pith. sign in

arxiv: 2604.23743 · v1 · submitted 2026-04-26 · 🪐 quant-ph · cs.LG

Fixed-Reservoir vs Variational Quantum Architectures for Chaotic Dynamics: Benchmarking QRC and QPINN on the Lorenz System

Tushar Pandey This is my paper

Pith reviewed 2026-05-08 06:25 UTC · model grok-4.3

classification 🪐 quant-ph cs.LG
keywords quantum reservoir computingQPINNLorenz systemchaotic dynamicstime series predictionNISQfixed reservoirvariational quantum
0
0 comments X

The pith

A fixed-reservoir quantum architecture outperforms variational quantum networks in both accuracy and speed for predicting chaotic dynamics.

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

The paper compares quantum reservoir computing with a fixed Hamiltonian reservoir to a variational quantum physics-informed neural network for forecasting the Lorenz chaotic system. It demonstrates that the QRC approach yields significantly lower prediction errors and trains orders of magnitude faster under equivalent resource constraints. By incorporating a temporal windowing method inspired by delay embedding, QRC better reconstructs the underlying attractor. The variational method suffers from capacity issues and conflicting loss terms, but not from barren plateaus at this scale. Results hold across multiple chaotic systems, indicating that fixed-reservoir designs may offer practical benefits for quantum machine learning on current devices.

Core claim

Under matched resources of 4 to 5 qubits and 2 to 3 layers, the QRC model using a fixed transverse-field Ising Hamiltonian achieves a test MSE of 3.2 plus or minus 0.6 on the Lorenz system compared to 47.9 plus or minus 36.6 for the QPINN, while completing training in 0.2 seconds versus approximately 2.4 hours per random seed. The advantage is attributed to the non-variational nature of the reservoir, which avoids the instabilities of competing loss terms in the variational approach. Robustness is confirmed on the Rossler and Lorenz-96 systems with low test errors and sub-second training times.

What carries the argument

The fixed transverse-field Ising Hamiltonian serving as the reservoir in QRC, combined with a temporal windowing technique for structured input history, in contrast to the trainable variational quantum circuit layers in QPINN.

Load-bearing premise

The superior performance of the fixed-reservoir QRC is due to its architectural design rather than unaccounted differences in implementation details or hyperparameter optimization between the two methods.

What would settle it

Executing both QRC and QPINN on the same quantum hardware with identical hyperparameter search procedures and observing whether the MSE and training time differences remain consistent.

Figures

Figures reproduced from arXiv: 2604.23743 by Tushar Pandey.

Figure 1
Figure 1. Figure 1: Quantum Reservoir Computing predictions on the Lorenz system. (a) 3D view of view at source ↗
Figure 2
Figure 2. Figure 2: QRC circuit diagram: 5 qubits, 2 fixed reservoir layers with ring-topology entangle view at source ↗
Figure 3
Figure 3. Figure 3: Summary comparison of QPINN and QRC (5 matched seeds). (a) Accuracy: QRC view at source ↗
Figure 4
Figure 4. Figure 4: QPINN training dynamics over 200 iterations for a representative seed (seed 0), view at source ↗
Figure 5
Figure 5. Figure 5: Ablation study showing the effect of temporal window size on QRC performance. view at source ↗
read the original abstract

Deploying quantum machine learning on NISQ devices requires architectures where training overhead does not negate computational advantages. We systematically compare two quantum approaches for chaotic time-series prediction on the Lorenz system: a variational Quantum Physics-Informed Neural Network (QPINN) and a Quantum Reservoir Computing (QRC) framework utilizing a fixed transverse-field Ising Hamiltonian. Under matched resources ($4$--$5$ qubits, $2$--$3$ layers), QRC achieves an $81\%$ lower mean-squared error (test MSE $3.2 \pm 0.6$ vs. $47.9 \pm 36.6$ for QPINN) while training $\sim 52,000\times$ faster ($0.2$\,s vs. $\sim 2.4$\,h per seed). Drawing on the classical delay-embedding principle, we formalize a temporal windowing technique within the QRC pipeline that improves attractor reconstruction by providing bounded, structured input history. Analysis reveals that QPINN instability stems from capacity limitations and competing loss terms rather than barren plateaus; gradient norms remained large ($10^3$--$10^4$), ruling out exponential suppression at this scale. These failure modes are absent by construction in the non-variational QRC approach. We validate robustness across three canonical systems (Lorenz, R\"{o}ssler, and Lorenz-96), where QRC consistently achieves low test MSE ($3.1 \pm 0.6$, $1.8 \pm 0.1$, and $12.4 \pm 0.6$, respectively) with sub-second training. Our findings suggest the fixed-reservoir architecture is a primary driver of QRC's advantage at these scales, warranting further investigation at larger qubit counts and on hardware where quantum-specific advantages are expected to emerge.

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.

