{"id":"baaf1db5-1c1c-4aea-b3b4-2ac92883d54f","arxiv_id":"1908.06287","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The convergence rate of wireless federated learning is governed by a scheduling-and-decoding success probability, with proportional fair best at high SINR and round robin best at low SINR.","lead":"This paper develops an analytical model for how federated learning converges over wireless networks when only some devices can be scheduled for each update and transmissions can fail. It compares three scheduling rules, random, round robin, and proportional fair, and finds the best rule depends on the signal quality threshold.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's proof conflates per-UE success probability U with E[η_t]=G U, so the printed rate expression is unsupported; the scheduling conclusions likely survive but need re-derivation.","rationale":"The reader's weakest-assumption identification is correct. The key step is eq. (26): it claims η_t converges to U^z_k, but the definition in (22)/(25) has N in the denominator. Since the total number of successful events in one round has mean K U^z_k, the limiting average is G U^z_k. Lemma 2 is the bridge between the wireless success statistics and the optimization progress, and it uses the η_t/U identification in the convexity inequality (60); once that identification is off by a factor G, the theorem's exponent is not the one proved. I considered whether this could be dismissed as a semantic choice, i.e., U meaning success probability conditional on being scheduled. It cannot, because eq. (16) defines U with the selection event, and Corollaries 1 and 3 explicitly divide by G, while Corollary 2 handles the periodicity separately; the mixed usage is the source of the factor. The paper has genuine independent support: the Section V SVM and CNN experiments show the claimed orderings, and the framework is clearly salvageable. A normalization fix plus a re-derivation of the corollaries is an appropriate condition rather than a rejection. Hence no change to the reader's CONDITIONAL verdict.","tokens_in":23912,"tokens_out":12580,"duration_ms":134505,"concrete_test":"Re-derive Appendix A and B with the corrected normalization: compute E[η_t] from (25) explicitly for RS, RR, and PF, then replace the coefficient U^z_k in (27) and (30) by G U^z_k, capped at 1. Re-plot Fig. 3 with both the printed and corrected rate expressions using the same parameters. If the corrected curves preserve the high/low SINR ordering, the qualitative central claim stands and the paper needs only a normalization fix; if the ordering flips, or the required rounds change by substantially more than the G factor, the comparison rests on the erroneous equality (26).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equations (22) and (25) define η_t as a time average of the number of successful UE updates per communication round divided by N. Each of the K UEs succeeds with probability U^z_k as defined in (16), so the expectation of the numerator in (25) is K U^z_k and therefore E[η_t] = (K/N)U^z_k = G U^z_k, not U^z_k as asserted in (26); the asserted equality holds only for K=N. This is not a harmless notational choice. The proof of Lemma 2 uses the value of η_t as a convex-combination coefficient when bounding R_k(a_t + η_t Δ a_t) in (60), and it substitutes U^z_k for η_t; with the actual averaging denominator the coefficient is G U^z_k. The same factor propagates through the contraction bound in Theorem 1, which should contain 1 - (1-β)G U^z_k rather than 1 - (1-β)U^z_k. Hence the round-count expression (30) and the corollary formulas that feed the figures are not established as written. The qualitative comparisons of PF over RS/RR at high SINR and RR over PF at low SINR may well survive once the factor is inserted, because in the log-linear regime the G factors are partly absorbed; but that is a repair, not a proof contained in the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a wireless federated learning (FL) algorithm (Algorithm 2) in which the local learning rate η_t is a running average of the number of successful UE parameter updates, adaptively coupling the optimization step to the wireless channel state. The authors model AP and UE locations as Poisson point processes, derive parameter update success probabilities under random scheduling (RS), round robin (RR), and proportional fair (PF) scheduling using stochastic geometry, and combine these quantities with a CoCoA-style dual coordinate ascent analysis to obtain explicit expressions for the number of communication rounds needed to reach an ε duality gap (Theorem 1). Based on these expressions, they conclude that PF outperforms RS and RR at high SINR thresholds, that RR is preferable at low SINR thresholds, and that the convergence rate