{"id":"deac5048-d78a-4888-bb7e-698618475674","arxiv_id":"2505.06268","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A cluster contribution threshold plus PPO-based power and update-frequency allocation improves simulated FL accuracy by roughly 3-6% over baselines in non-IID wireless settings.","lead":"Wireless federated learning slows down when devices have uneven data and channel quality. This paper groups similar devices into clusters and lets only high-contribution clusters train several rounds before uploading, then uses reinforcement learning to split energy between local training and transmission.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1 does not support the claimed smaller-A mechanism: condition (18) implies B_c<0, making N_c=1 increase A; the needed B_c>0 regime violates the convergence condition.","rationale":"The reader's conditional verdict was based mainly on the A4-trajectory gap and a claimed missing L/2 factor. The missing L/2 claim does not reproduce: Eq. (44) to Eq. (45) uses ||∇F(w[t])||² ≥ 2µ[F(w[t])−F(w*)] correctly, so the factor is present after multiplication. The A4 issue is real but secondary. The more serious problem is the sign of the per-cluster coefficient B_c defining A. The paper's 'smaller A' claim is true only if B_c > 0 for gated clusters, but Eq. (18) is precisely the condition B_c < 0. Therefore, under the theorem's own convergence condition, gating low-contribution clusters to N_c=1 makes A larger, not smaller. In the opposite regime where gating lowers A, the theorem's convergence condition fails and no A<1 guarantee is provided. This is not a minor constant error; it reverses the stated mechanism and makes P1's objective misaligned with the CAMU policy. The empirical gains are plausible and consistently reported, but they do not compensate for a central theoretical claim that its own equations contradict. I would therefore move from the reader's CONDITIONAL to REJECT in the current form, while noting that a revised theorem with a properly stated convergence condition and explicit sign analysis could make the paper salvageable.","tokens_in":18587,"tokens_out":11513,"duration_ms":116552,"concrete_test":"Evaluate Eq. (20) with a two-cluster instance satisfying (18): set λ=0.1, L=µ=1, δ=δ_c=1, G_c=0.5, Σ_k G_{k,c}=1, and Σ_k G_{k,c}²=0.5. Then B_c=-0.0975, A(N_c=1)=0.9025, and A(N_c=2)=0.805, so reducing N_c from 2 to 1 increases A, contradicting the Eq. (20) claim. Then choose a low-contribution cluster with δ_c=500 (so B_c>0) and verify that (18) is violated, so Theorem 1's convergence guarantee does not apply. Recomputing Eq. (20) with the paper's experimental parameters and reporting A_CAMU versus A_Benchmark2 across the used threshold values would settle whether the claimed 'smaller A' actually occurs.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that CAMU's threshold gating (N_c=1 for low-contribution clusters) reduces the convergence rate A in Theorem 1 and thereby accelerates convergence. From Eq. (46), A = 1 + Σ_c N_c B_c with B_c = µLλ²δ²G_c²δ_c²Σ_k G_{k,c}² − 2µλδG_cδ_cΣ_k G_{k,c}. The stated convergence condition (18) is LλδG_cδ_cΣ_k G_{k,c}² < 2Σ_k G_{k,c}, which is exactly B_c < 0. For any cluster satisfying (18), increasing N_c makes A smaller and reducing N_c to 1 makes A larger. Thus, under the paper's own sufficient convergence condition, the CAMU gating policy moves A in the opposite direction from the claimed one. If, instead, low-contribution clusters are in the B_c > 0 regime (where reducing N_c does lower A), then those clusters violate (18), so the theorem's A < 1 guarantee is unavailable; the paper does not prove that the sum Σ N_c B_c is negative in that regime. Consequently, Eq. (20)'s assertion that low-contribution clusters have larger δ_c and therefore N_c = 1 'obtains a smaller A' conflates the sign of δ_c with the sign of B_c: making δ_c large enough to make B_c positive also invalidates the convergence condition as written. This flaw is independent of the A4-to-local-trajectory gap in Eq. (41) and is more damaging, because it also means the P1 objective (23) could prefer increasing N_c for all clusters rather than enforcing the CAMU threshold.