betaFeedback

WorksPhysics and Astronomy

Spectra and pseudo-spectra of tridiagonal kk-Toeplitz matrices and the topological origin of the non-Hermitian skin effect

Abstract

We establish new results on the spectra and pseudo-spectra of tridiagonal kk-Toeplitz operators and matrices. In particular, we prove the connection between the winding number of the eigenvalues of the symbol function and the exponential decay of the associated eigenvectors (or pseudo-eigenvectors). Our results elucidate the topological origin of the non-Hermitian skin effect in general one-dimensional polymer systems of subwavelength resonators with imaginary gauge potentials, proving the observation and conjecture in arXiv:2307.13551. We also numerically verify our theory for these systems.

Introduction

Motivated by the study of the non-Hermitian skin effect in polymer systems of subwavelength resonators and its topological origin, in this paper we extend the theory on the spectra and pseudo-spectra of Toeplitz operators and matrices to tridiagonal kk-Toeplitz operators and matrices. Our main goal is to establish a connection between, on one hand, the winding number of the eigenvalues of the symbol function of a tridiagonal kk-Toeplitz operator and, on the other hand, the exponential decay of the associated eigenvectors or pseudo-eigenvectors. By doing so, we elucidate the topological origin of the eigenmode condensation at one edge of a polymer system of subwavelength resonators where an imaginary gauge potential is added inside the resonators to break Hermiticity. We show that, due to a nonzero total winding of the eigenvalues of the symbol of the kk-Toeplitz operator corresponding to the polymer system, the eigenmodes exhibit exponential decay and condensate at one of the edges of the structure.

In addition to advancing our understanding of the non-Hermitian skin effect, which can be seen as the classical analogue of the non-Hermitian Anderson model [16], one of our key contributions in this paper is to optimally characterize the spectra of tridiagonal kk-Toeplitz operators. Notably, we reveal the distinctive property that the eigenvectors of these operators undergo exponential decay. Using the notion of pseudo-spectrum, we rigorously justify the exponential decay of the pseudo-eigenvectors of the associated kk-Toeplitz matrices and significantly generalize the results presented in [9, 15, 18, 20, 21].

Our paper is organized as follows. In Section 2, we first provide an optimal characterization of the spectra of tridiagonal kk-Toeplitz and Laurent operators. Then we prove the existence of an exponentially decaying eigenvector for Toeplitz operators. In Section 3, we establish a connection between the eigenvectors of a tridiagonal kk-Toeplitz operator and the associated tridiagonal kk-Toeplitz matrix. In Section 4 we discuss the topological origin of the non-Hermitian skin effect in polymer systems of subwavelength resonators and numerically verify our findings on the spectra of tridiagonal kk-Toeplitz operators. In Section 5, we make some concluding remarks and give some possible extensions of our present work.

Spectra of tridiagonal kk-Toeplitz operators

In this section, we characterize the spectrum of tridiagonal kk-Toeplitz operators and prove that their associated eigenvectors exhibit exponential decay. This type of operators and their associated truncated operators, i.e. matrices, are crucial in the study of the one-dimensional polymer systems of subwavelength resonators [1, 5, 6, 11]. We define a tridiagonal kk-Toeplitz operator by

A=(a1b10c1a2b2⋱0⋱⋱⋱⋱⋱ck−2ak−1bk−1⋱⋱ck−1akbk⋱⋱cka1b1⋱⋱⋱⋱)(1)A=\left(\begin{array}{cccccccc} a_1 & b_1 & 0 & & & & & \\ c_1 & a_2 & b_2 & \ddots& & & & \\ 0 & \ddots& \ddots& \ddots& \ddots& & & \\ & \ddots& c_{k-2} & a_{k-1} & b_{k-1} & \ddots& & \\ & & \ddots& c_{k-1} & a_k & b_k & \ddots& \\ & & & \ddots& c_k & a_1 & b_1 & \ddots\\ & & & & & \ddots& \ddots& \ddots \end{array}\right) \tag*{(1)}

where ai,bi,ci∈Ca_i,b_i,c_i\in\mathbb{C} for i∈Ni\in\mathbb{N} and Aij=0A_{ij}=0 if ∣i−j∣>1|i-j|>1 for i,j∈Ni,j\in\mathbb{N}. Its finite, truncated, counterpart will be called a tridiagonal kk-Toeplitz matrix.

To characterize the spectrum of tridiagonal kk-Toeplitz operators, we introduce the notions of Fredholm operator and Fredholm index. Let XX be a Banach space and B(X)\mathcal{B}(X) the set of bounded linear operators on XX. We say that an operator A∈B(X)A\in\mathcal{B}(X) is a Fredholm operator if Im⁡A\operatorname{Im}A is a closed subspace in XX and

dim⁡Ker⁡(A)<∞,dim⁡Coker⁡A<∞,\dim\operatorname{Ker}(A)<\infty,\qquad\dim\operatorname{Coker}A<\infty,

where Coker⁡A:=X/Im⁡A\operatorname{Coker}A:=X/\operatorname{Im}A. The index of the Fredholm operator AA is defined as

Ind⁡A:=dim⁡Ker⁡A−dim⁡Coker⁡A.(2)\operatorname{Ind}A:=\dim\operatorname{Ker}A-\dim\operatorname{Coker}A. \tag*{(2)}

Tridiagonal kk-Toeplitz operators

Observe that the tridiagonal kk-Toeplitz operator AA presented in (2.1) can be reformulated as a tridiagonal block Toeplitz operator, where the blocks repeat in a 1-periodic way:

A=(A0A−1A1A0⋱⋱⋱)(3)A=\left(\begin{array}{cccc} A_0 & A_{-1} & & \\ A_1 & A_0 & \ddots& \\ & \ddots& \ddots& \end{array}\right) \tag*{(3)}

with

A0=(a1b10⋯0c1a2b2⋱⋮0⋱⋱⋱0⋮⋱ck−2ak−1bk−10⋯0ck−1ak),A−1=(0⋯⋯0⋮⋱⋮0⋱⋮bk0⋯0),A1=(0⋯0ck⋮⋱0⋮⋱⋮0⋯⋯0).A_0=\left(\begin{array}{ccccc} a_1 & b_1 & 0 & \cdots& 0\\ c_1 & a_2 & b_2 & \ddots& \vdots\\ 0 & \ddots& \ddots& \ddots& 0\\ \vdots& \ddots& c_{k-2} & a_{k-1} & b_{k-1}\\ 0 & \cdots& 0 & c_{k-1} & a_k \end{array}\right),\quad A_{-1}=\left(\begin{array}{ccccc} 0 & \cdots& \cdots& 0\\ \vdots& \ddots& & \vdots\\ 0 & & \ddots& \vdots\\ b_k & 0 & \cdots& 0 \end{array}\right),\quad A_1=\left(\begin{array}{ccccc} 0 & \cdots& 0 & c_k\\ \vdots& \ddots& & 0\\ \vdots& & \ddots& \vdots\\ 0 & \cdots& \cdots& 0 \end{array}\right).

The symbol of the tridiagonal block Toeplitz operator (2.3) (and thus also of the tridiagonal kk-Toeplitz operator (2.1)) is defined by

f:S1→Ck×kf:S^1\to\mathbb{C}^{k\times k}
z↦A−1z−1+A0+A1z.(4)z\mapsto A_{-1}z^{-1}+A_0+A_1z. \tag*{(4)}

or explicitly

f(z)=(a1b10⋯0ckzc1a2b20⋯00c2a3b3⋱⋮⋮⋱⋱⋱⋱00⋯0ck−2ak−1bk−1bkz−10⋯0ck−1ak).(5)f(z)=\begin{pmatrix} a_1 & b_1 & 0 & \cdots& 0 & c_kz \\ c_1 & a_2 & b_2 & 0 & \cdots& 0 \\ 0 & c_2 & a_3 & b_3 & \ddots& \vdots\\ \vdots& \ddots& \ddots& \ddots& \ddots& 0 \\ 0 & \cdots& 0 & c_{k-2} & a_{k-1} & b_{k-1} \\ b_kz^{-1} & 0 & \cdots& 0 & c_{k-1} & a_k \end{pmatrix}. \tag*{(5)}

We will write A=T(f)A=T(f). In the following sections, we will characterize the spectrum of the operator T(f)T(f) based on its symbol f(z)f(z).

Hardy Spaces

To get a better understanding of the connection between the symbol ff and the operator T(f)T(f) itself, it is necessary to introduce the so-called Hardy Spaces. Let T\mathbb{T} be the unit circle in C\mathbb{C}.

Definition 2.1. The Hardy spaces H2:=H2(T)H^2:=H^2(\mathbb{T}) and H−2:=H−2(T)H^2_-:=H^2_-(\mathbb{T}) are defined by

H2:={f∈L2(T):fn=0 for n<0},H^2:=\{f\in L^2(\mathbb{T}):f_n=0\text{ for }n<0\},
H−2:={f∈L2(T):fn=0 for n≥0},H^2_-:=\{f\in L^2(\mathbb{T}):f_n=0\text{ for }n\geq0\},

where fnf_n denotes the Fourier coefficients of ff.

The spaces H2H^2 and H−2H^2_- are closed, orthogonal subspaces of L2:=L2(T)L^2:=L^2(\mathbb{T}). Hence, one may write

L2=H2⊕H−2.L^2=H^2\oplus H^2_-.

Denote by PP the projection of L2L^2 onto H2H^2. The functions

{12πeinθ}n=0∞\left\{\frac{1}{\sqrt{2\pi}}e^{in\theta}\right\}_{n=0}^{\infty}

form an orthonormal basis of H2H^2. One important property of functions in H2H^2 is given by the following theorem [10 Theorem 6.13].

Theorem 2.2 (F. and M. Riesz). If ff is a nonzero function in H2H^2, then the set {eit∈T:f(eit)=0}\{e^{it}\in\mathbb{T}:f(e^{it})=0\} has measure zero.

It is well-known [21] that for a Toeplitz operator T(f)T(f) with symbol f∈L∞f\in L^\infty, T(f)T(f) is the matrix representation of the operator

H2→H2, g↦P(fg)H^2\to H^2,\ g\mapsto P(fg)

in the above orthonormal basis. A similar notion holds for block Toeplitz operators. We denote by (H2)k(H^2)^k the kk-dimensional vector space whose entries are elements of H2H^2. If P~\widetilde{P} is the orthogonal projection of (L2)k(L^2)^k onto (H2)k(H^2)^k and f∈(L∞)k×kf\in(L^\infty)^{k\times k}, then T(f)T(f) is the matrix representation of the operator

(H2)k→(H2)k,g↦P~(fg),(H^2)^k\to(H^2)^k,\quad\mathbf{g}\mapsto\widetilde{P}(f\mathbf{g}),

where fgf\mathbf{g} is the usual matrix-vector multiplication. Note that the (H2)k(H^2)^k equivalent of Theorem 2.2 holds verbatim with H2H^2 replaced by (H2)k(H^2)^k. With this in mind, we are now able to prove a generalization of Coburn’s lemma (for the Toeplitz case, see [10 Theorem 1.10]).

Theorem 2.3 (Coburn’s lemma; tridiagonal kk-Toeplitz version). Let f∈Ck×k(T)f\in C^{k\times k}(\mathbb{T}) be the symbol of a tridiagonal kk-Toeplitz operator such that det⁡(f(z))\det(f(z)) does not vanish identically on T\mathbb{T}. Then one of the following statements holds:

(i) T(f)T(f) has a trivial kernel;

(ii) T(f)T(f) has a dense range;

(iii) The leading (k−1)×(k−1)(k-1) \times(k-1) principal minor of A0A_0 has a nonzero kernel. In particular, there exists some z0∈C∪{∞}z_0 \in\mathbb{C} \cup\{\infty\} such that ker⁡(z0−1A−1+A0)∩ker⁡(A1)≠0\ker(z_0^{-1}A_{-1}+A_0) \cap\ker(A_1) \ne0.

