{"id":"473e2e80-799c-49e0-a91d-6046577c4600","arxiv_id":"2508.03327","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A hybrid quantum-classical neural network trained by reinforcement learning is claimed to beat a CNN for massive MIMO user scheduling in simulated Rician channels.","lead":"This paper combines a variational quantum circuit with classical neural networks to schedule users in massive MIMO systems using statistical channel information. It reports higher sum-rates than a CNN baseline in simulation, though the evidence is thin because no code, error bars, or hyperparameters are given.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central outperformance claim is contingent on the approximate sum-rate expression (6) being an unbiased proxy for true ergodic sum-rate; without Monte-Carlo validation of (6) in the tested regimes, the QNN-vs-CNN advantage may be an artifact of the surrogate reward.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the approximate sum-rate expression (6) is used both as the training reward and as the performance metric, so the central QNN-vs-CNN comparison is only as valid as that approximation. I agree this is the most load-bearing issue because if Eq. (6) is inaccurate or misorders policies, the headline 'outperforms in sum-rate' claim collapses even if the reported curves are self-consistent. The concrete Monte-Carlo test would settle the concern directly. I considered other issues such as the under-specified CNN baseline, missing seeds, and the ambiguous role of proportional fairness, but those affect confidence and reproducibility rather than the internal validity of the surrogate-metric concern. The appropriate verdict remains conditional: the paper should not be accepted as establishing the central claim until Eq. (6) is validated in the tested regimes or the evaluation is redone with the true ergodic sum-rate.","tokens_in":6968,"tokens_out":10691,"duration_ms":134434,"concrete_test":"Generate a held-out set of Rician channel realizations from the same model (same K-factor, M, K, SNR grid). For the trained QNN and CNN policies, compute (a) the approximate reward from Eq. (6) and (b) a Monte-Carlo estimate of the true ergodic sum-rate in Eq. (4) by averaging log2(1+SINR) over many H_l realizations. Report both metrics and their rankings. If the QNN's advantage over CNN under Eq. (6) is not reproduced under Eq. (4), or if Eq. (6) deviates from Eq. (4) by more than a small tolerance across the tested SNR and user configurations, the central claim is not established. As a secondary check, also report the approximation error of Eq. (6) against Monte Carlo for a fixed schedule.","verdict_should_be":"UNCHANGED","load_bearing_attack":"All learning and evaluation in the paper flows through Eq. (6): it is used as the reward in Algorithm 1 ('sumrate reward') and presumably as the y-axis metric in Figs. 4-6. Eq. (6) replaces the expectation over fading in Eq. (4) with a deterministic function of the beam-domain covariance matrix U_l, and the paper provides no evidence in these specific Rician scenarios (K-factor, M, K, SNR) that the approximation matches the true ergodic sum-rate or preserves the ordering of scheduling policies. If Eq. (6) is biased or misranks schedules, then the trained QNN is optimized for a surrogate objective, and the headline result that the hybrid-QNN outperforms the CNN benchmark in 'sum-rate' may not hold for the actual ergodic sum-rate. This is load-bearing because every figure that supports the central claim uses the same approximate quantity, so the comparison is self-consistent but not necessarily valid for the claimed objective. The reference to [8] is not sufficient unless the tested parameters fall in the validated regime, which the manuscript does not show.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid quantum-classical neural network (QNN) for user scheduling in a single-cell massive MIMO downlink system with statistical CSI. The scheduler is trained with a policy-gradient RL algorithm using a beam-domain approximate sum-rate expression as reward, and the architecture combines classical pre/post neural network layers with a variational quantum circuit. Simulation results report sum-rate comparisons against a CNN baseline across different user-antenna configurations and SNR values, and the authors claim faster convergence and higher sum-rate, especially in dense, high-SNR regimes. The manuscript is best understood as a proof-of-concept study: the problem formulation is sensible and the architectural idea is interesting, but the current evidence for the central outperformance claim is incomplete.","tokens_in":7235,"tokens_out":4931,"duration_ms":59390,"significance":"If the claims were fully validated, the work would be a useful proof-of-concept