λ a , P F L : X → Y \Lambda_{a,P}^{\mathrm{FL}}:\mathcal{X}\to\mathcal{Y} Λ a , P FL : X → Y by
⟨ ( Λ a , P F L − Λ 0 , ω ) f , g ⟩ = − t ω ∑ i Z i f s i g , B a , P Z f = s f . (86) \left\langle(\Lambda_{a,P}^{\mathrm{FL}}-\Lambda_{0,\omega})f,g\right\rangle=-t\omega\sum_i Z_i^f s_i^g,\qquad B_{a,P}Z^f=s^f. \tag*{(86)} ⟨ ( Λ a , P FL − Λ 0 , ω ) f , g ⟩ = − t ω i ∑ Z i f s i g , B a , P Z f = s f . (86)
The right-hand side is a continuous linear functional of g g g , because ∥ Z f ∥ ℓ ω 2 ≤ C ∥ f ∥ X \|Z^f\|_{\ell_\omega^2}\leq C\|f\|_{\mathcal{X}} ∥ Z f ∥ ℓ ω 2 ≤ C ∥ f ∥ X and ∥ s g ∥ ℓ ω 2 ≤ C ∥ g ∥ X \|s^g\|_{\ell_\omega^2}\leq C\|g\|_{\mathcal{X}} ∥ s g ∥ ℓ ω 2 ≤ C ∥ g ∥ X . It therefore defines an element of Y \mathcal{Y} Y , with the usual identification of the dual of X \mathcal{X} X . From (84) ,
⟨ ( Λ a , P − Λ a , P F L ) f , g ⟩ = − t ω ∑ i ( Y i f − Z i f ) s i g + ( α a + t ω ) ∑ i Y i f s i g + R a ( f , g ) . \left\langle(\Lambda_{a,P}-\Lambda_{a,P}^{\mathrm{FL}})f,g\right\rangle=-t\omega\sum_i(Y_i^f-Z_i^f)s_i^g+(\alpha_a+t\omega)\sum_iY_i^fs_i^g+R_a(f,g). ⟨ ( Λ a , P − Λ a , P FL ) f , g ⟩ = − t ω i ∑ ( Y i f − Z i f ) s i g + ( α a + t ω ) i ∑ Y i f s i g + R a ( f , g ) .
Cauchy–Schwarz in ℓ ω 2 \ell_\omega^2 ℓ ω 2 , the coefficient expansion (22) , and (77) give
∥ Λ a , P − Λ a , P F L ∥ L ( X , Y ) ≤ C ( t 2 a + t a h + t 2 d 2 ) . (87) \|\Lambda_{a,P}-\Lambda_{a,P}^{\mathrm{FL}}\|_{\mathcal{L}(\mathcal{X},\mathcal{Y})}\leq C(t^2a+ta^h+t^2d^2). \tag*{(87)} ∥ Λ a , P − Λ a , P FL ∥ L ( X , Y ) ≤ C ( t 2 a + t a h + t 2 d 2 ) . (87)
The factor in front of the coefficient remainder is ∣ α a + t ω ∣ / ω ≤ C a h |\alpha_a+t\omega|/\omega\leq Ca^h ∣ α a + t ω ∣/ ω ≤ C a h ; it is bounded by C t a h Cta^h C t a h since t ≥ 1 t\geq1 t ≥ 1 . The term C t a Cta C t a in (85) is absorbed in C t 2 a Ct^2a C t 2 a .
Comparison with the auxiliary map
Let U a f U_a^f U a f denote the restriction of u a a v , f u_a^{\mathrm{av},f} u a av , f to the support of ν a \nu_a ν a . Applying the background resolvent to (42) gives
( I + t T a ) U a f = S f ∣ supp ν a , ∥ U a f ∥ H a ≤ C ∥ f ∥ X . (88) (I+tT_a)U_a^f=S^f|_{\operatorname{supp}\nu_a},\qquad\|U_a^f\|_{\mathcal{H}_a}\le C\|f\|_{\mathcal{X}}. \tag*{(88)} ( I + t T a ) U a f = S f ∣ supp ν a , ∥ U a f ∥ H a ≤ C ∥ f ∥ X . (88)
For the norm bound, one may use either (43) or Lemma 4.1 and ∥ S f ∥ H a ≤ C ∥ S f ∥ L 2 ( Ω ) \|S^f\|_{\mathcal{H}_a}\le C\|S^f\|_{L^2(\Omega)} ∥ S f ∥ H a ≤ C ∥ S f ∥ L 2 ( Ω ) . Recall the map J J J defined in (50) . Since ( I + t K a ) J Z f = J S f (I+tK_a)JZ^f=JS^f ( I + t K a ) J Z f = J S f , subtraction yields
( I + t T a ) ( U a f − J Z f ) = ( S f − J S f ) − t ( T a − K a ) J Z f . (I+tT_a)(U_a^f-JZ^f)=(S^f-JS^f)-t(T_a-K_a)JZ^f. ( I + t T a ) ( U a f − J Z f ) = ( S f − J S f ) − t ( T a − K a ) J Z f .
Use (58) , (51) , the isometry of J J J , and the two inverse estimates. We obtain
∥ U a f − J Z f ∥ H a ≤ C ( d + t d ) ∥ f ∥ X ≤ C t d ∥ f ∥ X . (89) \|U_a^f-JZ^f\|_{\mathcal{H}_a}\le C(d+td)\|f\|_{\mathcal{X}}\le Ctd\|f\|_{\mathcal{X}}. \tag*{(89)} ∥ U a f − J Z f ∥ H a ≤ C ( d + t d ) ∥ f ∥ X ≤ C t d ∥ f ∥ X . (89)
The analogue of (83) for the auxiliary equation is
⟨ ( Λ a a v − Λ 0 , ω ) f , g ⟩ = − t ∫ Ω U f S g d ν a . (90) \langle(\Lambda_a^{\mathrm{av}}-\Lambda_{0,\omega})f,g\rangle=-t\int_{\Omega}U^fS^g\,\mathrm{d}\nu_a. \tag*{(90)} ⟨( Λ a av − Λ 0 , ω ) f , g ⟩ = − t ∫ Ω U f S g d ν a . (90)
Also, w ∑ i Z i f S i g = ∫ J Z f J S g d ν a w\sum_i Z_i^fS_i^g=\int JZ^fJS^g\,\mathrm{d}\nu_a w ∑ i Z i f S i g = ∫ J Z f J S g d ν a . Subtract (86) from (90) and use
U f S g − J Z f J S g = ( U f − J Z f ) S g + J Z f ( S g − J S g ) . U^fS^g-JZ^fJS^g=(U^f-JZ^f)S^g+JZ^f(S^g-JS^g). U f S g − J Z f J S g = ( U f − J Z f ) S g + J Z f ( S g − J S g ) .
Cauchy–Schwarz, (89) and (58) give
∥ Λ a a v − Λ a , P F L ∥ L ( X , Y ) ≤ C t 2 d . (91) \|\Lambda_a^{\mathrm{av}}-\Lambda_{a,P}^{\mathrm{FL}}\|_{\mathcal{L}(\mathcal{X},\mathcal{Y})}\le Ct^2d. \tag*{(91)} ∥ Λ a av − Λ a , P FL ∥ L ( X , Y ) ≤ C t 2 d . (91)
Combining (87) , (91) and (44) , we find
∥ Λ a , P − Λ ρ , P e f f ∥ ≤ C ( t 2 a + t a h + t 2 d 2 + t 2 d + P ϵ a ) . \|\Lambda_{a,P}-\Lambda_{\rho,P}^{\mathrm{eff}}\|\le C(t^2a+ta^h+t^2d^2+t^2d+P\epsilon_a). ∥ Λ a , P − Λ ρ , P eff ∥ ≤ C ( t 2 a + t a h + t 2 d 2 + t 2 d + P ϵ a ) .
Since a ≤ d ≤ 1 a\le d\le1 a ≤ d ≤ 1 , this is (15) . Existence and uniqueness were proved in Proposition 5.3 , and the uniform matrix inverse was proved in Lemma 4.2 . This completes the proof of Theorem 1.4 .
Finally, the limiting integral equation is
u ρ , P f ( x ) + P 2 ∫ Ω G ω ( x , y ) ρ ( y ) u ρ , P f ( y ) d y = S f ( x ) . (92) u_{\rho,P}^f(x)+P^2\int_{\Omega}G_\omega(x,y)\rho(y)u_{\rho,P}^f(y)\,\mathrm{d}y=S^f(x). \tag*{(92)} u ρ , P f ( x ) + P 2 ∫ Ω G ω ( x , y ) ρ ( y ) u ρ , P f ( y ) d y = S f ( x ) . (92)
Applying A 0 , ω A_{0,\omega} A 0 , ω gives precisely (12) . This also identifies the normalization of the effective perturbation: the measure w ∑ j δ z j w\sum_j\delta_{z_j} w ∑ j δ z j produces P 2 ρ P^2\rho P 2 ρ , without an additional factor involving the total mass of ρ \rho ρ .
A class of random configurations
We now verify the geometric and distribution assumptions for a concrete probability law. The construction permits random occupancies and random displacements, but retains a deterministic lower bound on the separation. Consequently, it is not the unrestricted sampling of independent centers from ρ ( x ) d x \rho(x)\,\mathrm{d}x ρ ( x ) d x . We call this the Bernoulli-rounded lattice model .
Occupancy and displacement laws
Let Q d \mathcal{Q}_d Q d be the collection of cubes of side length d d d in a fixed cubic grid which are contained in Ω \Omega Ω and whose distance from ∂ Ω \partial\Omega ∂ Ω is at least 2 d 2d 2 d . Put Ω d = ⋃ Q ∈ Q d Q \Omega_d=\bigcup_{Q\in\mathcal{Q}_d}Q Ω d = ⋃ Q ∈ Q d Q . The smoothness of the boundary gives
∣ Ω ∖ Ω d ∣ ≤ C d . (93) |\Omega\setminus\Omega_d|\le Cd. \tag*{(93)} ∣Ω ∖ Ω d ∣ ≤ C d . (93)
Indeed, a point at distance greater than ( 2 + 3 ) d (2+\sqrt{3})d ( 2 + 3 ) d from the boundary belongs to an admissible cube. The complement is therefore contained in a tubular neighborhood of ∂ Ω \partial\Omega ∂ Ω of width C d Cd C d , whose volume is at most C d Cd C d .
For Q ∈ Q d Q\in\mathcal{Q}_d Q ∈ Q d , set
λ Q = w − 1 ∫ Q ρ ( x ) d x , N Q = ⌊ λ Q ⌋ + B Q , P ( B Q = 1 ) = λ Q − ⌊ λ Q ⌋ . (94) \lambda_Q=w^{-1}\int_Q\rho(x)\,\mathrm{d}x,\qquad N_Q=\lfloor\lambda_Q\rfloor+B_Q,\qquad\mathbb{P}(B_Q=1)=\lambda_Q-\lfloor\lambda_Q\rfloor. \tag*{(94)} λ Q = w − 1 ∫ Q ρ ( x ) d x , N Q = ⌊ λ Q ⌋ + B Q , P ( B Q = 1 ) = λ Q − ⌊ λ Q ⌋ . (94)
The Bernoulli variables B Q B_Q B Q are independent. Since w = d 3 w=d^3 w = d 3 ,
ρ ∗ ≤ λ Q ≤ ρ ∗ , E N Q = λ Q , 0 ≤ N Q ≤ K : = ⌈ ρ ∗ ⌉ + 1. \rho_* \le\lambda_Q\le\rho^*,\qquad\mathbb{E}N_Q=\lambda_Q,\qquad0\le N_Q\le K:=\lceil\rho^*\rceil+1. ρ ∗ ≤ λ Q ≤ ρ ∗ , E N Q = λ Q , 0 ≤ N Q ≤ K := ⌈ ρ ∗ ⌉ + 1.
Fix K K K points in the interior of the unit cube, with positive mutual distance and positive distance from its boundary. In each cube Q Q Q , scale and translate these points to obtain K K K slots. Choose the radius of the displacement balls to be a sufficiently small fixed multiple of d d d . Conditional on N Q N_Q N Q , place one center in each of the first N Q N_Q N Q slots, independently according to any prescribed law supported in its displacement ball. Uniform distributions in these balls are one example. The displacement variables can be defined for all K K K slots in advance, independently of the occupancy variables; unused slots are then ignored. Here the occupancy law governs the number N Q N_Q N Q of centers in Q Q Q , whereas a displacement law specifies their random offsets from fixed slots.
All centers in a given cube, and all centers in neighboring cubes, are separated by at least c K d c_Kd c K d . They are also at distance at least c b d c_bd c b d from ∂ Ω \partial\Omega ∂ Ω . Moreover,
w M = w ∑ Q N Q ≤ K d 3 # Q d ≤ C . wM=w\sum_QN_Q\le Kd^3\#\mathcal{Q}_d\le C. w M = w Q ∑ N Q ≤ K d 3 # Q d ≤ C .
The preceding assertions can be stated as an exact probability-one statement. With the convention that a minimum over an empty set is + ∞ +\infty + ∞ , define
G a : = { ϖ : min i ≠ j ∣ z i ( ϖ ) − z j ( ϖ ) ∣ ≥ c K d , min j dist ( z j ( ϖ ) , ∂ Ω ) ≥ c b d , w M ( a , ϖ ) ≤ C M } . (95) \mathcal{G}_a:=\left\{\varpi:\min_{i\ne j}|z_i(\varpi)-z_j(\varpi)|\ge c_Kd,\quad\min_j\operatorname{dist}(z_j(\varpi),\partial\Omega)\ge c_bd,\quad wM(a,\varpi)\le CM\right\}. \tag*{(95)} G a := { ϖ : i = j min ∣ z i ( ϖ ) − z j ( ϖ ) ∣ ≥ c K d , j min dist ( z j ( ϖ ) , ∂ Ω ) ≥ c b d , w M ( a , ϖ ) ≤ C M } . (95)
The slot construction implies G a = Ξ \mathcal{G}_a=\Xi G a = Ξ (up to completion of the probability space), and therefore
P ( G a ) = 1. (96) \mathbb{P}(\mathcal{G}_a)=1. \tag*{(96)} P ( G a ) = 1. (96)
Thus Assumption 1.2 is satisfied with c 0 = c K c_0=c_K c 0 = c K and with probability exactly one, not only asymptotically one. The probability space is the product of the occupancy and displacement spaces over the cubes. With p Q = λ Q − ⌊ λ Q ⌋ p_Q=\lambda_Q-\lfloor\lambda_Q\rfloor p Q = λ Q − ⌊ λ Q ⌋ and independent offset laws ν Q , j \nu_{Q,j} ν Q , j , its distribution at scale a a a is P a = ⨂ Q ∈ Q d ( Ber ( p Q ) ⊗ ⨂ j = 1 K ν Q , j ) \mathbb{P}_a=\bigotimes_{Q\in\mathcal{Q}_d}\left(\operatorname{Ber}(p_Q)\otimes\bigotimes_{j=1}^K\nu_{Q,j}\right) P a = ⨂ Q ∈ Q d ( Ber ( p Q ) ⊗ ⨂ j = 1 K ν Q , j ) . The occupancy laws depend on Q Q Q when ρ \rho ρ is nonconstant; no stationarity assumption is imposed. For a prescribed countable sequence of scales, all these variables may be placed on one product probability space.
The event used simultaneously in Assumptions 1.2 and 1.3 will therefore be
E a ( m ) : = G a ∩ { D a ≤ C m d 1 / 2 } = { D a ≤ C m d 1 / 2 } up to a null set . (97) \mathcal{E}_a^{(m)}:=\mathcal{G}_a\cap\{\mathcal{D}_a\le C_md^{1/2}\}=\{\mathcal{D}_a\le C_md^{1/2}\}\quad\text{up to a null set}. \tag*{(97)} E a ( m ) := G a ∩ { D a ≤ C m d 1/2 } = { D a ≤ C m d 1/2 } up to a null set . (97)
Proposition 7.1 below proves that, for small d d d ,
P ( E a ( m ) ) ≥ 1 − d m . \mathbb{P}(\mathcal{E}_a^{(m)})\ge1-d^m. P ( E a ( m ) ) ≥ 1 − d m .
This is the precise probability entering (8) for the present random construction. Hence, one may take p a = d m p_a=d^m p a = d m and D a = C m d 1 / 2 D_a=C_md^{1/2} D a = C m d 1/2 .
