We establish new results on the spectra and pseudo-spectra of tridiagonal k-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 k-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 k-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 k-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 k-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 k-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 k-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 k-Toeplitz operator and the associated tridiagonal k-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 k-Toeplitz operators. In Section 5, we make some concluding remarks and give some possible extensions of our present work.
Spectra of tridiagonal k-Toeplitz operators
In this section, we characterize the spectrum of tridiagonal k-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 k-Toeplitz operator by
where ai,bi,ci∈C for i∈N and Aij=0 if ∣i−j∣>1 for i,j∈N. Its finite, truncated, counterpart will be called a tridiagonal k-Toeplitz matrix.
To characterize the spectrum of tridiagonal k-Toeplitz operators, we introduce the notions of Fredholm operator and Fredholm index. Let X be a Banach space and B(X) the set of bounded linear operators on X. We say that an operator A∈B(X) is a Fredholm operator if ImA is a closed subspace in X and
dimKer(A)<∞,dimCokerA<∞,
where CokerA:=X/ImA. The index of the Fredholm operator A is defined as
IndA:=dimKerA−dimCokerA.(2)
Tridiagonal k-Toeplitz operators
Observe that the tridiagonal k-Toeplitz operator A presented in (2.1) can be reformulated as a tridiagonal block Toeplitz operator, where the blocks repeat in a 1-periodic way:
We will write A=T(f). In the following sections, we will characterize the spectrum of the operator T(f) based on its symbol f(z).
Hardy Spaces
To get a better understanding of the connection between the symbol f and the operator T(f) itself, it is necessary to introduce the so-called Hardy Spaces. Let T be the unit circle in C.
Definition 2.1. The Hardy spaces H2:=H2(T) and H−2:=H−2(T) are defined by
H2:={f∈L2(T):fn=0 for n<0},
H−2:={f∈L2(T):fn=0 for n≥0},
where fn denotes the Fourier coefficients of f.
The spaces H2 and H−2 are closed, orthogonal subspaces of L2:=L2(T). Hence, one may write
L2=H2⊕H−2.
Denote by P the projection of L2 onto H2. The functions
{2π1einθ}n=0∞
form an orthonormal basis of H2. One important property of functions in H2 is given by the following theorem [10 Theorem 6.13].
Theorem 2.2 (F. and M. Riesz). If f is a nonzero function in H2, then the set {eit∈T:f(eit)=0} has measure zero.
It is well-known [21] that for a Toeplitz operator T(f) with symbol f∈L∞, T(f) is the matrix representation of the operator
H2→H2,g↦P(fg)
in the above orthonormal basis. A similar notion holds for block Toeplitz operators. We denote by (H2)k the k-dimensional vector space whose entries are elements of H2. If P is the orthogonal projection of (L2)k onto (H2)k and f∈(L∞)k×k, then T(f) is the matrix representation of the operator
(H2)k→(H2)k,g↦P(fg),
where fg is the usual matrix-vector multiplication. Note that the (H2)k equivalent of Theorem 2.2 holds verbatim with H2 replaced by (H2)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 k-Toeplitz version). Let f∈Ck×k(T) be the symbol of a tridiagonal k-Toeplitz operator such that det(f(z)) does not vanish identically on T. Then one of the following statements holds:
(i) T(f) has a trivial kernel;
(ii) T(f) has a dense range;
(iii) The leading (k−1)×(k−1) principal minor of A0 has a nonzero kernel. In particular, there exists some z0∈C∪{∞} such that ker(z0−1A−1+A0)∩ker(A1)=0.
