Abstract
NonHermitian systems exhibit markedly different phenomena than their conventional Hermitian counterparts. Several such features, such as the nonHermitian skin effect, are only present in spatially extended systems. Potential applications of these effects in manymode systems however remains largely unexplored. Here, we study how unique features of nonHermitian lattice systems can be harnessed to improve Hamiltonian parameter estimation in a fully quantum setting. While the quintessential nonHermitian skin effect does not provide any distinct advantage, alternate effects yield dramatic enhancements. We show that certain asymmetric nonHermitian tightbinding models with a \({{\mathbb{Z}}}_{2}\) symmetry yield a pronounced sensing advantage: the quantum Fisher information per photon increases exponentially with system size. We find that these advantages persist in regimes where nonMarkovian and nonperturbative effects become important. Our setup is directly compatible with a variety of quantum optical and superconducting circuit platforms, and already yields strong enhancements with as few as three lattice sites.
Introduction
Quantum metrology and sensing aim to improve measurement precision over classical devices by exploiting uniquely quantum phenomena such as entanglement and squeezing^{1,2,3}. It is interesting to ask whether distinct effects associated with nonHermitian dynamics can also be used to improve sensors operating in quantum regimes^{4,5,6,7,8}. In purely classical settings, mode degeneracies specific to nonHermitian systems (socalled exceptional points) have been suggested as a means for enhanced parametric sensing^{9}. Evidence for enhancement has been demonstrated in several classicaldomain experiments involving small coupled mode systems (see e.g., refs. ^{10,11,12,13}). Theory suggests that particular kinds of nonHermitian effects could also be useful in truly quantum settings^{5}.
To date, both theory and experiment have focused on nonHermitian sensing schemes that utilize at most a few coupled modes. It is however well known that unusual new phenomena appear when considering genuinely multimode nonHermitian dynamics. The paradigmatic example is the socalled “nonHermitian skin effect”^{14,15,16,17}, which occurs in several nonHermitian tightbinding models^{18,19,20,21,22}. In these systems, all eigenvalues and wavefunctions of the Hamiltonian exhibit a dramatic sensitivity to a change of boundary conditions. This extreme sensitivity would seem to be a potentially powerful resource for parametric sensing^{23}.
In this work, we show that nonHermitian lattice dynamics does indeed provide a unique means for constructing enhanced sensors; moreover, this advantage persists even when operating in truly quantum regimes. We study in detail Hamiltonian parameter estimation using a onedimensional lattice model with asymmetric tunneling (akin to the wellstudied Hatano–Nelson model^{24}). We find, somewhat surprisingly, that the nonHermitian skin effect does not provide any advantage over more traditional sensing protocols. Rather, we find another distinct nonHermitian mechanism that enables a dramatic enhancement of measurement sensitivity: the quantum Fisher information per photon exhibits an exponential scaling with system size. We find that the exponential enhancement of measurement sensitivity persists even when considering limitations associated with the finite propagation time of a large lattice. Even for parameters large enough to invalidate a full linear response analysis, we find that our scheme provides a strong advantage: it achieves a squareroot enhancement of the sensitivity (including noise effects). As we discuss, the underlying mechanism makes use of both nonreciprocity and an unusual kind of symmetry breaking.
Results
Key features
While our ideas are general, our analysis focuses on a system that uses parametric driving to realize nonHermitian dynamics; this has the strong advantage of not requiring any external dissipation or postselection^{17,25}. Further, we ultimately focus on dispersive sensing, where the parameter of interest shifts the frequency of a resonant mode. This is a ubiquitous sensing strategy, with applications ranging from superconducting qubit measurement^{26} to virus detection^{27}. Our proposal is also compatible with a number of different experimental platforms in superconducting quantum circuits and quantum optics, and ultimately requires one to make a standard homodyne measurement. Finally, while our discussion focuses on large lattices, the results we present are already interesting in a small system consisting of just three coupled resonators.
Basic ingredients for a nonHermitian lattice sensor
A key feature that we will exploit in our new sensor is the dramatically large and unidirectional response exhibited by certain nonHermitian lattice models: perturbing a single lattice site induces a large change at one end of the chain, but not the other (see e.g., refs. ^{23,28}). We start by providing a physicallytransparent explanation of this effect, based on interpreting nonHermitian asymmetry in tightbinding matrix elements as directional gain and loss.
The simplest relevant system is the wellknown Hatano–Nelson model^{24,29}. This is a 1D tightbinding chain with asymmetric nearestneighbour hoppings, \({\hat{H}}=iJ{\sum }_{n}({e}^{A}n+ 1\rangle \langle n  {e}^{A} n\rangle \langle n+1)\), where J, A are real and \(\leftn\right\rangle\) is a position eigenket. The corresponding singleparticle Schrödinger equation is (ℏ = 1 throughout)
where ψ_{n} = 〈n∣ψ〉. While A formally plays the role of an imaginary vector potential, it is more usefully thought of as an amplification factor. Assuming A is positive for definiteness, Eq. (1) describes a system where a wavefunction’s amplitude grows by e^{A} every time a particle hops one site to the right, and decays an equal amount e^{−A} as it travels to the left, regardless of its energy.
With this picture in mind, the form of the realspace susceptibility (i.e., single particle Green’s function) χ(n, m; t) for a finite open chain has an intuitive form. Letting \(\leftm(t)\right\rangle ={e}^{i\hat{H}t}\leftm\right\rangle\), a simple calculation yields (see “Methods” section):
Here, χ_{0}(n, m; t) is the susceptibility matrix when A = 0, i.e., the Green’s function of a Hermitian tightbinding chain. This quantity is reciprocal, in the sense that χ_{0}(n, m; t) = (−1)^{m−n}χ_{0}(m, n; t) (i.e., apart from a phase, there is no asymmetry in rightwards versus leftwards propagation). The Green’s function χ_{0}(n, m; t) both describes how particles propagate in the lattice, and also the response properties of the system (i.e., if you perturb site m at t = 0, how does site n respond at some later time?).
The simple factorization in Eq. (2) makes it clear that there are two basic processes determining the response. The first is a distance and directiondependent amplification/deamplification factor, whereas the second encodes the dynamics of the underlying (A = 0) Hermitian tightbinding model. We thus have a simple intuitive picture for the susceptibility, without having to make recourse to other seemingly more complicated nonHermitian features, such as exceptional points, the nonHermitian skin effect, or the Petermann factor^{30,31}. Further, interpreting nonreciprocity as directional amplification provides a physical explanation as to why nonHermitian Hamiltonian are so sensitive to small perturbations^{16,18,32}. Note that Eq. (2) can be easily derived via a similarity transformation, which is analogous to the gauge transformation one would make if A were imaginary (and hence a real synthetic gauge field)^{24,29}.
While nonHermitian Hamiltonians are generically sensitive to small parameter changes, to obtain a true sensing enhancement requires the “right” kind of perturbation. To achieve this, we will exploit a second basic ingredient in constructing our sensor: symmetry breaking. The Hatano–Nelson chain breaks reciprocity for any A ≠ 0; formally, it picks a preferred amplification direction, and does not remain invariant (up to a local gauge change) under a spatial inversion operation \(\leftn\right\rangle \to \leftn\right\rangle\). We can trivially restore this symmetry by considering a system with two uncoupled Hatano–Nelson chains indexed by σ = ↑, ↓ with amplification factors A_{↑}, A_{↓}. If we pick A_{↑} = − A_{↓}, then the composite system restores some of the lost symmetry. Formally, the twochain system is invariant up to a local gauge change under the combined operations \(\leftn\right\rangle \to \leftn\right\rangle\) (spatial inversion) and \(\sigma \to \bar{\sigma }\) (pseudospin inversion). Physically, this means that the two chains exhibit chiral propagation and amplification in opposite directions. While this may seem trivial, this kind of discrete symmetry can persist even for certain forms of interchain coupling, and has recently been interpreted as a formal \({{\mathbb{Z}}}_{2}\) symmetry class with its own distinct nonHermitian topological phenomena^{32}. We discuss this symmetry more formally in “Methods” section.