A common-event estimate for the empirical measure
Proposition 7.1. For every m > 0 m>0 m > 0 , there are C m , d m > 0 C_m,d_m>0 C m , d m > 0 such that the preceding construction satisfies
P { D a ≤ C m d 1 / 2 } ≥ 1 − d m , 0 < d < d m . (98) \mathbb{P}\{\mathcal{D}_a\le C_m d^{1/2}\}\ge1-d^m,\qquad0<d<d_m. \tag*{(98)} P { D a ≤ C m d 1/2 } ≥ 1 − d m , 0 < d < d m . (98) The event is independent of P P P (and of the boundary inputs).
Proof. Choose a point ξ Q ∈ Q \xi_Q\in Q ξ Q ∈ Q and put X Q = N Q − λ Q X_Q=N_Q-\lambda_Q X Q = N Q − λ Q . For ∥ φ ∥ C 1 ( Ω ‾ ) ≤ 1 \|\varphi\|_{C^1(\overline{\Omega})}\le1 ∥ φ ∥ C 1 ( Ω ) ≤ 1 , the displacement of centers within Q Q Q contributes at most C d w N Q CdwN_Q C d w N Q . Also,
∣ ∫ Q ρ ( x ) ( φ ( x ) − φ ( ξ Q ) ) d x ∣ ≤ C d ∫ Q ρ ( x ) d x . \left|\int_Q \rho(x)(\varphi(x)-\varphi(\xi_Q))\,dx\right|\le Cd\int_Q\rho(x)\,dx. ∫ Q ρ ( x ) ( φ ( x ) − φ ( ξ Q )) d x ≤ C d ∫ Q ρ ( x ) d x .
Together with (93) , these estimates give
∣ ∫ Ω φ d μ a − ∫ Ω ρ φ d x ∣ ≤ C d + w ∣ ∑ Q ∈ Q d X Q φ ( ξ Q ) ∣ . (99) \left|\int_\Omega\varphi\,d\mu_a-\int_\Omega\rho\varphi\,dx\right|\le Cd+w\left|\sum_{Q\in\mathcal{Q}_d}X_Q\varphi(\xi_Q)\right|. \tag*{(99)} ∫ Ω φ d μ a − ∫ Ω ρφ d x ≤ C d + w Q ∈ Q d ∑ X Q φ ( ξ Q ) . (99)
This estimate is uniform over the displacement variables.
To control the last term simultaneously for all φ \varphi φ , partition a fixed box containing Ω \Omega Ω into coarse cubes R R R of side length L ∈ [ d , 1 ] L\in[d,1] L ∈ [ d , 1 ] . Assign Q Q Q to R R R when ξ Q ∈ R \xi_Q\in R ξ Q ∈ R and write n R = # { Q : ξ Q ∈ R } n_R=\#\{Q:\xi_Q\in R\} n R = # { Q : ξ Q ∈ R } . The numbers of coarse cubes and fine cubes satisfy
# { R : n R > 0 } ≤ C L − 3 , ∑ R n R ≤ C d − 3 , n R ≤ C ( L / d ) 3 . (100) \#\{R:n_R>0\}\le CL^{-3},\qquad\sum_R n_R\le Cd^{-3},\qquad n_R\le C(L/d)^3. \tag*{(100)} # { R : n R > 0 } ≤ C L − 3 , R ∑ n R ≤ C d − 3 , n R ≤ C ( L / d ) 3 . (100)
Use a bounded C 1 C^1 C 1 extension of φ \varphi φ to the fixed box. Such an extension exists since Ω \Omega Ω is smooth; its C 1 C^1 C 1 norm is at most C Ω C_\Omega C Ω . Replace φ ( ξ Q ) \varphi(\xi_Q) φ ( ξ Q ) by the value of this extension at one point ξ R \xi_R ξ R of the corresponding coarse cube. As ∣ X Q ∣ ≤ 1 |X_Q|\le1 ∣ X Q ∣ ≤ 1 , this replacement has error at most C L w ∑ Q 1 ≤ C L CLw\sum_Q1\le CL C L w ∑ Q 1 ≤ C L . Consequently, from (99) , we get
D a ≤ C ( d + L ) + C w ∑ R ∣ ∑ Q : ξ Q ∈ R X Q ∣ . (101) \mathcal{D}_a\le C(d+L)+Cw\sum_R\left|\sum_{Q:\xi_Q\in R}X_Q\right|. \tag*{(101)} D a ≤ C ( d + L ) + C w R ∑ Q : ξ Q ∈ R ∑ X Q . (101)
The right-hand side no longer depends on φ \varphi φ .
We now make the concentration step explicit. Write
p Q : = λ Q − ⌊ λ Q ⌋ ∈ [ 0 , 1 ) , X Q = N Q − λ Q = B Q − p Q . p_Q:=\lambda_Q-\lfloor\lambda_Q\rfloor\in[0,1),\qquad X_Q=N_Q-\lambda_Q=B_Q-p_Q. p Q := λ Q − ⌊ λ Q ⌋ ∈ [ 0 , 1 ) , X Q = N Q − λ Q = B Q − p Q .
Thus
E X Q = 0 , ∣ X Q ∣ ≤ 1 , Var ( X Q ) = p Q ( 1 − p Q ) ≤ 1 4 , (102) \mathbb{E}X_Q=0,\qquad|X_Q|\le1,\qquad\operatorname{Var}(X_Q)=p_Q(1-p_Q)\le\frac14, \tag*{(102)} E X Q = 0 , ∣ X Q ∣ ≤ 1 , Var ( X Q ) = p Q ( 1 − p Q ) ≤ 4 1 , (102)
and the two possible values of X Q X_Q X Q are − p Q -p_Q − p Q and 1 − p Q 1-p_Q 1 − p Q ; hence the length of its range is exactly one. The variables X Q X_Q X Q remain independent because the B Q B_Q B Q are independent.
There are two equivalent ways to obtain the needed tail estimate. First, Hoeffding’s inequality for independent centered variables with range length one [25 ] gives, for a fixed coarse cube R R R with n R > 0 n_R>0 n R > 0 and S R : = ∑ Q : ξ Q ∈ R X Q S_R:=\sum_{Q:\xi_Q\in R}X_Q S R := ∑ Q : ξ Q ∈ R X Q ,
P { ∣ S R ∣ > u } ≤ 2 exp ( − 2 u 2 n R ) ≤ 2 exp ( − c u 2 / n R ) . (103) \mathbb{P}\{|S_R|>u\}\le2\exp\left(-\frac{2u^2}{n_R}\right)\le2\exp(-cu^2/n_R). \tag*{(103)} P { ∣ S R ∣ > u } ≤ 2 exp ( − n R 2 u 2 ) ≤ 2 exp ( − c u 2 / n R ) . (103)
For completeness, the same bound follows directly from exponential moments. For every s ∈ R s\in\mathbb{R} s ∈ R ,
E e s X Q = e − s p Q ( 1 − p Q + p Q e s ) ≤ e s 2 / 8 . \mathbb{E}e^{sX_Q}=e^{-sp_Q}(1-p_Q+p_Qe^s)\le e^{s^2/8}. E e s X Q = e − s p Q ( 1 − p Q + p Q e s ) ≤ e s 2 /8 .
where the last inequality is Hoeffding’s elementary lemma for a centered random variable supported in an interval of length one. Independence then gives
E e s S R = ∏ Q : ξ Q ∈ R E e s X Q ≤ e n R s 2 / 8 . \mathbb{E}e^{sS_R}=\prod_{Q:\xi_Q\in R}\mathbb{E}e^{sX_Q}\le e^{n_Rs^2/8}. E e s S R = Q : ξ Q ∈ R ∏ E e s X Q ≤ e n R s 2 /8 .
For s > 0 s>0 s > 0 , Markov’s inequality yields
P { S R > u } ≤ e − s u E e s S R ≤ exp ( − s u + n R s 2 8 ) . \mathbb{P}\{S_R>u\}\le e^{-su}\mathbb{E}e^{sS_R}\le\exp\left(-su+\frac{n_Rs^2}{8}\right). P { S R > u } ≤ e − s u E e s S R ≤ exp ( − s u + 8 n R s 2 ) .
Minimizing the exponent at s = 4 u / n R s=4u/n_R s = 4 u / n R gives P { S R > u } ≤ e − 2 u 2 / n R \mathbb{P}\{S_R>u\}\le e^{-2u^2/n_R} P { S R > u } ≤ e − 2 u 2 / n R . Applying the same computation to − S R -S_R − S R and adding the two tails gives (103) .
Let 0 < δ < 1 0<\delta<1 0 < δ < 1 and let N L = # { R : n R > 0 } ≤ C L − 3 N_L=\#\{R:n_R>0\}\le CL^{-3} N L = # { R : n R > 0 } ≤ C L − 3 . Choose
u R = C n R log ( 2 N L / δ ) u_R=C\sqrt{n_R\log(2N_L/\delta)} u R = C n R log ( 2 N L / δ )
with C C C large enough that (103) gives P { ∣ S R ∣ > u R } ≤ δ / N L \mathbb{P}\{|S_R|>u_R\}\le\delta/N_L P { ∣ S R ∣ > u R } ≤ δ / N L . Summing over the at most N L N_L N L occupied coarse cubes gives
P { ∣ S R ∣ ≤ u R for every occupied R } ≥ 1 − ∑ R : n R > 0 δ N L ≥ 1 − δ . \mathbb{P}\{|S_R|\le u_R\text{ for every occupied }R\}\ge1-\sum_{R:n_R>0}\frac{\delta}{N_L}\ge1-\delta. P { ∣ S R ∣ ≤ u R for every occupied R } ≥ 1 − R : n R > 0 ∑ N L δ ≥ 1 − δ .
Since N L ≤ C L − 3 N_L\le CL^{-3} N L ≤ C L − 3 , on this event
∣ S R ∣ ≤ C n R log ( C L − 3 / δ ) for every occupied coarse cube R . (104) |S_R|\le C\sqrt{n_R\log(CL^{-3}/\delta)}\quad\text{for every occupied coarse cube }R. \tag*{(104)} ∣ S R ∣ ≤ C n R log ( C L − 3 / δ ) for every occupied coarse cube R . (104)
Cauchy–Schwarz and (100) give
w ∑ R n R ≤ w ( # { R : n R > 0 } ) 1 / 2 ( ∑ R n R ) 1 / 2 ≤ C ( d / L ) 3 / 2 . w\sum_R\sqrt{n_R}\le w\bigl(\#\{R:n_R>0\}\bigr)^{1/2}\left(\sum_Rn_R\right)^{1/2}\le C(d/L)^{3/2}. w R ∑ n R ≤ w ( # { R : n R > 0 } ) 1/2 ( R ∑ n R ) 1/2 ≤ C ( d / L ) 3/2 .
Inserting these inequalities into (101) , we obtain, on the same event,
D a ≤ C [ d + L + ( d / L ) 3 / 2 log ( C L − 3 / δ ) ] . (105) \mathcal{D}_a\le C\left[d+L+(d/L)^{3/2}\sqrt{\log(CL^{-3}/\delta)}\right]. \tag*{(105)} D a ≤ C [ d + L + ( d / L ) 3/2 log ( C L − 3 / δ ) ] . (105)
Choose L = d 1 / 2 L=d^{1/2} L = d 1/2 and δ = d m \delta=d^m δ = d m . The last term is bounded by C m d 3 / 4 1 + ∣ log d ∣ ≤ C m d 1 / 2 C_m d^{3/4}\sqrt{1+|\log d|}\le C_m d^{1/2} C m d 3/4 1 + ∣ log d ∣ ≤ C m d 1/2 for small d d d . This proves (98) .
Corollary 7.2. For each fixed m > 0 m>0 m > 0 , the boundary maps for this construction satisfy, with probability at least 1 − d m 1-d^m 1 − d m ,
∥ Λ a , P − Λ ρ , P e f f ∥ ≤ C m ( P 2 a h + P 4 d + P d 1 / 10 ) , (106) \|\Lambda_{a,P}-\Lambda^{\mathrm{eff}}_{\rho,P}\|\le C_m(P^2a^h+P^4d+Pd^{1/10}), \tag*{(106)} ∥ Λ a , P − Λ ρ , P eff ∥ ≤ C m ( P 2 a h + P 4 d + P d 1/10 ) , (106)
simultaneously for all P ≥ P 0 P\ge P_0 P ≥ P 0 such that P d 1 / 10 ≤ c m Pd^{1/10}\le c_m P d 1/10 ≤ c m .
Proof. On the event in Proposition 7.1 , take D a = C m d 1 / 2 D_a=C_md^{1/2} D a = C m d 1/2 in Assumption 1.3 . Then ϵ a = ( D a + d ) 1 / 5 ≤ C m d 1 / 10 \epsilon_a=(D_a+d)^{1/5}\le C_md^{1/10} ϵ a = ( D a + d ) 1/5 ≤ C m d 1/10 . Theorem 1.4 gives the assertion after adjusting c m c_m c m .
We finally spell out the almost-sure statement. Fix m > 0 m>0 m > 0 and a countable sequence of scales d k ↓ 0 d_k\downarrow0 d k ↓ 0 , with the corresponding a k = d k 3 / ( 1 − h ) a_k=d_k^{3/(1-h)} a k = d k 3/ ( 1 − h ) , and realize all occupancies and displacements on the single product probability space described above. Define
E k : = { D a k ≤ C m d k 1 / 2 } . E_k:=\{\mathcal{D}_{a_k}\le C_md_k^{1/2}\}. E k := { D a k ≤ C m d k 1/2 } .
Proposition 7.1 gives P ( E k c ) ≤ d k m \mathbb{P}(E_k^c) \le d_k^m P ( E k c ) ≤ d k m for all sufficiently large k k k . Thus there is k 0 k_0 k 0 such that, for the concrete choice d k = 2 − k d_k = 2^{-k} d k = 2 − k ,
∑ k = k 0 ∞ P ( E k c ) ≤ ∑ k = k 0 ∞ 2 − m k < ∞ . \sum_{k=k_0}^{\infty} \mathbb{P}(E_k^c) \le\sum_{k=k_0}^{\infty} 2^{-mk} < \infty. k = k 0 ∑ ∞ P ( E k c ) ≤ k = k 0 ∑ ∞ 2 − mk < ∞.
Adding the finitely many terms with k < k 0 k < k_0 k < k 0 shows that ∑ k = 1 ∞ P ( E k c ) < ∞ \sum_{k=1}^{\infty} \mathbb{P}(E_k^c) < \infty ∑ k = 1 ∞ P ( E k c ) < ∞ . Hence the first Borel–Cantelli lemma yields
P ( E k c infinitely often ) = 0 , \mathbb{P}(E_k^c\ \text{infinitely often}) = 0, P ( E k c infinitely often ) = 0 ,
or, equivalently,
Ξ t y p : = ⋃ K = 1 ∞ ⋂ k ≥ K E k , P ( Ξ t y p ) = 1. (107) \Xi_{\mathrm{typ}} := \bigcup_{K=1}^{\infty}\bigcap_{k\ge K} E_k,\qquad\mathbb{P}(\Xi_{\mathrm{typ}}) = 1. \tag*{(107)} Ξ typ := K = 1 ⋃ ∞ k ≥ K ⋂ E k , P ( Ξ typ ) = 1. (107)
Thus for every ϖ ∈ Ξ t y p \varpi\in\Xi_{\mathrm{typ}} ϖ ∈ Ξ typ there is a finite K ( ϖ ) K(\varpi) K ( ϖ ) such that, for every k ≥ K ( ϖ ) k \ge K(\varpi) k ≥ K ( ϖ ) , (106) holds for the configuration at scale a k a_k a k . Because E k E_k E k depends only on the configuration and not on P P P , the same realization and the same index K ( ϖ ) K(\varpi) K ( ϖ ) work simultaneously for every P ≥ P 0 P \ge P_0 P ≥ P 0 satisfying P d k 1 / 10 ≤ c m P d_k^{1/10} \le c_m P d k 1/10 ≤ c m . In particular one may choose a scale-dependent deterministic P k P_k P k after the random realization has been fixed, provided P k d k 1 / 10 ≤ c m P_k d_k^{1/10} \le c_m P k d k 1/10 ≤ c m .
Nothing special is used about the dyadic sequence except summability. If ∑ k d k m < ∞ \sum_k d_k^m < \infty ∑ k d k m < ∞ , then ∑ k P ( E k c ) < ∞ \sum_k \mathbb{P}(E_k^c) < \infty ∑ k P ( E k c ) < ∞ and the same Borel–Cantelli argument proves (107) . No independence between different scales is required. This is the precise almost-sure form of the random extension; the analytic Foldy–Lax and boundary-map estimates remain deterministic statements for each realization belonging to the corresponding event.
