{"id":"785dd020-9911-4995-a41d-3dbd04ccd72f","arxiv_id":"2507.12492","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"SpoQFL applies sporadic learning to quantum federated learning by suppressing noisy client updates, claiming higher accuracy and more stable convergence in noisy-device simulations.","lead":"Quantum federated learning trains shared models across noisy quantum devices, and this paper adds a sporadic-learning layer that shrinks or skips client updates when noise is high. In simulations on CIFAR-10 and CIFAR-100, the method reports accuracy gains over a prior QFL baseline, but the mechanism relies on knowing the noise that real devices cannot directly observe.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central mechanism uses the true noise deviation ξ in Eq. (9), which clients cannot observe; no estimator is supplied, so the reported gains may depend on simulator-only knowledge.","rationale":"The reader's weakest-assumption match is exactly the load-bearing issue I identify: the algorithm as written requires each client to know |ξ|, the deviation of the noisy gradient from the noiseless gradient, but no client can observe this quantity on hardware and no estimation procedure is provided. If the simulation computes ξ by comparing against the known noiseless gradient, then the experimental comparison is not a test of a deployable algorithm; if instead ξ is estimated some other way, the paper does not say how, so the experiments are not reproducible from the text. The convergence argument in §3.4 is also only asserted: scaling by exp(−γ|ξ|) introduces a bias that depends on the noise distribution, and the claim that the variance term V in Eq. (8) is reduced requires a derivation, not a sentence. Secondary issues—single runs with no error bars, no released code, and test-set-based hyperparameter selection—further weaken the empirical support, but the ξ-access problem alone is sufficient to keep the verdict at rejection. I therefore recommend no change to the reader's verdict.","tokens_in":12947,"tokens_out":5098,"duration_ms":74732,"concrete_test":"Re-run Table 7's SpoQFL row after replacing Algorithm 1 line 9 with a client-accessible estimate of ξ, such as ξ̂ = ĝ_t − ĝ_{t−1} (the difference between two successive noisy gradient estimates) or the empirical residual from repeated parameter-shift evaluations; keep all other hyperparameters fixed. If the 4.87%/3.66% margins over wpQFL disappear or shrink below run-to-run variation, the published result relies on oracle knowledge of ξ rather than on an implementable noise-mitigation rule. As a minimal first step, the authors should state exactly how ξ was computed in the current torchquantum experiments.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Table 7, up to 4.87% and 3.66% over wpQFL) rests on Algorithm 1 lines 9–13 and Eq. (9), where the sporadic variable x = exp(−γ|ξ|) and the skip rule x < τ are functions of ξ_t^{n,k}, the noise-induced deviation in the gradient estimate (Eq. (6)). On a real NISQ client, the client observes only the noisy gradient ĝ; the noise-free gradient ∇f is the target being learned, so |ξ| = |ĝ − ∇f| is not available. The manuscript gives no estimation protocol for ξ, no repeated-measurement variance estimator, no zero-noise extrapolation procedure, and no sensitivity analysis to mis-specified ξ. In the torchquantum simulations, where Pauli noise is injected at a known level, the authors can compute ξ exactly by subtracting the simulator's noiseless gradient; if that is what Algorithm 1 line 9 does, Table 7 is an oracle experiment rather than an implementable algorithm. The convergence claim in §3.4 ('sporadic learning directly reduces the noise variance term V') is also asserted without proof: multiplying the noisy gradient by a factor that depends on ξ changes the mean as well as the variance, so it is not automatic that the V term in Eq. (8) decreases; a bias-variance trade-off analysis is missing. Because the practical mechanism and its theoretical justification both depend on this inaccessible quantity, the experimental superiority claim is not yet supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SpoQFL, a sporadic federated learning scheme for quantum neural network classifiers. In each local step, the noisy gradient estimate is decomposed as the noise-free gradient plus a noise-induced deviation xi, and the parameter update is scaled by exp(-gamma*|xi|) and skipped entirely when this factor falls below a threshold tau. The authors motivate the method by an appeal to a standard stochastic-convergence bound and evaluate it in torchquantum simulations on CIFAR-10 and CIFAR-100 with ten clients and heterogeneous encodings, reporting accuracy gains of up to 4.87% over wpQFL on CIFAR-10 and 3.66% on CIFAR-100, with corresponding loss reductions. The paper also contains ablations over qubit counts, layer counts, loss functions, learning rates, client counts, and noise levels.","tokens_in":13246,"tokens_out":3838,"duration_ms":48611,"significance":"The high-level idea of suppressing locally noisy quantum updates before federated aggregation is reasonable and could be valuable for heterogeneous NISQ devices. The paper is clearly organized, provides pseudocode for the algorithm, and includes broad ablations and comparisons against several classical and quantum baselines. However, the central mechanism depends on the noise-induced gradient