acceptodds
Under review as a conference paper at ICLR 2027

Learning to Regularize Variational Quantum Dynamics under Distribution Shift

Abstract

Variational quantum imaginary-time evolution (VarQITE) uses the quantum geometric tensor (QGT) to convert imaginary-time evolution into parameter-space updates, but solving the resulting QGT linear systems is sensitive to ill-conditioning, estimation error, and changes in problem geometry. We study whether regularization itself can be transferred across such changes. A single absolute Tikhonov value and a single spectrum-scaled Tikhonov value are selected on source validation and then frozen for every confirmatory target; the family-wise validation optima, which span more than six orders of magnitude, are used only as a diagnostic of heterogeneity. We introduce guarded learned regularization, in which a source-trained controller maps task-local QGT/force descriptors to either scalar damping or mode-wise spectral filtering, while a deterministic resolved-subspace verifier can revert an inadmissible proposal to a trusted QNG–LM reference. Across seven frozen non-floor confirmatory regimes under the predeclared full-source/leave-one-family training protocols, SpectralGuard attains the lowest descriptive method-level worst-case relative regret, 20.9%, without target-side hyperparameter selection. Under strong synthetic estimator perturbations, however, ScalarGuard is more stable and the spectral verifier falls back to the reference on most updates, exposing a clear expressivity–robustness boundary rather than uniform spectral superiority. The QEC branch should be read as transfer across stabilizer-derived Hamiltonian families, not as a decoding or fault-tolerance result. Overall, the evidence supports transferable regularization with explicit numerical verification, while also identifying the regimes in which learned spectral adaptation ceases to be the dominant deployed action.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
variational quantum algorithms, quantum geometric tensor, learned optimization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_sBu4OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

In VarQITE, each step solves G_t v_t = f_t with the quantum geometric tensor G_t, and the regularization choice is usually retuned per problem. The paper frames this as regularization transfer: learn, from source QGT/force systems, a map from local spectral descriptors to a regularization action, then deploy it on shifted targets without target-side tuning. It proposes two controllers. ScalarGuard predicts a log-residual on a fixed Tikhonov anchor λ_ref = 10⁻³. SpectralGuard predicts bounded per-mode filter gates. A deterministic verifier falls back to a QNG–Levenberg–Marquardt reference when a learned proposal scores worse on a resolved-subspace McLachlan model. Source tasks are random k-local, Sherrington–Kirkpatrick, and QEC-derived diagonal-Ising Hamiltonians. The evaluation covers clean in-distribution, held-out family/instance, size, depth, and synthetic estimator-noise shifts. The main established result is that SpectralGuard beats QNG–LM on pooled clean structural shifts. Descriptively, SpectralGuard also has the smallest worst-case relative regret (20.9%).

Critical strengths

None

Critical weaknesses

The scope is limited, and the ML contribution is modest for ICLR. Systems have ≤ 18 qubits, the simulations are statevector, the noise is synthetic, and the QEC families reduce to diagonal (classical) Ising Hamiltonians on which every method reaches the floor. The learning component is a small controller over spectral features trained by imitation. With effect sizes of about 0.006 infidelity against the weakest baseline, both the practical and the methodological contribution are thin.

Questions for authors

What does regularization actually correct in the clean regimes? Without estimation noise, the minimum-norm McLachlan solution is already correct. So what is the controller learning there: finite-step nonlinearity, ill-conditioning, or something else? And why should a fixed map from spectral features to gates transfer across Hamiltonian families?

4 comments
AuthorsOpenReview ↗

Response to Reviewer sBu4 (Part 1/4)

We thank the reviewer for the careful assessment and constructive suggestions. We appreciate the valuable suggestions regarding the machine-learning contribution, the role of regularization in the clean-QGT regime, the transferability of the learned policy across Hamiltonian families, and the experimental scale and validation methodology. We have further strengthened the manuscript through additional theoretical analysis, targeted experiments, expanded comparisons, and corresponding revisions to the main text and supplementary material. Our point-by-point responses are provided below.

1. Machine-learning contribution

We thank the reviewer for raising the question of whether the learning component constitutes a substantive machine-learning contribution beyond conventional numerical regularization. This is an important point for positioning the work.

