Pith. sign in

REVIEW 4 major objections 7 minor 61 references

Machine Learning-aided Optimal Control of a noisy qubit

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A transformer-based graybox model learns a four-parameter noise operator and acts as a differentiable emulator, yielding qubit gate fidelities above 99% at low coupling and above 90% at high coupling.

desk verdict A coherent but thin demonstration: the transformer-in-graybox pipeline is sound and independently re-validated, yet no baselines or error bars back the headline >90% claim. read the letter →

arxiv 2507.14085 v1 pith:J5CJQT7S submitted 2025-07-18 quant-ph cond-mat.other

Riccardo Cantone , Shreyasi Mukherjee , Luigi Giannelli , Elisabetta Paladino , Giuseppe Falci This is my paper
classification quant-phcond-mat.other
keywords grayboxquantumcontroltransformerneuralnetworkrandomtelegraphnoiseOrnstein-Uhlenbeckprocessfidelitysingle-qubitgatesoptimalnon-Markoviandephasing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a graybox machine-learning model—an exact unitary-evolution layer paired with a small transformer—can act as a fast surrogate for a single qubit under classical dephasing noise, accurate enough for gradient-based optimal control. The motivation is that non-Markovian and non-Gaussian noise is hard to handle analytically, so learning an effective noise operator from simulated data could replace explicit environment modeling. On random-telegraph and Ornstein-Uhlenbeck noise, the surrogate predicts gate fidelities with errors around $10^{-3}$ to $10^{-2}$, and the pulse sequences it designs implement a universal set of single-qubit gates with Monte Carlo-verified process fidelities above 99% at weak coupling and above 90% at the strongest coupling considered. The intended takeaway is that a fast, differentiable emulator of this kind makes high-fidelity control practical under realistic noise without a microscopic bath model.

What carries the argument

The load-bearing object is the four-parameter noise-averaged operator $V_O$ of Eq. (12). It sweeps all stochastic and system-bath coupling information into $D=\mathrm{diag}(\mu,-\mu)$ and a unitary $Q$ parameterized by $\theta,\psi,\Delta$, making the physics-informed part of the model exact while the transformer only has to learn four numbers per Pauli observable. The whitebox modules fix the control unitary $U_{\mathrm{ctrl}}$, the quantum-measurement layer, and the process-matrix fidelity computation, so the network's predictions are automatically consistent with unitary evolution and quantum measurement.

What would settle it

Re-run the emulator-training and pulse-optimization pipeline, and at the optimized pulses compare the surrogate's predicted fidelity with an exact Monte Carlo estimate using $K=2000$ noise realizations; if the discrepancy is systematically larger than the model's test-set prediction error (about 0.02 at $g=0.5$ MHz), the optimizer is exploiting regions where the four-number compression of the noise fails. The same comparison with a non-commuting quantum bath, for which the classical-noise formula cannot hold, would show where the method's validity ends.

Watch

Extended reading notes

Core claim

The central claim is that the entire noise-averaged effect of a classical environment on a driven qubit can be represented by a single operator $V_O = O^{-1}QDQ^{\dagger}$ with $D=\mathrm{diag}(\mu,-\mu)$ and $Q$ built from three angles $(\theta,\psi,\Delta)$, and that a lightweight transformer can learn the four real parameters $(\mu,\theta,\psi,\Delta)$ as functions of the control pulses. The whitebox layers then use Eq. (11), $E[O(T)]_\rho = \mathrm{tr}_S[V_O U_{\mathrm{ctrl}}(t)\rho(0)U_{\mathrm{ctrl}}^\dagger(t)O]$, to compute 18 expectation values that are converted into process fidelities for the gate set. The paper reports that this emulator, trained on Monte Carlo data for RTN and OU noise, predicts those fidelities with test MSE between about $8\times10^{-6}$ and $6\times10^{-4}$, and that gradient-based optimization of the emulator yields pulse trains whose Monte Carlo-verified fidelities exceed 0.99 for $g=0.1$ MHz and stay above 0.90 for $g=0.5$ MHz for both noise types.