Proof. It is not hard to see that the adjoint of T(f)T(f) is T(f‾T)T(\overline{f}^{\mathsf{T}}), where the superscript T\mathsf{T} denotes the transpose. Assume that T(f)T(f) has a nontrivial kernel and a nondense range, i.e., conditions (i) and (ii) do not hold. This implies that T(f‾T)T(\overline{f}^{\mathsf{T}}) has a nontrivial kernel as well. Hence there exist nonzero functions g+,h+∈(H2)k\mathbf{g}_+,\mathbf{h}_+ \in(H^2)^k such that

fg+=g−∈(H−2)kandf‾Th+=h−∈(H−2)k,(6)f\mathbf{g}_+ = \mathbf{g}_- \in(H^2_-)^k \quad\text{and} \quad\overline{f}^{\mathsf{T}}\mathbf{h}_+ = \mathbf{h}_- \in(H^2_-)^k, \tag*{(6)}

where g±=(g±,1,g±,2,⋯ ,g±,k)T\mathbf{g}_{\pm}=(g_{\pm,1},g_{\pm,2},\cdots,g_{\pm,k})^{\mathsf{T}} and h±=(h±,1,h±,2,⋯ ,h±,k)T\mathbf{h}_{\pm}=(h_{\pm,1},h_{\pm,2},\cdots,h_{\pm,k})^{\mathsf{T}}. Since g+,h+∈(H2)k\mathbf{g}_+,\mathbf{h}_+ \in(H^2)^k, there exist two sequences of vectors (vn)n,(wn)n∈Ck(\mathbf{v}_n)_n,(\mathbf{w}_n)_n \in\mathbb{C}^k such that

g+=∑n∈Nvnzn−1,h+=∑n∈Nwnzn−1.(7)\mathbf{g}_+ = \sum_{n\in\mathbb{N}} \mathbf{v}_n z^{n-1}, \qquad \mathbf{h}_+ = \sum_{n\in\mathbb{N}} \mathbf{w}_n z^{n-1}. \tag*{(7)}

for z=eiθz=e^{i\theta}, where we have absorbed the factor 12π\frac{1}{\sqrt{2\pi}} into vn\mathbf{v}_n and wn\mathbf{w}_n. This allows us to write

g−=z−1A−1v1+(A0v1+A−1v2)+∑n∈Nzn(A−1vn+2+A0vn+1+A1vn)=z−1A−1v1,(8)\begin{aligned} \mathbf{g}_-={}&z^{-1}A_{-1}\mathbf{v}_1+(A_0\mathbf{v}_1+A_{-1}\mathbf{v}_2)\\ &+\sum_{n\in\mathbb{N}}z^n(A_{-1}\mathbf{v}_{n+2}+A_0\mathbf{v}_{n+1}+A_1\mathbf{v}_n)\\ ={}&z^{-1}A_{-1}\mathbf{v}_1, \tag*{(8)} \end{aligned}

where we have used that g−∈(H−2)k\mathbf{g}_- \in(H^2_-)^k in the last equality. Similarly, using that h−∈(H−2)k\mathbf{h}_- \in(H^2_-)^k, we obtain the following expression for h−\mathbf{h}_-:

h−=z−1A‾1Tw1+(A‾0Tw1+A‾1Tw2)+∑n∈Nzn(A‾1Twn+2+A‾0Twn+1+A‾−1Twn)=z−1A‾1Tw1.(9)\begin{aligned} \mathbf{h}_-={}&z^{-1}\overline{A}_{1}^{\mathsf{T}}\mathbf{w}_1+(\overline{A}_0^{\mathsf{T}}\mathbf{w}_1+\overline{A}_{1}^{\mathsf{T}}\mathbf{w}_2)\\ &+\sum_{n\in\mathbb{N}}z^n(\overline{A}_{1}^{\mathsf{T}}\mathbf{w}_{n+2}+\overline{A}_0^{\mathsf{T}}\mathbf{w}_{n+1}+\overline{A}_{-1}^{\mathsf{T}}\mathbf{w}_n)\\ ={}&z^{-1}\overline{A}_{1}^{\mathsf{T}}\mathbf{w}_1. \tag*{(9)} \end{aligned}

By Theorem 2.2, it follows that g+,h+≠0\mathbf{g}_+,\mathbf{h}_+ \ne0 almost everywhere on T\mathbb{T}. We define

φ:=h+‾Tg−=h+‾Tfg+=(f‾Th+)Tg+=h−Tg+.\varphi:=\overline{\mathbf{h}_+}^{\mathsf{T}}\mathbf{g}_-=\overline{\mathbf{h}_+}^{\mathsf{T}}f\mathbf{g}_+=(\overline{f}^{\mathsf{T}}\mathbf{h}_+)^{\mathsf{T}}\mathbf{g}_+=\mathbf{h}_-^{\mathsf{T}}\mathbf{g}_+.

It holds that φ∈L1\varphi\in L^1. Moreover, φn=(h−Tg+)n=0\varphi_n=(\mathbf{h}_-^{\mathsf{T}}\mathbf{g}_+)_n=0 for n≤0n\le0 and φn=(h+‾Tg−)n=0\varphi_n=(\overline{\mathbf{h}_+}^{\mathsf{T}}\mathbf{g}_-)_n=0 for n≥0n\ge0 since by (2.7), (2.8), and (2.9),

h+‾Tg−=∑n∈Nz−nw‾nTA−1v1,\overline{\mathbf{h}_+}^{\mathsf{T}}\mathbf{g}_-=\sum_{n\in\mathbb{N}}z^{-n}\overline{\mathbf{w}}_n^{\mathsf{T}}A_{-1}\mathbf{v}_1,
h−Tg+=∑n∈Nzn(A‾1Tw1)Tvn.\mathbf{h}_-^{\mathsf{T}}\mathbf{g}_+=\sum_{n\in\mathbb{N}}z^n(\overline{A}_{1}^{\mathsf{T}}\mathbf{w}_1)^{\mathsf{T}}\mathbf{v}_n.

Hence φ=0\varphi=0. This implies A‾1Tw1⊥vn\overline{A}_{1}^{\mathsf{T}}\mathbf{w}_1 \perp\mathbf{v}_n and A−1v1⊥wnA_{-1}\mathbf{v}_1 \perp\mathbf{w}_n for n∈Nn\in\mathbb{N}. Thus either vn⊥ek\mathbf{v}_n\perp\mathbf{e}_k, n=1,⋯n=1,\cdots, or w1∈ker⁡A‾1T\mathbf{w}_1\in\ker\overline{A}_{1}^{\mathsf{T}}. However, if w1∈ker⁡A‾1T\mathbf{w}_1\in\ker\overline{A}_{1}^{\mathsf{T}}, then by (2.9)

f‾Th+=h−=z−1A‾1Tw1=0,\overline{f}^{\mathsf{T}}\mathbf{h}_+=\mathbf{h}_-=z^{-1}\overline{A}_{1}^{\mathsf{T}}\mathbf{w}_1=0,

which contradicts the fact that det⁡(f‾T)=det⁡(f)\det(\overline{f}^{\mathsf{T}})=\det(f) does not vanish identically on T\mathbb{T}. Therefore, vn⊥ek\mathbf{v}_n\perp\mathbf{e}_k for n∈N≥1n\in\mathbb{N}_{\ge1}. By the above arguments, we conclude that if T(f)T(f) has non-dense range and nontrivial kernel, then vn⊥ek\mathbf{v}_n\perp\mathbf{e}_k for n∈N≥1n\in\mathbb{N}_{\ge1} and for any vn\mathbf{v}_n defined by (2.7).

Now we prove that if T(f)T(f) has a non-dense range and a nontrivial kernel, then the condition (iii) must hold. Since fg+=g−∈(H−2)kf\mathbf{g}_+=\mathbf{g}_-\in(H^2_-)^k, by expansion (2.8) it holds that A−1vn+2+A0vn+1+A1vn=0A_{-1}\mathbf{v}_{n+2}+A_0\mathbf{v}_{n+1}+A_1\mathbf{v}_n=0 for n∈Nn\in\mathbb{N}. Since vn⊥ek\mathbf{v}_n\perp\mathbf{e}_k for n∈N≥1n\in\mathbb{N}_{\ge1} it follows that vn∈ker⁡(A1)\mathbf{v}_n\in\ker(A_1). Hence it holds that A−1vn+1+A0vn=0A_{-1}\mathbf{v}_{n+1}+A_0\mathbf{v}_n=0 for n∈N≥2n\in\mathbb{N}_{\ge2}. Note that the expansion (2.8) also gives A−1v2+A0v1=0A_{-1}\mathbf{v}_2+A_0\mathbf{v}_1=0. Therefore, A−1vn+1+A0vn=0A_{-1}\mathbf{v}_{n+1}+A_0\mathbf{v}_n=0 for n∈N≥1n\in\mathbb{N}_{\geq1}. Furthermore, since the image of A−1A_{-1} is confined to the span of ek\mathbf{e}_k and vn⊥ek\mathbf{v}_n\perp\mathbf{e}_k for n∈N≥1n\in\mathbb{N}_{\geq1}, we have

A0∣(k−1)×(k−1)vn∣1:k−1=0,A_0\big|_{(k-1)\times(k-1)}\mathbf{v}_n\big|_{1:k-1}=0,

where A0∣(k−1)×(k−1)A_0\big|_{(k-1)\times(k-1)} is the leading principal (k−1)×(k−1)(k-1)\times(k-1) submatrix of A0A_0. This also means that A0∣(k−1)×(k−1)A_0\big|_{(k-1)\times(k-1)} cannot have full rank. On the other hand, one can compute that

(z−1A−1+A0)vn=(0⋮0ck−1vn(k−1)+z−1bkvn(1)),n≥1(10)\left(z^{-1}A_{-1}+A_0\right)\mathbf{v}_n= \begin{pmatrix} 0\\ \vdots\\ 0\\ c_{k-1}v_n^{(k-1)}+z^{-1}b_kv_n^{(1)} \end{pmatrix},\qquad n\geq1 \tag*{(10)}

where vn(j)v_n^{(j)} is the jj-th element of vn\mathbf{v}_n. Considering the case when n=1n=1, by (2.8), we know that v1(1)≠0v_1^{(1)}\ne0, otherwise,

fg+=g−=z−1A−1v1=0,f g_+=g_-=z^{-1}A_{-1}\mathbf{v}_1=0,

which contradicts det⁡f\det f does not vanish identically on T\mathbb{T}. Similarly, we can also show that bk≠0b_k\ne0. Thus by (2.10) and bk≠0b_k\ne0 there exists some z0∈C∪{∞}z_0\in\mathbb{C}\cup\{\infty\} such that v1\mathbf{v}_1 lies in the kernel of z0−1A−1+A0z_0^{-1}A_{-1}+A_0. Therefore it holds that ker⁡(z0−1A−1+A0)∩ker⁡(A1)≠0\ker(z_0^{-1}A_{-1}+A_0)\cap\ker(A_1)\ne0, which concludes the proof. ■

Spectra of tridiagonal kk-Toeplitz operators

In this section, we will provide a full characterization of the spectrum of tridiagonal kk-Toeplitz operators. We begin by first recalling (and generalizing) a few theorems on Toeplitz operators. The first important theorem is a consequence of [15 Chapter 23, Theorem 4.3]. Its counterpart for Toeplitz operators was already established in [15 Chapter 23, Corollary 4.4].