for applying hybrid quantum-classical models to wireless scheduling, an area where quantum machine learning applications are still scarce. The use of statistical CSI to avoid instantaneous CSI overhead, the RL-based training formulation, and the explicit comparison to a CNN baseline are all appropriate components. However, the paper currently ships no code, no complete hyperparameter configuration, and no validation of the approximate reward surrogate that drives all reported results. Because the evaluation metric is the same as the training reward, the comparison is self-consistent but not yet convincing as an actual ergodic-sum-rate advantage. The contribution is potentially valuable, but the evidence presented is preliminary.","major_comments":[{"comment":"The approximate sum-rate expression in Eq. (6) is used both as the reward in Algorithm 1 and as the performance metric in Figs. 4-6. The manuscript does not validate this approximation against the exact ergodic sum-rate in Eq. (4) for the specific Rician scenarios tested (K-factor, M, K, and SNR ranges). If Eq. (6) systematically misranks candidate schedules relative to Eq. (4), the reported QNN advantage could be an artifact of the surrogate objective. Please add a Monte-Carlo comparison of Eq. (6) versus Eq. (4) over the tested parameter grid, or explicitly restrict the claim to the approximate sum-rate and justify the restriction from the cited source [8].","section":"Section II-B, Eq. (6), and Section IV"},{"comment":"No complete experimental configuration is reported: the number of qubits nq is only given in figure captions, and the number of variational layers N, learning rate, optimizer, batch size, number of epochs E, exact CNN architecture, epsilon schedule, and number of random seeds are all absent. The figures show single curves with no error bars or confidence intervals, so the performance gap in Fig. 6 cannot be distinguished from run-to-run randomness. Please provide a full hyperparameter table, report mean and standard deviation over at least several seeds, and include statistical significance tests for the QNN-versus-CNN comparison.","section":"Section IV, Figs. 3-6"},{"comment":"The policy gradient update is not precisely defined. Line 8 defines π = σ(logits), line 9 samples p ∼ Bernoulli(π) with ε-greedy exploration, and line 12 uses log π, but it is unclear whether the loss uses the log-probability of the sampled action p under the Bernoulli policy or something else. In addition, line 14 contains 'L ← L + L', which appears to be a typo for accumulating the loss, and the baseline b is updated with mean(r) while the loss uses (r - b) at the batch-element level; the variance-reduction property of the baseline depends on this dimension mismatch being resolved. Please write out the exact per-sample loss and gradient computation.","section":"Algorithm 1, lines 8-14"},{"comment":"The problem formulation includes proportional fairness through the update of R̄_l(t) in Eq. (7), but Algorithm 1 describes the reward only as 'sumrate reward(Gin, p)' and shows no update of the historical average rate R̄. If fairness is included in the reward, the algorithm should specify how; if it is not, the text overstates the objective. This matters because the fairness term changes the optimal policy and therefore affects the interpretation of the QNN-versus-CNN comparison.","section":"Section III-B, Eq. (7), and Algorithm 1"}],"minor_comments":[{"comment":"There are typographical issues: 'Convolutional Neural Networkss' in the abstract and 'Tertiary' in the related-work paragraph should be corrected.","section":"Abstract and Section I"},{"comment":"The figures do not label the axes or show units, and Fig. 3 has no axis labels. The text also refers to '20 training iterations' while Algorithm 1 uses E epochs; please clarify the correspondence.","section":"Figures 3-6"},{"comment":"The selection of x_l, the index of the maximum diagonal element of U_l used to choose the beamforming direction, is not defined precisely. Please state how x_i is selected for each user, since it directly enters the approximate rate expression in Eq. (6).","section":"Section II-A, Eqs. (3) and (6)"},{"comment":"The caption for Fig. 5 states 'qubits=8' while Fig. 6 states 'qubits=10'; if this is intentional, give the reason for using a different number of qubits in the comparison; otherwise the inconsistency should be corrected.","section":"Section IV, Figs. 5-6"},{"comment":"The conclusion says the model 'maximises ergodic sum-rate', but the actual objective is the proportional-fair approximate sum-rate of Eq. (7) with an approximation from Eq. (6). Please rephrase to match what is optimized and evaluated.