Linearization of the N-D map and reconstruction
The reference resolvent
The reference operator for the prescribed density is
L P = − Δ + P 2 ρ ( x ) . L_P = -\Delta+ P^2\rho(x). L P = − Δ + P 2 ρ ( x ) .
Let R P R_P R P be its inverse with zero Neumann condition. For f ∈ X f \in\mathcal{X} f ∈ X , denote by v P f v_P^f v P f and Λ P r e f \Lambda_P^{\mathrm{ref}} Λ P ref the solution and the boundary map of
L P v P f = 0 in Ω , ∂ ν v P f = f , Λ P r e f f = γ v P f . (108) L_P v_P^f = 0\quad\text{in }\Omega,\qquad\partial_\nu v_P^f = f,\qquad\Lambda_P^{\mathrm{ref}} f = \gamma v_P^f. \tag*{(108)} L P v P f = 0 in Ω , ∂ ν v P f = f , Λ P ref f = γ v P f . (108)
The operator L P L_P L P , including its variable coefficient ρ \rho ρ , is known.
Lemma 8.1. For P ≥ P 0 ≥ 1 P \ge P_0 \ge1 P ≥ P 0 ≥ 1 ,
∥ R P ∥ L 2 → L 2 ≤ ( ρ ∗ P 2 ) − 1 , ∥ R P ∥ L 2 → H 1 + ∥ γ R P ∥ L 2 → Y ≤ C P − 1 , ∥ v P f ∥ H 1 ≤ C ∥ f ∥ X , ∥ v P f ∥ L 2 ≤ C P − 1 ∥ f ∥ X . (109) \begin{aligned}
\lVert R_P\rVert_{L^2\to L^2} &\le(\rho_*P^2)^{-1},&
\lVert R_P\rVert_{L^2\to H^1} + \lVert\gamma R_P\rVert_{L^2\to Y} &\le CP^{-1},\\
\lVert v_P^f\rVert_{H^1} &\le C\lVert f\rVert_{\mathcal{X}},&
\lVert v_P^f\rVert_{L^2} &\le CP^{-1}\lVert f\rVert_{\mathcal{X}}. \tag*{(109)}
\end{aligned} ∥ R P ∥ L 2 → L 2 ∥ v P f ∥ H 1 ≤ ( ρ ∗ P 2 ) − 1 , ≤ C ∥ f ∥ X , ∥ R P ∥ L 2 → H 1 + ∥ γ R P ∥ L 2 → Y ∥ v P f ∥ L 2 ≤ C P − 1 , ≤ C P − 1 ∥ f ∥ X . (109) ∥ v P f ∥ H 1 ≤ C ∥ f ∥ X , ∥ v P f ∥ L 2 ≤ C P − 1 ∥ f ∥ X . (110) \lVert v_P^f\rVert_{H^1} \le C\lVert f\rVert_{\mathcal{X}},\qquad\lVert v_P^f\rVert_{L^2} \le CP^{-1}\lVert f\rVert_{\mathcal{X}}. \tag*{(110)} ∥ v P f ∥ H 1 ≤ C ∥ f ∥ X , ∥ v P f ∥ L 2 ≤ C P − 1 ∥ f ∥ X . (110) Proof. Let z = R P F z = R_PF z = R P F . Testing its equation with z ‾ \overline{z} z gives
∥ ∇ z ∥ 2 2 + P 2 ∫ Ω ρ ∣ z ∣ 2 = ∫ Ω F z ‾ . \lVert\nabla z\rVert_2^2 + P^2\int_{\Omega}\rho|z|^2 = \int_{\Omega}F\overline{z}. ∥ ∇ z ∥ 2 2 + P 2 ∫ Ω ρ ∣ z ∣ 2 = ∫ Ω F z . Dropping the gradient term and using Cauchy–Schwarz yields ∥ z ∥ 2 ≤ ( ρ ∗ P 2 ) − 1 ∥ F ∥ 2 \lVert z\rVert_2 \le(\rho_*P^2)^{-1}\lVert F\rVert_2 ∥ z ∥ 2 ≤ ( ρ ∗ P 2 ) − 1 ∥ F ∥ 2 . Substitution into the same identity gives ∥ ∇ z ∥ 2 2 ≤ C P − 1 ∥ F ∥ 2 \lVert\nabla z\rVert_2^2 \le CP^{-1}\lVert F\rVert_2 ∥ ∇ z ∥ 2 2 ≤ C P − 1 ∥ F ∥ 2 . The trace theorem proves the remaining estimate in (109) .
For the reference solution, its weak equation and the trace theorem give
c ∥ v P f ∥ P 2 ≤ ∣ ⟨ f , γ v P f ⟩ ∣ ≤ C ∥ f ∥ X ∥ v P f ∥ P . c\lVert v_P^f\rVert_P^2 \le\left\lvert\langle f,\gamma v_P^f\rangle\right\rvert\le C\lVert f\rVert_{\mathcal{X}}\lVert v_P^f\rVert_P. c ∥ v P f ∥ P 2 ≤ ⟨ f , γ v P f ⟩ ≤ C ∥ f ∥ X ∥ v P f ∥ P . Hence ∥ v P f ∥ P ≤ C ∥ f ∥ X \lVert v_P^f\rVert_P \le C\lVert f\rVert_{\mathcal{X}} ∥ v P f ∥ P ≤ C ∥ f ∥ X , which implies both estimates in (110) .
The linearized Neumann-to-Dirichlet map
Set Q ω = ω a , P 2 n 2 Q_\omega=\omega_{a,P}^{2}n^{2} Q ω = ω a , P 2 n 2 and
η P = q ∗ ρ ∗ P 2 . \eta_P=\frac{q_*}{\rho_*P^{2}}. η P = ρ ∗ P 2 q ∗ .
Increase P 0 P_0 P 0 so that η P ≤ 1 / 2 \eta_P\le1/2 η P ≤ 1/2 . The effective solution satisfies
u ρ , P f = v P f + R P ( Q ω u ρ , P f ) , u ρ , P f = ∑ j = 0 ∞ ( R P Q ω ) j v P f in L 2 ( Ω ) . (111) u_{\rho,P}^{f}=v_P^{f}+R_P(Q_\omega u_{\rho,P}^{f}),\qquad u_{\rho,P}^{f}=\sum_{j=0}^{\infty}(R_PQ_\omega)^jv_P^{f}\quad\text{in }L^{2}(\Omega). \tag*{(111)} u ρ , P f = v P f + R P ( Q ω u ρ , P f ) , u ρ , P f = j = 0 ∑ ∞ ( R P Q ω ) j v P f in L 2 ( Ω ) . (111)
The series follows from ∥ R P Q ω ∥ L 2 → L 2 ≤ η P < 1 \lVert R_PQ_\omega\rVert_{L^2\to L^2}\le\eta_P<1 ∥ R P Q ω ∥ L 2 → L 2 ≤ η P < 1 .
Theorem 8.2. Define, for b ∈ L ∞ ( Ω ) b\in L^\infty(\Omega) b ∈ L ∞ ( Ω ) ,
L P ( b ) f = γ R P ( b v P f ) . (112) \mathcal{L}_P(b)f=\gamma R_P(bv_P^f). \tag*{(112)} L P ( b ) f = γ R P ( b v P f ) . (112) Then
Λ ρ , P e f f = Λ P r e f + ω a , P 2 L P ( n 2 ) + R P , 1 , ∥ R P , 1 ∥ L ( X , Y ) ≤ C P − 4 . (113) \Lambda_{\rho,P}^{\mathrm{eff}}=\Lambda_P^{\mathrm{ref}}+\omega_{a,P}^{2}\mathcal{L}_P(n^{2})+R_{P,1},\qquad\lVert R_{P,1}\rVert_{\mathcal{L}(\mathcal{X},\mathcal{Y})}\le CP^{-4}. \tag*{(113)} Λ ρ , P eff = Λ P ref + ω a , P 2 L P ( n 2 ) + R P , 1 , ∥ R P , 1 ∥ L ( X , Y ) ≤ C P − 4 . (113) More generally, after retaining the terms of orders 1 , … , K 1,\ldots,K 1 , … , K in (111) , the boundary remainder satisfies
∥ R P , K ∥ L ( X , Y ) ≤ C P − 2 η P K 1 − η P ≤ C K P − 2 K − 2 , K ≥ 1. (114) \lVert R_{P,K}\rVert_{\mathcal{L}(\mathcal{X},\mathcal{Y})}\le CP^{-2}\frac{\eta_P^K}{1-\eta_P}\le C_KP^{-2K-2},\qquad K\ge1. \tag*{(114)} ∥ R P , K ∥ L ( X , Y ) ≤ C P − 2 1 − η P η P K ≤ C K P − 2 K − 2 , K ≥ 1. (114) The constants are uniform in the frequency and the geometric parameters of Theorem 1.4 .
Proof. For j ≥ 1 j\ge1 j ≥ 1 , take the trace of the j j j th term in (111) by separating its first resolvent:
γ ( R P Q ω ) j v P f = γ R P Q ω ( R P Q ω ) j − 1 v P f . \gamma(R_PQ_\omega)^jv_P^f=\gamma R_PQ_\omega(R_PQ_\omega)^{j-1}v_P^f. γ ( R P Q ω ) j v P f = γ R P Q ω ( R P Q ω ) j − 1 v P f .
Lemma 8.1 gives
∥ γ ( R P Q ω ) j v P f ∥ Y ≤ ∥ γ R P ∥ L 2 → Y ∥ Q ω ∥ ∞ η P j − 1 ∥ v P f ∥ L 2 ≤ C P − 2 η P j − 1 ∥ f ∥ X . \lVert\gamma(R_PQ_\omega)^jv_P^f\rVert_{\mathcal{Y}}\le\lVert\gamma R_P\rVert_{L^2\to\mathcal{Y}}\lVert Q_\omega\rVert_\infty\eta_P^{j-1}\lVert v_P^f\rVert_{L^2}
\le CP^{-2}\eta_P^{j-1}\lVert f\rVert_{\mathcal{X}}. ∥ γ ( R P Q ω ) j v P f ∥ Y ≤ ∥ γ R P ∥ L 2 → Y ∥ Q ω ∥ ∞ η P j − 1 ∥ v P f ∥ L 2 ≤ C P − 2 η P j − 1 ∥ f ∥ X .
Thus the series of traces converges in Y \mathcal{Y} Y , uniformly for $\lVert f\rVert_{\mathcal{X}}\le1. Its sum agrees with the trace of the effective solution: the sum of all nonzero-order terms is R P Q ω u ρ , P f R_PQ_\omega u_{\rho,P}^{f} R P Q ω u ρ , P f , which belongs to H 1 ( Ω ) H^1(\Omega) H 1 ( Ω ) by (109) . Summing the displayed estimate from j = K + 1 j=K+1 j = K + 1 proves (114) . The first term is ω a , P 2 L P ( n 2 ) f \omega_{a,P}^{2}\mathcal{L}_P(n^{2})f ω a , P 2 L P ( n 2 ) f , which proves (113) .
The map in (112) is the Fréchet derivative, at b = 0 b=0 b = 0 , of the Neumann-to-Dirichlet map for L P − b L_P-b L P − b . This follows from the same series, with Q ω Q_\omega Q ω replaced by b b b since the terms of degree at least two are bounded by C P − 4 ∥ b ∥ ∞ 2 CP^{-4}\lVert b\rVert_\infty^{2} C P − 4 ∥ b ∥ ∞ 2 on a fixed bounded set of b b b , for sufficiently large P P P . Thus the linearization concerns the unknown zeroth-order coefficient, while the large known potential P 2 ρ P^{2}\rho P 2 ρ is retained in the reference equation.
Suppose that a measured map satisfies
∥ Λ a , P m e a s − Λ a , P ∥ L ( X , Y ) ≤ δ a , P . (115) \lVert\Lambda_{a,P}^{\mathrm{meas}}-\Lambda_{a,P}\rVert_{\mathcal{L}(\mathcal{X},\mathcal{Y})}\le\delta_{a,P}. \tag*{(115)} ∥ Λ a , P meas − Λ a , P ∥ L ( X , Y ) ≤ δ a , P . (115)
Theorems 1.4 and 8.2 imply
∥ ω a , P − 2 ( Λ a , P m e a s − Λ P r e f ) − L P ( n 2 ) ∥ L ( X , Y ) ≤ C ( E ( a , P ) + δ a , P + P − 4 ) . (116) \left\lVert\omega_{a,P}^{-2}\left(\Lambda_{a,P}^{\mathrm{meas}}-\Lambda_P^{\mathrm{ref}}\right)-\mathcal{L}_P(n^{2})\right\rVert_{\mathcal{L}(\mathcal{X},\mathcal{Y})}\le C\left(E(a,P)+\delta_{a,P}+P^{-4}\right). \tag*{(116)} ω a , P − 2 ( Λ a , P meas − Λ P ref ) − L P ( n 2 ) L ( X , Y ) ≤ C ( E ( a , P ) + δ a , P + P − 4 ) . (116)
The factor ω a , P − 2 \omega_{a,P}^{-2} ω a , P − 2 is uniformly bounded because I I I is a compact subset of ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) .
For the random construction in § 8 , one can make the homogenization error smaller than the linearization remainder. Indeed, choose
P = d − σ , 0 < σ < 1 50 . (117) P=d^{-\sigma},\qquad0<\sigma<\frac{1}{50}. \tag*{(117)} P = d − σ , 0 < σ < 50 1 . (117)
Then P d 1 / 10 → 0 Pd^{1/10}\to0 P d 1/10 → 0 , and
P 4 ( P 2 a h + P 4 d + P d 1 / 10 ) = d 3 h / ( 1 − h ) − 6 σ + d 1 − 8 σ + d 1 / 10 − 5 σ ⟶ 0. (118) P^{4}\left(P^{2}a^{h}+P^{4}d+Pd^{1/10}\right)=d^{3h/(1-h)-6\sigma}+d^{1-8\sigma}+d^{1/10-5\sigma}\longrightarrow0. \tag*{(118)} P 4 ( P 2 a h + P 4 d + P d 1/10 ) = d 3 h / ( 1 − h ) − 6 σ + d 1 − 8 σ + d 1/10 − 5 σ ⟶ 0. (118)
All exponents are positive for h > 1 / 3 h>1/3 h > 1/3 . Consequently the physical-to-effective error is o ( P − 4 ) o(P^{-4}) o ( P − 4 ) on the common events. This comparison concerns operator norms; the effect of CGO boundary inputs is estimated separately below.
CGO solutions for the known reference operator
We recall the precise form of the CGO input which we use. Fix − 1 < β < 0 -1<\beta<0 − 1 < β < 0 and write
L β 2 ( R 3 ) = { u : ( 1 + ∣ x ∣ 2 ) β / 2 u ∈ L 2 ( R 3 ) } . L^{2}_{\beta}(\mathbb{R}^{3})=\{u:(1+|x|^{2})^{\beta/2}u\in L^{2}(\mathbb{R}^{3})\}. L β 2 ( R 3 ) = { u : ( 1 + ∣ x ∣ 2 ) β /2 u ∈ L 2 ( R 3 )} .
For ζ ∈ C 3 \zeta\in\mathbb{C}^{3} ζ ∈ C 3 with ζ ⋅ ζ = 0 \zeta\cdot\zeta=0 ζ ⋅ ζ = 0 and ∣ ζ ∣ = 2 s |\zeta|=\sqrt{2}s ∣ ζ ∣ = 2 s , the Faddeev inverse g ζ g_{\zeta} g ζ of − Δ − 2 ζ ⋅ ∇ -\Delta-2\zeta\cdot\nabla − Δ − 2 ζ ⋅ ∇ is a convolution operator and satisfies
∥ g ζ F ∥ L β 2 ≤ C s ∥ F ∥ L β + 1 2 , ∥ ∇ g ζ F ∥ L β 2 ≤ C ∥ F ∥ L β + 1 2 , s ≥ 1. (119) \|g_{\zeta}F\|_{L^{2}_{\beta}}\leq\frac{C}{s}\|F\|_{L^{2}_{\beta+1}},\qquad\|\nabla g_{\zeta}F\|_{L^{2}_{\beta}}\leq C\|F\|_{L^{2}_{\beta+1}},\qquad s\geq1. \tag*{(119)} ∥ g ζ F ∥ L β 2 ≤ s C ∥ F ∥ L β + 1 2 , ∥∇ g ζ F ∥ L β 2 ≤ C ∥ F ∥ L β + 1 2 , s ≥ 1. (119)
These are the weighted resolvent estimates used in the construction of [42 ] ; see also [29 Sections 2–3] . The local conjugated estimates and the perturbative construction are discussed in [39 Section 3.2] . We use the canonical convolution inverse in (119) , rather than an arbitrary local right inverse, since the boundary construction in Appendix § 10 requires its fundamental-solution property. Estimate (119) is the standard CGO resolvent theorem used as an input here. The dependence on P P P is obtained by the following argument.