We denote by σ(⋅)\sigma(\cdot), σess(⋅)\sigma_{\mathrm{ess}}(\cdot) the spectrum and the essential spectrum of the operator, respectively. We define

σdet(f)={λ∈C:det⁡(f(z)−λ)=0, ∃z∈T}.(11)\sigma_{\mathrm{det}}(f)=\{\lambda\in\mathbb{C}:\det(f(z)-\lambda)=0,\ \exists z\in\mathbb{T}\}. \tag*{(11)}

We first have the following theorem characterizing the essential spectrum of T(f)T(f).

Theorem 2.4. The operator T(f)T(f) in (2.3) is Fredholm if and only if det⁡(f)\det(f) has no zeros on T\mathbb{T}. Furthermore, the essential spectrum of T(f)T(f) is given by

σess(T(f))=σdet(f).\sigma_{\mathrm{ess}}(T(f))=\sigma_{\mathrm{det}}(f).

Proof. The first part has been proven in [15 Chapter 23, Theorem 4.3]. For the second part, note that λ∈σess(T(f))\lambda\in\sigma_{\mathrm{ess}}(T(f)) if and only if T(f)−λT(f)-\lambda is not a Fredholm operator. By the first part, this happens only if det⁡(f(z0)−λ)=0\det(f(z_0)-\lambda)=0 for some z0∈Tz_0\in\mathbb{T}. That is, λ\lambda is an eigenvalue of f(z0)f(z_0). On the other hand, if λ\lambda is an eigenvalue for some z0∈Tz_0\in\mathbb{T}, it holds that det⁡(f(z0)−λ)=0\det(f(z_0)-\lambda)=0, which concludes the proof. ■

We now recall the following Theorem due to Gohberg [15 Chapter 23, Theorem 5.1].

Theorem 2.5. Let f:T→Ck×kf:\mathbb{T}\to\mathbb{C}^{k\times k} be such that T(f)T(f) is a Toeplitz operator. Then, T(f)T(f) is Fredholm on the space ℓ2\ell^2 if and only if det⁡(f)\det(f) has no zeros on T\mathbb{T}, in which case

Ind⁡T(f)=−wind⁡(det⁡(f(T)),0),\operatorname{Ind}T(f)=-\operatorname{wind}(\det(f(\mathbb{T})),0),

with wind⁡(det⁡(f(T)),0)\operatorname{wind}(\det(f(\mathbb{T})),0) being the winding number about the origin of the determinant of f(z)f(z) with z∈Tz\in\mathbb{T}.

Utilizing this theorem together with Theorem 2.3, we are now able to establish one of the main results of our paper, an optimal characterization of the spectra of tridiagonal kk-Toeplitz operators. Before stating the theorem, we first define

σwind(f):={λ∈C∖σdet(f):wind⁡(det⁡(f(T)−λ),0)≠0}.(12)\sigma_{\mathrm{wind}}(f):=\{\lambda\in\mathbb{C}\setminus\sigma_{\mathrm{det}}(f):\operatorname{wind}(\det(f(\mathbb{T})-\lambda),0)\ne0\}. \tag*{(12)}

By Lemma A.1,

det⁡(f(z)−λ)=(−1)k+1(∏i=1kci)z+(−1)k+1(∏i=1kbi)z−1+g(λ)\det(f(z)-\lambda)=(-1)^{k+1}\left(\prod_{i=1}^{k}c_i\right)z+(-1)^{k+1}\left(\prod_{i=1}^{k}b_i\right)z^{-1}+g(\lambda)

with g(λ)g(\lambda) being given by (A.2), which yields

σwind(f)={λ∈C∖σdet(f):wind⁡((−1)k+1((∏i=1kci)z+(∏i=1kbi)z−1),−g(λ))≠0}.(13)\sigma_{\mathrm{wind}}(f)=\left\{\lambda\in\mathbb{C}\setminus\sigma_{\mathrm{det}}(f):\operatorname{wind}\left((-1)^{k+1}\left(\left(\prod_{i=1}^{k}c_i\right)z+\left(\prod_{i=1}^{k}b_i\right)z^{-1}\right),-g(\lambda)\right)\ne0\right\}. \tag*{(13)}

Furthermore, since

det⁡(f(z)−λ)=∏j=1k(λj(z)−λ),\det(f(z)-\lambda)=\prod_{j=1}^{k}(\lambda_j(z)-\lambda),

where λj(z)\lambda_j(z), 1≤j≤k1\le j\le k, are the eigenvalues of the matrix f(z)f(z), we obtain

σwind(f)={λ∈C∖σdet(f):∑j=1kwind⁡(λj(T)−λ,0)≠0}={λ∈C∖σdet(f):∑j=1kwind⁡(λj(T),λ)≠0}.(14)\sigma_{\mathrm{wind}}(f)=\left\{\lambda\in\mathbb{C}\setminus\sigma_{\mathrm{det}}(f):\sum_{j=1}^{k}\operatorname{wind}(\lambda_j(\mathbb{T})-\lambda,0)\ne0\right\} =\left\{\lambda\in\mathbb{C}\setminus\sigma_{\mathrm{det}}(f):\sum_{j=1}^{k}\operatorname{wind}(\lambda_j(\mathbb{T}),\lambda)\ne0\right\}. \tag*{(14)}

Note that the representations (12), (2.13) and (2.14) have their own pro and contra. For relating our results to the conjecture in [5], we utilize representation (14) in our main results. We now establish the following theorem for the spectrum of the operator T(f)T(f).

Theorem 2.6. Let f∈Ck×k(T)f\in\mathbb{C}^{k\times k}(\mathbb{T}) be the symbol of a tridiagonal kk-Toeplitz operator T(f)T(f). Denote by B0B_0 the leading (k−1)×(k−1)(k-1)\times(k-1) principal minor of A0A_0. It holds that

σdet(f)∪σwind(f)⊂σ(T(f))⊂σdet(f)∪σwind(f)∪σ(B0),\sigma_{\mathrm{det}}(f)\cup\sigma_{\mathrm{wind}}(f)\subset\sigma(T(f))\subset\sigma_{\mathrm{det}}(f)\cup\sigma_{\mathrm{wind}}(f)\cup\sigma(B_0),

where σdet(f),σwind(f)\sigma_{\mathrm{det}}(f),\sigma_{\mathrm{wind}}(f) are given by (11), (14), respectively.

Proof. The first inclusion is an immediate consequence of Theorems 2.4 and 2.5. For the second inclusion, note that if

λ∈σ(T(f))∖(σdet(f)∪σwind(f)),(15)\lambda\in\sigma(T(f))\setminus\left(\sigma_{\mathrm{det}}(f)\cup\sigma_{\mathrm{wind}}(f)\right), \tag*{(15)}

then it must hold that T(f)−λIT(f)-\lambda I is Fredholm and Ind⁡(T(f)−λI)=wind⁡(det⁡(f−λ),0)=0\operatorname{Ind}(T(f)-\lambda I)=\operatorname{wind}(\det(f-\lambda),0)=0. Based on the definition (2), this implies

dim⁡Ker⁡(T(f)−λI)=dim⁡(ℓ2∖Im⁡(T(f)−λI)).(16)\dim\operatorname{Ker}(T(f)-\lambda I)=\dim\left(\ell^2\setminus\operatorname{Im}(T(f)-\lambda I)\right). \tag*{(16)}

Moreover, since λ∈σ(T(f))\lambda\in\sigma(T(f)), T(f)−λIT(f)-\lambda I has a non-trivial kernel and non-dense image. By Theorem 2.3, this implies that λ\lambda must be an eigenvalue of B0B_0. ■

This characterization is optimal. It cannot be made more precise, as it cannot be guaranteed that an eigenvalue λ\lambda of B0B_0 with wind⁡(det⁡(f−λ),0)=0\operatorname{wind}(\det(f-\lambda),0)=0 is also an eigenvalue of T(f)T(f). To illustrate this, consider

f(z)=(01+12z1+12z−11).f(z)=\begin{pmatrix}0&1+\frac{1}{2}z\\1+\frac{1}{2}z^{-1}&1\end{pmatrix}.

The determinant of this symbol is given by det⁡(f)=−(1+12z)(1+12z−1)\det(f)=-(1+\frac{1}{2}z)(1+\frac{1}{2}z^{-1}). We have wind⁡(det⁡(f(T)),0)=0\operatorname{wind}(\det(f(\mathbb{T})),0)=0 and det⁡f(z)≠0, ∀z∈T\det f(z)\ne0,\ \forall z\in\mathbb{T}. In this case, B0=0B_0=0 is just the top left entry of f(z)f(z). Furthermore, by (2), if 0∈σ(T(f))0 \in\sigma(T(f)), then the kernel space of T(f)T(f) must be non-trivial. Since

T(f)=(010011120012010011⋱),T(f)=\begin{pmatrix} 0 & 1 & 0 & 0\\ 1 & 1 & \frac{1}{2} & 0\\ 0 & \frac{1}{2} & 0 & 1\\ 0 & 0 & 1 & 1\\ & & & \ddots \end{pmatrix},

one can check that, after scaling, the only possible eigenvector of T(f)T(f) corresponding to the eigenvalue 00 is u=(1,0,−2,0,4,…)T\mathbf{u}=(1,0,-2,0,4,\ldots)^{\mathsf{T}}. However, u\mathbf{u} does not lie in ℓ2\ell^2, which means that 0∉σ(T(f))0 \notin\sigma(T(f)).

On the other hand, consider

f~(z)=(01+2z1+2z−11).\widetilde{f}(z)=\begin{pmatrix} 0 & 1+2z\\ 1+2z^{-1} & 1 \end{pmatrix}.

The determinant is det⁡(f~)=−(1+2z)(1+2z−1)\det(\widetilde{f})=-(1+2z)(1+2z^{-1}), hence 0∉σdet⁡(f)0 \notin\sigma_{\det}(f) and wind⁡(det⁡(f~(T)),0)=0\operatorname{wind}(\det(\widetilde{f}(\mathbb{T})),0)=0. Therefore, by Theorem 2.3, the operator T(f~)T(\widetilde{f}) is either invertible or has a nontrivial kernel. By explicitly writing down T(f~)T(\widetilde{f})

T(f~)=(0100112002010011⋱),T(\widetilde{f})=\begin{pmatrix} 0 & 1 & 0 & 0\\ 1 & 1 & 2 & 0\\ 0 & 2 & 0 & 1\\ 0 & 0 & 1 & 1\\ & & & \ddots \end{pmatrix},

we find that u~=(1,0,−12,0,14,…)T\widetilde{\mathbf{u}}=(1,0,-\frac{1}{2},0,\frac{1}{4},\ldots)^{\mathsf{T}} satisfies T(f~)u~=0T(\widetilde{f})\widetilde{\mathbf{u}}=0. Then in this case, u~∈ℓ2\widetilde{\mathbf{u}}\in\ell^2 and 0∈σ(T(f))0\in\sigma(T(f)).

Remark 2.7. Note that (13) and (14) also provide ways to compute σwind(f)\sigma_{\mathrm{wind}}(f) explicitly. We will utilize (14) in Section 4 for numerical illustrations of the topological origin of the skin effect.

For the sake of completeness, we will devote the last part in this section to tridiagonal kk-Laurent operators that are defined as

L(f)=(⋱⋱⋱A1A0A−1A1A0A−1⋱⋱⋱)(17)L(f)=\begin{pmatrix} \ddots& \ddots& \ddots\\ & A_1 & A_0 & A_{-1}\\ & & A_1 & A_0 & A_{-1}\\ & & & \ddots& \ddots& \ddots \end{pmatrix} \tag*{(17)}

with the generating symbol ff being (5). Using the identification L(f−λ)=L(f)−λL(f-\lambda)=L(f)-\lambda, which we have used to characterize the spectrum of tridiagonal kk-Toeplitz operators, one can obtain an explicit characterization of the spectrum of tridiagonal kk-Laurent operators as a consequence of [15 Chapter 23, Corollary 2.5]. The following theorem holds.