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"The paper reads more like a workshop-style proof-of-concept than a complete journal article. The main concerns are experimental reproducibility and the lack of validation of the surrogate reward, both of which are fixable with additional experiments and clearer writing. I would also ask the editor to weigh the novelty against [5]-[8], since the problem formulation and the approximate sum-rate expression are taken from prior work and the new contribution is mainly the hybrid QNN architecture and its RL training. A careful revision that supplies the missing simulation details and a direct validation of Eq. (6) would make the central claim testable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a legitimate new application—hybrid QNN for user scheduling in massive MIMO with statistical CSI—but the evaluation as written doesn't back the headline claim. The paper is worth engaging, not because the results are convincing, but because the question is sensible and the approach is concrete.\n\nWhat's new: prior work used deep RL for scheduling or QNNs for beamforming; this is the first combination of statistical CSI with a variational quantum circuit for scheduling, trained with a policy-gradient RL loop. The architecture is standard (angle encoding, RY rotations, CNOT ring, Pauli-Z measurements, classical pre/post layers), but that's not a flaw—it's a reasonable starting point. The system model and problem formulation are clear, and Algorithm 1 is detailed enough to be reimplemented in principle.\n\nWhat undermines the central claim: the comparison in Fig. 6 measures sum-rate using Eq. (6), the same approximate expression used as the training reward. That means the QNN-vs-CNN gap could reflect which model better optimizes the surrogate, not necessarily the true ergodic sum-rate. The paper gives no Monte-Carlo validation of Eq. (6) in the tested Rician scenarios, so the stress-test concern is real. Additionally, there are no hyperparameters, no number of seeds, no error bars, and no comparison to any prior classical scheduler—only to a self-designed CNN. That's a weak evidential base for \"outperforms classical schedulers.\" There are also small but telling errors in Algorithm 1: the loss line uses E as if it were an expectation, and the update line \"L ← L + L\" looks like a typo. These suggest the manuscript wasn't carefully proofread.\n\nWhat's good: the authors acknowledge the intractability of the true ergodic rate and are explicit about using an approximation. The RL formulation with proportional fairness is sensible. The idea of using a quantum kernel for a combinatorial selection problem is not crazy—discrete scheduling over K users with L slots has a large search space, and quantum circuits can represent correlations that may help.\n\nBottom line: the paper is a preliminary idea, not a demonstrated result. With code, seeds, error bars, a validated reward proxy, and a comparison to an existing scheduler (e.g., the deep learning approach in [7]), it could be a solid contribution. As is, it deserves a serious referee—the topic is relevant and the approach is concrete—but the reviewer should push for substantial revision.\n\nRecommendation: send to peer review with a strong request for reproducibility details, and consider desk-rejecting only if the authors can't provide code/data upon request.","headline":"A plausible but under-supported first application of hybrid QNNs to mMIMO scheduling; the idea deserves scrutiny, but the evaluation needs far more rigor before the outperformance claim can be believed.","tokens_in":7724,"tokens_out":2440,"would_cite":false,"duration_ms":28566,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A hybrid quantum-classical neural network, trained by reinforcement learning on statistical channel information, can schedule users in massive MIMO downlinks with higher sum rate than a CNN benchmark, the paper claims.","keywords":["quantum neural network","massive MIMO","user scheduling","statistical CSI","variational quantum circuit","reinforcement learning","hybrid quantum-classical","sum-rate optimization"],"falsifier":"Evaluate the trained hybrid-QNN and CNN policies under the full Monte-Carlo ergodic sum-rate (4) on the same Rician channel realizations across SNR; if the QNN's sum-rate advantage over the CNN shrinks below the noise floor or reverses, the central claim that the quantum kernel improves scheduling is refuted.","tokens_in":6818,"feed_emoji":"📡","tokens_out":4619,"duration_ms":53556,"temperature":0.7,"pith_summary":"This paper claims that user scheduling in massive MIMO downlinks can be done by a hybrid quantum-classical neural network that reads only statistical channel information rather than full instantaneous CSI. The model is a three-layer architecture: a classical network compresses the beam-domain channel-gain