Choose a known real extension ρ ~ ∈ C c 1 ( R 3 ) \widetilde{\rho}\in C^{1}_{c}(\mathbb{R}^{3}) ρ ∈ C c 1 ( R 3 ) of ρ \rho ρ , with support in a fixed ball B 0 ⊃ Ω ‾ B_{0}\supset\overline{\Omega} B 0 ⊃ Ω , and put V P = P 2 ρ ~ V_{P}=P^{2}\widetilde{\rho} V P = P 2 ρ . Positivity of this extension outside Ω \Omega Ω is not needed.
Lemma 8.3. If s ≥ C ( 1 + P 2 ) s\geq C(1+P^{2}) s ≥ C ( 1 + P 2 ) , there exists a solution
v ζ ( x ) = e ζ ⋅ x ( 1 + r ζ ( x ) ) , ( − Δ + V P ) v ζ = 0 in R 3 , (120) v_{\zeta}(x)=e^{\zeta\cdot x}(1+r_{\zeta}(x)),\qquad(-\Delta+V_{P})v_{\zeta}=0\quad\text{in }\mathbb{R}^{3}, \tag*{(120)} v ζ ( x ) = e ζ ⋅ x ( 1 + r ζ ( x )) , ( − Δ + V P ) v ζ = 0 in R 3 , (120) with r ζ ∈ L β 2 ( R 3 ) r_{\zeta}\in L^{2}_{\beta}(\mathbb{R}^{3}) r ζ ∈ L β 2 ( R 3 ) and
∥ r ζ ∥ L β 2 ≤ C P 2 s , ∥ ∇ r ζ ∥ L β 2 ≤ C P 2 . (121) \|r_{\zeta}\|_{L^{2}_{\beta}}\leq C\frac{P^{2}}{s},\qquad\|\nabla r_{\zeta}\|_{L^{2}_{\beta}}\leq CP^{2}. \tag*{(121)} ∥ r ζ ∥ L β 2 ≤ C s P 2 , ∥∇ r ζ ∥ L β 2 ≤ C P 2 . (121) The solution is determined by the known potential V P V_{P} V P and the Faddeev normalization.
Proof. Multiplication by V P V_{P} V P , whose support is fixed, maps L β 2 L^{2}_{\beta} L β 2 into L β + 1 2 L^{2}_{\beta+1} L β + 1 2 with norm at most C P 2 CP^{2} C P 2 . Thus (119) implies
∥ g ζ V P ∥ L β 2 → L β 2 ≤ C P 2 / s . \|g_{\zeta}V_{P}\|_{L^{2}_{\beta}\to L^{2}_{\beta}}\leq CP^{2}/s. ∥ g ζ V P ∥ L β 2 → L β 2 ≤ C P 2 / s .
Choose the constant in s ≥ C ( 1 + P 2 ) s\geq C(1+P^{2}) s ≥ C ( 1 + P 2 ) so that this norm is at most 1 / 2 1/2 1/2 . Then
r ζ = − ( I + g ζ V P ) − 1 g ζ V P (122) r_{\zeta}=-(I+g_{\zeta}V_{P})^{-1}g_{\zeta}V_{P} \tag*{(122)} r ζ = − ( I + g ζ V P ) − 1 g ζ V P (122)
is well-defined; in the last factor V P V_{P} V P is viewed as a function. Its norm in L β + 1 2 L^{2}_{\beta+1} L β + 1 2 is at most C P 2 CP^{2} C P 2 , so the first estimate in (121) follows. Also r ζ = − g ζ [ V P ( 1 + r ζ ) ] r_{\zeta}=-g_{\zeta}[V_{P}(1+r_{\zeta})] r ζ = − g ζ [ V P ( 1 + r ζ )] and
∥ V P ( 1 + r ζ ) ∥ L β + 1 2 ≤ C P 2 ( 1 + ∥ r ζ ∥ L β 2 ) ≤ C P 2 . \|V_{P}(1+r_{\zeta})\|_{L^{2}_{\beta+1}}\leq CP^{2}(1+\|r_{\zeta}\|_{L^{2}_{\beta}})\leq CP^{2}. ∥ V P ( 1 + r ζ ) ∥ L β + 1 2 ≤ C P 2 ( 1 + ∥ r ζ ∥ L β 2 ) ≤ C P 2 .
The gradient estimate in (119) gives the second bound. Applying − Δ − 2 ζ ⋅ ∇ -\Delta-2\zeta\cdot\nabla − Δ − 2 ζ ⋅ ∇ to (122) proves the equation in (120) . Interior elliptic regularity gives v ζ ∈ H l o c 2 ( R 3 ) v_{\zeta}\in H^{2}_{\mathrm{loc}}(\mathbb{R}^{3}) v ζ ∈ H loc 2 ( R 3 ) .
For ξ ∈ R 3 \xi\in\mathbb{R}^{3} ξ ∈ R 3 , choose orthonormal vectors e 1 , e 2 e_{1},e_{2} e 1 , e 2 perpendicular to ξ \xi ξ ; when ξ = 0 \xi=0 ξ = 0 , choose any two orthonormal vectors. If s > ∣ ξ ∣ / 2 s>|\xi|/2 s > ∣ ξ ∣/2 , set
ζ 1 = s e 1 + i ( ξ 2 + s 2 − ∣ ξ ∣ 2 / 4 e 2 ) , \zeta_{1}=se_{1}+i\left(\frac{\xi}{2}+\sqrt{s^{2}-|\xi|^{2}/4}e_{2}\right), ζ 1 = s e 1 + i ( 2 ξ + s 2 − ∣ ξ ∣ 2 /4 e 2 ) ,
ζ 2 = − s e 1 + i ( ξ 2 − s 2 − ∣ ξ ∣ 2 / 4 e 2 ) . (123) \zeta_{2}=-se_{1}+i\left(\frac{\xi}{2}-\sqrt{s^{2}-|\xi|^{2}/4}e_{2}\right). \tag*{(123)} ζ 2 = − s e 1 + i ( 2 ξ − s 2 − ∣ ξ ∣ 2 /4 e 2 ) . (123)
A direct computation gives
ζ j ⋅ ζ j = 0 , ∣ ζ j ∣ = 2 s , ζ 1 + ζ 2 = i ξ . \zeta_{j}\cdot\zeta_{j}=0,\qquad|\zeta_{j}|=\sqrt{2}s,\qquad\zeta_{1}+\zeta_{2}=i\xi. ζ j ⋅ ζ j = 0 , ∣ ζ j ∣ = 2 s , ζ 1 + ζ 2 = i ξ .
Let v j = v ζ j v_{j}=v_{\zeta_{j}} v j = v ζ j and f j = ∂ ν v j ∣ ∂ Ω f_{j}=\partial_{\nu}v_{j}|_{\partial\Omega} f j = ∂ ν v j ∣ ∂ Ω . These fluxes are known and belong to X \mathcal{X} X . By uniqueness in (108) , v j ∣ Ω = v P f j v_{j}|_{\Omega}=v_{P}^{f_{j}} v j ∣ Ω = v P f j .
Proposition 8.4. For s ≥ C ( 1 + P 2 + ∣ ξ ∣ ) s\ge C(1+P^{2}+|\xi|) s ≥ C ( 1 + P 2 + ∣ ξ ∣ ) ,
∣ ⟨ L P ( n 2 ) f 1 , f 2 ⟩ − ∫ Ω n 2 ( x ) e i ξ ⋅ x d x ∣ ≤ C ( P 2 s + P 4 s 2 ) . (124) \left|\left\langle\mathcal{L}_{P}(n^{2})f_{1},f_{2}\right\rangle-\int_{\Omega}n^{2}(x)e^{i\xi\cdot x}\,dx\right|\le C\left(\frac{P^{2}}{s}+\frac{P^{4}}{s^{2}}\right). \tag*{(124)} ⟨ L P ( n 2 ) f 1 , f 2 ⟩ − ∫ Ω n 2 ( x ) e i ξ ⋅ x d x ≤ C ( s P 2 + s 2 P 4 ) . (124) In particular, the exact linearized map L P ( n 2 ) \mathcal{L}_{P}(n^{2}) L P ( n 2 ) determines n 2 n^{2} n 2 for every fixed P ≥ P 0 P\ge P_{0} P ≥ P 0 .
Proof. Put W = R P ( n 2 v 1 ) W=R_{P}(n^{2}v_{1}) W = R P ( n 2 v 1 ) . Its weak equation is L P W = n 2 v 1 \mathcal{L}_{P}W=n^{2}v_{1} L P W = n 2 v 1 with zero Neumann data. Test it with v 2 v_{2} v 2 and use the reference equation for v 2 v_{2} v 2 . The symmetry of the bilinear form gives
⟨ L P ( n 2 ) f 1 , f 2 ⟩ = ⟨ γ W , ∂ ν v 2 ⟩ = ∫ Ω n 2 v 1 v 2 d x . (125) \left\langle\mathcal{L}_{P}(n^{2})f_{1},f_{2}\right\rangle=\langle\gamma W,\partial_{\nu}v_{2}\rangle=\int_{\Omega}n^{2}v_{1}v_{2}\,dx. \tag*{(125)} ⟨ L P ( n 2 ) f 1 , f 2 ⟩ = ⟨ γ W , ∂ ν v 2 ⟩ = ∫ Ω n 2 v 1 v 2 d x . (125)
By (123) ,
v 1 v 2 = e i ξ ⋅ x ( 1 + r ζ 1 + r ζ 2 + r ζ 1 r ζ 2 ) . v_{1}v_{2}=e^{i\xi\cdot x}(1+r_{\zeta_{1}}+r_{\zeta_{2}}+r_{\zeta_{1}}r_{\zeta_{2}}). v 1 v 2 = e i ξ ⋅ x ( 1 + r ζ 1 + r ζ 2 + r ζ 1 r ζ 2 ) .
The linear remainder terms have integrals bounded by C ∥ n 2 ∥ ∞ ∣ Ω ∣ 1 / 2 P 2 / s C\|n^{2}\|_{\infty}|\Omega|^{1/2}P^{2}/s C ∥ n 2 ∥ ∞ ∣Ω ∣ 1/2 P 2 / s . Cauchy–Schwarz bounds the quadratic term by C P 4 / s 2 CP^{4}/s^{2} C P 4 / s 2 . This proves (124) . For fixed P P P and ξ \xi ξ , let s → ∞ s\to\infty s → ∞ . We obtain the Fourier transform, with positive exponential convention, of n 2 1 Ω n^{2}1_{\Omega} n 2 1 Ω at every ξ \xi ξ . Fourier inversion in L 2 ( R 3 ) L^{2}(\mathbb{R}^{3}) L 2 ( R 3 ) then determines n 2 n^{2} n 2 in Ω \Omega Ω .
The boundary-data amplification
Let R Ω R_{\Omega} R Ω be such that Ω ‾ ⊂ B ( 0 , R Ω ) \overline{\Omega}\subset B(0,R_{\Omega}) Ω ⊂ B ( 0 , R Ω ) . The CGO estimates and ∣ Re ζ j ∣ = s |\operatorname{Re}\zeta_{j}|=s ∣ Re ζ j ∣ = s give
∥ v j ∥ H 1 ( Ω ) ≤ C ( 1 + s + P 2 ) e R Ω s . (126) \|v_{j}\|_{H^{1}(\Omega)}\le C(1+s+P^{2})e^{R_{\Omega}s}. \tag*{(126)} ∥ v j ∥ H 1 ( Ω ) ≤ C ( 1 + s + P 2 ) e R Ω s . (126)
Since Δ v j = P 2 ρ v j \Delta v_{j}=P^{2}\rho v_{j} Δ v j = P 2 ρ v j in Ω \Omega Ω , the weak normal trace estimate of Lemma B.3 yields
∥ f j ∥ X ≤ C ( 1 + s + P 2 ) e R Ω s . (127) \|f_{j}\|_{\mathcal{X}}\le C(1+s+P^{2})e^{R_{\Omega}s}. \tag*{(127)} ∥ f j ∥ X ≤ C ( 1 + s + P 2 ) e R Ω s . (127)
Here we used ∥ v j ∥ 2 ≤ C e R Ω s \|v_{j}\|_{2}\le Ce^{R_{\Omega}s} ∥ v j ∥ 2 ≤ C e R Ω s , which follows from P 2 / s ≤ C P^{2}/s\le C P 2 / s ≤ C . Define B P , s , ξ = ∥ f 1 ∥ X ∥ f 2 ∥ X \mathcal{B}_{P,s,\xi}=\|f_{1}\|_{\mathcal{X}}\|f_{2}\|_{\mathcal{X}} B P , s , ξ = ∥ f 1 ∥ X ∥ f 2 ∥ X . For the approximation formed directly from the measured map,
F a , P lin ( ξ , s ) = ω a , P − 2 ⟨ ( Λ a , P m e a s − Λ P r e f ) f 1 , f 2 ⟩ , F_{a,P}^{\operatorname{lin}}(\xi,s)=\omega_{a,P}^{-2}\left\langle\left(\Lambda_{a,P}^{\mathrm{meas}}-\Lambda_{P}^{\mathrm{ref}}\right)f_{1},f_{2}\right\rangle, F a , P lin ( ξ , s ) = ω a , P − 2 ⟨ ( Λ a , P meas − Λ P ref ) f 1 , f 2 ⟩ ,
we have the complete estimate
∣ F a , P lin ( ξ , s ) − ∫ Ω n 2 e i ξ ⋅ x d x ∣ ≤ C ( P 2 s + P 4 s 2 ) + C B P , s , ξ ( E ( a , P ) + δ a , P + P − 4 ) . (128) \left|F_{a,P}^{\operatorname{lin}}(\xi,s)-\int_{\Omega}n^{2}e^{i\xi\cdot x}\,dx\right|\le C\left(\frac{P^{2}}{s}+\frac{P^{4}}{s^{2}}\right)+C\mathcal{B}_{P,s,\xi}\left(E(a,P)+\delta_{a,P}+P^{-4}\right). \tag*{(128)} F a , P lin ( ξ , s ) − ∫ Ω n 2 e i ξ ⋅ x d x ≤ C ( s P 2 + s 2 P 4 ) + C B P , s , ξ ( E ( a , P ) + δ a , P + P − 4 ) . (128)
Indeed, apply (116) to f 1 , f 2 f_{1},f_{2} f 1 , f 2 and then use (124) .
The underlying issue is a coupling between linearization accuracy and the amplification of boundary-map errors. Increasing P P P reduces the operator-level Born remainder, but the variable-density CGO construction used here is available for s ≥ C ( 1 + P 2 + ∣ ξ ∣ ) s \ge C(1+P^2+|\xi|) s ≥ C ( 1 + P 2 + ∣ ξ ∣ ) . In particular, (127) gives the generic estimate
∣ ⟨ R P , 1 f 1 , f 2 ⟩ ∣ ≤ C P − 4 B P , s , ξ ≤ C P − 4 ( 1 + s + P 2 ) 2 e 2 R Ω s . (129) |\langle R_{P,1}f_1,f_2\rangle| \le CP^{-4}B_{P,s,\xi} \le CP^{-4}(1+s+P^2)^2e^{2R_{\Omega}s}. \tag*{(129)} ∣ ⟨ R P , 1 f 1 , f 2 ⟩ ∣ ≤ C P − 4 B P , s , ξ ≤ C P − 4 ( 1 + s + P 2 ) 2 e 2 R Ω s . (129)
The last expression does not tend to zero as P P P increases in the admissible CGO regime. This shows a limitation of the operator-norm estimate, namely it does not give a lower bound on the actual tested remainder. The boundary correction removes the Born truncation, while (A.26) retains exponential amplification of the physical approximation error and measurement noise. Thus P P P , s s s , and the Fourier cutoff must be selected jointly relative to the available data accuracy. Subsection A.5 makes this selection quantitative.