Load-bearing premise

The load-bearing premise is that the accumulated effect of the classical noise on each measured Pauli direction can always be compressed into four numbers that depend on the controls and the noise but not on the qubit's initial state; if that compression is incomplete under the Gaussian-pulse drives used here, the emulator's predictions are systematically biased and the optimized pulses could be mis-calibrated.

Editorial extensions

If this is right

  • One trained model yields optimized pulse trains for all six gates in the universal set $\{I,X,Y,Z,H,R_X(\pi/4)\}$, because $V_O$ is independent of the qubit's initial state.
  • The same emulator handles Gaussian OU noise and non-Gaussian RTN at comparable accuracy, so memory effects alone do not force a heavier blackbox.
  • Monte Carlo-verified process fidelities stay above 90% even at $g=0.5$ MHz, setting a practical operating range for this Gaussian-pulse parametrization.
  • Because each optimization step uses only a few forward passes through the trained network, the approach replaces costly Monte Carlo averaging inside gradient-based search.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural next test is to apply the same architecture to a bath with quantum back-action; Eq. (12) should break there, and the failure mode would mark the boundary of the four-parameter compression.
  • The paper does not report the surrogate's own fidelity prediction at the optimized pulses; comparing that number with the Monte Carlo value would reveal whether the optimizer is exploiting model bias.
  • The learned parameters $(\mu,\theta,\psi,\Delta)$ of $V_O$ could be read as compressed noise-filter information, offering a route to noise spectroscopy from control data alone, beyond the gate-design use demonstrated here.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The manuscript presents a graybox machine-learning approach to model and control a single qubit subject to classical dephasing noise. The whitebox part computes the control unitary from Gaussian pulse amplitudes and imposes the structure of Eq. (11), while a transformer-based blackbox outputs the four parameters of the noise operator V_O in Eq. (12). The model is trained separately for Random Telegraph Noise and Ornstein-Uhlenbeck noise at coupling strengths g=0.1-0.5 MHz on Monte Carlo-generated datasets with K=2000 realizations, and predicts process fidelities for six single-qubit gates. The trained model is then used as a differentiable emulator for BFGS optimal control of the pulse amplitudes, and the resulting pulses are re-evaluated with independent Monte Carlo simulation. These simulations report fidelities above 99% at g=0.1 MHz and 90-91% at g=0.5 MHz.

Significance. If the quantitative claims survive scrutiny, the paper offers a useful proof-of-concept that a physics-informed transformer can serve as a fast surrogate for optimal control of a noisy qubit. A clear strength is that the final fidelities are computed by independent Monte Carlo simulations rather than by the model's own fitted output, so the headline numbers are not self-confirmed. The systematic comparison of RTN and OU noise under matched power spectra is also a nice feature. However, the paper does not yet establish the headline 'above 90%' claim with statistical error bars, and it does not demonstrate an advantage over existing control methods such as GRAPE or the original graybox framework. These gaps currently limit the significance of the work.

