Pith. sign in

REVIEW 4 major objections 6 minor 41 references

A physics-informed neural network, enhanced with Fourier features, per-epoch renormalization, and pretraining, matches Krotov and PRONTO in designing high-fidelity quantum gates while suppressing leakage to non-computational states.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-07-31 23:51 UTC pith:GSLGMGUG

load-bearing objection Useful three-way benchmark for QOC tooling, but the abstract's 'fidelities exceeding 99.9%' claim is contradicted by the paper's own tables and needs to be walked back. the 4 major comments →

arxiv 2607.23276 v1 pith:GSLGMGUG submitted 2026-07-25 quant-ph

Performance of Krotov, PRONTO and PINN for optimal control of quantum gates

Mart\'in D. Jim\'enez , Murilo D. Forlevesi , Emanuel F. de Lima , Leonardo K. Castelano This is my paper
classification quant-ph MSC 81Q9368T07
keywords quantum optimal controlphysics-informed neural networksKrotov methodPRONTOgate fidelitypopulation leakageFourier feature embeddingsfluxonium qubit
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper's goal is to show that a physics-informed neural network (PINN), after three specific training upgrades, can solve quantum optimal control problems as accurately as two established continuous-control solvers, Krotov and PRONTO. The upgrades—Fourier feature embeddings, per-epoch state renormalization, and pretraining on a known oscillatory field—overcome the spectral bias and non-convergence that plague standard PINNs on these problems. Benchmarks on a three-level fluxonium qubit and a four-level nitrogen-vacancy center yield gate fidelities above 99% (exceeding 99.9% in some cases), with PINNQOC sometimes achieving lower mean population leakage than the traditional solvers. The authors contend that although PINNQOC is slower in these small systems, its tensor-parallel structure and warm-starting potential make it a promising route to scalable, adaptive quantum hardware control.

Core claim

On its own terms, the central claim is that naive PINN training fails for gate optimization—loss functions oscillate and final-state populations plateau around 0.67–0.88—but that three modifications fix it: mapping the time input through Fourier features with random frequencies drawn from a Gaussian (σ=4), renormalizing the neural state vector to unit norm every epoch, and pretraining the network for 3000 epochs on a guess control field before optimization. With these changes, PINNQOC achieves a final population of 0.99999 for the three-level X gate and, across the benchmark set, fidelities comparable to Krotov and PRONTO (e.g., 99.97% vs 99.92% and 99.89% for the X gate). The paper also rep

What carries the argument

The load-bearing mechanism is the combination of three training modifications. First, Fourier feature embedding maps the scalar time input to a 32-dimensional sinusoidal feature vector with frequencies sampled from a Gaussian distribution of standard deviation 4, which lets the network represent high-frequency oscillatory quantum dynamics and alleviates spectral bias. Second, 'dynamic epoch normalization' projects the neural state vector to unit norm after every training epoch, enforcing probability conservation during training. Third, an informed pretraining routine first trains the network to reproduce the exact dynamics under a fixed guess field u₁(t)=cos(20πt/T) for 3000 epochs, providin

Load-bearing premise

The claim that PINNQOC matches traditional solvers rests on the assumption that the three modifications—Fourier features with σ=4, per-epoch renormalization, and pretraining with cos(20πt/T)—are the actual cause of convergence and will work on other systems without per-system tuning; this is not tested.

What would settle it

Run the same three-level fluxonium X-gate optimization with each modification ablated (e.g., no pretraining, or Fourier frequencies drawn from σ=1 or σ=8) and check whether gate fidelity stays above 99%. If any ablation causes fidelity to collapse, or if the method fails to converge on a Hamiltonian outside the paper's two models, the claim of parity with Krotov and PRONTO is not established.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • PINNQOC can be used as an alternative to Krotov and PRONTO for leakage-aware gate design in small multilevel systems, with comparable fidelity and sometimes better leakage suppression.
  • The method's continuous representation works with a moderate temporal grid (N_t=500), avoiding PRONTO's need for very fine grids (5000–50000 points) to maintain numerical stability.
  • A pretrained PINN is, in principle, warm-startable: when system parameters drift, the network could be recalibrated with far fewer epochs than training from scratch, unlike Krotov and PRONTO which must restart the whole optimization.
  • Because neural network operations are tensor-based and parallelizable, the authors expect the approach to exploit GPU/TPU hardware at larger Hilbert-space dimensions, although the paper does not demonstrate this scaling directly.

