# Wave Enhancement through Optimization of Boundary Conditions

It is well known that changing boundary conditions for the Laplacian from Dirichlet to Neumann can result in significant changes to the associated eigenmodes, while keeping the eigenvalues close. We present a new and efficient approach for optimizing the transmission signal between two points in a cavity at a given frequency, by changing boundary conditions. The proposed approach makes use of recent results on the monotonicity of the eigenvalues of the mixed boundary value problem and on the sensitivity of the Green s function to small changes in the boundary conditions. The switching of the boundary condition from Dirichlet to Neumann can be performed through the use of the recently modeled concept of metasurfaces which are comprised of coupled pairs of Helmholtz resonators. A variety of numerical experiments are presented to show the applicability and the accuracy of the proposed new methodology.

## Authors

• 7 publications
• 1 publication
• 3 publications
• 2 publications
• ### On the effect of boundary conditions on the scalability of Schwarz methods

In contrast with classical Schwarz theory, recent results have shown tha...
03/27/2021 ∙ by Gabriele Ciaramella, et al. ∙ 0

• ### Frequency-based tension assessment of an inclined cable with complex boundary conditions using the PSO algorithm

The frequency-based method is the most commonly used method for measurin...
08/11/2021 ∙ by Wen-ming Zhang, et al. ∙ 0

• ### Insights into the performance of loosely-coupled FSI schemes based on Robin boundary conditions

Robin boundary conditions are a natural consequence of employing Nitsche...
05/31/2021 ∙ by Chennakesava Kadapa, et al. ∙ 0

• ### Convolutional Sparse Coding: Boundary Handling Revisited

Two different approaches have recently been proposed for boundary handli...
07/20/2017 ∙ by Brendt Wohlberg, et al. ∙ 0

• ### Reflectionless propagation of Manakov solitons on a line:A model based on the concept of transparent boundary conditions

We consider the problem of absence of backscattering in the transport of...
02/20/2021 ∙ by K. K. Sabirov, et al. ∙ 0

• ### The Finite Cell Method with Least Squares Stabilized Nitsche Boundary Conditions

We apply the recently developed least squares stabilized symmetric Nitsc...
10/27/2021 ∙ by Karl Larsson, et al. ∙ 0

• ### Robin Pre-Training for the Deep Ritz Method

We compare different training strategies for the Deep Ritz Method for el...
06/11/2021 ∙ by Luca Courte, et al. ∙ 0

##### This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

## 1 Introduction

This paper develops a new and efficient approach for maximizing the transmission signal between two points at a chosen frequency through changes to specific eigenmodes of the cavity. These changes are achieved by changing the boundary conditions. The eigenmodes and the associated eigenfrequencies of a cavity are sensitively dependant on the geometric properties of the domains, as well as the location of Dirichlet and Neumann boundary conditions. Many recent works have been devoted to the understanding of the effect of changing the boundary condition on the eigenmodes and the eigenfrequencies [1, 2, 4, 5, 11, 13, 14, 16].

Through the use of a tunable reflecting metasurface, the boundary condition can be switched from Dirichlet to Neumann at some specific resonant frequencies [3]. In [3, Part I], the physical mechanism underlying the concept of tunable metasurfaces is modeled both mathematically and numerically. It is shown that an array of coupled pairs of Helmholtz resonators behaves as an equivalent surface with Neumann boundary condition at some specific subwavelength resonant frequencies, where the size of one pair of Helmholtz resonators is much smaller than the wavelengths at the resonant frequencies. The Green’s function of a cavity with mixed (Dirichlet and Neumann) boundary conditions (called also the Zaremba function) is also characterized. In [3, Part II], a one-shot optimization algorithm is proposed and used to obtain a good initial guess for the positions around which the boundary conditions should be switched from Dirichlet to Neumann.