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies wireless federated learning (FL) with heterogeneous devices and communication channels. It proposes a two-step clustering mechanism that groups devices by communication quality and data distribution, and then introduces a Cluster-Aware Multi-round Update (CAMU) strategy: clusters whose contribution to the global update, measured by |D_c|e^{1/W_c}, falls below a threshold are forced to N_c=1 local update per global round, while other clusters perform N_c=1+n_c updates. A convergence upper bound is derived in Theorem 1, and a joint optimization problem (P1) over local update frequencies and transmission powers is formulated and solved with PPO. Simulations on MNIST and Fashion-MNIST with a CNN compare the proposed method against several baselines and report accuracy improvements.","tokens_in":18955,"tokens_out":6802,"duration_ms":67143,"significance":"The problem is timely, and the idea of using a data-driven contribution measure to gate the local update frequency is practically motivated. The paper makes its theoretical dependency explicit by stating a convergence theorem and a proof, and the experimental setup is described in enough detail to be reproduced. If the convergence analysis and the claimed rate-reduction mechanism were correct, the resource-allocation formulation would be a useful contribution to wireless FL. However, as detailed below, the central theoretical claim about how N_c affects the convergence rate contradicts the paper's own sufficient condition, and the proof contains a gap in the use of the dissimilarity assumption. These are load-bearing issues for the main contribution, not local presentation problems.","major_comments":[{"comment":"The central claim that setting N_c=1 for low-contribution clusters reduces A is contradicted by the paper's own sufficient condition. Define B_c = µLλ²δ²G_c²δ_c²Σ_k G_{k,c}² − 2µλδG_cδ_cΣ_k G_{k,c}; then A = 1 + Σ_c N_c B_c (Eq. (46)). Condition (18) is equivalent to LλδG_cδ_cΣ_k G_{k,c}² < 2Σ_k G_{k,c}, i.e. exactly B_c<0. Under (18), A decreases as each N_c increases, so forcing N_c to 1 increases A, which is the opposite of the claimed effect. If instead low-contribution clusters have B_c>0, then they violate (18), and the A<1 guarantee and the bound (17) are not available. Eq. (20)'s statement that low-contribution clusters “obtain a smaller A” does not follow; the sign of δ_c does not determine the sign of B_c, and the P1 objective (23) could favor larger N_c for all clusters. This invalidates the theoretical rationale for the CAMU gating rule.","section":"Section III-C, Eqs. (18), (20), (46)"},{"comment":"Assumption A4 is stated at the global iterate w[t] for each local loss function, but Eq. (41) applies it to every local iterate w_c^{[t,n]} along the multi-round trajectory, bounding each term by δ²||∇F(w[t])||². This extension requires a uniform dissimilarity bound along the local paths, which is neither assumed nor proved. Without such an assumption, the intermediate bound after Eq. (41) and the convergence result in Theorem 1 are not established.","section":"Appendix I, Eq. (41)"},{"comment":"The Taylor expansion (37) produces, in Eq. (44), a noise term with coefficient L/2, namely (L/2)Σ_c G_c²σ_n²/(p_c²||h_c||²). This factor L/2 is missing from the statement of Theorem 1 in Eq. (17) and from the optimization objective in P1, Eq. (23a). Since P1 is derived from the convergence bound, the theorem and the optimization problem do not match the proof.","section":"Appendix I, Eqs. (44)-(45) and Theorem 1, Eq. (17)"},{"comment":"Assumption A2 requires each local loss F_k to be µ-strongly convex, and the text in Section III-C says that the cross-entropy function is strongly convex. The experiments, however, train a CNN with cross-entropy loss, which is neither strongly convex nor convex as a function of the network parameters. The convergence guarantee of Theorem 1 therefore does not apply to the evaluated system, and the statement that cross-entropy is strongly convex is incorrect.","section":"Section III-C, Assumptions A1-A4; Section V-A"}],"minor_comments":[{"comment":"The