When ρ ≡ ρ 0 \rho\equiv\rho_0 ρ ≡ ρ 0 is constant, the reference equation admits the explicit solutions e ζ j ⋅ x e^{\zeta_j\cdot x} e ζ j ⋅ x with
ζ 1 = P 2 ρ 0 + ∣ ξ ∣ 2 / 4 e 1 + i ξ / 2 , ζ 2 = − P 2 ρ 0 + ∣ ξ ∣ 2 / 4 e 1 + i ξ / 2. \zeta_1=\sqrt{P^2\rho_0+\lvert\xi\rvert^2/4}\,e_1+i\xi/2,\qquad\zeta_2=-\sqrt{P^2\rho_0+\lvert\xi\rvert^2/4}\,e_1+i\xi/2. ζ 1 = P 2 ρ 0 + ∣ ξ ∣ 2 /4 e 1 + i ξ /2 , ζ 2 = − P 2 ρ 0 + ∣ ξ ∣ 2 /4 e 1 + i ξ /2.
Their product is exactly e i ξ ⋅ x e^{i\xi\cdot x} e i ξ ⋅ x , recovering the constant reference construction of [21 Remark 1.7] . For nonconstant ρ \rho ρ , the remainders in Lemma 8.3 are required. In both cases, the complex vectors are spatial phases of auxiliary solutions and the incident frequency in (10) remains real.
Reconstruction from the full boundary map
This appendix gives a reconstruction procedure which uses the full effective map, rather than replacing it by its first-order term. It also transfers the procedure to the measured droplet map. The boundary equation is a reference-potential version of the constructive CGO equations in [34 , 36 ] . We include its derivation and the perturbation estimates, since their constants must be compared with the error in Theorem 1.4.
Dirichlet maps and reference fundamental solutions
We use Dirichlet-to-Neumann maps here because they enter the boundary integral equation. In the coercive regime they are the inverses of the Neumann-to-Dirichlet maps used above; see (A.1). Fix the same real frequency ω = ω P \omega=\omega_P ω = ω P and set Q = ω P 2 n 2 Q=\omega_P^2n^2 Q = ω P 2 n 2 . Denote the Dirichlet-to-Neumann maps for L P − Q L_P-Q L P − Q and L P L_P L P on Ω \Omega Ω by D P \mathcal{D}_P D P and D P 0 \mathcal{D}_P^0 D P 0 , respectively. Both Dirichlet problems are coercive for P ≥ P 0 P\ge P_0 P ≥ P 0 . Moreover,
∥ D P ∥ L ( Y , X ) + ∥ D P 0 ∥ L ( Y , X ) ≤ C ( 1 + P 2 ) , D P = ( Λ ρ , P e f f ) − 1 , D P 0 = ( Λ P r e f ) − 1 . (130) \lVert\mathcal{D}_P\rVert_{\mathcal{L}(\mathcal{Y},\mathcal{X})}+\lVert\mathcal{D}_P^0\rVert_{\mathcal{L}(\mathcal{Y},\mathcal{X})}\le C(1+P^2),\qquad\mathcal{D}_P=(\Lambda_{\rho,P}^{\mathrm{eff}})^{-1},\qquad\mathcal{D}_P^0=(\Lambda_P^{\mathrm{ref}})^{-1}. \tag*{(130)} ∥ D P ∥ L ( Y , X ) + ∥ D P 0 ∥ L ( Y , X ) ≤ C ( 1 + P 2 ) , D P = ( Λ ρ , P eff ) − 1 , D P 0 = ( Λ P ref ) − 1 . (130)
To prove the norm bound, take a bounded trace extension E : Y → H 1 ( Ω ) E:\mathcal{Y}\to H^1(\Omega) E : Y → H 1 ( Ω ) . The Dirichlet solution u h u_h u h minimizes the positive energy with trace h h h . Thus
a ρ ( u h , u h ) ≤ a ρ ( E h , E h ) ≤ C ( 1 + P 2 ) ∥ h ∥ Y 2 . a_\rho(u_h,u_h)\le a_\rho(Eh,Eh)\le C(1+P^2)\lVert h\rVert_{\mathcal{Y}}^2. a ρ ( u h , u h ) ≤ a ρ ( E h , E h ) ≤ C ( 1 + P 2 ) ∥ h ∥ Y 2 .
For traces h , k h,k h , k , symmetry and the weak equations give ⟨ D P h , k ˉ ⟩ = a P ( u h , u k ) \langle\mathcal{D}_P h,\bar{k}\rangle=a_P(u_h,u_k) ⟨ D P h , k ˉ ⟩ = a P ( u h , u k ) . Cauchy–Schwarz in this positive form proves the first bound in (A.1); the proof for D P 0 \mathcal{D}_P^0 D P 0 is the same. In particular,
∥ u h ∥ H 1 ( Ω ) + ∥ v h ∥ H 1 ( Ω ) ≤ C ( 1 + P 2 ) 1 / 2 ∥ h ∥ Y , (131) \lVert u_h\rVert_{H^1(\Omega)}+\lVert v_h\rVert_{H^1(\Omega)}\le C(1+P^2)^{1/2}\lVert h\rVert_Y, \tag*{(131)} ∥ u h ∥ H 1 ( Ω ) + ∥ v h ∥ H 1 ( Ω ) ≤ C ( 1 + P 2 ) 1/2 ∥ h ∥ Y , (131)
where v h v_h v h is the reference Dirichlet solution. The inverse relations in (A.1) follow from uniqueness of both the Dirichlet and Neumann problems. There is no restriction to mean-zero boundary data, because the positive zeroth-order term removes the constant Neumann kernel.
Retain the compactly supported extension V P = P 2 ρ ~ V_P=P^2\widetilde{\rho} V P = P 2 ρ from § 9 . On compactly supported volume sources define
G ζ 0 = e ζ ⋅ x g ζ e − ζ ⋅ x , \mathcal{G}^0_\zeta=e^{\zeta\cdot x}g_\zeta e^{-\zeta\cdot x}, G ζ 0 = e ζ ⋅ x g ζ e − ζ ⋅ x ,
G P , ζ = e ζ ⋅ x ( I + g ζ V P ) − 1 g ζ e − ζ ⋅ x . (132) \mathcal{G}_{P,\zeta}=e^{\zeta\cdot x}(I+g_\zeta V_P)^{-1}g_\zeta e^{-\zeta\cdot x}. \tag*{(132)} G P , ζ = e ζ ⋅ x ( I + g ζ V P ) − 1 g ζ e − ζ ⋅ x . (132)
Here g ζ g_\zeta g ζ is the canonical inverse in (119) . The operator G ζ 0 \mathcal{G}^0_\zeta G ζ 0 is convolution with a fundamental solution of − Δ -\Delta − Δ , whereas G P , ζ \mathcal{G}_{P,\zeta} G P , ζ is a fundamental-solution operator for − Δ + V P -\Delta+V_P − Δ + V P . Neither operator is the Neumann inverse R P R_P R P . They satisfy the resolvent identity
G P , ζ = G ζ 0 − G P , ζ V P G ζ 0 . (133) \mathcal{G}_{P,\zeta}=\mathcal{G}^0_\zeta-\mathcal{G}_{P,\zeta}V_P\mathcal{G}^0_\zeta. \tag*{(133)} G P , ζ = G ζ 0 − G P , ζ V P G ζ 0 . (133)
For ϕ ∈ X \phi\in\mathcal{X} ϕ ∈ X , let ϕ δ ∂ Ω \phi\delta_{\partial\Omega} ϕ δ ∂ Ω denote the surface distribution whose action on a test function is ⟨ ϕ , γ φ ⟩ \langle\phi,\gamma\varphi\rangle ⟨ ϕ , γ φ ⟩ . Define the reference single-layer potential and its boundary trace by
S P , ζ ϕ = G P , ζ ( ϕ δ ∂ Ω ) , S P , ζ ϕ = γ S P , ζ ϕ . (134) \mathcal{S}_{P,\zeta}\phi=\mathcal{G}_{P,\zeta}(\phi\delta_{\partial\Omega}),\qquad S_{P,\zeta}\phi=\gamma\mathcal{S}_{P,\zeta}\phi. \tag*{(134)} S P , ζ ϕ = G P , ζ ( ϕ δ ∂ Ω ) , S P , ζ ϕ = γ S P , ζ ϕ . (134)
The meaning of (A.5) for rough boundary data, and the bounds required below, follow from the next lemma.
Lemma A.1. Let B 1 B_1 B 1 be a fixed ball containing the support of V P V_P V P and Ω ‾ \overline{\Omega} Ω in its interior. There are constants C 1 , C 2 C_1,C_2 C 1 , C 2 , independent of P P P , s s s and of the direction of ζ \zeta ζ , such that, for s ≥ C 1 ( 1 + P 2 ) s\ge C_1(1+P^2) s ≥ C 1 ( 1 + P 2 ) ,
∥ G P , ζ F ∥ H 2 ( B 1 ) + ∥ G ζ 0 F ∥ H 2 ( B 1 ) ≤ C 2 e C 2 s ∥ F ∥ L 2 ( B 1 ) , (135) \lVert\mathcal{G}_{P,\zeta}F\rVert_{H^2(B_1)}+\lVert\mathcal{G}^0_\zeta F\rVert_{H^2(B_1)}\le C_2e^{C_2s}\lVert F\rVert_{L^2(B_1)}, \tag*{(135)} ∥ G P , ζ F ∥ H 2 ( B 1 ) + ∥ G ζ 0 F ∥ H 2 ( B 1 ) ≤ C 2 e C 2 s ∥ F ∥ L 2 ( B 1 ) , (135) ∥ S P , ζ ϕ ∥ H 1 ( B 1 ) + ∥ S P , ζ ϕ ∥ Y ≤ C 2 e C 2 s ∥ ϕ ∥ X . (136) \lVert\mathcal{S}_{P,\zeta}\phi\rVert_{H^1(B_1)}+\lVert S_{P,\zeta}\phi\rVert_Y\le C_2e^{C_2s}\lVert\phi\rVert_X. \tag*{(136)} ∥ S P , ζ ϕ ∥ H 1 ( B 1 ) + ∥ S P , ζ ϕ ∥ Y ≤ C 2 e C 2 s ∥ ϕ ∥ X . (136) In the first inequality, F F F is extended by zero outside B 1 B_1 B 1 . For compactly supported w ∈ H 1 ( R 3 ) w\in H^1(\mathbb{R}^3) w ∈ H 1 ( R 3 ) , one has
G P , ζ ( − Δ + V P ) w = w . (137) \mathcal{G}_{P,\zeta}(-\Delta+V_P)w=w. \tag*{(137)} G P , ζ ( − Δ + V P ) w = w . (137) The identity is understood through the continuous extension of these operators to the corresponding compactly supported distributions.
Proof. For a compactly supported source F F F , put z = ( I + g ζ V P ) − 1 g ζ F z=(I+g_\zeta V_P)^{-1}g_\zeta F z = ( I + g ζ V P ) − 1 g ζ F . The contraction estimate in Lemma 8.3 gives ∥ z ∥ L ζ 2 ≤ C s − 1 ∥ F ∥ 2 \lVert z\rVert_{L^2_\zeta}\le Cs^{-1}\lVert F\rVert_2 ∥ z ∥ L ζ 2 ≤ C s − 1 ∥ F ∥ 2 . As z = g ζ ( F − V P z ) z=g_\zeta(F-V_Pz) z = g ζ ( F − V P z ) , (119) also gives ∥ ∇ z ∥ L ζ 2 ≤ C ∥ F ∥ 2 \lVert\nabla z\rVert_{L^2_\zeta}\le C\lVert F\rVert_2 ∥ ∇ z ∥ L ζ 2 ≤ C ∥ F ∥ 2 . The equation
− Δ z = 2 ζ ⋅ ∇ z + F − V P z -\Delta z=2\zeta\cdot\nabla z+F-V_Pz − Δ z = 2 ζ ⋅ ∇ z + F − V P z
and interior elliptic regularity on a slightly larger fixed ball imply ∥ z ∥ H 2 ( B 1 ) ≤ C ( 1 + s ) ∥ F ∥ 2 \lVert z\rVert_{H^2(B_1)}\le C(1+s)\lVert F\rVert_2 ∥ z ∥ H 2 ( B 1 ) ≤ C ( 1 + s ) ∥ F ∥ 2 . Weights are equivalent to constants on each of these fixed balls. Applying this estimate to e − ζ ⋅ x F e^{-\zeta\cdot x}F e − ζ ⋅ x F and differentiating the factor e ζ ⋅ x e^{\zeta\cdot x} e ζ ⋅ x proves (A.6), after absorbing polynomial factors in s s s into e C s e^{Cs} e C s . The free case is obtained by setting V P = 0 V_P=0 V P = 0 .
For the boundary source, solve
( − Δ + 1 ) b = 0 in Ω , ∂ ν b = ϕ . (-\Delta+1)b=0\ \text{in }\Omega,\qquad\partial_\nu b=\phi. ( − Δ + 1 ) b = 0 in Ω , ∂ ν b = ϕ .
Lax–Milgram gives ∥ b ∥ H 1 ( Ω ) ≤ C ∥ ϕ ∥ X \|b\|_{H^1(\Omega)} \le C\|\phi\|_{\mathcal X} ∥ b ∥ H 1 ( Ω ) ≤ C ∥ ϕ ∥ X . In distributions on R 3 \mathbb{R}^3 R 3 ,
ϕ δ ∂ Ω = − div ( 1 Ω ∇ b ) + 1 Ω b . (138) \phi\delta_{\partial\Omega}=-\operatorname{div}(1_{\Omega}\nabla b)+1_{\Omega}b. \tag*{(138)} ϕ δ ∂ Ω = − div ( 1 Ω ∇ b ) + 1 Ω b . (138)
This follows by testing the weak Neumann equation with the restriction of a smooth function on R 3 \mathbb{R}^3 R 3 . Since the free operator is convolution, it commutes with derivatives of compactly supported distributions. Hence its potential of the source in (A.9) equals
G ζ 0 ϕ = − ∑ k = 1 3 ∂ k G ζ 0 ( 1 Ω ∂ k b ) + G ζ 0 ( 1 Ω b ) . \mathcal{G}^{0}_{\zeta}\phi=-\sum_{k=1}^{3}\partial_k\mathcal{G}^{0}_{\zeta}(1_{\Omega}\partial_kb)+\mathcal{G}^{0}_{\zeta}(1_{\Omega}b). G ζ 0 ϕ = − k = 1 ∑ 3 ∂ k G ζ 0 ( 1 Ω ∂ k b ) + G ζ 0 ( 1 Ω b ) .
The free part of (A.6) bounds this expression in H 1 ( B 1 ) H^1(B_1) H 1 ( B 1 ) by C e C s ∥ ϕ ∥ X Ce^{Cs}\|\phi\|_{\mathcal X} C e C s ∥ ϕ ∥ X . By (A.4),
G P , ζ ϕ = G ζ 0 ϕ − G P , ζ ( V P G ζ 0 ϕ ) . \mathcal{G}_{P,\zeta}\phi=\mathcal{G}^{0}_{\zeta}\phi-\mathcal{G}_{P,\zeta}(V_P\mathcal{G}^{0}_{\zeta}\phi). G P , ζ ϕ = G ζ 0 ϕ − G P , ζ ( V P G ζ 0 ϕ ) .
The product on the right is a compactly supported L 2 L^2 L 2 source. Another application of (A.6), using P 2 ≤ C s P^2\le Cs P 2 ≤ C s , proves the H 1 H^1 H 1 estimate in (A.7). Its boundary estimate follows from the trace theorem. This argument defines the potential for every ϕ ∈ X \phi\in\mathcal X ϕ ∈ X and agrees with the usual kernel definition on smooth densities.
Finally, convolution with a fundamental solution gives G ζ 0 ( − Δ ) w = w \mathcal{G}^{0}_{\zeta}(-\Delta)w=w G ζ 0 ( − Δ ) w = w for compactly supported smooth w w w . Using (A.4), we obtain
G P , ζ ( − Δ + V P ) w = G ζ 0 ( − Δ ) w − G P , ζ V P G ζ 0 ( − Δ ) w + G P , ζ V P w = w . \begin{aligned}
\mathcal{G}_{P,\zeta}(-\Delta+V_P)w=\mathcal{G}^{0}_{\zeta}(-\Delta)w-\mathcal{G}_{P,\zeta}V_P\mathcal{G}^{0}_{\zeta}(-\Delta)w+\mathcal{G}_{P,\zeta}V_Pw \\
&=w.