Where Pith is reading between the lines

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

  • The three fixes are presented as a general recipe, but the specific choices—Fourier frequency scale σ=4, the cosine pretraining field, and per-epoch renormalization—are fixed without a sensitivity study; if these need to be retuned per system, the claimed parity with traditional solvers would not transfer automatically.
  • The benchmark uses different temporal grids for each method (Krotov N_t=100, PRONTO 5000, PINNQOC 500), so some of the fidelity and leakage differences could be influenced by discretization rather than algorithmic merit.
  • A concrete test of the scalability narrative would be to measure warm-start speedup: if a pretrained PINN, after a Hamiltonian drift, reconverges in only a fraction of the original epochs, the closed-loop adaptive-control argument is supported; the paper does not provide such an experiment.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The manuscript benchmarks a physics-informed neural network method (PINNQOC), augmented with Fourier features, per-epoch normalization, and pretraining, against Krotov and PRONTO for single-qubit gate synthesis on a three-level fluxonium model and a four-level NV-center model. The authors report fidelity and leakage metrics and computational costs, claiming that PINNQOC matches Krotov/PRONTO with fidelities exceeding 99.9% while suppressing leakage.

Significance. If the claims were supported, the work would be a useful systematic benchmark and a step toward PINN-based warm-startable quantum control. The proposed modifications are concrete and the multi-method comparison across two models and three gate targets is a strength. However, the headline claims are contradicted by the paper's own tables, so the significance is currently not established; with corrected reporting and robustness analysis, the comparison could be a useful contribution.

major comments (4)
  1. [Abstract; Tables I–VI] The abstract claims 'fidelities exceeding 99.9%' and 'successfully suppresses population leakage,' but the paper's own Tables I–VI show PINNQOC mean fidelities of 0.99972 (Table I), 0.99819 (II), 0.98823 (III), 0.99701 (IV), 0.99861 (V), and 0.99241 (VI). Mean leakage is also worse than both Krotov and PRONTO in Table I (0.0381 vs 0.0270/0.0239), Table III (0.0325 vs 0.0294/0.0274), and Table V (0.2699 vs 0.1717/0.1206). The central quantitative claim is refuted by the manuscript's own data and must be reconciled.
  2. [VII A] Section VII A selectively highlights 'outstanding leakage suppression' for the four-level X and S gates (Tables IV, VI) while ignoring the four-level Hadamard case (Table V) where PINNQOC leakage is 0.2699, the worst of the three methods and roughly double PRONTO's 0.1206. The conclusion that all methods show parity in physical accuracy is therefore not supported across all benchmarks. The narrative should report every case without cherry-picking and qualify the general claims.
  3. [VI B; Eqs. (49), (51)] The three proposed fixes—Fourier features with σ=4 (Eq. 51), per-epoch normalization (Eq. 49), and pretraining with u1(t)=cos(20πt/T)—are presented as a general solution but no sensitivity analysis is given. The frequency vector B is sampled once, σ is fixed, and the pretraining field is fixed. The reader cannot tell whether these results are robust or tuned per system. Please provide variance across B draws, σ variation (or a physical rationale for σ=4), and at least one alternative pretraining field; also discuss whether the renormalization in Eq. (49) preserves the solution set of the original QOC problem.
  4. [VII C; Tables I–VI] The timing and iteration comparisons are confounded by unequal N_t values: Krotov uses N_t=100 (or 1000 for the three-level S gate), PRONTO N_t=5000, and PINNQOC N_t=500, and stopping criteria are not systematically varied. Time-to-solution comparisons are therefore not apples-to-apples. The authors should either match discretization and tolerance or provide a quantitative sensitivity analysis in N_t rather than only qualitative comments about grid resolution.
minor comments (6)
  1. [Eq. (38)] The loss notation 'L[1, θ2]' is a typo; it should read 'L[θ1, θ2]'.
  2. [Eq. (44)] The sum is over k but the expression uses j (|ψ_j^0⟩ and |φ_j⟩). The index should be consistent.
  3. [Figs. 5–6 vs. Sec. VII] Figure captions state 'no leakage penalization was included,' while the benchmark tables and Section VII use q=0.3 for the three-level system. Please clarify that these figures are preliminary runs without leakage penalization and that the tables use the penalized functional.
  4. [VI B] 'Dynamic epoch normalization' should be specified more precisely: Eq. (49) renormalizes each state at each collocation point at every epoch. The relation of this operation to the original optimization functional should be stated explicitly.
  5. [Abstract] Even after correcting the numerical claims, 'matching the efficacy of traditional solvers' should be qualified as holding for most—not all—benchmarked gates, given the S-gate fidelities and the Hadamard leakage.
  6. [IV] The four-level Hamiltonian is said to use atomic units and T=3 a.u., but the definition of the unit time is not given. A one-sentence definition would help reproducibility.

