{"id":"65cae20a-67e0-4fd8-b151-16709acd4f59","arxiv_id":"2506.02422","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A quantization-assisted Gaussian differential privacy mechanism and a min-max fair scheduling algorithm for wireless personalized federated learning, backed by convergence bounds and simulations showing up to 87% accuracy improvement.","lead":"The paper proposes a wireless personalized federated learning method that uses quantization and Gaussian noise to protect privacy while scheduling clients to improve convergence and fairness. Simulations show large gains in accuracy, worst-client test loss, and fairness over existing scheduling alternatives.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed optimality of Algorithm 2 rests on an unproved min-max/max-min conversion and an unverified convexity assertion in Theorem 5; without either, the 'optimal scheduling' claim is not established.","rationale":"The reader's verdict is CONDITIONAL, and the rationale already notes that Theorem 5 is incomplete and that the min-max to max-min conversion and decoupling are not rigorously justified. I agree with those observations, but the reader's weakest_assumption focuses on Assumption 1 (strong convexity/smoothness) rather than on the optimality chain. In my reading, the optimality chain is at least as load-bearing: even if Assumption 1 held, the algorithm's claim to solve Problem P optimally would still require the min-max/max-min equality, the decoupling under the coupling constraint C1, and the convexity of Problem P7. Each of these is asserted rather than proved. The experiments do provide real evidence that the proposed scheduling policy improves on the baselines, and the empirical gains do not depend on proving global optimality; however, the abstract and Section VI-D make the stronger claim of optimality, and that claim is what my concern targets. The proposed concrete test would settle the issue by checking whether the sequential decoupled solution actually attains the optimum of the original min-max problem in a small but representative instance, and by numerically testing the positivity of the second derivative that Theorem 5 asserts. If the test passes, the theoretical claim is supported; if it fails, the paper should be revised to present the scheduling method as a heuristic with strong empirical performance rather than as an optimal strategy. The reader's CONDITIONAL verdict already accommodates this uncertainty, so no verdict change is needed.","tokens_in":34922,"tokens_out":7271,"duration_ms":81654,"concrete_test":"Implement a one-round numerical audit of Section VI with N=2 clients, the CNN parameters reported in the Section VII footnote (L=0.29, µ=0.05), fixed channel realizations, and the same Γ constants. Grid-search ε_P∈[1−µ²/4,1), η_P,n∈(0,1), λ_n∈(0,2) satisfying C1, with fine resolution, and compare the true minimum of max_n Φ_n (Eq. 32a/34) to the value produced by Algorithm 2's sequential KM+P5+P7 procedure. If the grid optimum is lower, the claimed optimality/decoupling is false. In the same search, compute ∂²Φ_n/∂η² from Appendix F; any negative value on Ω0∪Ω1 directly refutes Theorem 5.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim of an optimal min-max fair scheduling strategy rests on the chain in Section VI: Problem P (32) is converted to the max-min Problem P1, then decoupled into P2/P3, P5, and P7, with P7 solved exactly by convex optimization per Theorem 5. The equality between min_x max_n Φ_n and max_n min_x Φ_n is asserted without proof in Section VI-A; in general max-min ≤ min-max, and while the additive structure of Φ_n in Θ_L helps, constraint C1 (ε_{P,n}=ε_P, a shared variable across clients) reintroduces coupling and the required equality is not immediate. The decoupling statement ('this min-max problem can be converted...') is not derived. Theorem 5's convexity proof is also incomplete: Appendix F states H1–H4 are positive 'by analyzing monotonicity' without giving the analysis; the displayed second derivative is not obviously nonnegative on Ω0∪Ω1. Since Algorithm 2's claimed global optimality and the abstract's 'optimal transmission scheduling strategy' depend on this chain, an unproven or false step would reduce the method to a heuristic that nonetheless performs well empirically. This is the most load-bearing concern because it attacks the theoretical guarantee that distinguishes the paper from a heuristic scheduling policy.