Pith. sign in

REVIEW 5 major objections 5 minor 23 references

A two-stage scheme — convex warm start followed by nonconvex Riemannian refinement — is claimed to overcome barren plateaus in variational quantum circuits by removing the coordinate-induced condition number from the convergence rate.

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 ↗

A two-stage convex/nonconvex least-squares algorithm is claimed to remove the condition-number barrier in variational quantum optimization and achieve high-fidelity BB84 quantum-state cloning.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection The paper's theory rests on an unproved geodesic-convexity assumption and a proof gap; the one defensible result is standard, and the experiments are too thin to carry the barren-plateau claim. the 5 major comments →

arxiv 2601.18060 v4 pith:KLDQHMVZ submitted 2026-01-26 quant-ph cs.ITmath.IT

Overcoming Barren Plateaus in Variational Quantum Circuits using a Two-Step Least Squares Approach

Francis Boabang , Samuel Asante Gyamerah This is my paper
classification quant-ph cs.ITmath.IT MSC 81P6868Q1290C26 PACS 03.67.Lx03.67.Dd
keywords barren plateauvariational quantum circuittwo-stage least squaresquantum least squaresRiemannian optimizationcondition numberBB84quantum cloning
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 reading

The paper aims to solve the barren-plateau problem in variational quantum algorithms, where gradients vanish exponentially as circuits grow. It proposes a two-stage optimizer: first fit a convex, regularized quantum least-squares surrogate to obtain a good warm start, then switch to the original nonconvex objective and refine using Riemannian gradient descent. The central theoretical claim is that this second stage eliminates the parameterization-induced condition number from the linear convergence bound, leaving only intrinsic geometric conditioning. To demonstrate, the method is used to clone BB84 quantum states with high fidelity, outperforming random-initialization counterparts and approaching the phase-covariant cloning limit.

Core claim

The authors claim that barren plateaus are worsened by the coordinate parameterization of quantum circuits: the pullback metric G(θ)=J(θ)^T J(θ) has a condition number κ_coord that inflates the effective condition number in Euclidean updates. By performing the unregularized least-squares refinement as Riemannian gradient descent on the unitary manifold, using grad L(θ)=G(θ)^{-1}∇L(θ) and a retraction step, the contraction constant becomes (1−ημ) with no explicit κ_coord factor (Theorem 3). Combined with a first stage that solves a regularized quantum least squares to enter a well-conditioned region, the algorithm is claimed to escape barren plateaus and to clone BB84 states with high fidelit

What carries the argument

The central object is the two-stage least-squares scheme: Stage 1 uses a convex ridge-regression surrogate f(w)=‖Φw−b‖²+λ‖w‖² to produce a warm start outside the barren plateau; Stage 2 optimizes the unregularized objective f(w)=‖Φw−b‖² on the Riemannian manifold induced by the state parameterization, via the retraction update θ⁺ = Retr_θ(−η G(θ)^{-1}∇L(θ)). The key identity is the contrast between μ_E = μ λ_min(G), L_E = L λ_max(G), giving κ_E = κ_M κ_coord for Euclidean updates, versus the Riemannian rate that cancels the κ_coord factor, reducing iteration complexity by Θ(κ_coord).

Load-bearing premise

The proof that the nonconvex refinement stage converges linearly with no condition-number term assumes, without evidence, that the unregularized least-squares objective is geodesically μ-strongly convex and L-smooth on a geodesically convex neighborhood of the optimum for the given variational circuit; no realistic ansatz is shown to satisfy this.

What would settle it

Train the two-stage algorithm on a 20-qubit, 8-layer hardware-efficient ansatz and measure the gradient variance of the nonconvex objective; if it decays as Θ(2^{-n}) rather than Ω(poly(n)), the claim that the convex warm start prevents exponential gradient collapse is falsified. Independently, estimate the geodesic strong-convexity constant μ in Eq. (23) by finite differences along geodesics; if μ ≤ 0 in the region visited, Theorem 3's linear contraction bound does not apply.

