Pith. sign in

REVIEW 4 major objections 7 minor 21 references

Hybrid Quantum Convolutional Neural Network-Aided Pilot Assignment in Cell-Free Massive MIMO Systems

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 hybrid quantum convolutional neural network with a single shared parameterized quantum circuit can assign pilots in cell-free massive MIMO with near-optimal total throughput.

desk verdict A plausible but under-evidenced QML-for-pilot-assignment paper: the empirical claim is not backed by a train/test split or error bars, and the Appendix proof of Lemma 2 is flawed, but the core idea is worth a serious referee. read the letter →

arxiv 2507.06585 v1 pith:6JT5GEMK submitted 2025-07-09 cs.IT eess.SPmath.IT

Doan Hieu Nguyen , Xuan Tung Nguyen , Seon-Geun Jeong , Trinh Van Chien , Lajos Hanzo , Won Joo Hwang This is my paper · ORCID
classification cs.ITeess.SPmath.IT
keywords pilotassignmentcell-freemassiveMIMOhybridquantumconvolutionalneuralnetworkparameterizedcircuitcontaminationergodicsumratemachinelearningparametersharing
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 pilot assignment in cell-free massive MIMO, a combinatorial problem whose optimal solution normally requires exhaustive search, can be learned by a hybrid quantum convolutional neural network whose quantum convolutional layers all share one parameterized quantum circuit. The authors claim this shared-PQC design maps large-scale fading coefficients to pilot-selection probabilities, reaching about 98 percent of the exhaustive-search sum-rate in small systems while outperforming greedy, master-AP, location-based, MLP, and classical CNN baselines and converging in roughly 7 supervised training epochs. A sympathetic reader should care because, if true, it shows quantum machine learning with aggressive parameter sharing can tackle a wireless resource-allocation task with few trainable parameters and near-optimal throughput.

What carries the argument

The argument is carried by the parameterized quantum circuit treated as a shared convolutional kernel: every quantum convolution layer uses the same PQC (8 qubits, depth 6), so the number of quantum trainable parameters drops from the layer-specific design's $T\prod_{i=1}^{\log_2 n}2^i$ to $T\log_2(n_0)$. Angle embedding converts classical LSF inputs into qubit rotations, Pauli-Z expectation values form the features passed to a classical post-processing layer, and gradients are computed by the parameter-shift rule rather than backpropagation. The learnability guarantee is assembled in the appendix by composing classical and quantum universal approximation theorems, with the total approximation error bounded by a weighted sum of per-stage errors.

What would settle it

Compute the two sides of inequality (14) for an actually trained HQCNN by measuring $\|f(\beta;\theta)-g^*(\beta)\|_F$ and the weighted per-stage errors; if the inequality fails, the theoretical guarantee is not established. Also test generalization by training on one set of random user locations and evaluating on a disjoint set; a large drop in sum-rate would show the reported throughputs are memorization rather than a learned mapping.

Watch

Extended reading notes

Core claim

At the center is the claim that a QCNN with one parameterized quantum circuit reused across all quantum convolutional layers can approximate the optimal mapping from large-scale fading coefficients to pilot assignments. Concretely, the paper reports that with $(M,L,K,\tau_p)=(10,2,6,3)$ the HQCNN yields 18.08 Mbps versus the exhaustive search's 18.43 Mbps, and at $(45,2,20,10)$ it yields 133.75 Mbps against 111.05 Mbps for location-based assignment. It also claims faster convergence than classical deep models and than a QCNN with layer-specific parameters, and that zero-noise extrapolation partly restores performance under depolarizing noise.

Load-bearing premise

Everything rests on the assumption that the total approximation error is no larger than the weighted sum of the three stages' individual errors, and that the trained network generalizes to user deployments it was not trained on.

Editorial extensions

If this is right

  • HQCNN can be used as a drop-in pilot-assignment policy for small and medium cell-free massive MIMO systems, achieving near-optimal ergodic sum-rate without exhaustive search.
  • Reusing one PQC across all quantum convolutional layers cuts quantum parameters exponentially compared with layer-specific QCNN designs, making training and inference less costly.
  • Supervised training converges in about 7 epochs, so the policy can be retrained quickly when the large-scale fading statistics change.
  • Unsupervised training with the negative sum-rate as loss lets the model directly maximize throughput without needing labels from a heuristic.
  • Noise mitigation such as zero-noise extrapolation is needed to preserve the near-optimal performance on near-term quantum hardware.

