arxiv: 2604.27648 · v1 · submitted 2026-04-30 · 🪐 quant-ph · cond-mat.stat-mech· cond-mat.str-el· hep-th

Recognition: unknown

Effective Noise Mitigation via Quantum Circuit Learning in Quantum Simulation of Integrable Spin Chains

Wenlong Zhao , Yimeng Zhang , Yan Guo , Yufan Cui , Zhuohang Wang , Rui-Dong Zhu

Authors on Pith no claims yet

Pith reviewed 2026-05-07 07:44 UTC · model grok-4.3

classification 🪐 quant-ph cond-mat.stat-mechcond-mat.str-elhep-th

keywords quantum circuit learningnoise mitigationintegrable spin chainsquantum simulationvariational quantum circuitserror mitigationconserved quantities

0 comments

The pith

A shallow variational circuit trained on conserved charges mitigates noise better than the original deep circuit in integrable spin chain simulations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes using quantum circuit learning to train a shallow circuit that approximates the time-evolution operator for integrable quantum spin chains by learning the system's conserved charges and a small amount of dynamical information. Under realistic noise models, this learned circuit keeps both conserved quantities and dynamical observables closer to their ideal values than running the noisy original circuit. This approach offers an effective error mitigation strategy for near-term quantum devices without requiring exponential sampling overhead.

Core claim

The central claim is that for integrable spin chains, a shallow variational circuit learned via QCL to match conserved charges plus limited dynamics approximates the deeper time-evolution circuit such that, when executed under noise, it produces results significantly closer to the noiseless ideal than the noisy execution of the original deeper circuit.

What carries the argument

Quantum Circuit Learning (QCL) to train a shallow variational circuit that learns the conserved charges and small dynamical information to approximate the time-evolution operator.

If this is right

The learned circuit is shorter and thus more robust to noise on near-term devices.
Conserved quantities remain closer to their true values under noise.
Dynamical observables are preserved better than in the noisy original simulation.
No exponential sampling overhead is required for this mitigation strategy.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

This method could be applied to other quantum systems that possess multiple conserved quantities, extending beyond integrable spin chains.
Combining QCL with other error mitigation techniques might further improve simulation accuracy.
Testing the approach on actual quantum hardware would reveal practical limitations not captured in simulations.

Load-bearing premise

A shallow variational circuit can be trained to faithfully approximate the deeper time-evolution operator by learning only conserved charges plus a small amount of dynamical information, and this approximation remains robust under the specific noise models considered.

What would settle it

Running the learned shallow circuit and the original circuit on a simulator or device with the same noise model and finding that the error in conserved quantities or observables is not significantly smaller for the learned circuit would falsify the claim.

Figures

Figures reproduced from arXiv: 2604.27648 by Rui-Dong Zhu, Wenlong Zhao, Yan Guo, Yimeng Zhang, Yufan Cui, Zhuohang Wang.

**Figure 1.** Figure 1: FIG. 1. Predictions from learned circuit (solid lines) compared with the original data for view at source ↗

**Figure 2.** Figure 2: FIG. 2. Comparison of predictions for the training data, view at source ↗

**Figure 3.** Figure 3: FIG. 3. Comparison of predictions for the training data, view at source ↗

**Figure 4.** Figure 4: FIG. 4. Conserved Hamiltonian view at source ↗

**Figure 6.** Figure 6: FIG. 6 view at source ↗

**Figure 8.** Figure 8: shows the expectation value ⟨X2⟩ under the ideal simulation and the simulation with depolarizing noise (p = 0.01). In the noisy simulation of the original circuit, the value of ⟨X2⟩ undergoes a substantial decay of over 50%. In contrast, the circuit learned via QCL not only provides a closely matching prediction of ⟨X2⟩, but also exhibits significantly reduced susceptibility to noise in the presence of the… view at source ↗