The central ML problem in this study is amortized, conditional regularization under distribution shift. Along variational quantum dynamics, the appropriate numerical action changes with the local quantum geometry, force orientation, and trajectory state. The learning problem is therefore to infer this action from source experience and to preserve useful behavior when the Hamiltonian family, circuit dimension, or circuit depth changes. This is distinct from identifying a single globally tuned regularization parameter for a fixed numerical problem.

There is a close methodological precedent in Learning to Relax (Khodak et al., 2024), which formulates parameter selection for a sequence of related linear systems as a learning problem and studies the transition from a globally fixed solver parameter to a context-dependent policy. More broadly, this perspective is consistent with amortized optimization (Amos, 2023), learned optimization across tasks (Andrychowicz et al., 2016; Wichrowska et al., 2017), and learning-augmented iterative numerical methods (Arisaka & Li, 2023). In inverse problems specifically, Afkham et al. (2021) learn instance-dependent regularization parameters rather than solving a new parameter-selection problem for every observation. These works establish the relevant methodological setting: repeated numerical decisions with shared structure can themselves define a learning and generalization problem.

Within that setting, our contribution is more specific than scalar hyperparameter prediction. The learned object is a structured spectral decision rule whose hypothesis class is shared across modes and parameter dimensions, while its supervision is defined by the downstream finite-step quantum-state consequence of the update. These two choices are coupled: dimension-shared spectral structure makes frozen transfer across different QGT sizes well-defined, while trajectory-level supervision makes the learned decision depend on the effect of regularization on executed variational dynamics rather than solely on an algebraic inverse criterion. The subsequent short-horizon rollout refinement is therefore methodologically relevant: the training procedure uses imitation to initialize a policy that is further optimized for finite-step trajectory behavior.

The transfer question is also empirically nontrivial. Source validation exhibits approximately 6.5 orders of magnitude variation in preferred scalar regularization across Hamiltonian families, while family identity is unavailable to the learned policy. Under the frozen evaluation protocol, the same learned spectral rule improves over the condition-adaptive QNG–LM reference in each of the four non-floor clean structural shifts. These observations connect the learning formulation to the quantity being evaluated: whether regularization experience acquired from source quantum-geometric systems can be reused when the inverse systems themselves change.

We have revised the manuscript to state this learning problem explicitly and to position SpectralGuard relative to amortized optimization, learned regularization, and learning-augmented numerical solvers, separately from questions of architectural complexity.

AuthorsOpenReview ↗

Response to Reviewer sBu4 (Part 2/4)

2. What does regularization correct in the clean-QGT regime?

An exact QGT determines the instantaneous McLachlan projection, while the executed variational update is evaluated through its finite-step state fidelity and numerical stability. Our analytic one-qubit example demonstrates that, even with a perfectly conditioned QGT (G=1G=1), two states with identical instantaneous McLachlan velocities can have different finite-step fidelity-optimal velocities.

This distinction is also supported empirically. On 20 matched Random/SK tasks, at h=1/32h=1/32, the learned update has a higher instantaneous McLachlan objective than source-fixed ridge (Δq=+0.034372\Delta q=+0.034372), yet achieves lower one-step infidelity (ΔI=−0.000186763\Delta I=-0.000186763), with this ranking reversal observed in 20/20 task averages. Separately, in 30 exact-QGT trajectory experiments, the unregularized pseudoinverse triggers numerical failure in 20 tasks; at h=1/32h=1/32, its failure-penalized infidelity is 0.66667, compared with 0.01700 for SpectralGuard and 0.01543 for fixed ridge.

These results identify two distinct mechanisms: finite-step state-fidelity mismatch and numerical amplification of weak QGT modes. Together, they establish the numerical role of regularization even when the QGT is exactly available, and provide a direct motivation for learning regularization decisions from finite-step state accuracy.

We have added the analytic example and exact-QGT trajectory experiments to Section 3.1 and Table 1, with the complete matched-state analysis provided in Supplementary Section H.1.

3. Why should the learned rule transfer across Hamiltonian families?

We thank the reviewer for raising this important question. The relevant transfer hypothesis is that regularization depends more directly on the local numerical geometry of the inverse problem than on Hamiltonian identity itself. This is consistent with data-driven numerical linear algebra and cross-instance algorithm configuration, where solver decisions can generalize when conditioned on structural information shared across problem instances (Khodak et al., 2024; Balcan et al., 2022; Bartlett et al., 2022).