degrades rapidly as the SINR threshold increases. The paper also includes SVM and CNN experiments that qualitatively support the scheduling-policy comparisons.","tokens_in":24206,"tokens_out":7036,"duration_ms":67530,"significance":"The paper addresses a timely and important problem: quantifying how wireless scheduling and inter-cell interference affect the convergence of FL. Its main contribution is a tractable analytical framework that connects physical-layer transmission success probabilities with an optimization-theoretic convergence rate. The success probabilities are derived from the channel model rather than fitted to the convergence goal, and the convergence analysis follows an established contraction framework, which is a genuine strength. If the proofs are repaired, the paper would provide a useful design tool for wireless FL systems and a clear basis for comparing scheduling policies. The experimental results are a further positive feature, although they do not by themselves compensate for the theoretical issue described below.","major_comments":[{"comment":"The assertion after Eq. (26) that U^z_k = E[η_t] is not correct as written. From Eq. (25), η_t = (1/(N t)) Σ_{l=0}^{t-1} Σ_{k=1}^K 1{S^z_{k,l}=1, γ_{k,l}>θ}. Since each indicator has expectation U^z_k by definition (16), the law of large numbers gives E[η_t] = (K/N) U^z_k = G U^z_k, not U^z_k. The equality U^z_k = E[η_t] holds only in the special case K = N (G = 1). This factor-of-G error is load-bearing because the proof of Lemma 2 in Appendix A uses E[η_t] as the convex-combination coefficient in Eq. (60). With the correct expectation, the contraction factor in Theorem 1, Eq. (30), should contain 1 − (1−β) G U^z_k rather than 1 − (1−β) U^z_k. Consequently, the round-count expression (30) and all corollary formulas that depend on it are not established as printed.","section":"Section IV-A, Eq. (26)"},{"comment":"There is an internal inconsistency in the treatment of RR. Definition (16) defines U^z_k as the joint probability P(γ_{k,t}>θ, S^z_{k,t}=1). For RR, a typical UE is scheduled in only one out of every G rounds, so the joint per-round success probability should be (1/G)(1/(1+V(θ,α))). However, Eq. (35) states U^RR_k = 1/(1+V(θ,α)) if scheduled and 0 otherwise, which is the success probability conditional on being scheduled, not the joint probability. Appendix D then multiplies the round count by G in Eq. (36) and in Eq. (70). This means the RR branch uses a different object than the RS and PF branches, so the comparison of RR with RS and PF in Theorem 1 is not well-defined on a common footing. The authors should either define U^z_k consistently as a per-round joint probability and remove the extra G factor, or explicitly re-derive the RR contraction over the G-round cycle.","section":"Corollary 2, Eq. (35), and Appendix D"},{"comment":"The numerical results in Section V are computed from the printed corollary formulas, which inherit the factor-of-G error from Theorem 1. For RS, the corrected contraction factor would be 1 − (1−β)/(1+V(θ,α)) rather than 1 − (1−β)/(G(1+V(θ,α))), assuming independent and identically distributed UEs. Similar corrections apply to the PF formula. The qualitative conclusions of the paper may survive this correction, since the G factors would partially cancel in the log-linear asymptotic comparisons of Remarks 6 and 7, but that is a repair that must be carried out explicitly. As it stands, Figures 3–5 do not validate the theorem as stated.","section":"Corollaries 1 and 3 and Section V"}],"minor_comments":[{"comment":"The indicator function is rendered as \"/BD\" in several equations (e.g., Eq. (22), Eq. (25), Eq. (26), Eq. (57)–(59)). This appears to be a typesetting artifact and should be replaced with the standard indicator notation 1{·}.","section":"Eqs. (22), (25), (26), and throughout the appendix"},{"comment":"Algorithm 2 initializes η_0 = K/(2N), but the time-average expression in Eq. (25) does not include this initial value. For t ≥ 1, the recurrence (22) would yield η_t = (1/(N t)) Σ_{l=0}^{t-1} Σ_k 1{...} only if η_0 = 0. The role of the initialization should be clarified, or Eq. (25) should be stated as valid for t ≥ 1 with an explicit note that η_0 is not used.","section":"Algorithm 2 and Eq. (25)"},{"comment":"The asymptotic expression after Eq. (56) writes T_NS ≳ log(n/ε)/log(β). Since β ∈ (0,1), both numerator and denominator would be negative, yielding a positive but confusing expression. This appears to be a sign typo and should read log(ε/n)/log(β), consistent with the paper's other round-count formulas.","section":"Eq. (56)"},{"comment":"The proof of Corollary 5 is only a sketch. The statement that “both the parameter update success probability and the required communication rounds are increased by a factor of C” needs a more explicit