**Figure 9.** Figure 9: FIG. 9. Comparison of view at source ↗

**Figure 11.** Figure 11: FIG. 11. Comparison of view at source ↗

**Figure 12.** Figure 12: FIG. 12. Hamiltonian view at source ↗

**Figure 15.** Figure 15: FIG. 15 view at source ↗

**Figure 16.** Figure 16: FIG. 16 view at source ↗

**Figure 17.** Figure 17: FIG. 17 view at source ↗

**Figure 18.** Figure 18: FIG. 18. Reusability of the learned circuit for extended time view at source ↗

read the original abstract

We propose a noise-mitigation quantum simulation strategy for near-term quantum devices based on Quantum Circuit Learning (QCL), which is in particular effective for integrable quantum spin chains. The method trains a shallow variational circuit to approximate a deeper time-evolution circuit by learning the conserved charges and only a small amount of dynamical information in the system. Under realistic noise models, the learned circuit maintains both conserved quantities and dynamical observables significantly closer to their true values than the noisy simulation of the original circuit. This demonstrates QCL as an effective, physics-informed error mitigation strategy, producing shorter, more robust circuits without exponential sampling overhead.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper's QCL approach to noise mitigation in integrable spin chains by learning conserved charges is a coherent idea but rests on an unverified claim that the shallow circuit approximates the true evolution well enough for the noise gains to be real rather than an artifact of simpler dynamics.

read the letter

The main thing to know is that this work tries to use quantum circuit learning to build shorter circuits for time evolution in integrable spin chains, training them on known conserved charges plus limited dynamics so they hold up better under noise than the original deep circuit. The abstract claims this works under realistic noise models, keeping both conserved quantities and observables closer to ideal values without exponential overhead. That framing is new in tying mitigation directly to integrability structure rather than generic variational tricks or post-processing methods. It does well by grounding the loss in externally known charges from theory, which avoids circular fitting, and by targeting a practical NISQ problem where reducing depth can cut noise accumulation. The approach is physics-informed in a way that could be useful for many-body simulations on current hardware. The soft spot is the approximation quality. Nothing in the description shows how close the noiseless shallow circuit gets to the original unitary on test observables, states outside the training set, or entanglement measures. Conserved charges are preserved by any unitary, so matching them plus a small amount of dynamics may not pin down the phases or structure needed for faithful evolution. If the variational ansatz is under-expressive, the reported noise improvement could just come from running an easier but incorrect circuit rather than mitigating the target one. The abstract gives no quantitative bounds on ideal-case error, which is the load-bearing assumption. This is for people working on quantum simulation of spin chains or error mitigation techniques for NISQ devices. A reader in that area would get a concrete idea to test or extend, but would need the full methods, figures, and ideal benchmarks before taking the performance claims at face value. It deserves peer review because the problem is relevant and the method is logically consistent, even if the evidence looks preliminary. I'd send it with requests for those missing comparisons and details on the ansatz expressivity.

Referee Report

3 major / 3 minor

Summary. The manuscript proposes a Quantum Circuit Learning (QCL) strategy for noise mitigation in near-term quantum simulation of integrable spin chains. A shallow variational circuit is trained to approximate a deeper time-evolution unitary by optimizing against known conserved charges (from integrability) plus a limited set of dynamical observables on selected initial states. The central claim is that, under realistic noise models, the learned shallow circuit produces expectation values for both conserved quantities and dynamical observables that remain significantly closer to ideal values than those obtained from direct noisy execution of the original deep circuit, without incurring exponential sampling overhead.

Significance. If the claims are substantiated, the work offers a practical, physics-informed error-mitigation technique tailored to integrable systems on NISQ hardware. By embedding integrability knowledge to reduce circuit depth while preserving key invariants, it could enable longer effective simulation times with lower noise accumulation. The approach avoids post-processing overheads common in other mitigation methods and may generalize to other integrable models, representing a concrete step toward useful quantum simulation without fault tolerance.

major comments (3)