deviation xi, which a client cannot observe on real hardware and for which no estimator is supplied; the convergence justification in Section 3.4 is asserted rather than proved; and the experimental section lacks error bars, seed counts, and key hyperparameter values. If the mechanism were implementable and the gains reproducible, this would be a useful contribution, but as written the central claim is not established.","major_comments":[{"comment":"The sporadic variable in Eq. (9) and the skip rule in Algorithm 1 (lines 9-11) are functions of xi_t^{n,k}, the difference between the noisy gradient estimate and the noise-free gradient defined in Eq. (6). A client on a real device observes only the noisy gradient estimate; the noise-free gradient is the quantity being learned. The manuscript provides no estimator for |xi|, no repeated-measurement variance estimator, no zero-noise extrapolation procedure, and no sensitivity analysis to mis-specified xi. In the torchquantum simulations, the authors can compute the true noiseless gradient and subtract it, so Table 7 appears to rely on simulator-only knowledge. Because this inaccessible quantity drives both the scaling and the skipping mechanism, the algorithmic proposal is not implementable as stated.","section":"Algorithm 1, Eq. (9)"},{"comment":"The claim that sporadic learning directly reduces the noise variance term V in Eq. (8) is not proved. Multiplying the noisy gradient by x_t^{n,k} = exp(-gamma*|xi|) changes the expectation of the update as well as its variance, so a bias-variance decomposition is required to show that the V term decreases and that the induced bias is controlled. Eq. (8) is a standard bound for unmodified stochastic gradients, and no argument is given that the modified update satisfies the same contraction with a smaller V. The notation in Eq. (8) also conflates the noise deviation xi with the loss suboptimality E[L(omega^T)] - L*, which further obscures the argument.","section":"Section 3.4, Eqs. (8)-(10)"},{"comment":"The layer-count ablation in Table 2 shows that l=3 is best for CIFAR-10 (89.36%) and l=1 is best for CIFAR-100 (55.63%), but the text states that l=1 is selected for subsequent evaluations without qualification. Since Table 7 reports a single configuration for both datasets, the choice contradicts the reported ablation and leaves the experimental setup ambiguous. In addition, the headline comparison in Table 7 is presented without error bars, without the number of random seeds, and without the values of gamma and tau; hyperparameters are selected on test accuracy, which risks overfitting the evaluation protocol.","section":"Section 4.3, Table 2 and Section 4.5, Table 7"},{"comment":"The noise level used for the Table 7 comparison is not stated. Figure 3 uses epsilon=0.001, while Table 6 spans epsilon from 0.001 to 0.5, and the main comparison table gives no indication of which noise regime is used or whether all baselines are evaluated under identical noise conditions. Without this information, the reported improvements over wpQFL and other baselines cannot be reproduced or interpreted.","section":"Sections 4.4-4.5, Tables 6-7"}],"minor_comments":[{"comment":"There is a typo 'wshere' instead of 'where', and the symbols N_h, D, H, and M are introduced without a precise definition of the Hermitian observable decomposition.","section":"Eq. (7)"},{"comment":"Panel (d) is captioned 'CIFAR-10 Non-IID accuracy', but it is placed after the CIFAR-100 IID panel and the text describes it as part of the CIFAR-100 comparison; the caption should be corrected.","section":"Figure 2"},{"comment":"The text says layer configurations 1, 2, 3, 5, and 10 were tested, but the table also contains a row for l=4; the list and table should be reconciled.","section":"Section 4.3, Table 2"},{"comment":"The notation switches between Dq and Dc for the number of qubits; these should be unified.","section":"Section 4.3"},{"comment":"The axis labels of the convergence plots are garbled in the typeset version, with values such as '0 25 50 75 10 0' appearing on the vertical axis; the figures need to be regenerated.","section":"Figure 3"},{"comment":"The abstract refers to 'real-world datasets', but the experiments are simulations using torchquantum with synthetic noise; the wording should be adjusted to avoid overclaiming.","section":"Abstract and Section 4.1"},{"comment":"The connection to the cited sporadic federated learning work [37] is mentioned only briefly; a short comparison of what is new in the quantum setting relative to [37] would help the reader.","section":"Related work, Section 2.2"}],"recommendation":"reject","confidential_remarks":"The manuscript does not provide code or detailed experimental configuration, and the central algorithmic quantity xi is unobservable without an estimation protocol. In my view this is a load-bearing correctness issue rather than a presentation issue: the reported superiority in Table 7 could be an artifact of the simulator giving the algorithm access to the true noise-free gradient. I therefore recommend rejection, although a substantially revised version that proposes and validates an estimator for the noise deviation, supplies a bias-variance analysis, and reports seeds and hyperparameters could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper transfers sporadic FL from classical to quantum federated learning and adds a noise-proportional scaling/skip rule. That combination is new in the QFL literature, and the idea is sensible: suppress updates from clients whose gradients are heavily corrupted by noise. The writing is clear enough, the related work is adequate, and the experiments cover a reasonable range of settings (datasets, IID/non-IID, noise levels, number of clients). The gain over wpQFL is modest but consistently positive in the reported tables, so there is something here worth examining.