In this paper, we present a new methodology for maximizing the Zaremba function between two points at a chosen frequency through specific eigemodes of the cavity. The paper is organized as follows. In Section 2, we first recall some useful results on the eigenvalues of the mixed boundary value problem (called Zaremba eigenvalue problem). Of particular interest is their monotonicity property with respect to the size of the Neumann part proven in [14]. Then we reformulate the eigenvalue problem using boundary integral operators. Based on this nonlinear formulation and the use of the generalized argument principle for the characterization of the characteristic values of finitely meromorphic operator-valued functions of Fredholm-type, we derive an accurate asymptotic formula of the changes of eigenfrequencies of a cavity with mixed boundary conditions in terms of the size of the part of the cavity boundary where the boundary condition is switched from Dirichlet to Neumann. Finally, we recall the asymptotic expansion of the Zaremba function in terms of the size of the Neumann part. The problem of changing a portion of a Dirichlet boundary to Neumann is more delicate than the converse. If a portion of the boundary is changed from having Neumann conditions to Dirichlet, the reverse consideration than in this paper, then an asymptotic expansion of the eigenvalues is easier to derive [5, 16]. The perturbation theory for the introduction of Neumann boundaries requires a careful consideration of the asymptotic behaviour of the Zaremba near the perturbation [3, Part II]. In Section 3, we derive a spectral decomposition of the Zaremba function. In Section 4, we consider the problem where we have a source in a bounded domain operating at a given frequency, and we want to determine, by exploiting the monotonicity property of the eigenvalues of the mixed boundary value problem, which part of the boundary to choose to be reflecting such that an eigenvalue of the mixed boundary value problem gets close enough to the operating frequency. In order to significantly enhance the signal at a given receiving point, both the emitter and the receiver should not belong to the nodal set corresponding to the eigenmode associated with the eigenvalue of the mixed boundary value problem.

There are two distinct issues: where to place the Neumann boundary condition, and how long it should be, to achieve the twin objectives of maximizing gain between a fixed source-receiver pair, and at a frequency close to a desired one.

Our main idea is to first nucleate the Neumann boundary conditions in order to maximize gain of the Zaremba function by making use of an asymptotic expansion of the Zaremba function in terms of the size of the Neumann part. Then the size of the Neumann part is changed in such away that an eigenvalue of the mixed boundary value problem gets close to the operating frequency by using the monotonicity property of the eigenvalues of the mixed eigenvalue problem. The optimization needs the high-accuracy evaluation of certain boundary integral operators, and this is done using techniques from [1, 2].

We present in Section 5 some numerical experiments to show the applicability and the accuracy of the proposed methodology.

## 2 Preliminaries

### 2.1 Laplace Eigenvalue with Mixed Boundary Conditions

Let be an open, bounded domain with a smooth boundary. We define as the topological closure of . We decompose the boundary into two parts, , where and are finite unions of open boundary sets. We define to be a partition of . Let and . The Zaremba function is the Green’s function to the Zaremba problem, also known as the fundamental Helmholtz equation with mixed boundary conditions,

 ⎧⎪ ⎪ ⎪ ⎪⎨⎪ ⎪ ⎪ ⎪⎩(△+k2)ZkxS(xS,y)=δ0(xS−y)fory∈Ω,ZkxS(xS,y)=0fory∈ΓD,∂νyZkxS(xS,y)=0fory∈ΓN. (2.1)

Here denotes the outer normal at and the normal derivative at . It is clear that we can write

 Zk(xs,⋅)=Γk(xs,⋅)+Rk(xs,⋅)

where is the fundamental solution of the Helmholtz problem with wavenumber , and is a smooth function satisfying the boundary value problem

 (2.2)

In Section 3, we will see that exists for all but countably many values of , which are related to the unique solvability of the problem for . These exceptional values of are the eigenvalues to the associated Laplace eigenvalue problem with mixed boundary conditions

 ⎧⎪⎨⎪⎩−△u=k2uinΩ,u=0onΓD,∂νyu=0onΓN. (2.3)

Equation (2.3) has a non-trivial solution for a countable set of real values of [15, Theorem 4.10], which we refer to as , so that . We know that and that for all partitions of .

We denote by the pure Dirichlet eigenvalues for , corresponding to the case . We let denote the Neumann eigenvalues associated to the case . Then we have

 0 <λ∂Ω1, λ∂Ω1<λ∂Ω2, λ∂Ω2≤λ∂Ωj,∀j≥3, 0 =λ∅1, λ∅1<λ∅2, λ∅2≤λ∅j,∀j≥3.

In [9], it is shown that , for all , for a very general class of domains .

###### Remark

Let be the unit circle, we have that is (up to sorting) equal to

 {α2∈(0,∞)∣∃n∈N0:α is positive root of Jn(x)},

where is a Bessel function of the first kind and order . The eigenvalues corresponding to the roots of appear as simple Dirichlet eigenvalues; all others have multiplicity two. and is (up to sorting) equal to

 {α2∈(0,∞)∣∃n∈N0:α is positive root of J′n(x)}.