matrix, a variational quantum circuit with angle encoding and entanglement learns a policy representation, and a classical layer maps measurement outcomes to a top-L user selection. Training uses a policy-gradient reinforcement learning algorithm whose reward is a proportional-fairness-weighted approximate sum rate. In simulated correlated Rician fading channels, the hybrid model is reported to converge faster and to achieve higher sum rates than a similarly sized CNN benchmark, with the gap growing in moderate-to-high SNR and denser user configurations. If these results hold, quantum-enhanced machine learning becomes a practical candidate for scheduling decisions in 5G/B5G systems where full CSI acquisition is too expensive.","feed_headline":"Quantum-classical net edges out CNN for massive MIMO scheduling","feed_subtitle":"A hybrid variational-circuit policy trained by RL raises sum-rate in simulated 5G downlinks, most at high SNR.","key_machinery":"The central object is the hybrid-QNN policy network: a variational quantum circuit (the QNN kernel) wrapped by classical pre- and post-layers. Each input is a K-by-K matrix of beam-domain gains from statistical CSI; the classical pre-layer reduces it to nq real angles, which are encoded by RY gates; the circuit applies N variational layers of RZ-RY-RZ single-qubit rotations interlaced with CNOT entangling gates in a ring; Pauli-Z measurement yields nq real expectations; and a final linear layer maps those to K scheduling scores, with the L largest scores selected. The design's work is to turn a high-dimensional, noisy statistical-CSI matrix into a discrete scheduling decision whose quality is measured by the approximate sum-rate expression (6). The reinforcement-learning loop (epsilon-greedy exploration, cosine-annealed momentum baseline, matrix-based reward, gradient clipping) is what tunes the classical and quantum parameters.","core_discovery":"On the paper's own terms, the central discovery is that a variational quantum circuit can serve as the trainable kernel of a user-scheduling policy in a single-cell multi-user massive MIMO system, and that this policy outperforms a classical CNN benchmark under the same channel realizations. The model takes the statistical-CSI beam-domain gain matrix, flattens and standardizes it, compresses it with a classical layer to a number of qubits, encodes the result via RY rotations, applies stacked variational layers with ZYZ single-qubit decompositions and ring-topology CNOT entanglement, and reads Pauli-Z expectations; a final linear layer produces scheduling scores for a top-L selection. The whole pipeline is trained end-to-end with a reinforcement-learning algorithm that maximizes a proportional-fairness-weighted approximate sum rate. The paper reports faster policy convergence, lower training loss at stable configurations, and higher sum rates than the CNN baseline, especially at SNR above roughly 10 dB and with larger antenna arrays.","pith_inferences":["A testable extension would be to use the same architecture with exact Monte-Carlo ergodic rate reward (4) instead of the approximation (6); if the QNN still beats CNN, the claimed advantage is not an artifact of the surrogate objective.","The comparison's fairness could be checked by scaling the CNN width and depth to match the QNN's parameter count and expressiveness; the paper does not report such an ablation, so part of the observed gap may come from an under-provisioned classical benchmark rather than a quantum effect.","Because the evaluation uses simulated Rician channels and idealized quantum circuit execution, the implicit assumption is that the variational unitary can be implemented at high gate fidelity; a hardware-in-the-loop test at current noisy-device error rates would clarify whether the advantage survives.","If equation (6) is accurate, the same reward can be used to train purely classical policies, and comparing those against the QNN would isolate whether the quantum feature map captures beam-domain interference patterns that classical convolutions miss."],"forward_implications":["If the hybrid QNN's advantage holds on real channels, user scheduling in dense 5G/B5G cells can be run off slowly varying statistical CSI, avoiding the overhead of instantaneous CSI acquisition.","The policy-gradient training with matrix-based reward gives a template for other discrete resource-allocation decisions such as beam selection, power control, and pilot assignment where the objective is a sum of rates.","Since the gap over CNN grows with user and antenna count, quantum kernels become more attractive exactly in the interference-rich regimes where classical scheduling is hardest.","The reported robustness in noisy channels suggests the variational circuit can tolerate imperfect statistical CSI, easing deployment.","The top-L selection from learned logits is a generic combinatorial-selection head that can be re-trained for different L and K without changing the circuit size."],"supporting_citations":[{"why":"Supplies the beam-domain system model, the ergodic sum-rate optimization problem, and the scheduling-indicator formulation that the paper builds on.","marker":"[7]"},{"why":"Provides the approximate ergodic sum-rate expression (6) used as the RL reward and performance metric, along with the correlated Rician channel model.","marker":"[8]"},{"why":"Justifies the DFT-based beamforming vectors through the diagonalized channel correlation matrix U_l.","marker":"[9]"},{"why":"Adds the proportional fairness weighting that turns the pure sum-rate objective into the weighted objective (7).","marker":"[10]"},{"why":"Defines quantum angle encoding and the Pauli-Z measurement readout used in the QNN kernel.","marker":"[11]"},{"why":"Supplies the parametrized variational circuit formalism, including the layered unitary U(θ).","marker":"[12]"},{"why":"Supports the ZYZ decomposition as a universal single-qubit parametrization in the variational layers.","marker":"[13]"},{"why":"Motivates the ring-topology CNOT entanglement for correlating qubits in the variational circuit.","marker":"[14]"}],"fun_headline_variants":["Hybrid quantum net beats CNN for MIMO scheduling","Quantum kernel speeds user scheduling in massive MIMO","Variational circuit scores better in MIMO scheduling","Quantum-classical model ups sum-rate in 5G scheduling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the approximate ergodic sum-rate formula in equation (6), taken from earlier work, is accurate enough to serve both as the RL reward and as the performance metric; if it diverges from the true channel behavior in the tested Rician scenarios, the trained policy is optimizing a surrogate objective and the reported advantage over the CNN may not transfer to the real sum rate.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid quantum net beats CNN for MIMO scheduling","Quantum kernel speeds user scheduling in massive MIMO","Variational circuit scores better in MIMO scheduling","Quantum-classical model ups sum-rate in 5G scheduling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000141,"raw_usage":{"total_tokens":1104,"prompt_tokens":827,"completion_tokens":277,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":443,"completion_tokens_details":{"reasoning_tokens":214}},"tokens_in":443,"tokens_out":277,"duration_ms":3733,"temperature":1.0,"reasoning_tokens":214,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T04:29:47.721040+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the trained hybrid-QNN and CNN policies under the full Monte-Carlo ergodic sum-rate (4) on the same Rician channel realizations across SNR; if the QNN's sum-rate advantage over the CNN shrinks below the noise floor or reverses, the central claim that the quantum kernel improves scheduling is refuted.","supporting_citations":[{"cited_title":"Deep learning based user schedul- ing for massive mimo downlink system,","cited_arxiv_id":null,"evidence_quote":"Supplies the beam-domain system model, the ergodic sum-rate optimization problem, and the scheduling-indicator formulation that the paper builds on."},{"cited_title":"Joint scheduling and deep learning-based beamforming for fd-mimo systems over correlated rician fading,","cited_arxiv_id":null,"evidence_quote":"Provides the approximate ergodic sum-rate expression (6) used as the RL reward and performance metric, along with the correlated Rician channel model."},{"cited_title":"Beam di- vision multiple access transmission for massive mimo communications,","cited_arxiv_id":null,"evidence_quote":"Justifies the DFT-based beamforming vectors through the diagonalized channel correlation matrix U_l."},{"cited_title":"Proportional fair space-time scheduling for wireless com- munications,","cited_arxiv_id":null,"evidence_quote":"Adds the proportional fairness weighting that turns the pure sum-rate objective into the weighted objective (7)."},{"cited_title":"Quantum angle encoding with learnable rotation applied to quantum–classical convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"Defines quantum angle encoding and the Pauli-Z measurement readout used in the QNN kernel."},{"cited_title":"Efficient decomposition of unitary matrices in quantum circuit compilers,","cited_arxiv_id":null,"evidence_quote":"Supports the ZYZ decomposition as a universal single-qubit parametrization in the variational layers."},{"cited_title":"Quantum entanglement in neural network states,","cited_arxiv_id":null,"evidence_quote":"Motivates the ring-topology CNOT entanglement for correlating qubits in the variational circuit."}],"review_version":1}