For our purposes, the interesting feature here will be to consider breaking this symmetry with an external perturbation whose magnitude we wish to estimate. As we will see, the response to this symmetry breaking can be exponentially large in system size, enabling a new kind of sensor.
Model and measurement protocol
With the motivation of the previous section, we now consider a sensor comprised of two Hatano–Nelson chains with an opposite chirality (see Fig. 1a). There are a variety of means for such realizing nonHermitian directional tightbinding models using dissipation^{33,34}; approaches based on feedback control are also possible and have been recently implemented^{35}. However, for optimal sensing properties in quantum settings, methods that are both autonomous and avoid the noise associated with dissipation are desirable. We thus focus on a dissipationfree method for realizing nonHermitian dynamics based on parametric driving^{17,25}. We stress that the response properties of our sensor will be independent of how the nonHermitian dynamics is implemented, and hence apply equally well to dissipative and feedbackbased strategies.
We consider an Nsite chain of driven, coupled bosonic modes described by the fully Hermitian Hamiltonian
Here \({\hat{a}}_{j}\) is the photon annihilation operator on site j, w is the nearestneighbour hopping term, Δ is the nearestneighbour twophoton drive, and we consider open boundary conditions. We take both w and Δ to be positive and w > Δ. This model describes a 1D cavity array subject to parametric drives on each bond (described in a rotating frame set by the external pump frequency). As discussed extensively in ref. ^{17}, this system could be realized in both quantum superconducting circuits or nonlinear quantum optical systems. Note the lack of any onsite terms corresponds to the parametric driving frequency matching the resonance frequency of each isolated cavity.
Although not immediately obvious, the dynamics generated by \({\hat{H}}_{B}\) corresponds to two copies of the Hatano–Nelson model. In the basis of local canonical quadrature operators \({\hat{x}}_{j}\) and \({\hat{p}}_{j}\), defined via \({\hat{a}}_{j}=({\hat{x}}_{j}+i{\hat{p}}_{j})/\sqrt{2}\), the Hamiltonian reads
This then yields the Heisenberg equations of motion
where the effective hopping amplitude J and imaginary vector potential A are related to w and Δ by
Comparing against Eq. (1), we see that the dynamics of each canonical quadrature corresponds to that of a Hatano Nelson model, with opposite chiralities for \(\hat{x}\) and \(\hat{p}\) (Fig. 1). These orthogonal quadratures correspond to different phases of photonic excitations, and hence the system exhibits phasedependent nonreciprocal amplification^{17}. Note that there is a constraint on our mapping: the complex wavefunction amplitudes in the Hatano–Nelson model have been replaced by Hermitian quadrature operators in our system. This will play no role in what follows.
We now demonstrate how this setup can be used for Hamiltonian parameter estimation. We add a Hermitian perturbation \(\epsilon \hat{V}\) to our Hamiltonian where \(\hat{V}\) is some system operator; the goal is to estimate ϵ. We also couple the first site of our lattice to an inputoutput waveguide as a means to probe its properties. The simplest protocol is to use this waveguide to drive the system with a classical tone (i.e., a coherent state), and then measure the outgoing light in the waveguide (see Fig. 1b). The full Hamiltonian becomes
\({\hat{H}}_{\kappa }\) describes damping of the first site at a rate κ, due to coupling to the modes of the waveguide which we treat using standard inputoutput theory^{36}. The last term corresponds to a classical drive with amplitude β = ∣β∣e^{iθ}. Note that we take the drive frequency to match the resonance frequency of the isolated cavities; this frequency is zero in our rotating frame.
Using the standard inputoutput boundary condition, the output field in the waveguide is given by
where \({\hat{B}}^{({\rm{in}})}\), the operator equivalent of Gaussian white noise, describes the noise entering the lattice through the waveguide. Our goal is to estimate ϵ by making an optimal measurement of the output field. In what follows, we take ϵ to have units of frequency and \(\hat{V}\) to be dimensionless.
We further specialize to the usual case where ϵ is so small that it can only be estimated by integrating the output field over a long timescale τ. If we turn on the drive tone at t = 0, the relevant temporal mode of the output field to consider is
Note that this is normalized to be a canonical bosonic lowering operator, satisfying \([{\hat{{\mathcal{B}}}}_{\tau }(N),{\hat{{\mathcal{B}}}}_{\tau }^{\dagger }(N)]=1\). We write an explicit dependence on the chain size N, as we will be interested in understanding how things scale as N is increased.
The maximum amount of information available in \({\hat{{\mathcal{B}}}}_{\tau }(N)\) on ϵ is quantified by the quantum Fisher information (QFI). The QFI provides a lower bound on the root mean square error of any (unbiased) estimate of ϵ regardless of how \({\hat{{\mathcal{B}}}}_{\tau }(N)\) is measured^{1}. Calculation of the QFI unfortunately does not in general tell one the form of the optimal measurement. However, in our linear Gaussian system, things are much simpler: for large ∣β∣, the optimal measurement will always correspond to a standard homodyne measurement^{37,38}. The relevant Hermitian measurement operator has the form
i.e., a quadrature of the output operator \({\hat{{\mathcal{B}}}}_{\tau }(N)\) along a direction in phase space determined by the angle ϕ.
We will focus throughout on the largedrive limit, and will be interested in characterizing the QFI to leading order in ∣β∣. In this limit, QFI is determined by the statistics of \({\hat{M}}_{\tau }(N)\) via^{5,37}
where
are the signal and the noise, respectively, associated with the measurement. Here, 〈⋅〉_{z} means an average with respect to a state whose dynamics are governed by \(\hat{H}[z]\).
This expression for the QFI coincides with the SNR of an optimal homodyne measurement, and scales as ∣β∣^{2}; the nextleading order term is independent of ∣β∣. Note that the QFI only depends on the noise \({{\mathcal{N}}}_{\tau }(N,\epsilon )\) calculated to zeroth order in ϵ. We stress that the expression for the QFI still depends on the drive phase θ, as well as the form of the operator \(\hat{V}\); in what follows, we will be interested in optimizing these as well.
Given its role as a fundamental performance metric, it is tempting to declare that a better sensor has been built if it increases the QFI. Different measurement strategies however use resources differently, and one must carefully consider which to constrain when making comparisons. In our case, we wish to distinguish a true sensing enhancement from a more trivial effect, where a different protocol simply results in there being more photons in the system available to interact with the perturbation \(\hat{V}\) (as occurs with standard exceptionalpoint based sensing schemes^{5,6}). For this reason, we will take as the relevant metric the QFI scaled by the total average photon number \({\bar{n}}_{{\rm{tot}}}\)^{5}:
As we consider throughout the largedrive limit, we only keep the leadingorderinβ contribution to \({\bar{n}}_{{\rm{tot}}}\). This is simply the photon number associated with the driveinduced displacement of each cavity annihilation operator. The additional contribution to \({\bar{n}}_{{\rm{tot}}}\) due to amplification of vacuum fluctuations is βindependent, hence plays no role in the largedrive limit we consider (see “Methods” section).