Circularity Check

0 steps flagged

No significant circularity: PINNQOC is benchmarked against independent external solvers and its reported metrics are re-evaluated from the optimized fields.

full rationale

Walking the paper's derivation chain, the central benchmark is not circular. Krotov and PRONTO are implemented as independent, external algorithms (Refs. [16, 18, 28]), and their control fields are not derived from or fitted to the PINN results. For PINNQOC, the control fields are obtained by minimizing the composite loss in Eqs. (38)–(46), but the reported figures of merit are not read out from the training loss: the mean fidelity F (Eq. 52) and mean leakage ℓ (Eq. 53) are evaluated by re-propagating the optimized fields over K=500 Haar-random initial states. This is an independent post-hoc evaluation, not a rename of the optimized objective. The pretraining step (Eq. 50) uses a generic guess field u1(t)=cos(20πt/T) and exact evolution data under that same field; it warm-starts the network but does not inject the target gate or the final optimized control. The Fourier feature embedding (Eq. 51) and per-epoch normalization (Eq. 49) are fixed hyperparameters/constraints, not quantities fitted to the benchmark outputs. The self-citations to the authors' earlier PINNverse work (Refs. [22, 23]) support background motivation about inverse problems but are not load-bearing for the present comparison, which is validated against independent Krotov and PRONTO implementations and against Haar-random-state metrics. The paper itself acknowledges that standard PINNQOC 'faces problems during the minimization of loss functions' (Sec. VI) and requires the proposed pretraining, normalization, and Fourier features; that is an honest limitation, not a circular step. Finally, the abstract's claim of 'gate fidelities exceeding 99.9%' and leakage suppression 'matching the efficacy of traditional solvers' appears contradicted by the paper's own Tables I–VI (e.g., Table III PINNQOC F=0.98823; Table V PINNQOC ℓ=0.2699 vs Krotov 0.1717 and PRONTO 0.1206). This is an internal-consistency or correctness problem, but it is not a circular-derivation problem: no reported quantity reduces by construction to an input of the same quantity.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The paper's numerical claims rest primarily on hand-chosen hyperparameters, domain truncations, and an unproven normalization trick. No new physical entities are introduced; the only new object is the PINN parameterization, which is a numerical method rather than an invented entity.

