{"id":"2e91c063-ae0d-4394-9a8e-ded94b04480f","arxiv_id":"2509.02538","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A stochastic post-coding step makes quantized over-the-air gradient transmissions unbiased, yielding federated SGD with near-centralized convergence guarantees at lower communication cost.","lead":"This paper designs federated learning algorithms that send gradient information through noisy wireless channels with low-resolution analog-to-digital converters, and corrects the bias those converters introduce. If the approach works as claimed, it gives convergence rates close to standard centralized SGD while using roughly one fifth of the communication symbols of coded transmission.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1 is stated with synchronization condition (9b), but its proof in §6.2 requires T(τ_i)−T(τ_{i−1}) < 1/[2L(v*+Δ²)], a condition only implied by (9b) when the unstated inequality v*+Δ² ≤ 1 holds.","rationale":"I read the paper in good faith: the post-coding, scale-adaptive transmission, and periodic synchronization framework is coherent, and Lemmas 1 and 2 are derived in real detail. The central risk is that the main theorem's proof uses a synchronization condition that the theorem statement does not state. The reader identified this same mismatch, so the core concern is shared. I would qualify the reader's wording: the proof's condition is not uniformly stronger than (9b) — for the q = 8 and q = 16 simulations, 5Δ² < 1, so (9b) actually implies the proof's requirement. The gap is real for coarse quantization (q = 4 or 5) and for unproven regimes where v* is large. The secondary concern about the low-SNR simulation lying outside Lemma 1's feasibility regime is legitimate and worth reporting: the paper should state whether LP (6) was verified feasible there. It does not by itself invalidate the conditional theory. Since the main gap is concrete, addressable, and does not appear to be fatal to the overall approach, the conditional verdict stands unchanged.","tokens_in":20989,"tokens_out":18501,"duration_ms":169451,"concrete_test":"Solve the LP (6) for q = 4 (Δ = 2/3) over σc ∈ (0, Δ/2] and compute v* + Δ². If for any such σc the value exceeds 1, then the proof's synchronization condition is strictly stronger than (9b) within Lemma 1's own feasibility regime; independently re-derive the D_k absorption step after Eq. (13) to confirm that Theorem 1's stated assumptions are then insufficient. If v* + Δ² ≤ 1 for all such σc, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof-to-statement mismatch in Theorem 1 is the most load-bearing gap. In §6.2, around Eq. (13), the disagreement term 6LΣη_kD_{k−1} is absorbed only under the condition T(τ_i)−T(τ_{i−1}) < 1/[2L(v*+Δ²)], together with a small-stepsize condition. The theorem statement instead assumes (9b), namely T(τ_i)−T(τ_{i−1}) ≤ 1/(2L). The proof's condition is implied by (9b) only when v*+Δ² ≤ 1, an inequality the paper neither assumes nor proves. Lemma 1 gives only v* ≤ 4Δ², so v*+Δ² ≤ 5Δ² = 20/(q−1)², which exceeds 1 for q = 4 or 5. Since the theorems are stated for general q, the proof does not cover the stated assumptions: for coarse quantization the negative gradient/function terms in Eq. (13) may fail to dominate the disagreement terms, so the claimed near-optimal bound is not established as written. The gap is fixable by stating the channel-dependent synchronization condition or by restricting the quantization resolution, but as written it is a real correctness issue in the central theorem.