betaFeedback

WorksMathematics

A spectral-element method for computing many eigenvalues and their asymptotics of the Schrödinger operator with Robin boundary condition

Abstract

We propose spectral and spectral-element methods for accurately computing many eigenvalues of the Schrödinger operator with Robin boundary condition on simple and complex geometries, respectively. Due to their spectral-type accuracy in approximating those eigenfunctions corresponding to high-index eigenvalues, which are usually highly oscillatory, the proposed approaches have excellent resolution in computing thousands of eigenvalues accurately and efficiently with the resolution property that the number of eigenvalues with reasonable accuracy is proportional to the number of degrees of freedom. Using a standard HPC node with 128 GB of memory, we can obtain numerically more than 5,000 reliable eigenvalues with relative errors below 10−810^{-8} for different complex domains in two dimensions (2D). Based on these computed eigenvalues, we first confirm some theoretical results on the Robin-to-Neumann (RtN) gaps of the Laplacian operator in 2D, which were recently studied by Rudnick et al. [Comm. Math. Phys. 388 (2021)]. Then we systematically study the RtN gaps of the Schrödinger operator and their convergence rates. Based on our extensive numerical results, we formulate a unified conjecture on the cumulative averages of the RtN gaps of the Schrödinger operator.

Introduction

Consider the following eigenvalue problem of the Schrödinger operator with Robin boundary condition, i.e. find λ∈R\lambda\in\mathbb{R} and a real-valued function u:=u(x)≠0u := u(x) \ne0 such that

LSOu:=[−Δ+V(x)]u(x)=λu(x),x∈Ω,(1)L_{\mathrm{SO}}u := \left[-\Delta+ V(\mathbf{x})\right]u(\mathbf{x}) = \lambda u(\mathbf{x}), \qquad\mathbf{x} \in\Omega, \tag*{(1)}
∂u(x)∂n+σu(x)=0,x∈∂Ω,(2)\frac{\partial u(\mathbf{x})}{\partial n} + \sigma u(\mathbf{x}) = 0, \qquad\mathbf{x} \in\partial\Omega, \tag*{(2)}

where x∈Rd\mathbf{x} \in\mathbb{R}^{d} is the spatial coordinate, Ω⊂Rd\Omega\subset\mathbb{R}^{d} is a connected bounded domain with piecewise smooth boundary, n\mathbf{n} is the unit outward normal vector, σ∈R\sigma\in\mathbb{R} is a given constant, and V:=V(x)∈L∞(Ω)V := V(\mathbf{x}) \in L^{\infty}(\Omega) is a given real-valued function which models external fields, trapping effects, impurities or material inhomogeneities.

The Schrödinger operator LSO:=−Δ+V(x)L_{\mathrm{SO}} := -\Delta+ V(\mathbf{x}) is one of the most fundamental operators in quantum mechanics, spectral theory and mathematical physics [4, 5, 16, 35]. The eigenvalues of the Schrödinger operator represent the admissible energy levels of a quantum system, while the corresponding eigenfunctions describe the associated stationary states or wave modes. When σ=0\sigma= 0, the boundary condition (2) collapses to the homogeneous Neumann boundary condition; and when σ→+∞\sigma\to+\infty, it collapses to the homogeneous Dirichlet boundary condition u(x)∣∂Ω=0u(x)|_{\partial\Omega} = 0 [7]. In fact, different boundary conditions encode different interactions between eigenfunctions and the boundary: (i) the homogeneous Dirichlet boundary condition corresponds to perfectly confining or impenetrable walls, (ii) the homogeneous Neumann boundary condition describes zero-flux or reflecting boundaries, and (iii) the Robin boundary condition models more general boundary interactions, such as partially reflecting walls, surface potentials and impedance-type effects [13]. The Schrödinger operator (1) with different boundary conditions leads to different spectra and distinct physical interpretations. Thus understanding the eigenvalues of the Schrödinger operator (1) with either homogeneous Dirichlet or homogeneous Neumann or Robin boundary condition is important in quantum confinement, wave propagation, resonant modes, spectral geometry and related problems in mathematical physics.

Without loss of generality, we assume that V(x)V(\mathbf{x}) is non-negative, i.e. V(x)≥0V(\mathbf{x}) \ge0 for x∈Ω\mathbf{x} \in\Omega. Since the self-adjoint Schrödinger operator with the Robin boundary condition has compact resolvent, its spectrum is real and discrete [2]. We can rank (or order) all eigenvalues (counting multiplicities) {λnσ∣n=1,2,…}\{\lambda_n^\sigma\mid n = 1,2,\ldots\} (with the superscript σ\sigma referring to the constant in the Robin boundary condition) of (1)-(2) as

λ1σ≤λ2σ≤⋯≤λnσ≤λn+1σ≤⋯(3)\lambda_1^\sigma\le\lambda_2^\sigma\le\cdots\le\lambda_n^\sigma\le\lambda_{n+1}^\sigma\le\cdots \tag*{(3)}

where the number of times that an eigenvalue λ\lambda of (1)-(2) appears in the above sequence (3) is the same as its algebraic multiplicity. Here we adopt {λn0∣n=1,2,…}\{\lambda_n^0 \mid n = 1,2,\ldots\} and {λn∞∣n=1,2,…}\{\lambda_n^\infty\mid n = 1,2,\ldots\} to denote the corresponding eigenvalues of the Schrödinger operator (1) with homogeneous Neumann (σ=0\sigma= 0 in (2)) and homogeneous Dirichlet (σ=+∞\sigma= +\infty in (2)) boundary conditions, respectively. Define the Robin-to-Neumann (RtN) gaps of the Schrödinger operator (1)-(2) as

gn(σ):=λnσ−λn0,n=1,2,…,(4)g_n(\sigma) := \lambda_n^\sigma- \lambda_n^0,\qquad n = 1,2,\ldots, \tag*{(4)}

and define the cumulative averages of the RtN gaps as

Gn(σ):=1n∑m=1ngm(σ)=1n∑m=1n(λmσ−λm0),n=1,2,….(5)G_n(\sigma) := \frac{1}{n}\sum_{m=1}^{n} g_m(\sigma) = \frac{1}{n}\sum_{m=1}^{n}(\lambda_m^\sigma-\lambda_m^0),\qquad n = 1,2,\ldots. \tag*{(5)}

When V(x)≡0V(\mathbf{x}) \equiv0 in (1), the Schrödinger operator LSOL_{\mathrm{SO}} collapses to the Laplacian operator. In this case, under suitable regularity assumptions on the domain Ω\Omega, for any fixed n∈Nn \in\mathbb{N}, one has [7, 26]

λnσ=λn0+O(∣σ∣),as σ→0;λnσ=λn∞+O(∣σ∣−1),as σ→+∞.(6)\lambda_n^\sigma= \lambda_n^0 + O(|\sigma|),\quad\text{as }\sigma\to0;\qquad\lambda_n^\sigma= \lambda_n^\infty+ O\left(|\sigma|^{-1}\right),\quad\text{as }\sigma\to+\infty. \tag*{(6)}

In addition, when the domain Ω\Omega is a simple geometry, all the eigenvalues of (1)-(2) can be obtained explicitly. Specifically, in one dimension (1D), e.g. taking d=1d = 1, V(x)≡0V(\mathbf{x}) \equiv0 and Ω=(0,π)\Omega= (0,\pi) in (1)-(2), all the eigenvalues are given as [15]