derivation, since Theorem 1 as stated has no parameter C and the interaction between collecting more updates per global aggregation and the per-round contraction is not immediate.","section":"Corollary 5"},{"comment":"The approximation (1+θ)^{1/N} − 1 ≈ θ/N used in the proof of Corollary 4 is stated without conditions. Since the corollary is presented as a general result, the authors should specify the parameter regime in which this approximation is valid, or provide the exact expression.","section":"Corollary 4"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper is worth reading, but the main theorem is not proven as stated. The setup is new: it puts federated learning on a wireless network with inter-cell interference, scheduling, and unreliable uplinks, and combines a CoCoA-style dual coordinate ascent with stochastic geometry to derive closed-form convergence rates for RS, RR, and PF. That combination is the real contribution, and the authors honestly cite the building blocks (CoCoA, PPP-based SINR coverage, Choi–Bahk for proportional fairness).\n\nThe core problem is a factor-of-G confusion between the per-UE update success probability U_k^z and the global learning rate η_t. Equation (25) defines η_t as the time average of the number of successful UE updates per round divided by N. Since each of the K UEs succeeds with probability U_k^z, the expectation is (K/N)·U_k^z = G·U_k^z, not U_k^z as asserted in (26). The proof of Lemma 2 then uses E[η_t] = U_k^z in the convexity bound (60), and that error propagates into the contraction factor of Theorem 1: it should contain 1 − (1−β)G·U_k^z, not 1 − (1−β)U_k^z. So the round-count expressions (30), (33), (36), and (39) are unsupported as written. This is not cosmetic; the correction changes the dependence on G and may alter some of the policy comparisons. The qualitative story—PF wins at high SINR, RR wins at low—might survive a re-derivation, but the paper as it stands does not prove it.\n\nThe experiments are helpful but under-controlled: in Fig. 8 the CNN uses K=30 in the high-SINR regime and K=100 in the low-SINR regime, so the regime comparison is confounded, and no error bars are reported despite only five trials.\n\nWhat the paper does well: the system model is clean, the stochastic geometry derivations are standard but competently executed, and the connection between scheduling and FL convergence fills a recognized gap in the literature. The failure is in the optimization-theoretic part, not the wireless part.\n\nThis deserves a serious referee. The contribution is meaningful enough that a major revision fixing the η_t normalization, reproving Lemma 2 and Theorem 1, and re-running the experiments with matched K would make it a solid paper. I would not trust the quantitative claims until that is done.","headline":"A useful framework with a real proof gap: the convergence-rate expressions are unsupported because η_t is equated to U_k^z when it should be G·U_k^z, but the core ideas and qualitative conclusions are salvageable.","tokens_in":24732,"tokens_out":7239,"would_cite":false,"duration_ms":70028,"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":"Federated learning's convergence rate in wireless networks is set by a single per-user quantity: the probability that an update is scheduled and successfully decoded.","keywords":["federated learning","wireless scheduling","proportional fair","round robin","random scheduling","convergence rate","stochastic geometry","SINR threshold"],"falsifier":"Run Algorithm 2 on a convex task with known data size $n$ and record $\\eta_t$ alongside the empirical fraction of successfully decoded scheduled updates; if $\\eta_t$ converges to $G$ times that fraction rather than to the fraction itself, the premise of Lemma 2 fails and the predicted round counts $T_z$ would need rescaling.","tokens_in":23691,"feed_emoji":"📡","tokens_out":7759,"duration_ms":72349,"temperature":0.7,"pith_summary":"Federated learning in a wireless edge network cannot count on every user's update arriving: only a subset of user devices is scheduled in each round, and transmissions can be lost to interference. This paper develops an analytical model, based on random spatial layouts of access points and users, that expresses how fast federated learning converges in terms of a single per-user quantity: the probability that a user is both scheduled and successfully decoded under a given scheduling policy. Using that quantity, the paper derives closed-form expressions for the number of communication rounds needed to reach any target duality gap under random scheduling, round robin, and proportional fair scheduling. The paper's conclusion is that the best policy depends on the SINR decoding threshold: proportional fair converges fastest when the threshold is high, round robin when the threshold is low, and convergence slows sharply as the threshold rises. The practical upshot is that scheduling and signal-quality design directly control federated-learning training time, and that reducing the decoding threshold via compression or quantization can sometimes be worth more than improving local computation accuracy.","feed_headline":"Federated learning's best wireless schedule flips with SINR","feed_subtitle":"PF wins when the decoding threshold is high, round robin when low; new bound ties speed to update-success probability.","key_machinery":"The load-bearing object is the parameter update success probability $U_k^z = P(\\gamma_{k,t} > \\theta, S_{k,t}^z = 1)$: the joint probability that user $k$ is selected by the scheduler and that its packet decodes above the SINR threshold $\\theta$. Algorithm 2 couples this to the optimization by setting the local learning rate $\\eta_t$ to the time-averaged fraction of successful update indicators, so that users advance more aggressively when the channel is reliable and more cautiously when it is not. Lemma 2 shows that the expected increase in the dual objective in any round is no less than $U_k^z$ times the sum of local subproblem improvements, which converts the scheduling question into a calculation of $U_k^z$; Theorem 1 then turns that into round counts. The stochastic-geometry quantity $V(\\theta,\\alpha)$, an interference-plus-noise-to-signal measure, supplies the closed forms for $U_k^z$ under each policy.","core_discovery":"The central claim is Theorem 1: for any convergence target $\\epsilon$, the wireless federated-learning procedure in Algorithm 2 — a distributed dual optimization method with an adaptive local learning rate — reaches an $\\epsilon$ duality gap after at least $T_z \\ge \\log(\\epsilon/n) / \\log(1 - (1-\\beta)U_k^z)$ communication rounds, where $\\beta$ is the local-solver error level, $n$ is the total data size, and $U_k^z$ is the parameter update success probability of a user under scheduling policy $z$. The paper computes $U_k^z$ under three policies using stochastic geometry, obtaining closed forms for random scheduling, round robin, and proportional fair. From these expressions, the paper concludes that proportional fair scheduling outperforms the other two when the SINR threshold is high, because it selects users with the strongest channels; round robin is preferable when the SINR threshold is low, because almost every transmission succeeds and guaranteeing each user a turn matters more; and the convergence rate degrades rapidly as the SINR threshold increases. The analysis also identifies a trade-off between the number of scheduled users and subchannel bandwidth when the total spectrum is fixed.","pith_inferences":["Editorial inference: because the bound depends on $U_k^z$ only through the product $(1-\\beta)U_k^z$, the framework suggests a general law of wireless federated learning — convergence speed is set by the probability that a user both computes and communicates successfully, so policies or protocols that raise either factor are interchangeable levers on training time.","Editorial inference: the SINR threshold $\\theta$ can be reinterpreted as an abstraction of the entire physical layer — modulation, coding, quantization, and model size — so the model predicts that anything that moves $\\theta$ down, such as stronger codes or fewer bits per parameter, buys convergence speed in a quantitatively predictable way.","Editorial inference: if the assumption that every access point sees statistically identical conditions is relaxed, the same per-cell success-probability machinery could be extended to study how uneven cell loads affect global convergence; the paper does not do this, but its single-cell analysis is a natural starting point.","Editorial inference: the three schedulers compared here are only a sample; the same analytic path would rate energy-aware, age-of-information-based, or delay-aware schedulers by computing their $U_k^z$, turning scheduler design for federated learning into a probabilistic optimization over one scalar."],"forward_implications":["If the central bound is correct, the communication-round count of wireless federated learning is determined by the scheduler through a single number $U_k^z$, so a new scheduling policy can be plugged into the same theorem by computing its update-success probability.","At high SINR thresholds, proportional fair scheduling can reduce required rounds by roughly a factor of $G$ relative to random or round robin, making channel-aware selection valuable exactly when transmissions are fragile.","At low SINR thresholds, round robin is preferable because per-transmission failures are rare and the binding constraint is guaranteeing each user a turn; fair time-sharing beats channel chasing.","Raising the decoding threshold degrades convergence quickly, so compressing or quantizing model updates to lower the required SINR is a direct lever on training time, sometimes outweighing gains from better local optimization — the paper's numerics show a 5 dB threshold reduction offsetting a six-fold larger error level.","A fixed spectrum budget forces a trade-off: scheduling more users per round increases the number of updates aggregated but splits the bandwidth and lowers success probability, and the analysis says the optimum lies at an interior number of subchannels."],"supporting_citations":[{"why":"Supplies the distributed primal-dual local-solver framework that Algorithm 2 adapts to the wireless setting.","marker":"[20]"},{"why":"Defines the proportional fair scheduler and the channel-gain selection result used to compute $U_k^{PF}$.","marker":"[29]"},{"why":"Supplies the stochastic-geometry tools — Poisson point process model, Slivnyak's theorem, and coverage probability — used to evaluate SINR success probabilities.","marker":"[33]"},{"why":"Gives the general communication-efficient distributed optimization framework whose convergence analysis is extended, including the bound $E[D(v_0)-D(v^*)] < n$ used in the proof of Theorem 1.","marker":"[37]"},{"why":"Motivates the scheduling-aware federated-learning problem and provides the empirical observation that packing more users into each round accelerates convergence.","marker":"[10]"},{"why":"Approximates interfering uplink user positions by a non-homogeneous Poisson point process, which is used in the derivations of the success-probability formulas.","marker":"[38]"}],"fun_headline_variants":["Scheduling choice for wireless FL flips with SINR","Proportional fair wins at high SINR for federated learning","Round robin beats PF at low SINR in wireless FL","FL convergence bound reveals optimal scheduling policy","Wireless FL: pick schedule by SINR threshold"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise, used in the proof of Lemma 2 through eq. (26), is that the adaptive learning rate $\\eta_t$ in Algorithm 2 converges exactly to each user's probability of a successful scheduled update, with no extra factor from the ratio of users to subchannels; the entire round-count bound of Theorem 1 rests on that alignment.","fun_headline_variants_meta":{"raw":{"variants":["Scheduling choice for wireless FL flips with SINR","Proportional fair wins at high SINR for federated learning","Round robin beats PF at low SINR in wireless FL","FL convergence bound reveals optimal scheduling policy","Wireless FL: pick schedule by SINR threshold"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000558,"raw_usage":{"total_tokens":2734,"prompt_tokens":1105,"completion_tokens":1629,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":721,"completion_tokens_details":{"reasoning_tokens":1551}},"tokens_in":721,"tokens_out":1629,"duration_ms":10096,"temperature":1.0,"reasoning_tokens":1551,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:52:36.231102+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 2 on a convex task with known data size $n$ and record $\\eta_t$ alongside the empirical fraction of successfully decoded scheduled updates; if $\\eta_t$ converges to $G$ times that fraction rather than to the fraction itself, the premise of Lemma 2 fails and the predicted round counts $T_z$ would need rescaling.","supporting_citations":[{"cited_title":"Distributed optimization with arbitrary l ocal solvers,","cited_arxiv_id":null,"evidence_quote":"Supplies the distributed primal-dual local-solver framework that Algorithm 2 adapts to the wireless setting."},{"cited_title":"Cell-throughput analysis of th e proportional fair scheduler in the single-cell environment,","cited_arxiv_id":null,"evidence_quote":"Defines the proportional fair scheduler and the channel-gain selection result used to compute $U_k^{PF}$."},{"cited_title":"CoCoA: A general framework for communication-efficient di stributed optimization,","cited_arxiv_id":null,"evidence_quote":"Gives the general communication-efficient distributed optimization framework whose convergence analysis is extended, including the bound $E[D(v_0)-D(v^*)] < n$ used in the proof of Theorem 1."},{"cited_title":"Client Selection for Federated Learning with Heterogeneous Resources in Mobile Edge","cited_arxiv_id":"1804.08333","evidence_quote":"Motivates the scheduling-aware federated-learning problem and provides the empirical observation that packing more users into each round accelerates convergence."},{"cited_title":"User point processes in cellular networks ,","cited_arxiv_id":null,"evidence_quote":"Approximates interfering uplink user positions by a non-homogeneous Poisson point process, which is used in the derivations of the success-probability formulas."}],"review_version":1}