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

If this is right

  • If the gradient-variance guarantee holds, exponentially vanishing gradients do not appear in the second stage, making deeper and more expressive circuits trainable.
  • The theoretical removal of κ_coord predicts a speed-up of the nonconvex refinement stage relative to Euclidean gradient descent, with no added quantum per-iteration cost beyond estimating the natural gradient.
  • The two-stage prescription is concrete: solve a regularized quantum least squares first, then an unregularized one, with the switch triggered by the normalized gradient norm.
  • For quantum cryptanalysis, the framework implies that variational cloning of QKD states can approach the phase-covariant cloning bound, so security analyses should assume near-optimal cloning attacks by variational adversaries.

Where Pith is reading between the lines

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

  • Editorial inference: The geodesic strong-convexity assumption in Theorem 3 could be verified for specific ansatz families by computing the spectral gap of the Riemannian Hessian on hardware-efficient circuits; without such verification, the theory remains conditional, though the BB84 experiments suggest the practical effect is real.
  • Editorial inference: The surrogate linearization in Eq. (7) invites a direct test: if φ_i^T w poorly approximates the true expectation for a given observable, the convex stage may place the warm start in a different basin than assumed, and the method's fidelity gains would need an alternative explanation.
  • Editorial inference: A security-relevant consequence: if variational attackers can clone BB84 states at the phase-covariant bound, then practical QKD implementations with limited coherence may face stronger attacks than standard security proofs assume, motivating more robust post-processing or additional verification rounds.
  • Editorial inference: The method may transfer to other nonconvex quantum objectives, such as ground-state preparation in naive VQE, provided a convex surrogate exists; identifying such surrogates for chemistry problems is a natural next step.
Share X Bluesky LinkedIn Reddit HN

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

5 major / 5 minor

Summary. The paper proposes a two-stage variational quantum optimization framework: a convex warm-start stage based on a regularized least-squares surrogate, followed by a nonconvex refinement stage that minimizes the unregularized Hamiltonian expectation. The central theoretical claim is that performing the refinement stage with Riemannian gradient descent (preconditioning by the pullback metric G(θ)) removes the coordinate-induced condition number κcoord from the convergence rate, yielding a linear contraction whose complexity depends only on intrinsic quantities (Theorem 3, Eqs. 32–33). The authors further derive an end-to-end cost expression (Eq. 38), claim that Stage 1 produces a polynomial-gradient-variance warm start (Corollary 1), and report PennyLane simulations for cloning BB84 states, comparing their two-stage method against random-initialization nonconvex training and a phase-covariant cloning baseline (Section 4, Figures 1–2).

Significance. If the theoretical claims were established, the paper would offer a practically useful prescription: a convex initialization that escapes barren-plateau regions and a manifold-based refinement that avoids coordinate-conditioning slowdowns. The application to variational quantum cloning of BB84 states is also timely and could be of interest to quantum cryptanalysis. The paper's strength is that it attempts to separate the costs of the regularized and unregularized least-squares stages and to articulate a geometric argument for preconditioning. However, the significance is currently only conditional: the load-bearing assumptions are not proved or numerically verified, and the proof of the main Euclidean convergence lemma contains a gap. The empirical section has no error bars or code release, so the reported comparison is suggestive rather than demonstrative.

