Abstract
A fundamental challenge in digital quantum simulation (DQS) is the control of an inherent error, which appears when discretizing the time evolution of a quantum many-body system as a sequence of quantum gates, called Trotterization. Here, we show that quantum localization-by constraining the time evolution through quantum interference-strongly bounds these errors for local observables, leading to an error independent of system size and simulation time. DQS is thus intrinsically much more robust than suggested by known error bounds on the global many-body wave function. This robustness is characterized by a sharp threshold as a function of the Trotter step size, which separates a localized region with controllable Trotter errors from a quantum chaotic regime. Our findings show that DQS with comparatively large Trotter steps can retain controlled errors for local observables. It is thus possible to reduce the number of gate operations required to represent the desired time evolution faithfully.
INTRODUCTION
Quantum computers promise to solve certain computational problems exponentially faster than any classical machine (1). A particularly promising application is the solution of quantum many-body problems (2), with large potential impact on quantum chemistry, material science, and fundamental physics. The devices used in this effort can be divided into two major classes: analog quantum simulators, where the Hamiltonian of interest is engineered to mimic the desired quantum many-body physics; and digital quantum simulators (DQSs), where a target time-evolution operator is represented by a sequence of elementary quantum gates. The digital approach is particularly flexible since a universal digital quantum simulator can be freely programmed to simulate the unitary evolution of any many-body Hamiltonian with local interactions (Fig. 1A) (3). Recent experiments have demonstrated remarkable progress in implementing digital quantum simulation (DQS), e.g., by simulating simple molecules in quantum chemistry (4–6), condensed-matter models (7–12), and lattice gauge theories (13).
(A) Gate sequence for the digital quantum simulation (DQS) of an Ising model. The desired evolution up to total simulation time t is split into n repeated sequences of length τ = t/n, each decomposed into fundamental quantum gates. The example shows a gate sequence for a four-qubit chain with Ising spin-spin interactions (ZZ) and transverse and longitudinal fields (simulated by single-qubit operations along the X and Z directions on the Bloch sphere). (B) Magnetization dynamics
The working principle of DQS is as follows. Suppose that the target Hamiltonian
This Trotterization comes inherently with an error that can be rigorously upper bounded via the accuracy of the global unitary time-evolution operator (3)
Here, ϵ subsumes terms of order t3/n2 and higher. According to Eq. 2, for the lowest-order corrections, the Trotter error on the full time-evolution operator may grow quadratically with total simulation time t and (in generic quantum many-body systems) linearly in the number of simulated degrees of freedom N. It is possible to improve this upper bound, but an error bound that scales less than linear in t is not possible if one is concerned with the entire unitary operator (14). Although the polynomial scaling with both t and N is efficient in a computational complexity sense, it poses a substantial challenge for practical computations (15, 16), seemingly preventing current technology from simulating all but small instances. As we show in this article, these generic bounds on the global many-body wave function overestimate by far the actual error on local observables such as magnetizations or low-order correlation functions. For example, in the DQS of a quantum Ising chain, the deviation of the magnetization dynamics from the ideal evolution can be considerably smaller and remain bounded even at long times (see Fig. 1B). It is the purpose of this article to explain this observation from physical grounds and thus assign a physical interpretation to Trotter errors.
We achieve this by linking Trotter errors to quantum localization. Localization is a ubiquitous phenomenon with many facets. Initially, it has been introduced to understand the absence of transport in systems of free particles with disorder (17). Since then, the concept has been generalized to various contexts such as many-body localization in Hilbert space as absence of quantum ergodicity (18) or energy localization in periodic time-dependent quantum many-body systems as absence of heating in continuously driven systems (19). As we show here, there occurs a related localization in Hilbert space at small Trotter steps that bounds time-discretization errors on local observables occurs. Our aim is to isolate the role of this universal error source, for which we focus in the following mostly on an idealized setting. We discuss the interplay of the Trotter errors with other, platform-dependent error sources in the concluding remarks and in the Supplementary Materials.
RESULTS
Trotter sequences as Floquet systems
In this work, we interpret the Trotterized evolution as a periodically time-dependent quantum many-body system with a period τ = t/n (see Fig. 1). The desired stroboscopic dynamics is therefore governed by an associated Floquet Hamiltonian HF, which we define for later convenience in the following form
The starting point of our considerations is an analytical expression for HF in the limit of sufficiently small Trotter steps τ
This form, which can be obtained from Eq. 3 via a Magnus expansion, quantifies the Trotterization error on a Hamiltonian level. There remain, however, two fundamental questions that we aim to address in this work: (i) What is the radius of convergence τ* of this expansion? [Mathematically rigorous bounds for the convergence radius of the Magnus expansion do exist, but their applicability to generic quantum many-body systems is not often evident (19, 20).] (ii) What is the influence of corrections to H that appear in HF on the long-time dynamics of local observables? Recent theoretical predictions for heating in generic quantum many-body systems subject to a periodic drive might leave a rather pessimistic impression (21–23). We show in this work that the errors on local observables can nevertheless be controlled for all practical purposes. Throughout this work, we adopt the notion of local observables to be operators that include a bounded number of constituents. This definition includes local order parameters and typical correlation functions. These are not only generically the measures to describe the properties of physical systems, they are also the quantities that can be experimentally measured in a scalable way. Moreover, our general argumentation holds for physical Hamiltonians with few-body interactions, which for almost all of our discussion may even be taken to be long-ranged, except if otherwise stated.
