Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Regularizing quantum loss landscapes by noise injection

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read This paper claims that inserting a tunable Pauli noise channel after each parameterized rotation of a variational quantum circuit exponentially suppresses the high-frequency Fourier modes of the loss, and that the resulting smoothed…

desk verdict A clean, honest heuristic paper: the noise-injection filtering formula is solid and the benchmarks show real improvement, but the central conjecture about high-frequency modes causing local minima is never directly tested and the gains might be generic annealing. read the letter →

arxiv 2505.08759 v1 pith:J3GUZQLT submitted 2025-05-13 quant-ph

Daniil S. Bagaev , Maxim A.Gavreev , Alena S. Mastiukova , Aleksey K. Fedorov , Nikita A. Nemkov This is my paper · ORCID
classification quant-ph
keywords noise-inducedregularizationquantumlosslandscapeslocalminimaFourierexpansionvariationalalgorithmsconvolutionalneuralnetworksWishartrandomfieldsheatequation
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

Training variational quantum algorithms is known to fail when the loss landscape is dominated by poor local minima, and this paper tries to show that a deliberate dose of noise can remove much of that obstruction. The protocol inserts a Pauli noise channel of strength $\mu$ after every parameterized rotation, which, by an exact algebraic identity, rescales each Fourier mode of order $m$ by $(1-\mu)^m$, exponentially damping high-frequency oscillations while preserving the low-frequency skeleton of the landscape. The paper's conjecture is that those damped high-frequency terms are what create most bad local minima, so a schedule that starts with strong smoothing and gradually removes the noise guides gradient descent into the right basin. In numerical experiments on random Wishart fields and on quantum convolutional neural networks, the regularized runs match or beat the best non-regularized solution several times more often than the baseline. If the conjecture holds, the method adds an inexpensive, hardware-implementable tool that combines with other mitigations such as the quantum natural gradient optimizer.

What carries the argument

The central object is the Pauli noise channel $E_P(\mu)$ with Kraus operators $\{\sqrt{1-\mu/2}\,I, \sqrt{\mu/2}\,P\}$, inserted after each parameterized Pauli rotation. In the Heisenberg picture the channel multiplies Pauli strings that commute with $P$ by $1$ and strings that anti-commute by $1-\mu$, so its only effect on the loss is to rescale every $\cos\phi_k$ and $\sin\phi_k$ by $1-\mu$; the homogeneity of Fourier modes then yields the closed form $L(\mu,\phi) = \sum_m (1-\mu)^m L_m(\phi)$ without ever computing the exponentially many coefficients. The same channel has a second, interpretive role: with $1-\mu = e^{-t}$ the smoothed loss satisfies the heat equation $\partial_t L = \Delta_\phi L$, so the regularization is exactly a diffusion of the landscape. In hardware the channel is implemented with one ancillary qubit per gate (a controlled-$P$ gate, with the noise strength set by the ancilla's rotation angle), and in simulation it only requires switching to density-matrix evolution.

What would settle it

Construct a small quantum loss function whose global minimum is produced by a high-order Fourier mode, for instance a one- or two-parameter loss combining a low-order term with a ripple of order $m \ge 3$ whose deepest well is the true minimum, so the smoothed loss $L(\mu,\phi)$ has its global minimum far from the true one; then run the paper's scheduled noise-injection optimization from many random starts on that family. If the regularized runs consistently find worse solutions than the non-regularized baseline, or the tracked basin jumps away from the true minimum as the noise is released, the conjecture that high-frequency modes are mainly responsible for poor minima fails for that family.