major comments (5)
  1. [Section 3.2.2, Assumption A1] Assumption A1 states that L_convex(θ) is µ-strongly convex and L_c-smooth as a function of circuit parameters θ. The text only proves convexity of the surrogate f(w) = ||Φw−b||^2 + λ||w||^2 in the auxiliary weight w (Section 3.2.1). The map θ ↦ ⟨ψ(θ)|O_i|ψ(θ)⟩ is generally nonconvex in θ, as the paper itself notes in Remark 1. Strong convexity of the parameterized landscape is therefore not established, and all Stage-1 convergence and finite-termination results (Lemmas B.1–B.2, Theorem 1) rest on this unverified premise.
  2. [Section 3.2.5, Theorem 3] Theorem 3 assumes that L_ref is geodesically µ-strongly convex and L-smooth on a geodesically convex neighborhood U of the optimum. This is the entire basis for the headline rate (1−ηµ)^t in Eq. (33) and for the claimed removal of κcoord. No argument is given that this holds for the PQC-3 ansatz or for the BB84 cloning Hamiltonian; geodesic convexity on the unitary manifold is a strong condition that does not follow from Stage-1 local-observable convexity. Without a circuit-dependent verification or a constructive condition implying the assumption, Eqs. (32)–(38) are conditional statements, not a demonstrated improvement over Euclidean descent.
  3. [Appendix A.4, Lemma B.4] The proof of summability of squared Stage-2 gradients is incomplete. It defines Δ_t(θ) = L_convex(θ) − E(θ) and writes L_refine^{(t+1)}(θ_{t+1}) = L_refine^{(t)}(θ_{t+1}) + Δ_t(θ_{t+1}), but never bounds Δ_t(θ_{t+1}) or its partial sums. The conclusion Σ_t ||∇L_refine^{(t)}(θ_t)||^2 < ∞ therefore does not follow from Lemma B.3. This is a load-bearing gap because Theorem 1's claim that gradients vanish and all accumulation points are stationary points of E depends on this lemma.
  4. [Section 3.2.1, Eq. (7)] The convex stage relies on the linear surrogate ⟨ψ(θ)|O_i|ψ(θ)⟩ ≈ φ_i^T w. The paper does not specify the feature map φ_i, the construction of the matrix Φ, or any accuracy bound for this approximation. Since Stage 1 is claimed to produce a warm start that avoids barren plateaus, the validity of this surrogate is essential; without an error estimate, the convex stage may not be related to the actual quantum landscape at all.
  5. [Corollary 1 and Section 3.2.5] Corollary 1 asserts that Stage 1 lands in a region where gradient variance scales as O(poly(n)) and that Stage 2 remains in this region. Appendix A.6 cites Cerezo et al. (2021a) for polynomial variance of local-observable gradients, but the refinement stage minimizes ⟨ψ(θ)|H|ψ(θ)⟩ with a general Hamiltonian H, not a local observable. No argument is given that the convex warm start keeps the iterates inside the polynomial-variance region for the nonconvex objective, so the barren-plateau avoidance claim is not supported by the stated proof.
minor comments (5)
  1. [General] The paper contains numerous typographical and formatting errors: 'VAQ' for VQA, 'reimannian' for Riemannian, 'dependance' for dependence, inconsistent spacing in equations, and unnumbered Appendix B sections. A thorough copyedit is needed.
  2. [Section 3.3, Eqs. (39)–(41)] Equations (39)–(41) are not well-formed: the bra-ket expressions contain ambiguous placements of brackets, and the definitions of ⟨Z_B⟩ and ⟨Z_E⟩ are written with unbalanced notation. The cloning fidelity formulas need to be rewritten precisely.
  3. [Section 4, Figure 1] The experimental section reports a single simulator run with no error bars, no standard deviation, and no code or data repository. Given that the theoretical claims are already conditional, the empirical evidence is too thin to independently establish that the two-stage method outperforms random initialization across circuit depths.
  4. [Algorithm 1] The stopping rule in the pseudocode uses ||∇L_convex(θ)||^2 / dim(θ) > τ_g, while the text in Section 3.2.1 and Appendix A uses the opposite inequality. This inconsistency must be resolved.
  5. [Section 3.2.3] The cost model introduces constants C1 and C2 and an approximate optimal λ* without defining how these constants relate to the actual quantum algorithm; the expression λ*_2-step ≈ sqrt(C1 σ²_max σ²_min / (C2 κ_A)) is heuristic. This should be clearly labeled as an order-of-magnitude estimate, not an exact analysis.

