betaFeedback

WorksPhysics and Astronomy

Complex mobility edges from competing localization channels in non-Hermitian quasicrystals

Abstract

Non-Hermitian localization transitions often accompany multiple mobility edges, but their physical origin can be ambiguous. To address the issue, we introduce an extension of the Ganeshan-Pixley-Das Sarma model with both asymmetric hopping and complex onsite potential, thereby incorporating a broad range of non-Hermitian quasiperiodic chains. Its solution reveals two energy-dependent localization contributions manifested as distinct Lyapunov growth scales, whose competition determines the structure of mobility edges. We derive exact Lyapunov exponents via Avila's global theory, obtain thresholds of the two channels on complex-energy plane, and formulate a selection rule that distinguishes physical mobility edges from suppressed analytical candidates. The channel that uniquely controls localization can switch with energy, producing complex mobility edges of non-Hermitian origins. Nonreciprocal hopping further shifts the thresholds and alters the bulk energy patterns. Numerical spectra and finite-size localization diagnostics confirm the theory. Our work thus provides a unified framework for characterizing phase transitions and mobility edges under competing localization mechanisms in non-Hermitian quasicrystals.

Introduction

Quasicrystals exhibit long-range order without periodicity and have long served as ideal platforms for studying topological phenomena [1, 2, 3] and localization transitions [4, 5, 6]. A paradigmatic example is the Aubry-André-Harper (AAH) model [7, 8, 9], in which a localization transition occurs when the quasiperiodic potential exceeds a critical value corresponding to the self-dual point. Recent studies of various extensions of the AAH model have revealed a richer variety of transitions: introducing long-range hopping [10, 11, 12], generalized onsite potentials [13, 14, 15, 16, 17], or more complex lattice structures, such as mosaic and network configurations [18, 19, 20, 21], can induce energy-dependent localization [22, 23], allowing localized and extended states to coexist in different energy intervals, with the boundaries between them known as mobility edges. In addition, quasiperiodic localization and mobility edges have also been experimentally observed in quantum simulators such as ultracold atoms [24, 25, 26, 27] and photonic platforms [28, 29, 30, 31, 32].

Non-Hermiticity [33, 34, 35], introduced via asymmetric hopping or complex potentials, further enriches localization transitions and mobility-edge phenomena in quasicrystals [36, 37, 38]. For example, asymmetric hopping can drive delocalization under periodic boundary conditions [39, 40, 41, 42, 43], whereas under open boundary conditions it can induce non-Hermitian skin effects [44, 45, 46, 47], causing eigenstates to accumulate near the system boundaries [48]. In addition, complex quasiperiodic potentials can induce parity-time (PTPT) symmetry breaking [49, 50, 51, 52] and produce elliptical complex-energy spectra [53] with mobility-edge loops [54, 55, 56, 57, 58]. Analytical results for mobility edges have been obtained for various non-Hermitian quasicrystals [59, 60, 61, 62, 63, 64, 65, 66]. To date, however, the issue regarding how multiple mobility-edge loops coexisting in the complex-energy plane jointly determine the localization properties of the system has yet to be properly addressed.

Here, we reveal how competing localization channels determine complex mobility edges in a generic class of non-Hermitian quasiperiodic chain. Using Avila’s global theory [67], we derive the exact spectral Lyapunov exponent for a generalized non-Hermitian Ganeshan–Pixley–Das Sarma (GPD) model and identify two localization contributions associated with opposite hopping directions in momentum space. The quasiperiodic modulation makes these contributions depend differently on eigenenergy, allowing the channel-controlling localization to switch across the complex-energy spectrum. This competition yields a selection rule for physical mobility edges. The stronger of the two localization contributions determines the actual transition boundary. Even when the spectrum crosses a candidate boundary, no localization transition occurs if the other, stronger channel dominates localization in that region. Physical mobility edges therefore consist of selected segments of the two candidate loops, with the dominant channel switching where these segments meet. Nonreciprocal hopping further shifts the localization thresholds and reshapes these boundaries. We verify this mechanism using numerical spectra and eigenstate localization diagnostics. These results show that the analytical loops jointly determine localization through channel competition, rather than each serving as an independent mobility edge.

Results

The non-Hermitian GPD model

We start with a non-Hermitian extension of the GPD model, whose non-Hermiticity originates from both hopping asymmetry and complex onsite potentials. This model incorporates a broad range of non-Hermitian quasiperiodic chains as specific realizations while retaining analytically accessible localization thresholds. In real space, the Schrödinger equation of our model reads

tegψn+1+te−gψn−1+V(θn)ψn=Eψn,(1)t e^{g}\psi_{n+1}+t e^{-g}\psi_{n-1}+V(\theta_n)\psi_n=E\psi_n, \tag*{(1)}

where

V(θ)=2λcos⁡θ+2iγsin⁡θ1−acos⁡θ−ibsin⁡θ,θn=2παn+φ.(2)V(\theta)=\frac{2\lambda\cos\theta+2i\gamma\sin\theta}{1-a\cos\theta-ib\sin\theta},\qquad\theta_n=2\pi\alpha n+\varphi. \tag*{(2)}

The symmetric part of the hopping amplitude t>0t>0 fixes the energy scale. g∈Rg\in\mathbb{R} controls the hopping asymmetry. The lattice index n∈Zn\in\mathbb{Z}, and the potential parameters λ\lambda, γ\gamma, aa and bb are all real. Quasiperiodicity is realized by an irrational α\alpha, set as (5−1)/2(\sqrt{5}-1)/2 and approximated by Fibonacci ratios in numerical calculations. The parameter φ∈[0,2π)\varphi\in[0,2\pi) represents a phase shift. We work in the pole-free regime a2+b2<1a^2+b^2<1, which is sufficient to keep the denominator of V(θ)V(\theta) nonzero for any θ∈R\theta\in\mathbb{R}. Figure 1(a) illustrates the real-space chain. All numerical spectra and right-eigenstate diagnostics in this work use periodic boundary conditions (PBC). In the Hermitian GPD model, the denominator 1−acos⁡θ1-a\cos\theta of V(θ)V(\theta) makes the effective momentum-space hopping energy dependent. States at different energies therefore reach the localization threshold at different potential strengths, allowing localized and extended states to coexist in the spectrum and be separated by a mobility edge [14]. This mobility edge is found to be greatly enriched by non-Hermitian effects in our generalized GPD model, as will be uncovered below.

Potential and hopping schemes in real and momentum spaces

Figure 1. Potential and hopping schemes of the generalized non-Hermitian GPD model in real and momentum spaces. (a) Real-space chain with onsite potential V(θn)V(\theta_n) and hopping amplitudes te±gte^{\pm g} in opposite directions. (b) Effective chain in momentum space at g=0g=0, obtained after multiplying by the denominator of V(θ)V(\theta). εm\varepsilon_m is the onsite term. J±(m,E)J_{\pm}(m,E) couple um∓1u_{m\mp1} to site mm [see Eqs. (3) and (4)]. The hopping profiles J±(m,E)J_{\pm}(m,E) depend on energy and could have distinct strengths.

Unequal dual hopping and competing Fourier channels

We now examine how the non-Hermitian modulation changes this energy-dependent localization mechanism. We first set g=0g=0, and restore nonreciprocal hopping in the final subsection. To simplify the notation, we introduce z=eiθz=e^{i\theta}, let A=(a+b)/2A=(a+b)/2, B=(a−b)/2B=(a-b)/2, and c±=λ±γc_{\pm}=\lambda\pm\gamma. We can then write the onsite potential as V(z)=v(z)/d(z)V(z)=v(z)/d(z) in (2), where v(z)=c+z+c−z−1v(z)=c_{+}z+c_{-}z^{-1} and d(z)=1−Az−Bz−1d(z)=1-Az-Bz^{-1}. In particular, Ed(z)−v(z)=E−p+(E)z−p−(E)z−1Ed(z)-v(z)=E-p_{+}(E)z-p_{-}(E)z^{-1}, where p+(E)=AE+c+p_{+}(E)=AE+c_{+} and p−(E)=BE+c−p_{-}(E)=BE+c_{-}. Thus, the denominator of V(z)V(z) introduces EE into both coefficients p±p_{\pm} before any further transformations are performed.

Let umu_m be the wave amplitude at the mmth lattice site in momentum space. Applying the dual transformation ψn=∑mumeimθn\psi_n=\sum_m u_m e^{im\theta_n} and multiplying both sides of (1) by the denominator of V(z)V(z), we arrive at the Schrödinger equation in momentum space

J+(m,E)um−1+J−(m,E)um+1+εmum=Eum.(3)J_{+}(m,E)u_{m-1}+J_{-}(m,E)u_{m+1}+\varepsilon_m u_m=Eu_m. \tag*{(3)}

Here, εm=2tcos⁡(2παm)\varepsilon_m=2t\cos(2\pi\alpha m) is the potential at site mm in momentum space. The effective hopping coefficients J±(m,E)J_{\pm}(m,E) are

J+(m,E)=p+(E)−Aεm−1=λ+γ+A(E−εm−1),J−(m,E)=p−(E)−Bεm+1=λ−γ+B(E−εm+1).(4)\begin{aligned} J_{+}(m,E)&=p_{+}(E)-A\varepsilon_{m-1}=\lambda+\gamma+A(E-\varepsilon_{m-1}),\\ J_{-}(m,E)&=p_{-}(E)-B\varepsilon_{m+1}=\lambda-\gamma+B(E-\varepsilon_{m+1}). \tag*{(4)} \end{aligned}

Figure 1(b) illustrates the chain in momentum space. Because both hopping terms J±(m,E)J_{\pm}(m,E) depend on EE explicitly, Eq. (3) constitutes an energy-dependent dual problem of Eq. (1).

In localized and extended phases related by the AAH/GPD duality [8, 14, 15, 68], extension/localization in momentum space accompanies localization/extension in real space. Changing the momentum-space hopping affects how strongly different momentum components are mixed and can therefore drive a real-space transition between extended and localized phases. In our generalized GPD model, the complex modulation makes the hopping J±(m,E)J_{\pm}(m,E) different in the two momentum-space directions. As the parameters or energy vary, the localization contribution associated with one direction can reach its transition threshold while the other remains below it. The two directional hopping processes can thus give rise to two distinct candidate mobility-edge contours in the complex-energy plane. Moreover, both hopping terms in Eq. (4) contain the eigenenergy EE, with different parameter combinations. Their relative strengths can therefore change across the complex-energy plane, allowing the dominant contribution to real-space localization to switch between the two directions. This is the physical origin of their competition. The resulting physical mobility edge can consequently be formed from selected portions of the two candidate contours. We next use the exact Lyapunov exponent to determine the two thresholds and their selection rule, and compare these predictions with numerical results.

Exact mobility-edge contours and numerical verification

For reciprocal hopping (g=0g=0), the transfer matrix of (1) maps the wave components (ψn,ψn−1)T(\psi_n,\psi_{n-1})^{\mathsf T} to (ψn+1,ψn)T(\psi_{n+1},\psi_n)^{\mathsf T} and is given by

TE(z)=([E−v(z)/d(z)]/t−110).(5)T_E(z)=\begin{pmatrix}[E-v(z)/d(z)]/t&-1\\1&0\end{pmatrix}. \tag*{(5)}

Multiplying by d(z)d(z) on both sides leads to the regularized transfer matrix T~E(z)=d(z)TE(z)\widetilde{T}_E(z)=d(z)T_E(z). Using Ed−v=E−p+z−p−z−1Ed-v=E-p_{+}z-p_{-}z^{-1}, we collect the terms proportional to zz, 11, and z−1z^{-1}, which define the coefficient matrices M+M_{+}, M0M_0, and M−M_{-} as

T~E(z)=z(−p+(E)/tA−A0)⏟M+(E)+(E/t−110)⏟M0(E)+z−1(−p−(E)/tB−B0)⏟M−(E).(6)\widetilde{T}_E(z)=z\underbrace{\begin{pmatrix}-p_{+}(E)/t&A\\-A&0\end{pmatrix}}_{M_{+}(E)}+\underbrace{\begin{pmatrix}E/t&-1\\1&0\end{pmatrix}}_{M_0(E)}+z^{-1}\underbrace{\begin{pmatrix}-p_{-}(E)/t&B\\-B&0\end{pmatrix}}_{M_{-}(E)}. \tag*{(6)}

In the dual transformation, the coefficient of eimθne^{im\theta_n} in zψnz\psi_n is um−1u_{m-1}, whereas in z−1ψnz^{-1}\psi_n it is um+1u_{m+1} with z=eiθnz=e^{i\theta_n}. These are the two shifts appearing in the effective hopping terms of (3). We therefore refer to the zM+zM_{+} and z−1M−z^{-1}M_{-} contributions as the two Fourier channels. Here, the indices τ=±\tau=\pm label the positive and negative Fourier harmonics z±1z^{\pm1} of the regularized transfer matrix. Their dimensionless scales ρτ(E)\rho_{\tau}(E) are the spectral radii of Mτ(E)M_{\tau}(E), namely the largest eigenvalue modulus.

To quantify the two directional hopping processes, let Jτ,geom(E)J_{\tau,\mathrm{geom}}(E) be the geometric mean of hopping magnitude ∣Jτ∣|J_{\tau}| over one period of its quasiperiodic modulation. Here, geom stands for geometric mean. Jτ,geomJ_{\tau,\mathrm{geom}} is the nonnegative strength of the full hopping profile, whereas Jτ(m,E)J_{\tau}(m,E) is the coefficient at a particular dual site. Since localization is governed by exponential growth or decay under repeated quasiperiodic transfer, we use a logarithmic phase average with tt the reference energy.

For τ=±\tau=\pm, we set s+=As_{+}=A and s−=Bs_{-}=B. The exact relation between the hopping strengths and dimensionless channel scales is

Jτ,geom(E)=texp⁡[∫02πdθ2πln⁡∣pτ(E)−2tsτcos⁡θt∣]=tρτ(E).(7)J_{\tau,\mathrm{geom}}(E)=t\exp\left[\int_{0}^{2\pi}\frac{d\theta}{2\pi}\ln\left|\frac{p_{\tau}(E)-2ts_{\tau}\cos\theta}{t}\right|\right]=t\rho_{\tau}(E). \tag*{(7)}

This identity follows from factoring the hopping profile using the channel eigenvalues and applying Jensen’s formula. Explicit channel scales and phase-average calculations are given in Methods [Eqs. (20)–(25)].

Multiplying the transfer matrix by d(z)d(z) also changes its accumulated growth. At each site, the same scalar d(eiθn)d(e^{i\theta_n}) multiplies the entire matrix, so a product of NN transfer matrices acquires the factor ∏n=0N−1d(eiθn)\prod_{n=0}^{N-1}d(e^{i\theta_n}). Let L~(E,0)\widetilde{L}(E,0) and L0(E)L_0(E) be the Lyapunov exponents of the regularized and original transfer matrices, respectively, evaluated with no imaginary phase shifts in θ\theta. Their difference is the mean logarithmic contribution of these scalar factors. Defining its exponential as DD, we find

L~(E,0)−L0(E)=lim⁡N→∞1N∑n=0N−1ln⁡∣d(eiθn)∣=∫02πdθ2πln⁡∣d(eiθ)∣≡ln⁡D.(8)\begin{aligned} \widetilde{L}(E,0)-L_0(E)=\lim_{N\to\infty}\frac{1}{N}\sum_{n=0}^{N-1}\ln|d(e^{i\theta_n})|=\int_{0}^{2\pi}\frac{d\theta}{2\pi}\ln|d(e^{i\theta})| \\ &\equiv\ln D. \tag*{(8)} \end{aligned}

In the pole-free regime, this gives

D=1+1−4AB2=1+1−a2+b22.(9)D = \frac{1+\sqrt{1-4AB}}{2} = \frac{1+\sqrt{1-a^2+b^2}}{2}. \tag*{(9)}

Recovering the original Lyapunov exponent requires subtracting ln⁡D\ln D. Accordingly, the channel terms enter as ln⁡(ρ±/D)\ln(\rho_\pm/D), so their critical scale is ρ±=D\rho_\pm= D, or equivalently J±,geom=tDJ_{\pm,\mathrm{geom}}=tD. Both channels share this normalization because they originate from the same denominator. The expression for DD is derived in Methods, Eq. (18). Its extension to a complex phase shift is given in Supplementary Note 2.

For reciprocal hopping, Avila’s global theory gives the result of Lyapunov exponent for the original transfer matrix as

L0(E)=max⁡{0,ln⁡ρ+(E)D,ln⁡ρ−(E)D}.(10)L_0(E)=\max\left\{0,\ln\frac{\rho_+(E)}{D},\ln\frac{\rho_-(E)}{D}\right\}. \tag*{(10)}

The derivation is given in Methods. If both ρ±≤D\rho_\pm\le D, the L0(E)L_0(E) vanishes. If either one of ρ±\rho_\pm exceeds DD, the L0(E)L_0(E) is positive. When both ρ±>D\rho_\pm> D, the larger one determines the Lyapunov exponent and controls the localization transition.

The individual thresholds ρ+=D\rho_+ = D and ρ−=D\rho_- = D, or J+,geom=tDJ_{+,\mathrm{geom}}=tD and J−,geom=tDJ_{-,\mathrm{geom}}=tD, give the candidate curves for a±b≠0a\pm b\ne0 (see Supplementary Note 2 for derivations), i.e.,

E±(0)(φ)=−2(λ±γ)a±b−2ta±b[Deiφ+(a±b)24De−iφ].(11)E_\pm^{(0)}(\varphi)=-\frac{2(\lambda\pm\gamma)}{a\pm b}-\frac{2t}{a\pm b}\left[De^{i\varphi}+\frac{(a\pm b)^2}{4D}e^{-i\varphi}\right]. \tag*{(11)}

The angle φ∈[0,2π)\varphi\in[0,2\pi) traces each complex-energy curve and differs from the lattice phase ϕ\phi. These curves are ellipses for real parameters, with centers and shapes determined by the different parameter combinations in p±(E)p_\pm(E). If a+b=0a+b=0 or a−b=0a-b=0, the corresponding radius is independent of energy and must be treated directly rather than with this ellipse formula.

Figure 2 compares these analytical boundaries with numerical spectra and illustrates how the two contributions select the physical mobility edge. In panel (a), the red and blue curves represent the thresholds ρ+=D\rho_+=D and ρ−=D\rho_-=D. The eigenstates under PBC are colored by the indicator D2D_2: values closer to zero indicate localization, while values closer to one indicate extended states. Panel (b) shows which contribution dominates above the threshold and explains why only certain portions of the two curves form the transition boundary.

Figure 2. Two localization channels and the selection of candidate boundaries. Parameters are t=1t=1, a=0.85a=0.85, b=0.1b=0.1, λ=0\lambda=0, γ=0.6\gamma=0.6, g=0g=0, N=377N=377, α=233/377\alpha=233/377, and φ=0\varphi=0. (a) Eigenenergies under PBC colored by the indicator D2D_2. Red and blue contours arise from the two analytical channel thresholds. Solid lines border the white region in (b) and satisfy Eq. (12). Dashed lines lie inside a colored region, where the stronger competing contribution sustains localization. (b) Analytical channel map: the white region has ρ+≤D\rho_+ \le D and ρ−≤D\rho_- \le D; red and blue regions indicate which channel dominates above threshold. Black dots mark the switching points ρ+=ρ−=D\rho_+ = \rho_- = D, where the selected red-white and blue-white boundary segments meet and exchange their controlling channel.

To understand which parts of these curves can describe a localization transition, we start from the white region in Figure 2(b). There, both contributions are at or below the threshold, and all eigenstates have L0=0L_0=0. In the red region, the ++ branch exceeds the threshold and becomes dominant, while the −- branch does so in the blue region. Therefore, the boundary between the white and a colored region is where the spectral Lyapunov exponent changes between zero and a positive value.

The solid red segment in Figure 2(b) divides the white and red regions. On this segment, the ++ branch just reaches the threshold, ρ+=D\rho_+=D, while the −- one is at or below it. Entering the red region makes the ++ branch sustain localization. Similarly, the solid blue segment separates the white and blue regions: the −- branch reaches the threshold while the ++ one does not exceed it. These two cases give the selection rule

ρτ(E)=D,ρ−τ(E)≤D,τ=±.(12)\rho_\tau(E)=D,\qquad\rho_{-\tau}(E)\le D,\qquad\tau=\pm. \tag*{(12)}

The spectrum in panel (a) shows where these selected boundaries are reached by eigenstates, and the changes in D2D_2 track the associated change in spatial localization.

The dashed segments in Figure 2(a) have a different meaning. Consider the red dashed curve inside the blue region. The ++ branch reaches its own threshold there, but the stronger −- branch has already exceeded the threshold. Eigenstates therefore have L0(E)=ln⁡[ρ−(E)/D]>0L_0(E)=\ln[\rho_-(E)/D]>0 on the red dashed curve and remain localized on both sides of it. Crossing this curve changes whether the weaker contribution is above or below its own threshold, but does not change the localized character of the state. This part of the red ellipse is suppressed as a mobility edge. A blue dashed segment inside the red region has the same interpretation with the labels exchanged. Figure 2(b) therefore explains why the spectral crossing of a dashed curve in Figure 2(a) needs not produce a localization transition.

Finally, we follow the edge of the white region through a black dot in Figure 2(b). The boundary changes from a solid red to a solid blue segment, or vice versa. At this junction, denoted by EsE_s, the contributions of both branches are exactly at the threshold:

ρ+(Es)=ρ−(Es)=D.(13)\rho_+(E_s)=\rho_-(E_s)=D. \tag*{(13)}

On one side of the dot, the ++ branch determines the transition boundary, while the −- one does on the other side. This is the exchange of control between the two candidate ellipses. The analytical map provides this geometric picture, while the eigenstates in Figure 2(a) reveal its spectral pattern.

Figure 3 verifies the analytical results. It compares Eq. (10) with Lyapunov exponents obtained by multiplying the original transfer matrices at sampled eigene­nergies of finite Fibonacci approximants, yielding perfect agreements.

Dual-hopping-based classification of AAH and GPD models

The exact results above allow us to classify various known AAH and GPD models within the same framework according to their directional hopping strengths J±,geom(E)J_{\pm,\mathrm{geom}}(E), as illustrated in Figure 4. We organize these models into three general groups, i.e., (i) both J±,geomJ_{\pm,\mathrm{geom}} are EE-independent, (ii) J+,geom(E)=J−,geom(E)J_{+,\mathrm{geom}}(E)=J_{-,\mathrm{geom}}(E) for all EE, and (iii) the ratio J+,geom(E)/J−,geom(E)J_{+,\mathrm{geom}}(E)/J_{-,\mathrm{geom}}(E) is fixed. We apply Eqs. (10), (12) and use ρτ=Jτ,geom/t\rho_\tau=J_{\tau,\mathrm{geom}}/t to explain the localization behavior in each case.

Figure 4. Classification of directional momentum-space hopping. The curves schematically represent the geometric-mean hopping strengths J+,geom(E)J_{+,\mathrm{geom}}(E) and J−,geom(E)J_{-,\mathrm{geom}}(E) of the reciprocal reference problem. The horizontal threshold is tDt_D. For complex energies, the horizontal axis follows a specified path in the complex plane. (a) Energy-independent strengths: equal in the AAH model and generally unequal in the complex AAH-type chains. (b) Identical energy-dependent strengths: the Hermitian GPD model and the related construction of Ref. [69]. The reciprocal reference of the nonreciprocal GPD model also belongs to this class. (c) Energy-dependent strengths with a fixed ratio J+,geom/J−,geom=e−2hJ_{+,\mathrm{geom}}/J_{-,\mathrm{geom}}=e^{-2h} in the common-shift complex GPD model. (d) Independently tunable channel strengths in our generalized model, allowing the order of J±,geomJ_{\pm,\mathrm{geom}} to reverse. The physical boundary is selected by Eq. (12). An exchange of control along that boundary occurs when both strengths reach tDt_D, as specified by Eq. (13).

Group (i). In the AAH limit a=b=γ=g=0a=b=\gamma=g=0, both dual hoppings are equal to λ\lambda, so J+,geom=J−,geom=∣λ∣J_{+,\mathrm{geom}}=J_{-,\mathrm{geom}}=|\lambda|. With D=1D=1, Eq. (10) predicts the common transition at ∣λ∣=t|\lambda|=t. The complex AAH-type chain with a=b=g=0a=b=g=0 and γ≠0\gamma\ne0 instead has unequal hoppings J±=λ±γJ_{\pm}=\lambda\pm\gamma and J±,geom=∣λ±γ∣J_{\pm,\mathrm{geom}}=|\lambda\pm\gamma|, but neither depends on EE. Our theory predicts the transition at max⁡{∣λ+γ∣,∣λ−γ∣}=t\max\{|\lambda+\gamma|,|\lambda-\gamma|\}=t. The horizontal curves in Figure 4(a) illustrate this energy independence: all eigenstates share the same localization threshold, showing no mobility edges.

Group (ii). In the GPD limit b=γ=g=0b=\gamma=g=0, both hopping profiles have the form λ+(a/2)(E−ε)\lambda+(a/2)(E-\varepsilon) with identical geometric means J+,geom(E)=J−,geom(E)J_{+,\mathrm{geom}}(E)=J_{-,\mathrm{geom}}(E), as represented by the overlapping curves in Figure 4(b). The common threshold J±,geom=tDJ_{\pm,\mathrm{geom}}=tD gives the unique localization condition ∣aE+2λ∣=2t|aE+2\lambda|=2t. Its two algebraic branches come from the same localization scale. A branch describes a physical mobility edge only where it separates spectral states with different localization character. Adding real-space nonreciprocity while keeping b=γ=0b=\gamma=0 preserves the identical strengths of the reciprocal reference problem and shifts their common threshold to tDe∣g∣tD e^{|g|} [61]. The generalized Hermitian GPD model of Ref. [69] adds a constant numerator term. This changes the spectrum without separating the two directional energy responses, placing it in the same class as Figure 4(b). In the bounded-potential regime, the two branches reported there are algebraic solutions of one localization condition, and the spectrum selects the physical branch. They therefore arise from a common localization scale rather than the two competing directional contributions considered here.

Group (iii). A common non-Hermitian GPD construction applies the same imaginary phase shift θ↦θ+ih\theta\mapsto\theta+ih to the numerator and denominator of V(θ)V(\theta) with h∈Rh\in\mathbb{R} [62]. The Λ\Lambda and β\beta in Figure 4(c) specify the numerator amplitude and denominator modulation in Vh(θ)=2Λcos⁡(θ+ih)/[1−βcos⁡(θ+ih)]V_h(\theta)=2\Lambda\cos(\theta+ih)/[1-\beta\cos(\theta+ih)], where Λ\Lambda is related to the λ\lambda in Eq. (2) by λ=Λcosh⁡h\lambda=\Lambda\cosh h. The two dual hopping profiles then share the same energy-independent form and multiplied by fixed factors e−he^{-h} and ehe^h at their respective shifted sites, yielding J+,geom(E)=e−2hJ−,geom(E)J_{+,\mathrm{geom}}(E)=e^{-2h}J_{-,\mathrm{geom}}(E) as illustrated in Figure 4(c). Changing EE maintains their ratio fixed. For h>0h>0 (h<0h<0) the (+)(+) direction remains stronger, while for h=0h=0 they coincide. Applying Eq. (12), only the stronger direction can supply the localization boundary. Two possible threshold contours can thus exist without exchanging control.

As a key advantage, our model allows aa, bb, λ\lambda, and γ\gamma to vary independently within the pole-free regime. The two hoppings in Eq. (4) thus contain independently adjustable coefficients of EE and offsets, rather than sharing an identical profile or a fixed proportionality. Their geometric-mean strengths can exchange dominance as energy varies, as illustrated in Figure 4(d). The larger one determines the dominant localization contribution in Eq. (10). Each direction can offer the selected boundary where J+,geom=tDJ_{+,\mathrm{geom}}=tD while J−,geom≤tDJ_{-,\mathrm{geom}}\le tD. When they both reach tDtD, the selected segments can join and exchange control, as illustrated in Figure 2. Solving the general model reveals this competition, which is absent in the special groups (i)–(iii) above. Table 1 summarizes our classification, with additional parameter reductions given in Supplementary Note 1.

ModelParameter choice or potentialRelation between hopping strengthsMobility-edge behavior
Group (i) [Figure 4(a)]
AAHa=b=γ=g=0a=b=\gamma=g=0J+,geom=J−,geom=∣λ∣J_{+,\mathrm{geom}}=J_{-,\mathrm{geom}}=|\lambda|No mobility edge; all energies share the same transition threshold ∣λ∣=t|\lambda|=t
Complex AAH-type chaina=b=g=0, γ≠0a=b=g=0,\ \gamma\ne0J±,geom=∣λ±γ∣J_{\pm,\mathrm{geom}}=|\lambda\pm\gamma|No mobility edge; unequal but energy-independent contributions give the same localization threshold at all energies
Group (ii) [Figure 4(b)]
Hermitian GPD [14]b=γ=g=0b=\gamma=g=0J+,geom=J−,geomJ_{+,\mathrm{geom}}=J_{-,\mathrm{geom}}One physical mobility edge; identical directional contributions give one localization condition, and the spectrum selects its physical root
Nonreciprocal GPDb=γ=0, g≠0b=\gamma=0,\ g\ne0Equal J±,geomJ_{\pm,\mathrm{geom}} in the reciprocal referenceOne mobility-edge boundary; hopping asymmetry shifts the common threshold of the identical reciprocal contributions
Generalized GPD [69]Additional constant numerator termEqual J±,geomJ_{\pm,\mathrm{geom}}; one common conditionOne physical mobility edge in the bounded-potential regime; the two reported algebraic branches solve the same localization condition, and the spectrum selects the physical branch
Group (iii) [Figure 4(c)]
Common-shift complex GPD [62]Shared shift θ→θ+ih\theta\to\theta+ihJ+,geom=e−2hJ−,geomJ_{+,\mathrm{geom}}=e^{-2h}J_{-,\mathrm{geom}}One mobility-edge boundary; a fixed ratio keeps the same channel stronger at every energy
General case [Figure 4(d)]
Our modelIndependent a,b,λ,γ,ga,b,\lambda,\gamma,gJ±,geom(E)J_{\pm,\mathrm{geom}}(E) can exchange orderTwo directional candidate curves; selected portions reached by the spectrum form the physical mobility edge, with the controlling channel able to switch between them

Table 1. AAH and GPD models classified by the energy dependence of their dual hopping, corresponding to Figure 4. The listed models represent limiting cases or constrained subfamilies of Eqs. (1) and (2), except for the related model of Ref. [69]. The geometric-mean hopping strengths J±,geomJ_{\pm,\mathrm{geom}} are defined in Eq. (7) for the reciprocal reference problem. Their thresholds are tDt_D at g=0g=0 and tDe∣g∣t_D e^{|g|} for the nonreciprocal criterion. Counts of physical mobility edges refer to regimes where localized and extended states coexist in the spectrum. Two algebraic branches of one localization condition are distinguished from the two separate directional candidate curves of the present model.

Nonreciprocal hopping and shifted mobility edges

Finally, we switch on the hopping asymmetry in real-space by setting g≠0g\ne0. In the dual equation, εm\varepsilon_m becomes 2tcos⁡(2παm−ig)2t\cos(2\pi\alpha m-ig), while the form of Eq. (4) is preserved. Physically, asymmetric hopping favors motion in one direction and competes with the confinement of quasiperiodic potential. Greater hopping asymmetry demands stronger disorder to localize a state spatially.

The imaginary gauge transformation for an open chain offers a way to understand this competition. It shows how asymmetric hopping changes the spatial profile of the wave function, motivating a comparison between ∣g∣|g| and the L0(E)L_0(E) of the reciprocal reference model. For the system under PBC, localization persists when L0(E)>∣g∣L_0(E)>|g|. The threshold L0(E)=∣g∣L_0(E)=|g| marks the balance between confinement and hopping asymmetry. This is the established imaginary gauge delocalization criterion of Hatano and Nelson [39, 40]. Applying Eq. (10) to the reciprocal localized spectrum then yields

max⁡{ρ+(E),ρ−(E)}=Rg,Rg≡De∣g∣.(14)\max\{\rho_{+}(E),\rho_{-}(E)\}=R_g,\qquad R_g\equiv D e^{|g|}. \tag*{(14)}

The comparison is thus unchanged in form, and we only need to replace the DD by RgR_g in the candidate ellipses and selection rule. Each channel controls a selected segment only if it reaches RgR_g while the other does not. Increasing ∣g∣|g| raises the disorder strength required to counteract hopping asymmetry, shifts candidate boundaries, and changes their intersections. The scales ρ±\rho_{\pm} remain those of the reciprocal reference problem.

Figure 5 compares the real part of PBC spectra for g=0g=0 and g=0.5g=0.5 as λ\lambda varies. The colors encode D2D_2. Solid curves pass the other channel comparison and dashed curves are suppressed by the other channel. The four stars mark the coincidences of the two candidates’ real-axis intercepts. Their coordinates are obtained by equating these intercepts, using DD in Figure 5(a) and RgR_g in Figure 5(b), as detailed in Supplementary eq:S31 and (S32). At each star, both channels simultaneously reach the threshold, providing analytical reference positions for comparing the candidate-branch structure with the spectrum.

Real parts of spectra under PBC for two values of $g$, with candidate channel boundaries and fractal dimension

Figure 5. Nonreciprocity shifts the channel thresholds. Real parts of spectra under PBC for (a) g=0g=0 and (b) g=0.5g=0.5, with t=1t=1, N=377N=377, α=233/377\alpha=233/377, γ=0.5\gamma=0.5, a=0.4a=0.4, and b=0.6b=0.6. Colors encode the fractal dimension D2D_2. Red and green curves show how the real-axis intercepts of the candidate boundaries for the ++ and −- channels, respectively, vary with λ\lambda. Solid (dashed) parts satisfy (fail) the selection inequality. Filled stars mark the four possible algebraic coincidences of real-axis intercepts of the two candidates, computed with R=DR=D in (a) and R=RgR=R_g in (b) [see Supplementary Eqs. (S31) and (S32) for details.]

Discussion

Our results reveal how complex quasiperiodic modulations can produce mobility edges controlled by different localization contributions. In momentum space, these contributions are associated with hopping along opposite directions. The generalized GPD potential makes both hopping terms energy dependent, but in different ways, so the direction that controls localization can change across the spectrum. The mobility edge can then combine portions of two different critical curves. Our model also incorporates representative AAH and GPD models as specific limits in which the two effective hopping terms are identical, dif-

fer only by a fixed factor, or are independent of energy. These extra constraints prevent the switch of control found in our general model.

The found competition implies that reaching the threshold of one Fourier channel does not necessarily imply a localization transition, as the other channel may dominate localization. The physical mobility edge is therefore formed by the selected spectral portions of the two candidate curves, as illustrated by the solid segments in Figure 2.

Beyond clarifying how multiple analytical contours jointly determine localization, our theoretical framework opens several avenues for future explorations. Potential experimental platforms include photonic lattices with engineered gain and loss, acoustic resonator arrays, ultracold atoms in bichromatic optical lattices, and active electrical circuits. Photonic quantum walks and electrical circuits have already demonstrated key ingredients of non-Hermitian lattice physics [29, 70]. Implementing the quasiperiodic modulation of the present model in such platforms could allow the predicted channel switching to be probed through spatially resolved intensity, transmission, or electrical-response measurements. Theoretically, extending our analysis to open boundaries would reveal how competing momentum-space channels interact with the non-Hermitian skin effect. Generalizing the mechanism to long-range hopping and interacting systems provides further directions for future work.

Methods

Analytical representation and channel scales

For reciprocal hopping (g=0g=0), we write d(θ)=d(eiθ)d(\theta)=d(e^{i\theta}). The transfer matrix and its regularized form are given by

TE(θ)=([E−V(θ)]/t−110),T~E(θ)=d(θ)TE(θ),det⁡TE=1.(15)T_E(\theta)=\begin{pmatrix} [E-V(\theta)]/t & -1 \\ 1 & 0 \end{pmatrix}, \qquad \widetilde{T}_E(\theta)=d(\theta)T_E(\theta),\qquad\det T_E=1. \tag*{(15)}

We use ϵ∈R\epsilon\in\mathbb{R} for an imaginary phase shift. In our parameter regime, the denominator of V(θ)V(\theta) is nonzero for all θ∈R\theta\in\mathbb{R} but may vanish after the shift θ→θ+iϵ\theta\to\theta+i\epsilon. The regularized transfer matrix has no such poles, and its scalar normalization can be removed exactly. Let T~E,ϵ\widetilde{T}_{E,\epsilon} and T~E,ϵ(N)\widetilde{T}^{(N)}_{E,\epsilon} be the ordered products of NN transfer matrices along the chain. Since the scalar factors commute with all matrices, the products satisfy

T~E,ϵ(N)(θ)=[∏j=0N−1d(θ+2πjα+iϵ)]TE,ϵ(N)(θ).(16)\widetilde{T}^{(N)}_{E,\epsilon}(\theta)=\left[\prod_{j=0}^{N-1}d(\theta+2\pi j\alpha+i\epsilon)\right]T^{(N)}_{E,\epsilon}(\theta). \tag*{(16)}

On a phase circle without zeros in the denominators, taking logarithmic norms and averaging over θ\theta gives the same scalar average for each factor. After division by NN and taking the limit N→∞N\to\infty, the phase-averaged Lyapunov exponents obey

L0(E,ϵ)=lim⁡N→∞1N∫02πdθ2πln⁡∥TE,ϵ(N)(θ)∥,L_0(E,\epsilon)=\lim_{N\to\infty}\frac{1}{N}\int_0^{2\pi}\frac{d\theta}{2\pi}\ln\left\|T^{(N)}_{E,\epsilon}(\theta)\right\|,
L~(E,ϵ)=lim⁡N→∞1N∫02πdθ2πln⁡∥T~E,ϵ(N)(θ)∥,\widetilde{L}(E,\epsilon)=\lim_{N\to\infty}\frac{1}{N}\int_0^{2\pi}\frac{d\theta}{2\pi}\ln\left\|\widetilde{T}^{(N)}_{E,\epsilon}(\theta)\right\|,
L0(E,ϵ)=L~(E,ϵ)−Id(ϵ),L_0(E,\epsilon)=\widetilde{L}(E,\epsilon)-I_d(\epsilon),
Id(ϵ)=∫02πdθ2πln⁡∣d(θ+iϵ)∣,ln⁡D=Id(0).(17)I_d(\epsilon)=\int_0^{2\pi}\frac{d\theta}{2\pi}\ln\left|d(\theta+i\epsilon)\right|, \qquad \ln D=I_d(0). \tag*{(17)}

Each regularized transfer matrix differs by a scalar dd, so the result picks up the product of these scalar factors. Its logarithmic average Id(ϵ)I_d(\epsilon) gives the subtraction in Eq. (17). For A≠0A\ne0, let z1z_1 and z2z_2 be the roots of Az2−z+B=0Az^2-z+B=0, so that d(z)=−A(z−z1)(z−z2)/zd(z)=-A(z-z_1)(z-z_2)/z. Because ∣A∣+∣B∣=max⁡{∣a∣,∣b∣}<1|A|+|B|=\max\{|a|,|b|\}<1, the term −Az2-Az^2 dominates Az2+BAz^2+ B on the unit circle. Therefore, exactly one root is inside and the other is outside. Jensen’s formula then gives

z1,2=1±1−4AB2A,z_{1,2}=\frac{1\pm\sqrt{1-4AB}}{2A},
D=∣A∣∏j=12max⁡{1,∣zj∣}=1+1−4AB2.(18)D=|A|\prod_{j=1}^{2}\max\{1,|z_j|\}=\frac{1+\sqrt{1-4AB}}{2}. \tag*{(18)}

For A=0A=0, ∣B∣<1|B|<1 gives D=1D=1, consistent with the same expression. The locations of two roots also imply D>∣A∣,∣B∣D>|A|,|B|, as required by the contour construction in Results. Since no roots cross the unit circle for a sufficiently small phase shift, Id(ϵ)=ln⁡DI_d(\epsilon)=\ln D near ϵ=0\epsilon=0. Supplementary Note 2 gives Id(ϵ)I_d(\epsilon) for arbitrary phase shifts.

Substituting Ed−v=E−p+z−p−z−1Ed-v=E-p_+z-p_-z^{-1} and d=1−Az−Bz−1d=1-Az-Bz^{-1} into the regularized transfer matrix gives the Fourier decomposition

T~E(z)=((Ed−v)/t−dd0)=z(−p+(E)/tA−A0)+(E/t−110)+z−1(−p−(E)/tB−B0)=zM+(E)+M0(E)+z−1M−(E).(19)\begin{aligned} \widetilde{T}_E(z)= \begin{pmatrix} (Ed-v)/t & -d\\ d & 0 \end{pmatrix} =z\begin{pmatrix} -p_+(E)/t & A\\ -A & 0 \end{pmatrix} +\begin{pmatrix} E/t & -1\\ 1 & 0 \end{pmatrix} +z^{-1}\begin{pmatrix} -p_-(E)/t & B\\ -B & 0 \end{pmatrix} \\ &= zM_+(E)+M_0(E)+z^{-1}M_-(E). \tag*{(19)} \end{aligned}

The upper-left entries of M±M_\pm are the coefficients −p±(E)/t-p_\pm(E)/t from Ed−vEd-v. Their off-diagonal entries A,−AA,-A and B,−BB,-B come from the multiplication by d(z)d(z). Thus, M±M_\pm are the coefficient matrices of the two Fourier harmonics. The complete regularized matrix has the determinant d2(z)d^2(z).

With s±=As_\pm=A and s−=Bs_-=B as defined in Results, let ητ,μ\eta_{\tau,\mu} (μ=±1\mu=\pm1) be the two eigenvalues of MτM_\tau for τ=±\tau=\pm. They are the roots of equation

det⁡(ηI−Mτ)=η2+pτ(E)tη+sτ2=0.(20)\det(\eta I-M_\tau)=\eta^2+\frac{p_\tau(E)}{t}\eta+s_\tau^2=0. \tag*{(20)}

Consequently, their larger modulus gives the channel scale:

Mτ(E)=(−pτ(E)/tsτ−sτ0),M_\tau(E)=\begin{pmatrix} -p_\tau(E)/t & s_\tau\\ -s_\tau& 0 \end{pmatrix},
ητ,μ(E)=−pτ(E)+μpτ2(E)−4t2sτ22t,\eta_{\tau,\mu}(E)=\frac{-p_\tau(E)+\mu\sqrt{p_\tau^2(E)-4t^2s_\tau^2}}{2t},
ρτ(E)=max⁡μ=±1∣ητ,μ(E)∣,\rho_\tau(E)=\max_{\mu=\pm1}|\eta_{\tau,\mu}(E)|,
ητ,+1+ητ,−1=−pτ(E)/t,ητ,+1ητ,−1=sτ2.(21)\eta_{\tau,+1}+\eta_{\tau,-1}=-p_\tau(E)/t, \qquad \eta_{\tau,+1}\eta_{\tau,-1}=s_\tau^2. \tag*{(21)}

Changing the square-root branch only interchanges the two eigenvalues and does not affect ρτ\rho_\tau. Repeated roots are also included in this definition. For g=0g = 0, we denote the logarithmic average of the magnitude of hopping in Eq. (4) over one period of its quasiperiodic modulation by Iτ(E)I_\tau(E), i.e.,

Iτ(E)=∫02πdθ2πln⁡∣pτ(E)−2tsτcos⁡θt∣.(22)I_\tau(E) = \int_0^{2\pi} \frac{\mathrm{d}\theta}{2\pi} \ln\left|\frac{p_\tau(E) - 2t s_\tau\cos\theta}{t}\right|. \tag*{(22)}

For sτ≠0s_\tau\ne0, we set z=eiθz = e^{i\theta}. The sum and product of eigenvalues in Eq. (21) give the factorization

pτ−tsτ(z+z−1)t=−sτz−1∏μ=±1(z+ητ,μsτ).(23)\frac{p_\tau- t s_\tau(z+z^{-1})}{t} = -s_\tau z^{-1}\prod_{\mu=\pm1}\left(z+\frac{\eta_{\tau,\mu}}{s_\tau}\right). \tag*{(23)}

Here, z−1z^{-1} has unit modulus on the unit circle. Jensen’s formula gives the phase average of the logarithm of each remaining factor as ln⁡max⁡{1,∣ητ,μ/sτ∣}\ln\max\{1,|\eta_{\tau,\mu}/s_\tau|\}. Hence

Iτ(E)=ln⁡∣sτ∣+∑μ=±1ln⁡max⁡{1,∣ητ,μ∣∣sτ∣}=ln⁡ρτ(E).(24)I_\tau(E) = \ln|s_\tau| + \sum_{\mu=\pm1}\ln\max\left\{1,\frac{|\eta_{\tau,\mu}|}{|s_\tau|}\right\} = \ln\rho_\tau(E). \tag*{(24)}

To see the last equality, the product of the eigenvalue moduli is sτ2s_\tau^2. The larger modulus is therefore at least ∣sτ∣|s_\tau|, and the smaller is at most ∣sτ∣|s_\tau|. Only the larger eigenvalue contributes a possibly nonzero Jensen term, namely ln⁡(ρτ/∣sτ∣)\ln(\rho_\tau/|s_\tau|). This argument also covers equal eigenvalue moduli. Isolated zeros of the hopping profile give integrable logarithmic singularities.

Exponentiating and restoring the hopping unit tt yields the geometric-mean strength defined in Eq. (7), i.e.,

Jτ,geom(E)=tρτ(E)=texp⁡[Iτ(E)]=texp⁡[∫02πdθ2πln⁡∣pτ(E)−2tsτcos⁡θt∣]=12max⁡μ=±1∣−pτ(E)+μpτ2(E)−4t2sτ2∣.(25)\begin{aligned} J_{\tau,\mathrm{geom}}(E) = t\rho_\tau(E) = t\exp[I_\tau(E)] \\ &= t\exp\left[\int_0^{2\pi}\frac{\mathrm{d}\theta}{2\pi}\ln\left|\frac{p_\tau(E)-2ts_\tau\cos\theta}{t}\right|\right] \\ &= \frac{1}{2}\max_{\mu=\pm1}\left|-p_\tau(E)+\mu\sqrt{p_\tau^2(E)-4t^2s_\tau^2}\right|. \tag*{(25)} \end{aligned}

The last line is tt times the spectral radius in Eq. (21). Thus, J±,geom(E)=tρ±(E)J_{\pm,\mathrm{geom}}(E)=t\rho_{\pm}(E) are the geometric-mean magnitudes of the full directional effective hopping profiles. For sτ=0s_\tau=0, the hopping profile is constant, Iτ=ln⁡(∣pτ∣/t)I_\tau=\ln(|p_\tau|/t), and the hopping strength reduces to ∣pτ∣=tρτ|p_\tau|=t\rho_\tau. If pτ=0p_\tau=0 as well, Iτ=−∞I_\tau=-\infty and the strength is zero, so the same identity includes a vanishing channel.

Complex phase shifts and channel growth scales

To separate the contributions of two Fourier channels, we first shift θ→θ+iϵ\theta\to\theta+i\epsilon, so that ∣z∣=e−ϵ|z|=e^{-\epsilon}. For large negative (positive) ϵ\epsilon, the term zM+zM_+ (z−1M−z^{-1}M_-) dominates the Laurent expansion. Removing the corresponding scalar power of zz leaves, in each limit, the constant matrix M+M_+ or M−M_-. The powers removed account for the linear −ϵ-\epsilon or +ϵ+\epsilon term in the Lyapunov exponent. Since a constant matrix grows at the logarithm of its spectral radius, the two limits are

L~(E,ϵ)=−ϵ+ln⁡ρ+(E)+o(1),ϵ→−∞,\widetilde{L}(E,\epsilon)=-\epsilon+\ln\rho_+(E)+o(1),\qquad\epsilon\to-\infty,
L~(E,ϵ)=+ϵ+ln⁡ρ−(E)+o(1),ϵ→+∞.(26)\widetilde{L}(E,\epsilon)=+\epsilon+\ln\rho_-(E)+o(1),\qquad\epsilon\to+\infty. \tag*{(26)}

These asymptotic expressions apply to nonzero channel radii. The o(1)o(1) terms vanish in the stated limits. A vanishing channel does not contribute finite logarithmic bounds, as detailed in the Supplementary Note 3. The convexity of the regularized Lyapunov exponent makes the two limiting straight lines lower bounded at ϵ=0\epsilon=0 [71, 72, 73]. The determinant-one condition also implies L0(E)≥0L_0(E)\geq0. Together, these facts give

L~(E,0)≥max⁡{ln⁡D,ln⁡ρ+(E),ln⁡ρ−(E)}.(27)\widetilde{L}(E,0)\geq\max\{\ln D,\ln\rho_+(E),\ln\rho_-(E)\}. \tag*{(27)}

To show that the bound is attained for spectral energies, we apply the quantization of acceleration in Avila’s global theory [67]. The denominator remains nonzero for θ+iϵ\theta+i\epsilon in a strip ∣ϵ∣<δ|\epsilon|<\delta around ϵ=0\epsilon=0, for some δ>0\delta>0, so the original matrix defines an analytic SL(2,C)\mathrm{SL}(2,\mathbb{C}) cocycle there. The function L0(E,ϵ)L_0(E,\epsilon) has integer-quantized one-sided slopes with respect to ϵ\epsilon in this strip. The scalar term Id(ϵ)I_d(\epsilon) is constant near ϵ=0\epsilon=0. Therefore, fE(ϵ)=L~(E,ϵ)f_E(\epsilon)=\widetilde{L}(E,\epsilon) also has integer-quantized slopes close to zero. The convexity and the two outer limits only allow slopes −1-1, 00, and +1+1.

If Eq. (27) were strict at a spectral energy, a slope −1-1 or +1+1 near zero would force the convex graph onto its corresponding outer supporting line, contradicting strictness. The only remaining local slope is zero. Since Re⁡(eiθ)>0\operatorname{Re}(e^{i\theta})>0 for ∣α∣<1|\alpha|<1, the denominator has zero winding and Id(ϵ)I_d(\epsilon) is constant near zero. The original Lyapunov exponent would thus be positive and locally constant under imaginary phase shifts. Avila’s regularity criterion then makes the cocycle uniformly hyperbolic, which places EE outside the spectrum of the reciprocal Schrödinger operator. This contradiction forces equality in Eq. (27), and a subtraction of ln⁡D\ln D yields Eq. (10).

Imaginary gauge choice and boundary conditions

For nonreciprocal hopping, the substitution ψn=e−gnϕn\psi_n=e^{-g n}\phi_n gives a reciprocal bulk equation. For a reciprocal localized state centered near n0n_0, this factor changes the decay rates of right and left tails to κR=L0(E)+g\kappa_R=L_0(E)+g and κL=L0(E)−g\kappa_L=L_0(E)-g. Both rates remain positive when L0(E)>∣g∣L_0(E)>|g|. The weaker one vanishes at L0(E)=∣g∣L_0(E)=|g|, yielding the threshold Rg=De∣g∣R_g=De^{|g|}. The finite-open-chain similarity transformation and the corresponding periodic-boundary relation are derived in Supplementary Note 4.

Numerical comparisons

The ratios of consecutive Fibonacci numbers approximate the irrational frequency α=(5−1)/2\alpha=(\sqrt{5}-1)/2. The system sizes and parameters are given in the figure captions. For the Lyapunov-exponent comparison in Figure 3, we first diagonalize the N=377N=377 periodic chain with α=233/377\alpha=233/377. We sort the eigenenergies by LanL_{\mathrm{an}} and select 60 entries approximately equally spaced in this order. For each selected energy, we iterate the original reciprocal transfer matrices for 30000 steps using the irrational frequency α=(5−1)/2\alpha=(\sqrt{5}-1)/2 and the single initial phase φ=0\varphi=0, without phase averaging. Starting from the vector (1,0)T(1,0)^{T}, we normalize the propagated vector after each step and accumulate the logarithms of its Euclidean norm. Their average gives LnumL_{\mathrm{num}}. The transfer-matrix iteration itself does not use the channel-radius formula, which is used only for the analytical comparison and the sampling order.

Figure 3. Numerical test of the exact Lyapunov exponent. Parameters are t=1t=1, a=0.85a=0.85, b=0.1b=0.1, λ=0\lambda=0, γ=0.6\gamma=0.6, g=0g=0, and φ=0\varphi=0. The 60 test energies are sampled from the periodic chain with N=377N=377 and α=233/377\alpha=233/377. At each energy, LnumL_{\mathrm{num}} is obtained from 30000 normalized transfer-matrix steps at α=(5−1)/2\alpha=(\sqrt{5}-1)/2, while LanL_{\mathrm{an}} is evaluated from Eq. (10). The dashed line indicates Lnum=LanL_{\mathrm{num}}=L_{\mathrm{an}}.

Right eigenstates are normalized by ∑n∣ψn∣2=1\sum_n|\psi_n|^2=1. We evaluate the inverse participation ratio (IPR) and the finite-size spread indicator (fractal dimension) as

IPR=∑n∣ψn∣4,D2=−ln⁡IPRln⁡N.(28)\mathrm{IPR}=\sum_n|\psi_n|^4,\qquad D_2=-\frac{\ln\mathrm{IPR}}{\ln N}. \tag*{(28)}

For increasingly large systems, an extended state tends towards D2=1D_2=1, while a localized state tends towards D2=0D_2=0. At finite sizes, D2D_2 is only a measure of how broadly a right eigenstate is distributed. Intermediate values do not demonstrate criticality. In particular, Figure 5 presents the real parts of eigenenergies, so an apparent crossing in that plot must not be mistaken for the exact membership of a complex contour.

Acknowledgements

L.Z. is supported by the National Natural Science Foundation of China (Grant No. 12275260), the Shandong Provincial Natural Science Foundation (Grant No. ZR2026QB21), and the Fundamental Research Funds for the Central Universities (Grant No. 23640008). W.Z. is supported by the National Natural Science Foundation of China (Grants No. 12404499).

References

  1. [1]Y. E. Kraus, Y. Lahini, Z. Ringel, M. Verbin, and O. Zilberberg, Topological States and Adiabatic Pumping in Quasicrystals, Phys. Rev. Lett. 109, 106402 (2012).pubmed.ncbi.nlm.nih.gov/23005308
  2. [2]Y. E. Kraus and O. Zilberberg, Topological Equivalence between the Fibonacci Quasicrystal and the Harper Model, Phys. Rev. Lett. 109, 116404 (2012).
  3. [3]M. Verbin, O. Zilberberg, Y. E. Kraus, Y. Lahini, and Y. Silberberg, Observation of Topological Phase Transitions in Photonic Quasicrystals, Phys. Rev. Lett. 110, 076403 (2013).pubmed.ncbi.nlm.nih.gov/25166388
  4. [4]D. B. Grempel, S. Fishman, and B. E. Prange, Localization in an Incommensurate Potential: An Exactly Solvable Model, Phys. Rev. Lett. 49, 833 (1982).DOI
  5. [5]J. B. Sokoloff, Unusual band structure, wave function and electrical conductance in crystals with incommensurate periodic potentials, Phys. Rep. 126, 189 (1985).DOI
  6. [6]B. Simon, Almost periodic Schrödinger operators I. V. The Maryland model, Ann. Phys. (NY) 159, 157 (1985).
  7. [7]P. G. Harper, Single band motion of conduction electrons in a uniform magnetic field, Proc. Phys. Soc. A 68, 874 (1955).DOI
  8. [8]S. Aubry and G. André, Analyticity breaking and Anderson localization in incommensurate lattices, Ann. Israel Phys. Soc. 3, 133 (1980).
  9. [9]S. Ya. Jitomirskaya, Metal-insulator transition for the almost Mathieu operator, Ann. Math. 150, 1159–1175 (1999).arxiv.org/abs/math/9911265
  10. [10]J. Biddle and S. Das Sarma, Predicted mobility edges in one-dimensional incommensurate optical lattices: An exactly solvable model of Anderson localization, Phys. Rev. Lett. 104, 070601 (2010).pubmed.ncbi.nlm.nih.gov/20366867
  11. [11]J. Biddle, D. J. Priour, Jr., B. Wang, and S. Das Sarma, Localization in one-dimensional lattices with non-nearest-neighbor hopping: Generalized Anderson and Aubry-André models, Phys. Rev. B 83, 075105 (2011).DOI
  12. [12]X. Deng, S. Ray, S. Sinha, G. V. Shlyapnikov, and L. Santos, One-Dimensional Quasicrystals with Power-Law Hopping, Phys. Rev. Lett. 123, 025301 (2019).DOI
  13. [13]C. M. Soukoulis and E. N. Economou, Localization in One-Dimensional Lattices in the Presence of Incommensurate Potentials, Phys. Rev. Lett. 48, 1043–1046 (1982).DOI
  14. [14]S. Ganeshan, J. H. Pixley, and S. Das Sarma, Nearest-neighbor hopping models with an exact mobility edge in one dimension, Phys. Rev. Lett. 114, 146601 (2015).pubmed.ncbi.nlm.nih.gov/25910146
  15. [15]Y. Wang, X. Xia, Y. Wang, Z. Zheng, and X.-J. Liu, Duality between two generalized Aubry–André models with exact mobility edges, Phys. Rev. B 103, 174205 (2021).DOI
  16. [16]X. Cai and Y.-C. Yu, Exact mobility edges in quasiperiodic systems without self-duality, J. Phys.: Condens. Matter 35, 035602 (2023).DOI
  17. [17]S. Banerjee, S. R. Padhi, and T. Mishra, Emergence of distinct exact mobility edges in a quasiperiodic chain, Phys. Rev. B 111, L220201 (2025).arxiv.org/abs/2503.19834
  18. [18]Y. Wang, X. Xia, L. Zhang, H. Yao, S. Chen, J. You, Q. Zhou, and X.-J. Liu, One-dimensional quasiperiodic mosaic lattice with exact mobility edges, Phys. Rev. Lett. 125, 196604 (2020).DOI
  19. [19]H.-Y. Cui, Y.-C. Yu, and X. Cai, Engineering exact mobility edges in quasiperiodic Aharonov–Bohm chains, Phys. Rev. B 114, 134202 (2026).DOI
  20. [20]H.-T. Hu, X. Lin, A.-M. Guo, G. Guo, Z. Lin, and M. Gong, Hidden self duality and exact mobility edges in quasiperiodic network models, Phys. Rev. Lett. 134, 246301 (2025).DOI
  21. [21]H.-T. Hu, Y. Chen, X. Lin, A.-M. Guo, Z. Lin, and M. Gong, Exact mobility edges in quasiperiodic network models with slowly varying potentials, Phys. Rev. B 112, 054201 (2025).DOI
  22. [22]X. Li and S. Das Sarma, Mobility edge and intermediate phase in one-dimensional incommensurate lattice potentials, Phys. Rev. B 101, 064203 (2020).DOI
  23. [23]S. Lee, T. Čadež, and K.-M. Kim, Structural constraints on mobility edges in one-dimensional quasiperiodic systems, Phys. Rev. B 113, 174202 (2026).DOI
  24. [24]G. Roati, C. D’Errico, L. Fallani, M. Fattori, C. Fort, M. Zaccanti, G. Modugno, M. Modugno, and M. Inguscio, Anderson localization of a non-interacting Bose–Einstein condensate, Nature 453, 895–898 (2008).
  25. [25]F. A. An, E. J. Meier, and B. Gadway, Engineering a Flux-Dependent Mobility Edge in Disordered Zigzag Chains, Phys. Rev. X 8, 031045 (2018).DOI
  26. [26]H. P. Lüschen, S. Scherg, T. Kohlert, M. Schreiber, P. Bordia, X. Li, S. Das Sarma, and I. Bloch, Single-Particle Mobility Edge in a One-Dimensional Quasiperiodic Optical Lattice, Phys. Rev. Lett. 120, 160404 (2018).DOI
  27. [27]F. A. An, K. Padavić, E. J. Meier, S. Hegde, S. Ganesha[n], J. H. Pixley, S. Vishveshwara, and B. Gadway, Interactions and mobility edges: Observing the generalized Aubry–André model, Phys. Rev. Lett. 126, 040603 (2021).DOI
  28. [28]Y. Lahini, R. Pugatch, F. Pozzi, M. Sorel, R. Morandotti, N. Davidson, and Y. Silberberg, Observation of a Localization Transition in Quasiperiodic Photonic Lattices, Phys. Rev. Lett. 103, 013901 (2009).pubmed.ncbi.nlm.nih.gov/19659147
  29. [29]S. Weidemann, M. Kremer, S. Longhi, and A. Szameit, Topological triple phase transition in non-Hermitian Floquet quasicrystals, Nature 601, 354–359 (2022).DOI
  30. [30]Q. Lin, C. Czedich, Q. Zhou, and P. Xue, Observation of Metal-Insulator and Spectral Phase Transitions in Aubry-André-Harper Models, Phys. Rev. Lett. 136, 206602 (2026).DOI
  31. [31]Q. Lin, T. Li, L. Xiao, K. Wang, W. Yi, and P. Xue, Topological Phase Transitions and Mobility Edges in Non-Hermitian Quasicrystals, Phys. Rev. Lett. 129, 113601 (2022).DOI
  32. [32]Y.-J. Chang, J.-H. Zhang, Y.-H. Lu, Y.-Y. Yang, F. Mei, J. Ma, S. Jia, and X.-M. Jin, Observation of Photonic Mobility Edge Phases, Phys. Rev. Lett. 134, 053601 (2025).DOI
  33. [33]Y. Ashida, Z. Gong, and M. Ueda, Non-Hermitian physics, Adv. Phys. 69, 249–435 (2020).
  34. [34]E. J. Bergholtz, J. C. Budich, and F. K. Kunst, Exceptional topology of non-Hermitian systems, Rev. Mod. Phys. 93, 015005 (2021).DOI
  35. [35]C. M. Bender and D. W. Hook, 𝒫𝒯-symmetric quantum mechanics, Rev. Mod. Phys. 96, 045002 (2024).
  36. [36]A. Jazaeri and I. I. Satija, Localization transition in incommensurate non-Hermitian systems, Phys. Rev. E 63, 036222 (2001).pubmed.ncbi.nlm.nih.gov/11308756
  37. [37]W. Han and L. Zhou, Dimerization-induced mobility edges and multiple reentrant localization transitions in non-Hermitian quasicrystals, Phys. Rev. B 105, 054204 (2022).
  38. [38]Y.-Q. Zheng, S.-Z. Li, and Z. Li, Emergent multiloop nested point gap in a non-Hermitian quasiperiodic lattice, Phys. Rev. B 111, 104204 (2025).DOI
  39. [39]N. Hatano and D. R. Nelson, Localization transitions in non-Hermitian quantum mechanics, Phys. Rev. Lett. 77, 570 (1996).DOI
  40. [40]N. Hatano and D. R. Nelson, Non-Hermitian delocalization and eigenfunctions, Phys. Rev. B 58, 8384 (1998).DOI
  41. [41]H. Jiang, L.-J. Lang, C. Yang, S.-L. Zhu, and S. Chen, Interplay of non-Hermitian skin effects and Anderson localization in nonreciprocal quasiperiodic lattices, Phys. Rev. B 100, 054301 (2019).DOI
  42. [42]S. Longhi, Phase transitions in a non-Hermitian Aubry–André–Harper model, Phys. Rev. B 103, 054203 (2021).DOI
  43. [43]L. Zhou and Y. Gu, Topological delocalization transitions and mobility edges in the nonreciprocal Maryland model, J. Phys.: Condens. Matter 34, 115402 (2022).DOI
  44. [44]F. K. Kunst, E. Edvardsson, J. C. Budich, and E. J. Bergholtz, Biorthogonal Bulk-Boundary Correspondence in Non-Hermitian Systems, Phys. Rev. Lett. 121, 026808 (2018).DOI
  45. [45]S. Yao and Z. Wang, Edge States and Topological Invariants of Non-Hermitian Systems, Phys. Rev. Lett. 121, 086803 (2018).DOI
  46. [46]K. Yokomizo and S. Murakami, Non-Bloch Band Theory of Non-Hermitian Systems, Phys. Rev. Lett. 123, 066404 (2019).
  47. [47]N. Okuma, K. Kawabata, K. Shiozaki, and M. Sato, Topological Origin of Non-Hermitian Skin Effects, Phys. Rev. Lett. 124, 086801 (2020).DOI
  48. [48]Y. Liu, Y. Wang, X.-J. Liu, Q. Zhou, and S. Chen, Exact mobility edges, 𝒫𝒯-symmetry breaking, and skin effect in one-dimensional non-Hermitian quasicrystals, Phys. Rev. B 103, 014203 (2021).DOI
  49. [49]S. Longhi, Topological Phase Transition in non-Hermitian Quasicrystals, Phys. Rev. Lett. 122, 237601 (2019).DOI
  50. [50]Y. Liu, X.-P. Jiang, J. Cao, and S. Chen, Non-Hermitian mobility edges in one-dimensional quasicrystals with parity-time symmetry, Phys. Rev. B 101, 174205 (2020).DOI
  51. [51]L. Zhou and W. Han, Driving-induced multiple 𝒫𝒯-symmetry breaking transitions and reentrant localization transitions in non-Hermitian Floquet quasicrystals, Phys. Rev. B 106, 054307 (2022).DOI
  52. [52]G. Zhang, B. Shao, L. Zhou, J. Gong, and W. Zhu, Topological metal-insulator transitions in one-dimensional non-Hermitian quasicrystals: Beyond 𝒫𝒯 symmetry, Phys. Rev. B 114, 305407 (2026).
  53. [53]S. Longhi, Metal-insulator phase transition in a non-Hermitian Aubry–André–Harper model, Phys. Rev. B 100, 125157 (2019).DOI
  54. [54]S. Longhi, Non-Hermitian Maryland model, Phys. Rev. B 103, 224206 (2021).DOI
  55. [55]S.-Z. Li and Z. Li, Ring structure in the complex plane: A fingerprint of a non-Hermitian mobility edge, Phys. Rev. B 110, L041102 (2024).DOI
  56. [56]L. Wang, J. Liu, Z. Wang, and S. Chen, Exact complex mobility edges and flagellate-like spectra for non-Hermitian quasicrystals with exponential hoppings, Phys. Rev. B 110, 144205 (2024).DOI
  57. [57]G.-X. Pang, Z. Li, S.-Z. Li, Y.-Y. Zhang, J.-F. Liu, and Y.-C. Zhang, Exact mobility line and mobility ring in the complex energy plane of a flat-band lattice with a non-Hermitian quasiperiodic potential, Phys. Rev. B 111, 214205 (2025).DOI
  58. [58]R.-J. Chen, G.-Q. Zhang, Z. Li, and D.-W. Zhang, Mobility rings in a non-Hermitian non-Abelian quasiperiodic lattice, Phys. Rev. A 112, 013320 (2025).DOI
  59. [59]T. Liu, H. Guo, Y. Pu, and S. Longhi, Generalized Aubry–André self-duality and mobility edges in non-Hermitian quasiperiodic lattices, Phys. Rev. B 102, 024205 (2020).DOI
  60. [60]Q.-B. Zeng and Y. Xu, Winding numbers and generalized mobility edges in non-Hermitian systems, Phys. Rev. Research 2, 033052 (2020).DOI
  61. [61]Y. Liu, Y. Wang, Z. Zheng, and S. Chen, Exact non-Hermitian mobility edges in one-dimensional quasicrystal lattice with exponentially decaying hopping and its dual lattice, Phys. Rev. B 103, 134208 (2021).DOI
  62. [62]Y. Liu, Q. Zhou, and S. Chen, Localization transition, spectrum structure, and winding numbers for one-dimensional non-Hermitian quasicrystals, Phys. Rev. B 104, 024201 (2021).DOI
  63. [63]X. Xia, K. Huang, S. Wang, and X. Li, Exact mobility edges in the non-Hermitian t₁–t₂ model: Theory and possible experimental realizations, Phys. Rev. B 105, 014207 (2022).
  64. [64]L. Wang, Z. Wang, J. Liu, and S. Chen, Exact multiple complex mobility edges and quantum state engineering in coupled one-dimensional quasicrystals, Phys. Rev. B 112, 104207 (2025).DOI
  65. [65]S.-R. He, J. Fan, L. Li, R. Wang, D.-H. Xu, and Z.-H. Wang, Exact mobility edges and hidden self-duality in non-Hermitian quasiperiodic network models, Phys. Rev. B 113, 224205 (2026).DOI
  66. [66]Z. Wang, N. Lu, D. Liu, X. Yang, and X. Tong, Non-Hermitian mosaic Maryland model, Phys. Rev. B 114, 094210 (2026).
  67. [67]A. Avila, Global theory of one-frequency Schrödinger operators, Acta Math. 215, 1–54 (2015).DOI
  68. [68]M. Gonçalves, B. Amorim, E. V. Castro, and P. Ribeiro, Hidden dualities in 1D quasiperiodic lattice models, SciPost Phys. 13, 046 (2022).DOI
  69. [69]Z. Wang, Y. Zhang, L. Wang, and S. Chen, Engineering mobility in quasiperiodic lattices with exact mobility edges, Phys. Rev. B 108, 174202 (2023).DOI
  70. [70]T. Helbig, T. Hofmann, S. Imhof, M. Abdelghany, T. Kiessling, L. W. Molenkamp, C. H. Lee, A. Szameit, M. Greiter, and R. Thomale, Generalized bulk–boundary correspondence in non-Hermitian topoelectrical circuits, Nat. Phys. 16, 747–750 (2020), doi:10.1038/s41567-020-0922-9.DOI
  71. [71]S. Jitomirskaya and C. A. Marx, Analytic quasi-periodic cocycles with singularities and the Lyapunov exponent of extended Harper’s model, Commun. Math. Phys. 316, 237–267 (2012).arxiv.org/abs/1010.0751
  72. [72]S.-Z. Li, E. Cheng, S.-L. Zhu, and Z. Li, Asymmetric transfer matrix analysis of Lyapunov exponents in one-dimensional nonreciprocal quasicrystals, Phys. Rev. B 110, 134203 (2024).DOI
  73. [73]X. Tong, Y. Zhang, B. Li, and X. Yang, Impact of nonreciprocal hopping on localization in non-Hermitian quasiperiodic systems, Phys. Rev. B 111, 214202 (2025).arxiv.org/abs/2501.14343

Paper details

vlnc:W_m8tbddez
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 8

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