Reading between the lines

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

  • A natural extension the paper leaves implicit is a train/test split over randomly generated user locations; the reported numbers appear to come from the same deployments used for training, so generalization to unseen deployments is untested.
  • The same shared-PQC convolution idea could be applied to other wireless combinatorial assignments, such as subchannel or power allocation, whenever the decision depends on an LSF-style channel-statistics matrix.
  • If the composition error bound in inequality (14) holds, the theory predicts that adding qubits should improve approximation accuracy in noiseless simulation but that hardware noise reverses this trend, which matches the paper's Fig. 3c and could be tested systematically.
  • The parameter-reduction scaling suggests HQCNN's per-parameter efficiency gap over heavy CNN should widen as $M$ and $K$ grow; plotting parameter efficiency against problem size would make that prediction explicit.
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 paper proposes a hybrid quantum convolutional neural network (HQCNN) for pilot assignment in cell-free massive MIMO systems. The architecture pre-processes large-scale fading coefficients with a classical linear layer, embeds them into quantum states, applies quantum convolution and pooling layers that share a single parameterized quantum circuit across layers, and post-processes expectation values into a pilot-assignment probability matrix. The authors train the model in both supervised mode (using Master-AP labels and cross-entropy) and unsupervised mode (using the negative sum-rate as the loss), and report numerical results showing near-exhaustive-search throughput, faster convergence than classical baselines, reduced parameter counts, and some robustness to depolarizing noise via zero-noise extrapolation.

Significance. If substantiated, the main contribution is an interesting and potentially useful reduction in trainable parameters for quantum convolutional models: using the same PQC across quantum convolution layers, as summarized in Table I, lowers the quantum parameter count from O(15n) to O(15). The numerical study also attempts a realistic comparison against classical CNNs, MLPs, heuristics, and exhaustive search, and the noise analysis with a depolarizing channel and ZNE is a valuable practical addition. However, the current evidence is not yet sufficient to support the headline claim that the HQCNN 'offers a total network throughput close to that of the excessive-complexity exhaustive search and outperforms the state-of-the-art benchmarks,' because the evaluation protocol lacks a train/test separation, error bars, and training hyperparameters, and the theoretical Lemma 2 proof has a gap. The central idea is worth pursuing, but the empirical and theoretical support need substantial strengthening.

major comments (4)
  1. [§IV-B, Table II] The evaluation does not report a train/test split, the number of independent LSF realizations, random seeds, or error bars. Since the input dimension is large relative to the number of reported scenarios, the model could memorize the training deployments rather than learn a generalizing pilot-assignment policy. Small reported differences, such as 18.08 Mbps for the proposed HQCNN versus 18.05 Mbps for HQCNN [11] and 18.43 Mbps for EPAS, are within the range where statistical fluctuation could change the conclusions. Please report test-set performance with confidence intervals over multiple seeds, and specify the number of training and test realizations, the optimizer, learning rate, batch size, and epochs.
  2. [Appendix, Eq. (14)] The composition error bound in step (a) of Eq. (14) does not follow from the cited triangle-inequality argument. For the composed function f = f2 ∘ f1 ∘ f0, bounding ∥f0 − g1*∥, ∥f1 − g2*∥, and ∥f2 − g3*∥ does not by itself bound ∥f − g*∥ unless Lipschitz constants and a decomposition of the ideal mapping g* are supplied. In addition, the applicability of the quantum universal approximation theorem [19] to this specific shared-PQC architecture with n = 8 qubits and depth-6 circuits is not established. Lemma 2 therefore currently lacks a valid proof; although the paper's central claim is empirical, the theorem should be corrected, weakened, or removed.
  3. [§III-B2, Eq. (9)] The parameter-shift rule as written uses a plus sign between the two shifted expectation values, so Eq. (9) reads ∂⟨M⟩(θ)/∂θ = (⟨M⟩(θ+s) + ⟨M⟩(θ−s))/(2 sin(s)). The standard parameter-shift rule is a difference, and Eq. (11)'s simplification to s = π/2 also presupposes the difference form. Please correct the sign and state explicitly whether the numerical results were obtained with the corrected rule; this is necessary for reproducibility of the training procedure.
  4. [§IV-A, Fig. 3a] The convergence comparison in Fig. 3a plots training-loss reduction only, which does not demonstrate that the proposed HQCNN generalizes better than the baselines. A model that converges quickly on training data can still overfit. Please show validation loss or test throughput as a function of training epoch, together with the final test performance for the supervised setting.
minor comments (7)
  1. [§II-A] There is a typo in 'Expliciting'; it should be 'Specifically' or 'Explicitly'.
  2. [§III-A2] The word 'containts' should be 'contains'.
  3. [Appendix] In the Appendix, 'the QCNN delighted as' should be 'the QCNN defined as'.
  4. [§III-A1] The notation θ1 is used both for the weight matrix of the preprocessing linear layer in the definition of h and for the QCNN parameters in Eq. (13) of the Appendix; please use distinct symbols to avoid ambiguity.
  5. [§IV-B] The phrase 'characterized the performance' should be 'characterize the performance'.
  6. [Introduction] The sentence 'converges faster than classical deep learning with a margin of 2% of the global optimum' is unclear: 2% of what quantity, and relative to which baseline? Please clarify.
  7. [General] No code or data availability statement is provided; making the simulation code and data-generation procedure public would substantially strengthen reproducibility, especially given the absence of error bars.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the HQCNN's reported throughput is an empirical optimization outcome, not a quantity that reduces to the training loss or to a self-citation by construction.

full rationale

The paper's central numerical claims are empirical evaluations of a trained model, not derivations that collapse into their own inputs. The system model and problem formulation in Section II are standard cell-free massive MIMO ingredients: the closed-form ergodic throughput in Lemma 1, the combinatorial pilot-assignment problem in (4), and the exhaustive-search baseline are all externally defined and do not presuppose the HQCNN's success. The unsupervised loss in (8) is explicitly the negative soft-probability version of the sum-rate objective, so it is intentionally aligned with the evaluation metric; however, Table II reports throughput computed from hard decoded pilot assignments via the discrete formula (3), not from the training loss itself, so the reported performance is not the loss by construction. No fitted parameter is baked into any derivation, and no quantity is defined in terms of the outcome it is supposed to predict. The Appendix proof of Lemma 2 contains an unproven composition bound in (14)(a), but that is a rigor or correctness concern, not circularity, and it is not used to generate the numerical results. The self-citations ([1], [10], [12]) are used for standard LMMSE formulas, the parameter-shift gradient rule, and prior QCNN designs; none is invoked as a uniqueness theorem or as a premise that already contains the paper's conclusion. The absence of an explicit train/test split and error bars is a legitimate concern about generalization evidence, but it is a methodological limitation rather than a circular reasoning step. Overall, the derivation chain is self-contained with respect to its external benchmarks, so the circularity burden is low.

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

The claims rest on the standard cell-free massive MIMO model and on the trained parameters of a small quantum-classical network. The model hyperparameters (8 qubits, depth 6, s=π/2) and the uniform power control are hand-chosen and affect the numerical results. The Appendix adds two approximation-theoretic assumptions that are not verified. No new physical entities are introduced.

free parameters (5)
  • Number of qubits n = 8
    Hand-chosen default; Fig. 3b/c vary n from 6 to 8 to test robustness, but all Table II results use 8 qubits.
  • PQC depth = 6
    Chosen 'to minimize decoherence risk' (Section IV.C) and used for all results.
  • Parameter shift value s = π/2
    Selected 'for simplicity' due to the periodic rotation structure (Section III.B.2, Eqs. 10-11).
  • Power control coefficients ηmk = 1/(L Σ_k γmk)
    Uniform power control assumed for all APs (Section IV); this choice affects the rate expression and is not optimized.
  • Trained HQCNN weights (θ0, θ1, θ2) = not disclosed
    The central empirical claim depends on the trained values of the pre-processing layer, PQC rotations, and post-processing layer, fitted to the training loss (7) or (8); values are not released.
assumptions (5)
  • domain assumption The i.i.d. Rayleigh fading channel model, LMMSE channel estimation, conjugate beamforming, and the closed-form ergodic throughput expression in Lemma 1 are taken as given from prior work [1], [4].
    Invoked in Section II to define the objective (4) and the unsupervised loss (8); the rate expression (3) is the foundation of all reported throughput values.
  • domain assumption The optimal pilot assignment depends only on the large-scale fading coefficients β, so the HQCNN input can be restricted to β.
    Stated in Section III.B.1: 'the achievable data rate of each user predominantly relies on its LSF with respect to other APs'; justifies the input representation and the problem mapping g*(β).
  • ad hoc to paper The classical and quantum universal approximation theorems [18], [19] apply to the specific shared-PQC HQCNN architecture with 8 qubits and depth-6 PQCs.
    Used in the Appendix proof of Lemma 2 (Eq. 14(b)) to claim the approximation error can be made arbitrarily small; the paper does not verify the conditions of [19] for this architecture.
  • ad hoc to paper The composition error bound (14)(a) is valid, i.e., the total error is bounded by the sum of per-stage errors weighted by w0 and w1.
    Invoked in the Appendix with citation [21, Ch. 5]; the inequality requires Lipschitz or boundedness conditions on f1 and f2 that are not established in the paper.
  • standard math The parameter-shift rule with s = π/2 provides unbiased gradients for the shared-PQC circuit.
    Standard QML result cited to [10] and used in Section III.B.2 (Eqs. 9-11) for training; accepted background, though the paper assumes the rule without proof.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hybrid Quantum Convolutional Neural Network-Aided Pilot Assignment in Cell-Free Massive MIMO Systems." pith.science (2026). https://pith.science/paper/6JT5GEMK

@misc{pith2026250706585,
  author       = {Pith},
  title        = {Pith review of: Hybrid Quantum Convolutional Neural Network-Aided Pilot Assignment in Cell-Free Massive MIMO Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6JT5GEMK}},
  note         = {Machine review of arXiv:2507.06585}
}
read the original abstract

A sophisticated hybrid quantum convolutional neural network (HQCNN) is conceived for handling the pilot assignment task in cell-free massive MIMO systems, while maximizing the total ergodic sum throughput. The existing model-based solutions found in the literature are inefficient and/or computationally demanding. Similarly, conventional deep neural networks may struggle in the face of high-dimensional inputs, require complex architectures, and their convergence is slow due to training numerous hyperparameters. The proposed HQCNN leverages parameterized quantum circuits (PQCs) relying on superposition for enhanced feature extraction. Specifically, we exploit the same PQC across all the convolutional layers for customizing the neural network and for accelerating the convergence. Our numerical results demonstrate that the proposed HQCNN offers a total network throughput close to that of the excessive-complexity exhaustive search and outperforms the state-of-the-art benchmarks.

Figures

Figures reproduced from arXiv: 2507.06585 by the authors.

Figure 1
Figure 1. Architecture of the proposed HQCNN model: the pre-processing layer for embedding classical data to quantum state; QCNN for processing quantum [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. (a) Illustration of a CNN layer; (b) QCNN layer; and (c) PQC design. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (a) Proportion of training loss reduction relative to the initial loss value versus the training epochs (supervised); (b) Parameter efficiency of learning [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 19 canonical work pages

  1. [19]

    Universal approximation theorem and error bounds for quantum neural networks and quantum reservoirs,

    L. Gonon et al., “Universal approximation theorem and error bounds for quantum neural networks and quantum reservoirs,” IEEE Trans. Neural Netw. Learn. Syst. , 2025

    work page 2025

  2. [11]

    Quantum convolutional neural network for classical data classification,

    T. Hur et al., “Quantum convolutional neural network for classical data classification,” Quantum Mach. Intell. , 2022

    work page 2022

  3. [1]

    Distributed Graph Neural Network Design for Sum Ergodic Spectral Efficiency Maximization in Cell-Free Massive MIMO,

    N. X. Tung et al., “Distributed Graph Neural Network Design for Sum Ergodic Spectral Efficiency Maximization in Cell-Free Massive MIMO,” IEEE Trans. Vehicular Tech, 2024

    work page 2024

  4. [2]