Exponential SNR and QFI enhancement
We now focus on computing the optimal SNR of the measurement operator \({\hat{M}}_{\tau }(N)\) for our N site chain in the ϵ → 0 limit; via Eq. (13), this directly yields the QFI. In this limit, a SNR ~ 1 will only be achieved for τ much longer than any internal dynamical timescale. We thus consider the longτ limit, effectively ignoring any transient behaviour and assuming the system is in its steady state. Note that our system is dynamically stable as long as w > Δ and κ > 0, ensuring that a steady state exists.
From Eqs. (14), (12), and (10), the first order in ϵ in this limit reads
where
is the steady state linear response of the site1 average amplitude to a nonzero ϵ. This response will be determined by the zerofrequency susceptibilities (Green’s functions) of the unperturbed system.
It will be convenient to split up \(\delta {\langle {\hat{a}}_{1}\rangle }_{\epsilon }^{{\rm{ss}}}\) into its real and imaginary parts, or equivalently to think of the dynamics in the quadrature picture. There are then four different types of susceptibilities: χ^{αβ}[n, m; ω] is the response of the α quadrature on site n to a force which directly drives the β quadrature on site m. From Eqs. (5–6) and Eq. (2), we find that the ϵ = 0 susceptibilities are
Here \({\tilde{\chi }}^{\alpha \beta }[n,m;\omega ]\) is the susceptibility of a Hermitian N site tightbinding chain with hopping iJ and amplitude decay rate κ/2 on the first site (see “Methods” section). The above structure reflects the fact that the dynamics of the \(\hat{x}\) and \(\hat{p}\) quadratures correspond to two uncoupled copies of the Hatano–Nelson chain with opposite signed imaginary vector potential A. Hence, \(\hat{x}\) quadrature signals are amplified as they propagate to the right, and deamplified as they traverse to the left, while the opposite is true for \(\hat{p}\) quadrature signals. Note that if we started with two explicit Hatano–Nelson chains, the discussion here would be identical; x and p would then just index the two different chains.
To proceed, we need to specify the form of the perturbation Hamiltonian \(\hat{V}\). Our system exhibits the the nonHermitian skin effect (NHSE), implying a strong sensitivity to changes in boundary conditions. As the unperturbed system is an open chain, this suggests that an optimal \(\hat{V}\) would induce tunneling between the first and last site, i.e.,
with φ an arbitrary phase. As we show in “Methods” section, this choice of \(\hat{V}\) does not result in an enhanced sensitivity if one uses the proper metric of QFI/\({\bar{n}}_{{\rm{tot}}}\) (or equivalently SNR/\(\sqrt{{\bar{n}}_{{\rm{tot}}}}\)). While the signal produced by \({\hat{V}}_{{\rm{NHSE}}}\) is large, this is simply because our system is an amplifier with a large endtoend gain. The number of photons on the last site (and hence \({\bar{n}}_{{\rm{tot}}}\)) will be amplified equally by this gain. As a result, QFI/\({\bar{n}}_{{\rm{tot}}}\) does not show any enhancement as one increases the system size N, nor any enhancement over a conventional, singlecavity dispersive detector. We are thus left with a depressing conclusion: the nonHermitian skin effect does not provide any true advantage in sensing. Note also that \({\hat{V}}_{{\rm{NHSE}}}\) does not break the \({{\mathbb{Z}}}_{2}\) symmetry of the unperturbed system (see “Methods” section).
Luckily, this is not the end of the story. Enhanced sensing is possible with our system, if we chose a \(\hat{V}\) that fully exploits the opposite chiralities of our two (effective) Hatano–Nelson chains. Consider the innocuouslooking purely local perturbation
so that ϵ now corresponds to a small change in the resonance frequency of the last site. This perturbation does indeed break the \({{\mathbb{Z}}}_{2}\) symmetry of the unperturbed system. To understand how \({\hat{V}}_{N}\) affects the dynamics of the lattice, it is best to reexamine the equations of motion in the \(\hat{x}\) and \(\hat{p}\) basis. They remain the same everywhere except the last site N, where they now read
Recall that without the perturbation present, the dynamics of the \(\hat{x}\) and \(\hat{p}\) quadratures are completely independent (c.f. Eqs. (5) and (6)). The dispersive shift ϵ on site N now effectively couples the two nonHermitian chains, thereby breaking phasedependent nonreciprocity (see Fig. 1). While the intuitive picture of directional amplification remains unchanged in the rest of the lattice, a wavepacket with a well defined global phase can now scatter off of the perturbation ϵ and change its phase in the process. The role of ϵ is reminiscent to that of a magnetic impurity in the quantum spin Hall effect: in both cases the propagation direction of a particle is determined by some internal degree of freedom, which the impurity can change^{39}.
We next judiciously choose the phase of the drive β to be real and the measurement angle ϕ = π/2. Equivalently, we apply a driving force \(\sqrt{2\kappa } \beta \) to \({\hat{x}}_{1}\) and consider the corresponding response of its canonically conjugate quadrature \({\hat{p}}_{1}\). When ϵ = 0, this offdiagonal susceptibility vanishes, see Eq. (21). To first order in ϵ, it becomes nonzero. We further take N to be odd in what follows, as this guarantees (via the chiral symmetry of our unperturbed system) that the lattice will have a resonant mode at zero frequency. This then provides a further resonant enhancement of our system’s zero frequency response properties. Note that for an even N, we would still have the same exponential enhancement quoted in Eqs. (19–20); in this case however, there is no resonant mode at zero frequency, causing a suppression of susceptibilities by a multiplicative factor of κ/(2J) (see Eqs. (96–97)).
With these optimized choices, first order perturbation theory yields:
Here \({\bar{n}}_{N}\) denotes the leadingorderinβ average photon number of the last site in the lattice, and is given by:
For large A, the average photon number on site N is exponentially larger than that on other sites. Writing \({\bar{n}}_{N}=Z(A){\bar{n}}_{{\rm{tot}}}\) we have \(Z(A)=1{\mathcal{O}}({e}^{4A})\) (see “Methods” section). We thus obtain:
Equation (28) is a central result of this work: it shows that even when the total photon number \({\bar{n}}_{{\rm{tot}}}\) is held fixed, our system exhibits a signal power that grows exponentially with system size.
For an intuitive picture, consider the propagation of xquadrature photons injected from the waveguide into site 1, as depicted in Fig. 2. These photons will propagate to the last site N, with an amplitude χ^{xx}[N, 1; ω] ∝ e^{A(N−1)}. Photons that then scatter off the perturbation \(\epsilon {\hat{V}}_{N}\) will change phase, so that they now correspond to the p quadrature (c.f. Eq. (25)). They can then propagate back to the first lattice site with an amplitude − ϵχ^{pp}[1, N; ω] ∝ e^{A(N−1)}. This simple scattering process (involving both x and p quadrature propagation) leads to a parametrically large signal in \({\hat{p}}_{1}\).
The above heuristic picture also explains why the signal is amplified more than the average photon number \({\bar{n}}_{{\rm{tot}}}\): the average photon number only involves amplification along one traversal of the chain, whereas the signal magnitude involves two traversals (forward and back). This directly explains the extra large factor of e^{A(N−1)} in Eq. (28). We stress that this exponential signal enhancement would also occur in dissipative realizations of our doubled Hatano–Nelson chain.
The final step in characterizing our sensor is to examine its noise properties. Naively, one might expect that the same dynamics responsible for our signal enhancement would also exponentially amplify fluctuations in the output field. This is not the case: as already discussed, calculating the QFI only requires computing the noise to zeroth order in ϵ, see Eq. (13). Without the perturbation, the two effective Hatano–Nelson chains are completely decoupled. Thus, any noise entering through the waveguide will undergo equal amounts of amplification and deamplification before exiting the lattice. For the ideal case of zero internal loss, this means that the noise temperature of the output field will be identical to that of the input field. As a result, the noise in the homodyne current is
with \({\bar{n}}_{{\rm{th}}}\) representing the number of thermal quanta in the input field.