We tested this hypothesis directly. In the prespecified family-exclusion benchmark, the evaluated Hamiltonian family is removed from both training and validation. SpectralGuard improves over condition-adaptive QNG–LM on all four non-floor structural shifts, with the 118-group pooled result decreasing trajectory infidelity from 0.05191 to 0.04585 (Δ=−0.00606\Delta=-0.00606, 95% CI [−0.00716,−0.00504][-0.00716,-0.00504]).

We further added a targeted post-review experiment on 40 entirely new size/depth-OOD groups. To separate transferred spectral direction from step-size effects, every comparator is rescaled to the same local QGT/Fubini–Study speed. At h=1/32h=1/32, SpectralGuard obtains 5.89×10−45.89\times10^{-4} one-step infidelity versus 9.41×10−49.41\times10^{-4} for matched fixed ridge, a 37.4% reduction, with 37/40 groups favorable. Positive-spectrum TSVD gives 6.35×10−46.35\times10^{-4}, while source-selected adaptive LM and a force-aware rule obtain 5.31×10−45.31\times10^{-4} and 5.36×10−45.36\times10^{-4}, respectively. These matched-speed comparisons provide direct evidence that the transferred policy encodes useful directional information beyond step-size control, and characterize its behavior relative to strong adaptive alternatives.

Together, the family-exclusion and matched-speed experiments establish two complementary aspects of transfer: the frozen policy retains its effectiveness across structural distribution shifts, and its learned spectral direction contributes to finite-step accuracy under controlled update speeds. We have added the family-exclusion transfer evidence to Section 4.1 and the new Fubini–Study-speed-matched experiment to Section 4.3, with the full protocol reported in Supplementary Section H.2.

AuthorsOpenReview ↗

Response to Reviewer sBu4 (Part 3/4)

4. Experimental scale, numerical simulation, and estimator noise

We thank the reviewer for the thoughtful and constructive suggestion concerning experimental scale, numerical simulation, and measurement realism. Our experimental setting is designed around the central question of this work: whether finite-step-informed spectral regularization can be learned from source quantum-geometric inverse problems and transferred to unseen Hamiltonians and circuit geometries. This setting enables a controlled evaluation of the relationship between quantum geometry, regularization decisions, and finite-step state accuracy.

Regarding system scale, our training systems contain 6–12 qubits, while the size-OOD evaluation extends to 14–18 qubits, with additional independent shifts in circuit depth. These ranges provide a meaningful test of quantum-geometric extrapolation, particularly because circuit geometry, parameter redundancy, and the conditioning of the quantum geometric tensor evolve with system size and ansatz depth (McArdle et al., 2019; Stokes et al., 2020; Gomes et al., 2021; Haug et al., 2021). On 30 independent larger-system groups, SpectralGuard achieves a mean trajectory infidelity of 0.08293 versus 0.09547 for the condition-adaptive QNG–LM reference, representing a 13.1% relative reduction with a paired 95% confidence interval of [−0.01434,−0.01084][-0.01434,-0.01084]. These results demonstrate the effectiveness of the learned regularization policy under increased system size and previously unseen quantum geometries.

Larger-scale variational quantum dynamics introduces a related but distinct challenge: the measurement and computational resources required to construct and evolve the quantum-geometric system. Accordingly, recent scalable approaches employ specialized formulations, including measurement-efficient metric estimation (van Straaten & Koczor, 2021), QGT-free dual evolution (Gacon et al., 2024), random-measurement imaginary-time evolution (Kolotouros et al., 2025), and operator-projected dynamics (Anuar et al., 2026). These developments primarily address resource-efficient quantum evolution, whereas our work focuses on learning transferable regularization decisions for the QGT-based inverse problem. The two directions therefore investigate complementary aspects of variational quantum dynamics.

Regarding statevector simulation, exact state references allow us to evaluate the finite-step consequences of different regularization decisions under matched initial conditions, identical QGT/force inputs, and controlled distribution shifts. This follows the established numerical evaluation methodology in variational quantum dynamics (McArdle et al., 2019; Gomes et al., 2021; Gacon et al., 2024). A corresponding hardware study would additionally require sufficiently accurate QGT/force measurements, reproducible comparisons across multiple Hamiltonian families and circuit configurations, and reliable reference-state characterization throughout complete trajectories. These requirements motivate a dedicated experimental design that integrates measurement efficiency, device calibration, and state-characterization protocols. Our controlled numerical evaluation directly isolates the transferable regularization mechanism and provides a rigorous basis for that subsequent experimental development.