free parameters (6)
  • Fourier feature frequency scale σ = 4
    Hand-set bandwidth for the Fourier embedding in Eq. (51); controls the frequency range of representable control pulses and is not varied or ablated.
  • Leakage penalty weight q = 0.3 (3-level), 0.2 (4-level)
    Chosen by hand and different for the two systems; directly trades off gate fidelity against leakage suppression for all three methods.
  • PRONTO control regularization R = 8e-4 * I
    Hand-set to keep the Riccati/linear-quadratic subproblem well conditioned; affects convergence and the resulting pulse.
  • Krotov step widths λ = 1 (3-level), 10 (4-level)
    Hand-set update step widths for Krotov iterations; influence convergence speed and final field.
  • Initial/pretraining guess fields = cos(20πt/T) for 3-level; 15(t/T)exp(-(3t/T)^2) and its negative for 4-level
    The same guess field is used for pretraining PINNQOC and as the initial guess for Krotov/PRONTO; results may depend on this choice, but no sensitivity study is reported.
  • NN architecture and training schedule = 2 hidden layers x 128, sine activation, lr=1e-3, 3000 pretraining epochs
    Architecture and learning rate are chosen by hand with no ablation; the reported performance may be sensitive to these.
axioms (6)
  • domain assumption The Schrödinger equation with ℏ=1 and closed-system evolution correctly describes the control dynamics.
    Used throughout Eq. (1); ignores decoherence, crosstalk, and drive noise, which are mentioned as motivations but not modeled.
  • domain assumption Truncating to three and four levels captures the relevant leakage physics of the fluxonium and NV-C13 systems.
    Hamiltonians in Section VI.A are truncated without a convergence check against higher-level truncations.
  • ad hoc to paper Per-epoch renormalization in Eq. (49) preserves the solution set of the original optimal control problem.
    Introduced as a fix; the paper does not prove that the normalized NN outputs correspond to physical Schrödinger trajectories during training.
  • ad hoc to paper Fourier features with fixed σ=4 provide sufficient frequency coverage for the optimal control pulses.
    Eq. (51) samples frequencies once from a Gaussian with hand-set scale; no analysis shows this band is adequate for other systems or gate times.
  • domain assumption PRONTO's quasi-Newton/Riccati formulation, as described in Refs. [18,28], converges on these quantum control problems.
    The paper relies on cited PRONTO theory rather than re-deriving or verifying its convergence guarantees for these specific systems.
  • domain assumption Mean fidelity over 500 Haar-random states and time-averaged mean leakage are appropriate performance metrics.
    Eqs. (52) and (53) define the reported metrics; these are not directly optimized, and the J0<1e-3 stopping criterion does not guarantee high Haar-average fidelity for phase gates.

pith-pipeline@v1.3.0-alltime-deepseek · 18798 in / 14858 out tokens · 146119 ms · 2026-07-31T23:51:26.530822+00:00 · methodology

0 comments
read the original abstract

Achieving scalable quantum computing demands high-fidelity operations capable of mitigating population leakage into non-computational states. Physics-Informed Neural Networks (PINNs) have recently emerged as a powerful paradigm to unify quantum hardware characterization (inverse problems) and pulse engineering (direct problems), laying the foundational architecture for autonomous quantum processors. However, standard PINN frameworks face severe numerical bottlenecks, such as spectral bias, when attempting to simultaneously solve highly oscillatory multi-level dynamics and optimize continuous control fields under strict global phase constraints. In this work, we propose an enhanced PINN scheme for quantum optimal control (PINNQOC) that circumvents these limitations by incorporating Fourier feature embeddings, dynamic epoch normalization, and an informed pre-training routine. To rigorously evaluate its performance, we systematically benchmark our framework against two premier continuous control solvers: the first-order Krotov method and the second-order Projection Operator Newton Method for Trajectory Optimization (PRONTO). These techniques are applied to implement multiple quantum gates on a truncated three-level fluxonium qubit and a four-level Nitrogen-Vacancy center coupled to a Carbon-13 nuclear spin. Our advanced PINNQOC approach successfully suppresses population leakage while achieving gate fidelities exceeding 99.9$\%$, matching the efficacy of traditional solvers. Finally, we provide a comprehensive analysis of computational times, iteration efficiency, and mean leakage, highlighting the distinct trade-offs and avenues for embedding physics-guided machine learning into automated quantum hardware pipelines.