Proof. It is not hard to see that the adjoint of T(f) is T(fT), where the superscript T denotes the transpose. Assume that T(f) has a nontrivial kernel and a nondense range, i.e., conditions (i) and (ii) do not hold. This implies that T(fT) has a nontrivial kernel as well. Hence there exist nonzero functions g+,h+∈(H2)k such that
fg+=g−∈(H−2)kandfTh+=h−∈(H−2)k,(6)
where g±=(g±,1,g±,2,⋯,g±,k)T and h±=(h±,1,h±,2,⋯,h±,k)T. Since g+,h+∈(H2)k, there exist two sequences of vectors (vn)n,(wn)n∈Ck such that
g+=n∈N∑vnzn−1,h+=n∈N∑wnzn−1.(7)
for z=eiθ, where we have absorbed the factor 2π1 into vn and wn. This allows us to write
By Theorem 2.2, it follows that g+,h+=0 almost everywhere on T. We define
φ:=h+Tg−=h+Tfg+=(fTh+)Tg+=h−Tg+.
It holds that φ∈L1. Moreover, φn=(h−Tg+)n=0 for n≤0 and φn=(h+Tg−)n=0 for n≥0 since by (2.7), (2.8), and (2.9),
h+Tg−=n∈N∑z−nwnTA−1v1,
h−Tg+=n∈N∑zn(A1Tw1)Tvn.
Hence φ=0. This implies A1Tw1⊥vn and A−1v1⊥wn for n∈N. Thus either vn⊥ek, n=1,⋯, or w1∈kerA1T. However, if w1∈kerA1T, then by (2.9)
fTh+=h−=z−1A1Tw1=0,
which contradicts the fact that det(fT)=det(f) does not vanish identically on T. Therefore, vn⊥ek for n∈N≥1. By the above arguments, we conclude that if T(f) has non-dense range and nontrivial kernel, then vn⊥ek for n∈N≥1 and for any vn defined by (2.7).
Now we prove that if T(f) has a non-dense range and a nontrivial kernel, then the condition (iii) must hold. Since fg+=g−∈(H−2)k, by expansion (2.8) it holds that A−1vn+2+A0vn+1+A1vn=0 for n∈N. Since vn⊥ek for n∈N≥1 it follows that vn∈ker(A1). Hence it holds that A−1vn+1+A0vn=0 for n∈N≥2. Note that the expansion (2.8) also gives A−1v2+A0v1=0. Therefore, A−1vn+1+A0vn=0 for n∈N≥1. Furthermore, since the image of A−1 is confined to the span of ek and vn⊥ek for n∈N≥1, we have
A0(k−1)×(k−1)vn1:k−1=0,
where A0(k−1)×(k−1) is the leading principal (k−1)×(k−1) submatrix of A0. This also means that A0(k−1)×(k−1) cannot have full rank. On the other hand, one can compute that
where vn(j) is the j-th element of vn. Considering the case when n=1, by (2.8), we know that v1(1)=0, otherwise,
fg+=g−=z−1A−1v1=0,
which contradicts detf does not vanish identically on T. Similarly, we can also show that bk=0. Thus by (2.10) and bk=0 there exists some z0∈C∪{∞} such that v1 lies in the kernel of z0−1A−1+A0. Therefore it holds that ker(z0−1A−1+A0)∩ker(A1)=0, which concludes the proof. ■
Spectra of tridiagonal k-Toeplitz operators
In this section, we will provide a full characterization of the spectrum of tridiagonal k-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(⋅), σess(⋅) the spectrum and the essential spectrum of the operator, respectively. We define
σdet(f)={λ∈C:det(f(z)−λ)=0,∃z∈T}.(11)
We first have the following theorem characterizing the essential spectrum of T(f).
Theorem 2.4. The operator T(f) in (2.3) is Fredholm if and only if det(f) has no zeros on T. Furthermore, the essential spectrum of T(f) is given by
σess(T(f))=σdet(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)) if and only if T(f)−λ is not a Fredholm operator. By the first part, this happens only if det(f(z0)−λ)=0 for some z0∈T. That is, λ is an eigenvalue of f(z0). On the other hand, if λ is an eigenvalue for some z0∈T, it holds that det(f(z0)−λ)=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×k be such that T(f) is a Toeplitz operator. Then, T(f) is Fredholm on the space ℓ2 if and only if det(f) has no zeros on T, in which case
IndT(f)=−wind(det(f(T)),0),
with wind(det(f(T)),0) being the winding number about the origin of the determinant of f(z) with z∈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 k-Toeplitz operators. Before stating the theorem, we first define
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).