Again, the eigenvalues corresponding to the roots of appear as simple Neumann eigenvalues; all others have multiplicity two. We refer to [10].

Recently, Lotoreichik and Rohleder [14, Proposition 2.3] showed the following monotonicity statement.

###### Proposition

Let be two partitions of , such that . If has a non-empty interior then

 λΓDj<λΓD′j for all j∈N.

With Proposition 2.1, we can readily infer that if , and , then

 λ∅j<λΓDj<λ∂Ωj, for all j∈N.

### 2.2 Boundary Integral Formulation of the Eigenvalue Problem

The solution of the eigenvalue(2.3) can be represented by a single layer potential

 u(x)=∫∂ΩΓk(x,y)ψ(y)dσy, (2.4)

with surface density .

We define then the operators , , and by

 SkΓD[ψ](x) \coloneqq∫ΓDΓk(x,y)ψ(y)dσy, SkΓN[ψ](x)\coloneqq∫ΓNΓk(x,y)ψ(y)dσy, ∂SkΓD[ψ](x) \coloneqq∫ΓD∂νxΓk(x,y)ψ(y)dσy, (KkΓN)∗[ψ](x)\coloneqq\vbox\raisebox−3.87ptp.v.∫ΓN∂νxΓk(x,y)ψ(y)dσy,

where the ’p.v.’ stands for the principle value integral. This actually is the standard (Lebesgue-) integral for a smooth curved , since is a bounded and sufficiently smooth integral operator kernel. From [17, Chapter 11] we have that is a Fredholm operator with index 0, we also readily infer that , , and are compact operator.

We then define in terms of these integral operators through

 A(k)[ψ∣ΓDψ∣ΓN]\coloneqq⎡⎢⎣SkΓDSkΓN∂SkΓD−12IL2ω(ΓN)+(KkΓN)∗⎤⎥⎦[ψ∣ΓDψ∣ΓN]. (2.5)

We readily see that is an analytic Fredholm operator of index in .

To locate the Zaremba eigenvalues, we have the following statement:

 “The real positive characteristics values of the operator-valued function k↦A(k) are the square roots of the % Zaremba eigenvalues''. (2.6)

In [2, Section 3] and [1], it is shown that every square root of a Zaremba eigenvalue is a real positive characteristic value of and every real positive characteristic value of is the square root of a Zaremba eigenvalue.

We see that is invertible for not a square root of a Zaremba eigenvalue.

We remark that the non-real characteristic values of cannot correspond to eigenvalues to the Laplace equation. This yields the undesirable, but avoidable, difficulty in choosing a neighbourhood to apply Proposition 2.3 in our algorithm, see also Section 4, comment on Line 13.

The Statement (2.2) allows for a discretization and thus a numerical approximation of the value . We will use this further on. For these facts, we refer to [2, Sections 3 and 5].

Let us also consider the regularity of the solution and the density near a Dirichlet-Neumann junction. The following result can be found in [1, Theorems 4.2 and 4.3].

###### Proposition

Let be non-empty. Let and satisfy the Statement (2.2). Let . Then there exists a neighborhood around such that for all and for all

 u(y) =Pny⋆(z1/2,¯¯¯z1/2)+o(zn), ψ∣ΓD(y) =|z|−1/2QnD,y⋆(|z|1/2)+o(|z|n−1), ψ∣ΓN(y) =|z|−1/2QnN,y⋆(|z|1/2)+o(|z|n−1),

where is the complexification of , that is with being the imaginary unit, and being its conjugate value, and where are polynomial functions of their respective arguments and of a degree such that none of their terms can be included in their respective error terms.

### 2.3 Approximation of the Zaremba Eigenvalue using the Generalized Argument Principle

In this section we derive asymptotic expressions for the perturbation of the Zaremba eigenvalues, when a small portion of the boundary is changed from Dirichlet to Neumann.

Let be a boundary interval of length . Let be a partition of . We associate the operator , defined via (2.5), to that partition. This corresponds to having a Dirichlet boundary condition. Then we define , also by obvious changes in the integrals in (2.5), to be the operator associated to the partition . This in turn corresponds to being a Neumann part. For ease of notation, we define and for all , and call those characteristic values to their respective operators. From [6, Lemma 3.8] we then have the following Lemma:

###### Lemma

Let be a simple characteristic value. Let be a neighbourhood of , such that . Assume further that no other square root of Zaremba eigenvalue to the partition of is in . Then is given by the contour integral

 kεj−k0j=12πitr∫∂V(ω−k0j)Aε(ω)−1∂ωAε(ω)dω.