Circularity Check

1 steps flagged

Theorem 3's κ-removal is definitional: the Riemannian gradient G^{-1}∇ cancels coordinate conditioning by construction, and the promised linear contraction is exactly the assumed geodesic strong-convexity/smoothness premise.

specific steps
  1. self definitional [Section 3.2.5, Theorem 3 (Eqs. 31–33); also Section 3.2.4, Eqs. (13)–(16)]
    "Assume that, restricted to a geodesically convex neighborhood U around θ⋆, the quantum unregularized least squares refinement objective L_ref is (i) geodesically µ-strongly convex and (ii) geodesically L-smooth on the manifold (M,⟨·,·⟩). Let quantum unregularized least squares performs Riemannian gradient descent with a retraction: θt+1 = Retr_θt(−η grad L_ref(θt)), 0<η≤1/L. Then the iterates remain in U and satisfy the linear contraction d(θt+1,θ⋆)≤(1−ηµ)d(θt,θ⋆)... Crucially, the contraction constant (1−ηµ) contains no explicit factor κ from the parameterization."

    The theorem's premise already states that L_ref is geodesically μ-strongly convex and L-smooth with intrinsic constants μ,L; the advertised contraction (1−ηµ) is the textbook consequence of that premise for Riemannian gradient descent. The 'no explicit κ' conclusion is wired into the definition of the Riemannian gradient as G^{-1}∇ (Eq. 13/29), which preconditions away the coordinate metric by construction. The paper then elevates this conditional, definitional statement into the headline result that the algorithm 'reduces the dependence on the condition number' and uses it in the end-to-end cost (38). Since the geodesic-convexity assumption is not verified for the PQC-3/BB84 experiments, the claimed reduction is equivalent to the input assumption rather than an independently established p

full rationale

The paper is mostly a conditional mathematical exposition, and its BB84 experiment is an independent comparison against random initialization and the phase-covariant cloning bound; there is no obvious data-fitting circularity and no load-bearing self-citation chain. However, the central theoretical contribution—the claimed removal of the parameterization-induced condition number by Riemannian manifold optimization—is definitional: the Riemannian gradient is defined as G^{-1}∇, so the absence of κ_coord in the contraction constant follows immediately from the assumed geodesic strong convexity/smoothness. The theorem's premise (geodesic convexity of L_ref on U) is not established for any concrete PQC and is exactly what guarantees the advertised rate. Thus the 'reduction' is a consequence of the setting rather than a derived property of the quantum least-squares problem. The unproven premise and the unbounded Δ_t gap in Appendix A.4 are correctness risks, but the main circularity is the construction-level identification of the conclusion with the definition/assumption. This warrants a partial circularity score rather than a clean bill.

Axiom & Free-Parameter Ledger

6 free parameters · 8 axioms · 1 invented entities

The paper introduces no new physics, but relies on several unverified assumptions and ad hoc modeling choices; the 'Hilmaton' term is an unexplained neologism.