major comments (4)
  1. [Sec. II E, Tables III and VI] The quantitative claim in the abstract that fidelities 'remain above 90% for the highest coupling' is not supported by the reported digits because the Monte Carlo results are quoted without statistical error bars. With K=2000 realizations, a single measured probability near p=0.90 has a binomial standard error of about sqrt(0.9*0.1/2000) ≈ 0.0067; the process fidelity is reconstructed from 18 independently averaged expectation values, so the propagated uncertainty on each entry is plausibly 0.01-0.03. Since every g=0.5 entry in Tables III and VI lies between 0.901 and 0.911, the 95% confidence intervals for most gates extend below 0.90. The paper should report bootstrap or analytic error bars for all simulation-based fidelity values and should restate the 'above 90%' claim in the abstract, Sec. IV A2, and Sec. IV B2 in terms of the lower confidence bound.
  2. [Sec. IV A2 and IV B2] No baseline control is reported. At g=0.5, optimized fidelities are only about 0.90, so without a comparison it is unclear whether the ML-optimized pulses outperform a constant-amplitude drive, a no-control evolution, or a standard GRAPE pulse by more than the Monte Carlo noise. The paper's central claim is that the ML emulator enables effective control; that claim requires at least a no-control/constant-drive baseline and ideally a comparison with an established method such as GRAPE or the original graybox of Ref. [18]. Reporting the difference in fidelity, with uncertainty, would make the advantage quantitative.
  3. [Sec. II C, Eq. (12)] The four-parameter form V_O = O^{-1}QDQ† is imported from Ref. [18] with the phrase 'It can be shown [18]', but the conditions under which this representation is exact in the present driven, finite-time setting are not stated. The whole blackbox output is constrained by this ansatz, so an incomplete or inapplicable ansatz would bias the emulator's predictions. Please provide a self-contained derivation, or at minimum a precise statement of why the result of Ref. [18] applies to Gaussian pulse trains and arbitrary coupling strengths considered here. The independent Monte-Carlo validation of the final pulses mitigates, but does not remove, the concern that the optimizer may exploit errors in the surrogate.
  4. [Sec. III B and III D] The architecture uses a separate refinement head for each target gate, each of which 'independently adjusts the raw expectation values' before the corresponding process fidelity is computed. If this is taken literally, the model produces six different sets of expectation values and therefore does not represent a single quantum channel, which conflicts with the claim that the whitebox layers construct a tomographically complete process matrix. Please clarify whether the refinement is a shared operation or six independent corrections, and explain why the resulting six fidelity outputs can still be interpreted as fidelities of one underlying process. If the heads are independent predictors, the 'greybox emulator' terminology should be qualified accordingly.
minor comments (7)
  1. [Sec. III D] Typographical: 'T ransformer encoder' should be 'Transformer encoder'.
  2. [Sec. V and Acknowledgments] Minor language issues: 'a system subject' should be 'a system', 'accross' should be 'across', and 'acknowlwdge' should be 'acknowledge'.
  3. [Sec. II A, Eq. (3)] In the sentence defining τk, the index 'i' should be 'k'.
  4. [Sec. II B a] In Eq. (4), the subscript on β should be 'βRTN(0)' rather than 'βRNT(0)'.
  5. [References] Reference [42] contains a stray '/suppress' in the author list and several other references have incomplete formatting; please clean the bibliography.
  6. [Sec. III A] The fixed pulse widths and positions are stated in prose, but the input vector is said to contain '10 real numbers ... together with the fixed τk and σk'; please state explicitly in the data-generation section that σk and τk are the same for all data points and are not part of the optimized parameter vector.
  7. [Figs. 2 and 3] Figures 2 and 3 plot the same fidelity values as Tables III and VI without error bars; adding confidence intervals to the figures, or removing the redundant figures, would improve clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the headline fidelities are independently recomputed by Monte Carlo simulation, and the imported V_O structure comes from external prior work, not from the authors' own self-citation chain.

full rationale

The central optimal-control claim is not circular. Although the trained transformer is used as an emulator to optimize pulse parameters, the reported fidelities in Tables III and VI are explicitly obtained by feeding the optimized pulses into a Monte Carlo simulation of the system (Sec. II E, K=2000 realizations), not by outputting the model's own fitted fidelities. This is a standard supervised surrogate loop: the model is trained on simulated labels, used inside a gradient-based optimizer, and then the proposed controls are re-evaluated with the forward simulator. No reported 'prediction' is equivalent to a fitted parameter by construction. The four-parameter form of V_O in Eq. (12) is imported from reference [18] (Youssry, Paz-Silva, and Ferrie), which is external to the present author list, so this is an ansatz/assumption carried from prior work rather than a self-citation that forces the result. The authors' own self-citations (e.g., refs [6], [15], [32]-[36]) are background, tutorial, or noncentral and do not carry the derivation. Concerns about whether the V_O parameterization is fully justified for driven finite-time classical noise, or about missing Monte Carlo error bars on the 'above 90%' claim, are correctness and statistical-support issues, not circularity. No equation or fitted quantity is shown to reduce to its own input.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central results are built on the V_O parameterization imported from Youssry et al. [18], the classical-noise and pure-dephasing model, and a hand-chosen Gaussian pulse parameterization. The only learned quantities are the neural network weights; no new physical entities, particles, fields, or conserved quantities are introduced.