Combining these two results, our signaltonoise ratio is
where SNR_{τ}(1, ϵ) is the signaltonoise ratio of a ubiquitous singlemode dispersive detector^{36,40}. As we have stressed, SNR_{τ}(N, ϵ) also represents the QFI of our system. We see that the SNR and QFI can be exponentially enhanced by either increasing system size N or amplification factor A, all the while maintaining a fixed total photon number \({\bar{n}}_{{\rm{tot}}}\). This is the central result of our work. The crucial ingredients here are the inherent chiral amplification present in a Hatano–Nelson chain, the effective symmetry breaking that occurs when coupling the two oppositechirality chains in our sensor, and the lack of any amplified output noise in the unperturbed system.
This result might give one pause and, hence, several comments are in order. First, note that the large SNR achieved here is not contingent on approaching a parametric instability: our system is dynamically stable for any value of ϵ and A. Our enhancement is also not due to a large change in the spectrum of our chain: the spectrum of our nonHermitian Hamiltonian is independent of A and not exponentially sensitive to ϵ. Rather, coupling the two chains with opposite chirality has caused a large change in the wavefunctions of order ϵe^{2A(N−1)} (see Supplementary Note 3). Second, the mechanism we discuss here is useful even in small systems, as the fixed photon number QFI has an exponential dependence on A; an arbitrarily large QFI can thus be achieved with only three lattice sites. We further emphasize that the spatiallydependent amplification is a crucial aspect of our scheme. Indeed, the signaltonoise ratio for a singlemode cavity amplifier can never achieve this sort of sensing enhancement, since the signal and noise are amplified in a similar manner^{40}. Finally, we stress that this enhanced QFI in no way requires or is even related to the existence of an exceptional point in our dynamical matrix.
It is also worth stressing that our mechanism is completely distinct from other recently introduced methods that use parametric amplifiers to enhance dispersive sensing^{40,41,42}. These works exploit noise squeezing as the basic mechanism for enhancing the SNR and QFI. Unfortunately, in many practical settings this squeezing is difficult to exploit, as one becomes extremely sensitive to the added noise of amplification stages that follow the primary measurement (i.e., one needs following amplifiers to be quantum limited). In contrast, our scheme does not rely on squeezing the measurement noise, but instead effectively amplifies the signal power at fixed total photon number. The output noise has the same magnitude as the input noise, and hence taking advantage of our enhanced QFI does not need following amplification stages to be quantum limited. This represents a significant practical advantage. We also note that Ref. ^{5} studied nonHermitian quantum sensing that exploits nonreciprocity, but in a very different physical setting and using completely distinct underlying mechanism. Reference ^{5} focused on a zerodimensional gainloss dimer system. As a result, the sensing schemes in^{5} do not exploit propagation in an extended system (as we do), and cannot have interesting scaling with system size (as there is no notion of system size).
We end this section by pointing out that the N dependence of the QFI in Eq. (30) does not violate standard Heisenberglimit constraints^{1}, as the setting here is different. The usual Heisenberg limit applies to N sensor systems which each interact independently with the parameter of interest; the QFI here scales as best as ∝ N^{2}, a result which requires entanglement. In contrast, each of the N modes in our system is not an independent sensor interacting independently with the dispersive perturbation, as the sites are coupled. The enhanced scaling we find is not the result of entanglement: we stress that the input light to our system is just a coherent state. Instead, the enhancement is a consequence of our system’s unusual mechanism for nonreciprocal amplification.
NonMarkovian effects
We now relax the assumption that the parameter ϵ is infinitely weak. For concreteness, we assume the sensing target is to distinguish the case ϵ = 0 from the case ϵ = ϵ_{0} ≠ 0. This kind of discrimination is relevant in many practical situations, for example the dispersive measurement of the state of a qubit^{36}. We assume that ϵ_{0} is small enough such that linear response is still valid, but not so small that measurement will be infinitely long compared to internal system timescales. We thus need to understand the finitefrequency response and noise properties of our nonHermitian lattice sensor, and how these nonMarkovian corrections impact overall sensitivity.
In this section, we will characterize our sensor by its measurement time τ_{M}: what is the minimum integration time to achieve a SNR of unity? Heuristically, τ_{M} is the minimum amount of time required to distinguish between ϵ = 0 and ϵ = ϵ_{0}. In the limit ϵ_{0} → 0, τ_{M} will be much longer than any internal sensor timescale, and we can use the longtime limit SNR expression derived in the previous section (c.f. Eq. (30)). We define \({\tau }_{M}^{* }(N)\) to be this ϵ_{0} → 0 expression for the measurement time. Assuming that the input field has only vacuum noise, we find:
The obviously attractive feature here is the exponential reduction of τ_{M} with increasing lattice size N (but at fixed total photon number).
As N or ϵ_{0} is increased, \({\tau }_{M}^{* }(N)\) will become increasing smaller, and at some point will become comparable to internal system timescales. At this point, the longtime limit assumption used to derive this expression becomes invalid. There are two distinct relevant timescales that govern the dynamics of our sensor. The first t_{rt}(N) determines the ballistic propagation time to traverse the lattice end to end:
The second t_{esc}(N) involves the coupling to the waveguide: how quickly does a particle that is delocalized in the lattice leak out to the waveguide. A simple Fermi’s Golden Rule estimate yields the scale:
Both these timescales increase with system size. As a result, nonMarkovian effects associated with internal dynamics become increasingly important with increasing N. The crucial question is how this physics modified or places a limit on the exponentialinN measurement enhancement predicted by Eqs. (30) and (31). For large enough N the measurement will be so fast that these internal timescales matter. Do they simply put a bound on the measurement time, or does performance continue to increase with increasing N?
We first consider the limit J ≫ κ; the only relevant dynamical timescale is then t_{esc}(N), the time it takes a delocalized photon to escape the lattice. In this regime, the level spacing of lattice resonances is much larger than their widths. We can thus accurately approximate the relevant lowfrequency behaviour of lattice susceptibilities by the contribution from the zerofrequency resonance (whose width is 1/t_{esc}(N)). Assuming as always that N is odd, we have:
Note crucially that the residue at the poles are exponentially large in system size; this directly reflects the amplification physics we have discussed previously. Because of these factors, the above response functions are not simply equivalent to those of a singlemode system with a very small linewidth.
With this approximation, we find that the SNR is given by (see “Methods” section for details)
The bracketed factor represents the nonMarkovian correction to the longtime limit expression. Note that the correction is only to the magnitude of the signal. As we continue to use linear response, we only need to compute the homodyne current noise to zeroth order in ϵ_{0}. This noise is thus always vacuum noise regardless of the choice of integration time τ.
Using the above expression, we can then directly compute the measurement time τ_{M} in the J → ∞ limit. While finding the measurement time analytically is unfeasible, we can describe its asymptotic behavior in the strong and weak measurement limit (see “Methods” section)
We thus find a surprising result: even for fast measurements where the escape time from the lattice plays a role, the measurement time continues to improve exponentially with lattice size N. Intuitively, this is because the deleterious effects of increasing the escape time t_{esc} = (N + 1)/κ with increasing N is more than offset by the exponentially large number of photons e^{2A(N−1)} that exit through the waveguide when ϵ = ϵ_{0}.