\end{aligned} G P , ζ ( − Δ + V P ) w = G ζ 0 ( − Δ ) w − G P , ζ V P G ζ 0 ( − Δ ) w + G P , ζ V P w = w .
Approximation by compactly supported smooth functions, together with the preceding volume and divergence-source bounds, extends this identity to compactly supported H 1 H^1 H 1 functions. This proves (A.8).
The boundary integral equation
Let v ζ v_\zeta v ζ be the known reference CGO solution and let u ζ u_\zeta u ζ be the CGO solution for the full potential V P − 1 Ω Q V_P-1_{\Omega}Q V P − 1 Ω Q . The proof of Lemma 8.3, applied to this bounded compactly supported potential, gives
u ζ = e ζ ⋅ x ( 1 + r ζ Q ) , ∥ r ζ Q ∥ L β 2 ≤ C 1 + P 2 s , s ≥ C ( 1 + P 2 ) . (139) u_\zeta=e^{\zeta\cdot x}(1+r^Q_\zeta),\qquad\|r^Q_\zeta\|_{L^2_\beta}\le C\frac{1+P^2}{s},\qquad s\ge C(1+P^2). \tag*{(139)} u ζ = e ζ ⋅ x ( 1 + r ζ Q ) , ∥ r ζ Q ∥ L β 2 ≤ C s 1 + P 2 , s ≥ C ( 1 + P 2 ) . (139)
Subtracting the conjugated reference and full equations gives
u ζ = v ζ + G P , ζ ( 1 Ω Q u ζ ) . (140) u_\zeta=v_\zeta+\mathcal{G}_{P,\zeta}(1_{\Omega}Qu_\zeta). \tag*{(140)} u ζ = v ζ + G P , ζ ( 1 Ω Q u ζ ) . (140)
Define
A P , ζ = I + S P , ζ ( D P − D P 0 ) : Y ⟶ Y . (141) A_{P,\zeta}=I+S_{P,\zeta}(D_P-D_P^0):\mathcal Y\longrightarrow\mathcal Y. \tag*{(141)} A P , ζ = I + S P , ζ ( D P − D P 0 ) : Y ⟶ Y . (141)
Proposition A.2. For s ≥ C ( 1 + P 2 ) s\ge C(1+P^2) s ≥ C ( 1 + P 2 ) , the operator A P , ζ A_{P,\zeta} A P , ζ is invertible. The trace of the full CGO solution is obtained from
A P , ζ h ζ = γ v ζ , h ζ = γ u ζ . (142) A_{P,\zeta}h_\zeta=\gamma v_\zeta,\qquad h_\zeta=\gamma u_\zeta. \tag*{(142)} A P , ζ h ζ = γ v ζ , h ζ = γ u ζ . (142)
Furthermore,
∥ A P , ζ − 1 ∥ Y → Y ≤ C e C s . (143) \|A_{P,\zeta}^{-1}\|_{\mathcal Y\to\mathcal Y}\le Ce^{Cs}. \tag*{(143)} ∥ A P , ζ − 1 ∥ Y → Y ≤ C e C s . (143)
Proof. We first derive the boundary identity for an arbitrary trace h ∈ Y h \in\mathcal{Y} h ∈ Y . Let u h u_h u h and v h v_h v h be the full and reference Dirichlet solutions with this trace. Set
w h = u h − v h ∈ H 0 1 ( Ω ) , ψ = ( D P − D P 0 ) h . w_h=u_h-v_h\in H^1_0(\Omega),\qquad\psi=(D_P-D_P^0)h. w h = u h − v h ∈ H 0 1 ( Ω ) , ψ = ( D P − D P 0 ) h .
Extend w h w_h w h by zero to R 3 \mathbb{R}^3 R 3 , obtaining a compactly supported H 1 H^1 H 1 function w ~ h \widetilde{w}_h w h . The weak equations imply
( − Δ + V P ) w ~ h = 1 Ω Q u h + ψ δ ∂ Ω . (144) (-\Delta+V_P)\widetilde{w}_h=\mathbf{1}_{\Omega}Qu_h+\psi\delta_{\partial\Omega}. \tag*{(144)} ( − Δ + V P ) w h = 1 Ω Q u h + ψ δ ∂ Ω . (144)
In particular, the boundary term has a plus sign: testing the left side and integrating the gradient term over Ω \Omega Ω produces ⟨ ∂ ν w h , γ φ ⟩ \langle\partial_\nu w_h,\gamma\varphi\rangle ⟨ ∂ ν w h , γ φ ⟩ . Apply (A.8) to (A.15):
w ~ h = G P , ζ ( 1 Ω Q u h ) + S P , ζ ψ . (145) \widetilde{w}_h=\mathcal{G}_{P,\zeta}(\mathbf{1}_{\Omega}Qu_h)+\mathcal{S}_{P,\zeta}\psi. \tag*{(145)} w h = G P , ζ ( 1 Ω Q u h ) + S P , ζ ψ . (145)
All terms have a common H 1 H^1 H 1 trace near ∂ Ω \partial\Omega ∂ Ω . Since γ w ~ h = 0 \gamma\widetilde{w}_h=0 γ w h = 0 , taking traces gives
γ G P , ζ ( 1 Ω Q u h ) = − S P , ζ ( D P − D P 0 ) h . (146) \gamma\mathcal{G}_{P,\zeta}(\mathbf{1}_{\Omega}Qu_h)=-\mathcal{S}_{P,\zeta}(D_P-D_P^0)h. \tag*{(146)} γ G P , ζ ( 1 Ω Q u h ) = − S P , ζ ( D P − D P 0 ) h . (146)
The function u h − G P , ζ ( 1 Ω Q u h ) u_h-\mathcal{G}_{P,\zeta}(\mathbf{1}_{\Omega}Qu_h) u h − G P , ζ ( 1 Ω Q u h ) solves the reference equation in Ω \Omega Ω and has trace A P , ζ h \mathcal{A}_{P,\zeta}h A P , ζ h . For u h = u ζ ∣ Ω u_h=u_\zeta|_\Omega u h = u ζ ∣ Ω , (A.11) identifies this function with v ζ v_\zeta v ζ , proving (A.13).
We next prove bijectivity, without assuming a boundary Fredholm alternative. Given g ∈ Y g\in\mathcal{Y} g ∈ Y , solve the reference Dirichlet problem with trace g g g , and call its solution v g v_g v g . Seek u u u on Ω \Omega Ω from
u = v g + G P , ζ ( 1 Ω Q u ) . (147) u=v_g+\mathcal{G}_{P,\zeta}(\mathbf{1}_{\Omega}Qu). \tag*{(147)} u = v g + G P , ζ ( 1 Ω Q u ) . (147)
After conjugation by e − ζ ⋅ x e^{-\zeta\cdot x} e − ζ ⋅ x , the operator on the right has L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) norm at most C ∥ Q ∥ ∞ / s C\lVert Q\rVert_\infty/s C ∥ Q ∥ ∞ / s . This follows from (119) and the inverse of I + g ζ V P I+g_\zeta V_P I + g ζ V P ; source and observation weights are bounded on Ω \Omega Ω . For sufficiently large s s s , this norm is at most 1 / 2 1/2 1/2 . Thus (A.18) has a unique solution. Lemma A.1 then gives u ∈ H 1 ( Ω ) u\in H^1(\Omega) u ∈ H 1 ( Ω ) . Applying the differential operator shows that u u u solves the full homogeneous equation. If h = γ u h=\gamma u h = γ u , the preceding boundary identity yields A P , ζ h = g \mathcal{A}_{P,\zeta}h=g A P , ζ h = g . This proves surjectivity.
If A P , ζ h = 0 \mathcal{A}_{P,\zeta}h=0 A P , ζ h = 0 , the reference solution u h − G P , ζ ( 1 Ω Q u h ) u_h-\mathcal{G}_{P,\zeta}(\mathbf{1}_{\Omega}Qu_h) u h − G P , ζ ( 1 Ω Q u h ) has zero Dirichlet trace, and hence is zero. The same volume contraction implies u h = 0 u_h=0 u h = 0 , so h = 0 h=0 h = 0 . This proves injectivity.
Finally, the volume contraction gives ∥ u ∥ L 2 ( Ω ) ≤ C e C s ∥ v g ∥ L 2 ( Ω ) \lVert u\rVert_{L^2(\Omega)}\leq Ce^{Cs}\lVert v_g\rVert_{L^2(\Omega)} ∥ u ∥ L 2 ( Ω ) ≤ C e C s ∥ v g ∥ L 2 ( Ω ) . Use (A.18), (A.6) and (A.2), and absorb powers of P P P into the exponential since P 2 ≤ C s P^2\leq Cs P 2 ≤ C s . This yields ∥ u ∥ H 1 ( Ω ) ≤ C e C s ∥ g ∥ Y \lVert u\rVert_{H^1(\Omega)}\leq Ce^{Cs}\lVert g\rVert_{\mathcal{Y}} ∥ u ∥ H 1 ( Ω ) ≤ C e C s ∥ g ∥ Y . Taking the trace proves (A.14).
The corrected Fourier identity
Choose ζ 1 , ζ 2 \zeta_1,\zeta_2 ζ 1 , ζ 2 as in (123) . Let g 1 = γ v ζ 1 g_1=\gamma v_{\zeta_1} g 1 = γ v ζ 1 , k 2 = γ v ζ 2 k_2=\gamma v_{\zeta_2} k 2 = γ v ζ 2 and h 1 = A P , ζ 1 − 1 g 1 h_1=\mathcal{A}_{P,\zeta_1}^{-1}g_1 h 1 = A P , ζ 1 − 1 g 1 . Define
F P ( ξ , s ) = ω − 2 ⟨ ( D P 0 − D P ) h 1 , k 2 ⟩ . (148) F_P(\xi,s)=\omega^{-2}\left\langle(D_P^0-D_P)h_1,k_2\right\rangle. \tag*{(148)} F P ( ξ , s ) = ω − 2 ⟨ ( D P 0 − D P ) h 1 , k 2 ⟩ . (148)
Green’s identity gives
F P ( ξ , s ) = ∫ Ω n 2 u ζ 1 v ζ 2 d x . (149) F_P(\xi,s)=\int_\Omega n^2u_{\zeta_1}v_{\zeta_2}\,dx. \tag*{(149)} F P ( ξ , s ) = ∫ Ω n 2 u ζ 1 v ζ 2 d x . (149)
Indeed the full weak equation gives the reference bilinear energy of u ζ 1 , v ζ 2 u_{\zeta_1},v_{\zeta_2} u ζ 1 , v ζ 2 minus ∫ Q u ζ 1 v ζ 2 \int Q u_{\zeta_1}v_{\zeta_2} ∫ Q u ζ 1 v ζ 2 as ⟨ D P h 1 , k 2 ⟩ \langle D_Ph_1,k_2\rangle ⟨ D P h 1 , k 2 ⟩ , while the reference equation gives that energy as ⟨ D P 0 h 1 , k 2 ⟩ \langle D_P^0h_1,k_2\rangle ⟨ D P 0 h 1 , k 2 ⟩ . Subtraction proves (A.20). Using (A.10) and (121) in the product, as in Proposition 8.4, gives
∣ F P ( ξ , s ) − ∫ Ω n 2 e i ξ ⋅ x d x ∣ ≤ C 1 + P 2 s , s ≥ C ( 1 + P 2 + ∣ ξ ∣ ) . (150) \left|F_P(\xi,s)-\int_\Omega n^2e^{i\xi\cdot x}\,dx\right|\leq C\frac{1+P^2}{s},\qquad s\geq C(1+P^2+|\xi|). \tag*{(150)} F P ( ξ , s ) − ∫ Ω n 2 e i ξ ⋅ x d x ≤ C s 1 + P 2 , s ≥ C ( 1 + P 2 + ∣ ξ ∣ ) . (150)
There is no Born remainder in (A.21) and the unknown full CGO trace is obtained through the boundary equation.
Perturbation of the measured map
Put
η N = ∥ Λ a , P m e a s − Λ ρ , P e f f ∥ L ( X , Y ) . \eta_N = \left\|\Lambda^{\mathrm{meas}}_{a,P}-\Lambda^{\mathrm{eff}}_{\rho,P}\right\|_{\mathcal{L}(X,Y)}. η N = Λ a , P meas − Λ ρ , P eff L ( X , Y ) .
On E a \mathcal{E}_a E a , Theorem 1.4 and (115) give η N ≤ C E ( a , P ) + δ a , P \eta_N \le C E(a,P)+\delta_{a,P} η N ≤ C E ( a , P ) + δ a , P . If ∥ D P ∥ η N ≤ 1 / 2 \|D_P\|\eta_N \le1/2 ∥ D P ∥ η N ≤ 1/2 , then the measured map is invertible and
D P m e a s = ( Λ a , P m e a s ) − 1 , η D : = ∥ D P m e a s − D P ∥ ≤ 2 ∥ D P ∥ 2 η N ≤ C ( 1 + P 2 ) 2 η N . (151) D_P^{\mathrm{meas}} = \left(\Lambda^{\mathrm{meas}}_{a,P}\right)^{-1}, \qquad\eta_D := \left\|D_P^{\mathrm{meas}}-D_P\right\| \le2\|D_P\|^2\eta_N \le C(1+P^2)^2\eta_N. \tag*{(151)} D P meas = ( Λ a , P meas ) − 1 , η D := ∥ D P meas − D P ∥ ≤ 2∥ D P ∥ 2 η N ≤ C ( 1 + P 2 ) 2 η N . (151)
For example, this follows by writing Λ a , P m e a s = Λ ρ , P e f f [ I + D P ( Λ a , P m e a s − Λ ρ , P e f f ) ] \Lambda^{\mathrm{meas}}_{a,P}=\Lambda^{\mathrm{eff}}_{\rho,P}[I+D_P(\Lambda^{\mathrm{meas}}_{a,P}-\Lambda^{\mathrm{eff}}_{\rho,P})] Λ a , P meas = Λ ρ , P eff [ I + D P ( Λ a , P meas − Λ ρ , P eff )] and inverting the second factor by a Neumann series.
For the first phase, abbreviate
κ = ∥ A P , ζ 1 − 1 ∥ , σ S = ∥ S P , ζ 1 ∥ , Δ D = ∥ D P − D P 0 ∥ . \kappa=\left\|A^{-1}_{P,\zeta_1}\right\|,\qquad\sigma_S=\left\|S_{P,\zeta_1}\right\|,\qquad\Delta_D=\left\|D_P-D_P^0\right\|. κ = A P , ζ 1 − 1 , σ S = ∥ S P , ζ 1 ∥ , Δ D = D P − D P 0 .
Replace D P D_P D P by D P m e a s D_P^{\mathrm{meas}} D P meas in (A.12), obtaining A m e a s A^{\mathrm{meas}} A meas . If κ σ S η D ≤ 1 / 2 \kappa\sigma_S\eta_D\le1/2 κ σ S η D ≤ 1/2 , it is invertible and, for
h 1 m e a s = ( A m e a s ) − 1 g 1 , h_1^{\mathrm{meas}}=\left(A^{\mathrm{meas}}\right)^{-1}g_1, h 1 meas = ( A meas ) − 1 g 1 ,
∥ h 1 m e a s ∥ Y ≤ 2 κ ∥ g 1 ∥ Y , \left\|h_1^{\mathrm{meas}}\right\|_Y\le2\kappa\left\|g_1\right\|_Y, ∥ h 1 meas ∥ Y ≤ 2 κ ∥ g 1 ∥ Y ,
∥ h 1 m e a s − h 1 ∥ Y ≤ 2 κ 2 σ S η D ∥ g 1 ∥ Y . (152) \left\|h_1^{\mathrm{meas}}-h_1\right\|_Y\le2\kappa^2\sigma_S\eta_D\left\|g_1\right\|_Y. \tag*{(152)} ∥ h 1 meas − h 1 ∥ Y ≤ 2 κ 2 σ S η D ∥ g 1 ∥ Y . (152)
The second estimate follows by subtracting the two boundary equations: A ( h 1 m e a s − h 1 ) = − S P , ζ 1 ( D P m e a s − D P ) h 1 m e a s A(h_1^{\mathrm{meas}}-h_1)=-S_{P,\zeta_1}(D_P^{\mathrm{meas}}-D_P)h_1^{\mathrm{meas}} A ( h 1 meas − h 1 ) = − S P , ζ 1 ( D P meas − D P ) h 1 meas .