Here denotes the variation of the operator in the wavenumber parameter . This expression is exact. Unfortunately, its use in a practical algorithm is limited, since it would entail inverting the operator for each used in an optimization. It is useful, therefore, to locate an expression in which this inverse is approximated by instead.

From [6, Theorem 3.12] we get the approximation

 kεj−k0j≈−12πitr∫∂VA0(ω)−1(Aε(ω)−A0(ω))dω, (2.7)

where we expect the error to be in . We can, in fact, obtain a faster and even more accurate approximation, which we describe in the following proposition.

###### Proposition

Let be the (sorted) characteristic value of corresponding to the decomposition and assume it is simple Then one can find a and a neighbourhood containing so that

• the characteristic value of the operator (obtained by changing to a Neumann boundary condition) is contained ;

• no other square root of the Laplace eigenvalues to the partition of are in .

• The characteristic value of the perturbed operator is given by

 kεj−k0j=−12πitr∫∂V(I+(ω−k0j)Aε(k0j)−1∂ωAε(k0j))−1dω ×[1+O(tr∫∂V(I+(ω−k0j)Aε(k0j)−1∂ωAε(k0j))−1dω)].

Here is the identity operator.

###### Proof

We first observe from Proposition 2.1 together with the fact that is a Fredholm analytic operator of index in , we can see that for . We now examine the perturbed operator . Its characteristic value is . Provided is sufficiently close to , we have the following Taylor expansion:

 Aε(ω)=Aε(k0j)+(ω−k0j)∂ωAε(k0j)+Bε(ω), (2.8)

where . This expansion holds only in a neighborhood of , and so must be small enough such that .

Then consider that we have in the operator-norm

 ∥∥∥(Aε(k0j)+(ω−k0j)∂ωAε(k0j))−1Bε(ω)∥∥∥<1,

for close enough to both and , because then the Taylor remainder . If is small enough, then . Then by the Generalization of Rouché’s Theorem [6, Theorem 1.15] we have that since and are close in operator norm, they both have the same number of characteristic values in . Thus has a simple characteristic value in . Now we can use Lemma 2.3, but replacing by :

to get

 k♯j−k0j =12πitr∫∂Vε(ω−k0j)(Aε(k0j)+(ω−k0j)∂ωAε(k0j))−1 =12πitr∫∂Vε(ω−k0j)(Aε(k0j)+(ω−k0j)∂ωAε(k0j))−1∂ωAε(k0j)dω,

and hence,

 k♯j−k0j =12πitr∫∂Vε(Aε(k0j)+(ω−k0j)∂ωAε(k0j))−1 ×(Aε(k0j)+(ω−k0j)∂ωAε(k0j)−Aε(k0j))dω =12πitr(∫∂VεIdω−∫∂Vε(Aε(k0j)+(ω−k0j)∂ωAε(k0j))−1Aε(k0j)dω) =−12πitr∫∂Vε(Aε(k0j)+(ω−k0j)∂ωAε(k0j))−1Aε(k0j)dω.

Moreover, by a standard perturbation argument [6, Section 5.2.4], we have at the leading-order term

 kεj−k♯j=−(Bε(k♯j)ψ♯j,ψ♯j),

where is the root function associated with the characteristic value evaluated at . Thus,

 kεj−k0j=(k♯j−k0j)(1+O(k♯j−k0j)),

and therefore, Proposition 2.3 holds.

We remark on the significance of this result from the point of view of computation, and which makes it a key ingredient in our algorithm. If one seeks a high-accuracy approximation of the characteristic value of , and one already has a good approximation of , the approximation in Proposition 2.5 allows us to proceed by assembling only one matrix, that corresponding to . The contour integrals can be effectively computed using the trapezoidal rule, making this an inexpensive but very accurate approximation of .

### 2.4 Approximation of the Zaremba Function

Let , and let be a boundary interval of length with center . Let be the partition of , and with it we associate the Zaremba function , for , defined via (2.1). This corresponds to having a Dirichlet boundary condition. Then we define , , also defined via (2.1), to be the Zaremba function associated to the partition . This in turn corresponds to having a Neumann boundary condition. We then have the following lemma.