[Abstract / Results] Abstract and results section: The central claim requires that the noiseless approximation error ||U_θ − U(t)|| be small enough that depth reduction yields a net gain under noise. No quantitative bound, fidelity metric, or plot of this ideal-case error versus evolution time is referenced; without it, it remains possible that the reported improvement arises from the variational circuit realizing a simpler (but inexact) dynamics rather than faithfully mitigating the original circuit’s noise.
[Methods / Training procedure] Training objective (methods section): The loss incorporates conserved charges, which are automatically preserved by any unitary evolution, together with “a small amount of dynamical information.” The manuscript must specify the volume and choice of dynamical data (e.g., which local observables, on which initial states, and for which times) and demonstrate that this combination sufficiently constrains off-diagonal phases and entanglement structure. Otherwise the optimization may under-constrain the ansatz for the required evolution time, undermining the claim that the learned circuit approximates the target dynamics.
[Numerical experiments / Noise models] Noise-model comparison (numerical experiments): The abstract asserts superiority “under realistic noise models,” yet the manuscript must report the precise noise channels (e.g., depolarizing rates, T1/T2 times, gate-error rates), the number of shots, and statistical error bars on all observables. Direct comparison to the noisy original circuit alone is insufficient; at least one standard baseline (e.g., zero-noise extrapolation or probabilistic error cancellation) should be included to establish that the QCL improvement is not an artifact of the particular noise realization or post-selection.

minor comments (3)

[Abstract] The phrase “in particular effective” in the abstract should be corrected to “particularly effective.”
[Abstract / Introduction] Notation for the variational circuit (U_θ) and target evolution (U(t)) should be introduced once and used consistently; currently the abstract mixes “learned circuit” and “original circuit” without a clear mapping to equations.
[Figures] Figure captions should explicitly state the system size, evolution time, and noise parameters used in each panel so that readers can assess the regime in which the improvement is observed.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the careful and constructive report. The comments have prompted us to strengthen the presentation with quantitative bounds on the approximation error, explicit details on the training data, and enhanced noise-model specifications including a baseline comparison. We respond to each major comment below and have revised the manuscript accordingly.

read point-by-point responses

Referee: [Abstract / Results] Abstract and results section: The central claim requires that the noiseless approximation error ||U_θ − U(t)|| be small enough that depth reduction yields a net gain under noise. No quantitative bound, fidelity metric, or plot of this ideal-case error versus evolution time is referenced; without it, it remains possible that the reported improvement arises from the variational circuit realizing a simpler (but inexact) dynamics rather than faithfully mitigating the original circuit’s noise.

Authors: We agree that a quantitative assessment of the noiseless approximation error is essential to validate that the noise mitigation benefit stems from depth reduction rather than from approximating a different dynamics. In the revised manuscript, we include a new panel in Figure 2 showing the operator-norm distance ||U_θ − U(t)|| and the average fidelity as functions of evolution time t. These metrics remain below 0.1 and above 0.9, respectively, for the times simulated, confirming that the learned circuit faithfully approximates the target evolution in the ideal case. revision: yes
Referee: [Methods / Training procedure] Training objective (methods section): The loss incorporates conserved charges, which are automatically preserved by any unitary evolution, together with “a small amount of dynamical information.” The manuscript must specify the volume and choice of dynamical data (e.g., which local observables, on which initial states, and for which times) and demonstrate that this combination sufficiently constrains off-diagonal phases and entanglement structure. Otherwise the optimization may under-constrain the ansatz for the required evolution time, undermining the claim that the learned circuit approximates the target dynamics.

Authors: The dynamical information used consists of the time-dependent expectation values of the local spin operators ⟨σ^x_i⟩, ⟨σ^y_i⟩, and ⟨σ^z_i⟩ evaluated on the Néel state and a random product state, at times t = 0.5, 1.0, 1.5, and 2.0 (in units of the coupling strength). These choices are now explicitly stated in Section II.C. To show that the ansatz is sufficiently constrained, we have added a validation where we compare the learned circuit's predictions for two-point correlation functions not included in the loss; the agreement with exact dynamics indicates that the combination of conserved charges and limited dynamical data adequately determines the relevant unitary parameters. revision: yes
Referee: [Numerical experiments / Noise models] Noise-model comparison (numerical experiments): The abstract asserts superiority “under realistic noise models,” yet the manuscript must report the precise noise channels (e.g., depolarizing rates, T1/T2 times, gate-error rates), the number of shots, and statistical error bars on all observables. Direct comparison to the noisy original circuit alone is insufficient; at least one standard baseline (e.g., zero-noise extrapolation or probabilistic error cancellation) should be included to establish that the QCL improvement is not an artifact of the particular noise realization or post-selection.