free parameters (6)
  • regularization parameter λ = λ*_2-step ≈ sqrt(C1 σ^2_max σ^2_min / (C2 κ_A)) (approximate formula)
    Chosen to balance conditioning and bias in the cost model; not fixed experimentally.
  • Stage-1 stopping threshold τ_g
    Used in Algorithm 1; value not specified.
  • Learning rates η_c, η_n
    Step sizes in Stages 1 and 2; not tuned systematically.
  • Epoch split (100 convex / 450 nonconvex) = 100/450
    Chosen by hand; no sensitivity analysis.
  • Cost-model constants C1, C2
    Introduced in Eq. (6) and the λ* derivation; no values.
  • Feature matrix Φ
    The transformation from θ to the linear surrogate in Eq. (7) is never defined; effectively a free modeling choice.
axioms (8)
  • domain assumption A1: L_convex is μ-strongly convex and L_c-smooth
    Assumption A1 (Section 3.2.2) is not proven for any quantum circuit; Remark 1 admits the circuit parameterization remains nonconvex.
  • domain assumption A2: Hamiltonian energy E is L_H-smooth
    Assumption A2, standard but unverified for generic Hamiltonians in the paper's setting.
  • standard math A3: Components L_convex, E are bounded below
    Common regularity condition for descent methods.
  • standard math A4: Step-size conditions η_c<2/L_c, η_n<2/L_H
    Standard descent step-size restrictions.
  • ad hoc to paper Theorem 3: L_ref is geodesically μ-strongly convex and L-smooth on a neighborhood U of the manifold
    This is the key unverified assumption; without it the contraction rate (1−ηµ) and κ-removal break down.
  • ad hoc to paper Linear surrogate: ⟨ψ(θ)|O_i|ψ(θ)⟩ ≈ φ_i^T w
    Eq. (7) assumes the quantum expectation values are linear in an auxiliary weight vector; no construction of Φ is given.
  • ad hoc to paper Corollary 1: Stage-1 warm start lands in a polynomial-gradient-variance region
    Asserts gradient variance is O(poly(n)) after Stage 1, based on a citation, but the mechanism is not proven.
  • domain assumption Runtime of quantum linear-system solvers is proportional to condition number κ
    Standard for quantum linear system algorithms; used to derive cost formulas.
invented entities (1)
  • Hilmaton landscape no independent evidence
    purpose: Rename the quantum energy landscape in which the two-stage optimization operates
    The word 'Hilmaton' is used throughout (abstract, Sections 1, 3) but never defined; it appears to be an invented name for the optimization landscape, not a physical entity, and provides no falsifiable handle.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Overcoming Barren Plateaus in Variational Quantum Circuits using a Two-Step Least Squares Approach." pith.science (2026). https://pith.science/paper/KLDQHMVZ

@misc{pith2026260118060,
  author       = {Pith},
  title        = {Pith review of: Overcoming Barren Plateaus in Variational Quantum Circuits using a Two-Step Least Squares Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KLDQHMVZ}},
  note         = {Machine review of arXiv:2601.18060}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Variational Quantum Algorithms are a vital part of quantum computing. It is a blend of quantum and classical methods for tackling tough problems in machine learning, chemistry, and combinatorial optimization. Yet as these algorithms scale up, they cannot escape the barren-plateau phenomenon. As systems grow, gradients can vanish so quickly that training deep or randomly initialized circuits becomes nearly impossible. To overcome the barren plateau problem, we introduce a two-stage optimization framework. First comes the convex initialization stage. Here, we shape the quantum energy landscape, the Hilmaton landscape, into a smooth, low-energy basin. This step makes gradients easier to spot and keeps noise from derailing the process. Once we have gotten a stable gradient flow, we move to the second stage: nonconvex refinement. In this phase, we let the algorithm wander through different energy minima, making the model more expressive. We show that our proposed algorithm theoretically reduces the dependence on the condition number of the underlying quantum least squares approximate matrix via Riemannian manifold optimization. Finally, we used our two-stage solution to perform quantum cryptanalysis of quantum key distribution protocol (i.e., BB84) to determine the optimal cloning strategies. The simulation results showed that our proposed two-stage solution outperforms its random initialization counterpart.

Figures

Figures reproduced from arXiv: 2601.18060 by Francis Boabang, Samuel Asante Gyamerah.

Figure 1
Figure 1. Figure 1: Average cloning fidelity vs number of layers [PITH_FULL_IMAGE:figures/full_fig_p014_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Average Fidelity vs iteration Next, we tested the optimal attack by fixing the number of qubit at 10 and using a 5-layer ansatz [PITH_FULL_IMAGE:figures/full_fig_p014_2.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

23 extracted references · 3 canonical work pages

  1. [1]

    Arrasmith, A

    A. Arrasmith, A. Holmes, M. Cerezo, and P. J. Coles. Equivalence of local and global costs for variational quantum algorithms. Quantum, 5: 0 558, 2021