channel notation is inconsistent: h_k appears as a vector in Eq. (6) and later as ||h_c|| in the convergence bound, while the SNR γ_k in Eq. (7) uses |h_k|² without clarifying whether this is the squared Euclidean norm. The noise power is denoted both σ²_{n0} and σ²_n; please unify the notation.","section":"Section II-B and Appendix I"},{"comment":"The reported curves appear to be single runs without error bars or confidence intervals; this makes it difficult to assess the statistical significance of the claimed performance improvements, which are often small (below 6%).","section":"Section V, Figs. 4-6"},{"comment":"Reference [29] duplicates reference [4] (same title and author list); this duplicate citation should be corrected.","section":"References"},{"comment":"The local update frequency is defined as N_c = 1 + S_c·n_c in Eq. (15), but the domain of n_c is not specified. Since P1 optimizes n_c, the paper should state integer or real bounds and explain how the PPO action projection enforces them.","section":"Section III-B and P1"}],"recommendation":"reject","confidential_remarks":"The sign error in the convergence-rate analysis is load-bearing and contradicts the paper's central claim, so I cannot recommend major revision in the present form; the theoretical basis of CAMU would need to be re-derived. The duplicate self-citation [4]/[29] should also be corrected in any future submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead this one for the experimental setup and the heuristic; don't rely on the theorem. The paper's real contribution is a sensible engineering idea: cluster devices by communication quality and data distribution, then gate multi-round local updates by a cluster-contribution threshold (Wasserstein-weighted), and use PPO to allocate local update counts and transmit power under an energy budget. That combination is new in the FL-systems literature, and the experiments show small but consistent accuracy gains (a few percent) over reasonable baselines on MNIST/Fashion-MNIST in heterogeneous settings. The convergence analysis is self-contained and the PPO reward is the derived bound, so no fitting to test accuracy is going on. That is real work and citable for the heuristic.\n\nThe soft spot is load-bearing. In Theorem 1, A = 1 + Σ N_c B_c, and condition (18) is exactly B_c < 0. So under the paper's own convergence condition, larger N_c makes A smaller. CAMU sets low-contribution clusters to N_c=1, which makes A larger, not smaller. The paper's statement that this 'obtains a smaller A' gets the sign backwards. If one instead assumes B_c > 0 for those clusters, then reducing N_c does shrink A, but those clusters violate (18), so the A<1 guarantee is lost and the whole bound unravels. The authors conflate a larger heterogeneity factor δ_c with a positive B_c; the two are not the same. This is not a nitpick: the optimization objective P1 and the entire motivation for threshold gating are built on that bound. The heuristic may still work empirically, but the theory as written does not support the claim.\n\nOther issues are smaller: the proof uses A4 at global iterates but applies it to every local iterate along the trajectory without justification; the noise term in the theorem statement is missing the L/2 factor that appears in the proof; the strong-convexity assumption doesn't match the CNN cross-entropy experiments; and the plotted curves are single runs with no error bars. Power/SNR units are also sloppy (pc ∈ C then |p_k|² in constraints).\n\nMy recommendation: send it to review, but require the authors to fix the sign argument or explicitly reframe the contribution as a heuristic without Theorem 1 as support. The experimental part deserves an outlet; the theory needs more work. I wouldn't cite the theorem in its current form, but I'd bring the paper to a reading group to discuss exactly this kind of sign error.","headline":"Useful heuristic and clean experiments, but the central convergence argument in Theorem 1 has a sign error that reverses the claimed mechanism.","tokens_in":19483,"tokens_out":6189,"would_cite":false,"duration_ms":57047,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that throttling low-contribution clusters to one local update per round suppresses bias accumulation and accelerates convergence of wireless federated learning.","keywords":["wireless federated learning","non-IID data","cluster contribution threshold","local update frequency","convergence analysis","resource allocation","reinforcement learning optimization"],"falsifier":"On the paper's reported MNIST or Fashion-MNIST setup, measure the ratio $\\|\\nabla F_c(w_c^{[t,n]})\\|^2 / \\|\\nabla F(w^{[t]})\\|^2$ over the $n$ local rounds; if any cluster's ratio exceeds $\\delta^2$ for $n>0$, inequality (41) fails and the claimed contraction factor $A$ is unsupported.","tokens_in":18374,"feed_emoji":"📡","tokens_out":6984,"duration_ms":73980,"temperature":0.7,"pith_summary":"Federated learning over wireless links degrades when devices have different data and communication quality. This paper groups similar devices into clusters and then lets only high-contribution clusters run several local training rounds per global aggregation, holding low-contribution clusters to a single round. The authors derive a convergence upper bound whose contraction factor is reduced when low-contribution clusters are throttled, and they jointly optimize each cluster's local-round count and transmit power under an energy budget. If the analysis holds, the scheme would improve model accuracy and convergence in exactly the heterogeneous, resource-limited settings where plain federated averaging degrades.","feed_headline":"Cap low-impact clusters at one local step to speed federated learning","feed_subtitle":"A cluster-aware update rule cuts non-IID bias and jointly budgets computation and transmit power.","key_machinery":"The load-bearing object is the cluster contribution $\\vartheta_c = |D_c| e^{1/W_c}$, where $W_c$ is the Wasserstein distance between the cluster's label distribution and the global label distribution. A threshold $\\vartheta_{\\mathrm{th}}$ switches each cluster's local update frequency $N_c = 1 + S_c n_c$, with $S_c=1$ if $\\vartheta_c \\geq \\vartheta_{\\mathrm{th}}$ and $S_c=0$ otherwise. This contribution score determines which clusters may spend computation on multiple local rounds; the convergence-rate constant $A$ in Theorem 1 is the mechanism through which that choice changes the contraction, and the energy constraint couples $N_c$ to transmit power $p_c$ in the optimization objective.","core_discovery":"The paper's central claim is Theorem 1: under L-smoothness, μ-strong convexity, and a δ-local dissimilarity assumption, the expected optimality gap after T aggregation rounds satisfies $$\\mathbb{E}[F(w[t+1])-F(w^*)] \\leq A^T \\mathbb{E}[F(w[0])-F(w^*)] + \\frac{1-A^T}{1-A}\\sum_{c} \\frac{$G_c^{2}$ \\$sigma_n^{2}$}{$p_c^{2}$ \\|h_c\\|^2},$$ with $$A = 1 + \\sum_{c} N_c\\left(\\mu L \\$lambda^{2}$ \\$delta^{2}$ $G_c^{2}$ \\$delta_c^{2}$ \\sum_{k} G_{k,c}^2 - 2\\mu \\$\\lambda$ \\delta G_c \\delta_c \\sum_{k} G_{k,c}\\right).$$ Because low-contribution clusters are taken to have larger dissimilarity factors, forcing their update frequency to $N_c=1$ removes their multi-round bias contribution from $A$; the paper argues this yields a smaller $A$ and faster convergence. The paper further claims that this makes the joint allocation of per-cluster transmit power and local update frequency under a total energy budget a well-posed optimization that can be solved with an actor-critic reinforcement-learning optimizer.","pith_inferences":["Because the contribution threshold is based only on label-distribution distance and sample count, a natural extension is to make the threshold adaptive over training or to define contribution from observed gradient norms rather than a priori data statistics.","The proof assumes error-free intra-cluster transmission; a testable extension is to relax that assumption and see where the convergence advantage erodes when cluster-leader links are noisy or bandwidth-limited.","The same contribution-threshold mechanism could be applied to client selection rather than update frequency, excluding low-contribution clusters entirely in some global rounds.","The convergence bound predicts a specific dependence of the optimality gap on $p_c$ and $N_c$, so a direct check is to fit the measured loss trajectory to the bound and compare the estimated contraction factor with the formula for $A$."],"forward_implications":["If the theorem is correct, a cluster below the contribution threshold contributes no multi-round term to the convergence-rate constant, so holding it to one local update reduces the bias-accumulation component of the global error bound.","The joint optimization of local update frequency and transmit power gives a concrete answer to the computation-communication tradeoff: within the same total energy budget, solving for both quantities outperforms fixing power and optimizing only local rounds.","The two-stage clustering, by SNR first and label distribution second, creates near-IID intra-cluster subsystems, so the corollary's condition $\\delta G_c < 1$ can be met by downweighting heterogeneous clusters during aggregation.","In highly non-IID settings, the strategy's advantage over unrestricted multi-round updates grows, matching the reported MNIST and Fashion-MNIST accuracy gains over the baselines."],"supporting_citations":[{"why":"Supplies the Wasserstein-distance formulation used to define cluster contribution and aggregation weights.","marker":"[27]"},{"why":"Supplies the affinity propagation algorithm used for SNR-based primary clustering.","marker":"[28]"},{"why":"Provides the theoretical convergence analysis of local SGD with multiple local updates that motivates the local multi-round framework.","marker":"[13]"},{"why":"Provides the cluster-based local multi-round update scheme used as Benchmark 2 in the experiments.","marker":"[22]"},{"why":"Provides the local-iteration-only optimization baseline, Benchmark 3, against which the joint optimization is compared.","marker":"[30]"}],"fun_headline_variants":["Cluster-aware local steps fix non-IID bias in wireless FL","One-step clusters cut update bias and power costs in FL","Jointly budget power and updates per cluster to speed FL","Heterogeneous FL speedup via cluster-aware update limits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The convergence bound assumes that every local update's gradient stays within a fixed multiple of the global gradient, even after several local rounds; if dissimilarity grows along the local trajectories, the bound and the optimization built on it are not justified.","fun_headline_variants_meta":{"raw":{"variants":["Cluster-aware local steps fix non-IID bias in wireless FL","One-step clusters cut update bias and power costs in FL","Jointly budget power and updates per cluster to speed FL","Heterogeneous FL speedup via cluster-aware update limits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00019,"raw_usage":{"total_tokens":1360,"prompt_tokens":989,"completion_tokens":371,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":303}},"tokens_in":605,"tokens_out":371,"duration_ms":4411,"temperature":1.0,"reasoning_tokens":303,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:59:22.147640+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the paper's reported MNIST or Fashion-MNIST setup, measure the ratio $\\|\\nabla F_c(w_c^{[t,n]})\\|^2 / \\|\\nabla F(w^{[t]})\\|^2$ over the $n$ local rounds; if any cluster's ratio exceeds $\\delta^2$ for $n>0$, inequality (41) fails and the claimed contraction factor $A$ is unsupported.","supporting_citations":[{"cited_title":"Clustering by passing messages between data points,","cited_arxiv_id":null,"evidence_quote":"Supplies the affinity propagation algorithm used for SNR-based primary clustering."},{"cited_title":"Petrel: Heterogeneity-aware distributed deep learning via hybrid synchroniza- tion,","cited_arxiv_id":null,"evidence_quote":"Provides the cluster-based local multi-round update scheme used as Benchmark 2 in the experiments."},{"cited_title":"Heterogeneous training in- tensity for federated learning: A deep reinforcement learning approach,","cited_arxiv_id":null,"evidence_quote":"Provides the local-iteration-only optimization baseline, Benchmark 3, against which the joint optimization is compared."}],"review_version":1}