We next consider the case where the hopping amplitude J is not infinitely larger than all other scales. In this case, we must also take into account the finite propagation speed v ∝ J of particles the lattice. Because an injected wavepacket must make a round trip before acquiring any information about the perturbation ϵ, for times less than 2N/v = N/J = t_{rt}(N) we expect the signal to be approximately zero. After this first round trip, the limiting factor in obtaining a large signal is once again the escape rate. Including the effects of a finite J, we find that the SNR is well approximated by simply adding a cutoff to the J → ∞ result in Eq. (36):
where Θ(t) is the Heaviside step function. This form reflects the basic intuition that it is impossible to make a measurement faster than the propagation time.
Combining these results, we finally find that including the effects of both internal timescales t_{rt}(N) and t_{esc}(N), the measurement time (to good approximation) is given by
where \({\tau }_{M}^{J = \infty }(N)\) is given in Eq. (31). Thus, as a function of increasing system size N, the measurement time first decreases exponentially until it reaches the roundtrip time in the lattice, after which it increases with N. The upshot of our analysis is that increasing the lattice size still provides an exponential sensing advantage when including nonMarkovian effects. This continues to be true until the measurement time is reduced to being on par with the roundtrip propagation time t_{rt}(N) = N/J.
In Fig. 3, we plot the numericallycalculated measurement time τ_{M}(N) versus lattice size N for a fixed total photon number \({\bar{n}}_{{\rm{tot}}}\) and perturbation size ϵ_{0}/κ; different curves correspond to different values of the hopping J/κ. We find an excellent agreement with the analytic approximation given in Eq. (39). The measurement time follows \({\tau }_{M}^{J = \infty }(N)\) (dark solid line) until it reaches the roundtrip time t_{rt}(N) (faint dashed lines), after which it increases linearly with N.
Beyond linear response
In this final section, we again consider the sensing problem of distinguishing ϵ = 0 from ϵ = ϵ_{0}; now however, we analyze the regime where (due to amplification effects) ϵ_{0} is too large for a linear response analysis to be valid. This is in contrast to the previous section, where ϵ_{0} was small enough that linear response was still valid, but large enough that nonMarkovian detector effects were important.
For any ϵ_{0} the output state of the light leaving the waveguide will be Gaussian, and the statistics of the measured homodyne current will be Gaussian. We can thus again quantify our sensor’s performance by calculating the signaltonoise ratio. We now however need to account for the fact that the homodyne current noise will also depend on ϵ_{0}. The definition of the signaltonoise ratio becomes:
This SNR quantifies the distinguishability between the Gaussian homodyne current distributions obtained for ϵ = 0 versus ϵ = ϵ_{0} (see e.g., refs. ^{36,43}).
As might be expected, the nonlinear dependence of SNR on ϵ_{0} will prevent one from indefinitely improving the measurement with increasing N. The key issue is that beyond linear response, noise amplification will also play a role. We show in what follows that even with this complication, our system yields a strong advantage, allowing one to fundamentally change the scaling of the SNR with ϵ_{0}.
We will focus on the most interesting situation where ϵ_{0}/κ ≪ 1, but where linear response breaks down because of a large amplification factor (i.e., e^{A(N−1)}ϵ_{0}/κ is not necessarily small). Further, we take the roundtrip time t_{rt}(N) = N/J to be small enough that we can ignore the transient dynamics and consider only the steadystate response. Formally, we now need to calculate the output field leaving the waveguide to all orders in ϵ_{0}. We thus expand the zero frequency quadratures of the output field as a power series in ϵ_{0}/κ:
To zeroth order in ϵ_{0}, there is no mixing of quadratures, and input signals are reflected with no net amplification (but just a trivial sign change):
Note that throughout this section, we associate the coherent drive tone amplitude β with the average value of \({\hat{X}}^{({\rm{in}})}[0]\).
In contrast, the first order contributions correspond to a process where input fields scatter once off the “impurity” before returning to the waveguide. This scattering converts one canonical quadrature to the other, and also results in a net amplification or deamplification
The amplification of \({\hat{X}}^{({\rm{in}})}\) is exactly the process we discussed earlier that is responsible for the exponentiallyenhanced signal. The attenuation of \({\hat{P}}^{({\rm{in}})}\) at this order can be understood analogously.
What about the second order in ϵ_{0} contributions? Heuristically, these correspond to input fields scattering off the impurity twice. While we expect such a process to preserve the identity of each canonical quadrature, it also has a more surprising feature: it results in no net amplification or deamplification:
This unexpected result can again be traced by to the chiral and quadraturedependent nature of gain and loss in our system. Interacting with the impurity twice implies that an input signal has performed at least two roundtrip traversals of the lattice (partially as an X, partially as a P). The gain and attenuation for each of these roundtrips necessarily cancel.
This pattern continues to higher order, and provides a simple explanation for the full expression we find for the output field: the net amplification/deamplification factor for each kind of quadrature to quadrature scattering process is independent of ϵ_{0}. We find
where
are elements of an orthogonal scattering matrix describing the conversion of quadratures (see “Methods” section for details). We see that quadraturepreserving scattering processes never come with amplification factors, whereas the amplification factors for quadraturechanging scattering are independent of ϵ_{0}. Crucially, there are no amplification factors in denominators in this expression. This result can be derived via a canonical squeezing transformation which eliminates the anomalous terms in Eq. (3); it also reflects the fact that our system is dynamically stable regardless of the strength of ϵ_{0}.
From these inputoutput relations, we can readily compute the SNR. Taking the noise of the input field to be vacuum, we have:
where \({\bar{n}}_{{\rm{tot}}}=({\bar{n}}_{{\rm{tot}}}(0)+{\bar{n}}_{{\rm{tot}}}({\epsilon }_{0}))/2\). Note that a finite ϵ_{0} now changes the number of coherent photons present on each lattice site m, the leadingorder correction being \(\delta {\bar{n}}_{m}\propto {({\epsilon }_{0}/\kappa )}^{2}{e}^{4A(Nm)}\). It is no longer necessarily the case that most of the photons reside on the last site. This ϵ_{0}induced change in photon number is incorporated in the definition of the correction factor Q(A, ϵ_{0}): in what follows we will make a careful choice of A such that Q(A, ϵ) is of order one.
We see that now, the denominator in Eq. (51) also depends on the amplification factor A, which corresponds to the amplification of noise. Because of this, increasing A and/or N indefinitely is no longer optimal. There remains nonetheless an advantage in using a carefully chosen amount of amplification. Ignoring Q(A, ϵ_{0}) and maximizing the SNR Eq. (51) with respect to the amplification, we see that the optimal choice corresponds to amplification that simply doubles the output noise over pure vacuum noise. In the ϵ_{0} ≪ κ limit of interest, the condition is:
With this optimized choice of A, the SNR written in terms of ϵ_{0} is then
We show in “Methods” section that \(Q({A}^{* },{\epsilon }_{0})=1{\mathcal{O}}\left({\left(\frac{{\epsilon }_{0}^{2}}{{\kappa }^{2}}\right)}^{\frac{1}{N1}}\right)\). Comparing against Eq. (26), we see that the optimized amplification has changed the fundamental scaling of the longtime SNR from being linear in the small parameter ϵ_{0}/κ to a squareroot dependence. Thus, by extending our analysis beyond a simple linearresponse treatment, we see that the exponential enhancement of the SNR predicted in Eq. (26) cannot extend indefinitely: the best one can do is to enhance the SNR (over a conventional dispersive measurement) by a large factor \(\sqrt{\kappa /{\epsilon }_{0}}\). This behaviour is plotted in Fig. 4. We again note that this predicted measurement enhancement does not require a large number of lattice sites; just three is already enough.