Theorem 2.6. Let f∈Ck×k(T) be the symbol of a tridiagonal k-Toeplitz operator T(f). Denote by B0 the leading (k−1)×(k−1) principal minor of A0. It holds that
where σdet(f),σwind(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)
then it must hold that T(f)−λI is Fredholm and Ind(T(f)−λI)=wind(det(f−λ),0)=0. Based on the definition (2), this implies
dimKer(T(f)−λI)=dim(ℓ2∖Im(T(f)−λI)).(16)
Moreover, since λ∈σ(T(f)), T(f)−λI has a non-trivial kernel and non-dense image. By Theorem 2.3, this implies that λ must be an eigenvalue of B0. ■
This characterization is optimal. It cannot be made more precise, as it cannot be guaranteed that an eigenvalue λ of B0 with wind(det(f−λ),0)=0 is also an eigenvalue of T(f). To illustrate this, consider
f(z)=(01+21z−11+21z1).
The determinant of this symbol is given by det(f)=−(1+21z)(1+21z−1). We have wind(det(f(T)),0)=0 and detf(z)=0,∀z∈T. In this case, B0=0 is just the top left entry of f(z). Furthermore, by (2), if 0∈σ(T(f)), then the kernel space of T(f) must be non-trivial. Since
T(f)=010011210021010011⋱,
one can check that, after scaling, the only possible eigenvector of T(f) corresponding to the eigenvalue 0 is u=(1,0,−2,0,4,…)T. However, u does not lie in ℓ2, which means that 0∈/σ(T(f)).
On the other hand, consider
f(z)=(01+2z−11+2z1).
The determinant is det(f)=−(1+2z)(1+2z−1), hence 0∈/σdet(f) and wind(det(f(T)),0)=0. Therefore, by Theorem 2.3, the operator T(f) is either invertible or has a nontrivial kernel. By explicitly writing down T(f)
T(f)=0100112002010011⋱,
we find that u=(1,0,−21,0,41,…)T satisfies T(f)u=0. Then in this case, u∈ℓ2 and 0∈σ(T(f)).
Remark 2.7. Note that (13) and (14) also provide ways to compute σwind(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 k-Laurent operators that are defined as
L(f)=⋱⋱A1⋱A0A1A−1A0⋱A−1⋱⋱(17)
with the generating symbol f being (5). Using the identification L(f−λ)=L(f)−λ, which we have used to characterize the spectrum of tridiagonal k-Toeplitz operators, one can obtain an explicit characterization of the spectrum of tridiagonal k-Laurent operators as a consequence of [15 Chapter 23, Corollary 2.5]. The following theorem holds.
Theorem 2.8. Let f∈(L(T)∞)k×k and denote the associated tridiagonal k-Laurent operator L(f). Then, it holds that
Now that we have characterized the spectra of tridiagonal k-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 and ∏j=1kbj=0. Let f(z)∈Ck×k be the symbol (5) and let λ∈C∖σess(T(f)). If ∑j=1kwind(λj(T),λ)<0 for λj being the eigenvalues of f(z), then there exists an eigenvector x∈ℓ2 of T(f) associated to λ and some ρ<1 such that
maxi∣xi∣∣xj∣≤Cρ⌈j/k⌉−1,j≥1,(18)
where C>0 is a constant depending only on λ,ap,bp,cp for 1≤p≤k.
Proof. Let λ∈C be such that ∑j=1kwind(λj(T),λ)=wind(det(f(T)−λ),0)<0. Since ∏i=1kbi=0,∏i=1kci=0, by Lemma A.1, det(f(z)−λ) is a meromorphic function in C, with poles at 0 and ∞. Then the argument principle implies that det(f(z)−λ)=0 for two values z1,z2 with 1<ρ1≤∣zj∣, j=1,2, counted with multiplicity. Moreover, since det(f(zi)−λ)=0, we find that λ is an eigenvalue to f(zi) and hence there exists an associated eigenvector vi. Assume for now that z1=z2. Consider the vectors defined by
ui=(viT,zi−1viT,zi−2viT,…)T,i∈{1,2}.(19)
The vector ui satisfies the eigenvalue equation T(f)ui=λui in all but the first row. Since u1,u2 are linearly independent, there exists a linear combination of them that is indeed an eigenvector of T(f).