Regarding estimator noise, our perturbations are applied directly to the estimated QGT and force, providing controlled tests of the numerical quantities governing variational updates. This experimental interface is motivated by established analyses of quantum-geometric measurement uncertainty and its influence on variational dynamics (van Straaten & Koczor, 2021; Koczor & Benjamin, 2022; Zhang et al., 2026). We have further strengthened the evidence through two additional comparisons. Table 2 reports results on 108 clean structural-OOD groups and 16 moderate-noise OOD groups. Under moderate estimator perturbations, the frozen SpectralGuard policy without verification achieves 0.05286 versus 0.09543 mean trajectory infidelity for raw-QGT TSVD, a 44.6% reduction. Table 3 additionally examines severe estimator perturbations. At noise strengths 0.003 and 0.01, the verification mechanism reduces mean infidelity by 30.8% and 39.2%, respectively, with the associated fallback behavior explicitly characterized. These experiments provide complementary evidence for the learned spectral response and the numerical verification mechanism across different estimator regimes.

AuthorsOpenReview ↗

Response to Reviewer sBu4 (Part 4/4)

We find the reviewer's suggestion of extending this investigation to substantially larger systems and physical quantum hardware particularly interesting and constructive. We are highly interested in pursuing a dedicated large-scale formulation that combines scalable quantum-geometric estimation, resource-aware spectral regularization, and hardware-specific validation. The revised manuscript now presents the larger-system results in Table 1, the additional clean and moderate-noise comparisons in Table 2, and the severe-noise verification results in Table 3, strengthening the experimental evidence for finite-step-aware spectral regularization and its transfer across quantum-geometric distribution shifts.

References

Afkham, B. M., Chung, J., & Chung, M. (2021). Learning Regularization Parameters of Inverse Problems via Deep Neural Networks. Inverse Problems, 37(10), 105017.

Amos, B. (2023). Tutorial on Amortized Optimization. Foundations and Trends in Machine Learning, 16(5), 592–732.

Andrychowicz, M., et al. (2016). Learning to Learn by Gradient Descent by Gradient Descent. Advances in Neural Information Processing Systems (NeurIPS 2016).

Anuar, A. A., et al. (2026). Operator-Projected Variational Quantum Imaginary Time Evolution. Journal of Physics A: Mathematical and Theoretical, 59, 245301.

Arisaka, S., & Li, Q. (2023). Principled Acceleration of Iterative Numerical Methods Using Machine Learning. Proceedings of the 40th International Conference on Machine Learning (ICML 2023), PMLR 202, 1041–1059.

Balcan, M.-F., et al. (2022). Provably Tuning the ElasticNet Across Instances. Advances in Neural Information Processing Systems (NeurIPS 2022).

Bartlett, P., Indyk, P., & Wagner, T. (2022). Generalization Bounds for Data-Driven Numerical Linear Algebra. Proceedings of the 35th Conference on Learning Theory (COLT 2022), PMLR 178, 2013–2040.

Gacon, J., et al. (2024). Variational Quantum Time Evolution Without the Quantum Geometric Tensor. Physical Review Research, 6, 013143.

Gomes, N., et al. (2021). Adaptive Variational Quantum Imaginary Time Evolution Approach for Ground State Preparation. Advanced Quantum Technologies, 4(12), 2100114.

Haug, T., Bharti, K., & Kim, M. S. (2021). Capacity and Quantum Geometry of Parametrized Quantum Circuits. PRX Quantum, 2, 040309.

Khodak, M., et al. (2024). Learning to Relax: Setting Solver Parameters Across a Sequence of Linear System Instances. International Conference on Learning Representations (ICLR 2024).

Koczor, B., & Benjamin, S. C. (2022). Quantum Natural Gradient Generalized to Noisy and Nonunitary Circuits. Physical Review A, 106, 062416.

Kolotouros, I., Joseph, D., & Narayanan, A. K. (2025). Accelerating Quantum Imaginary-Time Evolution with Random Measurements. Physical Review A, 111, 012424.

McArdle, S., et al. (2019). Variational Ansatz-Based Quantum Simulation of Imaginary Time Evolution. npj Quantum Information, 5, 75.

Stokes, J., Izaac, J., Killoran, N., & Carleo, G. (2020). Quantum Natural Gradient. Quantum, 4, 269.

van Straaten, B., & Koczor, B. (2021). Measurement Cost of Metric-Aware Variational Quantum Algorithms. PRX Quantum, 2, 030324.