Benchmark model: Quantum Ising chain
In the following, we illustrate our discussion with a generic, experimentally relevant model, the quantum Ising chain with Hamiltonian H = HZ + HX, with
Quantum many-body chaos threshold
As the central result of this work, we connect Trotter errors in DQS with a threshold separating a many-body quantum chaotic region from a localized regime, thus linking the intrinsic accuracy of DQS with a quantum many-body phenomenon. For that purpose, we first investigate the inverse participation ratio (IPR)
(A) Rate function λIPR of the IPR, normalized to the maximally achievable value
A strong fingerprint of quantum chaos can also be found in out-of-time–ordered (OTO) correlators, which quantify how fast quantum information scrambles through a many-body system. A typical OTO correlator is of the form
In Fig. 2B, we present numerical evidence that this quantity detects the many-body quantum chaos threshold that we have seen in the IPR. There is a clear threshold that separates a localized region at small Trotter steps τ, where ℱ > 0, from a quantum chaotic region at large τ, where ℱ → 0. The vanishing OTO correlator in the many-body quantum chaotic regime can be understood directly from the results obtained for the IPR. Consider the spectral decomposition of a local Hermitian operator V = ∑αλα|α〉〈α|, with λα as the eigenvalues and |α〉 as the eigenvectors of V (for the considered magnetization, these are equivalent to the set of spin configurations). The effective Floquet dynamics yields after n periods
Within the localized phase, the amplitudes Cvα contain more structure than only the phase information, which yields a nonzero value for the OTO correlator. For small systems, such as for N = 10 in Fig. 2B, one can observe additional structures in the crossover region, which vanish for larger N. We attribute these to individual quantum many-body resonances, which can be resolved in small systems but which merge for large N.
Robustness of local observables
While the corrections due to time discretization are weak on a Hamiltonian level, as seen in the Magnus expansion in Eq. 4, there is a priori no guarantee that the long-time dynamics is equally well reproduced. It is, e.g., well known for classical chaotic systems that even weak perturbations can grow quickly in time. Here, we provide numerical evidence that in the localized regime, the dynamics of local observables remains constrained and controlled, even in the long-time limit. This asymptotic long-time dynamics is a worst-case scenario for DQS: When Trotter errors on local observables can be controlled in this limit, so can they on shorter times. In the Supplementary Materials, we also discuss the buildup of Trotter errors on short to intermediate times in more detail.
In Fig. 3A, we show the asymptotic long-time value ℳ of the magnetization,
Both the magnetization ℳ (A) and simulation accuracy QE (C) exhibit a sharp crossover from a regime of controllable Trotter errors for small Trotter steps τ to a regime of strong heating at larger τ. The dashed line in (A) refers to the desired case of the ideal evolution. The Trotter error exhibits a quadratic scaling at small τ for both the deviation of the magnetization, Δℳ = ℳ − ℳτ=0, (B) and QE (D). The solid lines in (B) and (D) represent analytical results obtained perturbatively in the limit of small Trotter steps τ. These results indicate the controlled robustness of digital quantum simulation against Trotter errors, in the long-time limit and largely independent of N.
Simulation accuracy
In the previous sections, we have provided evidence for a sharp threshold between a delocalized and a localized regime with controllable Trotter errors. We now aim to understand the influence of the regular regions onto the dynamics of local observables. We identify as the underlying reason for the weak Trotter errors a dynamical constraint due to an emergent stroboscopic constant of motion in the effective time-periodic problem, which is the Floquet Hamiltonian HF. Although this integral of motion is different from the desired energy conservation of the target Hamiltonian H, the perturbative expansion in Eq. 4 suggests a close connection. It is therefore natural to quantify the accuracy of DQS by measuring how far the system deviates from the desired constant of motion H via
Here, we have introduced Eτ(nτ) = 〈H(nτ)〉τ and E0 = Eτ → 0(nτ) = 〈ψ0|H|ψ0〉, where the subindex τ refers to the used Trotter step for the dynamics. In QE(t), we normalize the errors using the system’s energy at infinite temperature,
In Fig. 3C, we show numerical data for the long-time average QE. Again, we find a sharp threshold between the localized and quantum chaotic regimes. For small Trotter steps, QE acquires only a weak quadratic dependence on τ (see Fig. 3D), yielding
To obtain an analytical understanding for the observations of weak Trotter errors on local observables, let us start by considering the Magnus expansion for the Floquet Hamiltonian in Eq. 4, which quantifies the leading-order corrections due to time discretization on a Hamiltonian level. From our numerical results for QE, we anticipate that the target Hamiltonian H is an almost conserved quantity, which motivates us to study the perturbative corrections to strict energy conservation. Using time-dependent perturbation theory up to second order in the Trotter step size τ, we find
The explicit derivation and the final formula for qE are given in Materials and Methods. For the considered parameters, we estimate qE = 0.18. As it can be seen in Fig. 3D, this analytical value matches well the numerical results.