Figures

Figures reproduced from arXiv: 2607.23276 by Emanuel F. de Lima, Leonardo K. Castelano, Mart\'in D. Jim\'enez, Murilo D. Forlevesi.

Figure 1
Figure 1. Figure 1: FIG. 1. (a) Loss functions according to Eqs. (39) and (40) as [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. (a) Loss functions according to Eqs. (43) and (44) as [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. (a) Loss functions according to Eqs. (39) and (40) as [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. (a) Loss functions according to Eq. (46) as a function [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. (a) Gate functional according to Eq. (3) as a function [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

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

41 extracted references · 5 linked inside Pith

  1. [1]

    For simplicity, we choosef T (tn) = (1−t n/T) and f0(tn) =t n/Tand the loss function given in Eq. (39) becomes L4[θ1, θ2] = NtX n=1 NX j=1 1 Nt f0(tn)fT (tn) d|φ[θ1,θ2] j (tn)⟩ dt + fT (tn)−f 0(tn) T |φ[θ1,θ2] j (tn)⟩+ ψtgt j T − ψ0 j T +i(H 0 +H c[θ2, tn]) h fT (tn) ψ0 j +f 0(tn) ψtgt j +f 0(tn)fT (tn) φ[θ1,θ2] j (tn) Ei 2 .(46) The numerical procedure t...

    2025

  2. [2]

    M. A. Nielsen and I. L. Chuang,Quantum Computation and Quantum Information(Cambridge University Press, 2000)

    2000

  3. [3]

    Breuer and F

    H.-P. Breuer and F. Petruccione,The Theory of Open Quantum Systems(Oxford University Press, Oxford, 2002)

    2002

  4. [4]

    J. M. Gambetta, A. D. C´ orcoles, S. T. Merkel, and et al., Characterization of addressability by simultaneous ran- domized benchmarking, Physical Review Letters109, 240504 (2012)

    2012

  5. [5]

    D. C. McKay, C. J. Wood, S. Sheldon, and et al., Three- qubit randomized benchmarking, Physical Review Let- ters122, 200502 (2019)

    2019

  6. [6]

    Jiang, J

    X. Jiang, J. Scott, M. Friesen, and M. Saffman, Sensitiv- ity of quantum gate fidelity to laser phase and intensity noise, Phys. Rev. A107, 042611 (2023)

    2023

  7. [7]

    C. J. Wood and J. M. Gambetta, Quantification and characterization of leakage errors, Phys. Rev. A97, 032306 (2018)

    2018

  8. [8]

    Hyypp¨ a, A

    E. Hyypp¨ a, A. Veps¨ al¨ ainen, M. Papiˇ c, C. F. Chan, S. Inel, A. Landra, W. Liu, J. Luus, F. Marxer, C. Ockeloen-Korppi, S. Orbell, B. Tarasinski, and J. Heinsoo, Reducing leakage of single-qubit gates for su- perconducting quantum processors using analytical con- trol pulse envelopes, PRX Quantum5, 030353 (2024)

    2024

  9. [9]

    Chiaro and Y

    B. Chiaro and Y. Zhang, Active leakage cancellation in single qubit gates, Phys. Rev. Lett.135, 130601 (2025)

    2025

  10. [10]

    Werninghaus, D

    M. Werninghaus, D. J. Egger, F. Roy, S. Machnes, F. K. Wilhelm, and S. Filipp, Leakage reduction in fast super- conducting qubit gates via optimal control, npj Quantum Information7, 14 (2021)

    2021

  11. [11]