    2021

  2. [2]

    Pennylane: Automatic differentiation of hybrid quantum-classical computations

    Ville Bergholm, Josh Izaac, Maria Schuld, Christian Gogolin, Shahnawaz Ahmed, Vishnu Ajith, M Sohaib Alam, Guillermo Alonso-Linaje, B AkashNarayanan, Ali Asadi, et al. Pennylane: Automatic differentiation of hybrid quantum-classical computations. arXiv preprint arXiv:1811.04968, 2018

    Pith/arXiv arXiv 2018

  3. [3]

    Refining Optimization Methods for Training Machine Learning Models: A Case Study in Robotic Surgical Procedures

    Francis Boabang. Refining Optimization Methods for Training Machine Learning Models: A Case Study in Robotic Surgical Procedures. PhD thesis, Concordia University, 2024

    2024

  4. [4]

    Mauro D'Ariano, and Chiara Macchiavello

    Dagmar Bru , Mirko Cinchetti, G. Mauro D'Ariano, and Chiara Macchiavello. Phase-covariant quantum cloning. Phys. Rev. A, 62: 0 012302, Jun 2000. doi:10.1103/PhysRevA.62.012302. URL https://link.aps.org/doi/10.1103/PhysRevA.62.012302

    work page doi:10.1103/physreva.62.012302 2000

  5. [5]

    Cerezo, A

    M. Cerezo, A. Arrasmith, R. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, K. Mitarai, X. Yuan, L. Cincio, and P. J. Coles. Cost function-dependent barren plateaus in shallow parametrized quantum circuits. Nature Communications, 12 0 (1): 0 1791, 2021 a . doi:10.1038/s41467-021-21728-w. URL https://doi.org/10.1038/s41467-021-21728-w

    doi:10.1038/s41467-021-21728-w 2021

  6. [6]

    Variational quantum algorithms

    Marco Cerezo, Andrew Arrasmith, Ryan Babbush, Simon C Benjamin, Suguru Endo, Keisuke Fujii, Jarrod R McClean, Kosuke Mitarai, Xiao Yuan, Lukasz Cincio, et al. Variational quantum algorithms. Nature Reviews Physics, 3 0 (9): 0 625--644, 2021 b

    2021

  7. [7]

    Quantum regularized least squares

    Shantanav Chakraborty, Aditya Morolia, and Anurudh Peduri. Quantum regularized least squares. Quantum, 7: 0 988, 2023

    2023

  8. [8]

    Variational quantum cloning: Improving practicality for quantum cryptanalysis

    Brian Coyle, Mina Doosti, Elham Kashefi, and Niraj Kumar. Variational quantum cloning: Improving practicality for quantum cryptanalysis. arXiv preprint arXiv:2012.11424, 2020

    Pith/arXiv arXiv 2012

  9. [9]

    Quantum key distribution as a quantum machine learning task

    Thomas Decker, Marcelin Gallezot, Sven Florian Kerstan, Alessio Paesano, Anke Ginter, and Wadim Wormsbecher. Quantum key distribution as a quantum machine learning task. npj Quantum Information, 11 0 (1): 0 140, 2025

    2025

  10. [10]

    Quantum-inspired support vector machine

    Chen Ding, Tian-Yi Bao, and He-Liang Huang. Quantum-inspired support vector machine. IEEE Transactions on Neural Networks and Learning Systems, 33 0 (12): 0 7210--7222, 2022. doi:10.1109/TNNLS.2021.3084467

    arXiv 2022

  11. [11]

    Quantum machine learning: A hands-on tutorial for machine learning practitioners and researchers

    Yuxuan Du, Xinbiao Wang, Naixu Guo, Zhan Yu, Yang Qian, Kaining Zhang, Min-Hsiu Hsieh, Patrick Rebentrost, and Dacheng Tao. Quantum machine learning: A hands-on tutorial for machine learning practitioners and researchers. arXiv preprint arXiv:2502.01146, 2025