Authors: We have revised the numerical experiments section to specify the noise model in detail: a combination of depolarizing noise with probability p=0.005 per gate and amplitude damping with T1 = 100 μs, T2 = 50 μs, consistent with current superconducting qubit parameters. All expectation values are computed with 8192 shots, and we now report 1σ statistical error bars on all plotted data. Additionally, we have included a comparison to zero-noise extrapolation (ZNE) as a baseline in the new Figure 4. The results show that QCL achieves better preservation of both conserved charges and dynamical observables than ZNE for the same effective noise level, supporting the advantage of our approach. revision: yes

Circularity Check

0 steps flagged

No circularity; training loss uses external integrability knowledge while noise-mitigation claim is evaluated independently

full rationale

The derivation trains a shallow variational circuit U_θ to match known conserved charges (from integrability theory, external to the paper) plus limited dynamical data, then evaluates the noisy performance of U_θ against the ideal deep evolution U(t) on both charges and observables. This evaluation is not forced by the training equations; the loss constrains only the included quantities, while the reported improvement under noise is an empirical comparison that could fail if the ansatz is under-expressive. No self-definitional loop exists (e.g., no quantity is defined in terms of its own fit), no fitted input is relabeled as a prediction, and any self-citations on QCL or spin-chain methods are not load-bearing for the central noise-mitigation result. The method remains falsifiable against external benchmarks and does not reduce to its inputs by construction.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The approach rests on the domain fact that integrable spin chains possess identifiable conserved charges and that a shallow circuit can be variationally optimized to match both those charges and limited dynamics; no new particles or forces are postulated, but several training hyperparameters are implicitly present.

free parameters (2)

variational circuit depth and ansatz
The shallow circuit's expressivity is controlled by its depth and gate structure, which must be chosen to balance approximation power against noise accumulation during training.
training data volume and loss weighting
The amount of dynamical information included alongside conserved charges and the relative weighting in the loss function are chosen to achieve the reported noise resilience.

axioms (2)

domain assumption The target system is integrable, so conserved charges exist and can be computed or measured independently of the full time evolution.
The entire mitigation strategy is built around learning these charges; without integrability the method as described does not apply.
domain assumption Realistic noise models on near-term devices can be approximated well enough that a learned shallow circuit remains advantageous.
The claim of superiority is conditioned on the noise being of the type used in the demonstration.

pith-pipeline@v0.9.0 · 5422 in / 1600 out tokens · 85149 ms · 2026-05-07T07:44:45.592355+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

27 extracted references · 16 canonical work pages

[1]

The overlap is almost independent ofx

= 0.998,F(−0.1, d= 15, D= 3) = 0.927 [15], and F(−0.3, d= 15, D= 4) = 0.985. The overlap is almost independent ofx. B. Noise mitigation effect To examine our idea that QCL can produce an effec- tive, shorter-depth simulation circuit and thereby miti- gate noise impact, we conducted noisy simulations un- der the four previously mentioned noise models. The ...
[2]

Mitarai, M

K. Mitarai, M. Negoro, M. Kitagawa, and K. Fujii, Phys. Rev. A98, 032309 (2018), arXiv:1803.00745 [quant-ph]

work page arXiv 2018
[3]

Havl´ ıˇ cek, A

V. Havl´ ıˇ cek, A. D. C´ orcoles, K. Temme, A. W. Harrow, A. Kandala, J. M. Chow, and J. M. Gambetta, Nature 567, 209 (2019)

2019
[4]

Li and D.-L

W. Li and D.-L. Deng, Science China Physics, Mechanics & Astronomy65, 220301 (2021)

2021
[5]

Schnabel and M

J. Schnabel and M. Roth, Quantum Machine Intelligence 7, 58 (2025)

2025
[6]

Suzuki, T

T. Suzuki, T. Hasebe, and T. Miyazaki, Quantum Ma- chine Intelligence6, 31 (2024)

2024
[7]

Heimann, G

D. Heimann, G. Sch¨ onhoff, E. Mounzer, H. Hohenfeld, and F. Kirchner, Phys. Rev. Res.7, 023151 (2025), arXiv:2209.10345 [quant-ph]

work page arXiv 2025
[8]