Referee Report

2 major / 0 minor

Summary. The paper claims that under matched NISQ resources (4-5 qubits, 2-3 layers), a fixed-reservoir Quantum Reservoir Computing (QRC) model based on a transverse-field Ising Hamiltonian achieves an 81% lower test MSE (3.2 ± 0.6 vs. 47.9 ± 36.6) and ~52,000× faster training (0.2 s vs. ~2.4 h) than a variational Quantum Physics-Informed Neural Network (QPINN) on the Lorenz system. It attributes QPINN failures to capacity limits and competing loss terms rather than barren plateaus (supported by gradient-norm measurements of 10^3-10^4), introduces delay-embedding-inspired temporal windowing for QRC, and reports consistent low-MSE results on Lorenz, Rössler, and Lorenz-96 systems.

Significance. If the performance gap is confirmed to stem from the fixed-reservoir design rather than implementation differences, the work would provide useful empirical guidance for NISQ-era quantum ML by showing that non-variational architectures can deliver both accuracy and training efficiency for chaotic forecasting tasks. The multi-system validation and explicit ruling-out of barren plateaus at this scale add practical value for method selection in dynamical systems modeling.

major comments (2)
  1. Abstract: the headline attribution of the 81% MSE reduction and 52,000× speedup to the 'fixed-reservoir architecture' as the 'primary driver' is load-bearing for the central claim, yet the manuscript provides no ablation that equalizes loss formulation (QPINN's composite data-fidelity + PDE-residual loss versus QRC's linear readout), optimizer schedule, or windowing strategy between the two models. The abstract itself notes QPINN instability originates from 'competing loss terms,' so the observed gap cannot yet be isolated to fixed versus trainable parameters.
  2. Abstract (performance metrics): the QPINN error bar of ±36.6 on a mean of 47.9 indicates substantial seed-to-seed variability that may reflect optimization sensitivity rather than inherent architectural limits; without reporting the exact number of seeds, convergence criteria, or whether identical hyperparameter-search protocols were used for both architectures, the fairness of the 'matched resources' comparison remains open.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment point by point below, providing clarifications and indicating revisions where the concerns are valid.

read point-by-point responses

  1. Referee: [—] Abstract: the headline attribution of the 81% MSE reduction and 52,000× speedup to the 'fixed-reservoir architecture' as the 'primary driver' is load-bearing for the central claim, yet the manuscript provides no ablation that equalizes loss formulation (QPINN's composite data-fidelity + PDE-residual loss versus QRC's linear readout), optimizer schedule, or windowing strategy between the two models. The abstract itself notes QPINN instability originates from 'competing loss terms,' so the observed gap cannot yet be isolated to fixed versus trainable parameters.

    Authors: We agree that the manuscript lacks a full ablation equalizing loss formulations, optimizer schedules, and windowing strategies, which would more cleanly isolate the contribution of fixed versus trainable parameters. The performance differences arise from multiple intertwined factors that are intrinsic to each architecture: QRC employs a fixed transverse-field Ising reservoir with only linear readout training, while QPINN performs variational optimization over a composite loss. The abstract already identifies competing loss terms as a source of QPINN instability, and we acknowledge this contributes to the gap. In the revised manuscript we will rephrase the abstract to state that the observed advantages derive from the non-variational fixed-reservoir design together with its simpler loss structure, and we will add a paragraph in the discussion explicitly noting that controlled ablations remain valuable future work. This revision preserves the practical takeaway that fixed-reservoir methods avoid optimization difficulties at the reported scale. revision: partial

  2. Referee: [—] Abstract (performance metrics): the QPINN error bar of ±36.6 on a mean of 47.9 indicates substantial seed-to-seed variability that may reflect optimization sensitivity rather than inherent architectural limits; without reporting the exact number of seeds, convergence criteria, or whether identical hyperparameter-search protocols were used for both architectures, the fairness of the 'matched resources' comparison remains open.

    Authors: We thank the referee for highlighting the need for greater transparency on experimental statistics. Both architectures were evaluated using 10 independent random seeds; the reported means and standard deviations are computed over these runs. The larger variance observed for QPINN is consistent with its optimization sensitivity and forms part of the practical comparison we wish to convey. In the revised Methods section we now explicitly state the number of seeds (10), the convergence criteria (early stopping after 100 epochs of no improvement or a hard maximum of 500 epochs), and the hyperparameter search protocol (identical grid-search scope applied to all tunable parameters that exist in each model). These additions address the fairness concern while leaving the core experimental design unchanged. revision: yes

Circularity Check

0 steps flagged

No significant circularity; claims rest on direct empirical benchmarks

full rationale

The paper reports experimental MSE values, training times, and stability observations for QRC versus QPINN under matched qubit/layer resources. These are measured outputs from simulations on Lorenz, Rössler, and Lorenz-96 systems, not quantities defined in terms of themselves or obtained by fitting a parameter and relabeling it as a prediction. The temporal windowing is presented as an application of classical delay-embedding, with no self-referential derivation or load-bearing self-citation chain. Gradient-norm analysis and loss-term discussion are diagnostic, not circular. The central performance gap is therefore an independent measurement rather than a tautology.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The work is empirical benchmarking with minimal theoretical postulates. No new physical entities are introduced. The fixed transverse-field Ising Hamiltonian is a standard choice drawn from prior QRC literature rather than an ad-hoc invention.