free parameters (5)
  • Neural network weights (transformer encoder, output branches, refinement heads) = trained by Adam to minimize MSE
    These are the fitted degrees of freedom that determine how accurately V_O parameters and gate fidelities are predicted from input pulses.
  • Noise rate gamma = k/2 = 1 MHz
    Fixed for every dataset in Section III.E; the correlation time and non-Markovian character of the noise are set by this choice.
  • OU power-spectrum matching constants (D, k) = k = 2 MHz, D not stated
    Section II.B says D and k are chosen so the OU spectrum matches RTN, but D is never reported, so the OU datasets are not exactly reproducible.
  • Gaussian pulse geometry (N, T, sigma) = N=5, T=3.2 us, sigma=T/(12N)
    Chosen by hand in Section II.A; this restricts the control search space and shapes the achievable fidelities.
  • Training amplitude ranges for A_k,alpha = not specified
    The random amplitudes used to generate the synthetic training sets are not given; the model's generalization and the optimized pulse scale depend on this distribution.
assumptions (5)
  • domain assumption The environment is a purely classical stochastic process beta(t); quantum bath operators are set to zero (Appendix A).
    The Hamiltonian (1) and the V_O formalism treat noise as a scalar classical process; this excludes genuine quantum back-action and is not justified beyond the chosen noise models.
  • domain assumption Driftless qubit, Omega=0, with pure dephasing along sigma_z and control only along x and y (Section II.A, Appendix A).
    The whole derivation of Eq. (1) and the interaction-picture simplification rely on Omega=0, which removes free precession and makes the results specific to this regime.
  • domain assumption For classical noise and traceless O, V_O = O^{-1} Q D Q† with four real parameters (Section II.C, Eq. 12, from ref [18]).
    The completeness and validity of this four-parameter ansatz for the driven, finite-time setting is not proved in this paper; the model's predictive accuracy is conditional on it.
  • domain assumption Monte Carlo averaging with K=2000 realizations and M=3000 time steps produces converged ground-truth fidelities (Section III.E).
    No convergence study or statistical uncertainty is reported, so the labels used for training and final validation are assumed accurate.
  • domain assumption The 18 Pauli expectation values are tomographically complete for reconstructing the single-qubit process matrix (Section III.C.e).
    This is standard for single-qubit process tomography, but the reconstruction formula is not shown and the refinement heads may break consistency across gate-specific outputs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Machine Learning-aided Optimal Control of a noisy qubit." pith.science (2026). https://pith.science/paper/J5CJQT7S

@misc{pith2026250714085,
  author       = {Pith},
  title        = {Pith review of: Machine Learning-aided Optimal Control of a noisy qubit},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J5CJQT7S}},
  note         = {Machine review of arXiv:2507.14085}
}
read the original abstract

We apply a graybox machine-learning framework to model and control a qubit undergoing Markovian and non-Markovian dynamics from environmental noise. The approach combines physics-informed equations with a lightweight transformer neural network based on the self-attention mechanism. The model is trained on simulated data and learns an effective operator that predicts observables accurately, even in the presence of memory effects. We benchmark both non-Gaussian random-telegraph noise and Gaussian Ornstein-Uhlenbeck noise and achieve low prediction errors even in challenging noise coupling regimes. Using the model as a dynamics emulator, we perform gradient-based optimal control to identify pulse sequences implementing a universal set of single-qubit gates, achieving fidelities above 99% for the lowest considered value of the coupling and remaining above 90% for the highest.