To test whether the errors on other local observables are also controlled by the emergent constant of motion in the localized regime, we exemplarily study the corrections to the targeted magnetization dynamics. From time-dependent perturbation theory we obtain
Our observations give a smaller error on local observables than suggested by general considerations on Floquet dynamics in high-frequency regimes (corresponding to small Trotter steps) (32, 33). In these works, it is shown that there exists a static local Hamiltonian
DISCUSSION
As we have shown, intrinsic Trotter errors in DQS are controllable for local observables, with a sharp threshold separating a localized from a many-body quantum chaotic regime. We have achieved this by identifying the Trotterized time evolution on general grounds with an effective time-periodic Floquet problem. As a consequence, the dynamics is constrained by an emergent conserved quantity given by the Floquet Hamiltonian HF in Eq. 3. While the target Hamiltonian H is not conserved in the Trotterized dynamics, in the localized regime, H remains almost conserved up to perturbative corrections for small Trotter steps τ. This finding does not hold anymore in the many-body quantum chaotic regime, where Trotter errors proliferate and become uncontrollable. While we present data here for a specific model and a specific initial state, our arguments remain general; we find similar properties also for other initial conditions and other model systems such as the recently experimentally realized lattice Schwinger model, which we discuss in the Supplementary Materials. Furthermore, analogous behavior is also found in long-ranged spin models (34). Our numerical studies are based on up to N = 20 qubits, which is within realized and expected size ranges of digital quantum simulators (4–13, 35–38).
For experiments, it is of particular interest to assess the precise value of the threshold scale τ*. Theoretically predicting τ* is in general as difficult as solving the desired time evolution. Nevertheless, one can estimate τ* as follows. Before running an experiment, one can numerically calculate QE for small N, yielding a first estimate on τ*. From this starting point, experiments can find an optimal Trotter step at larger N by decreasing τ until sufficient convergence is reached. Once in the perturbative regime, one can use data at nonzero τ to extrapolate to the ideal dynamics in a well-defined way.
For concrete experimental realizations, it is furthermore of relevance how Trotter errors behave on short to intermediate time scales. From Fig. 1B, one can anticipate that the long-time Trotter error emerges already on rather short time scales. In the Supplementary Materials, we study this transient dynamics in more detail. As we show there for the considered Ising chain, the buildup of the sharp threshold between controllable Trotter errors and quantum chaotic behavior can be observed already on experimentally relevant time scales.
While our results appear to be robust upon increasing the number of degrees of freedom, a quantitative extrapolation to N → ∞ would require the numerical study of larger systems. In this context, recent works have argued that generic periodically driven systems will eventually heat up indefinitely in the thermodynamic limit (21–23). This might leave a rather pessimistic impression, but, as we explain now, time discretization errors still remain controllable. Even in the worst-case scenario where such an indefinite heating takes place, the energy growth can still be bounded as long as the Hamiltonian has only short-ranged interactions, via
Therefore, the accuracy of DQS experiments on local observables is limited mainly by extrinsic error sources. While these may in the future be eliminated by error correction (39, 40), for relevant system sizes to solve many-body problems, full error correction is still out of reach with currently available resources. In the Supplementary Materials, we discuss in detail two typical extrinsic error sources, timing errors on individual gates and slow drifts of gate couplings over various shots of the experiment. The slow drifts turn out to be relatively benign, leading only to an effective average over an ensemble of target Hamiltonians. Individual timing errors, however, induce in the limit of small τ, a time scale beyond which the accuracy of DQS is severely affected. In addition, a realistic implementation on a physical device will suffer from other potential imperfections, many of which can be very device specific. Typical error sources include qubit decoherence and faulty pulses such as imperfect swaps between internal levels. Both of these make it highly preferable to use as few gates as possible. In view of these, our results become particularly relevant: As they show, intrinsic errors in DQS remain controlled even with relatively large Trotter steps. This makes it possible to reach a desired simulation time with a reduced number of gates, thus diminishing the influence of extrinsic errors and enhancing the accuracy in DQS for local observables.