  • for the Neumann boundary condition, i.e. σ=0\sigma= 0,

λn0=(n−1)2,n=1,2,…;(7)\lambda_n^0 = (n-1)^2,\qquad n = 1,2,\ldots; \tag*{(7)}
  • for the Dirichlet boundary condition, i.e. σ=+∞\sigma= +\infty,

λn∞=n2,n=1,2,…;(8)\lambda_n^\infty= n^2,\qquad n = 1,2,\ldots; \tag*{(8)}
  • for the Robin boundary condition, i.e. σ≠0\sigma\ne0,

λnσ=κn2(σ)π2={2πσ+O(σ2),n=1,σ→0,(n−1)2+4πσ+O(σ2),n≥2,σ→0,n2−4n2πσ+O(σ−2),n≥1,σ→+∞;(9)\lambda_n^\sigma= \frac{\kappa_n^2(\sigma)}{\pi^2} = \begin{cases} \frac{2}{\pi}\sigma+ O(\sigma^2), & n=1,\quad\sigma\to0,\\ (n-1)^2+\frac{4}{\pi}\sigma+O(\sigma^2), & n\ge2,\quad\sigma\to0,\\ n^2-\frac{4n^2}{\pi\sigma}+O(\sigma^{-2}), & n\ge1,\quad\sigma\to+\infty; \end{cases} \tag*{(9)}

where, when nn is odd, κn(σ)\kappa_n(\sigma) is the unique solution in ((n−1)π,nπ)((n-1)\pi,n\pi) of the nonlinear equation κn(σ)tan⁡(κn(σ)2)−σπ=0\kappa_n(\sigma)\tan\left(\frac{\kappa_n(\sigma)}{2}\right)-\sigma\pi=0, and; when nn is even, it is the unique solution in ((n−1)π,nπ)((n-1)\pi,n\pi) of the nonlinear equation κn(σ)+σπtan⁡(κn(σ)2)=0\kappa_n(\sigma)+\sigma\pi\tan\left(\frac{\kappa_n(\sigma)}{2}\right)=0. The first case in (9) is separated because the first Robin eigenvalue bifurcates from the zero Neumann eigenvalue.

Combining (7), (8) and (9), (6) is confirmed immediately in this special case. In addition, for any σ∈R\sigma\in\mathbb{R}, we can get

lim⁡n→∞gn(σ)=lim⁡n→∞(λnσ−λn0)=4πσ,\lim_{n\to\infty}g_n(\sigma)=\lim_{n\to\infty}(\lambda_n^\sigma-\lambda_n^0)=\frac{4}{\pi}\sigma,
lim⁡n→∞Gn(σ)=lim⁡n→∞1n∑m=1n(λmσ−λm0)=4πσ.(10)\lim_{n\to\infty}G_n(\sigma)=\lim_{n\to\infty}\frac{1}{n}\sum_{m=1}^{n}(\lambda_m^\sigma-\lambda_m^0)=\frac{4}{\pi}\sigma. \tag*{(10)}

In two dimensions (2D), e.g. taking d=2d=2, V(x)≡0V(\mathbf{x})\equiv0 and Ω⊂R2\Omega\subset\mathbb{R}^2 to be a connected bounded domain with piecewise smooth boundary in (1.1)-(1.2), Rudnick et al. [31] studied the RtN gaps and their cumulative averages. Based on analytical and numerical results for rectangles, the hemisphere, and the equilateral triangle [28, 29, 30], for each fixed σ>0\sigma>0, they proved the following result [31]

lim⁡n→∞Gn(σ)=lim⁡n→∞1n∑m=1n(λmσ−λm0)=2∣∂Ω∣∣Ω∣σ.(11)\lim_{n\to\infty}G_n(\sigma)=\lim_{n\to\infty}\frac{1}{n}\sum_{m=1}^{n}(\lambda_m^\sigma-\lambda_m^0)=\frac{2|\partial\Omega|}{|\Omega|}\sigma. \tag*{(11)}

where ∣Ω∣|\Omega| and ∣∂Ω∣|\partial\Omega| denote the area of Ω\Omega and the length of ∂Ω\partial\Omega, respectively. The above formula (11) reveals an interesting connection between an algebraic spectral quantity, i.e. the cumulative averages of the RtN gaps, and the geometric quantities of the domain Ω\Omega, i.e., its area and boundary length.

In general domains, especially for the Schrödinger operator with nontrivial potentials, exact eigenvalues are rarely available, thus efficient and accurate as well as high resolution numerical approximation is necessary and important as well as challenging. In fact, the finite element method provides a mature and geometrically flexible framework, supported by general Galerkin spectral approximation theory [1, 2, 6]. Most concrete algorithms and analyses have been developed for computing the eigenvalues of the Schrödinger operator with homogeneous Dirichlet boundary condition, including adaptive eigensolvers [9, 11, 12, 40], two-grid schemes [21, 39, 41], multilevel correction methods [24], and rigorous error estimates and eigenvalue bounds [10, 22]. For computing the eigenvalues of the Schrödinger operator with homogeneous Neumann condition, finite element error analysis on curved domains was given in [20]. For computing the eigenvalues of the Schrödinger operator with Robin boundary condition, far fewer results are available in the literature [25, 36]. Recently, machine learning methods have also been developed for computing eigenvalues of Schrödinger and fractional Schrödinger operators [17, 18]. Specifically, all these existing works focus mainly on algorithmic and theoretical analysis, or on a relatively small number of low-index eigenpairs.

In order to study the Weyl law on the asymptotics of the eigenvalues of the Schrödinger operator with Robin boundary condition (1.1)-(1.2) and the asymptotics of the Robin-to-Neumann (RtN) gaps in (1.4) and (1.5), one needs to compute numerically thousands of eigenvalues with high accuracy. To achieve this very challenging computational task, under a fixed number of degrees of freedom (DOFs), low-order finite element or finite difference discretizations typically resolve only a limited portion of the spectrum [3, 32, 42], and thus they cannot handle it successfully. When the domain Ω\Omega is a simple geometry, e.g. a rectangle or a disc in 2D, and a cube or a ball in 3D, spectral methods can provide a much larger range of accurate high-index eigenvalues with the resolution of the number of computed eigenvalues being proportional to the degrees of freedom of the corresponding discretization [3, 19, 23, 38, 42]. In contrast, when the domain Ω\Omega is complicated, spectral methods cannot be extended directly. As we know, spectral-element methods can address this geometric limitation by combining high-order spectral approximation with element-wise decomposition and geometric mappings [14, 27, 34]. Representative applications to eigenvalue problems include the Schrödinger operator with singular potentials and homogeneous Dirichlet boundary condition [23, 37], and the Stokes eigenvalue problems with homogeneous Dirichlet boundary condition [33].