Wichrowska, O., et al. (2017). Learned Optimizers That Scale and Generalize. Proceedings of the 34th International Conference on Machine Learning (ICML 2017), PMLR 70, 3751–3760.

Zhang, F., et al. (2026). Sampling Noise and Optimized Measurement Distribution in Imaginary-Time Quantum Dynamics Simulations. Advanced Quantum Technologies, 9(9), e70418.

Reviewer_h2FcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method for learning regularization rules for linear inverse problems in variational quantum dynamics and transferring them across different problems. Specifically, the authors propose ScalarGuard, which predicts the parameters of Tikhonov regularization, and SpectralGuard, which learns spectral filters for each eigenvector. The authors also introduce Verifier, which assesses the validity of the learned updates. In the experiments, the authors evaluate the performance across different quantum systems and under distribution shifts.

Critical strengths
  • The authors are addressing the practically important problem of the transferability of regularization schemes in linear inverse problems.
  • The design, which combines adaptive regularization through learning with a fallback to conventional methods, is intriguing.
  • The authors evaluate the method under various distribution shifts and noise conditions, analyzing not only the effectiveness of the proposed method but also its limitations.
Critical weaknesses
  • The essence of the proposed method lies in learning-based regularization for general linear inverse problems, and it appears to make little use of techniques that actively exploit structures specific to quantum computing. The differences from existing learning-based regularization methods are also not sufficiently clear.
  • The training data, loss function, and learning procedures for the neural network are difficult to understand from the main text, making it a challenging read for machine learning researchers.
  • The paper focuses primarily on comparisons with conventional methods that use fixed regularization parameters, and lacks sufficient comparison with existing methods that adaptively select regularization based on condition numbers and spectra.
Questions for authors
  • This paper is structured primarily around knowledge of quantum computing, making it somewhat difficult for machine learning researchers to read. While the proposed method appears applicable to general linear inverse problems, what design innovations utilize properties specific to quantum computing? Furthermore, what is the technical novelty of this method compared to existing learning-based regularization techniques?
  • Regarding the neural network training method, please provide a clearer explanation in the main text of the training data generation method, the loss function, and the optimization procedure.
  • Is it possible to compare this method with non-learning-based adaptive regularization techniques, such as adaptive Tikhonov regularization based on condition numbers and eigenvalue distributions, or Truncated SVD? Please demonstrate that learning via a neural network is necessary.
2 comments
AuthorsOpenReview ↗

Response to Reviewer h2Fc (Part 1/2)

We thank the reviewer for the careful assessment and constructive suggestions. We appreciate the concerns regarding the quantum-specific structure, the transparency and reproducibility of training, and comparisons with adaptive classical regularization. We have revised the corresponding figures, manuscript sections, supplementary material, and experimental comparisons to address these points. Our responses are organized below.

1. Quantum-specific structure and relation to learned regularization

We thank the reviewer for highlighting the need to clarify the role of quantum structure. To make the paper accessible to machine-learning readers, we initially kept the quantum-geometric background concise. Quantum structure is central to our method. The QGT, induced by the Fubini–Study geometry of parameterized quantum circuits, characterizes physically distinguishable tangent directions; the Hamiltonian-induced force determines how these directions contribute to imaginary-time evolution [1,2,7]. Together, they define the spectral decisions whose consequences are evaluated through the fidelity of an executed finite circuit step [2,3]. In matched initial-state tests at step size 1/32, fixed ridge achieves a better instantaneous McLachlan score, whereas SpectralGuard achieves lower one-step quantum-state infidelity in 20/20 task averages (Section 4.1). This reversal shows how the finite-step quantum-state objective changes the preferred numerical update. We have clarified these connections in Sections 2–3.1 and Supplementary Section A.

Regarding the reviewer's question about existing learned regularization, we have expanded Section 2 to compare learned scalar, Tikhonov, and spectral methods [8–10], with emphasis on supervision from recurrent quantum-state evolution. In the direct learned-controller comparison, SpectralGuard reduces trajectory infidelity relative to ScalarGuard by 15.6% on size shifts and 12.7% on depth shifts (Section 4.2), supporting the value of mode-wise decisions under the evaluated structural shifts.