    A. P. Babu, J. Tuorila, and T. Ala-Nissila, State leak- age during fast decay and control of a superconducting transmon qubit, npj Quantum Information7, 30 (2021)

    2021

  12. [12]

    J. O. del Moral, O. S. Larrarte, R. Dastbasteh, J. E. Martinez, and R. M. Otxoa, Impact of leakage on the dynamics of an st 0 qubit implemented in a double quan- tum dot device, Phys. Rev. B113, 094310 (2026)

    2026

  13. [13]

    S. J. Glaser, U. Boscain, T. Calarco, C. P. Koch, W. K¨ ockenberger, R. Kosloff, I. Kuprov, B. Luy, S. Schirmer, T. Schulte-Herbr¨ uggen, D. Sugny, and F. K. Wilhelm, Training schr¨ odinger’s cat: quantum opti- mal control, The European Physical Journal D69, 279 (2015)

    2015

  14. [14]

    Khaneja, T

    N. Khaneja, T. Reiss, C. Kehlet, T. Schulte-Herbr¨ uggen, and S. J. Glaser, Optimal control of coupled spin dy- namics: design of nmr pulse sequences by gradient as- cent algorithms, Journal of Magnetic Resonance172, 296 (2005)

    2005

  15. [15]

    Caneva, T

    T. Caneva, T. Calarco, and S. Montangero, Chopped random-basis quantum optimization, Phys. Rev. A84, 022326 (2011)

    2011

  16. [16]

    Doria, T

    P. Doria, T. Calarco, and S. Montangero, Optimal control technique for many-body quantum dynamics, Phys. Rev. Lett.106, 190501 (2011)

    2011

  17. [17]

    M. H. Goerz, D. Basilewitsch, F. Gago-Encinas, M. G. Krauss, K. P. Horn, D. M. Reich, and C. P. Koch, Kro- tov: A Python implementation of Krotov’s method for quantum optimal control, SciPost Phys.7, 080 (2019)

    2019

  18. [18]

    M. H. Goerz, D. M. Reich, and C. P. Koch, Optimal control theory for a unitary operation under dissipative evolution, New Journal of Physics16, 055012 (2014)

    2014

  19. [19]

    J. Shao, M. Naris, J. Hauser, and M. M. Nicotra, Solv- ing quantum optimal control problems using projection- operator-based newton steps, Phys. Rev. A109, 012609 (2024)

    2024

  20. [20]

    M. E. F. Fernandes, F. F. Fanchini, E. F. de Lima, and L. K. Castelano, Effectiveness of the krotov method in finding controls for open quantum systems, Journal of Physics A: Mathematical and Theoretical56, 495303 (2023)

    2023

  21. [21]

    Raissi, P

    M. Raissi, P. Perdikaris, and G. Karniadakis, Physics- informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, Journal of Computational Physics378, 686 (2019). 15

    2019

  22. [22]

    De Ryck and S

    T. De Ryck and S. Mishra, Numerical analysis of physics- informed neural networks and related models in physics- informed machine learning, Acta Numerica33, 633–713 (2024)

    2024

  23. [23]

    G. G. De Lima, I. Cunha, and L. K. Castelano, Inverse physics-informed neural networks for noise detection in open quantum systems, Physics Letters A582, 131548 (2026)

    2026

  24. [24]

    L. K. Castelano, I. Cunha, F. S. Luiz, R. de Jesus Napoli- tano, M. V. d. S. Prado, and F. F. Fanchini, Combin- ing physics-informed neural networks with the freezing mechanism for general hamiltonian learning, Phys. Rev. A110, 032607 (2024)

    2024

  25. [25]

    Norambuena, M

    A. Norambuena, M. Mattheakis, F. J. Gonz´ alez, and R. Coto, Physics-informed neural networks for quantum control, Phys. Rev. Lett.132, 010801 (2024)

    2024

  26. [26]

    Bukov, A

    M. Bukov, A. G. R. Day, D. Sels, P. Weinberg, A. Polkovnikov, and P. Mehta, Reinforcement learning in different phases of quantum control, Phys. Rev. X8, 031086 (2018)

    2018

  27. [27]

    Magesan, J

    E. Magesan, J. M. Gambetta, and J. Emerson, Scalable and robust randomized benchmarking of quantum pro- cesses, Phys. Rev. Lett.106, 180504 (2011)

    2011

  28. [28]

    Carroll, S

    M. Carroll, S. Rosenblatt, P. Jurcevic, I. Lauer, and A. Kandala, Dynamics of superconducting qubit relax- ation times (2022), arXiv:2105.15201 [quant-ph]

    Pith/arXiv arXiv 2022

  29. [29]

    J. Shao, J. Combes, J. Hauser, and M. M. Nicotra, Projection-operator-based newton method for the trajec- tory optimization of closed quantum systems, Phys. Rev. A105, 032605 (2022)

    2022

  30. [30]

    C. C. Margossian, A review of automatic differentiation and its efficient implementation, WIREs Data Mining and Knowledge Discovery9, e1305 (2019)

    2019

  31. [31]

    L. B. Nguyen, Y.-H. Lin, A. Somoroff, R. Mencia, N. Grabon, and V. E. Manucharyan, High-coherence flux- onium qubit, Phys. Rev. X9, 041041 (2019)

    2019

  32. [32]

    D. D. Awschalom, R. Hanson, J. Wrachtrup, and B. B. Zhou, Quantum technologies with optically interfaced solid-state spins, Nature Photonics12, 516 (2018)

    2018

  33. [33]

    F. J. Gonz´ alez and R. Coto, Decoherence-protected quan- tum register of nuclear spins in diamond, Quantum Sci- ence and Technology7, 025015 (2022)

    2022

  34. [34]

    Sitzmann, J

    V. Sitzmann, J. Martel, A. Bergman, D. Lindell, and G. Wetzstein, Implicit neural representations with peri- odic activation functions, inAdvances in Neural Informa- tion Processing Systems, Vol. 33, edited by H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin (Curran Associates, Inc., 2020) pp. 7462–7473

    2020

  35. [35]

    Lauten and M

    S. Lauten and M. Otten, Physics-informed neural net- works for gate design using quantum optimal control (2025), arXiv:2511.09463 [quant-ph]

    arXiv 2025

  36. [36]

    Basri, M

    R. Basri, M. Galun, A. Geifman, D. Jacobs, Y. Kasten, and S. Kritchman, Frequency bias in neural networks for input of non-uniform density (2020), arXiv:2003.04560 [cs.LG]

    Pith/arXiv arXiv 2020

  37. [37]

    Tancik, P

    M. Tancik, P. P. Srinivasan, B. Mildenhall, S. Fridovich- Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. T. Barron, and R. Ng, Fourier features let networks learn high frequency functions in low dimensional domains (2020), arXiv:2006.10739 [cs.CV]

    Pith/arXiv arXiv 2020

  38. [38]

    Ash and R

    J. Ash and R. P. Adams, On warm-starting neural net- work training, Advances in neural information processing systems33, 3884 (2020)

    2020

  39. [39]

    Perrone, R

    V. Perrone, R. Jenatton, M. W. Seeger, and C. Ar- chambeau, Scalable hyperparameter transfer learning, Advances in neural information processing systems31 (2018)

    2018

  40. [40]

    W. Cho, M. Jo, H. Lim, K. Lee, D. Lee, S. Hong, and N. Park, Parameterized physics-informed neu- ral networks for parameterized pdes, arXiv preprint arXiv:2408.09446 (2024)

    Pith/arXiv arXiv 2024

  41. [41]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhat- tacharya, A. Stuart, and A. Anandkumar, Fourier neu- ral operator for parametric partial differential equations (2021), arXiv:2010.08895 [cs.LG]

    Pith/arXiv arXiv 2021