Define the measured version of (A.19) by
F a , P m e a s ( ξ , s ) = ω − 2 ⟨ ( D P 0 − D P m e a s ) h 1 m e a s , k 2 ⟩ . F_{a,P}^{\mathrm{meas}}(\xi,s)=\omega^{-2}\left\langle(D_P^0-D_P^{\mathrm{meas}})h_1^{\mathrm{meas}},k_2\right\rangle. F a , P meas ( ξ , s ) = ω − 2 ⟨ ( D P 0 − D P meas ) h 1 meas , k 2 ⟩ .
Subtracting (A.19) and using (A.23) yields the explicit data estimate
∣ F a , P m e a s ( ξ , s ) − ∫ Ω n 2 e i ξ ⋅ x d x ∣ ≤ C 1 + P 2 s + 2 ω − 2 κ ∥ g 1 ∥ Y ∥ k 2 ∥ Y ( 1 + Δ D κ σ S ) η D . (153) \left|F_{a,P}^{\mathrm{meas}}(\xi,s)-\int_\Omega n^2e^{i\xi\cdot x}\,dx\right|
\le C\frac{1+P^2}{s}+2\omega^{-2}\kappa\|g_1\|_Y\|k_2\|_Y(1+\Delta_D\kappa\sigma_S)\eta_D. \tag*{(153)} F a , P meas ( ξ , s ) − ∫ Ω n 2 e i ξ ⋅ x d x ≤ C s 1 + P 2 + 2 ω − 2 κ ∥ g 1 ∥ Y ∥ k 2 ∥ Y ( 1 + Δ D κ σ S ) η D . (153)
No conditioning factor has been included in an unspecified P P P -dependent constant in this inequality.
Theorem A.3. There are constants C , c > 0 C,c>0 C , c > 0 such that, on E a \mathcal{E}_a E a , the measured boundary correction is well-defined if (14) holds, s ≥ C ( 1 + P 2 + ∣ ξ ∣ ) s\ge C(1+P^2+|\xi|) s ≥ C ( 1 + P 2 + ∣ ξ ∣ ) , and
( 1 + P 2 ) 2 e C s ( E ( a , P ) + δ a , P ) ≤ c . (154) (1+P^2)^2e^{C s}\bigl(E(a,P)+\delta_{a,P}\bigr)\le c. \tag*{(154)} ( 1 + P 2 ) 2 e C s ( E ( a , P ) + δ a , P ) ≤ c . (154) In that case,
∣ F a , P m e a s ( ξ , s ) − ∫ Ω n 2 e i ξ ⋅ x d x ∣ ≤ C 1 + P 2 s + C ( 1 + P 2 ) 2 e C s ( E ( a , P ) + δ a , P ) . (155) \left|F_{a,P}^{\mathrm{meas}}(\xi,s)-\int_\Omega n^2e^{i\xi\cdot x}\,dx\right|
\le C\frac{1+P^2}{s}+C(1+P^2)^2e^{C s}\bigl(E(a,P)+\delta_{a,P}\bigr). \tag*{(155)} F a , P meas ( ξ , s ) − ∫ Ω n 2 e i ξ ⋅ x d x ≤ C s 1 + P 2 + C ( 1 + P 2 ) 2 e C s ( E ( a , P ) + δ a , P ) . (155) The constants are independent of a a a , P P P , s s s , ξ \xi ξ and the realization in the stated ranges.
Proof. Lemmas 8.3 and A.1, the trace theorem, and Proposition A.2 give
κ + σ S + ∥ g 1 ∥ Y + ∥ k 2 ∥ Y ≤ C e C s . \kappa+\sigma_S+\|g_1\|_Y+\|k_2\|_Y\le Ce^{C s}. κ + σ S + ∥ g 1 ∥ Y + ∥ k 2 ∥ Y ≤ C e C s .
Also Δ D ≤ C ( 1 + P 2 ) \Delta_D\le C(1+P^2) Δ D ≤ C ( 1 + P 2 ) by (A.1). Since P 2 ≤ C s P^2\le Cs P 2 ≤ C s , products of these quantities are bounded by C e C ′ s Ce^{C's} C e C ′ s after increasing the constant in the exponent. Condition (A.25), with a sufficiently large C C C and sufficiently small c c c , implies both inverse conditions preceding (A.22) and (A.23). Substitution of (A.22) into (A.24) proves (A.26). All estimates for the phases are uniform in their directions, so the same constants apply to all ξ \xi ξ in the stated range.
Quantitative trade-off and recoverable Fourier band
We extract a finite-accuracy criterion from Theorem A.3 . Put M P = 1 + P 2 M_P=1+P^2 M P = 1 + P 2 and η a , P = E ( a , P ) + δ a , P \eta_{a,P}=E(a,P)+\delta_{a,P} η a , P = E ( a , P ) + δ a , P , and, for K ≥ 0 K\ge0 K ≥ 0 , define
E F ( a , P , s , K ) = sup ∣ ξ ∣ ≤ K ∣ F a , P m e a s ( ξ , s ) − ∫ Ω n 2 ( x ) e i ξ ⋅ x d x ∣ . \mathcal{E}_F(a,P,s,K)=\sup_{|\xi|\le K}\left|F_{a,P}^{\mathrm{meas}}(\xi,s)-\int_{\Omega}n^2(x)e^{i\xi\cdot x}\,\mathrm{d}x\right|. E F ( a , P , s , K ) = ∣ ξ ∣ ≤ K sup F a , P meas ( ξ , s ) − ∫ Ω n 2 ( x ) e i ξ ⋅ x d x .
Since the estimates are uniform in the directions of the CGO phases, there are constants C 0 , C 1 , C 2 , C e , c i n v > 0 C_0,C_1,C_2,C_e,c_{\mathrm{inv}}>0 C 0 , C 1 , C 2 , C e , c inv > 0 , independent of the parameters and the realization, such that, on E a \mathcal{E}_a E a and under (14) , the boundary correction is well-defined throughout ∣ ξ ∣ ≤ K |\xi|\le K ∣ ξ ∣ ≤ K whenever
s ≥ C 0 ( M P + K ) , M P 2 e C e s η a , P ≤ c i n v . (156) s\ge C_0(M_P+K),\qquad M_P^2e^{C_es}\eta_{a,P}\le c_{\mathrm{inv}}. \tag*{(156)} s ≥ C 0 ( M P + K ) , M P 2 e C e s η a , P ≤ c inv . (156)
In this range,
E F ( a , P , s , K ) ≤ C 1 M P s + C 2 M P 2 e C e s η a , P . (157) \mathcal{E}_F(a,P,s,K)\le C_1\frac{M_P}{s}+C_2M_P^2e^{C_es}\eta_{a,P}. \tag*{(157)} E F ( a , P , s , K ) ≤ C 1 s M P + C 2 M P 2 e C e s η a , P . (157)
Corollary A.4 (A sufficient Fourier-band criterion). Assume (14) and η a , P > 0 \eta_{a,P}>0 η a , P > 0 . For a prescribed Fourier error tolerance τ > 0 \tau>0 τ > 0 , set
m τ = min { c i n v , τ 2 C 2 } , S τ ( a , P ) = 1 C e log m τ M P 2 η a , P . (158) m_\tau=\min\left\{c_{\mathrm{inv}},\frac{\tau}{2C_2}\right\},\qquad S_\tau(a,P)=\frac{1}{C_e}\log\frac{m_\tau}{M_P^2\eta_{a,P}}. \tag*{(158)} m τ = min { c inv , 2 C 2 τ } , S τ ( a , P ) = C e 1 log M P 2 η a , P m τ . (158)
If the interval
max { C 0 ( M P + K ) , 2 C 1 M P τ } ≤ s ≤ S τ ( a , P ) (159) \max\left\{C_0(M_P+K),\frac{2C_1M_P}{\tau}\right\}\le s\le S_\tau(a,P) \tag*{(159)} max { C 0 ( M P + K ) , τ 2 C 1 M P } ≤ s ≤ S τ ( a , P ) (159)
is nonempty, then every s s s in this interval gives a well-defined measured boundary correction on E a \mathcal{E}_a E a and E F ( a , P , s , K ) ≤ τ \mathcal{E}_F(a,P,s,K)\le\tau E F ( a , P , s , K ) ≤ τ .
Proof. The lower bound in (A.30) gives CGO admissibility and C 1 M P / s ≤ τ / 2 C_1M_P/s\le\tau/2 C 1 M P / s ≤ τ /2 . The upper bound gives M P 2 e C e s η a , P ≤ m τ M_P^2e^{C_es}\eta_{a,P}\le m_\tau M P 2 e C e s η a , P ≤ m τ , which ensures the perturbation condition in (A.27) and bounds the second term of (A.28) by τ / 2 \tau/2 τ /2 .
Consequently, if
S τ ( a , P ) ≥ max { C 0 M P , 2 C 1 M P τ } , (160) S_\tau(a,P)\ge\max\left\{C_0M_P,\frac{2C_1M_P}{\tau}\right\}, \tag*{(160)} S τ ( a , P ) ≥ max { C 0 M P , τ 2 C 1 M P } , (160)
every Fourier radius in the range
0 ≤ K ≤ K c e r t ( a , P , τ ) : = 1 C 0 C e log m τ ( 1 + P 2 ) η a , P − ( 1 + P 2 ) (161) 0\le K\le K_{\mathrm{cert}}(a,P,\tau):=\frac{1}{C_0C_e}\log\frac{m_\tau}{(1+P^2)\eta_{a,P}}-(1+P^2) \tag*{(161)} 0 ≤ K ≤ K cert ( a , P , τ ) := C 0 C e 1 log ( 1 + P 2 ) η a , P m τ − ( 1 + P 2 ) (161)
is guaranteed at tolerance τ \tau τ . The feasibility condition is part of this conclusion. If η a , P = 0 \eta_{a,P}=0 η a , P = 0 , there is no upper restriction on s s s in this criterion. These are sufficient bounds for certified Fourier accuracy.
The window (A.30) displays the parameter trade-off: the reference potential, Fourier radius, and desired accuracy impose its lower endpoint, while the physical and measurement errors impose its upper endpoint. Equivalently, (A.30) is nonempty exactly when
η a , P ≤ m τ M P 2 exp [ − C e max { C 0 ( M P + K ) , 2 C 1 M P τ } ] . \eta_{a,P}\le\frac{m_\tau}{M_P^2}\exp\left[-C_e\max\left\{C_0(M_P+K),\frac{2C_1M_P}{\tau}\right\}\right]. η a , P ≤ M P 2 m τ exp [ − C e max { C 0 ( M P + K ) , τ 2 C 1 M P } ] .
This explicit compatibility condition is achieved in the joint regime of Subsection A.6.
For a joint limit, write L a , P = log ( 1 / ( M P 2 η a , P ) ) → ∞ L_{a,P}=\log(1/(M_P^2\eta_{a,P}))\to\infty L a , P = log ( 1/ ( M P 2 η a , P )) → ∞ . Choosing s = θ L a , P / C e s=\theta L_{a,P}/C_e s = θ L a , P / C e , with fixed 0 < θ < 1 0<\theta<1 0 < θ < 1 , gives
E F ( a , P , s , K ) ≤ C 1 C e θ M P L a , P + C 2 e − ( 1 − θ ) L a , P , K ≤ θ L a , P C 0 C e − M P , (162) \begin{aligned}
\mathcal{E}_F(a,P,s,K)&\le\frac{C_1C_e}{\theta}\frac{M_P}{L_{a,P}}+C_2e^{-(1-\theta)L_{a,P}},\\
K&\le\frac{\theta L_{a,P}}{C_0C_e}-M_P, \tag*{(162)}
\end{aligned} E F ( a , P , s , K ) K ≤ θ C 1 C e L a , P M P + C 2 e − ( 1 − θ ) L a , P , ≤ C 0 C e θ L a , P − M P , (162)
for all sufficiently large L a , P L_{a,P} L a , P in the physical admissible regime. Thus M P = o ( L a , P ) M_P=o(L_{a,P}) M P = o ( L a , P ) permits vanishing uniform Fourier error on an expanding band. This is a sufficient compatibility condition for the present estimates. Convergence of the reconstructed coefficient also requires control of the accumulated error over that band, as shown in Subsection A.6.
A sufficient joint reconstruction regime
We give one choice of parameters which meets all the preceding conditions. For the random construction of § 8 , fix ν > 0 \nu>0 ν > 0 and take, along integers k → ∞ k\to\infty k → ∞ ,
P k = k + P 0 , s k = P k 2 + ν , a k = exp ( − P k 2 + 2 ν ) , δ a k , P k ≤ exp ( − P k 2 + 2 ν ) . (163) P_k=k+P_0,\qquad s_k=P_k^{2+\nu},\qquad a_k=\exp(-P_k^{2+2\nu}),\qquad\delta_{a_k,P_k}\le\exp(-P_k^{2+2\nu}). \tag*{(163)} P k = k + P 0 , s k = P k 2 + ν , a k = exp ( − P k 2 + 2 ν ) , δ a k , P k ≤ exp ( − P k 2 + 2 ν ) . (163)
Then d k = a k ( 1 − h ) / 3 d_k=a_k^{(1-h)/3} d k = a k ( 1 − h ) /3 and P k d k 1 / 10 → 0 P_kd_k^{1/10}\to0 P k d k 1/10 → 0 . On the events of Corollary 7.2, the sum of the forward and measurement errors is bounded by a polynomial in P k P_k P k times exp ( − c h P k 2 + 2 ν ) \exp(-c_hP_k^{2+2\nu}) exp ( − c h P k 2 + 2 ν ) , where one may take
c h = min { h , ( 1 − h ) / 3 , ( 1 − h ) / 30 } > 0. c_h=\min\{h,(1-h)/3,(1-h)/30\}>0. c h = min { h , ( 1 − h ) /3 , ( 1 − h ) /30 } > 0.
This decay dominates the factor e C e s k e^{C_es_k} e C e s k . Thus (A.25) holds for large k k k , and
sup ∣ ξ ∣ ≤ R k ∣ F a k , P k m e a s ( ξ , s k ) − ∫ Ω n 2 e i ξ ⋅ x d x ∣ ≤ C P k − ν , R k = P k 1 / 3 . (164) \sup_{|\xi|\le R_k}\left|F^{\mathrm{meas}}_{a_k,P_k}(\xi,s_k)-\int_\Omega n^2e^{i\xi\cdot x}\,dx\right|\le CP_k^{-\nu},\qquad R_k=P_k^{1/3}. \tag*{(164)} ∣ ξ ∣ ≤ R k sup F a k , P k meas ( ξ , s k ) − ∫ Ω n 2 e i ξ ⋅ x d x ≤ C P k − ν , R k = P k 1/3 . (164)
One can choose the orthonormal vectors in (123) measurably in ξ \xi ξ , using finitely many coordinate choices. This makes the reconstructed Fourier data measurable.
Define
n k 2 ( x ) = ( 2 π ) − 3 ∫ ∣ ξ ∣ ≤ R k e − i ξ ⋅ x F a k , P k m e a s ( ξ , s k ) d ξ . n_k^2(x)=(2\pi)^{-3}\int_{|\xi|\le R_k}e^{-i\xi\cdot x}F^{\mathrm{meas}}_{a_k,P_k}(\xi,s_k)\,d\xi. n k 2 ( x ) = ( 2 π ) − 3 ∫ ∣ ξ ∣ ≤ R k e − i ξ ⋅ x F a k , P k meas ( ξ , s k ) d ξ .
By Plancherel’s theorem and (A.35),
∥ n k 2 − n 2 1 Ω ∥ L 2 ( R 3 ) ≤ C R k 3 / 2 P k − ν / 2 + C ∥ 1 { ∣ ξ ∣ > R k } n 2 1 Ω ^ ∥ L 2 ( R 3 ) ⟶ 0. \left\|n_k^2-n^2\mathbf{1}_\Omega\right\|_{L^2(\mathbb{R}^3)}
\le CR_k^{3/2}P_k^{-\nu/2}+C\left\|\mathbf{1}_{\{|\xi|>R_k\}}\widehat{n^2\mathbf{1}_\Omega}\right\|_{L^2(\mathbb{R}^3)}
\longrightarrow0. n k 2 − n 2 1 Ω L 2 ( R 3 ) ≤ C R k 3/2 P k − ν /2 + C 1 { ∣ ξ ∣ > R k } n 2 1 Ω L 2 ( R 3 ) ⟶ 0.
The first term is C P k − ν / 2 CP_k^{-\nu/2} C P k − ν /2 and the second is a Fourier tail of an L 2 L^2 L 2 function. The exceptional probabilities d k m d_k^m d k m are summable. By Corollary 1.5, this reconstruction holds for every realization in a set of probability one, for all sufficiently large k k k . The physical frequencies throughout this procedure are still those in (10) ; only the auxiliary CGO phases and the spatial resolution parameters tend to infinity.
Auxiliary estimates
We give the details which connect the background spectral assumption to the kernel and solution estimates used in the point-interaction proof. The underlying elliptic regularity and trace results can be found in [22 , 32 ] ; for Green kernels and boundary potential theory, see also [33 ] .
Lemma B.1. Under (1) ,
∥ S f ∥ H 1 ( Ω ) ≤ C ∥ f ∥ X , ω ∈ I , (165) \|S^f\|_{H^1(\Omega)} \le C\|f\|_{\mathcal{X}}, \qquad\omega\in I, \tag*{(165)} ∥ S f ∥ H 1 ( Ω ) ≤ C ∥ f ∥ X , ω ∈ I , (165) and (4) –(5) hold with uniform constants.
Proof. Let A + = − Δ + 1 A_+ = -\Delta+ 1 A + = − Δ + 1 with zero Neumann condition, let R + = A + − 1 R_+ = A_+^{-1} R + = A + − 1 , and put W ω = 1 + q ω W_\omega= 1+q_\omega W ω = 1 + q ω . The operator W ω W_\omega W ω is multiplication by a uniformly bounded Lipschitz function, and A 0 , ω = A + − W ω A_{0,\omega}=A_+ - W_\omega A 0 , ω = A + − W ω .
First solve A + S + f = 0 A_+S_+^f=0 A + S + f = 0 with flux f f f . By Lax–Milgram and the trace theorem, ∥ S + f ∥ H 1 ≤ C ∥ f ∥ X \|S_+^f\|_{H^1}\le C\|f\|_{\mathcal{X}} ∥ S + f ∥ H 1 ≤ C ∥ f ∥ X . The background solution has the representation
S f = S + f + A 0 , ω − 1 ( W ω S + f ) . S^f=S_+^f+A_{0,\omega}^{-1}(W_\omega S_+^f). S f = S + f + A 0 , ω − 1 ( W ω S + f ) .
The second term has zero Neumann data and cancels A 0 , ω S + f = − W ω S + f A_{0,\omega}S_+^f=-W_\omega S_+^f A 0 , ω S + f = − W ω S + f . Global Neumann regularity gives
∥ A 0 , ω − 1 F ∥ H 2 ≤ C ( ∥ F ∥ 2 + ∥ A 0 , ω − 1 F ∥ 2 ) ≤ C ∥ F ∥ 2 , \|A_{0,\omega}^{-1}F\|_{H^2}\le C(\|F\|_2+\|A_{0,\omega}^{-1}F\|_2)\le C\|F\|_2, ∥ A 0 , ω − 1 F ∥ H 2 ≤ C ( ∥ F ∥ 2 + ∥ A 0 , ω − 1 F ∥ 2 ) ≤ C ∥ F ∥ 2 ,
where the final constant is uniform by (1) . This proves (B.1).
We next prove the Green-kernel value estimate without assuming that A 0 , ω A_{0,\omega} A 0 , ω is positive. The positive Neumann kernel G + G_+ G + of A + A_+ A + has the standard bound
∣ G + ( x , y ) ∣ ≤ C ∣ x − y ∣ − 1 , sup x ∈ Ω ∥ G + ( x , ⋅ ) ∥ L 2 ( Ω ) ≤ C . |G_+(x,y)|\le C|x-y|^{-1},\qquad\sup_{x\in\Omega}\|G_+(x,\cdot)\|_{L^2(\Omega)}\le C. ∣ G + ( x , y ) ∣ ≤ C ∣ x − y ∣ − 1 , x ∈ Ω sup ∥ G + ( x , ⋅ ) ∥ L 2 ( Ω ) ≤ C .
The second bound follows from the first by integration in dimension three. The two resolvent identities imply
A 0 , ω − 1 = R + + R + W ω R + + R + W ω A 0 , ω − 1 W ω R + . (166) A_{0,\omega}^{-1}=R_++R_+W_\omega R_++R_+W_\omega A_{0,\omega}^{-1}W_\omega R_+. \tag*{(166)} A 0 , ω − 1 = R + + R + W ω R + + R + W ω A 0 , ω − 1 W ω R + . (166)
For fixed x , y x,y x , y , the kernel of the second term has modulus at most
∥ W ω ∥ ∞ ∥ G + ( x , ⋅ ) ∥ 2 ∥ G + ( ⋅ , y ) ∥ 2 ≤ C . \|W_\omega\|_\infty\|G_+(x,\cdot)\|_2\|G_+(\cdot,y)\|_2\le C. ∥ W ω ∥ ∞ ∥ G + ( x , ⋅ ) ∥ 2 ∥ G + ( ⋅ , y ) ∥ 2 ≤ C .
The kernel of the third term is the L 2 L^2 L 2 pairing of W ω G + ( x , ⋅ ) W_\omega G_+(x,\cdot) W ω G + ( x , ⋅ ) with A 0 , ω − 1 ( W ω G + ( ⋅ , y ) ) A_{0,\omega}^{-1}(W_\omega G_+(\cdot,y)) A 0 , ω − 1 ( W ω G + ( ⋅ , y )) and is also bounded by C C C , using (1) . Thus ∣ G ω ( x , y ) − G + ( x , y ) ∣ ≤ C |G_\omega(x,y)-G_+(x,y)|\le C ∣ G ω ( x , y ) − G + ( x , y ) ∣ ≤ C . Since Ω \Omega Ω is bounded, this proves ∣ G ω ( x , y ) ∣ ≤ C ∣ x − y ∣ − 1 |G_\omega(x,y)|\le C|x-y|^{-1} ∣ G ω ( x , y ) ∣ ≤ C ∣ x − y ∣ − 1 . The kernel is real and symmetric because the Neumann resolvent is real and self-adjoint.
To obtain the gradient estimate, take r = ∣ x − y ∣ r=|x-y| r = ∣ x − y ∣ and apply a scaled interior gradient estimate, or a homogeneous Neumann boundary gradient estimate, on a neighborhood of x x x of radius comparable to r r r which does not contain y y y . On that neighborhood the kernel solves the homogeneous equation, and its supremum is at most C / r C/r C / r . Consequently ∣ ∇ x G ω ( x , y ) ∣ ≤ C / r 2 |\nabla_xG_\omega(x,y)|\le C/r^2 ∣ ∇ x G ω ( x , y ) ∣ ≤ C / r 2 . For radii larger than the size of a boundary chart, use a fixed chart radius and then adjust the constant, since r ≤ diam Ω r\le\operatorname{diam}\Omega r ≤ diam Ω . The estimates are uniform for the bounded family q ω q_\omega q ω and for the finite collection of C 2 , α C^{2,\alpha} C 2 , α boundary charts. Symmetry gives the same bound for ∇ y G ω \nabla_yG_\omega ∇ y G ω . This proves (4) .
It remains to estimate the regular part near the diagonal. Fix y y y at distance at least c b cb c b from ∂ Ω \partial\Omega ∂ Ω , and choose a ball B ( y , R ) ⋐ Ω B(y,R)\Subset\Omega B ( y , R ) ⋐ Ω with R R R a fixed sufficiently small multiple of b b b . The local parametrix
K y ( x ) = 1 4 π ∣ x − y ∣ − q ω ( y ) 8 π ∣ x − y ∣ K_y(x)=\frac{1}{4\pi|x-y|}-\frac{q_\omega(y)}{8\pi}|x-y| K y ( x ) = 4 π ∣ x − y ∣ 1 − 8 π q ω ( y ) ∣ x − y ∣
satisfies, distributionally,
A 0 , ω K y = δ y + [ q ω ( y ) − q ω ( x ) ] Φ 0 ( x , y ) + q ω ( x ) q ω ( y ) 8 π ∣ x − y ∣ . A_{0,\omega}K_y=\delta_y+[q_\omega(y)-q_\omega(x)]\Phi_0(x,y)+\frac{q_\omega(x)q_\omega(y)}{8\pi}|x-y|. A 0 , ω K y = δ y + [ q ω ( y ) − q ω ( x )] Φ 0 ( x , y ) + 8 π q ω ( x ) q ω ( y ) ∣ x − y ∣.
Here we used Δ ∣ x − y ∣ = 2 / ∣ x − y ∣ \Delta|x-y| = 2/|x-y| Δ∣ x − y ∣ = 2/∣ x − y ∣ away from y y y ; this identity also holds distributionally without a point mass. Therefore F y = G ω ( ⋅ , y ) − K y F_y = G_\omega(\cdot,y) - K_y F y = G ω ( ⋅ , y ) − K y solves
A 0 , ω F y = [ q ω ( x ) − q ω ( y ) ] Φ 0 ( x , y ) − q ω ( x ) q ω ( y ) 8 π ∣ x − y ∣ . (167) A_{0,\omega}F_y = [q_\omega(x)-q_\omega(y)]\Phi_0(x,y) - \frac{q_\omega(x)q_\omega(y)}{8\pi}|x-y|. \tag*{(167)} A 0 , ω F y = [ q ω ( x ) − q ω ( y )] Φ 0 ( x , y ) − 8 π q ω ( x ) q ω ( y ) ∣ x − y ∣. (167)
The right-hand side is uniformly bounded because q ω q_\omega q ω is Lipschitz. Initially F y ∈ L 2 F_y \in L^2 F y ∈ L 2 locally, and interior elliptic regularity applied to this distributional equation gives F y ∈ H l o c 2 F_y \in H^2_{\mathrm{loc}} F y ∈ H loc 2 , so that it is continuous near y y y . On ∂ B ( y , R ) \partial B(y,R) ∂ B ( y , R ) , (4) gives ∣ F y ∣ ≤ C / R |F_y| \le C/R ∣ F y ∣ ≤ C / R .
Rescale the Dirichlet problem on this ball. For R R R sufficiently small, the norm of the zero-Dirichlet Newtonian solution operator followed by multiplication by q ω q_\omega q ω is at most C R 2 ∥ q ω ∥ ∞ < 1 / 2 CR^2\|q_\omega\|_\infty< 1/2 C R 2 ∥ q ω ∥ ∞ < 1/2 on L ∞ L^\infty L ∞ . Splitting into the harmonic extension of the boundary values and a zero-Dirichlet correction in (B.3) gives
∥ F y ∥ L ∞ ( B ( y , R ) ) ≤ C ( R − 1 + R 2 ) ≤ C / R . \|F_y\|_{L^\infty(B(y,R))} \le C(R^{-1}+R^2) \le C/R. ∥ F y ∥ L ∞ ( B ( y , R )) ≤ C ( R − 1 + R 2 ) ≤ C / R .
The scaled interior gradient estimate then yields ∥ ∇ F y ∥ L ∞ ( B ( y , R / 2 ) ) ≤ C / R 2 \|\nabla F_y\|_{L^\infty(B(y,R/2))} \le C/R^2 ∥∇ F y ∥ L ∞ ( B ( y , R /2 )) ≤ C / R 2 . Since
H ω ( x , y ) = F y ( x ) − q ω ( y ) 8 π ∣ x − y ∣ , H_\omega(x,y) = F_y(x) - \frac{q_\omega(y)}{8\pi}|x-y|, H ω ( x , y ) = F y ( x ) − 8 π q ω ( y ) ∣ x − y ∣ ,
we obtain the asserted value and Lipschitz bounds when ∣ x − y ∣ < R / 2 |x-y| < R/2 ∣ x − y ∣ < R /2 . For ∣ x − y ∣ ≥ R / 2 |x-y| \ge R/2 ∣ x − y ∣ ≥ R /2 in the region specified in (5) , the estimates follow directly by subtracting Φ 0 \Phi_0 Φ 0 in (4) . Finally, symmetry gives the same Lipschitz estimate in the second variable. This proves (5) , including the existence of the diagonal value.
A block-operator estimate
Lemma B.2. Let V \mathcal{V} V be a Hilbert space and let A i j : V → V A_{ij}:\mathcal{V}\to\mathcal{V} A ij : V → V satisfy ∥ A i j ∥ ≤ c i j \|A_{ij}\|\le c_{ij} ∥ A ij ∥ ≤ c ij with c i j ≥ 0 c_{ij}\ge0 c ij ≥ 0 . Suppose
sup i ∑ j c i j ≤ R , sup j ∑ i c i j ≤ C 0 . \sup_i \sum_j c_{ij} \le R,\qquad\sup_j \sum_i c_{ij} \le C_0. i sup j ∑ c ij ≤ R , j sup i ∑ c ij ≤ C 0 . On the product space with norm ∥ v ∥ w 2 = w ∑ j ∥ v j ∥ V 2 \|v\|_w^2 = w\sum_j\|v_j\|_{\mathcal{V}}^2 ∥ v ∥ w 2 = w ∑ j ∥ v j ∥ V 2 , one has
∥ ( ∑ j A i j v j ) i ∥ w ≤ R C 0 ∥ v ∥ w . \left\|\left(\sum_j A_{ij}v_j\right)_i\right\|_w \le\sqrt{RC_0}\,\|v\|_w. ( j ∑ A ij v j ) i w ≤ R C 0 ∥ v ∥ w . Proof. For each row, the triangle inequality and weighted Cauchy–Schwarz give
∥ ∑ j A i j v j ∥ V 2 ≤ ( ∑ j c i j ∥ v j ∥ V ) 2 ≤ ( ∑ j c i j ) ∑ j c i j ∥ v j ∥ V 2 . \left\|\sum_j A_{ij}v_j\right\|_{\mathcal{V}}^2
\le\left(\sum_j c_{ij}\|v_j\|_{\mathcal{V}}\right)^2
\le\left(\sum_j c_{ij}\right)\sum_j c_{ij}\|v_j\|_{\mathcal{V}}^2. j ∑ A ij v j V 2 ≤ ( j ∑ c ij ∥ v j ∥ V ) 2 ≤ ( j ∑ c ij ) j ∑ c ij ∥ v j ∥ V 2 . Multiply by w w w , sum in i i i , and interchange the finite sums. The result is at most R C 0 w ∑ j ∥ v j ∥ V 2 RC_0w\sum_j\|v_j\|_{\mathcal{V}}^2 R C 0 w ∑ j ∥ v j ∥ V 2 . Taking square roots proves the assertion.
The weak normal trace
Lemma B.3. If u ∈ H 1 ( Ω ) u\in H^1(\Omega) u ∈ H 1 ( Ω ) and Δ u ∈ L 2 ( Ω ) \Delta u\in L^2(\Omega) Δ u ∈ L 2 ( Ω ) , its weak normal trace belongs to X X X and satisfies
∥ ∂ ν u ∥ X ≤ C ( ∥ u ∥ H 1 ( Ω ) + ∥ Δ u ∥ L 2 ( Ω ) ) . (168) \|\partial_\nu u\|_X \le C\left(\|u\|_{H^1(\Omega)}+\|\Delta u\|_{L^2(\Omega)}\right). \tag*{(168)} ∥ ∂ ν u ∥ X ≤ C ( ∥ u ∥ H 1 ( Ω ) + ∥Δ u ∥ L 2 ( Ω ) ) . (168)
Proof. For h ∈ Y h \in\mathcal{Y} h ∈ Y , choose a bounded H 1 H^1 H 1 extension E h E_h E h of the trace and define
⟨ ∂ ν u , h ⟩ = ∫ Ω ∇ u ⋅ ∇ ( E h ) + Δ u E h . \langle\partial_\nu u,h\rangle= \int_{\Omega} \nabla u \cdot\nabla(E_h) + \Delta u E_h. ⟨ ∂ ν u , h ⟩ = ∫ Ω ∇ u ⋅ ∇ ( E h ) + Δ u E h .
The definition is independent of the extension: the difference of two extensions belongs to H 0 1 ( Ω ) H^1_0(\Omega) H 0 1 ( Ω ) , and the distributional definition of Δ u \Delta u Δ u makes the corresponding integral vanish. Cauchy–Schwarz and ∥ E h ∥ H 1 ≤ C ∥ h ∥ Y \lVert E_h\rVert_{H^1} \le C\lVert h\rVert_{\mathcal{Y}} ∥ E h ∥ H 1 ≤ C ∥ h ∥ Y prove (B.4). For smooth functions it agrees with the classical normal derivative, by integration by parts. □ \square □