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies wireless personalized federated learning (WPFL) over OFDMA links and makes three intertwined claims: (i) a quantization-assisted Gaussian differential privacy mechanism that exploits quantization error as part of the privacy budget; (ii) convergence upper bounds for both the FL global model and each client's PL model under DP noise, quantization, and imperfect channels; and (iii) a min-max fair scheduling and configuration strategy that, by decoupling the min-max problem into client selection, channel allocation, power control, and learning-rate/weight adjustment, is claimed to be optimal. The experimental section reports large improvements over round-robin, random selection, and non-adjustment baselines on MNIST, FMNIST, and CIFAR10 with DNN, MLR, and CNN models, as well as comparisons against pFedMe, APPLE, FedAMP, and FedALA. The paper is well structured and contains substantial derivations in the appendices.","tokens_in":35197,"tokens_out":6995,"duration_ms":67002,"significance":"If the theoretical chain is fully established, the paper would be a valuable integrated contribution: it combines privacy accounting, convergence analysis, and scheduling in a single WPFL framework, and the idea of treating quantization error as a privacy resource is timely and potentially useful. The paper also ships a fairly extensive evaluation against both DP baselines and PFL baselines, including communication-overhead measurements. The main value of the paper, however, rests on the claim that Algorithm 2 provides an optimal min-max fair scheduling strategy; that claim is currently not fully supported by the proofs as written. The convergence and privacy analyses are long and not machine-checked, and the experimental validation does not by itself fill the gap because the scheduling objective is derived from the theoretical bounds.","major_comments":[{"comment":"The conversion of the min-max problem P into the max-min problem P1 is asserted without proof. In general, max over variables of min over clients is at most min over clients of max over variables, and equality requires a saddle-point or minimax-theorem condition that is not established. The clients are coupled not only through the common term Theta_L^t and the shared channel variables c^t, p^t, but also through constraint C1, which forces epsilon_{P,n}^{t+1} = epsilon_P^{t+1} for all n; this coupling is precisely the kind of structure that can prevent a trivial interchange of min and max. The subsequent decoupling into P2 and P4, and the statement in Section VI-D that 'the solution to Problem P1 is optimal', inherit this gap. The authors should either prove the equality under their assumptions or downgrade the optimality claim to a heuristic policy whose performance is validated empirically.","section":"Section VI-A, Eq. (32) and Problem P1"},{"comment":"The convexity proof of Phi_n^{t+1} on Omega_0^{t+1} union Omega_1^{t+1} is incomplete. The displayed second derivative has a denominator a0 * eta^4 * (b0 - mu*eta + eta^2)^3, and the sign of this denominator over the feasible set is not discussed. The four terms H1-H4 are asserted to be positive 'by analyzing the monotonicity', but that analysis is not shown. Since Theorem 5 is the basis for the claim that Problem P7 is solved to global optimality by convex optimization methods (Section VI-D), this is a load-bearing gap. A complete analytical proof, or an explicit statement that parts of the convexity verification are numerical rather than analytical, is required.","section":"Appendix F, Theorem 5"},{"comment":"The convergence theorems and the scheduling objective depend on Assumption 1, which requires every local loss function F_n to be mu-strongly convex and L-smooth with bounded stochastic gradients. However, the experimental evaluation in Section VII uses ReLU-based CNNs with cross-entropy loss, which are neither strongly convex nor L-smooth. The footnote that estimates L and mu empirically provides numerical constants, but it does not establish that Assumption 1 holds for these models. Consequently, Theorems 2-4 and the bound Phi_n^{t+1} in Eq. (30c) may not describe the actual training dynamics in the CNN experiments. The authors should either restrict the formal claims to models satisfying Assumption 1, justify the assumption through a smoothed or regularized surrogate objective, or provide direct evidence, such as tracking the derived bound against the measured E[||tilde{varpi}_n^t - varpi_n^*||^2], that the analysis remains predictive for CNNs.","section":"Section V, Assumption 1, and Section VII"}],"minor_comments":[{"comment":"The definition of E_max^G uses Delta_L/2; it should be Delta_G/2 = beta_G * C.","section":"Eq. (7)"},{"comment":"The quantifier order in P1 is written as max_{n in N} min_{...}, which is not the intended max-min reformulation; the maximization should be over the optimization variables and the minimization over clients. Please correct the notation and define the strategy sets explicitly.","section":"Section VI-A, Problem P1"},{"comment":"The legend entry 'Error-free (t_n, U = 0)' is undefined; it presumably means rho_{n,L}^t = 0 and rho_{n,G}^{t+1} = 0.","section":"Section VII, Fig. 3"},{"comment":"Reference [18] contains the typo 'inn Proc.', and several sentences in Related Work are ungrammatical, e.g., 'None has taken fairness among the PL models of different participants.'","section":"References and Related Work"},{"comment":"The subsequent '-DP' expression with log(psi/psi_1) and log(psi'/psi'_1) is undefined if psi_1 or psi'_1 vanishes; the authors should specify conditions on sigma_DP, C, and R that exclude this case.","section":"Theorem 1"},{"comment":"Algorithm 2 sets epsilon_P^{t+1} = epsilon_P, but Theorem 5's convexity guarantee holds only for epsilon_P in specific intervals; the algorithm should either