\n\nThe soft spots are real and load-bearing. The method's sporadic variable x = exp(-gamma|xi|) and the skip rule depend on |xi|, the noise-induced deviation in the gradient. On a real device a client only sees the noisy gradient; the noise-free gradient is the target being learned. The paper never says how xi is estimated. In the torchquantum simulation, the authors can compute xi exactly because they know the injected noise and can compute the noiseless gradient. If that is what Algorithm 1 line 9 does, Table 7 is an oracle experiment, not an implementable algorithm. The convergence claim is also asserted: 'sporadic learning directly reduces the noise variance term V' in Eq (8) is not proven. Scaling the gradient by a factor that depends on xi changes the bias as well as the variance, and the bound in Eq (8) does not obviously apply. There are no error bars, no seeds, no code, and hyperparameters (gamma, tau, learning rate) are selected on test accuracy, so the reported numbers may overstate the effect.\n\nThat said, the paper is not incoherent. The mechanism is a natural heuristic, and the cited classical sporadic FL work gives it pedigree. The central flaw is the missing estimator for xi and the missing bias-variance analysis. Both are fixable: the authors could propose a practical estimator (e.g., repeated measurements or noise learning) or reframe the method as using an observable proxy for noise, and they could provide a proper convergence analysis. The experiments would be much more convincing with multiple seeds and error bars.\n\nWho is this for? Researchers working on QFL noise mitigation would want to know about this direction, but they should not treat the current numbers as evidence for a deployable method. I would send it to peer review because the idea is worth developing and the paper is honestly written, but I would expect major revision before acceptance.","headline":"SpoQFL is a plausible heuristic for noise-aware QFL, but the central algorithm needs access to an unobservable quantity and the evidence is single-run simulation; worth a referee but not acceptance as-is.","tokens_in":13772,"tokens_out":627,"would_cite":false,"duration_ms":9404,"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":"Scaling updates by noise intensity boosts quantum federated learning","keywords":["sporadic federated learning","quantum federated learning","quantum noise mitigation","noise heterogeneity","quantum neural networks","NISQ","convergence stability","CIFAR classification"],"falsifier":"A benchmark that removes the oracle access to the true gradient—for example, running SpoQFL on a real NISQ device or in a simulator where clients must estimate $|\\xi|$ from shot statistics—and shows the accuracy advantage over QFL vanishes or reverses would falsify the paper's central claim.","tokens_in":12724,"feed_emoji":"⚛️","tokens_out":6359,"duration_ms":60771,"temperature":0.7,"pith_summary":"SpoQFL is a federated learning method for quantum devices that targets a specific obstacle: quantum noise levels differ across devices, and that heterogeneity destabilizes the shared model. The paper's central claim is that if each client shrinks its gradient update by a factor that depends on the local noise deviation, and skips the update when that deviation is large, the global model becomes more accurate and converges more smoothly. Experiments on CIFAR-10 and CIFAR-100 show accuracy gains over the strongest prior QFL baseline by up to 4.87% and 3.66%, and a reduction in loss by 16.84% and 4.15%. If correct, this gives QFL a simple client-side rule for robustness to heterogeneous hardware noise that requires no change to the server-side aggregation.","feed_headline":"Scaling updates by noise intensity boosts quantum federated learning","feed_subtitle":"SpoQFL beats the best QFL baseline by up to 4.87% accuracy on CIFAR-10.","key_machinery":"The central object is the sporadic variable $x_{n,k}^t = \\exp(-\\gamma |\\xi_{n,k}^t|)$, defined in Eq. (9) from the norm of the noise-induced gradient deviation. Each client multiplies its gradient estimate by this factor before the local update (Eq. (10)), and the update is skipped entirely when the factor falls below the threshold $\\tau$ (Algorithm 1, lines 10-11). The paper ties this mechanism to the convergence bound in Eq. (8) by arguing that the scaling shrinks the noise variance term $V$, which accelerates convergence and stabilizes training across clients with heterogeneous noise.","core_discovery":"The paper claims that SpoQFL, a sporadic learning strategy, significantly outperforms conventional QFL approaches under simulated Pauli noise, achieving superior training performance and more stable convergence. On the reported benchmarks, SpoQFL reaches 91.92% accuracy on CIFAR-10 and 57.60% on CIFAR-100, up to 4.87% and 3.66% above the best prior quantum federated method (wpQFL). The mechanism is the sporadic variable $x = \\exp(-\\gamma |\\xi|)$, which scales the noisy gradient so that noisy clients contribute less to the global model, and a threshold $\\tau$ that skips the most unstable updates. The paper also argues this reduces the noise variance term in the existing convergence bound, improving both stability and final performance.","pith_inferences":["The method as presented assumes clients can measure the noise-induced deviation $|\\xi|$; a practical implementation would need a proxy for this quantity, such as the variance of measurement shots or a separate calibration round, and testing with such a proxy is a direct extension the paper does not report.","The main comparison is run at a single noise level (epsilon = 0.001); a natural extension is to sweep the noise level and check whether the relative advantage of SpoQFL grows or shrinks as noise increases.","The same scaling idea could be applied server-side by weighting clients according to estimated noise, which would relax the client-side assumption that the true gradient deviation is known."],"forward_implications":["SpoQFL needs no change to the server aggregation equation, so it can be dropped into existing QFL frameworks that use FedAvg-style averaging.","The reported accuracy gains (up to 4.87% on CIFAR-10 and 3.66% on CIFAR-100) suggest that sporadic attenuation is a competitive alternative to personalized QFL baselines like wpQFL.","Because the method reduces the noise variance term in the convergence bound, it should also help in other noisy-gradient settings, such as classical FL with heterogeneous communication noise.","The scaling parameter $\\gamma$ and threshold $\\tau$ give practitioners two tunable knobs for trading off update informativeness against noise suppression."],"supporting_citations":[{"why":"Supplies the basic QFL framework that SpoQFL extends and compares against.","marker":"[5]"},{"why":"The strongest quantum baseline (wpQFL) whose accuracy and loss SpoQFL is benchmarked against.","marker":"[9]"},{"why":"Provides the FedAvg aggregation rule used in Eq. (4) and the classical FL baseline.","marker":"[19]"},{"why":"The classical sporadic federated learning methodology that this work adapts to quantum noise mitigation.","marker":"[37]"},{"why":"Supplies the noisy-gradient formulation and convergence analysis used in Section 3.3.","marker":"[13]"},{"why":"Prior work on noise-robust QFL that motivates the problem setting.","marker":"[6]"}],"fun_headline_variants":["Sporadic learning adapts to quantum noise, boosting QFL accuracy by 4.87%","Quantum federated learning tamed by noise-aware sporadic updates","SpoQFL: skipping noisy steps makes quantum federated learning stable","Noise-adaptive sporadic training reduces quantum noise impact in QFL"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The core mechanism requires each client to know the size of the noise-induced error in its own gradient, but a real client only has access to the noisy gradient itself, so the scaling and skip rules cannot be computed on actual hardware without some estimate for the noise-free gradient.","fun_headline_variants_meta":{"raw":{"variants":["Sporadic learning adapts to quantum noise, boosting QFL accuracy by 4.87%","Quantum federated learning tamed by noise-aware sporadic updates","SpoQFL: skipping noisy steps makes quantum federated learning stable","Noise-adaptive sporadic training reduces quantum noise impact in QFL"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000613,"raw_usage":{"total_tokens":2797,"prompt_tokens":840,"completion_tokens":1957,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":456,"completion_tokens_details":{"reasoning_tokens":1888}},"tokens_in":456,"tokens_out":1957,"duration_ms":16042,"temperature":1.0,"reasoning_tokens":1888,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:03:45.954419+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A benchmark that removes the oracle access to the true gradient—for example, running SpoQFL on a real NISQ device or in a simulator where clients must estimate $|\\xi|$ from shot statistics—and shows the accuracy advantage over QFL vanishes or reverses would falsify the paper's central claim.","supporting_citations":[{"cited_title":"Quantum federated learn- ing with quantum data","cited_arxiv_id":null,"evidence_quote":"Supplies the basic QFL framework that SpoQFL extends and compares against."},{"cited_title":"A personalized quantum federated learning","cited_arxiv_id":null,"evidence_quote":"The strongest quantum baseline (wpQFL) whose accuracy and loss SpoQFL is benchmarked against."},{"cited_title":"Communication- efficient learning of deep networks from decentralized data","cited_arxiv_id":null,"evidence_quote":"Provides the FedAvg aggregation rule used in Eq. (4) and the classical FL baseline."},{"cited_title":"Decentralized sporadic federated learning: A unified methodology with generalized convergence guarantees","cited_arxiv_id":null,"evidence_quote":"The classical sporadic federated learning methodology that this work adapts to quantum noise mitigation."},{"cited_title":"Error-mitigation- aided optimization of parameterized quantum circuits: Con- vergence analysis","cited_arxiv_id":null,"evidence_quote":"Supplies the noisy-gradient formulation and convergence analysis used in Section 3.3."},{"cited_title":"Robust quan- tum federated learning with noise","cited_arxiv_id":null,"evidence_quote":"Prior work on noise-robust QFL that motivates the problem setting."}],"review_version":1}