Figures

Figures reproduced from arXiv: 2507.14085 by the authors.

Figure 1
Figure 1. FIG. 1. The transformer-based graybox architecture we used [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Process matrix fidelity versus RTN noise coupling [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. Process matrix fidelity versus OU noise coupling [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 45 canonical work pages

  1. [18]

    H. Y. Carr and E. M. Purcell, Effects of diffusion on free precession in nuclear mag- netic resonance experiments, Physical Review https://doi.org/10.1103/PhysRev.94.630 (1954)

    doi:10.1103/physrev.94.630 1954

  2. [1]

    The final train and test loss quantifying the performance of the prediction of each gate fidelity is summarized in Table I

    Training We trained separate instances of the graybox model on each of the datasets generated for the various values ofg. The final train and test loss quantifying the performance of the prediction of each gate fidelity is summarized in Table I. It shows that the graybox attention-based model main- tains low training and testing MSE values across all gate...

    work page

  3. [2]

    The resulting optimized fidelities are reported in Fig

    Optimal Control After training, the graybox model was used as a quan- tum system emulator in the optimal control protocol de- scribed in Subsection II D. The resulting optimized fidelities are reported in Fig. 2 and Tab. III for the various values of g. The fidelity val- ues reported here are obtained by feeding the pulse se- quences generated by the opti...

    work page

  4. [3]

    The prediction performance for each gate and coupling value is summarized in Table IV, which reports the final training and testing MSE values for each gate

    Training Following the same strategy used for the RTN noise we train a different instance of the graybox model on each of the datasets generated for the OU case for different couplings g. The prediction performance for each gate and coupling value is summarized in Table IV, which reports the final training and testing MSE values for each gate. Results sho...

    work page

  5. [4]

    Figure 3 and Tab

    Optimal Control After training, the attention-based graybox model is used as an emulator for the optimal control stage. Figure 3 and Tab. VI report the optimized process matrix fidelities obtained as result of the optimal control procedure. FIG. 3. Process matrix fidelity versus OU noise coupling strength g for each gate in the universal set S. g = 0.1 g ...

    work page

  6. [5]

    Stronger noise regimes demand larger datasets to capture complex system-environment correlations accurately

    Scaling with noise strength and dataset size. Stronger noise regimes demand larger datasets to capture complex system-environment correlations accurately. Future work can explore systematic strategies for dataset generation and optimal pa- rameter sampling, ensuring that the model size and training costs remain manageable even in with higher values of the...

    work page

  7. [6]

    Prac- tical quantum hardware may impose constraints on accessible control pulses

    Adaptation to real-device constraints. Prac- tical quantum hardware may impose constraints on accessible control pulses. The graybox framework can be adapted to these limitations by embedding hardware-specific constraints during training, en- abling device-tailored optimal control solutions

    work page

  8. [7]

    Although this work focused on a single qubit, the method naturally extends to multiqubit registers and multilevel systems

    Generalization to multiqubit and multilevel systems. Although this work focused on a single qubit, the method naturally extends to multiqubit registers and multilevel systems. The software de- veloped here can be adapted for two-qubit gates constructed from two-level unitary blocks [36]. Fu- ture work should explore scaling properties and how additional w...

    work page 2024

Show all 61 references
  1. [8]

    H. Ball, M. J. Biercuk, A. R. Carvalho, J. Chen, M. Hush, L. A. De Castro, L. Li, P. J. Liebermann, H. J. Slatyer, C. Edmunds,et al., Software tools for quan- tum control: Improving quantum computer performance through noise and error suppression, Quantum Science and Technolog...