Theorem 2.8. Let f∈(L(T)∞)k×kf\in(\mathcal{L}(\mathbb{T})^\infty)^{k\times k} and denote the associated tridiagonal kk-Laurent operator L(f)L(f). Then, it holds that

σ(L(f))=σess(L(f))=σdet⁡(f)\sigma(L(f))=\sigma_{\mathrm{ess}}(L(f))=\sigma_{\det}(f)

with σdet⁡(f)\sigma_{\det}(f) being defined by (11).

Eigenvectors of tridiagonal kk-Toeplitz operators

Now that we have characterized the spectra of tridiagonal kk-Toeplitz operators, we will venture onwards to explore their eigenvectors. The following theorem is the second main result of this paper.

Theorem 2.9. Suppose ∏j=1kcj≠0\prod_{j=1}^{k}c_j \ne0 and ∏j=1kbj≠0\prod_{j=1}^{k}b_j \ne0. Let f(z)∈Ck×kf(z) \in\mathbb{C}^{k \times k} be the symbol (5) and let λ∈C∖σess(T(f))\lambda\in\mathbb{C} \setminus\sigma_{\mathrm{ess}}(T(f)). If ∑j=1kwind⁡(λj(T),λ)<0\sum_{j=1}^{k}\operatorname{wind}(\lambda_j(T),\lambda)<0 for λj\lambda_j being the eigenvalues of f(z)f(z), then there exists an eigenvector x∈ℓ2x \in\ell^2 of T(f)T(f) associated to λ\lambda and some ρ<1\rho<1 such that

∣xj∣max⁡i∣xi∣≤Cρ⌈j/k⌉−1,j≥1,(18)\frac{|x_j|}{\max_i |x_i|} \le C \rho^{\lceil j/k\rceil-1}, \qquad j \ge1, \tag*{(18)}

where C>0C>0 is a constant depending only on λ,ap,bp,cp\lambda,a_p,b_p,c_p for 1≤p≤k1\le p\le k.

Proof. Let λ∈C\lambda\in\mathbb{C} be such that ∑j=1kwind⁡(λj(T),λ)=wind⁡(det⁡(f(T)−λ),0)<0\sum_{j=1}^{k}\operatorname{wind}(\lambda_j(T),\lambda)=\operatorname{wind}(\det(f(T)-\lambda),0)<0. Since ∏i=1kbi≠0,∏i=1kci≠0\prod_{i=1}^{k}b_i\ne0,\prod_{i=1}^{k}c_i\ne0, by Lemma A.1, det⁡(f(z)−λ)\det(f(z)-\lambda) is a meromorphic function in C\mathbb{C}, with poles at 00 and ∞\infty. Then the argument principle implies that det⁡(f(z)−λ)=0\det(f(z)-\lambda)=0 for two values z1,z2z_1,z_2 with 1<1ρ≤∣zj∣1<\frac{1}{\rho}\le|z_j|, j=1,2j=1,2, counted with multiplicity. Moreover, since det⁡(f(zi)−λ)=0\det(f(z_i)-\lambda)=0, we find that λ\lambda is an eigenvalue to f(zi)f(z_i) and hence there exists an associated eigenvector vi\mathbf{v}_i. Assume for now that z1≠z2z_1\ne z_2. Consider the vectors defined by

ui=(viT,zi−1viT,zi−2viT,…)T,i∈{1,2}.(19)\mathbf{u}_i=\left(\mathbf{v}_i^{\mathsf T},z_i^{-1}\mathbf{v}_i^{\mathsf T},z_i^{-2}\mathbf{v}_i^{\mathsf T},\ldots\right)^{\mathsf T},\qquad i\in\{1,2\}. \tag*{(19)}

The vector ui\mathbf{u}_i satisfies the eigenvalue equation T(f)ui=λuiT(f)\mathbf{u}_i=\lambda\mathbf{u}_i in all but the first row. Since u1,u2\mathbf{u}_1,\mathbf{u}_2 are linearly independent, there exists a linear combination of them that is indeed an eigenvector of T(f)T(f).

Next, we consider the case where z1=z2z_1=z_2 and first suppose that dim⁡Eig⁡(f(z1),λ)>1\dim\operatorname{Eig}(f(z_1),\lambda)>1, where Eig⁡(f(z1),λ)\operatorname{Eig}(f(z_1),\lambda) denotes the eigenspace of f(z1)f(z_1) associated to the eigenvalue λ\lambda. Then there exist linearly independent eigenvectors v1,v2\mathbf{v}_1,\mathbf{v}_2 of f(z1)f(z_1) associated to λ\lambda and we can use

ui=(viT,z1−1viT,z1−2viT,…)T,i∈{1,2},(20)\mathbf{u}_i=\left(\mathbf{v}_i^{\mathsf T},z_1^{-1}\mathbf{v}_i^{\mathsf T},z_1^{-2}\mathbf{v}_i^{\mathsf T},\ldots\right)^{\mathsf T},\qquad i\in\{1,2\}, \tag*{(20)}

to construct an exponentially decaying eigenvector of T(f)T(f). Finally, it remains to treat the case when z1=z2z_1=z_2 and dim⁡Eig⁡(f(z1),λ)=1\dim\operatorname{Eig}(f(z_1),\lambda)=1. Denote v1\mathbf{v}_1 the eigenvector of f(z1)f(z_1) associated to λ\lambda. We will construct a vector v2\mathbf{v}_2 such that

u2=(v2T,z1−1v2T+v1T,z1−2v2T+2z1−1v1T,…)T(21)\mathbf{u}_2=\left(\mathbf{v}_2^{\mathsf T},z_1^{-1}\mathbf{v}_2^{\mathsf T}+\mathbf{v}_1^{\mathsf T},z_1^{-2}\mathbf{v}_2^{\mathsf T}+2z_1^{-1}\mathbf{v}_1^{\mathsf T},\ldots\right)^{\mathsf T} \tag*{(21)}

satisfies the eigenvalue problem in all but the first row. Similarly to the previous arguments, this is enough to construct the eigenvector of T(f)T(f). By explicitly writing down (T(f)−λ)u2=0(T(f)-\lambda)\mathbf{u}_2=0 (except for the first row), we find that v2\mathbf{v}_2 must satisfy the condition

(A1+(A0−λ)z1−1+A−1z1−2)v2=−((A0−λ)+2A−1z1−1)v1,(22)\left(A_1+(A_0-\lambda)z_1^{-1}+A_{-1}z_1^{-2}\right)\mathbf{v}_2=-\left((A_0-\lambda)+2A_{-1}z_1^{-1}\right)\mathbf{v}_1, \tag*{(22)}

which is also a sufficient condition. From Lemma A.1, it is not hard to see that for all λ~\widetilde{\lambda} in a small neighborhood of λ\lambda, the equation det⁡(f(z)−λ~)\det(f(z)-\widetilde{\lambda}) has two distinct roots. Let (εj)j(\varepsilon_j)_j be a sequence converging to zero. We define λ(εj)=λ+εj\lambda(\varepsilon_j)=\lambda+\varepsilon_j. Then by the previous observation, we may assume that to each λ(εj)\lambda(\varepsilon_j) there exist two distinct roots z1(εj),z2(εj)z_1(\varepsilon_j),z_2(\varepsilon_j) of the equation det⁡(f(z)−λ(εj))=0\det(f(z)-\lambda(\varepsilon_j))=0. Let w1(εj),w2(εj)\mathbf{w}_1(\varepsilon_j),\mathbf{w}_2(\varepsilon_j) be the corresponding unit eigenvectors. We may assume that

w1(εj)→v0,w2(εj)→v0.(23)\mathbf{w}_1(\varepsilon_j)\to\mathbf{v}_0,\qquad\mathbf{w}_2(\varepsilon_j)\to\mathbf{v}_0. \tag*{(23)}

The limit of wi(εj)\mathbf{w}_i(\varepsilon_j) (or a subsequence thereof) exists by a standard compactness argument and it must be an eigenvector of f(z1)f(z_1) associated to λ\lambda. In particular, for each jj it holds that

(A1+(A0−λ(εj))z1(εj)−1+A−1z1(εj)−2)w1(εj)=0,\left(A_1+(A_0-\lambda(\varepsilon_j))z_1(\varepsilon_j)^{-1}+A_{-1}z_1(\varepsilon_j)^{-2}\right)\mathbf{w}_1(\varepsilon_j)=0,
(A1+(A0−λ(εj))z2(εj)−1+A−1z2(εj)−2)w2(εj)=0.\left(A_1+(A_0-\lambda(\varepsilon_j))z_2(\varepsilon_j)^{-1}+A_{-1}z_2(\varepsilon_j)^{-2}\right)\mathbf{w}_2(\varepsilon_j)=0.

We define

Δz(εj)=z1(εj)−1−z2(εj)−1.\Delta z(\varepsilon_j)=z_1(\varepsilon_j)^{-1}-z_2(\varepsilon_j)^{-1}.

Then we have

A1w1(εj)−w2(εj)Δz(εj)+(A0−λ(εj))z1(εj)−1w1(εj)−z2(εj)−1w2(εj)Δz(εj)+A−1z1(εj)−2w1(εj)−z2(εj)−2w2(εj)Δz(εj)=0.A_1\frac{\mathbf{w}_1(\varepsilon_j)-\mathbf{w}_2(\varepsilon_j)}{\Delta z(\varepsilon_j)} +(A_0-\lambda(\varepsilon_j))\frac{z_1(\varepsilon_j)^{-1}\mathbf{w}_1(\varepsilon_j)-z_2(\varepsilon_j)^{-1}\mathbf{w}_2(\varepsilon_j)}{\Delta z(\varepsilon_j)} +A_{-1}\frac{z_1(\varepsilon_j)^{-2}\mathbf{w}_1(\varepsilon_j)-z_2(\varepsilon_j)^{-2}\mathbf{w}_2(\varepsilon_j)}{\Delta z(\varepsilon_j)}=0.

::: Further expansion of this expression then yields

A1w1(εj)−w2(εj)Δz(εj)+(A0−λ(εj))z1(εj)−1(w1(εj)−w2(εj))Δz(εj)+A−1z1(εj)−2(w1(εj)−w2(εj))Δz(εj)=−((A0−λ(εj))(z1(εj)−1−z2(εj)−1)w2(εj)Δz(εj)+A−1(z1(εj)−2−z2(εj)−2)w2(εj)Δz(εj)).(24)\begin{aligned} A_{1}\frac{\mathbf{w}_{1}(\varepsilon_j)-\mathbf{w}_{2}(\varepsilon_j)}{\Delta z(\varepsilon_j)} + (A_{0}-\lambda(\varepsilon_j))\frac{z_{1}(\varepsilon_j)^{-1}(\mathbf{w}_{1}(\varepsilon_j)-\mathbf{w}_{2}(\varepsilon_j))}{\Delta z(\varepsilon_j)} + A_{-1}\frac{z_{1}(\varepsilon_j)^{-2}(\mathbf{w}_{1}(\varepsilon_j)-\mathbf{w}_{2}(\varepsilon_j))}{\Delta z(\varepsilon_j)} \\ = -\left((A_{0}-\lambda(\varepsilon_j))\frac{(z_{1}(\varepsilon_j)^{-1}-z_{2}(\varepsilon_j)^{-1})\mathbf{w}_{2}(\varepsilon_j)}{\Delta z(\varepsilon_j)} + A_{-1}\frac{(z_{1}(\varepsilon_j)^{-2}-z_{2}(\varepsilon_j)^{-2})\mathbf{w}_{2}(\varepsilon_j)}{\Delta z(\varepsilon_j)}\right). \tag*{(24)} \end{aligned}

Consider now the sequence w1(εj)−w2(εj)Δz(εj)\frac{\mathbf{w}_{1}(\varepsilon_j)-\mathbf{w}_{2}(\varepsilon_j)}{\Delta z(\varepsilon_j)}. If a bounded subsequence exists, we may define v2\mathbf{v}_{2} as the limit of this subsequence. In this case, it is not hard to see that v2\mathbf{v}_{2} satisfies the condition (2.22). Otherwise, there exists a subsequence of w1(εj)−w2(εj)Δz(εj)\frac{\mathbf{w}_{1}(\varepsilon_j)-\mathbf{w}_{2}(\varepsilon_j)}{\Delta z(\varepsilon_j)} whose norm tends to infinity. Define

Δw(εj)=∥w1(εj)−w2(εj)∥.\Delta w(\varepsilon_j)=\left\|\mathbf{w}_{1}(\varepsilon_j)-\mathbf{w}_{2}(\varepsilon_j)\right\|.