Next, we consider the case where z1=z2 and first suppose that dimEig(f(z1),λ)>1, where Eig(f(z1),λ) denotes the eigenspace of f(z1) associated to the eigenvalue λ. Then there exist linearly independent eigenvectors v1,v2 of f(z1) associated to λ and we can use
ui=(viT,z1−1viT,z1−2viT,…)T,i∈{1,2},(20)
to construct an exponentially decaying eigenvector of T(f). Finally, it remains to treat the case when z1=z2 and dimEig(f(z1),λ)=1. Denote v1 the eigenvector of f(z1) associated to λ. We will construct a vector v2 such that
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). By explicitly writing down (T(f)−λ)u2=0 (except for the first row), we find that v2 must satisfy the condition
which is also a sufficient condition. From Lemma A.1, it is not hard to see that for all λ in a small neighborhood of λ, the equation det(f(z)−λ) has two distinct roots. Let (εj)j be a sequence converging to zero. We define λ(εj)=λ+εj. Then by the previous observation, we may assume that to each λ(εj) there exist two distinct roots z1(εj),z2(εj) of the equation det(f(z)−λ(εj))=0. Let w1(εj),w2(εj) be the corresponding unit eigenvectors. We may assume that
w1(εj)→v0,w2(εj)→v0.(23)
The limit of wi(εj) (or a subsequence thereof) exists by a standard compactness argument and it must be an eigenvector of f(z1) associated to λ. In particular, for each j it holds that
Consider now the sequence Δz(εj)w1(εj)−w2(εj). If a bounded subsequence exists, we may define v2 as the limit of this subsequence. In this case, it is not hard to see that v2 satisfies the condition (2.22). Otherwise, there exists a subsequence of Δz(εj)w1(εj)−w2(εj) whose norm tends to infinity. Define
Δw(εj)=∥w1(εj)−w2(εj)∥.
Then (upon passing to said subsequence) it holds that
In the limit, the right-hand side tends to 0. Thus v2=limεj→0Δw(εj)w1(εj)−w2(εj) satisfies
(A1+(A0−λ)z1−1+A−1z1−2)v2=0.
This indicates that v2 is an eigenvector. But from the construction of v2, it becomes apparent that v2 must be orthogonal to v1, which contradicts the assumption that the eigenspace of f(z1) associated with λ is one-dimensional. Therefore, the sequence Δz(εj)w1(εj)−w2(εj) is uniformly bounded and arguments for the first case already demonstrate the existence of such v2 satisfying (2.22).
The exponential bound in (2.18) is clear by the construction of the vectors u1 and u2. ■
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, λ∈σ(T(f)) is because T(f) has a non-dense image. In particular, one can construct the eigenvector x of T(f)−λ in the same spirit of Theorem 2.9, which implies T(f)−λ has a non-dense image.
Pseudo-spectra of tridiagonal k-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. Then, λ∈C is an ε-pseudoeigenvalue of A∈CN×N if one of the following conditions is satisfied:
λ is a proper eigenvalue of A+E for some E∈CN×N such that ∥E∥≤ε;
∥(A−λI)u∥<ε for some vector u with ∥u∥=1;
∥(A−λI)−1∥−1≤ε.
The set of all ε-pseudoeigenvalues of A, the ε-pseudospectrum, is denoted by σε(A). If some nonzero u satisfies ∥(λ−A)u∥<ε, then we say that u is an ε-pseudoeigenvector of 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 k-Toeplitz matrices {AN} of various dimensions obtained as N×N finite sections of the infinite operator 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 k-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)
ΩR:={λ∈C:wind(det(f(TR)−λ),0)<0},(27)
where Tr={z∈C:∣z∣=r} and let σε(AN) be the set of ε-pseudoeigenvalues of AN.