Watch

Extended reading notes

Core claim

The central claim is that smoothing the quantum loss function by noise injection is both exactly describable and practically beneficial. For a circuit built from Clifford gates and parameterized Pauli rotations $U_P(\phi) = e^{-i\phi P/2}$, inserting the dephasing-style channel $E_P(\mu)$ with Kraus operators $\{\sqrt{1-\mu/2}\,I, \sqrt{\mu/2}\,P\}$ after each rotation transforms the loss into $L(\mu,\phi) = \sum_m (1-\mu)^m L_m(\phi)$, where $L_m$ is the homogeneous degree-$m$ Fourier mode of the original loss. The reason is that each Pauli string in the Hamiltonian either commutes with $P$, in which case the noise leaves it alone, or anti-commutes, in which case both $\cos\phi_k$ and $\sin\phi_k$ acquire a factor $1-\mu$; by homogeneity, a mode of order $m$ acquires $(1-\mu)^m$. Setting $1-\mu = e^{-t}$ shows the smoothed loss obeys the heat equation $\partial_t L = \Delta_\phi L$, so regularization is thermalization of the landscape. The paper argues that the highly oscillatory higher-order terms are responsible for the majority of poor local minima, and that optimizing a schedule of decreasing noise therefore lands more often at good solutions; its experiments on Wishart random fields and on quantum convolutional neural networks up to ten qubits show consistent, several-fold increases in the probability of matching or beating the best non-regularized solution.

Load-bearing premise

The load-bearing premise is the conjecture that the global minimum of the smoothed loss sits close to the global minimum of the original loss, so following the smoothed landscape into a basin and then releasing the noise leads to a good solution; if high-frequency modes instead create or hide the true minimum, the smoothing systematically steers optimization away from it.

Editorial extensions

If this is right

  • The regularized loss is available through a single tunable parameter $\mu$; no access to the exponentially many Fourier coefficients is required, so the smoothing works wherever the noisy circuit itself can be run.
  • The protocol is resource-cheap: one ancillary qubit per channel in principle, resettable and reusable, which makes the method implementable on current hardware and in density-matrix simulation.
  • Because the method deforms the landscape rather than the optimizer's trajectory, it composes with other mitigation techniques such as the quantum natural gradient optimizer.
  • If the main conjecture is right, the method pays off precisely in the regime the paper probes: shallow circuits and models whose landscapes are rough but not dominated by barren plateaus.

Reading between the lines

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

  • Because the identity $L(\mu,\phi) = \sum_m (1-\mu)^m L_m(\phi)$ is purely algebraic, it transfers unchanged to any setting where parameters enter as $e^{i\phi P/2}$ with $P^2 = I$, so the same exponential-smoothing effect could apply to classical non-convex losses over orthogonal-group parameterizations whenever the noise channel can be simulated.
  • One testable implication the paper does not run is to track how the Hessian's eigenvalue distribution changes along the heat-equation flow; if the high-frequency conjecture is right, the spurious negative-curvature directions should collapse at small $\mu$ while the curvature near the true minimum is barely affected.
  • The paper's data show the improvement factor is roughly constant in the overparameterization ratio $\gamma$ even where $\gamma \ll 1$ makes local minima plentiful, hinting that the benefit may come partly from shortening the optimizer's trajectory rather than only from removing minima; this could be checked by comparing loss-decay curves under matched iteration budgets.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a noise-injection protocol for variational quantum algorithms and quantum machine learning models. For circuits consisting of Clifford gates and parameterized Pauli rotations, inserting after each rotation the associated Pauli noise channel E_P(µ) yields an exact spectral deformation of the loss: L(µ, φ) = Σ_m (1−µ)^m L_m(φ), exponentially suppressing high-frequency Fourier terms. The protocol can be implemented with an ancilla qubit on hardware or with density-matrix simulation in software. The authors present a toy QAOA example, large-scale statistical experiments on random Wishart hypertoroidal fields, and teacher–student QCNN experiments, reporting consistent improvements in solution quality. The central heuristic is that high-frequency Fourier terms are responsible for most poor local minima, so smoothing the landscape guides optimization to better basins; the paper explicitly labels this as a conjecture and discusses caveats about noise levels, overhead, and scaling.

Significance. If the empirical claims hold, this is a useful and inexpensive addition to the toolbox for training variational quantum circuits: it gives an exact, tunable spectral filter on the loss landscape, is implementable with a single ancilla, and is complementary to optimizer-level mitigations. The derivation of Eq. (11) is clean and is the strongest part of the paper, and the numerical study is extensive, with code provided. However, the central mechanism is a conjecture that the experiments do not isolate from generic homotopy/annealing effects, and the statistical evidence lacks confidence intervals and significance tests. With these issues addressed, the protocol would be a solid contribution; the current manuscript is defensible but needs revision before publication.

major comments (3)
  1. [II.E, Eq. (12)] The claim that L(t, φ) = Σ_m e^{−mt} L_m(φ) satisfies the heat equation ∂_t L = Δ_φ L is incorrect. For a one-parameter example, take L(φ) = cos²φ, so L(t, φ) = e^{−2t} cos²φ. Then ∂_t L = −2e^{−2t} cos²φ, while Δ_φ L = e^{−2t}(2 − 4 cos²φ), which are not equal. The factor e^{−mt} corresponds to the homogeneity generator D = Σ_k (cos φ_k ∂/∂ cos φ_k + sin φ_k ∂/∂ sin φ_k), not to the Laplacian. This is a mathematical error in a stated theoretical result, so the heat-equation interpretation should be corrected or removed.
  2. [III.C–D] The claim of 'significant and robust improvements' is not supported by statistical tests or confidence intervals. The WHRF experiments use 100 instances and 2000 optimizations each, which would allow bootstrap or permutation tests, but none are reported; the QCNN experiments use only 30 student circuits per size. In addition, the schedule hyperparameters µ_max = 0.9 and a = 10, and the learning rate, are selected 'through numerical experiments' without a held-out validation protocol, so the reported gains may partly reflect tuning on the same model families. Please report confidence intervals, significance tests, and a clear statement of how hyperparameters were chosen.
  3. [II.B and III] The central optimization mechanism is not directly tested. The paper's main conjecture is that high-frequency Fourier terms are responsible for the majority of poor local minima, so suppressing them brings the regularized global minimum close to the true one. The numerical comparisons of final loss and accuracy do not verify this: they do not show that the basin of the true global minimum is preferentially preserved, nor that the improvement is due to spectral suppression rather than to the generic annealing effect of a time-dependent loss. A concrete test would be, e.g., to compare against a control deformation with the same schedule but a different spectral filtering, or to track, in small instances where the true global minimum is known (as in the QCNN teacher–student setup), whether the regularized optimizer's endpoint approaches the teacher parameters. Without such evidence, the abstract's causal claim that the protocol 'works by exponentially suppressing' high-frequency components is stronger than what the numerics establish.
minor comments (5)
  1. [II.B, Eq. (6)] Equation (6) has a typo: 'L_m(λ cos φ_k, λ cos φ_k)' should read 'L_m(λ cos φ_k, λ sin φ_k)'.
  2. [III.C, Fig. 5 caption] In the caption of Fig. 5, 'correpond' should be 'correspond'.
  3. [III.C] The sentence beginning 'Our show that...' is incomplete; it should be 'Our results show that...'.
  4. [Abstract and IV] The abstract claims the method 'can be combined with the quantum natural gradient optimizer', but no numerical demonstration of this combination is provided. This is plausible and the paper explicitly frames it as a complementarity argument, but the claim should be either supported by an experiment or softened to a statement of expected compatibility.
  5. [References] Reference [15] is misformatted: it is assigned to ICML 2016 volume 4, pages 2726–2739 but with a 2023 arXiv number; this reference should be corrected or consolidated with [14].

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: Eq. (11) is a direct algebraic identity and the optimization benchmarks are external; the main conjecture is openly labeled as such.

full rationale

Eq. (11) is not a fit or a renamed input: it follows by applying the Heisenberg-picture action of the noise channel (Eqs. 8-10) to the Fourier decomposition (Eqs. 5-7) and using the homogeneity of the order-m terms (Eq. 6). The suppression of high-frequency modes is therefore an identity, and no parameter is tuned to make it true. The optimization claims are tested on random Wishart fields and teacher-student QCNNs, which are external benchmarks; the schedule parameters (mu_max=0.9, a=10) are hyperparameters and do not force the observed improvements. The central heuristic is explicitly stated as a conjecture (Sec. II.B) and the paper discusses the counter-case in Fig. 1b, so it is not disguised as a theorem. The only self-citation, [17], supplies an elementary Fourier recursion rule; the load-bearing derivation and the numerical tests are independent of it. I therefore find no circular reduction, only the acknowledged conjecture that high-frequency modes dominate bad local minima.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central theoretical result is derived from standard circuit and Pauli channel assumptions plus an explicitly labeled conjecture about local minima. No new physical entities are introduced. The freely chosen schedule hyperparameters (mu_max, a, learning rate) are hand-tuned and shared across experiments, which is a minor source of selection pressure in the empirical results.

free parameters (3)
  • mu_max = 0.9
    Maximum noise strength in the exponential schedule (Eq. 13), hand-selected through numerical experiments; controls how strongly high-frequency modes are suppressed and affects the reported improvements.
  • a = 10
    Decay rate in the schedule mu(i) = mu_max e^{-a i/imax}; chosen by numerical experiments and used across all reported experiments.
  • learning_rate = 0.5e-2
    ADAM learning rate chosen for all experiments; influences convergence and could affect comparisons between regularized and non-regularized runs.
assumptions (4)
  • domain assumption The parameterized circuit consists only of constant Clifford gates and parameterized Pauli rotations (Eq. 4)
    The derivation of spectral suppression relies on this circuit structure, stated at the start of Section II B; most common VQA circuits satisfy it.
  • standard math The Hamiltonian can be decomposed into a sum of Pauli strings, each of which either commutes or anti-commutes with each rotation generator
    The Heisenberg propagation rule (Eq. 7) assumes this Pauli decomposition; standard and generally valid for qubit Hamiltonians.
  • ad hoc to paper High-frequency Fourier terms are primarily responsible for superfluous local minima, and suppressing them brings the regularized global minimum close to the true one
    This is the main conjecture in Section II B and Fig. 1; it is not proven and is the reason regularization is expected to improve optimization.
  • domain assumption The injected noise is an ideal trace-preserving Pauli channel with no additional decoherence
    The exact form of Eq. (11) holds only for this idealized channel; the paper notes that real hardware noise may differ.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Regularizing quantum loss landscapes by noise injection." pith.science (2026). https://pith.science/paper/J3GUZQLT

@misc{pith2026250508759,
  author       = {Pith},
  title        = {Pith review of: Regularizing quantum loss landscapes by noise injection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J3GUZQLT}},
  note         = {Machine review of arXiv:2505.08759}
}
read the original abstract

The difficulty of training variational quantum algorithms and quantum machine learning models is well established. In particular, quantum loss landscapes are often highly non-convex and dominated by poor local minima. While this renders their training NP-hard in general, efficient heuristics that work well for typical instances may still exist. Here, we propose a protocol that uses a targeted noise injection to smooth and regularize quantum loss landscapes. It works by exponentially suppressing the high-frequency components in the Fourier expansion of the quantum loss function. The protocol can be efficiently implemented both in hardware and in simulations. We observe significant and robust improvements of solution quality across various problem types. Our method can be combined with existing techniques mitigating the local minima, such as the quantum natural gradient optimizer, and adds to the toolbox of methods for optimizing quantum loss functions.

Figures

Figures reproduced from arXiv: 2505.08759 by the authors.

Figure 1
Figure 1. FIG. 1: A sketch showing how high-frequency terms [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. FIG. 3: A single-layer QAOA loss landscape without (a) [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. FIG. 4: Several optimization trajectories traced by the [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (3 more)
Figure 7
Figure 7. Figure 7: FIG. 7: Optimization results for QCNN. The highest [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 6
Figure 6. Figure 6: FIG. 6: Quality of the regularized optimization as a [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 9
Figure 9. Figure 9: FIG. 9: Fraction of loss values in the fifth percentile as a [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. QUBO-based training for VQAs on Quantum Annealers

    quant-ph 2025-09 reject novelty 4.0 of 10

    A QUBO-based annealer training scheme with recursive refinement is tested on Iris, Heart Disease, and Diabetes, but the QUBO derivation has a critical gap.

Reference graph

Works this paper leans on

50 extracted references · 18 canonical work pages · cited by 1 Pith paper

  1. [1]

    Variational quantum algorithms

    M. Cerezo, Andrew Arrasmith, Ryan Babbush, Simon C. Benjamin, Suguru Endo, Keisuke Fujii, Jarrod R. Mc- Clean, Kosuke Mitarai, Xiao Yuan, Lukasz Cincio, and Patrick J. Coles. “Variational quantum algorithms”. Nature Reviews Physics 2021 3:9 3, 625–644 (2021). arXiv:2012.09265

    arXiv 2021

  2. [2]

    Quantum machine learning

    Jacob Biamonte, Peter Wittek, Nicola Pancotti, Patrick Rebentrost, Nathan Wiebe, and Seth Lloyd. “Quantum machine learning”. Nature549, 195–202 (2017)

    2017

  3. [3]

    The effect of data encoding on the expressive power of variational quantum machine learning models

    Maria Schuld, Ryan Sweke, and Johannes Jakob Meyer. “The effect of data encoding on the expressive power of variational quantum machine learning models”. Physical Review A103 (2020). arXiv:2008.08605v2

    arXiv 2020

  4. [4]

    The theory of variational hybrid quantum-classical algorithms

    Jarrod R. McClean, Jonathan Romero, Ryan Babbush, and Alán Aspuru-Guzik. “The theory of variational hybrid quantum-classical algorithms”. New Journal of Physics 18, 1–20 (2016). arXiv:1509.04279

    arXiv 2016

  5. [5]

    A Review of Barren Plateaus in Variational Quantum Computing

    Martin Larocca, Supanut Thanasilp, Samson Wang, Ku- nal Sharma, Jacob Biamonte, Patrick J. Coles, Lukasz Cincio, Jarrod R. McClean, Zoë Holmes, and M. Cerezo. “A Review of Barren Plateaus in Variational Quantum Computing” (2024). arXiv:2405.00781

    arXiv 2024

  6. [6]

    Cerezo, Samson Wang, Tyler Volkoff, Andrew T

    Arthur Pesah, M. Cerezo, Samson Wang, Tyler Volkoff, Andrew T. Sornborger, and Patrick J. Coles. “Absence 9 FIG. 9: Fraction of loss values in the fifth percentile as a function ofγ computed under discretization in 100 bins. Pastγ = 1 local minima disappear from the landscape. of barren plateaus in quantum convolutional neural net- works”. Physical Review...

    work page 2021

  7. [7]

    Does provable absence of barren plateaus imply classical simulability? Or, why we need to rethink variational quantum computing

    M. Cerezo, Martin Larocca, Diego García-Martín, N. L. Diaz, PaoloBraccia, EnricoFontana, ManuelS.Rudolph, Pablo Bermejo, Aroosa Ijaz, Supanut Thanasilp, Eric R. Anschuetz, and Zoë Holmes. “Does provable absence of barren plateaus imply classical simulability? Or, why we need to rethink variational quantum computing” (2023). arXiv:2312.09121

    arXiv 2023

  8. [8]

    Critical Points in Quantum Gener- ative Models

    Eric R. Anschuetz. “Critical Points in Quantum Gener- ative Models” (2021). arXiv:2109.06957

    arXiv 2021

Show all 50 references
  1. [9]

    Beyond Bar- ren Plateaus: Quantum Variational Algorithms Are Swamped With Traps

    Eric R. Anschuetz and Bobak T. Kiani. “Beyond Bar- ren Plateaus: Quantum Variational Algorithms Are Swamped With Traps” (2022). arXiv:2205.05786

    2022 arXiv

  2. [10]

    Training variational quantum algorithms is NP-hard – even for logarithmi- cally many qubits and free fermionic systems

    Lennart Bittel and Martin Kliesch. “Training variational quantum algorithms is NP-hard – even for logarithmi- cally many qubits and free fermionic systems” (2021). arXiv:2101.07267

    2021 arXiv

  3. [11]

    Deep learning

    Ian Goodfellow, Yoshua Bengio, and Aaron Courville. “Deep learning”. MIT Press. (2016). url: http://www.deeplearningbook.org

    2016

  4. [12]

    Visual recon- struction

    Andrew Blake and Andrew Zisserman. “Visual recon- struction”. The MIT Press. (1987)

    1987

  5. [13]

    On the Link be- tween Gaussian Homotopy Continuation and Convex En- velopes

    Hossein Mobahi and John W. Fisher. “On the Link be- tween Gaussian Homotopy Continuation and Convex En- velopes”. In Lecture Notes in Computer Science (includ- ing subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Volume 8932, pages 43–56...

    2015

  6. [14]

    On Graduated Optimization for Stochastic Non-Convex Problems

    Elad Hazan, Kfir Y. Levy, and Shai Shalev-Shwartz. “On Graduated Optimization for Stochastic Non-Convex Problems”. 33rd International Conference on Ma- chine Learning, ICML 2016 4, 2726–2739 (2015). arXiv:1503.03712

    2015 arXiv

  7. [15]

    Stochas- tic Gradient Descent in the Viewpoint of Gradu- ated Optimization

    Da Li, Jingjing Wu, and Qingrun Zhang. “Stochas- tic Gradient Descent in the Viewpoint of Gradu- ated Optimization”. 33rd International Conference on Machine Learning, ICML 2016 4, 2726–2739 (2023). arXiv:2308.06775

    2023 arXiv

  8. [16]

    Explicit and Im- plicit Graduated Optimization in Deep Neural Net- works

    Naoki Sato and Hideaki Iiduka. “Explicit and Im- plicit Graduated Optimization in Deep Neural Net- works” (2024). arXiv:2412.11501

    2024 arXiv

  9. [17]

    Fourier expansion in variational quantum al- gorithms

    Nikita A. Nemkov, Evgeniy O. Kiktenko, and Aleksey K. Fedorov. “Fourier expansion in variational quantum al- gorithms”. Phys. Rev. A108, 032406 (2023)

    2023

  10. [18]

    Learning Unitaries by Gradient Descent

    Bobak Toussi Kiani, Seth Lloyd, and Reevu Maity. “Learning Unitaries by Gradient Descent” (2020). arXiv:2001.11897

    2020 arXiv

  11. [19]

    Training Saturation in Layerwise Quantum Approxi- mate Optimisation

    E. Campos, D. Rabinovich, V. Akshay, and J. Biamonte. “Training Saturation in Layerwise Quantum Approxi- mate Optimisation” (2021). arXiv:2106.13814

    2021 arXiv

  12. [20]

    Efficient vari- ational synthesis of quantum circuits with coherent multi-start optimization

    Nikita A. Nemkov, Evgeniy O. Kiktenko, Ilia A. Luchnikov, and Aleksey K. Fedorov. “Efficient vari- ational synthesis of quantum circuits with coherent multi-start optimization”. Quantum 7, 993 (2023). arXiv:2205.01121

    2023 arXiv

  13. [21]

    An Empirical Review of Optimization Techniques for Quantum Variational Circuits

    Owen Lockwood. “An Empirical Review of Optimization Techniques for Quantum Variational Circuits” (2022). arXiv:2202.01389

    2022 arXiv

  14. [22]

    Performance com- parison of optimization methods on variational quantum algorithms

    Xavier Bonet-Monroig, Hao Wang, Diederick Vermetten, Bruno Senjean, Charles Moussa, Thomas Bäck, Vedran Dunjko, and Thomas E. O’Brien. “Performance com- parison of optimization methods on variational quantum algorithms”. Physical Review A 107, 032407 (2023). arXiv:2111.13454v3

    2023 arXiv

  15. [23]

    Quantum Natural Gradient

    JamesStokes, JoshIzaac, NathanKilloran, andGiuseppe Carleo. “Quantum Natural Gradient”. Quantum4(2020). arXiv:1909.02108

    2020 arXiv

  16. [24]

    Avoiding local minima in variational quantum eigensolvers with the natural gradient optimizer

    David Wierichs, Christian Gogolin, and Michael Kasto- ryano. “Avoiding local minima in variational quantum eigensolvers with the natural gradient optimizer”. Phys- ical Review Research2 (2020). arXiv:2004.14666

    2020 arXiv

  17. [25]

    Using Differential Evolu- tion to avoid local minima in Variational Quantum Al- gorithms

    Daniel Faílde, José Daniel Viqueira, Mariamo Mussa Juane, and Andrés Gómez. “Using Differential Evolu- tion to avoid local minima in Variational Quantum Al- gorithms”. Scientific Reports 2023 13:113, 1–10 (2023). arXiv:2303.12186

    2023 arXiv

  18. [26]

    Diffusion-Enhanced Optimization of Variational Quan- tum Eigensolver for General Hamiltonians

    Shikun Zhang, Zheng Qin, Yongyou Zhang, Yang Zhou, Rui Li, Chunxiao Du, and Zhisong Xiao. “Diffusion-Enhanced Optimization of Variational Quan- tum Eigensolver for General Hamiltonians” (2025). arXiv:2501.05666

    2025

  19. [27]

    Avoiding local minima in Variational Quantum Algorithms with Neural Net- works

    Javier Rivera-Dean, Patrick Huembeli, Antonio Acín, and Joseph Bowles. “Avoiding local minima in Variational Quantum Algorithms with Neural Net- works” (2021). arXiv:2104.02955

    2021 arXiv

  20. [28]

    Quantum Learning with Noise and Decoherence: A Robust Quantum Neural Network

    Nam H. Nguyen, Elizabeth C. Behrman, and James E. Steck. “Quantum Learning with Noise and Decoherence: A Robust Quantum Neural Network”. Quantum Machine Intelligence 2, 1 (2016). arXiv:1612.07593

    2016 arXiv

  21. [29]

    Method for noise-induced regularization in quantum neural net- works

    Wilfrid Somogyi, Ekaterina Pankovets, Viach- eslav Kuzmin, and Alexey Melnikov. “Method for noise-induced regularization in quantum neural net- works” (2024). arXiv:2410.19921

    2024

  22. [30]

    In [19] it was observed that a small amount of noise may prevent layer-wise training saturation in quan- tum approximate optimization algorithms

    made a similar observation for quantum kernel meth- ods. In [19] it was observed that a small amount of noise may prevent layer-wise training saturation in quan- tum approximate optimization algorithms. Ref. [31] ar- gued that statistical sampling noise can help avoid sad- dle...

  23. [31]

    Stochastic noise can be helpful for variational quantum algorithms

    Junyu Liu, Frederik Wilde, Antonio Anna Mele, Liang Jiang, and Jens Eisert. “Stochastic noise can be helpful for variational quantum algorithms” (2022). arXiv:2210.06723

    2022 arXiv

  24. [32]

    Noisy quantum ker- nel machines

    Valentin Heyraud, Zejian Li, Zakari Denis, Alexandre Le Boité, and Cristiano Ciuti. “Noisy quantum ker- nel machines”. Physical Review A106, 052421 (2022). arXiv:2204.12192

    2022 arXiv

  25. [33]

    Low-rank density-matrix evolution for noisy quan- tum circuits

    Yi-Ting Chen, Collin Farquhar, and Robert M. Par- rish. “Low-rank density-matrix evolution for noisy quan- tum circuits”. npj Quantum Information 7, 61 (2021). arXiv:2009.06657

    2021 arXiv

  26. [34]

    Classical sim- 10 ulations of noisy variational quantum circuits

    Enrico Fontana, Manuel S Rudolph, Ross Duncan, Ivan Rungger, and Cristina Cîrstoiu. “Classical sim- 10 ulations of noisy variational quantum circuits” (2023). arXiv:2306.05400

    2023 arXiv

  27. [35]

    Simulatingquantumcircuitexpectationvaluesby Clifford perturbation theory

    Tomislav Begušić, Kasra Hejazi, and Garnet Kin-Lic Chan. “Simulatingquantumcircuitexpectationvaluesby Clifford perturbation theory” (2023). arXiv:2306.04797

    2023 arXiv

  28. [36]

    Non-zero noise extrapolation: accu- rately simulating noisy quantum circuits with tensor net- works

    Anthony P. Thompson, Arie Soeteman, Chris Cade, and Ido Niesen. “Non-zero noise extrapolation: accu- rately simulating noisy quantum circuits with tensor net- works” (2025). arXiv:2501.13237

    2025

  29. [37]

    Bagaev (2024)

    D. Bagaev (2024). url: github.com/quantumoon/noise- induced-optimization

    2024

  30. [38]

    Adam: A method for stochastic optimization

    Diederik P. Kingma and Jimmy Lei Ba. “Adam: A method for stochastic optimization”. 3rd Interna- tional Conference on Learning Representations, ICLR 2015 - Conference Track ProceedingsPages 1–15 (2015). arXiv:1412.6980

    2015 arXiv

  31. [39]

    Orqviz: Visualiz- ing high-dimensional landscapes in variational quantum algorithms

    Manuel S. Rudolph, Sukin Sim, Asad Raza, Michal Stechly, Jarrod R. McClean, Eric R. Anschuetz, Luis Ser- rano, and Alejandro Perdomo-Ortiz. “Orqviz: Visualiz- ing high-dimensional landscapes in variational quantum algorithms” (2021). arXiv:2111.04695

    2021 arXiv

  32. [40]

    A quantum approximate optimization algorithm

    Edward Farhi, Jeffrey Goldstone, and Sam Gutmann. “A quantum approximate optimization algorithm” (2014). arXiv:1411.4028

    2014 arXiv

  33. [41]

    is a popular QML model that was shown to be free of BPs [6]. Although QCNN were recently shown to be classically simulable [7, 42], which undermines their practical utility, they are still an excellent test bed for our optimization technique as they were shown to feature numer...

    2000

  34. [42]

    Connecting the Hamil- tonian structure to the QAOA energy and Fourier land- scape structure

    MichałSt¸ echły, LanruoGao, BonifaceYogendran, Enrico Fontana, and Manuel Rudolph. “Connecting the Hamil- tonian structure to the QAOA energy and Fourier land- scape structure” (2023). arXiv:2305.13594

    2023 arXiv

  35. [43]

    Quan- tum convolutional neural networks

    Iris Cong, Soonwon Choi, and Mikhail D. Lukin. “Quan- tum convolutional neural networks”. Nature Physics15, 1273–1278 (2019)

    2019

  36. [44]

    Quantum Con- volutional Neural Networks are (Effectively) Classically Simulable

    Pablo Bermejo, Paolo Braccia, Manuel S. Rudolph, Zoë Holmes, Lukasz Cincio, and M. Cerezo. “Quantum Con- volutional Neural Networks are (Effectively) Classically Simulable” (2024). arXiv:2408.12739

    2024 arXiv

  37. [45]

    Kiani (2024)

    B. Kiani (2024). url: github.com/bkiani/Beyond-Barren- Plateaus

    2024

  38. [46]

    Quantum computing in the NISQ era and beyond

    John Preskill. “Quantum computing in the NISQ era and beyond”. Quantum2, 1–20 (2018). arXiv:1801.00862

    2018 arXiv

  39. [47]

    Quantum computing at the quantum ad- vantage threshold: a down-to-business review

    A. K. Fedorov, N. Gisin, S. M. Beloussov, and A. I. Lvovsky. “Quantum computing at the quantum ad- vantage threshold: a down-to-business review” (2022). arXiv:2203.17181

    2022 arXiv

  40. [48]

    Mitigating Quantum Gate Errors for Vari- ational Eigensolvers Using Hardware-Inspired Zero- Noise Extrapolation

    Alexey Uvarov, Daniil Rabinovich, Olga Lakhmanskaya, Kirill Lakhmanskiy, Jacob Biamonte, and Soumik Ad- hikary. “Mitigating Quantum Gate Errors for Vari- ational Eigensolvers Using Hardware-Inspired Zero- Noise Extrapolation”. Physical Review A 110 (2023). arXiv:2307.11156v3

    2023 arXiv

  41. [49]

    Noise-Mitigated Variational Quantum Eigensolver with Pre-training and Zero-Noise Extrapolation

    Wanqi Sun, Jungang Xu, and Chenghua Duan. “Noise-Mitigated Variational Quantum Eigensolver with Pre-training and Zero-Noise Extrapolation” (2025). arXiv:2501.01646

    2025 arXiv

  42. [50]

    Myths around quantum computation before full fault tolerance: What no-go theorems rule out and what they don’t

    Zoltán Zimborás, Bálint Koczor, Zoë Holmes, Elsi-Mari Borrelli, András Gilyén, Hsin-Yuan Huang, Zhenyu Cai, Antonio Acín, Leandro Aolita, Leonardo Banchi, Fer- nando G. S. L. Brandão, Daniel Cavalcanti, Toby Cu- bitt, Sergey N. Filippov, Guillermo García-Pérez, John Goold, Ors...

    2025 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.