    2021

  2. [9]

    A. A. Abbott, J. Wechs, D. Horsman, M. Mhalla, and C. Branciard, Communication through coherent control of quantum channels, Quantum 4, 333 (2020)

    2020

  3. [10]

    Poggiali, P

    F. Poggiali, P. Cappellaro, and N. Fabbri, Optimal con- trol for one-qubit quantum sensing, Physical Review X 8, 021059 (2018)

    2018

  4. [11]

    Soare, H

    A. Soare, H. Ball, D. Hayes, J. Sastrawan, M. Jarratt, J. McLoughlin, X. Zhen, T. Green, and M. Biercuk, Ex- perimental noise filtering by quantum control, Nature Physics 10, 825 (2014)

    2014

  5. [12]

    Dong and I

    D. Dong and I. R. Petersen, Quantum control theory and applications: a survey, IET control theory & applications 4, 2651 (2010)

    2010

  6. [13]

    Giannelli, S

    L. Giannelli, S. Sgroi, J. Brown, G. S. Paraoanu, M. Pa- ternostro, E. Paladino, and G. Falci, A tutorial on op- timal control and reinforcement learning methods for quantum technologies, Physics Letters A 434, 128054 (2022)

    2022

  7. [14]

    Mabuchi, Coherent-feedback quantum control with a dynamic compensator, Physical Review A—Atomic, Molecular, and Optical Physics 78, 032323 (2008)

    H. Mabuchi, Coherent-feedback quantum control with a dynamic compensator, Physical Review A—Atomic, Molecular, and Optical Physics 78, 032323 (2008)

    2008

  8. [15]

    L. A. Clark, A. Stokes, and A. Beige, Quantum-enhanced metrology with the single-mode coherent states of an op- tical cavity inside a quantum feedback loop, Physical Re- view A 94, 023840 (2016)

    2016

  9. [16]

    Karmakar, P

    T. Karmakar, P. Lewalle, Y. Zhang, and K. B. Wha- ley, Noise-canceling quantum feedback: non-hermitian dynamics with applications to state preparation and magic state distillation, arXiv preprint arXiv:2507.05611 (2025)

    2025

  10. [17]

    Viola and S

    L. Viola and S. Lloyd, Dynamical suppression of deco- herence in two-state quantum systems, Physical Review A https://doi.org/10.1103/PhysRevA.58.2733 (1998)

    1998 doi

  11. [19]

    Viola, E

    L. Viola, E. Knill, and S. Lloyd, Dynamical decou- pling of open quantum systems, Physical Review Letters https://doi.org/10.1103/PhysRevLett.82.2417 (1999)

    1999 doi

  12. [20]

    Turyansky, Y

    D. Turyansky, Y. Zolti, Y. Cohen, and A. Pick, Pulse op- timization in adiabatic quantum computation and con- trol, arXiv preprint arXiv:2507.09770 (2025)

    2025

  13. [21]

    Y.-X. Zeng, T. Gebremariam, M.-S. Ding, and C. Li, Adi- abatic evolution: The influence of non-markovian charac- ters on quantum adiabatic evolution (ann. phys. 1/2019), Annalen der Physik 531, 1970010 (2019)

    2019

  14. [22]

    Mukherjee, D

    S. Mukherjee, D. Penna, F. Cirinn` a, M. Paternostro, 11 E. Paladino, G. Falci, and L. Giannelli, Noise classi- fication in three-level quantum networks by machine learning, Machine Learning: Science and Technology 5, 045049 (2024)

    2024

  15. [23]

    Delben, M

    G. Delben, M. Beims, and M. da Luz, Control of a qubit under markovian and non-markovian noise, Physical Re- view A 108, 012620 (2023)

    2023

  16. [24]

    Ortega-Taberner, E

    C. Ortega-Taberner, E. O’Neill, E. Butler, G. E. Fux, and P. Eastham, Unifying methods for optimal control in non-markovian quantum systems via process tensors, The Journal of Chemical Physics 161 (2024)

    2024

  17. [25]

    Youssry, G

    A. Youssry, G. A. Paz-Silva, and C. Ferrie, Characteriza- tion and control of open quantum systems beyond quan- tum noise spectroscopy, npj (Nature Partner Journals) Quantum Information https://doi.org/10.1038/s41534- 020-00332-8 (2020)

    2020 doi

  18. [26]