The enhanced squareroot dependence of the SNR on ϵ_{0} is superficially reminiscent of the behaviour found in nonHermitian exceptional point (EP) sensors^{9}. We stress that these phenomena are completely distinct. For EP sensors, it is the frequency of a resonance that exhibits a square root dependence, and not the SNR of a specific measurement (or other metric that also quantifies fluctuations). Further, EP sensing is based on operating near a point where the system’s dynamical matrix becomes defective and normal modes coalesce. In contrast, our system is not near such a special operating point. As we have stressed, the mechanism for enhanced SNR in our system is based on its directional amplification, and its ability to amplify signals and noise differently.
Discussion
In this work, we have shown how the unique features of nonHermitian lattice dynamics can be used for highly enhanced Hamiltonian parameter estimation and parametric sensing. We analyzed a concrete setup involving two copies of the Hatano–Nelson model and a symmetrybreaking perturbation. The response to the perturbation grows exponentially with system size, even when the total system photon number is kept fixed. Our analysis focused on a specific realization of this idea using a chain of parametrically driven cavities and a standard dispersing coupling to the parameter of interest. Here, even in the presence of quantum noise effects, the SNR and quantum Fisher information per photon both grow exponentially with system size. The system we described could be achieved in a variety of superconducting circuit and quantum optical platforms, and only requires one to make a homodyne measurement of the output field leaving the sensor. We also analyzed effects that go beyond standard linearresponse and Markovian assumptions. Even including higherorder effects, we show that our scheme allows one to dramatically enhance the SNR so that it depends on the square root of the sensing parameter.
Our work highlights the usefulness of multimode nonHermitian features that go beyond the mere existence of exceptional points. An open question is whether other unique features attributed to nonHermiticity, such as exotic topological phases or chiral mode switching, are also advantageous to quantum sensing problems.
Methods
\({{\mathbb{Z}}}_{2}\) symmetry of a nonHermitian tightbinding model
We discuss in more detail the \({{\mathbb{Z}}}_{2}\) symmetry that we wish to break to order to obtain an exponentially large response. We consider two finite, Nsite open Hatano–Nelson lattices with opposite chiralities (i.e., oppositely signed imaginary vector potentials A). The timedependent Schrödinger equation reads
where σ indexes the two chains. These equations of motion are invariant under a combination of time reversal \({\dot{\psi }}_{n}^{\sigma }\to {\dot{\psi }}_{n}^{\sigma }\), spatial inversion \({\psi }_{n}^{\sigma }\to {\psi }_{N+1n}^{\sigma }\) and pseudospin inversion \(\sigma \to \bar{\sigma }\).
While this is a seemingly trivial symmetry, we note that the Heisenberg equations of motion of our dissipationfree realization of the same model
are invariant under the same set of symmetries, where \({\hat{x}}_{n}\) and \({\hat{p}}_{n}\) play the role of pseudospin. To describe these symmetries requires using operators acting on the bosonic Hilbert space. To this end, we consider the antiunitary time reversal operator \({\mathcal{T}}\), a unitary rotation operator \({\mathcal{R}}\) and the unitary spatial inversion operator \({\mathcal{S}}\) whose action on the quadratures reads
Equivalently, we have
With these definitions in hand, it is easy to verify that the Hermitian Hamiltonian which gives the equations of motion Eqs. (56) and (57)
is invariant under the combination of timereversal, rotation, and spatial inversion. Similarly, the nonlocal perturbation
is invariant under the same combination of symmetries. Timereversal changes the phase φ → −φ, \({\hat{V}}_{{\rm{NHSE}}}\) commutes with \({\mathcal{R}}\) and spatial inversion sends \({\hat{a}}_{1}^{\dagger }{\hat{a}}_{N}\to {\hat{a}}_{N}^{\dagger }{\hat{a}}_{1}\).
While this \({{\mathbb{Z}}}_{2}\) symmetry can be made formal as we have just shown, it is perhaps best explained in words: each chain must have a preferred direction for amplification/propagation, with the direction being opposite in each chain. Hence, the symmetry only involves the chirality in each chain. One can then optimally sense a perturbation that couples the two chains (as we have discussed). Because the forward amplification simply contributes to the total photon number, the backwards amplification is most crucial as it directly contributes to the signal. Exponentiallyenhanced sensitivity is then obtained in a larger class of sensors by lifting the constraint of an exact \({{\mathbb{Z}}}_{2}\) symmetry. For instance, one could have two chains of unequal length with ∣A_{↑}∣ ≠ ∣A_{↓}∣ whose QFI per photon is still exponentially large in the length of the chain where backwards amplification occurs.
Quadrature susceptibility matrices
We first compute the susceptibilities for the \({\hat{x}}_{n}\) and \({\hat{p}}_{n}\) quadratures, defined as \({\hat{a}}_{n}=({\hat{x}}_{n}+i{\hat{p}}_{n})/\sqrt{2}\). The HeisenbergLangevin equations of motion in this basis are
where \({\hat{X}}^{({\rm{in}})}\) and \({\hat{P}}^{({\rm{in}})}\) are the operator equivalent of Gaussian white noise. They average to zero, and their second moment is
where \({\bar{n}}_{{\rm{th}}}\) is the number of thermal quanta in the input field. We now focus on the case where \({\bar{n}}_{{\rm{th}}}=0\), with generalizations to finitetemperature inputs being straightforward.
An immense simplification arises by making a local Bogoliubov (squeezing) transformation, such that the Hamiltonian preserves the total number of these new quasiparticles. The dynamical matrix of in this new basis is then explicitly Hermitian. Defining new canonically conjugate quadrature operators \({\hat{\tilde{x}}}_{n}\) and \({\hat{\tilde{p}}}_{n}\) by
with n_{0} an arbitrary real number, we have
with \({\hat{\tilde{a}}}_{n}=({\hat{\tilde{x}}}_{n}+i{\hat{\tilde{p}}}_{n})/\sqrt{2}\) a transformed canonical annihilation opreator. The parameter n_{0} does not enter the Hamiltonian since \({\hat{H}}_{B}\) is invariant under a uniform local squeezing operation that does not mix quadratures \({\hat{x}}_{n}\to {e}^{A{n}_{0}}{\hat{x}}_{n},{\hat{p}}_{n}\to {e}^{A{n}_{0}}{\hat{p}}_{n}\). In this section, it will be convenient to set n_{0} = 1, so that the annihilation operators on the first stie remain unchanged \({\hat{\tilde{a}}}_{1}={\hat{a}}_{1}\).
The HeisenbergLangevin equations of motion in this new basis read
As expected, the response properties of \({\hat{\tilde{x}}}_{n}\) and \({\hat{\tilde{p}}}_{n}\) are then determined by a completely Hermitian matrix (other than the waveguideinduced decay on the first site).
Using the squeezing transformations Eqs. (71–72) and the fact that the dynamics of the \(\hat{x}\) and \(\hat{p}\) quadratures are uncoupled, the relevant quadrature–quadrature susceptibilities read
where \({\tilde{\chi }}^{\alpha \beta }(n,m;t)\) are quadrature response functions of a standard (i.e., reciprocal and particleconserving) tightbinding chain with a waveguide attached to the first site. Note that our convention differs from that used in the condensed matter community, where χ^{αβ}(n, m; t) is the response of quadrature α to a force which couples to β in the Hamiltonian. Computing the quadraturequadrature susceptibilities χ^{αβ}(n, m; t) of our nonreciprocal system is then no more complicated than finding the susceptibilities of a reciprocal tightbinding chain \({\tilde{\chi }}^{xx}(n,m;t)\) and \({\tilde{\chi }}^{pp}(n,m;t)\).