The connection between the eigenvectors of a tridiagonal k-Toeplitz operator and those of the associated tridiagonal k-Toeplitz matrix can be made with an argument similar to [21 Theorem 7.2]. The first N components of an eigenvector of the Toeplitz operator can be used as the ε-pseudoeigenvector of the tridiagonal k-Toeplitz matrix AN. Since we have proven the existence of an exponentially decaying eigenvector in Theorem 2.9, we can conclude with the following statement on tridiagonal k-Toeplitz matrices.
Theorem 3.2. Let {AN} be a family of tridiagonal k-Toeplitz matrices such that ∏i=1kbi=0, ∏i=1kci=0. Then, for any r<1 and ρ>r, we have
where Δε is a closed unit disk of radius ε and C1,C2 are constants depending on r, ρ, and aj, bj, cj for 1≤j≤k but independent of N. In particular, for the corresponding λ∈Ωr∪Ω1/r, there exist nonzero pseudoeigenvectors v(N) satisfying
Proof. Firstly, we note that the inclusion σ(T(f))+Δε⊆σε(AN) is triviality valid for any matrix or operator. Secondly, by symmetry, Ωr must satisfy an estimate of the type (28) if Ω1/r does. Thus all that we have to prove is Ω1/r⊆σε(AN).
Given any r<1, let λ∈Ω1/r be arbitrary. Since wind(det(f(T1/r)−λ),0)<0, similarly to the arguments in the proof of Theorem 2.9, we can show that det(f(z)−λ)=0 for two values z1,z2 with 1/ρ≤∣zj∣, j=1,2, counted with multiplicity. Also, since det(f(z)−λ)=0, we find that λ is an eigenvalue of f(zi) and hence there exists an associated eigenvector vi. Assume for now that z1=z2. Consider the vectors defined by
where without loss of generality we assume that N is divisible by k. This vector satisfies the eigenvalue equation ANui=λui in all but the first and last rows. Since u1,u2 are linearly independent, there exists a normalized linear combination v(N)=a1u1+a2u2 so that (AN−λ)v(N)=0 in all but the last row. It is not hard to see that
to construct the pseudoeigenvector. This yields a similar bound except with C3ρ⌈N/k⌉−1 in (3.4) replaced by an algebraically growing factor at worst ⌈N/k⌉C4ρ⌈N/k⌉−1. Then since the function
N≥1sup⌈N/k⌉r⌈N/k⌉−1σN(λI−AN)
is a continuous function of λ on the compact set Ω1/r, its maximum provides the constant C2 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 k 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 N disjoint one-dimensional resonators D:=(xiL,xiR), where (xiL,xiR)1≤i≤N⊂R are the 2N extremities such that xiL<xiR<xi+1L for any 1≤i≤N. The lengths of the i-th resonators will be denoted by ℓi=xiR−xiL and the spacings between the i-th and the (i+1)-th resonators will be denoted by si=xi+1L−xiR. Furthermore, we assume that k resonators are periodically repeated, that is, si+k=si. The resulting system is illustrated in Figure 4.1.
Image not extracted
Figure 4.1. A chain of N times periodically repeated k subwavelength resonators, with lengths (ℓi)1≤i≤N and spacings (si)1≤i≤N−1.