    A. Auza, A. Youssry, G. Paz-Silva, and A. Pe- ruzzo, Quantum control in the presence of strongly coupled non-markovian noise, arXiv https://doi.org/10.48550/arXiv.2404.19251

    doi arXiv

  19. [27]

    Youssry and H

    A. Youssry and H. I. Nurdin, Multi-axis con- trol of a qubit in the presence of unknown non- markovian quantum noise, Quantum Sci. Technol. https://doi.org/10.1088/2058-9565/aca711 (2023)

    2023 doi

  20. [28]

    Youssry, Y

    A. Youssry, Y. Yang, and R. J. C. et al., Experimental graybox quantum system identification and control, npj Quantum Information https://doi.org/10.1038/ (2024)

    2024

  21. [29]

    Gurney, An introduction to neural networks (CRC press, 2018)

    K. Gurney, An introduction to neural networks (CRC press, 2018)

    2018

  22. [30]

    Marquardt, Machine learning and quantum devices, SciPost Physics Lecture Notes , 029 (2021)

    F. Marquardt, Machine learning and quantum devices, SciPost Physics Lecture Notes , 029 (2021)

    2021

  23. [31]

    G´ eron,Hands-on machine learning with Scikit-Learn, Keras, and TensorFlow (” O’Reilly Media, Inc

    A. G´ eron,Hands-on machine learning with Scikit-Learn, Keras, and TensorFlow (” O’Reilly Media, Inc. ”, 2022)

    2022

  24. [32]

    Nasteski, An overview of the supervised machine learning methods, Horizons

    V. Nasteski, An overview of the supervised machine learning methods, Horizons. b 4, 56 (2017)

    2017

  25. [33]

    Perrier, A

    E. Perrier, A. Youssri, and C. Ferrie, Qdataset, quantum datasets for machine learning, Nature - scientific data https://doi.org/10.1038/s41597-022-01639-1 (2022)

    2022 doi

  26. [34]

    C. W. Gardiner, Handbook of Stochastic Methods - For Physics, Chemistry and the Natural Sciences (Springer, 1983)

    1983

  27. [35]

    Mandel and E

    L. Mandel and E. Wolf, Optical coherence and quantum optics (Cambridge University Press, 1995)

    1995

  28. [36]

    Papoulis and S

    A. Papoulis and S. U. Pillai, Probability, Random Vari- ables and Stochastic Processes (McGraw-Hill, 2002)

    2002

  29. [37]

    G. E. Uhlenbeck and L. S. Ornstein, On the theory of the brownian motion, Physical Review https://doi.org/10.1103/PhysRev.36.823 (1930)

    1930 doi

  30. [38]

    D. T. Gillespie, Exact numerical simulation of the ornstein-uhlenbeck process and its integral, Physical Review E https://doi.org/10.1103/PhysRevE.54.2084 (1996)

    1996 doi

  31. [39]

    Paladino, L

    E. Paladino, L. Faoro, A. D’Arrigo, and G. Falci, Back- ground charges induced stochastic fluctuations in joseph- son qubits, Journal of the Physical Society of Japan https://doi.org/10.1143/JPSJS.72SA.165 (2002)

    2002 doi

  32. [40]

    Paladino, L

    E. Paladino, L. Faoro, and G. Falci, Decoherence due to discrete noise in josephson qubits, Chapter in Advances in Solid State Physics (ASSP, volume 43) https://doi.org/10.1007/978-3-540-44838-9 53 (2003)

    2003 doi

  33. [41]

    Paladino, L

    E. Paladino, L. Faoro, G. Falci, and R. Fazio, Decoher- ence and 1/f noise in josephson qubits, Physical Review Letters https://doi.org/10.1103/PhysRevLett.88.228304 (2002)

    2002 doi

  34. [42]

    J. J. G. Ripoll, Quantum Information and Quantum Op- tics with Superconducting Circuits (Cambridge Univer- sity Press, 2022)

    2022

  35. [43]

    D’Arrigo, G

    A. D’Arrigo, G. Piccitto, G. Falci, and E. Paladino, Open-loop quantum control of small-size networks for high-order cumulants and cross-correlations sensing, Na- ture Scientific Reports https://doi.org/10.1038/s41598- 024-67503-x (2024)

    2024 doi

  36. [44]

    C. J. Wood, J. D. Biamonte, and D. G. Cory, Tensor net- works and graphical calculus for open quantum systems https://doi.org/10.48550/arXiv.1111.6950 (2015)

    2015 doi arXiv

  37. [45]

    Giannelli, P

    L. Giannelli, P. Sgroi, J. Brown, G. S. Paraoanu, M. Paternostro, E. Paladino, and G. Falci, A tu- torial on optimal control and reinforcement learning methods for quantum technologies, Physics Letters A https://doi.org/10.1016/j.physleta.2022.128054 (2022)

    2022

  38. [46]

    F. K. Wilhelm, S. Kirchhoff, S. Machnes, N. Wittler, and D. Sugny, An introduction into optimal control for quan- tum technologies, arXiv:2003.10132 [quant-ph] (2020)

    2020 arXiv

  39. [47]

    Liberzon, Calculus of Variations and Optimal Control Theory: A Concise Introduction (Princeton University Press, 2012)

    D. Liberzon, Calculus of Variations and Optimal Control Theory: A Concise Introduction (Princeton University Press, 2012)

    2012

  40. [48]

    Nocedal and S

    J. Nocedal and S. J. Wright, Numerical Optimization (Springer, 2006)

    2006

  41. [49]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, /suppress Lukasz Kaiser, and I. Polosukhin, Attention is all you need, Proceedings of the 31st Con- ference on Neural Information Processing Systems (NIPS

  42. [50]

    D. P. Kingma and J. Ba, Adam: A method for stochastic optimization, 3rd International Conference on Learning Representations, Conference Track Proceedings (2015)

    2015

  43. [51]

    Saha, Doing Math with Python (No Starch Press, Inc., 2015)

    A. Saha, Doing Math with Python (No Starch Press, Inc., 2015)

    2015

  44. [52]

    M. A. et al., Tensorflow: Large-scale machine learning on heterogeneous systems, (2015)

    2015

  45. [53]

    F. C. et al., Keras, (2015)

    2015

  46. [54]

    G. A. Paz-Silva, L. M. Norris, F. Beaudoin, and L. Viola, Extending comb-based spectral estimation to multiaxis quantum noise, Phys. Rev. Lett. A https://doi.org/10.1103/PhysRevA.100.042334 (2019)

    2019 doi

  47. [55]

    Z. Shen, M. Zhang, H. Zhao, S. Yi, and H. Li, Efficient attention: Attention with linear complexities, 2021 IEEE Winter Conference on Applications of Computer Vision (WACV) 10.1109/WACV48630.2021.00357 (2021)

    2021

  48. [56]

    K. He, X. Zhang, S. Ren, and J. Sun, Deep residual learn- ing for image recognition, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016) pp. 770–778

    2016

  49. [57]

    J. L. Ba, J. R. Kiros, and G. E. Hinton, Layer normal- ization, arXiv preprint arXiv:1607.06450 (2016)

    2016 arXiv

  50. [58]

    Hendrycks and K

    D. Hendrycks and K. Gimpel, Gaussian error linear units (gelus), arXiv preprint arXiv:1606.08415 (2016)

    2016 arXiv

  51. [59]

    M. Lin, Q. Chen, and S. Yan, Network in network, in International Conference on Learning Representations (ICLR) (2014)

    2014

  52. [60]

    Srivastava, G

    N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, Dropout: A simple way to pre- vent neural networks from overfitting, Journal of Ma- chine Learning Research 15, 1929 (2014)

    2014

  53. [61]

    Loshchilov and F

    I. Loshchilov and F. Hutter, Sgdr: Stochastic gradient descent with warm restarts, in International Conference on Learning Representations (ICLR) (2017)

    2017

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.