The susceptibilities of the Hatano–Nelson model Eq. (1) are computed in a similar manner. There, instead of a local squeezing transformation, one makes a so called imaginary gauge transformation \(\leftn\right\rangle \to {e}^{A(n{j}_{0})}\leftn\right\rangle\) and \(\left\langle n\right\to {e}^{A(n{j}_{0})}\left\langle n\right\). In this new gauge, the Hamiltonian is Hermitian and completely independent of A. The factorization of Eq. (2) as \(\chi (n,m;t)={e}^{A(nm)}\tilde{\chi }(n,m;t)\) immediately follows.
Particleconserving susceptibilities
Although so far we have only considered quadraturequadrature response functions, the fact that we can map our Hamiltonian onto a particle conserving one makes it much simpler to keep track of the dynamics of the single squeezed mode \({\hat{\tilde{a}}}_{n}\). Indeed, we have
where
Because our Hamiltonian is quadratic in boson opeators and conserves total quasiparticle number, we can readily use the singleparticle formalism to find the relevant susceptibilities. If we let \(\leftn\right\rangle\) denote a position eigenket, we then have
with
It is more convenient to write the susceptibilities in the frequency domain:
We’ll first compute the susceptibilities without the effects of ϵ or κ, that is
Written out explicitly, the matrix elements of the susceptibility for a finite open chain then satisfy the difference equation
with boundary conditions \({\tilde{\chi }}_{0}[0,m;\omega ]={\tilde{\chi }}_{0}[N+1,m;\omega ]=0\). The exact form of the susceptibility matrix is known; here for the sake of completeness we quickly sketch how to obtain it. First, we note that Eq. (90) has the form of a translationally invariant Green’s function problem in the index space n, with −iδ_{nm}/J acting as a source term. The general solution will then consist of a linear combination of the source free solution and a convolution (in the index space n) of the source with the homogeneous solution.
The source free solution, which satisfies
is precisely (up to a factor of i) the recursion relation that defines T_{n}(ω/2J) and U_{n}(ω/2J), the Chebyshev polynomials of the first and second kind, respectively. Since U_{−1}(ω/2J) = 0, and given our boundary condition \({\tilde{\chi }}_{0}[0,m;\omega ]=0\), we conclude that the source free solution is
with c_{m} a constant that will be used to satisfy the second boundary condition. The full solution to Eq. (90) is then
with Θ(n − m) the Heaviside step function (where Θ(0) = 0). Enforcing the second boundary condition \({\tilde{\chi }}_{0}[N+1,m;\omega ]=0\) yields
We now turn our attention to computing the response functions in the presence of the waveguide on the first site. Formally, this introduces a local term −κ/2δ_{n,1}δ_{m,1} to the the dynamical matrix. The full susceptibilities \(\tilde{\chi }[n,m;\omega ]\) can then readily be solved algebraically using Dyson’s equation
Since there is only a driving force on the first site and we are only interested in the response on the first site, we must only compute \(\tilde{\chi }[n,1;\omega ]\) and \(\tilde{\chi }[1,m;\omega ]\):
Because \({\tilde{\chi }}^{px}(n,m;t)={\tilde{\chi }}^{xp}(n,m;t)=0\), from Eq. (82) we conclude that \(\tilde{\chi }[n,m;\omega ]={\tilde{\chi }}^{xx}[n,m;\omega ]={\tilde{\chi }}^{pp}[n,m;\omega ]\). With this result and Eqs. (77–78), we now have all the relevant quadrature–quadrature susceptibilities.
QFI for \({\hat{{\boldsymbol{V}}}}_{{\boldsymbol{NHSE}}}\)
We are now in a position to compute \({{\rm{QFI}}}_{\tau }(N)/{\bar{n}}_{{\rm{tot}}}\) for any choice of perturbation \(\hat{V}\). Recall that that in the large β limit of interest, the QFI coincides with SNR squared, optimizing over the homodyne angle ϕ (see Eqs. (13)). As is written in the main text, see Eq. (17), the steadystate signal takes the form
with \(\delta {\langle {\hat{x}}_{1}\rangle }^{{\rm{ss}}}\) and \(\delta {\langle {\hat{p}}_{1}\rangle }^{{\rm{ss}}}\) the steady state linear response of the site1 average quadrature amplitude to a nonzero ϵ. The signal will depend on the phase of the coherent drive β which we take to be real, as in the main text. Our conclusion that \({\hat{V}}_{{\boldsymbol{NHSE}}}\) does not have an exponentially large QFI/\({\bar{n}}_{{\rm{tot}}}\) is independent of the phase of β, as will become evident. This choice of phase is equivalent to driving the \({\hat{x}}_{1}\) quadrature with a force \(\sqrt{2\kappa }\beta\), so that the signal is
The form of the responses δχ^{xx}[1, 1; 0] and δχ^{px}[1, 1; 0] will depend on \(\hat{V}\). For the nonlocal hopping perturbation
the change to the equations of motion induced by ϵ to the quadratures on the first site read:
First order perturbation theory then yields
where χ^{αα}[n, m; ω] the susceptibilities of the unperturbed system, which we computed in the previous sections. The salient feature is that χ^{xx}[n, m; ω] ∝ e^{A(n−m)} and χ^{pp}[n, m; ω] ∝ e^{−A(n−m)} due to the phasedependent chiral propagation. With the factor of χ^{xx}[N, ; 0], it would appear that we have we do in fact have an exponentially large response. Yet it precisely this terms which controls the number of coherent photons on site N, since \({\bar{n}}_{N}=\kappa  \beta { }^{2} {\chi }^{xx}[N,1,0]{ }^{2}\). Expressing the signal in terms of \({\bar{n}}_{N}\) gives
where we have used χ^{xx}[1, 1; 0] = χ^{pp}[1, 1; 0] = 2/κ for a chain with a odd number of sites. The form of Eq. (105) makes it evident that SNR\(/\sqrt{{\bar{n}}_{{\rm{tot}}}}\) does not scale exponentially with system size, and therefore neither does QFI\(/{\bar{n}}_{{\rm{tot}}}\).
Despite the perturbation having coupled the two effective Hatano–Nelson chains with an amplitude of \(\epsilon \cos \varphi\) (see Eqs. (101) and (102)), this is not enough to ensure a large SNR\(/\sqrt{{\bar{n}}_{{\rm{tot}}}}\). The nonlocal form of \({\hat{V}}_{{\rm{NHSE}}}\) implies that a wavepacket only experiences unidirectional amplification before exiting the waveguide. In contrast, the perturbation \({\hat{V}}_{N}={\hat{a}}_{N}^{\dagger }{\hat{a}}_{N}\) studied throughout the main text allows for amplification before and after interacting with ϵ.
Data availability
The numerical data generated in this work is available from the authors upon reasonable request.
References
 1.
Giovannetti, V., Lloyd, S. & Maccone, L. Advances in quantum metrology. Nat. Photon. 5, 222 (2011).
 2.
Degen, C. L., Reinhard, F. & Cappellaro, P. Quantum sensing. Rev. Mod. Phys. 89, 035002 (2017).
 3.
Pezzè, L., Smerzi, A., Oberthaler, M. K., Schmied, R. & Treutlein, P. Quantum metrology with nonclassical states of atomic ensembles. Rev. Mod. Phys. 90, 035005 (2018).
 4.
Langbein, W. No exceptional precision of exceptionalpoint sensors. Phys. Rev. A 98, 023805 (2018).
 5.
Lau, H.K. & Clerk, A. A. Fundamental limits and nonreciprocal approaches in nonhermitian quantum sensing. Nat. Commun. 9, 1–13 (2018).
 6.
Zhang, M. et al. Quantum noise theory of exceptional point amplifying sensors. Phys. Rev. Lett. 123, 180501 (2019).
 7.
Chen, C., Jin, L. & Liu, R.B. Sensitivity of parameter estimation near the exceptional point of a nonhermitian system. N. J. Phys. 21, 083002 (2019).
 8.
Naghiloo, M., Abbasi, M., Joglekar, Y. N. & Murch, K. W. Quantum state tomography across the exceptional point in a single dissipative qubit. Nat. Phys. 15, 1232–1236 (2019).
 9.
Wiersig, J. Enhancing the sensitivity of frequency and energy splitting detection by using exceptional points: Application to microcavity sensors for singleparticle detection. Phys. Rev. Lett. 112, 203901 (2014).
 10.
Chen, W., Özdemir, Ş. K., Zhao, G., Wiersig, J. & Yang, L. Exceptional points enhance sensing in an optical microcavity. Nature 548, 192–196 (2017).
 11.
Hodaei, H. et al. Enhanced sensitivity at higherorder exceptional points. Nature 548, 187–191 (2017).
 12.
Xiao, Z., Li, H., Kottos, T. & Alù, A. Enhanced sensing and nondegraded thermal noise performance based on \({\mathcal{P}}{\mathcal{T}}\)symmetric electronic circuits with a sixthorder exceptional point. Phys. Rev. Lett. 123, 213901 (2019).
 13.
Lai, Y.H., Lu, Y.K., Suh, M.G., Yuan, Z. & Vahala, K. Observation of the exceptionalpointenhanced sagnac effect. Nature 576, 65–69 (2019).
 14.
Yao, S. & Wang, Z. Edge states and topological invariants of nonhermitian systems. Phys. Rev. Lett. 121, 086803 (2018).
 15.
Lee, T. E. Anomalous edge state in a nonhermitian lattice. Phys. Rev. Lett. 116, 133903 (2016).
 16.
Kunst, F. K., Edvardsson, E., Budich, J. C. & Bergholtz, E. J. Biorthogonal bulkboundary correspondence in nonhermitian systems. Phys. Rev. Lett. 121, 026808 (2018).
 17.
McDonald, A., PeregBarnea, T. & Clerk, A. A. Phasedependent chiral transport and effective nonhermitian dynamics in a bosonic kitaevmajorana chain. Phys. Rev. X 8, 041031 (2018).
 18.
Xiong, Y. Why does bulk boundary correspondence fail in some nonhermitian topological models. J. Phys. Commun. 2, 035043 (2018).
 19.
Lee, C. H. & Thomale, R. Anatomy of skin modes and topology in nonhermitian systems. Phys. Rev. B 99, 201103 (2019).
 20.
Kawabata, K., Shiozaki, K., Ueda, M. & Sato, M. Symmetry and topology in nonhermitian physics. Phys. Rev. X 9, 041015 (2019).
 21.
Borgnia, D. S., Kruchkov, A. J. & Slager, R.J. Nonhermitian boundary modes and topology. Phys. Rev. Lett. 124, 056802 (2020).
 22.
Lieu, S. Topological phases in the nonhermitian suschriefferheeger model. Phys. Rev. B 97, 045106 (2018).
 23.
Schomerus, H. Nonreciprocal response theory of nonhermitian mechanical metamaterials: response phase transition from the skin effect of zero modes. Phys. Rev. Res. 2, 013058 (2020).
 24.
Hatano, N. & Nelson, D. R. Vortex pinning and nonhermitian quantum mechanics. Phys. Rev. B 56, 8651–8673 (1997).
 25.
Wang, Y.X. & Clerk, A. A. Nonhermitian dynamics without dissipation in quantum systems. Phys. Rev. A 99, 063834 (2019).
 26.
Blais, A., Huang, R.S., Wallraff, A., Girvin, S. M. & Schoelkopf, R. J. Cavity quantum electrodynamics for superconducting electrical circuits: An architecture for quantum computation. Phys. Rev. A 69, 062320 (2004).
 27.
Vollmer, F., Arnold, S. & Keng, D. Single virus detection from the reactive shift of a whisperinggallery mode. Proc. Natl Acad. Sci. USA 105, 20701–20704 (2008).
 28.
Wanjura, C. C., Brunelli, M. & Nunnenkamp, A. Topological framework for directional amplification in drivendissipative cavity arrays. Nat. Commun. 11, 3149 (2020).
 29.
Hatano, N. & Nelson, D. R. Localization transitions in nonhermitian quantum mechanics. Phys. Rev. Lett. 77, 570–573 (1996).
 30.
Petermann, K. Calculated spontaneous emission factor for doubleheterostructure injection lasers with gaininduced waveguiding. IEEE J. Quantum Electron. 15, 566–570 (1979).
 31.
Grangier, P. & Poizat, J.P. A simple quantum picture for the petermann excess noise factor. Eur. Phys. J. D.At. Mol. Optical Plasma Phys. 1, 97–104 (1998).
 32.
Okuma, N., Kawabata, K., Shiozaki, K. & Sato, M. Topological origin of nonhermitian skin effects. Phys. Rev. Lett. 124, 086801 (2020).
 33.
Metelmann, A. & Clerk, A. A. Nonreciprocal photon transmission and amplification via reservoir engineering. Phys. Rev. X 5, 021025 (2015).
 34.
Metelmann, A. & Türeci, H. E. Nonreciprocal signal routing in an active quantum network. Phys. Rev. A 97, 043833 (2018).
 35.
Brandenbourger, M., Locsin, X., Lerner, E. & Coulais, C. Nonreciprocal robotic metamaterials. Nat. Commun. 10, 4608 (2019).
 36.
Clerk, A. A., Devoret, M. H., Girvin, S. M., Marquardt, F. & Schoelkopf, R. J. Introduction to quantum noise, measurement, and amplification. Rev. Mod. Phys. 82, 1155–1208 (2010).
 37.
Banchi, L., Braunstein, S. L. & Pirandola, S. Quantum fidelity for arbitrary gaussian states. Phys. Rev. Lett. 115, 260501 (2015).
 38.
Pinel, O., Jian, P., Treps, N., Fabre, C. & Braun, D. Quantum parameter estimation using general singlemode gaussian states. Phys. Rev. A 88, 040102 (2013).
 39.
Maciejko, J., Hughes, T. L. & Zhang, S.C. The quantum spin hall effect. Annu. Rev. Condens. Matter Phys. 2, 31–53 (2011).
 40.
Eddins, A. et al. Highefficiency measurement of an artificial atom embedded in a parametric amplifier. Phys. Rev. X 9, 011004 (2019).
 41.
Barzanjeh, S., DiVincenzo, D. P. & Terhal, B. M. Dispersive qubit measurement by interferometry with parametric amplifiers. Phys. Rev. B 90, 134515 (2014).
 42.
Peano, V., Schwefel, H. G. L., Marquardt, C. & Marquardt, F. Intracavity squeezing can enhance quantumlimited optomechanical position detection through deamplification. Phys. Rev. Lett. 115, 243603 (2015).
 43.
Laflamme, C. & Clerk, A. A. Weak qubit measurement with a nonlinear cavity: beyond perturbation theory. Phys. Rev. Lett. 109, 123602 (2012).
Acknowledgements
We thank Kero Lau for a careful reading of the manuscript. This material is based upon work supported by the Air Force Office of Scientific Research under award number FA95501910362.
Author information
Affiliations
Contributions
A.M. and A.A.C. contributed equally in the development of idea, derivation of results, and writing of manuscript.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Peer review information Nature Communications thanks the anonymous reviewer(s) for their contribution to the peer review of this work.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
McDonald, A., Clerk, A.A. Exponentiallyenhanced quantum sensing with nonHermitian lattice dynamics. Nat Commun 11, 5382 (2020). https://doi.org/10.1038/s41467020190904
Received:
Accepted:
Published:
Further reading

Impurity induced scalefree localization
Communications Physics (2021)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.