Connection to quantum computing. SpectralGuard is designed for hybrid quantum–classical variational algorithms, where system size and circuit depth are central factors governing expressivity, quantum geometry, and computational constraints [7,11]. Specifically, our method operates on the quantum geometric tensor and Hamiltonian-induced force arising from parameterized quantum circuits, following the geometric formulation of variational quantum evolution [1,2]. To assess its adaptability to these quantum-specific structural variations, we evaluated the frozen policy on larger systems (14–18 qubits) and deeper circuits (14 and 18 layers) without target-side retraining. Compared with the learned ScalarGuard baseline, SpectralGuard reduces mean trajectory infidelity by 15.6% under size shifts and 12.7% under depth shifts. These results demonstrate the transferability of learned spectral decisions across the evaluated quantum-circuit scales and architectures (Section 4.2).

2. Training data, objective, and optimization

We thank the reviewer for identifying where the training procedure needed greater clarity. We have revised Fig. 2(e) to show source-trajectory evaluation, teacher-guided initialization, and short-trajectory refinement as a single training sequence, with the corresponding descriptions aligned in Section 3.

Supplementary Appendix B now provides two reproducibility tables specifying the controllers' input dimensions, network widths, parameter counts, losses, optimizers, learning rates, and training schedules. We have also clarified how source trajectories and candidate evaluations generate the supervisory targets in Appendix B, while Appendix J records the task partitions, checkpoint protocols, and execution steps.

An anonymized core-code package accompanies the submission, and we plan to release the core implementation publicly after publication to facilitate independent reproduction.

AuthorsOpenReview ↗

Response to Reviewer h2Fc (Part 2/2)

3. Comparison with adaptive non-learning regularization

We thank the reviewer for requesting stronger comparisons with adaptive non-learning regularization. We have added Fig. 1 to contrast criterion-based classical updates with learned spectral control. Beyond fixed and scaled Tikhonov, we evaluate condition-adaptive QNG–LM, spectrum-based truncated SVD (TSVD), and residual–solution-norm-based L-curve selection (Section 4.2; Supplementary Sections C and I).

On the prespecified 118-group structural-shift benchmark, verified SpectralGuard achieves 0.04585 mean trajectory infidelity versus 0.05191 for condition-adaptive QNG–LM (11.7% lower). In a separate exploratory 108-group clean-shift comparison, the respective results are 0.05109 for SpectralGuard, 0.05671 for QNG–LM, 0.05909 for TSVD, and 0.11380 for L-curve. To assess the learned proposal independently of the optional verifier, a separate exploratory 16-group moderate-noise experiment evaluates SpectralGuard without verification: 0.05286 versus 0.09543 for TSVD (44.6% lower, favorable in 16/16 groups). These results concern distinct protocols and are not pooled.

We appreciate the reviewer's constructive suggestion, which has helped us strengthen the adaptive-baseline evaluation and clarify the scope of the empirical evidence.

References

[1] Stokes, J., Izaac, J., Killoran, N., and Carleo, G. (2020). Quantum Natural Gradient. Quantum, 4, 269. https://doi.org/10.22331/q-2020-05-25-269

[2] McArdle, S., Jones, T., Endo, S., Li, Y., Benjamin, S. C., and Yuan, X. (2019). Variational ansatz-based quantum simulation of imaginary time evolution. npj Quantum Information, 5, 75. https://doi.org/10.1038/s41534-019-0187-2

[3] Gacon, J., Nys, J., Rossi, R., Woerner, S., and Carleo, G. (2024). Variational quantum time evolution without the quantum geometric tensor. Physical Review Research, 6, 013143. https://doi.org/10.1103/PhysRevResearch.6.013143

[4] Andrychowicz, M., et al. (2016). Learning to learn by gradient descent by gradient descent. Advances in Neural Information Processing Systems, 29. https://proceedings.neurips.cc/paper/2016/hash/fb87582825f9d28a8d42c5e5e5e8b23d-Abstract.html

[5] Wichrowska, O., et al. (2017). Learned Optimizers that Scale and Generalize. Proceedings of the 34th International Conference on Machine Learning, PMLR 70, 3751–3760. https://proceedings.mlr.press/v70/wichrowska17a.html

[6] Adler, J., and Öktem, O. (2018). Learned Primal-Dual Reconstruction. IEEE Transactions on Medical Imaging, 37(6), 1322–1332. https://doi.org/10.1109/TMI.2018.2799231