check or enforce that epsilon_P lies in the required range.","section":"Algorithm 2, line 8"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely salvageable: the empirical results are promising, and the overall design is coherent. However, the two theoretical pillars of the optimality claim, namely the min-max/max-min equivalence in Section VI-A and the convexity proof in Appendix F, are currently too incomplete to support the abstract's 'optimal transmission scheduling strategy'. If the authors can repair these proofs, or alternatively weaken the claims to a heuristic with strong empirical validation, the paper could be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the quantization-assisted Gaussian DP accounting and the min-max scheduling design form a plausible and useful package for wireless PFL, but the paper's 'optimal scheduling' claim is not established as written. The privacy mechanism is the strongest piece: using deterministic uniform quantization's error to reduce the Gaussian noise needed is a real idea, and the composition analysis in Theorem 1 looks like a legitimate contribution. The convergence bounds are standard but carefully assembled to include quantization, DP, and channel errors, and the scheduling chain (KM for selection/allocation, then convex optimization for learning rates and weights) is a sensible design flow. The related work is well placed; the distinction from stochastic-quantization DP works [28]-[30] is accurate.\n\nThe soft spots are genuine. The min-max to max-min conversion in Section VI-A is asserted, not proved; in general max-min <= min-max, and the shared eps_P constraint across clients reintroduces coupling, so equality is not obvious. Theorem 5's convexity proof in Appendix F is a sketch that says H1-H4 are positive 'by analyzing monotonicity' without showing the analysis, and the second derivative is stated without derivation. Since Algorithm 2's global optimality rests on this chain, the theoretical guarantee is currently not supportable. Assumption 1's strong convexity and L-smoothness obviously do not hold for DNNs/CNNs with ReLU, and the authors estimate L and mu empirically, which makes the bounds design heuristics rather than hard guarantees, common in this literature but worth stating plainly. The lack of error bars is a real, fixable oversight given the size of the reported gains. To the paper's credit, the bound is not fitted to experimental curves; it is derived and then tested, which is a legitimate flow.\n\nMy take: the paper deserves a serious referee, with the expectation of heavy revision. The referee should either demand a complete proof of the min-max/max-min equality and a full Appendix F, or insist that the optimality language be softened to 'proposed scheduling strategy with empirical gains'. Also ask for error bars and a discussion of the strong-convexity assumption's scope. This is a paper for people in wireless personalized FL and communication-efficient privacy; they will get value from the quantization-assisted DP mechanism and the fairness-aware scheduling design. I would accept it for peer review, not desk-reject.","headline":"Quantization-assisted DP for wireless PFL is a real idea, but the optimal scheduling guarantee is not proved as written.","tokens_in":35746,"tokens_out":3747,"would_cite":false,"duration_ms":35384,"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":"Wireless personalized federated learning can treat the rounding errors of low-bit quantization as part of its differential-privacy budget, and a min-max scheduler built on the resulting convergence bounds improves accuracy, worst-client…","keywords":["personalized federated learning","wireless federated learning","differential privacy","quantization","min-max fairness","scheduling","convergence analysis","OFDMA"],"falsifier":"Audit the quantization-assisted Gaussian mechanism by computing the empirical maximum divergence between outputs produced from two datasets that differ in one sample, using the paper's stated clipping threshold $C$, quantization bits $R$, and noise scale $\\sigma_{DP}$; if the measured privacy loss exceeds the claimed $(\\epsilon_Q,\\delta_Q)$ budget, the central claim that quantization errors can be counted toward the differential-privacy budget is wrong.","tokens_in":34701,"feed_emoji":"🔒","tokens_out":18768,"duration_ms":154102,"temperature":0.7,"pith_summary":"The paper tries to establish that in wireless personalized federated learning (WPFL) over OFDMA subchannels, the rounding errors of low-bit quantization can be counted as a privacy resource rather than a nuisance: by adding Gaussian noise and then quantizing, a given differential-privacy budget can be met with less injected noise than noise-only schemes. The paper also derives per-client convergence upper bounds that fold in quantization error, DP noise, and imperfect wireless channels, and it uses the worst of those bounds as a scheduling objective. Minimizing that worst bound—through client selection, subchannel allocation, power control, and adaptive choice of learning rates and the personalization weight—decouples into a polynomial-time matching problem plus a convex per-client optimization, which is what makes a min-max fair scheduler practical. Reported experiments on MNIST, FMNIST, and CIFAR-10 show the scheduler beating round-robin, random, and fixed-configuration alternatives by 87.08% in accuracy, 16.21% in worst-client test loss, and 38.37% in a standard fairness index under a CNN model. A reader should care because privacy, communication efficiency, and fairness usually pull against each other in federated learning; the claim is that quantization error, which is already present for communication efficiency, can be made to help with privacy and fairness at the same time.","feed_headline":"Quantization errors strengthen wireless federated learning privacy","feed_subtitle":"A min-max scheduler built on the idea raises accuracy by 87% and fairness by 38% in wireless federated learning tests.","key_machinery":"The load-bearing object is the quantization-assisted Gaussian mechanism $M_Q(u_n,D_n)=Q(u_n(D_n)+z_n)$, whose output-distribution analysis (Theorem 1) turns the quantization step size—equivalently the maximum quantization error $E^L_{\\max}$—into a term of the differential-privacy budget, so coarser quantization buys a smaller required $\\sigma_{DP}$. The second load-bearing object is the per-round personalized-model convergence bound $\\Phi_n^{t+1}$ from Theorem 3, which separates the shared aggregation error $\\Theta_L^t$ from client-specific terms in the federated-learning rate $\\eta_{F,n}^t$, the personalized-learning rate $\\eta_{P,n}^{t+1}$, and the FL-PL weighting coefficient $\\lambda_n^{t+1}$. That separation creates the nested structure the scheduler exploits: minimizing $\\Theta_L^t$ over client selection, subchannel allocation, and power control is a maximum-weight bipartite matching, while minimizing each $\\Phi_n^{t+1}$ becomes a convex program in the personalized-learning rate over the feasible regions $\\Omega_0^{t+1}\\cup\\Omega_1^{t+1}$ when $\\mu<2$ and the common convergence rate $\\epsilon_P^{t+1}$ is chosen in $[1-\\mu^2/4,1)$. Together these pieces convert the min-max fairness objective into a sequentially decoupled policy with polynomial overall complexity.","core_discovery":"The paper's central claim is that quantization error is not merely tolerated but actively useful: the quantization-assisted Gaussian mechanism $M_Q(u,D)=Q(u(D)+z)$ satisfies $(\\epsilon_Q,\\delta_Q)$-differential privacy with a budget that explicitly depends on the maximum quantization error $E^L_{\\max}$, so for a fixed privacy target the required Gaussian noise scale $\\sigma_{DP}$ can be smaller than in a noise-only Gaussian mechanism. The paper further claims that the per-round convergence of each personalized model is bounded by a quantity $\\Phi_n^{t+1}$ depending on transmit powers, subchannel assignments, transmission error probabilities, quantization resolutions, the federated-learning and personalized-learning rates, and the FL-PL weighting coefficient, and that the worst of these bounds across clients can be minimized. The minimization is solved by revealing a nested structure: once the shared aggregation-error term $\\Theta_L^t$ is minimized by a maximum-weight matching over clients and subchannels, each client's remaining bound depends only on its own learning rate and weighting coefficient, so the min-max problem separates into a matching step and a per-client convex optimization. Under the assumption $\\mu<2$ on the strong-convexity constant, the resulting policy is claimed optimal, and experiments on DNN, MLR, and CNN models over MNIST, FMNIST, and CIFAR-10 report large improvements over round-robin, random-selection, and fixed-configuration scheduling in accuracy, worst-client test loss, and fairness.","pith_inferences":["The paper does not claim it, but the same accounting principle should extend to any compression scheme with a bounded error: if sparsification or low-rank compression admits a maximum-error bound, that bound could be substituted into the Theorem 1 reasoning to reduce $\\sigma_{DP}$ further.","A testable extension the paper leaves open is the boundary $\\mu\\geq 2$: the convex-region argument in Theorem 5 is restricted to $\\mu<2$, so running the scheduler on a model whose empirical strong-convexity constant is at least 2 would delimit where the optimality claim holds.","The fairness objective is defined through the worst convergence bound; the same machinery could be re-pointed at worst-case accuracy, tail performance, or group-fairness constraints, but the paper only evaluates the loss-based fairness metric."],"forward_implications":["For a fixed privacy budget $(\\epsilon_Q,\\delta_Q)$, the required Gaussian noise scale $\\sigma_{DP}$ decreases as quantization error grows, so compressing model updates and protecting privacy become complementary goals rather than competing ones.","The convergence bound $\\Phi_n^{t+1}$ gives a computable per-client objective for fairness: scheduling decisions that minimize the worst bound directly target the worst-client expected loss.","Because the min-max problem decouples, the joint scheduling-and-configuration problem is solvable in polynomial time instead of as a mixed-integer program, with overall complexity $O(N^3 + N\\log(1/\\alpha))$.","Within the stated strongly-convex regime, the policy is claimed optimal, meaning no other joint choice of client selection, channel allocation, power control, learning rates, and weighting coefficients can strictly improve the worst-client convergence bound while keeping convergence rates equal across clients."],"supporting_citations":[{"why":"It states the max-divergence definition and the composition theorem used to derive the cumulative privacy budget in Theorem 1.","marker":"[37]"},{"why":"It provides the moments-accountant and composition tools used to convert multivariate Gaussian noise into a one-dimensional privacy bound.","marker":"[21]"},{"why":"It supplies the personalized FL objective and the bounded-gradient and bounded-distance assumptions used throughout the convergence analysis.","marker":"[1]"},{"why":"It supplies the DP-FL framework, clipping convention, and the mu/L constants that Assumption 1 builds on.","marker":"[22]"},{"why":"It gives the bit-error-rate expression used to model how imperfect channels corrupt quantized model parameters.","marker":"[38]"},{"why":"It supplies the polynomial-time optimal matching algorithm used for client selection and channel allocation.","marker":"[43]"},{"why":"It describes the empirical estimation of the smoothness and strong-convexity constants L and mu used to instantiate Assumption 1 in the experiments.","marker":"[46]"}],"fun_headline_variants":["Quantization boosts privacy and fairness in wireless FL","Min-max scheduler uses quantization for private fair FL","Wireless federated learning gains privacy and fairness from quantization","Quantization-assisted scheduling improves wireless FL accuracy by 87%","Turning quantization noise into privacy and fairness wins"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole argument assumes every client's loss function is a well-behaved, bowl-shaped objective (strongly convex and smooth) with bounded gradients and with each client's ideal model close to the global ideal; the tested neural networks with ReLU activations and cross-entropy loss do not actually satisfy these conditions, so the convergence bounds and the scheduling decisions derived from them are only strictly valid under that idealized assumption.","fun_headline_variants_meta":{"raw":{"variants":["Quantization boosts privacy and fairness in wireless FL","Min-max scheduler uses quantization for private fair FL","Wireless federated learning gains privacy and fairness from quantization","Quantization-assisted scheduling improves wireless FL accuracy by 87%","Turning quantization noise into privacy and fairness wins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000751,"raw_usage":{"total_tokens":3416,"prompt_tokens":1092,"completion_tokens":2324,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":708,"completion_tokens_details":{"reasoning_tokens":2249}},"tokens_in":708,"tokens_out":2324,"duration_ms":15219,"temperature":1.0,"reasoning_tokens":2249,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:25:02.757018+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Audit the quantization-assisted Gaussian mechanism by computing the empirical maximum divergence between outputs produced from two datasets that differ in one sample, using the paper's stated clipping threshold $C$, quantization bits $R$, and noise scale $\\sigma_{DP}$; if the measured privacy loss exceeds the claimed $(\\epsilon_Q,\\delta_Q)$ budget, the central claim that quantization errors can be counted toward the differential-privacy budget is wrong.","supporting_citations":[{"cited_title":"Deep learning with differential privacy,","cited_arxiv_id":null,"evidence_quote":"It provides the moments-accountant and composition tools used to convert multivariate Gaussian noise into a one-dimensional privacy bound."},{"cited_title":"Ditto: Fair and robust federated learning through personalization,","cited_arxiv_id":null,"evidence_quote":"It supplies the personalized FL objective and the bounded-gradient and bounded-distance assumptions used throughout the convergence analysis."},{"cited_title":"Federated learning with differential pri- vacy: Algorithms and performance analysis,","cited_arxiv_id":null,"evidence_quote":"It supplies the DP-FL framework, clipping convention, and the mu/L constants that Assumption 1 builds on."},{"cited_title":"On the general BER expression of one-and two- dimensional amplitude modulations,","cited_arxiv_id":null,"evidence_quote":"It gives the bit-error-rate expression used to model how imperfect channels corrupt quantized model parameters."},{"cited_title":"The Hungarian method for the assignment problem,","cited_arxiv_id":null,"evidence_quote":"It supplies the polynomial-time optimal matching algorithm used for client selection and channel allocation."},{"cited_title":"Adaptive federated learning in resource constrained edge computing systems,","cited_arxiv_id":null,"evidence_quote":"It describes the empirical estimation of the smoothness and strong-convexity constants L and mu used to instantiate Assumption 1 in the experiments."}],"review_version":1}