Let D be the domain consisting of the union of all resonators
D=i=1⋃N(xiL,xiR)⊂R.(35)
The underlying wave equation that governs the system is given by
⎩⎨⎧u′′(x)+γu′(x)+vb2ω2u=0,u′′(x)+v2ω2u=0,u∣R(xiL,R)−u∣L(xiL,R)=0,dxduR(xiL)=δdxduL(xiL),dxduL(xiR)=δdxduR(xiR),d∣x∣du−ivωu=0,x∈D,x∈R\D,for all 1≤i≤N,for all 1≤i≤N,for all 1≤i≤N,at x=x1L,xNR, respectively.(36)
where 0<δ≪1, the wave speeds inside the resonator and in the background are respectively vb and v and are of order one, and γu′ for γ∈R∗ denotes the imaginary gauge potential. Here, for a function w we denote by
We are interested in nontrivial frequencies ω such that (4.2) is satisfied and ω→0 as δ→0. Such ω 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∗, we define the gauge capacitance matrixCγ∈RN×N by
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. For the remainder of this paper, we assume that γ=ℓi=1. Without loss of generality, we consider a system of k resonators and obtain a gauge capacitance matrix of the tridiagonal k-Toeplitz form (perturbed):
It has been observed and conjectured [5] that the exponential decay of the eigenvectors of k-Toeplitz matrices for k≥2 is due to the winding of the eigenvalues of the symbol being nontrivial. Now, by the spectral theory for the tridiagonal k-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 and ∏j=1kbj=0. Let f(z)∈Ck×k be the symbol (5) and let λ∈C∖σess(Ta(f)). If ∑j=1kwind(λj(T),λ)<0, then there exists an eigenvector x of Ta(f) associated to λ and some ρ<1 such that
maxi∣xi∣∣xj∣≤C⌈kj⌉ρ⌈j/k⌉−1,(39)
where C>0 is a constant depending only on λ, aj, bj, cj, j=1,…,k. If ∑j=1kwind(λj(T),λ)>0, then the above results hold for the left eigenvectors.
Proof. Although with a perturbation a 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 in the region
G:={λ∈C∖σdet(f):j=1∑kwind(λj(T),λ)=0}.(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 G in (40) are localised and the only non-decreasing eigenvector (gray eigenvector) corresponds to the eigenvalue 0 in the boundary of the region G. On the other hand, in [5] it is observed and conjectured that the non-trivial winding of the eigenvalues λj(z) predicts the exponential decay of the eigenmodes. This is due to wind(λj(T),λ)≤0, j=1,2 in the example
G={λ∈C∖σdet(f):j=1∏2wind(λj(T),λ)=0}.
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 G, 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 that satisfy ∑j=13Re(λj(T),λ)=0. The black dots along the real line denote the spectrum of the gauge capacitance matrix Cγ and the solid blue and orange lines around the spectrum are the ε-pseudospectra for ε=10k and k=−5,−2. Figures B and D show the eigenvectors of Cγ.
Image not extracted
Figure 4.2. The region of λ so that ∑j=1kwind(λj(T),λ)=0 and the localization of the eigenvectors. Computation performed for s1=1,s2=2, and N=50.
Remark 4.3. We remark that the green regions for k=1,2,3,4 can be computed analytically through the formula (2.14). For the case when k≥5, 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 k-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)), the determinant det(f(z)−λ), where f is the symbol of the tridiagonal Toeplitz operator T given in (2.5), has the form
with A=(−1)k+1(∏i=1kci), B=(−1)k+1(∏i=1kbi). By multiplication with z, we obtain that
zdet(f(z)−λ)=Az2+B+g(λ)z.
Thus only when
g(λ)2−4AB=0,
we have a double root. Since g(λ) is a polynomial of order k, we have at most 2k solutions λ 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]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]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]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]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]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]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]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]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]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]Douglas Ronald G., Banach Algebra Techniques in Operator Theory, Springer New York, 1998, doi: 10.1007/978-1-4612-1656-8.DOI
[11]Feppon, F., & Ammari, H. (2022). Subwavelength Resonant Acoustic Scattering in Fast Time-Modulated Media.
[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]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]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
[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]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]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]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]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]Trefethen Lloyd N. and Embree Mark, Spectra and Pseudospectra, Princeton University Press, Princeton, NJ, 2005, doi: 10.1515/9780691213101.DOI
Yuxuan Lin, Zafer Mutlu, Gabriela Borin Barin, Jenny Hong, Juan Pablo Llinas, Akimitsu Narita, Hanuman Singh, Klaus Müllen, Pascal Ruffieux, Roman Fasel et al. ·