Then (upon passing to said subsequence) it holds that

Δz(εj)Δw(εj)→0.(25)\frac{\Delta z(\varepsilon_j)}{\Delta w(\varepsilon_j)}\to0. \tag*{(25)}

Multiplying (2.24) with this yields

A1w1(εj)−w2(εj)Δw(εj)+(A0−λ(εj))z1(εj)−1(w1(εj)−w2(εj))Δw(εj)+A−1z1(εj)−2(w1(εj)−w2(εj))Δw(εj)=−((A0−λ(εj))(z1(εj)−1−z2(εj)−1)w2(εj)Δw(εj)+A−1(z1(εj)−2−z2(εj)−2)w2(εj)Δw(εj)).\begin{aligned} A_{1}\frac{\mathbf{w}_{1}(\varepsilon_j)-\mathbf{w}_{2}(\varepsilon_j)}{\Delta w(\varepsilon_j)} + (A_{0}-\lambda(\varepsilon_j))\frac{z_{1}(\varepsilon_j)^{-1}(\mathbf{w}_{1}(\varepsilon_j)-\mathbf{w}_{2}(\varepsilon_j))}{\Delta w(\varepsilon_j)} + A_{-1}\frac{z_{1}(\varepsilon_j)^{-2}(\mathbf{w}_{1}(\varepsilon_j)-\mathbf{w}_{2}(\varepsilon_j))}{\Delta w(\varepsilon_j)} \\ = -\left((A_{0}-\lambda(\varepsilon_j))\frac{(z_{1}(\varepsilon_j)^{-1}-z_{2}(\varepsilon_j)^{-1})\mathbf{w}_{2}(\varepsilon_j)}{\Delta w(\varepsilon_j)} + A_{-1}\frac{(z_{1}(\varepsilon_j)^{-2}-z_{2}(\varepsilon_j)^{-2})\mathbf{w}_{2}(\varepsilon_j)}{\Delta w(\varepsilon_j)}\right). \end{aligned}

In the limit, the right-hand side tends to 0. Thus v2=lim⁡εj→0w1(εj)−w2(εj)Δw(εj)\mathbf{v}_{2}=\lim_{\varepsilon_j\to0}\frac{\mathbf{w}_{1}(\varepsilon_j)-\mathbf{w}_{2}(\varepsilon_j)}{\Delta w(\varepsilon_j)} satisfies

(A1+(A0−λ)z1−1+A−1z1−2)v2=0.\left(A_{1}+(A_{0}-\lambda)z_{1}^{-1}+A_{-1}z_{1}^{-2}\right)\mathbf{v}_{2}=0.

This indicates that v2\mathbf{v}_{2} is an eigenvector. But from the construction of v2\mathbf{v}_{2}, it becomes apparent that v2\mathbf{v}_{2} must be orthogonal to v1\mathbf{v}_{1}, which contradicts the assumption that the eigenspace of f(z1)f(z_{1}) associated with λ\lambda is one-dimensional. Therefore, the sequence w1(εj)−w2(εj)Δz(εj)\frac{\mathbf{w}_{1}(\varepsilon_j)-\mathbf{w}_{2}(\varepsilon_j)}{\Delta z(\varepsilon_j)} is uniformly bounded and arguments for the first case already demonstrate the existence of such v2\mathbf{v}_{2} satisfying (2.22).

The exponential bound in (2.18) is clear by the construction of the vectors u1\mathbf{u}_{1} and u2\mathbf{u}_{2}. ■

Remark 2.10. Note that the constructive approach used in the proof of [21 Theorem 7.2] does not work here. We have used a new approximation approach to construct the eigenvectors.

Remark 2.11. For the case when ∑j=1kwind⁡(λj(T),λ)>0\sum_{j=1}^{k}\operatorname{wind}(\lambda_j(T),\lambda)>0, λ∈σ(T(f))\lambda\in\sigma(T(f)) is because T(f)T(f) has a non-dense image. In particular, one can construct the eigenvector x\mathbf{x} of T(f)−λT(f)-\lambda in the same spirit of Theorem 2.9, which implies T(f)−λT(f)-\lambda has a non-dense image.

Pseudo-spectra of tridiagonal kk-Toeplitz matrices

In this section, we will apply the results of the previous section to matrices. As already described by Trefethen in [20], the spectrum of non-Hermitian matrices is very sensitive to small perturbations. For this reason, it is often advisable to study their pseudo-spectra, rather than their spectra. For the convenience of the reader, we recall the definition of pseudoeigenvalues and pseudoeigenvectors.

Definition 3.1. Let ε>0\varepsilon>0. Then, λ∈C\lambda\in\mathbb{C} is an ε\varepsilon-pseudoeigenvalue of A∈CN×NA\in\mathbb{C}^{N\times N} if one of the following conditions is satisfied:

  • λ\lambda is a proper eigenvalue of A+EA+E for some E∈CN×NE\in\mathbb{C}^{N\times N} such that ∥E∥≤ε\lVert E\rVert\leq\varepsilon;

  • ∥(A−λI)u∥<ε\lVert(A-\lambda I)\mathbf{u}\rVert<\varepsilon for some vector u\mathbf{u} with ∥u∥=1\lVert\mathbf{u}\rVert=1;

  • ∥(A−λI)−1∥−1≤ε\lVert(A-\lambda I)^{-1}\rVert^{-1}\leq\varepsilon.

The set of all ε\varepsilon-pseudoeigenvalues of A\mathbf{A}, the ε\varepsilon-pseudospectrum, is denoted by σε(A)\sigma_{\varepsilon}(\mathbf{A}). If some nonzero u\mathbf{u} satisfies ∥(λ−A)u∥<ε\lVert(\lambda-\mathbf{A})\mathbf{u}\rVert<\varepsilon, then we say that u\mathbf{u} is an ε\varepsilon-pseudoeigenvector of A\mathbf{A}.

By the equivalence of norms in finite dimensions, any norm can be used. For a general treatment of pseudospectra, see [21].

We consider the family of tridiagonal kk-Toeplitz matrices {AN}\{\mathbf{A}_{N}\} of various dimensions obtained as N×NN\times N finite sections of the infinite operator T(f)T(f). Results on the pseudospectra of the family of Toeplitz matrices can be found in [20]. On the other hand, the spectrum of tridiagonal kk-Toeplitz matrices is fully understood and closed form characteristic polynomials have been found in [13].

Let us also introduce the sets

Ωr:={λ∈C:wind⁡(det⁡(f(Tr)−λ),0)>0},(26)\Omega_{r}:=\{\lambda\in\mathbb{C}:\operatorname{wind}(\det(f(\mathbb{T}_{r})-\lambda),0)>0\}, \tag*{(26)}
ΩR:={λ∈C:wind⁡(det⁡(f(TR)−λ),0)<0},(27)\Omega^{R}:=\{\lambda\in\mathbb{C}:\operatorname{wind}(\det(f(\mathbb{T}_{R})-\lambda),0)<0\}, \tag*{(27)}

where Tr={z∈C:∣z∣=r}\mathbb{T}_{r}=\{z\in\mathbb{C}:|z|=r\} and let σε(AN)\sigma_{\varepsilon}(\mathbf{A}_{N}) be the set of ε\varepsilon-pseudoeigenvalues of AN\mathbf{A}_{N}.

The connection between the eigenvectors of a tridiagonal kk-Toeplitz operator and those of the associated tridiagonal kk-Toeplitz matrix can be made with an argument similar to [21 Theorem 7.2]. The first NN components of an eigenvector of the Toeplitz operator can be used as the ε\varepsilon-pseudoeigenvector of the tridiagonal kk-Toeplitz matrix AN\mathbf{A}_{N}. Since we have proven the existence of an exponentially decaying eigenvector in Theorem 2.9, we can conclude with the following statement on tridiagonal kk-Toeplitz matrices.

Theorem 3.2. Let {AN}\{\mathbf{A}_{N}\} be a family of tridiagonal kk-Toeplitz matrices such that ∏i=1kbi≠0\prod_{i=1}^{k}b_{i}\ne0, ∏i=1kci≠0\prod_{i=1}^{k}c_{i}\ne0. Then, for any r<1r<1 and ρ>r\rho>r, we have

Ωr∪Ω1/r∪(σ(T(f))+Δε)⊆σε(AN)forε=max⁡(C1,⌈N/k⌉C2)ρ⌈N/k⌉−1,(28)\Omega_{r}\cup\Omega^{1/r}\cup\left(\sigma(T(f))+\Delta_{\varepsilon}\right)\subseteq\sigma_{\varepsilon}(\mathbf{A}_{N})\qquad\text{for}\qquad\varepsilon=\max(C_{1},\lceil N/k\rceil C_{2})\rho^{\lceil N/k\rceil-1}, \tag*{(28)}

where Δε\Delta_{\varepsilon} is a closed unit disk of radius ε\varepsilon and C1,C2C_{1},C_{2} are constants depending on rr, ρ\rho, and aja_{j}, bjb_{j}, cjc_{j} for 1≤j≤k1\le j\le k but independent of NN. In particular, for the corresponding λ∈Ωr∪Ω1/r\lambda\in\Omega_{r}\cup\Omega^{1/r}, there exist nonzero pseudoeigenvectors v(N)\mathbf{v}^{(N)} satisfying

∥(AN−λ)v(N)∥∥v(N)∥≤max⁡(C3,⌈N/k⌉C4)ρ⌈N/k⌉−1(29)\frac{\lVert(\mathbf{A}_{N}-\lambda)\mathbf{v}^{(N)}\rVert}{\lVert\mathbf{v}^{(N)}\rVert}\le\max(C_{3},\lceil N/k\rceil C_{4})\rho^{\lceil N/k\rceil-1} \tag*{(29)}

such that

∣vj(N)∣max⁡i∣vi(N)∣≤{C5⌈N/k⌉ρ⌈j/k⌉−1,if λ∈Ω1/r,C5⌈N/k⌉ρ⌈(N−j)/k⌉−1,if λ∈Ωr,1≤j≤N,(30)\frac{\lvert v_{j}^{(N)}\rvert}{\max_{i}\lvert v_{i}^{(N)}\rvert}\le \begin{cases} C_{5}\lceil N/k\rceil\rho^{\lceil j/k\rceil-1}, & \text{if }\lambda\in\Omega^{1/r},\\ C_{5}\lceil N/k\rceil\rho^{\lceil(N-j)/k\rceil-1}, & \text{if }\lambda\in\Omega_{r}, \end{cases} \qquad1\le j\le N, \tag*{(30)}

where C3,C4,C5C_{3},C_{4},C_{5} are constants independent of NN.