###### Lemma

Let and be defined as described above. Let be small enough. Let , such that , and , for all . Then for all ,

Lemma 2.4 follows readily from combining the results in [3, Theorem 5.4] and [3, Equation (6.24)]

Numerical experiments confirm that is of order of , as long as is far enough away from the boundary.

## 3 Spectral Decomposition of the Zaremba Function

Let us again consider the more general setup at the beginning of Section 2, that is let be a partition of , let be the Zaremba eigenvalues and let be an

-orthonormal basis of associated eigenfunctions. Then we have the following statement about the Zaremba function

, , defined by (2.1).

###### Theorem

For all , and for all which are not in the spectrum, ie, of the Zaremba eigenvalue problem, the Zaremba function , given by (2.1), exists and is in . Furthermore, we can write it as

 ZkxS(y)=∞∑j=1uj(xS)uj(y)k2−λΓDj.

Next, we will consider the proof of Theorem 3. To this end, we define

 H10,ΓD(Ω)\coloneqq{v∈H1(Ω)∣v∣ΓD=0}.

Consider that the solution to the Laplace eigenvalue-equation is element of .

 dom(−△)\coloneqq{w∈H10,ΓD(Ω)∣△w∈L2(Ω),∂νw∣ΓN=0}.

The operator is selfadjoint in , which we readily see using Green’s identity, and it has thus a discrete spectrum. Moreover, corresponds to the sesquilinear form with domain , since for all , see [7, 12, 18] for more details on semi-bounded self-adjoint operators and corresponding quadratic forms. And the form is closed, non-negative and symmetric. This allows us to use the min-max principle. Thus we can write for all ,

 (3.1)

This leads us to the following lemma.

###### Lemma

For all , we have that

 ∥∥ ∥∥f−N∑j=1cjuj∥∥ ∥∥2L2(Ω)=∫Ω∣∣f−N∑j=1cjuj∣∣2dxN→∞−−−−→0, (3.2)

where , that is the linear subset spanned by eigenfunctions of the Laplace eigenvalue-equation with mixed boundary conditions (2.3) is dense in .

###### Proof

Let . Then for all , we have that

 (rN,ui)L2(Ω) =(f−N∑j=1cjuj,ui)L2=(f,ui)L2−ci(ui,uj)L2=0, (∇rN,∇ui)L2(Ω) =(∇f,∇ui)L2−N∑j=1cj(∇uj,∇ui)L2 =λΓDi(f,ui)L2−λΓDicj(ui,ui)L2=0,

where we used Green’s identity and the fact that . Next, we want to show that

 λΓDN≤∥∇rN∥2L2(Ω)∥rN∥2L2(Ω). (3.3)

To this end, consider the min-max principle (3.1), it tells us that

 λΓDj

Thus, we can infer from , which in turn is given by an induction argument, whose induction basis follows trivially from the min-max principle (3.1). Using the definition of , we have that

 ∥∇rN∥2L2(Ω) =∥∇f∥2L2(Ω)−N∑j=1λΓDj(f,uj)2L2(Ω) ≤∥∇f∥2L2(Ω).

Thus, using (3.3), we have that

 ∥rN∥2L2≤∥∇f∥2L2λΓDN. (3.4)

Since , is bounded. Using the fact that , we have that . This completes the proof of Lemma 3.

###### Proof (Theorem 3)

To show the existence of the Zaremba function , we write , for all as

 ZkxS(y)=Γk(xS,y)+Rk(xS,y), (3.5)

where is the fundamental solution to the Helmholtz equation, and satisfies

 ⎧⎪ ⎪⎨⎪ ⎪⎩(△+k2)Rk(xS,y)=0inΩ,Rk(xS,y)=−Γk(xS,y)onΓD,∂νyRk(xS,y)=−∂νyΓk(xS,y)onΓN. (3.6)

The solution to (3.6) does exist, for those values of specified in the theorem, and it is in , see [15, Theorem 4.10]. Using that , we have that . Thus from Lemma 3 and the density of in , we have that for all , ,

 ZkxS(y)=∞∑j=1ajuj(y),

for some , depending on . Let us give an expression for the . Using Green’s identity, we have that

 ui(xS) =∫Ω(△+k2)ZkxS(y)ui(y)dy=∫ΩZkxS(y)(△+k2)ui(y)dy =(k2−λΓDi)∫ΩZkxS(y)ui(y)dy=(k2