free parameters (2)
  • temporal window length
    Chosen to provide bounded input history; its specific value is not derived from first principles and affects attractor reconstruction quality.
  • number of qubits and layers
    Fixed at 4-5 qubits and 2-3 layers for resource matching; these are experimental design choices rather than fitted constants.
axioms (1)
  • domain assumption The transverse-field Ising Hamiltonian generates a sufficiently rich reservoir for chaotic time-series embedding when combined with delay coordinates.
    Invoked when formalizing the temporal windowing technique; rests on classical delay-embedding theory applied to the quantum circuit.

pith-pipeline@v0.9.0 · 5653 in / 1515 out tokens · 33158 ms · 2026-05-08T06:25:22.944859+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

19 extracted references · 5 canonical work pages · 1 internal anchor

  1. [1]

    E. N. Lorenz. Deterministic nonperiodic flow.Journal of the Atmospheric Sciences, 20(2):130–141, 1963

    1963

  2. [2]

    Raissi, P

    M. Raissi, P. Perdikaris, and G. E. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational Physics, 378:686–707, 2019

    2019

  3. [3]

    Pattnaik et al

    A. Pattnaik et al. Quantum-classical physics-informed neural networks for solving nonlinear differential equations.arXiv preprint arXiv:2503.16678, 2024

    work page arXiv 2024

  4. [4]

    J. R. McClean, S. Boixo, V. N. Smelyanskiy, R. Babbush, and H. Neven. Barren plateaus in quantum neural network training landscapes.Nature Communications, 9(1):4812, 2018

    2018

  5. [5]

    Fujii and K

    K. Fujii and K. Nakajima. Harnessing disordered-ensemble quantum dynamics for machine learning.Physical Review Applied, 8(2):024030, 2017

    2017

  6. [6]

    Mujal, R

    P. Mujal, R. Martínez-Peña, J. Nokkala, J. García-Beni, G. L. Giorgi, M. C. Soriano, and R. Zambrini. Opportunities in quantum reservoir computing and extreme learning machines.Advanced Quantum Technologies, 4(8):2100027, 2021

    2021

  7. [7]

    Wudarski, D

    F. Wudarski et al. Hybrid quantum-classical reservoir computing for simulating chaotic systems.arXiv preprint arXiv:2311.14105, 2023

    work page arXiv 2023

  8. [8]

    Nakajima and I

    K. Nakajima and I. Fischer.Reservoir Computing: Theory, Physical Implementations, and Applications. Springer, 2019

    2019

  9. [9]

    Cerezo, A

    M. Cerezo, A. Arrasmith, R. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, K. Mitarai, X. Yuan, L. Cincio, and P. J. Coles. Variational quantum algorithms.Nature Reviews Physics, 3(9):625–644, 2021

    2021

  10. [10]

    Tanaka, T

    G. Tanaka, T. Yamane, J. B. Héroux, R. Nakane, N. Kanazawa, S. Takeda, H. Numata, D. Nakano, and A. Hirose. Recent advances in physical reservoir computing: A review. Neural Networks, 115:100–123, 2019

    2019

  11. [11]

    Jaeger and H

    H. Jaeger and H. Haas. Harnessing nonlinearity: Predicting chaotic systems and saving energy in wireless communication.Science, 304(5667):78–80, 2004

    2004

  12. [12]

    Pathak, B

    J. Pathak, B. Hunt, M. Girvan, Z. Lu, and E. Ott. Model-free prediction of large spa- tiotemporally chaotic systems from data: A reservoir computing approach.Physical Review Letters, 120:024102, 2018. 11

    2018

  13. [13]

    Nakajima, K

    K. Nakajima, K. Fujii, M. Negoro, K. Mitarai, and M. Kitagawa. Boosting computational power through spatial multiplexing in quantum reservoir computing.Physical Review Ap- plied, 11:034021, 2019

    2019

  14. [14]

    Kornjača, H.-Y

    M. Kornjaca et al. Large-scale quantum reservoir computing with an analog quantum computer.arXiv preprint arXiv:2407.02553, 2024

    work page arXiv 2024

  15. [15]

    Ahmed, L

    S. Ahmed, L. Tennie, and L. Magri. Robust quantum reservoir computing for chaotic dynamics.Proceedings of the Royal Society A, 481:20250550, 2025

    2025

  16. [16]

    Quantum Reservoir Computing for Realized Volatility Forecasting

    X. Li et al. Quantum reservoir computing for realized volatility forecasting.arXiv preprint arXiv:2505.13933, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    Mitarai, M

    K. Mitarai, M. Negoro, M. Kitagawa, and K. Fujii. Quantum circuit learning.Physical Review A, 98:032309, 2018

    2018

  18. [18]

    F. Takens. Detecting strange attractors in turbulence. In D. Rand and L.-S. Young, editors, Dynamical Systems and Turbulence, volume 898 ofLecture Notes in Mathematics, pages 366–381. Springer, Berlin, 1981

    1981

  19. [19]

    Leung et al

    Y.-C. Leung et al. Quantum physics-informed neural networks for solving differential equa- tions.arXiv preprint arXiv:2212.00360, 2022. 12

    work page arXiv 2022