[7] Haug, T., Bharti, K., and Kim, M. S. (2021). Capacity and Quantum Geometry of Parametrized Quantum Circuits. PRX Quantum, 2, 040309. https://doi.org/10.1103/PRXQuantum.2.040309

[8] Afkham, B. M., Chung, J., and Chung, M. (2021). Learning regularization parameters of inverse problems via deep neural networks. Inverse Problems, 37(10), 105017. https://doi.org/10.1088/1361-6420/ac245d

[9] Alberti, G. S., De Vito, E., Lassas, M., Ratti, L., and Santacesaria, M. (2021). Learning the optimal Tikhonov regularizer for inverse problems. Advances in Neural Information Processing Systems, 34, 25205–25216. https://proceedings.neurips.cc/paper_files/paper/2021/hash/d3e6cd9f66f2c1d3840ade4161cf7406-Abstract.html

[10] Dong, H.-C., Cheng, P., and Li, S. (2026). Interpretable operator learning for inverse problems via adaptive spectral filtering: Convergence and resolution transfer. Neural Networks, article 109794. https://doi.org/10.1016/j.neunet.2026.109794

[11] Cerezo, M., et al. (2021). Variational quantum algorithms. Nature Reviews Physics, 3, 625–644. https://doi.org/10.1038/s42254-021-00348-9

Reviewer_SH3mOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Variational Quantum Imaginary-Time Evolution (VarQITE) proposes a family of parametrized states and we aim to generate a sequence of the parameters to mimic the imaginary-time evolution. The sequence is obtained by an iterative method and this is precisely the place, where a certain matrix inversion problem, which is the main focus of this paper, appears. To obtain the step v_t, VarQITE repeatedly solves G_t v_t = f_t, where the matrix G_t might have very small eigenvalues, which brings certain numerical issues when calculating the inverse. A standard method is to add a regularization term, a small \epsilon multiple of identity, before the inverse, called the Tikhonov regularization. The choice of the regularization significantly affects the behavior of the inverse. The main idea of the paper is to use machine learning to adress this problem.

Section 3.2 introduces two ideas. First, Scalar Guard attempts to choose the best value of the regularization \epsilon_t possible for the time step t. Second, Spectral Guard chooses a fixed regularization \epsilon but then has a controller that filters certain directions corresponding to ill-conditioned eigenvectors. Section 3.3 contains a mechanism that either accepts or rejects the learnt proposed update. Section 4 then gives the numerical results.

Critical strengths

0.) The authors explore an interesting application of machine learning to tackle this particular matrix inverse problem.

1.) In the experimental results, the authors transparently report regimes, for which classical Tikhonov (or other methods) outperform their Spectral and Scalar Guard methods.

2.) Spectral and Scalar guard have the best worst-case relative regret. This suggests that the proposed approaches may offer improved robustness across different deployment conditions, even though they do not consistently outperform classical baselines on individual tasks.

Critical weaknesses

1.) In my opinion, the authors do not sufficiently describe the central machine-learning components of their proposed methods. For example, in Equation (3), ScalarGuard is defined through the learned function \phi_g(z_t), but the architecture of \phi_g is never specified, either in the main text or in the supplementary material. Similarly, the input z_t is only described qualitatively as summarizing trajectory progress, conditioning, effective rank, and other numerical properties, without specifying its exact components or dimensionality. Standard implementation details such as the network architecture, depth, width, number of trainable parameters, optimizer, learning rate, and exact training loss are missing. Similar architectural details are absent for SpectralGuard.

2.) Standard or scaled Tikhonov achieves the lowest mean trajectory infidelity in six out of eight benchmark regimes, while ScalarGuard performs best only under strong noise and SpectralGuard does not win any individual regime. Given the additional training and complexity required by the proposed methods, their practical advantage over simpler classical regularization techniques remains unclear.

Questions for authors

1.) Could you provide more information about the model \phi_g(z_t)?

2.) How was the training set for Spectral Guard obtained? How were the target coefficients (g_{t,i}) are generated?

Future work suggestion: 3.) Scalar Guard predicts \epsilon_t and Spectral Guard predicts g_{t,i}. Have the authors considered learning \epsilon_t jointly with the gates (g_{t,i}), and could this provide further improvements?

Future work suggestion: 4.) How would the authors decide which of the proposed methods to use?