    Learning decentralized power control in cell-free massive MIMO networks,

    D. Yu et al., “Learning decentralized power control in cell-free massive MIMO networks,” IEEE Trans. Vehicular Tech, 2023

    work page 2023

  5. [3]

    Pilot Assignment with Approximation Ratio in Cell- Free Massive MIMO Systems,

    Z. Wang et al., “Pilot Assignment with Approximation Ratio in Cell- Free Massive MIMO Systems,” IEEE Trans. Vehicular Tech, 2024

    work page 2024

  6. [4]

    Cell-free massive MIMO versus small cells,

    H. Q. Ngo et al., “Cell-free massive MIMO versus small cells,” IEEE Trans. Wirel. Commun., 2017

    work page 2017

  7. [5]

    Learning-based downlink power allocation in cell-free massive MIMO systems,

    M. Zaher et al., “Learning-based downlink power allocation in cell-free massive MIMO systems,” IEEE Trans. Wirel. Commun. , 2022

    work page 2022

  8. [6]

    An efficient location-based pilot assignment in cell-free massive MIMO,

    T. H. Nguyen et al., “An efficient location-based pilot assignment in cell-free massive MIMO,” ICT Express, 2023

    work page 2023

Show all 21 references
  1. [7]

    Deep CNN-based pilot allocation scheme in massive mimo systems,

    K. Kim et al., “Deep CNN-based pilot allocation scheme in massive mimo systems,” KSII T Internet Info , 2020

    2020

  2. [8]

    Hybrid quantum-classical convolutional neural network model for image classification,

    F. Fan et al., “Hybrid quantum-classical convolutional neural network model for image classification,” IEEE Trans. Neural Netw. Learn. Syst. , 2023

    2023

  3. [9]

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

    Y . Du et al., “Quantum machine learning: A hands-on tutorial for ma- chine learning practitioners and researchers,” arXiv:2502.01146, 2025

    2025 arXiv

  4. [10]

    Quantum information processing, sensing, and com- munications: Their myths, realities, and futures,

    L. Hanzo et al., “Quantum information processing, sensing, and com- munications: Their myths, realities, and futures,” Proc. IEEE, 2025

    2025

  5. [12]

    Hybrid quantum convolutional neural networks for UWB signal classification,

    S.-G. Jeong et al., “Hybrid quantum convolutional neural networks for UWB signal classification,” IEEE Access, 2023

    2023

  6. [13]

    A tutorial on quantum convolutional neural networks (QCNN),

    S. Oh, J. Choi, and J. Kim, “A tutorial on quantum convolutional neural networks (QCNN),” in Proc. ICTC, 2020

    2020

  7. [14]

    Comparing quantum encoding techniques,

    N. Munikote, “Comparing quantum encoding techniques,” arXiv:2410.09121, 2024

    2024 arXiv

  8. [15]

    Towards quantum-native communication systems: New developments, trends, and challenges,

    X. Zhou et al., “Towards quantum-native communication systems: New developments, trends, and challenges,” arXiv:2311.05239, 2023

    2023 arXiv

  9. [16]

    Choose your quantum hardware,

    Google Quantum AI, “Choose your quantum hardware,” 2024. [Online]. Available: https://quantumai.google/qsim/choose hw

    2024

  10. [17]

    Mitiq: A software package for error mitigation on noisy quantum computers,

    R. LaRose et al., “Mitiq: A software package for error mitigation on noisy quantum computers,” Quantum, 2022

    2022

  11. [18]

    Multilayer feedforward networks are universal ap- proximators,

    K. Hornik et al., “Multilayer feedforward networks are universal ap- proximators,” Neural Networks, 1989

    1989

  12. [20]

    Saltelli, Global Sensitivity Analysis: the Primer

    A. Saltelli, Global Sensitivity Analysis: the Primer. John Wiley & Sons, 2008

    2008

  13. [21]

    Shalev-Shwartz et al., Understanding machine learning: From theory to algorithms

    S. Shalev-Shwartz et al., Understanding machine learning: From theory to algorithms. Cambridge University Press, 2014

    2014

Pith tools

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