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies federated optimization over AWGN physical channels with low-resolution ADC/DAC quantization. It proposes a stochastic post-coding linear program that makes the quantized noisy channel unbiased on interior quantization levels, a scale-adaptive transformation that transmits normalized values over the air while sending scale indices over a coded channel, and periodic synchronization of worker models. For strongly convex and non-convex losses it proves convergence rates comparable to centralized SGD up to a channel-dependent term (v*+Δ²)ω²d, and it reports MNIST experiments with a CNN showing accuracy close to coded transmission at substantially lower symbol cost.","tokens_in":21289,"tokens_out":12162,"duration_ms":112747,"significance":"If the proofs are completed, this is a useful contribution: the paper gives a concrete construction (LP (6)) that turns a biased quantized AWGN channel into an approximately unbiased stochastic gradient oracle with an explicit variance bound, and the convergence analysis covers state-dependent gradient noise and periodic synchronization. The explicit constants in Lemmas 1 and 2 and the self-contained proof structure are strengths, as is the clear separation of the unbiasedness construction from the optimization analysis. The advertised near-optimal risk, matching centralized SGD up to the channel-dependent term (v*+Δ²)ω²d, would be of interest to the federated-learning-over-wireless-channels community. However, the main theorem as stated is not fully supported by the proof as written, and one simulation operating point lies outside the proved feasibility regime of the post-coding LP; these issues require correction before the central claims can be accepted.","major_comments":[{"comment":"Theorem 1 is stated with the synchronization condition T(τ_i)-T(τ_{i-1}) ≤ 1/(2L), but the proof absorbs the disagreement term 6LΣη_kD_{k-1} only under the stronger condition T(τ_i)-T(τ_{i-1}) < 1/[2L(v*+Δ²)], together with a small-stepsize condition η_t ≤ 1/(24ec_1(L+ℓ²)). Lemma 1 provides only v* ≤ 4Δ², and with the natural spacing Δ=2/(q-1) one has v*+Δ² ≤ 5Δ² = 20/(q-1)², which exceeds 1 for q=4,5. Hence (9b) does not imply the proof's condition for general q≥4, and the stated bound in Theorem 1 is not established as written. Please state the channel-dependent synchronization condition in the theorem, or restrict the quantization resolution so that v*+Δ² ≤ 1, and reconcile the stepsize constants c0 and 1/(24ec_1). Separately, the proof concludes the bound at synchronization times τ_r, while the theorem statement promises the bound for any n≥1; a final partial-block argument for n between synchronization times should be supplied or explicitly referenced.","section":"§4.2 and §6.2, Eq. (9b) and Eq. (13)"},{"comment":"The low-SNR simulation uses σ_c = 0.2 and q = 8, which with Δ = 2/(q-1) = 2/7 gives σ_c ≈ 0.2 > Δ/2 ≈ 0.1429, so the feasibility guarantee of Lemma 1 does not cover this operating point. Since the 'Ours' and 'Postcode' schemes require an optimal solution of LP (6), the paper should either report numerically verified feasibility of (6) for this channel (for example, the optimal value and constraint slacks), or provide a feasibility condition that covers the low-SNR experiment. Without this, the empirical claim in the low-SNR regime is not fully supported.","section":"§5, low-SNR experiment"},{"comment":"The statement of Theorem 2 imposes the stepsize condition η_k ≤ c0/(ℓ²+L), but under Assumption 3 the quantity ℓ is not defined and the proof in §6.3 uses η_k ≤ c0/[L(1+λ)]. Please replace the displayed condition with the condition actually used in the proof, or define ℓ consistently; as stated, the non-convex result cannot be checked because the stepsize condition is ill-posed.","section":"§4.3, Theorem 2"}],"minor_comments":[{"comment":"The definition of βω(x) is undefined for x = 0, since log2(ω^{-1}|x|) = -∞; please specify βω(0) = 0 or otherwise define the boundary case.","section":"§3.2, Eq. (7a)"},{"comment":"The condition (9b) uses T(τ_0), but T(0) is never defined; please state explicitly that T(0) = 0.","section":"§4.2, Eq. (9b)"},{"comment":"The text says the algorithms provide 'last-iterate and average-iterate guarantees', but only last-iterate results are stated and proved; please either add the average-iterate statement or rephrase the sentence.","section":"§3.3"},{"comment":"The same symbol H is used both for the stochastic post-coding mapping in (5) and for the matrix variable in LP (6); using a different symbol for the matrix (for example W) would avoid confusion in the proof of Lemma 1.","section":"§3.1 and §6.1.1"},{"comment":"The proof introduces an additional stepsize condition η_t ≤ 1/(24ec_1(L+ℓ²)) that is not listed in (9a); please make the relationship between this constant and c0 explicit so that the theorem statement and proof are consistent.","section":"§6.2, after Eq. (13)"},{"comment":"There is a typo 'syncrhonization', and the sentence 'the test accuracy of our method and the coded transmission results differs by only 0.07%' should be reworded for grammatical agreement.","section":"§5"},{"comment":"The notation eO(·) is used without definition; please define it or replace it with standard big-O notation with explicit logarithmic factors.","section":"§4.2 and §4.3"}],"recommendation":"major_revision","confidential_remarks":"The synchronization-condition mismatch in Theorem 1 is the most important issue: it is a proof-to-statement gap in the central strongly convex guarantee. The low-SNR simulation feasibility issue is checkable and should be easy to resolve by reporting the LP solver output. The paper is otherwise promising and the fixes appear local rather than requiring a new analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read on arXiv:2509.02538. The genuinely new piece is the post-coding matrix H: solve an LP to make the quantized-over-AWGN channel unbiased on interior quantization levels, then feed that through a scale-adaptive transformation and sync periodically. Lemma 2, the variance bound for this debiased oracle, is derived carefully and is the right kind of object for the convergence proof. The paper also makes an honest claim to the first exact unbiased gradient oracle under ADC/DAC quantization, and the convergence analysis is mostly conventional but competently executed. The 5x communication saving in the MNIST experiments is plausible and would be useful for wireless edge systems.\n\nThe soft spots are concrete. First, Theorem 1 is stated with synchronization condition (9b), T(τ_i)-T(τ_{i-1}) ≤ 1/(2L), but the proof in Section 6.2 actually needs T(τ_i)-T(τ_{i-1}) < 1/[2L(v*+Δ²)] to absorb the disagreement terms D_k. Since Lemma 1 only gives v* ≤ 4Δ², for coarse quantizers (q=4 or 5) v*+Δ² can exceed 1, so (9b) is not enough. The theorem as written is not proved for those cases. This is fixable by adding the channel-dependent condition or restricting the theorem to fine enough quantization, but it is a real correctness gap in the central result. Second, the low-SNR simulation (σc=0.2, q=8) lies outside the regime σc≤Δ/2 where Lemma 1 proves the LP is feasible, so the experiments don't validate the theory in that regime. It is possible the LP is feasible there anyway, but the paper doesn't say so. Third, the empirical section has no code, no error bars, and doesn't report key hyperparameters (ω, step size schedule, synchronization frequency beyond \"100\"), which makes the strong empirical claim hard to judge.\n\nNone of this is fatal. The construction is interesting, the lemmas are mostly clean, and the citation pattern looks fair. But the central theorem needs a corrected statement or an extra assumption, and the experiments need to either stay inside the proved regime or be accompanied by a feasibility check. This is a paper for people working on over-the-air and quantized federated learning. I'd send it to peer review with a request for major revision; it deserves a serious referee.","headline":"A genuinely new debiasing construction for over-the-air federated learning with a clean variance lemma, but Theorem 1 as stated needs a stronger synchronization condition than (9b), and the low-SNR experiments sit outside the proved feasibility region.","tokens_in":21776,"tokens_out":4655,"would_cite":true,"duration_ms":39218,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C15","68T05","68W40"],"pacs":[],"model":"deepseek-v4-flash","headline":"Over-the-air federated learning can match the accuracy of coded transmission.","keywords":["over-the-air computation","federated learning","stochastic gradient descent","quantization","unbiased stochastic oracle","adaptive algorithm","AWGN channel","communication-efficient learning"],"falsifier":"Solve the post-coding linear program for the low-SNR simulation parameters $\\sigma_c=0.2$, $\\Delta\\approx 0.286$, $q=8$; if it is infeasible, or if a direct measurement of the composed channel on a fixed vector shows the received expectation differs from the transmitted value, the central unbiasedness claim fails in that regime.","tokens_in":20786,"feed_emoji":"📡","tokens_out":10557,"duration_ms":81351,"temperature":0.7,"pith_summary":"Federated learning normally assumes error-free coded communication, which is expensive. This paper asks whether workers can instead send gradient information over physical over-the-air channels, with quantization, analog-to-digital conversion, and Gaussian noise, without losing the convergence guarantees of centralized SGD. The authors construct a stochastic post-coding scheme that makes the noisy quantized channel an unbiased estimator of the transmitted gradient, and they analyze adaptive federated SGD built on that oracle. They prove the statistical risk matches centralized SGD up to a channel-dependent term that can be made small by tuning a scale parameter at the cost of more communication. In image-classification simulations, the full method reaches the same test accuracy as coded transmission while sending about one fifth of the symbols.","feed_headline":"Over-the-air federated learning matches coded-channel accuracy","feed_subtitle":"Post-coding removes quantization bias; the paper's tests use a fifth of the symbols at equal accuracy.","key_machinery":"Three components carry the argument. First, a stochastic post-coding matrix $H$: the server solves a linear program over transition matrices so that $H\\circ Q_C\\circ C$ is unbiased on interior quantization levels and has worst-case variance $v^*$, and combining it with dithered randomized quantization $Q_D$ makes the whole channel unbiased. Second, a scale-adaptive transformation $(\\beta_\\omega,\\Psi_\\omega,A_\\omega)$ that rescales each coordinate onto a dyadic grid before transmission, keeping signals inside the range where post-coding is unbiased and making the variance depend on $\\|u\\|^2$ rather than on the signal range. Third, a periodic synchronization schedule: the server broadcasts the global parameter over the coded channel at times $\\tau_i$ spaced so that $T(\\tau_i)-T(\\tau_{i-1})\\le 1/(2L)$, which controls the disagreement $D_k$ between worker and server models. Lemma 2 is the hinge: it turns the channel into an unbiased oracle satisfying the moment condition needed for stochastic first-order methods.","core_discovery":"The central claim is that the composition of the over-the-air channel, the ADC/DAC quantization, and a carefully chosen random post-processing matrix can be made an unbiased stochastic gradient oracle, with mean squared error bounded by $(4v^*+\\Delta^2)(4\\|u\\|^2+\\omega^2 d)$ for a transmitted vector $u$. With this oracle, adaptive over-the-air SGD, using coded channels only for occasional synchronization of the global model, achieves a last-iterate risk bound of the form $e^{-\\mu T(n)/2}\\|\\theta_0-\\theta_*\\|^2 + (c\\eta_n/\\mu)(\\sigma_*^2/m + (v^*+\\Delta^2)\\omega^2 d)$ under strong convexity, and a matching gradient-norm bound in the non-convex setting. These bounds coincide with the centralized SGD rate up to the extra channel term $(v^*+\\Delta^2)\\omega^2 d$, which can be made negligible by choosing the scale parameter $\\omega$ small at the price of more communication. The paper claims this is the first exact unbiased gradient oracle under quantization and hardware noise.","pith_inferences":["Editorial extension: the proof of Theorem 1 in Section 6.2 invokes the stronger synchronization spacing $T(\\tau_i)-T(\\tau_{i-1}) < 1/[2L(v^*+\\Delta^2)]$ to absorb the worker-model disagreement terms $D_k$, not merely the stated $1/(2L)$; whether Theorem 1 holds exactly as stated depends on closing that gap.","Editorial extension: the low-SNR simulation uses $\\sigma_c=0.2$ with $\\Delta\\approx 0.286$, which falls outside the $\\sigma_c\\le\\Delta/2$ regime where Lemma 1 proves the post-coding linear program is feasible, so the simulation implicitly assumes feasibility beyond the proven range; a direct feasibility check of the linear program at those parameters would settle it.","Editorial extension: a natural testable prediction is that the same post-coding construction remains unbiased for sub-Gaussian or heavy-tailed channel noise with a modified variance bound, since the argument only uses symmetry and concentration of the noise through the transition matrix $P$."],"forward_implications":["If the unbiased-oracle analysis is correct, over-the-air transmission can replace coded channels for gradient exchange in federated learning without changing the order of the convergence rate.","Choosing $\\omega$ small drives the extra statistical error $(v^*+\\Delta^2)\\omega^2 d$ toward zero, so the method can approach the centralized minimax rate $\\sigma_*^2/(\\mu m)$ while still using physical channels for the bulk of the traffic.","The synchronization condition is mild with decreasing stepsizes: it permits geometrically spaced synchronization rounds, and with constant stepsize it requires synchronization every $O(1/(L\\eta))$ steps.","The non-convex guarantee gives the standard $O(\\epsilon^{-4})$ sample complexity for finding an $\\epsilon$-stationary point, with only $O(\\sqrt{n})$ broadcasting steps over $n$ iterations under the $\\eta_k\\asymp 1/\\sqrt{n}$ schedule.","In the paper's MNIST experiments, the full method matches coded transmission in test accuracy, differing by 0.07%, while transmitting about one fifth of the symbols."],"supporting_citations":[{"why":"Supplies the over-the-air SGD baseline that this paper extends to quantized hardware and unbiased post-coding.","marker":"[AG20b]"},{"why":"Establishes the over-the-air computation model for federated learning that the transmission scheme builds on.","marker":"[YJSD20]"},{"why":"Prior noisy-channel federated learning analysis whose requirement of fast noise decay the paper improves on.","marker":"[WS22]"},{"why":"Provides the randomized quantization idea used in the scale-adaptive transmission.","marker":"[AGKP20]"},{"why":"Supplies the dithering theory behind the stochastic post-coding construction.","marker":"[WLVW02]"},{"why":"Frames the state-dependent noise condition used in Assumption 2.","marker":"[MB11]"},{"why":"Supplies the state-dependent noise analysis tools used to prove the risk bounds.","marker":"[IJLL24]"},{"why":"Shows the benefit of periodic synchronization in distributed optimization, which the algorithm uses to limit coded communication.","marker":"[Sti18]"}],"fun_headline_variants":["First unbiased over-the-air gradient oracle for federated SGD","Adaptive over-the-air SGD achieves centralized rates","Near-optimal federated learning over physical channels","Unbiased gradients over the air: near-optimal federated SGD","Over-the-air federated learning with near-optimal adaptive SGD"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole argument rests on the post-coding linear program being solvable for the channel actually used, and the paper proves solvability only when the channel noise is at most half the quantization step.","fun_headline_variants_meta":{"raw":{"variants":["First unbiased over-the-air gradient oracle for federated SGD","Adaptive over-the-air SGD achieves centralized rates","Near-optimal federated learning over physical channels","Unbiased gradients over the air: near-optimal federated SGD","Over-the-air federated learning with near-optimal adaptive SGD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000613,"raw_usage":{"total_tokens":2799,"prompt_tokens":845,"completion_tokens":1954,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":461,"completion_tokens_details":{"reasoning_tokens":1874}},"tokens_in":461,"tokens_out":1954,"duration_ms":13535,"temperature":1.0,"reasoning_tokens":1874,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:37:21.569309+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Solve the post-coding linear program for the low-SNR simulation parameters $\\sigma_c=0.2$, $\\Delta\\approx 0.286$, $q=8$; if it is infeasible, or if a direct measurement of the composed channel on a fixed vector shows the received expectation differs from the transmitted value, the central unbiasedness claim fails in that regime.","supporting_citations":[],"review_version":2}