1 comment
AuthorsOpenReview ↗

Response to Reviewer SH3m

Response to Reviewer SH3m

We thank the reviewer for the careful assessment and constructive suggestions. We have revised the manuscript and supplementary material to provide additional methodological details and clarify the experimental comparisons and method-selection guidance. Our responses are organized below.

1. Architectures, features, and training settings

We have substantially expanded the methodological details in the revised main text and Supplementary Appendix B. In particular, we now explicitly specify the model ϕg(zt)\phi_g(z_t) and provide detailed architecture and training tables covering the inputs, parameterizations, normalization, optimization objectives, and training configurations of both ScalarGuard and SpectralGuard.

2. Source data and generation of target gates

Training uses 100 independent source dependency groups: 40 random kk-local, 40 Sherrington–Kirkpatrick, 10 repetition-derived, and 10 surface-derived Hamiltonian groups. At source-trajectory snapshots, we compute the quantum-geometric tensor (QGT) and force in the standard variational-dynamics setting (McArdle et al., 2019; Stokes et al., 2020), and construct scalar Tikhonov, spectrum-scaled Tikhonov, force-aware modal, and truncated-spectrum candidate responses. Each candidate is scored by its finite-step source-state consequence, including state infidelity, energy rebound, and update-size terms.

The best scalar candidate supplies an initial modal shrinkage profile. A structured candidate supplies the teacher gates gt,i⋆g_{t,i}^{\star} when it improves the source-only score by the prescribed margin. Structured non-Tikhonov teachers are selected in 12.0% of training and 11.9% of validation snapshots. Task- and mode-weighted imitation is followed by three-step source-rollout refinement. The revised manuscript now makes this teacher construction explicit; source-only references determine the training targets, while the deployed policy remains frozen on target tasks.

3. Why learn when classical methods sometimes win?

The eight-regime table identifies the lowest mean infidelity separately in each regime, whereas the central learning question concerns the performance of one source-trained policy across changing conditions. This is the cross-instance numerical decision setting studied in learned optimization and solver-parameter selection (Andrychowicz et al., 2016; Wichrowska et al., 2017; Khodak et al., 2024). Across the seven original non-floor regimes, SpectralGuard has the lowest descriptive worst-case relative excess infidelity among the principal methods: 20.9%, compared with 28.3% for ScalarGuard, 31.7% for condition-adaptive QNG–LM, and 47.9% for fixed Tikhonov. In the prespecified 118-group structural-transfer benchmark, verified SpectralGuard achieves 0.04585 mean trajectory infidelity versus 0.05191 for condition-adaptive QNG–LM, an 11.7% relative reduction.

The added exploratory comparison in Table 2 further evaluates the learned spectral response under estimator perturbations. On 16 moderate-noise OOD groups, SpectralGuard without verification achieves 0.05286 mean trajectory infidelity versus 0.09543 for raw-QGT TSVD, a 44.6% reduction, favorable in all 16 group averages. These distinct protocols establish complementary evidence for frozen structural transfer and learned mode-wise regularization under estimator shift. The measured controller cost is 0.0585 s per simulated trajectory, or 0.276% of the measured simulation workload.

4. Joint learning and method selection

We thank the reviewer for these constructive suggestions. Learning-to-optimize and cross-instance solver configuration provide a useful methodological basis for jointly modeling scalar and modal decisions and for selecting methods based on source-side validation (Andrychowicz et al., 2016; Khodak et al., 2024). We believe that developing a unified, regime-aware policy would be an interesting direction for future research.

References.

Andrychowicz, M., et al. (2016). Learning to Learn by Gradient Descent by Gradient Descent. Advances in Neural Information Processing Systems (NeurIPS 2016).

Khodak, M., Chow, E., Balcan, M.-F., & Talwalkar, A. (2024). Learning to Relax: Setting Solver Parameters Across a Sequence of Linear System Instances. International Conference on Learning Representations (ICLR 2024).

McArdle, S., et al. (2019). Variational Ansatz-Based Quantum Simulation of Imaginary Time Evolution. npj Quantum Information, 5, 75.

Stokes, J., Izaac, J., Killoran, N., & Carleo, G. (2020). Quantum Natural Gradient. Quantum, 4, 269.

Wichrowska, O., et al. (2017). Learned Optimizers That Scale and Generalize. Proceedings of the 34th International Conference on Machine Learning (ICML 2017), PMLR 70, 3751–3760.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.