To the best of our knowledge, efficient and accurate numerical methods for computing many eigenvalues of the Schrödinger operator with Robin boundary condition are very limited in the literature. The purpose of this paper is: (i) to develop spectral and spectral-element methods for accurately computing many eigenvalues of the Schrödinger operator with Robin boundary condition on simple and complex geometries, respectively; and (ii) to adopt the proposed numerical method for obtaining numerically more than 5,000 reliable eigenvalues with relative errors below 10−810^{-8} for different complex domains in two dimensions (2D) using a standard HPC node with 128 GB of memory and thus to study numerically the Robin-to-Neumann gaps and their asymptotics. Specifically, based on our extensive numerical results and observations, we speculate the following:

RtN Gap Conjecture (Robin-to-Neumann gaps of the Schrödinger operator (1.1)-(1.2)) Assume that Ω⊂Rd\Omega\subset\mathbb{R}^{d} is a connected bounded domain with piecewise smooth boundary and V∈L∞(Ω)V \in L^{\infty}(\Omega). For any σ∈R\sigma\in\mathbb{R}, we have

lim⁡n→∞Gn(σ)=2∣∂Ω∣∣Ω∣σ,Gn(σ)=2∣∂Ω∣∣Ω∣σ+O(n−1/d),n≫1.(12)\lim_{n\to\infty}G_n(\sigma)=\frac{2|\partial\Omega|}{|\Omega|}\sigma,\qquad G_n(\sigma)=\frac{2|\partial\Omega|}{|\Omega|}\sigma+O(n^{-1/d}),\qquad n\gg1. \tag*{(12)}

In addition, when d=1d=1 and Ω=(a,b)\Omega=(a,b), we also have

lim⁡n→∞gn(σ)=lim⁡n→∞(λnσ−λn0)=4b−aσ,λnσ=λn0+4b−aσ+O(n−2),n≫1.(13)\lim_{n\to\infty}g_n(\sigma)=\lim_{n\to\infty}\left(\lambda_n^\sigma-\lambda_n^0\right)=\frac{4}{b-a}\sigma,\qquad\lambda_n^\sigma=\lambda_n^0+\frac{4}{b-a}\sigma+O(n^{-2}),\qquad n\gg1. \tag*{(13)}

The rest of the paper is organized as follows. In Section 2, we introduce the scaling property and the variational formulation of the Schrödinger eigenvalue problem (1.1)-(1.2). In Section 3, we describe the spectral and spectral-element methods for solving the problem (1.1)-(1.2) on simple and complex geometries, respectively. Section 4 is devoted to accuracy tests and comparisons of different combinations of the number of sub-elements and the polynomial degree of basis functions under fixed numbers of DOFs. In Section 5, we investigate the Robin-to-Neumann gaps in 1D, 2D and three dimensions (3D). Finally, some concluding remarks are given in Section 6.

Scaling property and the variational formulation

In this section, we present a scaling property of the Schrödinger eigenvalue problem (1.1)-(1.2) and its variational (weak) formulation.

Scaling property

Denote

R:=diam⁡(Ω)=max⁡x,z∈Ω‾∣x−z∣,y:=xR,Ω~:={y:=x/R∣x∈Ω},R := \operatorname{diam}(\Omega) = \max_{x,z\in\overline{\Omega}} |\mathbf{x}-\mathbf{z}|,\qquad\mathbf{y}:=\frac{\mathbf{x}}{R},\qquad\widetilde{\Omega}:=\{\mathbf{y}:=\mathbf{x}/R\mid\mathbf{x}\in\Omega\},
V~(y):=R2V(x)=R2V(Ry),v(y):=u(x)=u(Ry),λ~=R2λ,σ~=Rσ.(14)\widetilde{V}(\mathbf{y}):=R^2V(\mathbf{x})=R^2V(R\mathbf{y}),\qquad v(\mathbf{y}):=u(\mathbf{x})=u(R\mathbf{y}),\qquad\widetilde{\lambda}=R^2\lambda,\qquad\widetilde{\sigma}=R\sigma. \tag*{(14)}

Plugging (2.1) into the Schrödinger eigenvalue problem (1.1)-(1.2), we obtain the following Schrödinger eigenvalue problem defined on a bounded domain Ω~\widetilde{\Omega} with unit diameter, i.e. find λ~∈R\widetilde{\lambda}\in\mathbb{R} and a real-valued function v:=v(y)≢0v:=v(\mathbf{y})\not\equiv0 such that

[−Δ+V~(y)]v(y)=λ~v(y),y∈Ω~,(15)\left[-\Delta+\widetilde{V}(\mathbf{y})\right]v(\mathbf{y})=\widetilde{\lambda}v(\mathbf{y}),\qquad\mathbf{y}\in\widetilde{\Omega}, \tag*{(15)}
∂v(y)∂n+σ~v(y)=0,y∈∂Ω~.(16)\frac{\partial v(\mathbf{y})}{\partial n}+\widetilde{\sigma}v(\mathbf{y})=0,\qquad\mathbf{y}\in\partial\widetilde{\Omega}. \tag*{(16)}

Then it is straightforward to have the following lemma:

Lemma 1. Let λ~\widetilde{\lambda} be an eigenvalue of (2.2)-(2.3) and v:=v(y)v:=v(\mathbf{y}) be the corresponding eigenfunction, than λ=λ~/R2\lambda=\widetilde{\lambda}/R^2 is an eigenvalue of (1.1)-(1.2) and u:=u(x)=v(y)=v(x/R)u:=u(\mathbf{x})=v(\mathbf{y})=v(\mathbf{x}/R) is the corresponding eigenfunction, and the converse also holds.

From this lemma, we can easily obtain that: assume that λ~1σ~≤λ~2σ~≤⋯≤λ~nσ~≤⋯\widetilde{\lambda}^{\widetilde{\sigma}}_1\leq\widetilde{\lambda}^{\widetilde{\sigma}}_2\leq\cdots\leq\widetilde{\lambda}^{\widetilde{\sigma}}_n\leq\cdots are all eigenvalues of (2.2)-(2.3), then λ1σ≤λ2σ≤⋯≤λnσ≤⋯\lambda^\sigma_1\leq\lambda^\sigma_2\leq\cdots\leq\lambda^\sigma_n\leq\cdots are all eigenvalues of (1.1)-(1.2) with λ=λ~/R2\lambda=\widetilde{\lambda}/R^2 and σ=σ~/R\sigma=\widetilde{\sigma}/R.

A variational (weak) formulation

Multiplying a test function w∈H1(Ω)w\in H^1(\Omega) in both sides of (1.1) and then integrating over Ω\Omega and using integration by parts, we obtain a variational formulation of the problem (1.1)-(1.2) as: find λ∈R\lambda\in\mathbb{R} and a real-valued function 0≠u∈H1(Ω)0\neq u\in H^1(\Omega) such that

aσ(u,w)=λb(u,w),∀w∈H1(Ω),(17)a_\sigma(u,w)=\lambda b(u,w),\qquad\forall w\in H^1(\Omega), \tag*{(17)}

where

aσ(u,w)=∫Ω(∇u⋅∇w+Vuw) dx+σ∫∂Ωuw ds,b(u,w)=∫Ωuw dx,∀w∈H1(Ω).(18)a_\sigma(u,w)=\int_\Omega(\nabla u\cdot\nabla w+Vuw)\,\mathrm{d}x+\sigma\int_{\partial\Omega}uw\,\mathrm{d}s,\qquad b(u,w)=\int_\Omega uw\,\mathrm{d}x,\qquad\forall w\in H^1(\Omega). \tag*{(18)}

When σ≥0\sigma\geq0, then aσ(v,v)≥a0(v,v)a_\sigma(v,v)\geq a_0(v,v); and respectively, when σ<0\sigma<0, from the trace inequality [8], ∀ ε>0\forall\ \varepsilon>0, we have

∥u∥L2(∂Ω)2≤C∥u∥L2(Ω)∥u∥H1(Ω)≤Cε∥u∥L2(Ω)2+ε∥∇u∥L2(Ω)2.(19)\lVert u\rVert_{L^2(\partial\Omega)}^2\leq C\lVert u\rVert_{L^2(\Omega)}\lVert u\rVert_{H^1(\Omega)}\leq C_\varepsilon\lVert u\rVert_{L^2(\Omega)}^2+\varepsilon\lVert\nabla u\rVert_{L^2(\Omega)}^2. \tag*{(19)}

Since ε>0\varepsilon>0 can be chosen sufficiently small in (2.6), the negative boundary term in aσ(v,v)a_\sigma(v,v) can be controlled by ∥∇u∥L2(Ω)2\lVert\nabla u\rVert_{L^2(\Omega)}^2. Assume V∈L∞(Ω)V\in L^\infty(\Omega), then the smallest eigenvalue of (2.4) is bounded below and all eigenvalues of (2.4) are discrete. Specifically, from the Rayleigh quotient [6], the smallest eigenvalue λ1σ\lambda^\sigma_1 is defined as

λ1σ=min⁡0≠v∈H1(Ω)aσ(v,v)b(v,v)=min⁡v∈H1(Ω)b(v,v)=∥v∥L2(Ω)2=1aσ(v,v).(20)\lambda^\sigma_1=\min_{0\neq v\in H^1(\Omega)}\frac{a_\sigma(v,v)}{b(v,v)} =\min_{\substack{v\in H^1(\Omega)\\ b(v,v)=\lVert v\rVert_{L^2(\Omega)}^2=1}}a_\sigma(v,v). \tag*{(20)}

For any n≥1n \ge1, assume that u1(x),…,un(x)u_1(x),\ldots,u_n(x) being the orthonormal eigenfunctions corresponding to the eigenvalues λ1σ≤…≤λnσ\lambda_1^\sigma\le\ldots\le\lambda_n^\sigma, and define Un=span⁡{u1,…,un}U_n=\operatorname{span}\{u_1,\ldots,u_n\}, then from the Rayleigh quotient [6], the next eigenvalue λn+1σ\lambda_{n+1}^\sigma is defined as

λn+1σ=min⁡0≠v∈Un⊥aσ(v,v)b(v,v)=min⁡En+1⊂H1(Ω)0≠v∈En+1max⁡aσ(v,v)b(v,v).(21)\lambda_{n+1}^\sigma=\min_{0\ne v\in U_n^\perp}\frac{a_\sigma(v,v)}{b(v,v)}=\min_{\substack{E_{n+1}\subset H^1(\Omega)\\0\ne v\in E_{n+1}}}\max\frac{a_\sigma(v,v)}{b(v,v)}. \tag*{(21)}

where the dimension of the subspace En+1E_{n+1} is n+1n+1. Combining eq:2.5, eq:2.7 and (21), we obtain immediately: (i) when σ>0\sigma>0, we have

λn0≤λnσ≤λn∞,n=1,2,…;(22)\lambda_n^0\le\lambda_n^\sigma\le\lambda_n^\infty,\qquad n=1,2,\ldots; \tag*{(22)}

and respectively, (ii) when σ<0\sigma<0, we have

λnσ≤λn0≤λn∞,n=1,2,….(23)\lambda_n^\sigma\le\lambda_n^0\le\lambda_n^\infty,\qquad n=1,2,\ldots. \tag*{(23)}

Spectral and spectral-element methods

In this section, we briefly introduce the spectral and spectral-element methods for solving the eigenvalue problem (1)-(2) on simple and complex/irregular geometries, respectively. For a more detailed description of the implementation, we refer the reader to Chapter 4 of [34] and Section 4.5 of [14].

A Galerkin approximation

Let XN=span⁡{ϕ1,…,ϕN}⊂H1(Ω)X_N=\operatorname{span}\{\phi_1,\ldots,\phi_N\}\subset H^1(\Omega) be a conforming finite-dimensional subspace with the degree of freedoms (DOFs) as NN. Then a Galerkin approximation of eq:2.4 reads: find λN∈R\lambda_N\in\mathbb{R} and a real-valued function 0≠uN∈XN0\ne u_N\in X_N such that

aσ(uN,wN)=λNb(uN,wN),∀wN∈XN.(24)a_\sigma(u_N,w_N)=\lambda_N b(u_N,w_N),\qquad\forall w_N\in X_N. \tag*{(24)}

Assume

uN=∑j=1Nujϕj=ΦTU∈XN,with Φ=(ϕ1,…,ϕN)T,U=(u1,…,uN)T∈RN,(25)u_N=\sum_{j=1}^{N}u_j\phi_j=\Phi^T U\in X_N,\quad\text{with }\Phi=(\phi_1,\ldots,\phi_N)^T,\quad U=(u_1,\ldots,u_N)^T\in\mathbb{R}^N, \tag*{(25)}

Plugging (25) into (24) and taking wN=ϕiw_N=\phi_i for i=1,…,Ni=1,\ldots,N respectively, we obtain the following generalized matrix eigenvalue problem

AU=λNMU,(26)A U=\lambda_N M U, \tag*{(26)}

where A=(aij)∈RN×NA=(a_{ij})\in\mathbb{R}^{N\times N} and M=(mij)∈RN×NM=(m_{ij})\in\mathbb{R}^{N\times N} are two N×NN\times N matrices given by

aij=aσ(ϕj,ϕi),mij=b(ϕj,ϕi),i,j=1,…,N.(27)a_{ij}=a_\sigma(\phi_j,\phi_i),\qquad m_{ij}=b(\phi_j,\phi_i),\qquad i,j=1,\ldots,N. \tag*{(27)}

Then spectral and spectral-element methods can be designed by properly choosing the basis functions {ϕ1,…,ϕN}\{\phi_1,\ldots,\phi_N\}, efficiently and accurately evaluating the integrals aσ(ϕj,ϕi)a_\sigma(\phi_j,\phi_i) and b(ϕj,ϕi)b(\phi_j,\phi_i), and properly assembling the two matrices AA and MM. Based on different domains Ω\Omega, we will present details in the next two subsections.

Spectral methods for simple geometries

When the domain Ω\Omega is a simple geometry, e.g. an interval in 1D, a rectangle or a disk in 2D, a cube or a ball or a cylinder in 3D, spectral methods can be easily designed for the Galerkin approximation (3.1).

For simplicity of notation, here we only present the spectral method for Ω=(−1,1)\Omega=(-1,1) in 1D. In this case, we adopt the Legendre polynomials to construct the basis in XNX_N. Define

ϕ1(x)=1−x2,ϕ2(x)=1+x2,ϕi(x)=Li−1(x)−Li−3(x), i≥3,x∈[−1,1],(28)\phi_1(x)=\frac{1-x}{2},\qquad\phi_2(x)=\frac{1+x}{2},\qquad\phi_i(x)=L_{i-1}(x)-L_{i-3}(x),\ i\ge3,\quad x\in[-1,1], \tag*{(28)}

where LjL_j denotes the Legendre polynomial of degree j≥0j\ge0 [34], then we can construct a subspace XN=span⁡{ϕ1,…,ϕN}⊂H1(Ω)X_N=\operatorname{span}\{\phi_1,\ldots,\phi_N\}\subset H^1(\Omega) with spectral accuracy. Plugging (3.5) into (3.1), (2.5) and (3.4), we obtain the corresponding generalized matrix eigenvalue problem (3.3) with

aij=∫−11[ϕj′(x)ϕi′(x)+V(x)ϕj(x)ϕi(x)] dx+σ[ϕj(−1)ϕi(−1)+ϕj(1)ϕi(1)],mij=∫−11ϕj(x)ϕi(x) dx,i,j=1,…,N.(29)\begin{aligned} a_{ij}&=\int_{-1}^{1}\left[\phi_j'(x)\phi_i'(x)+V(x)\phi_j(x)\phi_i(x)\right]\,\mathrm{d}x+\sigma\left[\phi_j(-1)\phi_i(-1)+\phi_j(1)\phi_i(1)\right],\\ m_{ij}&=\int_{-1}^{1}\phi_j(x)\phi_i(x)\,\mathrm{d}x,\qquad i,j=1,\ldots,N. \tag*{(29)} \end{aligned}

In practical computations, the above integrals will be further approximated by numerical quadrature (pseudospectral method) [34].

The above spectral method for discretizing the eigenvalue problems (2.4) can be extended straightforwardly to a rectangle or a disk in 2D, and a cube or a ball or a cylinder in 3D, by adopting the standard spectral basis functions in the literature [34].

Spectral-element methods for complex geometries

For complex/irregular geometries in 2D and 3D, we adapt the spectral-element methods. Again, for simplicity of notations, here we only present the spectral-element method for a planar domain in 2D [14]. Extension to 3D is straightforward and the details are omitted here for brevity [14].

In the case of a planar domain Ω\Omega in 2D [14], we first decompose the domain into NEN_E non-overlapping curvilinear quadrilateral elements [14] as Ω‾=⋃e=1NEΩ‾e\overline{\Omega}=\bigcup_{e=1}^{N_E}\overline{\Omega}_e, and introduce smooth element maps (cf. Figure 1)

Sketch of coordinate transformation from physical domain to computational domain

Figure 1. Sketch of coordinate transformation from physical domain Ωe\Omega_e to computational domain K^\widehat{K}.

Fe:K^=(−1,1)2⟶Ωe.F_e:\widehat{K}=(-1,1)^2\longrightarrow\Omega_e.

On the reference square K^\hat{K}, we can use tensor-product Gauss-Lobatto-Legendre nodes and the associated Lagrange basis with the degree of polynomials in each direction up to NpN_p. The local trial functions on Ωe\Omega_e are obtained by composition with Fe−1F_e^{-1}. Let u^=u∘Fe\hat{u}=u\circ F_e and w^=w∘Fe\hat{w}=w\circ F_e. The local volume is

aevol(u,w)=∫K^(∇x^u^)TGe(∇x^w^) dx^+∫K^V^e(x^)u^(x^)w^(x^)∣det⁡(DFe(x^))∣ dx^,(30)a_e^{\mathrm{vol}}(u,w)=\int_{\hat{K}}(\nabla_{\hat{x}}\hat{u})^T G_e(\nabla_{\hat{x}}\hat{w})\,\mathrm{d}\hat{x}+\int_{\hat{K}}\hat{V}_e(\hat{x})\hat{u}(\hat{x})\hat{w}(\hat{x})\left|\det(D F_e(\hat{x}))\right|\,\mathrm{d}\hat{x}, \tag*{(30)}

where

Ge=∣det⁡(DFe)∣(DFe)−1((DFe)T)−1,V^e=V∘Fe.G_e=\left|\det(D F_e)\right|(D F_e)^{-1}\left((D F_e)^T\right)^{-1},\qquad\hat{V}_e=V\circ F_e.

The local mass is

me(u,w)=∫K^u^(x^)w^(x^)∣det⁡(DFe(x^))∣ dx^.m_e(u,w)=\int_{\hat{K}}\hat{u}(\hat{x})\hat{w}(\hat{x})\left|\det(D F_e(\hat{x}))\right|\,\mathrm{d}\hat{x}.

If an edge γ\gamma of Ωe\Omega_e lies on the physical boundary ∂Ω\partial\Omega, the Robin term adds

σ∫γ^u^w^∣∂τFe∣ ds^,\sigma\int_{\hat{\gamma}}\hat{u}\hat{w}\left|\partial_\tau F_e\right|\,\mathrm{d}\hat{s},

where γ^\hat{\gamma} is the corresponding edge of the reference square and ∂τ\partial_\tau is derivative along the tangent direction. Finally, the assembly procedure is as the following: Each element is mapped to the reference square, high-order spectral matrices are computed locally with the appropriate Jacobian and metric factors, boundary-edge contributions are added only on ∂Ω\partial\Omega, and the local matrices are assembled by enforcing continuity across neighboring elements. The final algebraic eigenvalue problem is the same as (16), but with a basis adapted to the geometry by element-wise mappings. For more details, we refer to Section 4.5 of [14].

Accuracy test

In this section, we report accuracy of the spectral and spectral-element methods proposed in the previous section for computing the Schrödinger eigenvalue problem eq:1-(15). Let NN be the degree of freedoms (DOFs) of an approximation, and λn,Nσ\lambda^\sigma_{n,N} (n=1,2,…,Nn=1,2,\ldots,N) be all the eigenvalues of the numerical approximation (24), satisfying (or ranked as)

λ1,Nσ≤λ2,Nσ≤…≤λN,Nσ.\lambda^\sigma_{1,N}\leq\lambda^\sigma_{2,N}\leq\ldots\leq\lambda^\sigma_{N,N}.

We adopt the relative errors which are defined as

Relative error⁡(λnσ):=∣λnσ−λn,Nσ∣∣λnσ∣,n=1,2,…,N,(31)\operatorname{Relative\ error}\left(\lambda^\sigma_n\right):=\frac{\left|\lambda^\sigma_n-\lambda^\sigma_{n,N}\right|}{\left|\lambda^\sigma_n\right|},\qquad n=1,2,\ldots,N, \tag*{(31)}

if λnσ≠0\lambda^\sigma_n\neq0, and otherwise we use the absolute error as ∣0−λn,Nσ∣\left|0-\lambda^\sigma_{n,N}\right| if λnσ=0\lambda^\sigma_n=0. In the case that the eigenvalues λnσ\lambda^\sigma_n are not given analytically, the numerical ‘exact’ eigenvalues are obtained by choosing a sufficient large N=NexN=N_{\mathrm{ex}}.

Accuracy in 1D

We take d=1d=1 and Ω=(0,π)\Omega=(0,\pi) in (1)-(2). When V(x)≡0V(x)\equiv0, all the eigenvalues are given analytically in (7), (8) and (9) for homogeneous Neumann boundary condition, homogeneous Dirichlet boundary condition and Robin boundary condition, respectively [15]. We adopt the spectral method presented in the previous section to solve this problem numerically for obtaining the corresponding eigenvalues.

Figure 2 displays the relative errors of the eigenvalues when V(x)≡0V(x)\equiv0 for different σ\sigma and DOFs NN. Figure 3 plots the relative errors of the eigenvalues when V(x)=x22V(x)=\frac{x^2}{2} for different σ\sigma and DOFs NN, while the numerical reference solutions are obtained by taking Nex=5000N_{\mathrm{ex}}=5000.

Accuracy test in 1D for Neumann and Robin boundary conditions

Figure 2. Accuracy test in 1D with Ω=(0,π)\Omega=(0,\pi) and V(x)≡0V(x)\equiv0 in (1.1)-(1.2) for different NN and σ\sigma: (a) Neumann (σ=0\sigma=0), and (b) Robin (σ=1\sigma=1).

Accuracy test in 1D for Neumann and Robin boundary conditions with quadratic potential

Figure 3. Accuracy test in 1D with Ω=(0,π)\Omega=(0,\pi) and V=x22V=\frac{x^2}{2} in (1.1)-(1.2) for different NN and σ\sigma: (a) Neumann (σ=0\sigma=0), and (b) Robin (σ=1\sigma=1).

From Figs. 2&3, we can conclude that, with a given DOFs NN, the proposed spectral method can compute up to approximately 2πN≈0.6N\frac{2}{\pi}N\approx0.6N eigenvalues with relative errors below 10−810^{-8} (c.f. Figs. 2&3) for the Schrödinger eigenvalue problem (1.1)-(1.2) in 1D. With this kind of resolution capacity, we can obtain numerically the RtN gaps of the Schrödinger operator and their convergence rates in 1D, which will be reported in the next section.

Accuracy in 2D

We take d=2d=2 and Ω\Omega a unit square or a unit disk in (1.1)-(1.2). When V(x,y)≡0V(x,y)\equiv0, all the eigenvalues are given analytically under different boundary conditions [15, 31]. When Ω\Omega is taken as a unit square (0,1)2(0,1)^2 and a unit disk, we adopt the spectral method and the spectral-element method presented in the previous section to solve this problem numerically for obtaining the corresponding eigenvalues, respectively. In the latter case, we also investigate the performance of different combinations of the number of sub-elements and the polynomial degree of basis functions within each sub-element, while keeping the total number of degrees of freedom (DOFs) fixed.

Figure 4 shows the relative errors of the first 5,000 eigenvalues when V(x,y)≡0V(x,y)\equiv0 and Ω=(0,1)2\Omega=(0,1)^2 in (1.1)-(1.2) for different σ\sigma and DOFs N=Np×NpN=N_p\times N_p with Np=50,80,120N_p=50,80,120.

Accuracy test plots for Neumann and Robin boundary conditions on a square domain

Figure 4. Accuracy test in 2D with Ω=(0,1)2\Omega= (0,1)^2 and V(x,y)≡0V(x,y) \equiv0 in (1.1)-(1.2) for different N=Np×NpN = N_p \times N_p and σ\sigma: (a) Neumann (σ=0\sigma= 0), and (b) Robin (σ=1\sigma= 1).

Figure 5 plots the relative errors of the first 5,000 eigenvalues when V(x,y)≡0V(x,y) \equiv0 and Ω\Omega taken as a unit disk in (1.1)-(1.2) and NE=12N_E = 12 for different σ\sigma and NpN_p. In addition, Figure 7 displays the relative errors of the eigenvalues when V(x,y)≡0V(x,y) \equiv0 and Ω\Omega taken as a unit disk in (1.1)-(1.2) for different σ\sigma, NEN_E and NpN_p such that the total DOFs N≈25,000N \approx25,000 (c.f. Figure 6). Finally, Figure 8 shows the relative errors of the first 5000 eigenvalues when σ=1\sigma= 1 and V(x,y)=1+sin⁡(x+y)V(x,y) = 1 + \sin(x+y) in (1.1)-(1.2) when Ω\Omega is taken as a unit square and a unit disk to be solved numerically by the spectral method with different N=NP×NpN = N_P \times N_p and the spectral-element method with different NEN_E and NPN_P, respectively.

Accuracy test plots for Neumann and Robin boundary conditions on a unit disk

Figure 5. Accuracy test in 2D with Ω\Omega taken as a unit disk and V(x,y)≡0V(x,y) \equiv0 in (1.1)-(1.2) and NE=12N_E = 12 for different NpN_p and σ\sigma: (a) Neumann (σ=0\sigma= 0), and (b) Robin (σ=1\sigma= 1).

Figure 6. Schematic illustration of two different spectral-element partitions of the unit disk: NE=5N_E = 5 (left), and NE=12N_E = 12 (right).

Accuracy test in 2D for Neumann and Robin boundary conditions

Figure 7. Accuracy test in 2D with Ω\Omega taken as a unit disk and V(x,y)≡0V(x,y) \equiv0 in (1)-(2) for different NEN_E, NpN_p and σ\sigma: (a) Neumann (σ=0\sigma= 0), and (b) Robin (σ=1\sigma= 1).

Accuracy test in 2D with a nonzero potential

Figure 8. Accuracy test in 2D with σ=1\sigma= 1 and V(x,y)=1+sin⁡(x+y)V(x,y) = 1 + \sin(x+y) in (1)-(2) for different Ω\Omega: (a) a square with different N=Np×NpN = N_p \times N_p, and (b) a disk with NE=5N_E = 5 for different NpN_p.

From Figs. 4-8, we can draw the following conclusions: (i) With the same number of DOFs, we observe that a larger number of reliable eigenvalues can generally be obtained by increasing the polynomial degree NPN_P and decreasing the number of elements NEN_E (c.f. Fig. 7), which demonstrates the significant advantage of high-order spectral methods in computing a large number of eigenvalues with high accuracy. (ii) With a given DOFs NN, the

proposed spectral method can compute up to approximately 4π2N≈0.4N\frac{4}{\pi^2}N \approx0.4N eigenvalues with relative errors below 10−810^{-8} (c.f. Figs. 4&8 (left one)) for the Schrödinger eigenvalue problem (1.1)-(1.2) in 2D. (iii) With a given DOFs NN, the proposed spectral-element method can compute up to approximately 0.25N0.25N eigenvalues with relative errors below 10−810^{-8} (c.f. Figs. 5&8 (right one)) for the Schrödinger eigenvalue problem (1.1)-(1.2) in 2D. (iv) In 2D, with about DOFs at N≈25,000N \approx25,000, we can obtain numerically at least 5,000 reliable eigenvalues (with relative error less than 10−810^{-8}) (c.f. Figs. 4, 5 and 8). Again, with this kind of resolution capacity of the proposed spectral and spectral-element methods, we can obtain numerically the RtN gaps of the Schrödinger operator and their convergence rates in 2D, which will be reported in the next section.

Convergence rates of the RtN gaps in 1D

Figure 10. Convergence rates of the RtN gaps of (1.1)-(1.2) in 1D with σ=1\sigma= 1 for different potentials V(x)V(x): (a) convergence rate of the cumulative averages Gn(σ)G_n(\sigma), and (b) convergence rate of the RtN gaps gn(σ)g_n(\sigma).

Numerical studies on the Robin-to-Neumann (RtN) gaps

In this section, we investigate numerically the Robin-to-Neumann (RtN) gaps in 1D, 2D and 3D by using the spectral and spectral-element methods proposed in Section 3. In each case, we take enough large DOFs NN such that we have enough reliable eigenvalues for obtaining conclusions on the RtN gaps.

The RtN gaps in 1D

We take d=1d=1 and Ω=(−1,1)\Omega=(-1,1) in (1.1)–(1.2). The problem is solved numerically by the spectral method. Figure 9 shows the RtN gaps with V(x)=x22V(x)=\frac{x^2}{2} for different σ\sigma.

The RtN gaps in 1D for two values of sigma

Figure 9. The RtN gaps of (1.1)-(1.2) in 1D with V=x22V = \frac{x^2}{2} for different σ\sigma: (a) σ=1\sigma= 1, and (b) σ=−1\sigma= -1.