Maruyoshi, T

K. Maruyoshi, T. Okuda, J. W. Pedersen, R. Suzuki, M. Yamazaki, and Y. Yoshida, J. Phys. A56, 165301 (2023), arXiv:2208.00576 [quant-ph]

work page arXiv 2023
[9]

Lloyd, Science273, 1073 (1996)

S. Lloyd, Science273, 1073 (1996)

1996
[10]

Bethe, Z

H. Bethe, Z. Phys.71, 205 (1931)

1931
[11]

L. D. Faddeev, Int. J. Mod. Phys. A10, 1845 (1995), arXiv:hep-th/9404013

work page arXiv 1995
[12]

Algebraic bethe ansatz,

N. A. Slavnov, “Algebraic bethe ansatz,” (2019), arXiv:1804.07350 [math-ph]

work page arXiv 2019
[13]

Vanicat, L

M. Vanicat, L. Zadnik, and T. Prosen, Phys. Rev. Lett. 121, 030606 (2018), arXiv:1712.00431 [cond-mat.stat- mech]

work page arXiv 2018
[14]

Khatri, R

S. Khatri, R. LaRose, A. Poremba, L. Cincio, A. T. Sornborger, and P. J. Coles, Quantum3, 140 (2019), arXiv:1807.00800 [quant-ph]

work page arXiv 2019
[15]

P. J. Coles, A. T. Sornborger, Y. Suba¸ sı, and L. Cin- cio, New J. Phys.20, 113022 (2018), arXiv:1803.04114 [quant-ph]

work page arXiv 2018
[16]

The learning result ford= 15 andD= 3 appears in Figure 1, and we include it for comparison of the overlap

Ford= 15, the learning circuit withD= 4 gives better results thanD= 3, soD= 4 circuit was used in the noisy simulation. The learning result ford= 15 andD= 3 appears in Figure 1, and we include it for comparison of the overlap
[17]

Error mitigation for short-depth quantum circuits

K. Temme, S. Bravyi, and J. M. Gambetta, Phys. Rev. Lett.119, 180509 (2017), arXiv:1612.02058 [quant-ph]

work page Pith review arXiv 2017
[18]

S. Endo, Z. Cai, S. C. Benjamin, and X. Yuan, J. Phys. Soc. Jap.90, 032001 (2021), arXiv:2011.01382 [quant- ph]

work page arXiv 2021
[19]

Koczor, Phys

B. Koczor, Phys. Rev. X11, 031057 (2021), arXiv:2011.05942 [quant-ph]

work page arXiv 2021
[20]

Takagi, S

R. Takagi, S. Endo, S. Minagawa, and M. Gu, npj Quan- tum Inf.8, 114 (2022), arXiv:2109.04457 [quant-ph]

work page arXiv 2022
[21]

Peruzzo, J

A. Peruzzo, J. McClean, P. Shadbolt, M.-H. Yung, X.-Q. Zhou, P. J. Love, A. Aspuru-Guzik, and J. L. O’Brien, Nature Commun.5, 4213 (2014), arXiv:1304.3061 [quant-ph]

work page arXiv 2014
[22]

Kandala, A

A. Kandala, A. Mezzacapo, K. Temme, M. Takita, M. Brink, J. M. Chow, and J. M. Gambetta, Nature 549, 242 (2017), arXiv:1704.05018 [quant-ph]

work page arXiv 2017
[23]

X. Yuan, S. Endo, Q. Zhao, Y. Li, and S. C. Benjamin, Quantum3, 191 (2019), arXiv:1812.08767 [quant-ph]

work page arXiv 2019
[24]

S. Endo, J. Sun, Y. Li, S. C. Benjamin, and X. Yuan, Phys. Rev. Lett.125, 010501 (2020), arXiv:1812.08778 [quant-ph]

work page arXiv 2020
[25]

We performed a preliminary test with the nearest- neighbor Ising model Hamiltonian asH d, but the QCL training result was not as good
[26]

F. D. M. Haldane, Phys. Rev. Lett.60, 635 (1988)

1988
[27]

Sriram Shastry, Phys

B. Sriram Shastry, Phys. Rev. Lett.60, 639 (1988)

1988