    Pith/arXiv arXiv 2025

  12. [12]

    Grant, L

    E. Grant, L. Wossnig, M. Ostaszewski, and M. Benedetti. An initialization strategy for addressing barren plateaus in parametrized quantum circuits. Quantum, 3: 0 214, 2019

    2019

  13. [13]

    Cerezo, and Patrick J

    Alex Holmes, Kunal Sharma, M. Cerezo, and Patrick J. Coles. Connecting ansatz expressibility to gradient magnitudes and barren plateaus. PRX Quantum, 3 0 (1): 0 010313, 2022 a

    2022

  14. [14]

    Cerezo, and Patrick J

    Zachary Holmes, Kunal Sharma, M. Cerezo, and Patrick J. Coles. Connecting expressibility to trainability in quantum machine learning. Physical Review Letters, 129 0 (9): 0 090502, 2022 b . doi:10.1103/PhysRevLett.129.090502. URL https://doi.org/10.1103/PhysRevLett.129.090502

    doi:10.1103/physrevlett.129.090502 2022

  15. [15]

    Kiani, and Patrick J

    Carlos Ortiz Marrero, Benyamin T. Kiani, and Patrick J. Coles. Entanglement-induced barren plateaus. Quantum Science and Technology, 6 0 (4): 0 045024, 2021

    2021

  16. [16]

    McClean, Sergio Boixo, Vadim N

    Jarrod R. McClean, Sergio Boixo, Vadim N. Smelyanskiy, Ryan Babbush, and Hartmut Neven. Barren plateaus in quantum neural network training landscapes. Nature Communications, 9 0 (1): 0 4812, 2018. doi:10.1038/s41467-018-07090-4. URL https://doi.org/10.1038/s41467-018-07090-4

    doi:10.1038/s41467-018-07090-4 2018

  17. [17]

    Breaking through barren plateaus: Reinforcement learning initializations for deep variational quantum circuits

    Yifeng Peng, Xinyi Li, Zhemin Zhang, Samuel Yen-Chi Chen, Zhiding Liang, and Ying Wang. Breaking through barren plateaus: Reinforcement learning initializations for deep variational quantum circuits. arXiv preprint arXiv:2508.18514, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Sparse regularization via convex analysis

    Ivan Selesnick. Sparse regularization via convex analysis. IEEE Transactions on Signal Processing, 65 0 (17): 0 4481--4494, 2017

    2017

  19. [19]

    On a quantum inspired approach to train machine learning models

    Jean Michel Sellier. On a quantum inspired approach to train machine learning models. Applied AI Letters, 4 0 (4): 0 e89, December 2023. doi:10.1002/ail2.89. URL https://doi.org/10.1002/ail2.89

    work page doi:10.1002/ail2.89 2023

  20. [20]

    Introduction to variational quantum algorithms

    Micha St e ch y. Introduction to variational quantum algorithms. arXiv preprint arXiv:2402.15879, 2024

    Pith/arXiv arXiv 2024

  21. [21]

    Wang and D

    N. Wang and D. Suter. Provable and robust matrix completion via parsumi: A proximal alternating scheme. International Journal of Computer Vision, 111: 0 28--55, 2013

    2013

  22. [22]

    Direct robust matrix factorizatoin for anomaly detection

    Liang Xiong, Xi Chen, and Jeff Schneider. Direct robust matrix factorizatoin for anomaly detection. In 2011 IEEE 11th International Conference on Data Mining, pages 844--853, 2011. doi:10.1109/ICDM.2011.52

    work page doi:10.1109/icdm.2011.52 2011

  23. [23]

    Escaping from the barren plateau via gaussian initializations in deep variational quantum circuits

    Kaifeng Zhang, Li Liu, Min-Hsiu Hsieh, and Dacheng Tao. Escaping from the barren plateau via gaussian initializations in deep variational quantum circuits. In Advances in Neural Information Processing Systems, volume 35, pages 18612--18627, 2022

    2022

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.