MATERIALS AND METHODS
Numerical methods and gate sequences
The numerical data shown in this work was obtained for a quantum Ising chain with the Hamiltonian
Many of the involved contributions in this model Hamiltonian mutually commute. Therefore, only a small set of elementary quantum gates is required to simulate the Trotterized dynamics. We used the following sequence of two gates
For the presented simulations of observables, we computed the real-time evolution for 2 × 104 periods, except otherwise noted, using a Lanczos algorithm with full reorthogonalization. Because, for a finite-size system, observables still show remaining temporal fluctuations, we extracted the asymptotic long-time limit of the presented quantities by performing a stroboscopic time average over the last 104 periods. This large number of Trotter steps is far beyond realistic current-day implementations and serves here only as a worst-case scenario. However, when the Trotter errors on local observables can be controlled even in this idealized limit, one can expect the same to hold true on shorter times relevant for current experiments. In the Supplementary Materials, we illustrate in more detail how the Trotter error builds up on short time scales.
The IPR shown in Fig. 2 can, in principle, be obtained either by exact diagonalization or by use of a dynamical evolution. We chose the latter because it allows us to reach larger systems and is, in principle, an experimentally accessible approach. Dynamically, the IPR can be obtained by a stroboscopic mean
For the computation of the OTO correlator ℱ(t) defined in Eq. 6, we decompose ℱ(nτ) as
Trotter errors on local observables from perturbation theory
As mentioned before, the Trotter errors for local observables can be captured using time-dependent perturbation theory in the limit of sufficiently small τ. In the following, we outline how to obtain the analytical expressions for the coefficients qE and m for QE and ℳ, respectively. First, we consider the simulation accuracy QE and, afterward, the Trotter errors on the magnetization ℳ.
For the derivation of the corrections appearing in QE, we utilize the energy of the target Hamiltonian H, and therefore, the simulation accuracy QE exhibits a substantial overlap with the emergent conserved quantity HF
Here,
To obtain all corrections to the desired order, we first express HF using the Magnus expansion up to second order in the Trotter step size
For convenience, we restrict the presentation from now on to a sequence of two elementary gates within one period, as we have for the case of the simulated quantum Ising chain. Using the above expansion for HF in combination with the conservation of HF, one obtains for the energy deviation
As a next step, we use time-dependent perturbation theory to determine the leading order in τ corrections of ΔCv(nτ). For this purpose, we write
For the corrections to ΔE(nτ) quadratic in τ, we need to perform the time-dependent perturbation theory to first order in τ for
Let us first consider
The time integral can be conveniently evaluated by recognizing that
This gives
In the limit of n → ∞, we can use the general property that expectation values of operators are governed by the so-called diagonal ensemble (31)
For the contributions to ΔE(nτ) that are second order in τ stemming from
Collecting all contributions, we lastly obtain
For estimating the lowest-order corrections in τ for other observables such as the magnetization ℳ, we can not make direct use of the emergent conserved quantity HF as for the energy of the target Hamiltonian. Still, we can perform time-dependent perturbation theory, which we now have to carry out up to second order. Following the same steps as before, we obtain the following expression for the magnetization
Here, {A, B} = AB + BA denotes the anticommutator, and EZ is given by HZ|ψ0〉 = EZ|ψ0〉. In the limit n → ∞, we can again use that expectation values can be evaluated in the diagonal ensemble. In addition, the expression involving the time integral can be formally solved using the Lehman representation. Last, we obtain
SUPPLEMENTARY MATERIALS
Supplementary material for this article is available at http://advances.sciencemag.org/cgi/content/full/5/4/eaau8342/DC1
Section S1. Temporal buildup of the Trotter error threshold
Section S2. Trotter errors in the Ising model for alternative initial condition
Section S3. Trotter errors in the lattice Schwinger model
Section S4. Imperfections
Fig. S1. Temporal buildup of Trotter errors on transient time scales.
Fig. S2. Simulation accuracy QE for initial Neel state.
Fig. S3. IPR for the DQS of the lattice Schwinger model.
Fig. S4. Trotter errors for the DQS of the lattice Schwinger model.
Fig. S5. Timing errors in the dynamics of the simulation accuracy QE(t) for the Ising model.
Reference (41)
This is an open-access article distributed under the terms of the Creative Commons Attribution-NonCommercial license, which permits use, distribution, and reproduction in any medium, so long as the resultant use is not for commercial advantage and provided the original work is properly cited.
REFERENCES AND NOTES
- Copyright © 2019 The Authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original U.S. Government Works. Distributed under a Creative Commons Attribution NonCommercial License 4.0 (CC BY-NC).