Proof. Firstly, we note that the inclusion σ(T(f))+Δε⊆σε(AN)\sigma(T(f))+\Delta_{\varepsilon}\subseteq\sigma_{\varepsilon}(\mathbf{A}_{N}) is triviality valid for any matrix or operator. Secondly, by symmetry, Ωr\Omega_{r} must satisfy an estimate of the type (28) if Ω1/r\Omega^{1/r} does. Thus all that we have to prove is Ω1/r⊆σε(AN)\Omega^{1/r}\subseteq\sigma_{\varepsilon}(\mathbf{A}_{N}).

Given any r<1r<1, let λ∈Ω1/r\lambda\in\Omega^{1/r} be arbitrary. Since wind⁡(det⁡(f(T1/r)−λ),0)<0\operatorname{wind}(\det(f(\mathbb{T}_{1/r})-\lambda),0)<0, similarly to the arguments in the proof of Theorem 2.9, we can show that det⁡(f(z)−λ)=0\det(f(z)-\lambda)=0 for two values z1,z2z_{1},z_{2} with 1/ρ≤∣zj∣1/\rho\le|z_{j}|, j=1,2j=1,2, counted with multiplicity. Also, since det⁡(f(z)−λ)=0\det(f(z)-\lambda)=0, we find that λ\lambda is an eigenvalue of f(zi)f(z_{i}) and hence there exists an associated eigenvector vi\mathbf{v}_{i}. Assume for now that z1≠z2z_{1}\ne z_{2}. Consider the vectors defined by

ui=(viT,zi−1viT,zi−2viT,…,zi−⌈N/k⌉+1viT)Ti∈{1,2},(31)\mathbf{u}_{i}=\left(\mathbf{v}_{i}^{\mathsf{T}},z_{i}^{-1}\mathbf{v}_{i}^{\mathsf{T}},z_{i}^{-2}\mathbf{v}_{i}^{\mathsf{T}},\ldots,z_{i}^{-\lceil N/k\rceil+1}\mathbf{v}_{i}^{\mathsf{T}}\right)^{\mathsf{T}}\qquad i\in\{1,2\}, \tag*{(31)}

where without loss of generality we assume that NN is divisible by kk. This vector satisfies the eigenvalue equation ANui=λui\mathbf{A}_{N}\mathbf{u}_{i}=\lambda\mathbf{u}_{i} in all but the first and last rows. Since u1,u2\mathbf{u}_{1},\mathbf{u}_{2} are linearly independent, there exists a normalized linear combination v(N)=a1u1+a2u2\mathbf{v}^{(N)}=a_{1}\mathbf{u}_{1}+a_{2}\mathbf{u}_{2} so that (AN−λ)v(N)=0(\mathbf{A}_{N}-\lambda)\mathbf{v}^{(N)}=\mathbf{0} in all but the last row. It is not hard to see that

∥(AN−λ)v(N)∥≤C3max⁡{∣z1∣−⌈N/k⌉+1,∣z2∣−⌈N/k⌉+1},(32)\lVert(\mathbf{A}_{N}-\lambda)\mathbf{v}^{(N)}\rVert\le C_{3}\max\{|z_{1}|^{-\lceil N/k\rceil+1},|z_{2}|^{-\lceil N/k\rceil+1}\}, \tag*{(32)}

where C3C_3 depends on aja_j, bjb_j, cjc_j, λ\lambda but not on NN. Then since 1<1ρ≤∣zj∣1 < \frac{1}{\rho} \leq\lvert z_j\rvert, j=1,2j=1,2, we obtain the bound

∥(AN−λ)v(N)∥≤C3ρ⌈N/k⌉−1,(33)\left\lVert(\boldsymbol{A}_N-\lambda)\boldsymbol{v}^{(N)}\right\rVert\leq C_3 \rho^{\lceil N/k\rceil-1}, \tag*{(33)}

which proves the claim (3.4) if z1≠z2z_1 \ne z_2. If the roots z1,z2z_1,z_2 are distinct for all λ∈Ω1/r‾\lambda\in\overline{\Omega^{1/r}}, then the function

sup⁡N≥1σN(λI−AN)r⌈N/k⌉−1\sup_{N\geq1}\frac{\sigma_N(\lambda I-\boldsymbol{A}_N)}{r^{\lceil N/k\rceil-1}}

is a continuous function of λ\lambda on the compact set Ω1/r‾\overline{\Omega^{1/r}}; its maximum provides the constant C1C_1 in (3.3) that independent of λ\lambda, NN, and we can take ρ=r\rho=r.

In the case where z1=z2z_1=z_2, one can resort to a similar argument as the one in Theorem 2.9 to utilize

u1=(v1T,z1−1v1T,z1−2v1T,…,z1−⌈N/k⌉+1v1T)T,\boldsymbol{u}_1=\left(\boldsymbol{v}_1^{\mathsf{T}},z_1^{-1}\boldsymbol{v}_1^{\mathsf{T}},z_1^{-2}\boldsymbol{v}_1^{\mathsf{T}},\ldots,z_1^{-\lceil N/k\rceil+1}\boldsymbol{v}_1^{\mathsf{T}}\right)^{\mathsf{T}},
u2=(v2T,z1−1v2T+v1T,z1−2v2T+2z1−1v1T,…,z1−⌈N/k⌉+1v2T+(⌈Nk⌉−1)z1−⌈N/k⌉+2v1T)T.(34)\boldsymbol{u}_2=\left(\boldsymbol{v}_2^{\mathsf{T}},z_1^{-1}\boldsymbol{v}_2^{\mathsf{T}}+\boldsymbol{v}_1^{\mathsf{T}},z_1^{-2}\boldsymbol{v}_2^{\mathsf{T}}+2z_1^{-1}\boldsymbol{v}_1^{\mathsf{T}},\ldots,z_1^{-\lceil N/k\rceil+1}\boldsymbol{v}_2^{\mathsf{T}}+\left(\left\lceil\frac{N}{k}\right\rceil-1\right)z_1^{-\lceil N/k\rceil+2}\boldsymbol{v}_1^{\mathsf{T}}\right)^{\mathsf{T}}. \tag*{(34)}

to construct the pseudoeigenvector. This yields a similar bound except with C3ρ⌈N/k⌉−1C_3\rho^{\lceil N/k\rceil-1} in (3.4) replaced by an algebraically growing factor at worst ⌈N/k⌉C4ρ⌈N/k⌉−1\lceil N/k\rceil C_4\rho^{\lceil N/k\rceil-1}. Then since the function

sup⁡N≥1σN(λI−AN)⌈N/k⌉r⌈N/k⌉−1\sup_{N\geq1}\frac{\sigma_N(\lambda I-\boldsymbol{A}_N)}{\lceil N/k\rceil r^{\lceil N/k\rceil-1}}

is a continuous function of λ\lambda on the compact set Ω1/r‾\overline{\Omega^{1/r}}, its maximum provides the constant C2C_2 in (3.3). Finally, (3.5) is deduced from construction (3.6) and (3.9). ■

Topological origin of the skin effect in polymer systems of subwavelength resonators

The non-Hermitian skin effect is the phenomenon whereby in the subwavelength regime a large proportion of the bulk eigenmodes of a non-Hermitian open system of subwavelength resonators are localised at one edge of the structure [8, 19]. It has been realized experimentally in topological photonics, phononics, and other condensed matter systems [14, 17, 19]. Its significance lies in its substantial contribution to advancing the realm of active metamaterials, paving the way for novel opportunities to guide and control energy on subwavelength scales [3, 6, 7, 12].

Recently, the non-Hermitian skin effect has been demonstrated in one-dimensional systems of subwavelength resonators using first-principle mathematical analysis. In particular, [1] demonstrates the skin effect for the monomer case when an imaginary gauge potential is introduced inside the resonators to break Hermiticity and [5] generalizes the results in [1] to dimer systems. We also refer to [2] for the skin effect in a three-dimensional setting and to [4] for stability analysis.

By the results obtained in the previous sections, we provide the topological origin of the skin effect in the one-dimensional polymer system, i.e., in systems with periodically repeated cells of kk resonators. This also elucidates the numerical findings in [5] for the topological origin of the skin effect.

Let us first introduce the model setting. We consider a chain of NN disjoint one-dimensional resonators D:=(xiL,xiR)D:=\left(x_i^{L},x_i^{R}\right), where (xiL,xiR)1≤i≤N⊂R\left(x_i^{L},x_i^{R}\right)_{1\leq i\leq N}\subset\mathbb{R} are the 2N2N extremities such that xiL<xiR<xi+1Lx_i^{L}<x_i^{R}<x_{i+1}^{L} for any 1≤i≤N1\leq i\leq N. The lengths of the ii-th resonators will be denoted by ℓi=xiR−xiL\ell_i=x_i^{R}-x_i^{L} and the spacings between the ii-th and the (i+1)(i+1)-th resonators will be denoted by si=xi+1L−xiRs_i=x_{i+1}^{L}-x_i^{R}. Furthermore, we assume that kk resonators are periodically repeated, that is, si+k=sis_{i+k}=s_i. The resulting system is illustrated in Figure 4.1.

Image not extracted

Figure 4.1. A chain of NN times periodically repeated kk subwavelength resonators, with lengths (ℓi)1≤i≤N(\ell_i)_{1\le i\le N} and spacings (si)1≤i≤N−1(s_i)_{1\le i\le N-1}.

Let DD be the domain consisting of the union of all resonators

D=⋃i=1N(xiL,xiR)⊂R.(35)D=\bigcup_{i=1}^{N}\left(x_i^{L},x_i^{R}\right)\subset\mathbb{R}. \tag*{(35)}

The underlying wave equation that governs the system is given by

{u′′(x)+γu′(x)+ω2vb2u=0,x∈D,u′′(x)+ω2v2u=0,x∈R\D,u∣R(xiL,R)−u∣L(xiL,R)=0,for all 1≤i≤N,dudx∣R(xiL)=δdudx∣L(xiL),for all 1≤i≤N,dudx∣L(xiR)=δdudx∣R(xiR),for all 1≤i≤N,dud∣x∣−iωvu=0,at x=x1L,xNR, respectively.(36)\begin{cases} u''(x)+\gamma u'(x)+\frac{\omega^2}{v_b^2}u=0, & x\in D,\\ u''(x)+\frac{\omega^2}{v^2}u=0, & x\in\mathbb{R}\backslash D,\\ u|_{\mathrm{R}}\left(x_i^{\mathrm{L,R}}\right)-u|_{\mathrm{L}}\left(x_i^{\mathrm{L,R}}\right)=0, & \text{for all }1\le i\le N,\\ \left.\frac{\mathrm{d}u}{\mathrm{d}x}\right|_{\mathrm{R}}\left(x_i^{\mathrm{L}}\right)=\delta\left.\frac{\mathrm{d}u}{\mathrm{d}x}\right|_{\mathrm{L}}\left(x_i^{\mathrm{L}}\right), & \text{for all }1\le i\le N,\\ \left.\frac{\mathrm{d}u}{\mathrm{d}x}\right|_{\mathrm{L}}\left(x_i^{\mathrm{R}}\right)=\delta\left.\frac{\mathrm{d}u}{\mathrm{d}x}\right|_{\mathrm{R}}\left(x_i^{\mathrm{R}}\right), & \text{for all }1\le i\le N,\\ \frac{\mathrm{d}u}{\mathrm{d}|x|}-\mathrm{i}\frac{\omega}{v}u=0, & \text{at }x=x_1^{\mathrm{L}},x_N^{\mathrm{R}},\text{ respectively}. \tag*{(36)} \end{cases}

where 0<δ≪10<\delta\ll1, the wave speeds inside the resonator and in the background are respectively vbv_b and vv and are of order one, and γu′\gamma u' for γ∈R∗\gamma\in\mathbb{R}^* denotes the imaginary gauge potential. Here, for a function ww we denote by

w∣L(x):=lim⁡s→0,s>0w(x−s)andw∣R(x):=lim⁡s→0,s>0w(x+s)w|_{\mathrm{L}}(x):=\lim_{s\to0,s>0}w(x-s)\quad\text{and}\quad w|_{\mathrm{R}}(x):=\lim_{s\to0,s>0}w(x+s)

if the limits exist.