Convergence rates of the RtN gaps in 1D

Figure 11. Convergence rates of the RtN gaps of (1.1)–(1.2) in 1D with V(x)=x22V(x)=\frac{x^2}{2} for different σ\sigma: (a) convergence rate of the cumulative averages Gn(σ)G_n(\sigma), and (b) convergence rate of the RtN gaps gn(σ)g_n(\sigma).

From Figs. 9–11, and additional results with several other potentials, such as V(x)=1+sin⁡(x)V(x)=1+\sin(x), or V(x)=1−cos⁡(x)V(x)=1-\cos(x), or

V(x)={1,−12≤x≤13,0,otherwise,−1≤x≤1;V(x)= \begin{cases} 1, & -\frac{1}{2}\le x\le\frac{1}{3},\\ 0, & \text{otherwise}, \end{cases} \qquad-1\le x\le1;

which are not shown here for brevity, we can immediately conclude (1.12) when d=1d=1 and (1.13).

The RtN gaps in 2D

We take d=2d=2 in (1)-(2) with several different domains, such as right triangle with an angle π5\frac{\pi}{5}, L-shaped domain, quarter stadium, mushroom (c.f. Figure 12. The spectral-element method is adopted with NEN_E, NpN_p and DOFs NN given below (c.f. Figure 12):

Figure 12. Schematics and partitioning of the different domains Ω\Omega in 2D: a right triangle with an angle π/5\pi/5, an L-shaped domain, a quarter stadium, a mushroom (from left to right).

  • Triangle: NE=3N_E = 3, Np=128N_p = 128 and the DOF N=49,537N = 49,537;

  • L-shaped domain: NE=3N_E = 3, Np=128N_p = 128 and the DOF N=49,665N = 49,665;

  • Quarter stadium: NE=5N_E = 5, Np=100N_p = 100 and the DOF N=50,401N = 50,401;

  • Mushroom: NE=12N_E = 12, Np=70N_p = 70 and the DOF N=59,221N = 59,221.

Figure 13 shows the RtN gaps gn(σ)g_n(\sigma) and Gn(σ)G_n(\sigma) of (1.1)-(1.2) in 2D with V(x,y)=x2+y2V(x,y) = x^2 + y^2 and σ=1\sigma= 1 for different domains Ω\Omega. Figure 14 shows convergence rates of the cumulative averages Gn(σ)G_n(\sigma) of (1.1)-(1.2) in 2D with V(x,y)≡0V(x,y) \equiv0 and σ=1\sigma= 1 for different domains Ω\Omega.

The RtN gaps for different two-dimensional domains

Figure 13. The RtN gaps of (1)-(2) in 2D with V(x,y)=x2+y2V(x,y)=x^2+y^2 and σ=1\sigma=1 for different domains Ω\Omega: (a) the unit square, (b) the annulus with inner and outer radius as 1 and 2, respectively, (c) the right triangle with an angle π5\frac{\pi}{5}, (d) the L-shaped domain, (e) the quarter stadium, and (f) the mushroom.

Convergence-rate plots for the cumulative averages in two dimensions, with different domains and different values of sigma

Figure 14. Convergence rate of the cumulative averages Gn(σ)G_n(\sigma) of (1.1)-(1.2) in 2D with V(x,y)≡0V(x,y) \equiv0: (a) with σ=1\sigma= 1 for different domains Ω\Omega, and (b) with Ω\Omega taken as a quarter-stadium domain for different σ\sigma.

Convergence-rate plots for the cumulative averages in two dimensions, with different domains and different potentials

Figure 15. Convergence rate of the cumulative averages Gn(σ)G_n(\sigma) of (1.1)-(1.2) in 2D with σ=1\sigma= 1: (a) with V(x,y)=x2+y2V(x,y) = x^2 + y^2 for different domains Ω\Omega, and (b) with Ω\Omega taken as a quarter-stadium domain for different potentials.

From Figs. 13-15, and additional results not shown here for brevity, we can immediately conclude (1.12) when d=2d = 2.

The RtN gaps in 3D

We take d=3d = 3 and V(x)≡0V(\mathbf{x}) \equiv0 in (1)-(2) with two different domains Ω\Omega: one is a cube, i.e. Ω=(0,π)3\Omega= (0,\pi)^3, and another is a unit ball. The eigenvalues can be computed semi-analytically from separated transcendental equations.

Figure 16 shows the RtN gaps gn(σ)g_n(\sigma) and Gn(σ)G_n(\sigma) of (1)-(2) in 3D with σ=1\sigma= 1 for two different domains. Figure 17 displays convergence rates of the cumulative averages Gn(σ)G_n(\sigma) of (1)-(2) in 3D for different σ\sigma and domains Ω\Omega.

The RtN gaps for the cube and unit ball in 3D

Figure 16. The RtN gaps of (1.1)-(1.2) in 3D with V(x)≡0V(\mathbf{x}) \equiv0 and σ=1\sigma= 1 for different domains Ω\Omega: (a) the cube (0,π)3(0,\pi)^3, and (b) the unit ball.

Convergence rates of the cumulative averages for the cube and unit ball in 3D

Figure 17. Convergence rate of the cumulative averages Gn(σ)G_n(\sigma) of (1.1)-(1.2) in 3D with V(x)≡0V(\mathbf{x}) \equiv0 for different σ\sigma and different domains Ω\Omega: (a) the cube (0,π)3(0,\pi)^3, and (b) the unit ball.

From Figs. 16-17, and additional results not shown here for brevity, we can immediately conclude (1.12) when d=3d = 3.

Conclusion

We proposed and implemented spectral and spectral-element methods for accurately computing many eigenvalues of the Schrödinger operator with Robin boundary condition on simple and complex geometries, respectively. The proposed framework combines the geometric flexibility of finite element partitions with the high accuracy of spectral methods and pp-version finite element methods, allowing many reliable eigenvalues to be obtained under a fixed number of degrees of freedom. Using thousands of reliable eigenvalues obtained numerically by the proposed numerical methods, we investigated systematically the Robin-to-Neumann (RtN) gaps of Schrödinger operator with the Robin boundary condition in one dimension, two dimensions and three dimensions with different potentials and different geometries. Based on our extensive numerical results, we formulated a unified conjecture on the cumulative averages of the RtN gaps of the Schrödinger operator.

Acknowledgments

The authors would like to thank Professor Zeev Rudnick for his valuable suggestions and comments. This work was partially supported by the Ministry of Education of Singapore under its AcRF Tier 1 funding grant A-8003584-00-00 (W. Bao) and by the National Natural Science Foundation of China Grant No. 12501554 (F. Huang).

References

References

  1. [1]I. Babuška and J. E. Osborn. Finite element-Galerkin approximation of the eigenvalues and eigenvectors of selfadjoint problems. Math. Comp., 52(186):275–297, 1989.
  2. [2]I. Babuška and J. E. Osborn. Eigenvalue problems. In P. G. Ciarlet and J.-L. Lions, editors, Handbook of Numerical Analysis, volume 2, pages 641–787. North-Holland, Amsterdam, 1991.
  3. [3]W. Bao, L. Chen, X. Jiang, and Y. Ma. A Jacobi spectral method for computing eigenvalue gaps and their distribution statistics of the fractional Schrödinger operator. J. Comput. Phys., 421:109733, 2020.DOI
  4. [4]W. Bao, X. Ruan, J. Shen, and C. Sheng. Fundamental gaps of the fractional Schrödinger operator. Commun. Math. Sci., 17(2):447–471, 2019.
  5. [5]F. A. Berezin and M. A. Shubin. The Schrödinger Equation, volume 66 of Mathematics and Its Applications. Kluwer Academic Publishers, Dordrecht, 1991.
  6. [6]D. Boffi. Finite element approximation of eigenvalue problems. Acta Numer., 19:1–120, 2010.DOI
  7. [7]S. Böglі, J. B. Kennedy, and R. Lang. On the eigenvalues of the Robin Laplacian with a complex parameter. Anal. Math. Phys., 12(1):39, 2022.
  8. [8]S. C. Brenner and L. R. Scott. The Mathematical Theory of Finite Element Methods, volume 15 of Texts in Applied Mathematics. Springer, New York, 3 edition, 2008.
  9. [9]C. Carstensen and J. Gedicke. An adaptive finite element eigenvalue solver of asymptotic quasi-optimal computational complexity. SIAM J. Numer. Anal., 50(3):1029–1057, 2012.
  10. [10]C. Carstensen and J. Gedicke. Guaranteed lower bounds for eigenvalues. Math. Comp., 83(290):2605–2629, 2014.DOI
  11. [11]H. Chen, L. He, and A. Zhou. Finite element approximations of nonlinear eigenvalue problems in quantum physics. Comput. Methods Appl. Mech. Engrg., 200(21–22):1846–1865, 2011.DOI
  12. [12]X. Dai, J. Xu, and H. Zhou. Convergence and optimal complexity of adaptive finite element eigenvalue computations. Numer. Math., 110(3):313–355, 2008.DOI
  13. [13]E. B. Davies. Spectral Theory and Differential Operators, volume 42 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 1995.DOI
  14. [14]M. O. Deville, P. F. Fischer, and E. H. Mund. High-Order Methods for Incompressible Fluid Flow, volume 9 of Cambridge Monographs on Applied and Computational Mathematics. Cambridge University Press, Cambridge, 2002.
  15. [15]K. Gittins and B. Helffer. Courant-sharp Robin eigenvalues for the square: The case with small Robin parameter. Ann. Math. Qué., 44(1):91–123, 2020.
  16. [16]D. J. Griffiths and D. F. Schroeter. Introduction to Quantum Mechanics. Cambridge University Press, Cambridge, 3 edition, 2018.
  17. [17]Y. Guo and P. Ming. A deep learning method for computing eigenvalues of the fractional Schrödinger operator. J. Syst. Sci. Complex., 37(2):391–412, 2024.
  18. [18]Y. Guo, P. Ming, and H. Yu. Generalization error estimates of a machine learning method for solving high-dimensional Schrödinger eigenvalue problems. Sci. China Math., 69(4):1063–1114, 2026.DOI
  19. [19]B. Hashemi and Y. Nakatsukasa. Least-squares spectral methods for ODE eigenvalue problems. SIAM J. Sci. Comput., 44(5):A3244–A3264, 2022.
  20. [20]E. Hernández and R. Rodríguez. Finite element approximation of spectral problems with Neumann boundary conditions on curved domains. Math. Comp., 72(243):1099–1115, 2003.DOI
  21. [21]X. Hu and X. Cheng. Acceleration of a two-grid method for eigenvalue problems. Math. Comp., 80(275):1287–1301, 2011.DOI
  22. [22]M. G. Larson. A posteriori and a priori error analysis for finite element approximations of self-adjoint elliptic eigenvalue problems. SIAM J. Numer. Anal., 38(2):608–625, 2000.DOI
  23. [23]H. Li and Z. Zhang. Efficient spectral and spectral element methods for eigenvalue problems of Schrödinger equations with an inverse square potential. SIAM J. Sci. Comput., 39(1):A114–A140, 2017.
  24. [24]Q. Lin and H. Xie. A multi-level correction scheme for eigenvalue problems. Math. Comp., 84(291):71–88, 2015.
  25. [25]A. López-Yela and J. M. Pérez-Pardo. Finite element method to solve the spectral problem for arbitrary self-adjoint extensions of the Laplace–Beltrami operator on manifolds with a boundary. J. Comput. Phys., 347:235–260, 2017.
  26. [26]R. Ognibene. On asymptotics of Robin eigenvalues in the Dirichlet limit. Commun. Partial Differ. Equ., 50(9):1174–1210, 2025.
  27. [27]A. T. Patera. A spectral element method for fluid dynamics: Laminar flow in a channel expansion. J. Comput. Phys., 54(3):468–488, 1984.DOI
  28. [28]Z. Rudnick and I. Wigman. The Robin problem on rectangles. J. Math. Phys., 62(11):113503, 2021.DOI
  29. [29]Z. Rudnick and I. Wigman. On the Robin spectrum for the equilateral triangle. J. Phys. A: Math. Theor., 55(25):254004, 2022.DOI
  30. [30]Z. Rudnick and I. Wigman. On the Robin spectrum for the hemisphere. Ann. Math. Qué., 46(1):121–137, 2022.arxiv.org/abs/2008.12964
  31. [31]Z. Rudnick, I. Wigman, and N. Yesha. Differences between Robin and Neumann eigenvalues. Commun. Math. Phys., 388:1603–1635, 2021.DOI
  32. [32]S. A. Sauter. Finite elements for elliptic eigenvalue problems: Error estimates which are explicit with respect to λ, h, and p. SIAM J. Numer. Anal., 48(1):95–108, 2010.DOI
  33. [33]W. Shan and H. Li. The triangular spectral element method for Stokes eigenvalues. Math. Comp., 86(308):2579–2611, 2017.DOI
  34. [34]J. Shen, T. Tang, and L.-L. Wang. Spectral Methods: Algorithms, Analysis and Applications, volume 41 of Springer Series in Computational Mathematics. Springer, Berlin, Heidelberg, 2011.DOI
  35. [35]G. Teschl. Mathematical Methods in Quantum Mechanics: With Applications to Schrödinger Operators, volume 157 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 2 edition, 2014.
  36. [36]M. Vanmaele and R. Van Keer. An operator method for a numerical quadrature finite element approximation for a class of second-order elliptic eigenvalue problems in composite structures. ESAIM Math. Model. Numer. Anal., 29(3):339–365, 1995.DOI
  37. [37]W. Wang and Z. Zhang. Spectral element methods for eigenvalue problems based on domain decomposition. SIAM J. Sci. Comput., 44(2):A689–A719, 2022.DOI
  38. [38]J. A. C. Weideman and L. N. Trefethen. The eigenvalues of second-order spectral differentiation matrices. SIAM J. Numer. Anal., 25(6):1279–1298, 1988.DOI
  39. [39]J. Xu and A. Zhou. A two-grid discretization scheme for eigenvalue problems. Math. Comp., 70(233):17–25, 2001.
  40. [40]B. Yang and A. Zhou. Eigenfunction behavior and adaptive finite element approximations of nonlinear eigenvalue problems in quantum physics. ESAIM Math. Model. Numer. Anal., 55(1):209–227, 2021.
  41. [41]Y. Yang and H. Bi. Two-grid finite element discretization schemes based on shifted-inverse power method for elliptic eigenvalue problems. SIAM J. Numer. Anal., 49(4):1602–1624, 2011.DOI
  42. [42]Z. Zhang. How many numerical eigenvalues can we trust? J. Sci. Comput., 65(2):455–466, 2015.

Paper details

vlnc:W_wjyn6xyx
Version
v1 · latest
Published
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents