WorksPhysics and Astronomy
Bypass Transition via the Non-Hermitian Skin Effect
Abstract
Non-Hermitian skin lattices can directionally amplify propagating signals. Here we show that this amplification and nonlinear feedback together drive a bypass transition to self-sustained motion, familiar from shear flows. A finite local excitation can trigger sustained periodic or chaotic motion while the resting state remains linearly stable. We develop a return-map theory that predicts the triggering threshold, its scaling with directional gain, and the instability governing near-critical dynamics. The threshold decreases exponentially with chain length. Weak noise triggers sustained motion as well, with a characteristic ignition length that grows logarithmically as noise weakens at fixed observation time. Our work establishes a skin-effect route to bypass transitions, opening a new avenue for predicting and controlling nonlinear wave dynamics.
The non-Hermitian skin effect (NHSE) causes an extensive number of eigenmodes to accumulate near a boundary [1, 2, 3, 4, 5, 6, 7, 8, 9] and has been observed across diverse platforms [10, 11, 12, 13, 14, 15, 16]. It can produce direction-dependent amplification or attenuation of propagating waves [17, 19, 20, 21] [18]. Potential applications include directional amplifiers [22, 23], sensors [24, 25], and cooling schemes [26]. Remarkably, interference between nonorthogonal eigenmodes can transiently amplify an excitation even when every mode decays [27, 28, 29, 30, 31]. Such transient growth ultimately gives way to decay without additional driving.
Nonlinearity is becoming increasingly important in skin physics [32, 33], giving rise to solitons [34, 35], self-trapping [36], and spatial bifurcations of stationary states [37, 38]. Nonlinear lattice dynamics can also support synchronization [39] and temporal chaos [40], and enable single-mode lasing [41] and mode locking [42]. In realistic systems, nonlinearity is often unavoidable, while skin amplification can magnify weak signals or fluctuations until nonlinear effects become important. A fundamental question is how such systems enter self-sustained motion, and whether this onset can be predicted quantitatively. This question is also directly relevant to the stability and performance of active devices. A predictive theory linking skin amplification to self-sustained motion is still lacking.
Here we identify a bypass transition to self-sustained motion via the NHSE. A finite excitation can initiate periodic or chaotic motion while the resting state remains stable, analogous to the bypass route from a laminar state to turbulence in shear flows [43, 44, 45, 46, 47]. We demonstrate the mechanism in a minimal setting of two open chains with opposite skin directions, coupled by intensity-dependent conversion Figure 1(a). A local excitation is amplified along one chain and converted into the other, where it is amplified again on its return. We develop a return-map theory that predicts the triggering threshold and its scaling with directional gain. The threshold decreases exponentially with chain length despite an unchanged linear decay rate. Near threshold, trajectories follow an unstable circulating orbit whose dominant instability is quantified. Weak noise
Figure 1. Bypass transition via the non-Hermitian skin effect. (a) Two open chains A and B with opposite skin directions and uniform loss at rate on every site. Their complex field amplitudes at site are and , respectively. The local conversion rate is , where . Initially, only is nonzero. (b) Color maps of for (top to bottom), showing decay, critical circulation, and sustained chaos. Here is the single-site triggering threshold. Parameters: , , , and .
can initiate sustained motion, with a characteristic triggering length that grows logarithmically as noise weakens. Our work reveals a skin-effect mechanism for sustained motion, offering a fresh perspective on nonlinear wave dynamics.
Model and finite-amplitude triggering.—We consider two open chains, A and B, with sites each, as sketched in Figure 1(a). Their complex field amplitudes and obey
Here , with open boundaries . The mean hopping is set to unity. The parameter controls the hopping asymmetry, and denotes the uniform loss rate. The local intensity determines the conversion rate , with . For real , the conversion redistributes intensity between the two chains while conserving . Unless stated otherwise, , , and .
The zero-field configuration, , defines the resting state. The intensity-dependent conversion is cubic in the fields and vanishes upon linearization, leaving two independent Hatano–Nelson chains. With , the imaginary gauge transformation and maps the linearized equations to reciprocal chains with hopping [1, 48]. Their eigenfrequencies are
The eigenmodes accumulate at the right edge of chain A and the left edge of chain B, with amplitude localization length . Every eigenmode decays at the same rate , independent of .
We prepare a single-site excitation by setting at the leftmost site of chain A, with all other amplitudes zero, and let the system evolve without further driving. For this preparation, a threshold separates decay from sustained motion. Figure 1(b) illustrates the dynamics across this threshold at . Below threshold, successive passages weaken and the lattice returns to rest. At threshold, the packet circulates repeatedly with nearly constant peak intensity over the displayed time interval. Just above threshold, irregular motion spreads throughout the lattice and persists at late times. A positive largest Lyapunov exponent confirms the chaotic character of the sustained motion [49]. The threshold is determined by bisection and decreases exponentially with chain length [Fig. 2(a)], while the modal decay rate remains .
Return-map theory.— To predict the triggering threshold, we calculate the field generated during the first passage and return. Since only chain A is seeded, conversion produces a field in chain B at order and feeds back into chain A at order . We write and . Let be the hopping matrices and their propagators. The
Figure 2. Return-map prediction of the triggering threshold. (a) Single-site triggering threshold versus chain length for the parameter pairs shown. Symbols denote the numerical threshold , and lines the return-map prediction from Eq. (7). (b) Edge amplitudes , , and through the first passage and return at the numerical threshold for . Solid curves show the full nonlinear dynamics and open circles the leading responses and from Eqs. (3)–(4). The dashed line marks the fixed-point amplitude .
leading responses are
The first expression propagates a unit seed along chain A; the second sums the cubic sources generated during the passage and propagates them through chain B.
Let denote the first-arrival amplitude gain and the nonlinear return coefficient,
where the maxima are taken over the first arriving and returning packets, respectively. For a seed , the first arrival at the right edge of chain A has peak , while the return to the left edge of chain B has peak . The reflection-exchange symmetry of Eq. (1) makes the two propagation legs equivalent. Approximating successive packets as having the same normalized shape up to reflection, the successive arrival peaks at the right edge of chain A obey
The map has a nonzero fixed point . Since , packets below weaken on successive circuits, whereas those above it grow away from the fixed point. Requiring the first arrival to reach yields
Figure 2 (PDF p. 3).
Figure Fig. 2(a) compares from Eq. (7) with the thresholds obtained from the full dynamics over several lengths and parameter sets. They agree closely across the range shown, with no fit to the threshold data. In Fig. 2(b), the leading responses and closely reproduce the full evolution through the first return. The converted field is roughly two orders of magnitude weaker than the arriving field , but amplification along chain B regenerates a return peak of comparable height. Both the arriving and returning peaks lie close to , as given by Eq. (7).
The return map further predicts a distinctive scaling of the triggering threshold with the directional gain . An arriving packet of amplitude generates a nonlinear source proportional to , while propagation along the opposite chain contributes another factor . Thus , up to factors that vary subexponentially with . It follows that and
For comparison, a single-pass amplifier requires an input proportional to to reach a fixed output amplitude. Here nonlinear regeneration also lowers the circulating amplitude needed to close the return loop, , producing the additional factor in the triggering threshold. More generally, for a conversion of field order ( here), the same argument gives [49].
The length dependence of follows from the first-passage response. This response combines the directional amplification factor with wave-packet spreading and loss. Maximizing this response gives [49]
The same exponent governs the zero-frequency end-to-end Green’s function [19] [18]. Combining Eqs. (8) and (9) gives . For , and the triggering threshold falls exponentially with length, while every eigenmode still decays at rate .
Critical orbit and near-threshold dynamics
The nonzero fixed point of the return map suggests a critical circulating orbit, whose instability is also predicted by the map. Writing and expanding Eq. (6) around gives
around gives
A small deviation from the fixed point is thus amplified by a factor of nine after each circuit. The coefficient cancels, leaving a factor set solely by the two successive cubic conversions. More generally, for a conversion of field order , the full-circuit map has degree and gives .
We test this prediction by computing the corresponding periodic orbit independently from Eq. (1) using Newton shooting [49]. Its period is the time between successive arrivals at the same edge, and its arriving peak agrees closely with . Its stability is characterized by the Floquet multipliers [50]; apart from the neutral directions, there is a single unstable multiplier . Figure 3(a) shows that approaches the return-map prediction 9 with increasing length across the parameter sets considered. The threshold trajectory approaches this orbit, while nearby trajectories follow it temporarily before departing toward decay or sustained motion. The critical orbit thus lies on the boundary between the two outcomes, analogous to unstable solutions on the laminar–turbulent boundary in shear flows [51].

Figure 3. Critical-orbit signatures. (a) Unstable Floquet multiplier , normalized by the return-map prediction , versus chain length for the parameter pairs in Fig. 2(a). (b) Departure time near threshold at , with . The line uses the slope fixed by the independently computed Floquet multiplier. Only the offset is fitted.
For a near-threshold seed , the unstable displacement is proportional to and grows after circuits as . Defining as the departure time from the critical orbit gives
where is an offset. Figure 3(b) confirms this logarithmic dependence on both sides of threshold. The independently computed fixes the slope, while only is fitted. The logarithmic delay reflects repeated circuits near the unstable orbit rather than a slowing of the resting-state decay.
Power balance and the feedback loop.—The local conversion redistributes intensity between the chains but supplies no net power. The total power obeys
In sustained motion, the power supplied by nonreciprocal hopping balances the loss on average. In the chaotic regime, the mean power grows approximately linearly with and depends only weakly on the above-threshold seed amplitude [49]. The seed triggers the motion but does not power it.
For sufficiently strong loss, nonreciprocal hopping can no longer sustain the field. Equation (12) gives a global decay bound: every finite open lattice decays for , regardless of the initial amplitude [49]. This boundary also has a simple bulk interpretation. The periodic dispersions have maximum growth rate , so , equivalently , is the regime of directional amplification [17, 27] [18], even though all open-boundary eigenmodes decay. Directional amplification alone, however, cannot sustain motion. If both chains amplify toward the same end, the resting-state spectrum and optimal transient gain remain unchanged, yet the weighted power obeys , forcing every field to decay. With opposite skin directions, conversion closes the return loop by routing the field into another amplifying channel.
Noise-induced ignition.—Noise can also supply the finite disturbance required to trigger sustained motion. Starting from the resting state, we add independent complex Gaussian white noise to each of the field components, with , where is the noise strength. We record whether ignition occurs within an observation time . Figure 4(a) shows a sharp crossover with chain length: at fixed , the ignition probability rises from nearly zero to nearly one over only a few sites, while weaker noise shifts the crossover to longer chains. Once triggered, the motion persists after the noise is removed [49].
Figure 4. Noise-controlled ignition length. (a) Ignition probability versus chain length for independent complex Gaussian white noise applied to each field component, with observation time . Ignition is recorded when . Each point represents 128 realizations, with shading indicating 95% confidence intervals. (b) Half-ignition length versus noise strength for the observation times shown. The line follows Eq. (13), with only the offset fitted to the data.
Directional amplification controls this strong length dependence. Fluctuations generated near the left edge of chain A and the right edge of chain B experience the longest amplified passage and dominate the opposite-edge variance [49]. To leading exponential order, the critical circulating intensity scales as , while the endpoint noise variance scales as . Their ratio scales as .
Let denote the chain length at which half the realizations ignite within . At fixed , we estimate the ignition probability primarily from this ratio. The half-ignition condition then corresponds approximately to a fixed value of . Together with , this gives
Figure 4(b) confirms the predicted logarithmic dependence, with only the offset fitted. Changing shifts the crossover much less than changing the noise strength. The same directional gain that lowers the deterministic threshold also amplifies noise, so reducing the noise increases the characteristic ignition length only logarithmically.
Discussion.—We have shown that NHSE-induced amplification and nonlinear feedback can produce a bypass transition to self-sustained motion. The triggering threshold falls exponentially with system size, while a return-map theory captures both the threshold and the near-critical dynamics. The transition concerns the route into sustained motion rather than the final attractor, which may be chaotic or periodic [49]. Notably, the triggering process is tractable in the skin setting: for the specified local excitation, the spatially organized amplification and nonlinear return can be calculated explicitly, without nonlinear optimization or edge tracking [54, 55]. More broadly, our results reveal skin amplification not merely as a transient response, but as a dynamical resource that nonlinear feedback can repeatedly harness.
The mechanism involves transient directional amplification, nonlinear regeneration, and an energy source that balances dissipation. These ingredients are structural rather than specific to the minimal two-chain model: the same mechanism survives, e.g., with unequal propagation channels, dimerized lattices, additional Kerr nonlinearities, and different nonlinear orders [49]. This points to a broader class of nonreciprocal and nonnormal systems in which transient amplification can be converted into autonomous dynamics. The predicted thresholds and scaling laws could be tested in active electrical networks, where nonreciprocal couplings [11, 56] and nonlinear circuit elements [57] are available. For active non-Hermitian devices, these results provide guidance for stability and nonlinear performance, and may also inform the design of switches and locally triggered oscillators.
Acknowledgments
This work is supported by the National Key Research and Development Program of China (Grants No. 2023YFA1406704 and No. 2022YFA1405800) and National Natural Science Foundation of China (Grant No. 12474496 and No. 12547107).
References
Supplemental Material for ``Bypass Transition via the Non-Hermitian Skin Effect''
This Supplement specifies the numerical methods, derives the return prediction and its length dependence, and connects it to the critical circulating orbit. It then examines the power source and decay controls, noise-induced ignition, periodic sustained states, dimerized chains, and residual linear coupling.
Preparation and numerical methods
We prepare with all other fields zero and let the lattice evolve without further driving. Unless stated otherwise, , , and . Along this family of initial states, denotes the first threshold between decay and sustained motion, which may be chaotic or regular. We use for the threshold predicted by the return condition.
The real single-site seed belongs to the fixed subspace of the antilinear symmetry
Within this subspace, is real on odd sites and imaginary on even sites, with the opposite pattern for . Hopping and conversion connect fields of opposite real–imaginary type, so the factor makes each time derivative match the type of its field. Uniform loss preserves this pattern as well, so a trajectory starting in the subspace remains there. We use this same single-site preparation for thresholds, power statistics, and periodic states, while testing stability against perturbations both within and outside the subspace.
Deterministic calculations. Deterministic threshold scans use fourth-order Runge–Kutta integration with and bisection between decay and sustained motion. Runs initially extend to . We sample the total power once per unit time and average it in blocks of duration 100 over the final units. A trajectory is classified as decaying if every block mean is below and as sustained if every block mean exceeds 0.05. Otherwise, the integration time is doubled before assigning an outcome. Bisection reduces the relative threshold bracket to below . For the near-threshold departure times, we refine it to below , well below the smallest relative seed displacement used. Halving the step and doubling the duration preserve the outcomes just below and above threshold.
Response and stability calculations. We integrate the response equations with the adaptive eighth-order Runge–Kutta method DOP853, using relative and absolute tolerances of and and sampling endpoint fields every 0.02 time units. Floquet multipliers and Lyapunov spectra [58] are obtained from the full-field variational equations, as described in § 0.7 and S4.
From propagation to the triggering threshold
The initial field and its linear propagation. Writing the site amplitudes as and , the equations of the Letter become
Here contains the local intensities. The dot denotes multiplication at each site, . Inner products are written as . The matrix describes hopping and loss in chain , and is the identity. The hopping matrix has entries and , with all other entries zero. The opposite hopping direction in chain gives . We label site of chain () by (). The single-site preparation is and , where . Its total power is , with .
The conversion vanishes at linear order about the resting state, leaving two independently propagating fields. Their propagators are , so that and . The matrix element is the field at site and time produced by a unit initial field at site in chain . For the single-site preparation, this gives , while chain remains empty.
To evaluate these propagators, we use the imaginary gauge transformation and , where and . Since , with , both transformed chains have symmetric hopping, . Here is the open chain with unit nearest-neighbor hopping in both directions. Its standing waves are proportional to , giving the eigenvalues of and as [1, 3]
These growth rates satisfy , where are the mode frequencies in the Letter. Expanding in these standing waves and restoring the gauge factors gives
Replacing by gives . The factor favors propagation toward increasing in chain , while its inverse favors the opposite direction in chain . Both propagators also contain the uniform decay factor .
The amplified field becomes a source in the other chain. Conversion couples these propagating fields by supplying a local source to each chain. Applying the variation-of-constants formula to Eq. (15), with and , gives the exact relations
The integrals propagate each source from its generation time to the observation time . Since the initial field in chain is of order and its intensity is of order , the source first generates a field of order in chain . Feeding this field back into chain gives a correction of order . We therefore expand
The superscripts label orders in the seed amplitude, not powers of the fields. Substituting this expansion into Eq. (15) gives the leading response equations
The first equation propagates the unit seed, . Substituting it into the source for chain gives
The return is the coherent sum of the fields generated along chain and propagated back through chain by . For a seed of amplitude , the leading return is . The conversion strength is already included in .
Closing the return. Let be the first arrival peak of the unit-seed response at , and the first return peak of the cubic response at ,
The windows and select the first arriving and returning packets, respectively. The endpoint coefficient includes sources throughout the chain, as specified by Eq. (23).
The seed produces an arrival peak and a return peak . With , the return is related to the arrival by . We extend this relation to successive passages by approximating the arriving packets as having the same normalized shape. Reflection and exchange of the two identical chains then give the same coefficient for both directions. Two successive conversions give
Here is the next arrival at , after the field has returned through chain and traveled along chain again. A packet that repeats without growing or shrinking satisfies . Its nonzero peak and the seed needed to reach it are therefore
Since , peaks below decrease on successive circuits and those above it grow. Equivalently, is the seed for which , so the returning peak equals the arriving peak. The response coefficients and thus determine the predicted threshold without a nonlinear threshold search. Its circulating amplitude and instability are compared with the full-lattice periodic orbit in § 0.7.
In the numerical response calculation, we take and . The reference transit time is the end-to-end distance divided by the maximum group velocity of the reciprocal chain obtained by the imaginary gauge transformation.
Gain scaling and comparison with simulations. The arriving field has amplitude , so it supplies a cubic source of size . Propagation back along the mirror-related chain contributes one further gain factor , giving and hence . The linear dependence on is exact in the cubic response, while estimates the leading amplification. The long-chain gain, derived below by maximizing the linear response, is
When conversion is dominated by the ends and source contributions do not cancel at leading exponential order, the four gain factors give
Here denotes corrections that grow more slowly than , including the logarithms of algebraic spreading factors. They account for the difference between finite-length slopes and the asymptotic slope . Unlike the length asymptote, the dependence on is exact at every . Multiplying all fields by reduces the full equations to those with , so thresholds scale as and corresponding powers as .
The corrections to the gain law can be collected in a single factor,
This rewrites Eq. (26) exactly, with calculated from the responses rather than fitted to the threshold. It accounts for packet shape and conversion over space and time.
The same argument applies to conversion of field order , . An arriving field of amplitude generates a source of order . Return propagation gives , with to leading exponential order. For identical chains, the circulating peak satisfies , so
Within the same packet-shape approximation, the conversion order therefore sets the power of the gain, while propagation sets its exponential dependence on length.
At the reference parameters, the thresholds span eleven orders of magnitude up to , with local length slopes approaching from below [Fig. S1(a,b)]. Across the eleven parameter pairs in Figure S1, the largest prediction errors occur at small . This includes short chains and parameters near the amplification boundary . Stronger gain lowers the required seed and improves the first-passage approximation.
Figure S1. Single-site threshold prediction and gain scaling. (a) Nonlinear thresholds and endpoint predictions at the reference parameters. (b) Local length slopes, with asymptote . Symbols and solid curves in (a,b) denote nonlinear calculations and response predictions. (c) Signed relative prediction error, . (d) Unequal-chain tests of Eq. (31), at . The diagonal denotes equality. (e) Nonlinear thresholds versus gain . The dashed line indicates the slope. (f) Response factor from Eq. (29). The same colors identify three parameter pairs in (c,e,f). Gray points in (c,e) show the other eight pairs listed in Table S1.
Figure S1 compares the thresholds with the leading gain dependence across all eleven parameter pairs. The response factor varies more slowly than the exponential gain [Fig. S1(f)], as required for the leading exponential law. Table S1 lists its finite-length ranges together with the prediction errors.
| maximum error (%) | range of | |||
| 0.1 | 0.05 | 12–80 | 6.1 | 0.57–0.70 |
| 0.2 | 0.05 | 12–64 | 4.0 | 0.64–0.77 |
| 0.2 | 0.1 | 12–128 | 2.1 | 0.56–0.77 |
| 0.2 | 0.15 | 12–64 | 1.6 | 0.61–0.76 |
| 0.2 | 0.2 | 12–64 | 2.5 | 0.60–0.76 |
| 0.2 | 0.3 | 20–80 | 14.3 | 0.54–0.70 |
| 0.3 | 0.1 | 16–64 | 1.3 | 0.64–0.76 |
| 0.3 | 0.2 | 12–80 | 1.6 | 0.59–0.77 |
| 0.3 | 0.3 | 16–64 | 1.7 | 0.59–0.74 |
| 0.3 | 0.4 | 12–80 | 5.8 | 0.54–0.76 |
| 0.4 | 0.2 | 16–64 | 1.4 | 0.61–0.75 |
Table S1. Parameter coverage of the single-endpoint prediction. is calculated from the response coefficients using Eq. (29). The error column is the largest absolute relative error within the listed length range.
Unequal chains. If the chains have different hopping asymmetries or losses , their propagation directions are still opposite but the two return coefficients need not be equal. The response equations determine each coefficient separately. A unit seed at gives an arrival at and a return at , defining . For the reverse passage, a unit seed at gives an arrival at and a return at , defining . These auxiliary unit-seed responses are used to predict the threshold for the original preparation, with all other fields zero. Applying the two return relations in sequence gives
The circulating peaks and are obtained by imposing and . The predicted threshold is again the seed whose first arrival reaches , now . Starting from the reference parameters, we either reduce to or increase to , and test . The resting-state decay margin remains independently of length. The predictions differ from the numerical thresholds by at most [Fig. S1(d)], with threshold brackets checked as in Sec. S1. For identical chains, and , recovering the symmetric-chain result .
The long-chain gain exponent. Equation (S4) gives the gain at finite length. Its large-length exponent is most easily obtained from the corresponding infinite-chain propagator , where is the distance from a unit seed at site to the observation site . In the gauge-transformed chain, a single-site seed has equal Fourier weight at all wave numbers, whose dispersion is . Adding the propagation factors and restoring the gauge factor and loss gives
At fixed distance , the directional factor is constant, so the peak occurs where reciprocal propagation and loss combine to give the largest response. At and after the reciprocal chain’s ballistic arrival, , the unit modulus of the integrand on the real axis gives
The earlier tail suffers less loss and can exceed this late-time bound. For and large , contributions from different wave numbers largely cancel as their phases vary rapidly. The remaining response can be evaluated by steepest descent [59, 60]. Writing the exponent as , its stationary points satisfy
Before the ballistic arrival, this equation has no real solution. The early-time tail is governed by the complex saddle
The integration path can be shifted upward by to pass through this saddle. No singularities are crossed, and the vertical sides of the connecting contour cancel because the integrand is -periodic for integer . Cauchy’s theorem therefore leaves the integral unchanged. On the shifted path , with real , the integrand’s modulus falls away from by the factor . Near the saddle,
For large at fixed , the contributing region has width of order . Across this narrow region, the phase changes only by order , so the contributions add nearly in phase. The rest of the path is exponentially suppressed, so this Gaussian neighborhood dominates the integral. Its width supplies an algebraic prefactor, while the leading exponential dependence is set by , so
The prefactor contributes only to the term. The peak follows by maximizing Eq. (33) over time at fixed . Although depends on , the terms containing cancel because , leaving
The derivative decreases from to , giving a unique maximum at
At the peak, , so the exponent reduces to . For , , making this exponent larger than the late-time upper bound. The early peak therefore gives the global maximum at leading exponential order, . This recovers in Eq. (27). At finite distance, prefactors can shift the peak time without changing its asymptotic exponent.
To impose open boundaries, we add image sources to the gauge-transformed propagator so that the field vanishes at the two fictitious sites and . The leading terms near the first arrival have propagation distances , , and , which differ only by fixed offsets and therefore share the same large- exponent. Their sum retains this exponent at fixed , while contributions from additional traversals are exponentially smaller at these times. The boundaries thus change the prefactor, not . For the finite-length value of , we use Eq. (17), which retains these boundary effects and packet spreading.
The exponent also appears in the fixed-frequency response. For a real frequency , let be the complex amplitude ratio between neighboring sites. Substituting into the linear bulk equation gives
At the temporal peak found above, the saddle satisfies . Restoring the gauge factor gives , the smaller-magnitude root at . Thus maximizing the pulse response selects the same spatial exponent as the zero-frequency response [19] [18].
The critical circulating orbit
A full-field solution and its instability. The nonzero fixed point of the return map describes a packet that reproduces its peak after each circuit. To find the corresponding orbit of the full lattice, let denote the field at one instant and its value after evolving Eq. (15) for a time . An orbit that repeats up to a common phase satisfies
A near-threshold trajectory provides an initial guess for this orbit. We take a reference field from one arrival and estimate from the interval between successive, nearly identical arrivals. Evaluating Eq. (15) at gives its time derivative . Small changes of time origin and common phase move this field along and . We fix the two freedoms by making orthogonal to these directions in real field coordinates,
The second expression uses . These conditions choose a representative with a fixed time origin and common phase, without changing the physical orbit. Newton shooting adjusts all real field components, , and to solve them together with Eq. (35). The converged is zero on the branch studied here, so the full field repeats after , not just its intensity.
To determine stability, we write the real and imaginary field components as , with equations . Along the periodic solution , small deviations obey , where . The matrix maps an initial deviation to its later value, . It starts from the identity, since at it maps each deviation to itself, and obeys
where is the identity matrix. The one-period matrix applies on every circuit because the orbit and repeat after . Its eigenvalues are the Floquet multipliers. For , a deviation along is multiplied by each circuit, shrinking if and growing if .
The equations are unchanged by shifting the time origin or rotating all fields by a common phase. For the periodic field , these operations produce small deviations proportional to and , respectively. Both repeat after and give unit multipliers. They shift the timing or common phase of the same motion, so stability is determined by the remaining modes. The computed spectrum has one growing mode, with , while all other nonneutral modes decay [Fig. S2(a)].
Initial fields that approach the orbit, up to time and common phase shifts, form its stable manifold. The single growing direction makes this a surface with one fewer dimension than the full initial-condition space. Varying in , with all other fields initially zero, traces a line through this space, and marks its crossing of the manifold. In full-lattice simulations, seeds just below and above approach the orbit before decaying or igniting, respectively. These trajectories identify the manifold as the local boundary between the two outcomes.
The peak map in Eq. (25) predicts growth over a complete circuit with two conversions. Differentiating at its circulating peak gives
A small peak deviation obeys . Along the full-field unstable mode, the peak deviation grows by a factor over the same circuit. The two calculations can therefore be compared directly, with nine as the return map’s approximate prediction and obtained independently from the full dynamics. The coefficient cancels, leaving a prediction fixed by the conversion order. For the order- conversion of Eq. (30), two successive returns give . Using , the derivative at its nonzero fixed point is . The full-lattice multipliers approach these values for cubic and quintic conversion as length increases [Fig. S2(b)].
Figure S2. Critical circulation and the response prediction. (a) Complete Floquet magnitudes at . (b) Unstable multiplier divided by the predicted value for cubic and quintic conversion. (c) Predicted circulating endpoint amplitude divided by the independently computed orbit amplitude.
The amplitude and multiplier test different aspects of the return map. The orbit’s endpoint peaks remain close to the predicted [Fig. S2(c)], whose coefficient depends on packet shape and distributed conversion. This dependence cancels from the derivative at the fixed point. The multiplier can therefore approach nine even when the peak-amplitude ratio remains different from unity and the circulating packet differs in shape from the first arriving packet.
Why the departure time grows logarithmically. For a seed , measures the relative distance from threshold. After the trajectory approaches the orbit, the decaying perturbation modes diminish and the growing mode determines when it leaves. The single-site family crosses the boundary with a growing-mode coefficient , where measures its sensitivity to the seed amplitude. After circuits its magnitude is
If departure is recorded at a fixed deviation , the condition gives . Since each circuit takes time , the departure time satisfies
The offset includes the initial approach and the chosen departure criterion. Recording departure only at discrete arrival peaks gives the bounded correction. In the Letter, departure is the first peak of after outside the interval from one-half to twice the orbit peak intensity. Changing this criterion shifts the offset but not the leading logarithmic slope . The long delay therefore reflects extra circuits near the unstable orbit, with the slope fixed by its independently computed multiplier.
Sustained motion and its power source
Variational dynamics and the chaotic state. The critical orbit describes the transient near threshold. The long-time state is instead characterized by its Lyapunov spectrum. For infinitesimal deviations and from a trajectory , linearizing Eq. (15) gives
Here is the intensity variation. Because it contains complex conjugates, we solve these linear equations in the real and imaginary field components. A basis of perturbations is evolved alongside the trajectory and periodically orthonormalized by QR decomposition. The accumulated logarithms of the stretching factors, divided by elapsed time, give the Lyapunov exponents . After a transient of 5000 time units and a further 1000 for the perturbation basis to align, we accumulate the spectrum over 6000 time units, with QR steps every 0.2 and integration step 0.02. This auxiliary calculation leaves the underlying single-site trajectory unchanged. Positive exponents show that nearby trajectories separate exponentially and confirm chaos at [Fig. S3(a)].
Hopping and conversion preserve phase-space volume, while loss contracts each real coordinate at rate . Their total contraction fixes , providing a check on the spectrum.
Figure S3. Chaotic motion and its power from single-site seeds. (a) Full Lyapunov spectra at . (b) Largest exponents within and transverse to the exact preparation subspace. (c) Long-time power at three representative amplitudes. Error bars are standard errors of time-block averages.
At the reference parameters, positive exponents occur for perturbations both within and outside the real–imaginary preparation subspace of Sec. S1 [Fig. S3(b)]. The motion is chaotic within this subspace and also has growing transverse perturbations. Across six seed amplitudes spanning and seven lengths from to , the mean power varies little with seed amplitude and grows approximately linearly with length [Fig. S3(c)]. We average over , using blocks of duration to estimate statistical error. Twenty longer runs at remain active to , with positive largest Lyapunov exponents in the final measurement windows.
Power balance and exact decay controls. The source of this sustained motion can be identified from its power balance. Conversion at any real local rate makes equal and opposite contributions to the two intensities,
Conversion therefore conserves total power . Adding hopping and loss gives
The term is the power supplied by active nonreciprocal hopping. In statistically stationary motion, this supply balances loss, . Hopping therefore powers the motion after the seed has initiated it, while conversion returns the field to an amplifying channel.
Sufficiently strong loss prevents sustained motion. Since conversion conserves total power, differentiating leaves only the hopping and loss matrices , with ,
Thus instantaneous power growth is governed by the Hermitian matrices . For chain , has diagonal entries and nearest-neighbor entries and . A site-dependent phase change makes both hopping entries real and equal to , without changing the eigenvalues. Its spectrum is therefore that of a reciprocal open chain shifted by . Chain has the same spectrum, with largest eigenvalue . For any field , , so
The coefficient on the right is strictly negative for every finite chain when , so all fields decay regardless of their initial amplitude. For reciprocal hopping, , the hopping contribution vanishes and exactly. The corresponding bulk field-growth rate is the largest imaginary part of the periodic-boundary dispersions . Positive bulk growth coincides with , allowing directional amplification despite open-boundary spectral stability [17, 27].
With weaker loss, amplification is possible, but gain alone need not sustain motion. To test this, we orient both chains the same way, setting . They share the positive weight matrix , with and . Hopping therefore conserves the weighted power, as does conversion between equally weighted sites. Hence
Positive weights ensure that every finite initial field decays. Reflection of one uncoupled chain is unitary, so this control retains the same resting-state spectrum and maximum linear transient gain in total power. Yet conversion no longer feeds an amplifying return path.
A linear return path with onsite Kerr nonlinearity. The return path can also be made linear by replacing with a constant . Adding onsite Kerr terms then gives a nonlinear lattice with a different feedback mechanism,
Here , is real, and is the real Kerr coefficient. Kerr changes local phases without directly changing intensities, while transfers intensity between chains.
To establish decay at weak coupling, write and , with . This centered gauge makes both intrachain hopping matrices Hermitian. The weighted power is conserved by hopping and Kerr phase shifts. Linear conversion does not cancel under the opposite weights, giving
The inequality follows from and . For the diagonal matrix , this norm is its largest absolute diagonal entry. Loss dominates for , forcing every initial field to decay for any real .
For positive Kerr coefficients, the rescaling reduces the equations to , which we use below. We also test linearly stable cases beyond the sufficient decay bound at , with – and the reentrant stable region –. All tested single-site seeds between and decay. The sustained states found in this control appear after the resting state becomes linearly unstable. They are coherent at and chaotic at , and are reached from seeds down to . Smaller seeds delay onset by , where is the linear growth rate. In both chaotic cases, the long-time mean power shows no systematic growth in time and varies little with seed amplitude at fixed parameters. The sustained motion found in this control thus follows linear instability, rather than bypassing it.
Conversion with additional onsite phase shifts. To test Kerr phase shifts without replacing the nonlinear return path, we retain and add and , with . At the reference parameters and , shift the threshold by less than and pass the step and duration checks of Sec. S1. Both give sustained chaos at , with positive largest Lyapunov exponents. The leading return response is unchanged because Kerr first corrects the field in chain at order . In the conversion source , this correction is multiplied by two fields of order , giving a change only at order . The source that generates the leading response in chain is therefore unaffected.
Noise-induced ignition
Noise and ignition protocol. Noise can provide the initial disturbance without a prepared seed. Starting from the resting state, we add independent random fields at every site in both chains. After each deterministic RK4 step of duration , every complex amplitude receives the increment
where is the noise strength and the are independent Gaussian variables with zero mean and unit variance, drawn anew for every field and time step. The real and imaginary parts therefore have equal noise strength. In continuous time, this prescription is written as
Here contains the field amplitudes and is the deterministic rate of change given by Eq. (15). The complex Wiener increment has zero mean, and brackets denote averaging over noise realizations. Its first correlation specifies the variance and independence between fields. The second gives equal, uncorrelated real and imaginary parts. Increments over disjoint time intervals are independent as well.
We monitor ignition during an observation window , with noise applied continuously until ignition or the end of the window. Ignition is recorded when the total power first exceeds , checking once per unit time. In the tested range, this level lies above the critical-orbit power and below the sustained-state power. We evaluate ignition probabilities at , using 128 independent realizations at each sampled length and noise strength. Runs that have not ignited are followed to the corresponding , including the longest window at . We use binomial Wilson confidence intervals and interpolate between adjacent lengths to obtain , where half the realizations ignite.
We check time-step convergence by halving from to using the same 128 noise realizations. The ignition probabilities and half-ignition lengths agree within sampling uncertainty at all three observation times.
A separate test checks whether crossing the power threshold marks self-sustained chaos. We continue 48 crossing states at for 6000 further time units, once with noise left on and once with it switched off at the crossing. All retain positive largest Lyapunov exponents over the final 4000 units, and halving the step preserves these outcomes. One noise-free trajectory per length remains chaotic to . The motion therefore survives removal of the noise that triggered it.
Amplification of noise. Before ignition, the noise-induced field can be estimated from linear propagation, neglecting conversion. Consider noise injected at site of chain over a short interval around time . This increment has variance , and its field is multiplied by on reaching the right end at time . Here is the propagator defined in Sec. S2. Its contribution to the endpoint variance is consequently . Independent contributions from different sites and times add, giving
The subscript denotes linear evolution, for which the mean field is zero and the averaged intensity equals the variance. Since each increment eventually decays, the variance approaches a finite long-time value. The variance contains the full time-integrated response from every injection site, whereas measures the peak response from site 1. Writing for the elapsed propagation time separates the peak gain from the remaining integral,
The subscript denotes the long-time limit, and collects the response duration and the relative contributions of the injection sites.
For long chains with , noise near the left end dominates the right-end fluctuations because it traverses the longest amplified path. The propagation estimate in Sec. S2 gives . Pulse widths and the sum over injection sites change more slowly with length and contribute no additional exponential factor, so . Consequently,
The squared gain therefore sets the leading length dependence. Mirror symmetry gives the same result for , where fluctuations generated near the right end of chain undergo the longest amplified passage.
The logarithmic length scale. These amplified fluctuations must be compared with the circulating peak . The required intensity scales as , while the noise variance scales as . Their ratio therefore scales as , with a larger ratio requiring a rarer fluctuation to trigger motion.
We estimate the ignition probability at fixed as a function of this ratio, treating changes in the spatial structure, duration, and recurrence of fluctuations as subleading corrections. Within this scaling approximation, half-ignition corresponds to a fixed . Using from Eq. (28) then gives
The offset depends on the observation time and noise protocol. Packet spreading and fluctuation duration introduce subleading finite-length corrections to this leading logarithmic law. As in the deterministic problem, the rescaling sets the conversion strength to one. It also changes the noise strength to and the detection level to .
In the noise simulations, the fitted slopes of versus are 5.18, 5.20, and 5.14 sites per decade for , , , respectively. Their near equality across the observation windows and agreement with the leading prediction support the predicted logarithmic dependence.
Periodic motion and dimerized chains
A regular sustained state in the main model. As loss approaches the amplification boundary from below, the sustained motion in the uniform chains can become periodic. At and , the threshold is , and a single-site seed produces oscillating local intensities.
The field and local intensities repeat after . Varying the single-site seed amplitude over five values between and yields periodic motion with the same period and Floquet spectrum [Fig. S5(a,c)]. The one-period matrix , defined in § 0.7, has two neutral multipliers associated with time and global phase shifts. All others have modulus at most $0.8312, establishing stability also outside the preparation subspace. Halving the time step leaves these values unchanged at the displayed precision.
A nearest-neighbor dimerized extension. Replacing uniform hopping by alternating bonds lets us test the return condition in a two-band lattice. Each chain contains two-site cells, or physical sites. The fields are indexed by site, so cell contains sites and . The alternating bonds are
The local conversion remains , and only is initially nonzero. Figure S4(a) shows this lattice, and Fig. S4(b) illustrates decay and sustained motion on opposite sides of its triggering threshold.
With ,
where is the real symmetric Su–Schrieffer–Heeger hopping matrix [61]. Every resting-state mode therefore decays at exactly , for every . The two dimerizations are topological () and trivial () in this reciprocal representative. Both support triggering. The uniform case recovers the main model with .
Threshold prediction and periodic output. Using the dimerized hopping matrices in Eqs. (21) and (22) gives the arrival at and return at . Their peak coefficients enter the same prediction , without fitted coefficients. We extract the first arrival and return over and , respectively, with reference time .
For and , the predictions agree with the numerical thresholds to within 4% at and 2.4% at [Fig. S4(c,d)]. The latter set includes . The threshold brackets and the two trajectories in Figure S4 retain their outcomes under the duration and step-size checks of Sec. S1.
Figure S4. Triggering in dimerized chains. (a) Alternating bonds have stronger hopping along the thick arrows and weaker hopping in reverse. Green rungs denote local intensity-dependent conversion. (b) Site intensity below and above threshold at , , and , with a common logarithmic color scale. (c,d) Single-site thresholds (symbols) and endpoint predictions (lines) for two losses. Filled diamonds and open triangles distinguish the two dimerizations. Here , , and each chain has cells or sites.
The dimerized chains also support regular motion near their amplification boundary. For both dimerizations, the periodic-boundary spectrum of Eq. (50) has maximum imaginary part approximately 0.286 before loss is added. Increasing loss toward this value can produce stable periodic motion, as in the uniform chains.
For , , and , the numerical threshold is . A single-site seed reaches
The field changes sign after , giving an intensity period , whereas for the uniform example. The one-period matrix has two neutral multipliers and all other moduli at most 0.99774. Varying the single-site seed amplitude over five values between and leads to the same periodic motion. The one-period matrices for both examples were also checked against finite differences and a halved integration step. Both periodic states therefore attract nearby trajectories, including those outside the preparation subspace Figure S5,Figure S5). This distinguishes them from the critical orbit at threshold, whose growing mode drives nearby trajectories away.
Figure S5. Stable periodic motion from a single-site seed . (a,b) Local intensities over five intensity periods after transients, with the time origin reset. Subscripts label physical sites, with sites per chain in (b). Here in (a) and in (b), where is the field period. (c,d) Floquet multiplier magnitudes over . All lie below unity except the two symmetry-neutral multipliers, confirming stability of the periodic states. Insets enlarge the leading multipliers. The uniform example has , , and the dimerized example has , , .
Residual linear coupling
The stable length window. Adding a real constant coupling gives . Unlike the intensity-dependent conversion, remains in the linearization about the resting state. The linear equation is , with and
This linear return path makes the spectrum length dependent, as in the critical NHSE [62, 63]. Stability is first lost when a linear mode stops decaying. Write this mode as , where is real and are its spatial field profiles on the two chains. Substitution gives
Each chain is driven by the field in the other. With response matrix , the two relations become and . Substituting the second relation into the first gives . A nonzero field can satisfy this equation only if is singular, hence
The field must therefore return with both its amplitude and phase reproduced. The first instability occurs at the smallest satisfying this condition for real , denoted by . Direct diagonalization confirms the crossing at for the reference even-length chains.
For , each directed leg has a zero-frequency gain scaling as (Sec. S2). Two propagation legs and two conversions therefore give a loop gain scaling as . Setting it to unity gives
Here is a prefactor set by the channel parameters. At fixed residual coupling, defines the length at which the resting state first loses stability.
The return below the linear crossing. Below , the resting state remains linearly stable. For weak residual coupling, we approximate the field in chain during its first passage by , where is the uncoupled chain’s linear response to a unit excitation at , defined in Sec. S2. The constant coupling supplies chain with the source . Factoring out the seed amplitude and coupling strength, we write the generated field as , where the response obeys
where describes propagation and loss in chain .
Cubic conversion generates the additional field from Sec. S2. At this order, both fields propagate linearly in chain and add coherently, . Equating the peak magnitude of this sum at to the arrival peak at , and dividing by , gives
Here selects the first return. Solving for predicts the shifted threshold and recovers at . Figure S6 compares this prediction with full dynamics. For small , the threshold changes approximately linearly with signed , decreasing for positive coupling and increasing for negative coupling as the two return fields reinforce or oppose each other. Near the linear crossing, repeated linear circulation changes the wave packet, and the numerical thresholds depart from the first-return prediction.
Figure S6. Triggering below the residual-coupling instability. (a) Threshold normalized to its value at . Circles, squares, and triangles denote , , , respectively. Red and blue distinguish positive and negative coupling. Dotted lines indicate the associated critical lengths. (b) The same data versus . Solid curves use the coherent first-return approximation, Eq. (56).
Beyond the first crossing, the resting state is linearly unstable, so the threshold no longer separates decay from sustained motion. At the reference parameters, we sampled , 32, 40 and couplings in the range . Positive coupling leads to chaos from very small single-site seeds in these tests. For negative coupling, small seeds reach low-power nonzero motion and larger seeds can reach chaos.
References
- [1]N. Hatano and D. R. Nelson, Localization transitions in non-Hermitian quantum mechanics, Phys. Rev. Lett. 77, 570 (1996).DOI
- [2]S. Yao and Z. Wang, Edge states and topological invariants of non-Hermitian systems, Phys. Rev. Lett. 121, 086803 (2018).DOI
- [3]K. Yokomizo and S. Murakami, Non-Bloch band theory of non-Hermitian systems, Phys. Rev. Lett. 123, 066404 (2019).DOI
- [4]N. Okuma, K. Kawabata, K. Shiozaki, and M. Sato, Topological origin of non-Hermitian skin effects, Phys. Rev. Lett. 124, 086801 (2020).DOI
- [5]K. Zhang, Z. Yang, and C. Fang, Correspondence between winding numbers and skin modes in non-Hermitian systems, Phys. Rev. Lett. 125, 126402 (2020).DOI
- [6]C. H. Lee and R. Thomale, Anatomy of skin modes and topology in non-Hermitian systems, Phys. Rev. B 99, 201103(R) (2019).DOI
- [7]V. M. Martinez Alvarez, J. E. Barrios Vargas, and L. E. F. Foa Torres, Non-Hermitian robust edge states in one dimension: Anomalous localization and eigenspace condensation at exceptional points, Phys. Rev. B 97, 121401(R) (2018).DOI
- [8]H. Hu, Topological origin of non-Hermitian skin effect in higher dimensions and uniform spectra, Sci. Bull. 70, 51 (2025).DOI
- [9]E. J. Bergholtz, J. C. Budich, and F. K. Kunst, Exceptional topology of non-Hermitian systems, Rev. Mod. Phys. 93, 015005 (2021).DOI
- [10]S. Weidemann, M. Kremer, T. Helbig, T. Hofmann, A. Stegmaier, M. Greiter, R. Thomale, and A. Szameit, Topological funneling of light, Science 368, 311 (2020).DOI
- [11]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 topolectrical circuits, Nat. Phys. 16, 747 (2020).DOI
- [12]L. Xiao, T. Deng, K. Wang, G. Zhu, Z. Wang, W. Yi, and P. Xue, Non-Hermitian bulk–boundary correspondence in quantum dynamics, Nat. Phys. 16, 761 (2020).DOI
- [13]A. Ghatak, M. Brandenbourger, J. van Wezel, and C. Coulais, Observation of non-Hermitian topology and its bulk–edge correspondence in an active mechanical metamaterial, Proc. Natl. Acad. Sci. U.S.A. 117, 29561 (2020).DOI
- [14]L. Zhang, Y. Yang, Y. Ge, Y.-J. Guan, Q. Chen, Q. Yan, F. Chen, R. Xi, Y. Li, D. Jia, S.-Q. Yuan, H.-X. Sun, H. Chen, and B. Zhang, Acoustic non-Hermitian skin effect from twisted winding topology, Nat. Commun. 12, 6297 (2021).DOI
- [15]Q. Liang, D. Xie, Z. Dong, H. Li, H. Li, B. Gadway, W. Yi, and B. Yan, Dynamic signatures of non-Hermitian skin effect and topology in ultracold atoms, Phys. Rev. Lett. 129, 070401 (2022).DOI
- [16]H. Gao, W. Zhu, H. Xue, G. Ma, and Z. Su, Controlling acoustic non-Hermitian skin effect via synthetic magnetic fields, Appl. Phys. Rev. 11, 031410 (2024).DOI
- [17]C. C. Wanjura, M. Brunelli, and A. Nunnenkamp, Topological framework for directional amplification in driven-dissipative cavity arrays, Nat. Commun. 11, 3149 (2020).DOI
- [19]H. Li and S. Wan, Exact formulas of the end-to-end Green’s functions in non-Hermitian systems, Phys. Rev. B 105, 045122 (2022).DOI
- [20]Q. Wang, C. Zhu, Y. Wang, B. Zhang, and Y. D. Chong, Amplification of quantum signals by the non-Hermitian skin effect, Phys. Rev. B 106, 024301 (2022).DOI
- [21]Y. Li, C. Zhang, and C. Qiu, Observation of Aharonov–Bohm-caged directional amplification in acoustic lattices, Adv. Funct. Mater. 36, e20857 (2026).DOI
- [22]McDonald, A., Percegona, T., Carnea, A., & Clerk, A. A. (2018). Phase-dependent chiral transport and effective non-Hermitian dynamics in a bosonic Kitaev-Majorana chain. Phys. Rev. X, 8, 041031.
- [23]D. Porras and S. Fernández-Lorenzo, Topological amplification in photonic lattices, Phys. Rev. Lett. 122, 143901 (2019).DOI
- [24]J. C. Budich and E. J. Bergholtz, Non-Hermitian topological sensors, Phys. Rev. Lett. 125, 180403 (2020).DOI
- [25]M. Parto, C. Leefmans, J. Williams, R. M. Gray, and A. Marandi, Enhanced sensitivity via non-Hermitian topology, Light Sci. Appl. 14, 6 (2025).DOI
- [26]H. Xu, U. Delić, G. Wang, C. Li, P. Cappellaro, and J. Li, Exponentially enhanced non-Hermitian cooling, Phys. Rev. Lett. 132, 110402 (2024).DOI
- [27]B. Midya, Topological directed amplification, Phys. Rev. A 106, 053513 (2022).DOI
- [28]Trefethen, L. N., & Embree, M. (2005). Spectra and Pseudospectra: The Behavior of Nonnormal Matrices and Operators. In Princeton University Press.
- [29]K. G. Makris, L. Ge, and H. E. Türeci, Anomalous transient amplification of waves in non-normal photonic media, Phys. Rev. X 4, 041014 (2014).arxiv.org/abs/1410.4626
- [30]Y. Nakai, N. Okuma, D. Nakamura, K. Shimomura, and M. Sato, Topological enhancement of nonnormality in non-Hermitian skin effects, Phys. Rev. B 109, 144203 (2024).DOI
- [31]P. J. Schmid, Nonmodal stability theory, Annu. Rev. Fluid Mech. 39, 129 (2007).DOI
- [32]C. Yuce, Nonlinear non-Hermitian skin effect, Phys. Lett. A 408, 127484 (2021).DOI
- [33]B. Many Manda, R. Carretero-González, P. G. Kevrekidis, and V. Achilleos, Skin modes in a nonlinear Hatano–Nelson model, Phys. Rev. B 109, 094308 (2024).DOI
- [34]Wang, S., Wang, B., Liu, C., Qin, L., Liu, W., Long, S., & Lu, P. (2025). Nonlinear non-Hermitian skin effect and skin solitons in temporal photonic feedforward lattices. Phys. Rev. Lett., 134, 243805.
- [35]E. T. Kokkinakis, I. Komis, and K. G. Makris, Self-trapping and skin solitons in two-dimensional non-Hermitian lattices, Commun. Phys. 9, 22 (2026).DOI
- [36]M. Ezawa, Dynamical nonlinear higher-order non-Hermitian skin effects and topological trap-skin phase, Phys. Rev. B 105, 125421 (2022).DOI
- [37]K. Kawabata and D. Nakamura, Hopf bifurcation of nonlinear non-Hermitian skin effect, Phys. Rev. Lett. 135, 126610 (2025).DOI
- [38]H. Lin, Y. Qi, and G.-L. Long, Global bifurcations and basin geometry of the nonlinear non-Hermitian skin effect, arXiv:2602.17439 (2026).arXiv
- [39]W. Zhang, F. Di, and X. Zhang, Non-Hermitian global synchronization, Adv. Sci. 12, 2408460 (2025).arxiv.org/abs/2501.14169
- [40]L. Zhao, B. Li, and N. Uchida, Nonreciprocal amplification toward chaos in a chain of Duffing oscillators, Phys. Rev. E 112, 014216 (2025).DOI
- [41]B. Zhu, Q. Wang, D. Leykam, H. Xue, Q. J. Wang, and Y. D. Chong, Anomalous single-mode lasing induced by nonlinearity and the non-Hermitian skin effect, Phys. Rev. Lett. 129, 013903 (2022).DOI
- [42]C. R. Leefmans, M. Parto, J. Williams, G. H. Y. Li, A. Dutt, F. Nori, and A. Marandi, Topological temporally mode-locked laser, Nat. Phys. 20, 852 (2024).arxiv.org/abs/2209.00762
- [43]L. N. Trefethen, A. E. Trefethen, S. C. Reddy, and T. A. Driscoll, Hydrodynamic stability without eigenvalues, Science 261, 578 (1993).pubmed.ncbi.nlm.nih.gov/17758167
- [44]T. Gebhardt and S. Grossmann, Chaos transition despite linear stability, Phys. Rev. E 50, 3705 (1994).pubmed.ncbi.nlm.nih.gov/9962424
- [45]J. M. Hamilton, J. Kim, and F. Waleffe, Regeneration mechanisms of near-wall turbulence structures, J. Fluid Mech. 287, 317 (1995).DOI
- [46]F. Waleffe, On a self-sustaining process in shear flows, Phys. Fluids 9, 883 (1997).DOI
- [47]S. Grossmann, The onset of shear flow turbulence, Rev. Mod. Phys. 72, 603 (2000).DOI
- [48]S. Longhi, D. Gatti, and G. Della Valle, Non-Hermitian transparency and one-way transport in low-dimensional lattices by an imaginary gauge field, Phys. Rev. B 92, 094204 (2015).DOI
- [49][bibliography entry not printed on this page]
- [50][bibliography entry not printed on this page]
- [51][51–53] [bibliography entries not printed on this page]
- [54]C. C. T. Pringle and R. R. Kerswell, Using nonlinear transient growth to construct the minimal seed for shear flow turbulence, Phys. Rev. Lett. 105, 154502 (2010).pubmed.ncbi.nlm.nih.gov/21230910
- [55]R. R. Kerswell, Nonlinear nonmodal stability theory, Annu. Rev. Fluid Mech. 50, 319 (2018).DOI
- [56]S. Liu, R. Shao, S. Ma, L. Zhang, O. You, H. Wu, Y. J. Xiang, T. J. Cui, and S. Zhang, Non-Hermitian skin effect in a non-Hermitian electrical circuit, Research 2021, 5608038 (2021).DOI
- [57]Y. Hadad, J. C. Soric, A. B. Khanikaev, and A. Alù, Self-induced topological protection in nonlinear circuit arrays, Nat. Electron. 1, 178 (2018).DOI
- [58]G. Benettin, L. Galgani, A. Giorgilli, and J.-M. Strelcyn, Lyapunov characteristic exponents for smooth dynamical systems and for Hamiltonian systems; a method for computing all of them. Part 2: Numerical application, Meccanica 15, 21 (1980).DOI
- [59]S. Longhi, Probing non-Hermitian skin effect and non-Bloch phase transitions, Phys. Rev. Research 1, 023013 (2019).DOI
- [60]W.-T. Xue, Y.-M. Hu, F. Song, and Z. Wang, Non-Bloch edge dynamics of non-Hermitian systems, Phys. Rev. Lett. 137, 096403 (2026).DOI
- [61]W. P. Su, J. R. Schrieffer, and A. J. Heeger, Solitons in polyacetylene, Phys. Rev. Lett. 42, 1698 (1979).DOI
- [62]L. Li, C. H. Lee, S. Mu, and J. Gong, Critical non-Hermitian skin effect, Nat. Commun. 11, 5491 (2020).DOI
- [63]Y. Yi and Z. Yang, Anomalous scaling behavior of Green’s function in critical skin effects, Phys. Rev. B 112, 174303 (2025).DOI