We are interested in nontrivial frequencies ω\omega such that (4.2) is satisfied and ω→0\omega\to0 as δ→0\delta\to0. Such ω\omega are called subwavelength resonances.

In the high-contrast, low-frequency setting described above, gauge capacitance matrices provide a discrete approximation of the subwavelength eigenfrequencies induced in the structure.

Definition 4.1. For γ∈R∗\gamma\in\mathbb{R}^*, we define the gauge capacitance matrix Cγ∈RN×NC^\gamma\in\mathbb{R}^{N\times N} by

Ci,jγ={γs1ℓ11−e−γℓ1,i=j=1,γsi−1ℓi1−e−γℓi−γsiℓi1−e−γℓi,1<i=j<N,−γsiℓi1−e−γℓi,1≤i=j−1≤N−1,γsi−1ℓi1−e−γℓi,2≤i=j+1≤N,−γsN−1ℓN1−e−γℓN,i=j=N,(37)C^\gamma_{i,j}= \begin{cases} \frac{\gamma}{s_1}\frac{\ell_1}{1-\mathrm{e}^{-\gamma\ell_1}}, & i=j=1,\\ \frac{\gamma}{s_{i-1}}\frac{\ell_i}{1-\mathrm{e}^{-\gamma\ell_i}}-\frac{\gamma}{s_i}\frac{\ell_i}{1-\mathrm{e}^{-\gamma\ell_i}}, & 1<i=j<N,\\ -\frac{\gamma}{s_i}\frac{\ell_i}{1-\mathrm{e}^{-\gamma\ell_i}}, & 1\le i=j-1\le N-1,\\ \frac{\gamma}{s_{i-1}}\frac{\ell_i}{1-\mathrm{e}^{-\gamma\ell_i}}, & 2\le i=j+1\le N,\\ -\frac{\gamma}{s_{N-1}}\frac{\ell_N}{1-\mathrm{e}^{-\gamma\ell_N}}, & i=j=N, \end{cases} \tag*{(37)}

and all the other entries are zero.

In [1 Corollary 2.6], one can find a formal argumentation that the gauge capacitance matrix gives indeed a valid approximation of the eigenmodes in a subwavelength regime as δ→0\delta\to0. For the remainder of this paper, we assume that γ=ℓi=1\gamma=\ell_i=1. Without loss of generality, we consider a system of kk resonators and obtain a gauge capacitance matrix of the tridiagonal kk-Toeplitz form (perturbed):

Ank(a,b)=(a1+ab10⋯⋯0c1a2b2⋱⋮0⋱⋱⋱⋱⋮⋮⋱ck−3ak−2bk−20⋮⋱ck−2ak−1bk−10⋯⋯0ck−1ak+b).(38)A_{n k}^{(a,b)}= \begin{pmatrix} a_1+a & b_1 & 0 & \cdots& \cdots& 0\\ c_1 & a_2 & b_2 & \ddots& & \vdots\\ 0 & \ddots& \ddots& \ddots& \ddots& \vdots\\ \vdots& \ddots& c_{k-3} & a_{k-2} & b_{k-2} & 0\\ \vdots& & \ddots& c_{k-2} & a_{k-1} & b_{k-1}\\ 0 & \cdots& \cdots& 0 & c_{k-1} & a_k+b \end{pmatrix}. \tag*{(38)}

For the semi-infinite systems we obtain a perturbed tridiagonal kk-Toeplitz operator

Ta(f)=(a1+ab1c1a2b2⋱⋱⋱ck−1akbkcka1b1⋱⋱⋱)T^{a}(f)=\begin{pmatrix} a_{1}+a & b_{1} \\ c_{1} & a_{2} & b_{2} \\ & \ddots& \ddots& \ddots\\ && c_{k-1} & a_{k} & b_{k} \\ &&& c_{k} & a_{1} & b_{1} \\ &&&& \ddots& \ddots& \ddots\end{pmatrix}

with ff being the symbol (5).

It has been observed and conjectured [5] that the exponential decay of the eigenvectors of kk-Toeplitz matrices for k≥2k \ge2 is due to the winding of the eigenvalues of the symbol being nontrivial. Now, by the spectral theory for the tridiagonal kk-Toeplitz operators introduced in the previous sections, we have the following results for the topological origin of the skin effect in subwavelength resonator systems, validating the conjecture made in [5].

Theorem 4.2. Suppose ∏j=1kcj≠0\prod_{j=1}^{k}c_{j}\ne0 and ∏j=1kbj≠0\prod_{j=1}^{k}b_{j}\ne0. Let f(z)∈Ck×kf(z)\in\mathbb{C}^{k\times k} be the symbol (5) and let λ∈C∖σess(Ta(f))\lambda\in\mathbb{C}\setminus\sigma_{\mathrm{ess}}(T^{a}(f)). If ∑j=1kwind⁡(λj(T),λ)<0\sum_{j=1}^{k}\operatorname{wind}(\lambda_{j}(\mathbb{T}),\lambda)<0, then there exists an eigenvector xx of Ta(f)T^{a}(f) associated to λ\lambda and some ρ<1\rho<1 such that

∣xj∣max⁡i∣xi∣≤C⌈jk⌉ρ⌈j/k⌉−1,(39)\frac{|x_{j}|}{\max_{i}|x_{i}|}\le C\left\lceil\frac{j}{k}\right\rceil\rho^{\left\lceil j/k\right\rceil-1}, \tag*{(39)}

where C>0C>0 is a constant depending only on λ\lambda, aja_{j}, bjb_{j}, cjc_{j}, j=1,…,kj=1,\ldots,k. If ∑j=1kwind⁡(λj(T),λ)>0\sum_{j=1}^{k}\operatorname{wind}(\lambda_{j}(\mathbb{T}),\lambda)>0, then the above results hold for the left eigenvectors.

Proof. Although with a perturbation aa on the first element, the proof of Theorem 2.9 can still be applied. ■

Theorem 4.2 elucidates the topological origin of the skin effect in the polymer system of subwavelength resonators. In particular, the skin effect holds for all λ\lambda in the region

G:={λ∈C∖σdet(f):∑j=1kwind⁡(λj(T),λ)≠0}.(40)G:=\left\{\lambda\in\mathbb{C}\setminus\sigma_{\mathrm{det}}(f):\sum_{j=1}^{k}\operatorname{wind}(\lambda_{j}(\mathbb{T}),\lambda)\ne0\right\}. \tag*{(40)}

This is a generalization for the topological origin of the skin effect in the Toeplitz operator case given in [1].

The last part of the section is devoted to illustrating numerically the skin effect and its topological origin in chains of 2 and 3 periodically repeated resonators. We start by illustrating in Figure 4.2 the results of a system of 2 periodically repeated resonators as in [5]. In Figure 4.2A, we show the spectrum and pseudospectrum of the gauge capacitance matrix of a system of 25 dimers together with the winding of the two eigenvalues of the symbol of the corresponding 2-Toeplitz operator. Figure 4.2B shows that all the eigenvectors (black eigenvectors) associated with eigenvalues inside the region GG in (40) are localised and the only non-decreasing eigenvector (gray eigenvector) corresponds to the eigenvalue 0 in the boundary of the region GG. On the other hand, in [5] it is observed and conjectured that the non-trivial winding of the eigenvalues λj(z)\lambda_{j}(z) predicts the exponential decay of the eigenmodes. This is due to wind⁡(λj(T),λ)≤0\operatorname{wind}(\lambda_{j}(\mathbb{T}),\lambda)\le0, j=1,2j=1,2 in the example

G={λ∈C∖σdet(f):∣∏j=12wind⁡(λj(T),λ)∣≠0}.G=\left\{\lambda\in\mathbb{C}\setminus\sigma_{\mathrm{det}}(f):\left|\prod_{j=1}^{2}\operatorname{wind}(\lambda_{j}(\mathbb{T}),\lambda)\right|\ne0\right\}.

Figure 4.3 illustrates the results for 3 periodically repeated resonators. We numerically verify that indeed all the eigenvectors, except the one associated with eigenvalue 0 on the boundary of GG, are localised at the left edge of the structure.

Image not extracted

Figure 4.3. Figures A and C show the spectrum of the operator. The green regions consist of all the eigenvalues λ\lambda that satisfy ∑j=13Re⁡(λj(T),λ)≠0\sum_{j=1}^{3}\operatorname{Re}(\lambda_j(\mathcal{T}),\lambda) \ne0. The black dots along the real line denote the spectrum of the gauge capacitance matrix CγC^\gamma and the solid blue and orange lines around the spectrum are the ε\varepsilon-pseudospectra for ε=10k\varepsilon= 10^k and k=−5,−2k=-5,-2. Figures B and D show the eigenvectors of CγC^\gamma.

Image not extracted

Figure 4.2. The region of λ\lambda so that ∑j=1kwind⁡(λj(T),λ)≠0\sum_{j=1}^{k}\operatorname{wind}(\lambda_j(\mathrm{T}),\lambda) \ne0 and the localization of the eigenvectors. Computation performed for s1=1,s2=2s_1=1,s_2=2, and N=50N=50.

Remark 4.3. We remark that the green regions for k=1,2,3,4k=1,2,3,4 can be computed analytically through the formula (2.14). For the case when k≥5k\ge5, one can utilize (2.14) to compute numerically the green region.

Concluding remarks

We have developed new theories of spectra and pseudo-spectra for tridiagonal kk-Toeplitz operators and matrices. Specifically, we have established the relationship between the winding number of the symbol function’s eigenvalues and the exponential decay property of the corresponding eigenvectors (or pseudo-eigenvectors). This discovery sheds light on the topological origin of the non-Hermitian skin effect observed in one-dimensional polymer systems of subwavelength resonators. This paper also opens the door to the study of the spectral theory of block Toeplitz operators.

Acknowledgements

This work was partially supported by Swiss National Science Foundation grant number 200021-200307.

Code Availability

The data that support the findings of this work are openly available at https://doi.org/10.5281/zenodo.10438679.

Conflict of interest

The authors have no competing interests to declare that are relevant to the content of this article.

Appendix A. Computation of the determinant

Lemma A.1. For λ∈C∖σess(T(f))\lambda\in\mathbb{C}\setminus\sigma_{\mathrm{ess}}(\mathrm{T}(f)), the determinant det⁡(f(z)−λ)\det(f(z)-\lambda), where ff is the symbol of the tridiagonal Toeplitz operator T\mathrm{T} given in (2.5), has the form

det⁡(f(z)−λ)=(−1)k+1(∏i=1kci)z+(−1)k+1(∏i=1kbi)z−1+g(λ),(41)\det(f(z)-\lambda)=(-1)^{k+1}\left(\prod_{i=1}^{k}c_i\right)z+(-1)^{k+1}\left(\prod_{i=1}^{k}b_i\right)z^{-1}+g(\lambda), \tag*{(41)}

where gg is a polynomial in λ\lambda of degree kk, defined by (A.2). In particular, there are at most 2k2k λ∈C\lambda\in\mathbb{C} so that det⁡(f(z)−λ)=0\det(f(z)-\lambda)=0 admits a double root.

Proof. The case when k=1,2k=1,2 is easy to verify. For the case when k≥3k \ge3, by repeated Laplace expansion, one can show that

det⁡(f(z)−λ)=(−1)k+1b1bk∣b20⋯⋯0a3−λ⋱⋱⋯⋮c3⋱⋱⋱⋮⋮⋱⋱⋱00⋯ck−2ak−1−λbk−1∣z−1+(−1)k+1ckck−1∣c1a2−λb20⋯00c2a3−λb3⋱⋮⋮⋱⋱⋱⋱0⋮⋱⋱⋱⋱bk−3⋮⋱⋱ck−3ak−2−λ0⋯⋯⋯0ck−2∣z+(ak−λ)∣a1−λb10⋯0c1a2−λb2⋱⋮0⋱⋱⋱0⋮⋱⋱⋱bk−10⋯0ck−2ak−1−λ∣−bk−1ck−1∣a1−λb10⋯0c1a2−λb2⋱⋮0⋱⋱⋱0⋮⋱⋱⋱bk−30⋯0ck−3ak−2−λ∣−bkck∣a2−λb20⋯0c3a3−λb3⋱⋮0⋱⋱⋱0⋮⋱⋱⋱bk−20⋯0ck−1ak−1−λ∣.\begin{aligned} \det(f(z)-\lambda)=(-1)^{k+1}b_1b_k \begin{vmatrix} b_2 & 0 & \cdots& \cdots& 0 \\ a_3-\lambda& \ddots& \ddots& \cdots& \vdots\\ c_3 & \ddots& \ddots& \ddots& \vdots\\ \vdots& \ddots& \ddots& \ddots& 0 \\ 0 & \cdots& c_{k-2} & a_{k-1}-\lambda& b_{k-1} \end{vmatrix}z^{-1} +(-1)^{k+1}c_kc_{k-1} \begin{vmatrix} c_1 & a_2-\lambda& b_2 & 0 & \cdots& 0 \\ 0 & c_2 & a_3-\lambda& b_3 & \ddots& \vdots\\ \vdots& \ddots& \ddots& \ddots& \ddots& 0 \\ \vdots& \ddots& \ddots& \ddots& \ddots& b_{k-3} \\ \vdots& \ddots& \ddots& c_{k-3} & a_{k-2}-\lambda\\ 0 & \cdots& \cdots& \cdots& 0 & c_{k-2} \end{vmatrix}z \\ &+(a_k-\lambda) \begin{vmatrix} a_1-\lambda& b_1 & 0 & \cdots& 0 \\ c_1 & a_2-\lambda& b_2 & \ddots& \vdots\\ 0 & \ddots& \ddots& \ddots& 0 \\ \vdots& \ddots& \ddots& \ddots& b_{k-1} \\ 0 & \cdots& 0 & c_{k-2} & a_{k-1}-\lambda \end{vmatrix} -b_{k-1}c_{k-1} \begin{vmatrix} a_1-\lambda& b_1 & 0 & \cdots& 0 \\ c_1 & a_2-\lambda& b_2 & \ddots& \vdots\\ 0 & \ddots& \ddots& \ddots& 0 \\ \vdots& \ddots& \ddots& \ddots& b_{k-3} \\ 0 & \cdots& 0 & c_{k-3} & a_{k-2}-\lambda \end{vmatrix} -b_kc_k \begin{vmatrix} a_2-\lambda& b_2 & 0 & \cdots& 0 \\ c_3 & a_3-\lambda& b_3 & \ddots& \vdots\\ 0 & \ddots& \ddots& \ddots& 0 \\ \vdots& \ddots& \ddots& \ddots& b_{k-2} \\ 0 & \cdots& 0 & c_{k-1} & a_{k-1}-\lambda \end{vmatrix}. \end{aligned}

Hence (A.1) holds for

g(λ)=det⁡(A0−λ)−bkckp(λ),(42)g(\lambda)=\det(A_0-\lambda)-b_kc_kp(\lambda), \tag*{(42)}

where

p(λ)={0,k=1,1,k=2,∣a2−λb20⋯0c3a3−λb3⋱⋮0⋱⋱⋱0⋮⋱⋱⋱bk−20⋯0ck−1ak−1−λ∣,k≥3.p(\lambda)= \begin{cases} 0, & k=1,\\ 1, & k=2,\\ \begin{vmatrix} a_2-\lambda& b_2 & 0 & \cdots& 0 \\ c_3 & a_3-\lambda& b_3 & \ddots& \vdots\\ 0 & \ddots& \ddots& \ddots& 0 \\ \vdots& \ddots& \ddots& \ddots& b_{k-2} \\ 0 & \cdots& 0 & c_{k-1} & a_{k-1}-\lambda \end{vmatrix}, & k\ge3. \end{cases}

To demonstrate the last claim, we write

det⁡(f(z)−λ)=(Az+Bz−1)+g(λ)(43)\det(f(z)-\lambda)=(Az+Bz^{-1})+g(\lambda) \tag*{(43)}

with A=(−1)k+1(∏i=1kci)A=(-1)^{k+1}\left(\prod_{i=1}^{k}c_i\right), B=(−1)k+1(∏i=1kbi)B=(-1)^{k+1}\left(\prod_{i=1}^{k}b_i\right). By multiplication with zz, we obtain that

zdet⁡(f(z)−λ)=Az2+B+g(λ)z.z\det(f(z)-\lambda)=Az^2+B+g(\lambda)z.

Thus only when

g(λ)2−4AB=0,g(\lambda)^2-4AB=0,

we have a double root. Since g(λ)g(\lambda) is a polynomial of order kk, we have at most 2k2k solutions λ\lambda of the above equation so that the root is a double root.

References

References

References

Habib Ammari ETH Zürich, Department of Mathematics, Rämistrasse 101, 8092 Zürich, Switzerland. Email address: [email protected]

Silvio Barandun ETH Zürich, Department of Mathematics, Rämistrasse 101, 8092 Zürich, Switzerland. Email address: [email protected]

Yannick De Bruijn ETH Zürich, Department of Mathematics, Rämistrasse 101, 8092 Zürich, Switzerland. Email address: [email protected]

Ping Liu ETH Zürich, Department of Mathematics, Rämistrasse 101, 8092 Zürich, Switzerland. Email address: [email protected]

Clemens Thalhammer ETH Zürich, Department of Mathematics, Rämistrasse 101, 8092 Zürich, Switzerland. Email address: [email protected]

References

  1. [1]Ammar Habib, Barandun Silvio, Cao Jinghao, Davies Bryn and Hiltunen Erik Orvehed, “Mathematical foundations of the non-Hermitian skin effect”, in: arXiv pre-print arXiv:2306.15587 (2023), DOI: 10.48550/arXiv.2306.15587.DOI
  2. [2]Habib, A., Silvio, B., Jinghao, C., Bryn, D., Erik, H. O., & Ping, L. (2023). The non-Hermitian skin effect in three dimensions. https://doi.org/10.48550/arXiv.2311.10521
  3. [3]Ammari Habib, Barandun Silvio, Cao Jinghao and Feppon Florian, “Edge Modes in Subwavelength Resonators in One Dimension”, in: Multiscale Model. Simul. 21.3 (2023), pp. 964–992, doi: 10.48550/ARXIV.2301.06747.DOI
  4. [4]Ammari Habib, Barandun Silvio, Davies Bryn, Hiltunen Erik Orvehed and Liu Ping, Stability of the non-Hermitian skin effect, 2023, doi: 10.48550/arXiv.2308.06124.DOI
  5. [5]Ammari Habib, Barandun Silvio and Liu Ping, “Perturbed Block Toeplitz matrices and the non-Hermitian skin effect in dimer systems of subwavelength resonators”, in: arXiv preprint arXiv:2307.13551 (2023), doi: 10.48550/arXiv.2307.13551.DOI
  6. [6]Ammari Habib, Davies Bryn and Hiltunen Erik Orvehed, Functional Analytic Methods for Discrete Approximations of Subwavelength Resonator Systems, 2021, doi: 10.48550/ARXIV.2106.12301.DOI
  7. [7]Ammari Habib, Davies Bryn, Hiltunen Erik Orvehed and Yu Sanghyeon, “Topologically Protected Edge Modes in One-Dimensional Chains of Subwavelength Resonators”, in: Journal de Mathématiques Pures et Appliquées 144 (2020), pp. 17–49, doi: 10.1016/j.matpur.2020.08.007.DOI
  8. [8]Ashida Yuto, Gong Zongping and Ueda Masahito, “Non-Hermitian Physics”, in: Advances in Physics 69.3 (2nd July 2020), pp. 249–435, doi: 10.1080/00018732.2021.1876991, (visited on 28/09/2022).DOI
  9. [9]Böttcher Albrecht and Silbermann Bernd, Introduction to Large Truncated Toeplitz Matrices, Universitext, Springer-Verlag, New York, 1999, doi: 10.1007/978-1-4612-1426-7.DOI
  10. [10]Douglas Ronald G., Banach Algebra Techniques in Operator Theory, Springer New York, 1998, doi: 10.1007/978-1-4612-1656-8.DOI
  11. [11]Feppon, F., & Ammari, H. (2022). Subwavelength Resonant Acoustic Scattering in Fast Time-Modulated Media.
  12. [12]Feppon Florian, Cheng Zijian and Ammari Habib, “Subwavelength Resonances in One-Dimensional High-Contrast Acoustic Media”, in: SIAM Journal on Applied Mathematics 83.2 (2023), pp. 625–665, doi: 10.1137/22M1503841.DOI
  13. [13]Fonseca, C. M. da . (2007). The characteristic polynomial of some perturbed tridiagonal k-Toeplitz matrices. Appl. Math. Sci. (Ruse), 1(1-4), 59–67.
  14. [14]Ghatak Ananya, Brandenbourger Martin, Van Wezel Jasper and Coulais Corentin, “Observation of Non-Hermitian Topology and Its Bulk–Edge Correspondence in an Active Mechanical Metamaterial”, in: Proceedings of the National Academy of Sciences 117.47 (2020), pp. 29561–29568, doi: 10.1073/pnas.2010580117.DOI
  15. [15]Gohberg, I., Kaashoek, M. A., & Goldberg, S. (1993). Classes of Linear Operators. https://doi.org/10.1007/978-3-0348-8558-4
  16. [16]Hatano Naomichi and Nelson David R., “Localization Transitions in Non-Hermitian Quantum Mechanics”, in: Phys. Rev. Lett. 77 (3 1996), pp. 570–573, doi: 10.1103/PhysRevLett.77.570.DOI
  17. [17]Longhi Stefano, “Non-Hermitian skin effect beyond the tight-binding models”, in: Phys. Rev. B 104 (2021), p. 125109, doi: 10.1103/PhysRevB.104.125109.DOI
  18. [18]Lumsdaine, A., & Weyl, W. D. (1998). Spectra and pseudospectra of block Toeplitz matrices. Linear Algebra and its Applications, 272.1, 103–130. https://doi.org/10.1016/S0024-3795(97)00311-X
  19. [19]Okuma, N., Kawabata, K., Shozaki, K., & Sato, M. (2020). Topological Origin of Non-Hermitian Skin Effects. Physical Review Letters, 124(8), 086801. https://doi.org/10.1103/physrevlett.124.086801
  20. [20]Reichel, L., & Trefethen, L. N. (1992). Eigenvalues and pseudo-eigenvalues of Toeplitz matrices. Linear Algebra and its Applications, 162-164, 153–185. https://doi.org/10.1016/0024-3795(92)90374-J
  21. [21]Trefethen Lloyd N. and Embree Mark, Spectra and Pseudospectra, Princeton University Press, Princeton, NJ, 2005, doi: 10.1515/9780691213101.DOI

Paper details

vlnc:W_mrh79xt7
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
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.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents