{"id":"4df913e5-b942-4930-bbd9-7db095b3cb98","arxiv_id":"2501.03392","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"OTA-FFL uses an epsilon-constrained Chebyshev objective to set adaptive client weights and closed-form transmit scalars for over-the-air federated learning.","lead":"This paper proposes OTA-FFL, a wireless federated learning method that combines a fairness-aware Chebyshev weighting rule with over-the-air gradient aggregation. The authors report that it produces fairer accuracy across clients than existing over-the-air baselines on four image datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Under the heterogeneous-data setting that motivates OTA-FFL, the normalized gradient vector in §V.B has variance (v_t,k+(m_t,k−m_t)^2)/v_t, not 1; Lemma 2's power constraint (13) is therefore violated and the claimed optimal scalars are not justified.","rationale":"I read the paper as trying to show that fairness via a modified Chebyshev scheme can be implemented over a MAC with optimal linear encoders. The load-bearing piece is Lemma 2. The proof in the appendix rests entirely on E[s s^H]=I_d to replace the power constraint by |b|^2≤P0. Under the paper's own heterogeneous data premise, that moment identity fails. This is not a matter of differing from consensus; it is an internal mismatch between the stated client statistics and the normalization used in the proof. The fairness loop itself—modified Chebyshev with λ constrained to an l∞ ball—is plausible and could be salvaged by defining normalized symbols with per-client means and variances, or by incorporating r_t,k into the power constraint. The experiments are suggestive but are averaged over only five seeds, and Table I contains an impossible Fashion MNIST entry (worst 10% 76.28 > best 10% 76.15), which further limits confidence. Because the central OTA claim is unsupported, I agree with the reader's rejection.","tokens_in":9078,"tokens_out":4648,"duration_ms":44623,"concrete_test":"Run the two-client numerical check: K=2, d=1, P0=1, h1=h2=1, lambda1=lambda2=0.5, client gradients with mean 1 and −1 and variance 1 each. Then v_t=1, m_t=0, and r_k=2. Lemma 2 gives c_t=2 and b_t,k=1, but E|b_t,k s_t,k|^2 = 1^2·(1+(±1)^2)/1 = 2 > P0, violating the paper's own constraint (13). Recompute the variance-minimizing c_t under |b|^2 r_k≤P0; if it is sqrt(2) rather than 2, the claimed Lemma 2 optimum and E* are off by the factor max r_k.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In §V.B, after defining s_t,k=(g_t,k−m_t 1_d)/sqrt(v_t), the paper asserts E[s_t,k s_t,k^H]=I_d for each client. This would require every gradient entry of client k to have mean m_t and variance v_t. But the paper's heterogeneous setting gives client k its own mean m_t,k and variance v_t,k; m_t and v_t are only the weighted averages from (12a). Hence E|s_t,k[i]|^2=(v_t,k+(m_t,k−m_t)^2)/v_t. Unless the per-client statistics happen to coincide, this is not 1. The transmit-power bound (13) is then |b_t,k|^2 r_t,k ≤ P0 with r_t,k>1 for off-mean clients, not |b_t,k|^2≤P0. Lemma 2's c_t=min_k sqrt(P0)|h_t,k|/lambda_t,k ignores these r_t,k, so the resulting b_t,k can exceed the per-symbol power limit. The correct feasibility bound is c_t≤min_k sqrt(P0)|h_t,k|/(lambda_t,k sqrt(r_t,k)), and the minimized variance would include max_k lambda^2 r/|h|^2. The central OTA design—equations (18)–(19)—is thus not established for the heterogeneous regime the algorithm targets.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes OTA-FFL, a federated learning algorithm that combines a multi-objective formulation with over-the-air (OTA) computation. The PS computes adaptive client weights by solving a modified Chebyshev scalarization with a constraint that keeps the weights close to the FedAvg proportions. Clients then transmit normalized local gradients over a fading Gaussian MAC, and the PS applies a linear decoder. The paper derives closed-form transmit scalars and a denoising scalar that are claimed to yield an unbiased estimate of the weighted gradient sum with minimum variance. Experiments on CIFAR-10, CINIC-10, FEMNIST, and Fashion MNIST compare the proposed method against OTA-FedAvg, OTA-TERM, and OTA-q-FFL, reporting improved fairness metrics.","tokens_in":9288,"tokens_out":6590,"duration_ms":57817,"significance":"The paper addresses a timely problem: integrating fairness into wireless federated learning under the communication-efficiency constraint of OTA computation. The modified Chebyshev weighting is a reasonable mechanism for trading off average performance and fairness, and the paper provides code and extensive comparisons on four datasets. The weighting coefficients are computed from current losses and used in the same round, so the approach is not circular. However, the central analytical contribution, Lemma 2, relies on a statistical assumption about normalized gradients that is false in the heterogeneous setting the paper targets. Since the optimality of the transmit and denoising scalars is the main technical novelty, the significance of the paper is not established without correcting this derivation.","major_comments":[{"comment":"The claim that the normalized gradient vector s_{t,k} = (g_{t,k} - m_t 1_d)/sqrt(v_t) satisfies E[s_{t,k} s_{t,k}^H] = I_d is not true under heterogeneous per-client statistics. For client k, each entry of g_{t,k} has mean m_{t,k} and variance v_{t,k}, whereas m_t and v_t are weighted averages over the selected clients. The per-entry second moment of s_{t,k} is r_{t,k} := (v_{t,k} + (m_{t,k} - m_t)^2)/v_t, which equals 1 only when m_{t,k}=m_t and v_{t,k}=v_t. Consequently, the transmit power constraint (13) should be |b_{t,k}|^2 r_{t,k} <= P0, not |b_{t,k}|^2 <= P0. The feasibility condition in the appendix, |c_t|^2 <= P0 |h_{t,k}|^2 / (lambda*_{t,k})^2, is therefore invalid; the correct bound is |c_t|^2 <= P0 |h_{t,k}|^2 / (lambda*_{t,k}^2 r_{t,k}). The optimal scalars in (18) and the variance expression in (19) are not established, and the minimized variance would generally involve max_k lambda*_{t,k}^2 r_{t,k}/|h_{t,k}|^2. This is a load-bearing error because the analytical OTA encoding-decoding solution is the paper's core contribution.","section":"Section V.B, Eq. (12a)-(14), and Appendix (Lemma 2)"}],"minor_comments":[{"comment":"There is a typo: 'can leas to' should be 'can lead to'.","section":"Section IV.A"},{"comment":"The phrase \"proof is differed to the Appendix\" should read \"proof is deferred to the Appendix\".","section":"Section V and Appendix"},{"comment":"In the Fashion MNIST columns, the reported Worst 10% and Best 10% for OTA-FFL are 76.28 and 76.15, respectively, which would imply the worst decile is better than the best decile; the numbers appear to be inconsistent and should be checked.","section":"Table I"},{"comment":"The sentence \"We further set batch-size is set to 64\" contains a redundant verb and should be rephrased.","section":"Section VI.A"},{"comment":"The fairness metric in Definition 3 is defined through the standard deviation of test losses, but Table I reports standard deviations of accuracies; the paper should clarify that lower standard deviation of accuracy is used as the operational fairness measure and relate it to the definition.","section":"Definition 3 and Table I"},{"comment":"The hyperparameters epsilon (the Chebyshev constraint radius) and zeta (the reference loss vector) are not specified for the experiments, and no sensitivity analysis is provided; reporting chosen values or a sensitivity study would improve reproducibility.","section":"Section V and VI"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about Section V.B is accurate and central: the unit-covariance assumption on normalized gradients is violated by the paper's own heterogeneous-data setup, which invalidates the power-constraint derivation and Lemma 2. The error appears fixable within the manuscript's scope by re-deriving the OTA scalars with per-client second moments, but this will change the analytical results and likely the experimental implementation. I therefore recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the fairness idea is real, but the wireless half doesn't hold together as written. The epsilon-constrained Chebyshev weighting is a clean way to interpolate between FedAvg and AFL, and putting it together with OTA gradient aggregation appears new as far as the citations show. The authors also ship code and run four benchmark datasets. That is the good part.\n\nThe problem is Section V.B. After defining s_t,k = (g_t,k - m_t 1_d) / sqrt(v_t), the paper claims E[s_t,k s_t,k^H] = I_d and cross-client covariance zero. That only holds if every client's gradient entries have the same mean m_t and variance v_t. But the paper's own heterogeneous setting gives client k its own m_t,k and v_t,k; m_t and v_t are just weighted averages. The actual per-entry variance is (v_t,k + (m_t,k - m_t)^2) / v_t, so the power bound in (13) is |b_t,k|^2 r_t,k <= P0, not |b_t,k|^2 <= P0. Lemma 2 and the variance expression in (19) ignore r_t,k, so the scalars it outputs can exceed the per-symbol power limit for off-mean clients. This is a load-bearing flaw, not a typo.\n\nThere are smaller issues. Table I has an impossible entry: on Fashion MNIST the worst 10% for OTA-FFL (76.28) is above the best 10% (76.15), which suggests the columns were swapped or the numbers were misreported. Epsilon and zeta are never given, and there are no seed-to-seed error bars. The alternating inner/outer loop has no convergence analysis; that might be acceptable for a letter, but it is worth stating. The citation pattern is mostly a working group citing each other's prior OTA-FL work, which is normal for this subfield, not a red flag.\n\nMy overall take: the modified Chebyshev fairness mechanism is plausible and the empirical trend may survive, but the paper's main claim—optimal OTA scalars for fair aggregation—rests on a false statistical assumption. A repair is conceivable: redo Lemma 2 with r_t,k and adjust the power feasibility bound. But as submitted, the wireless contribution is not sound.\n\nWho is this for: people working on fair FL over wireless might learn the epsilon-constrained Chebyshev idea and the OTA problem setup, but they should not trust equations (18)–(19). I would still send it to a serious referee: the idea is clear, the flaw is specific and checkable, and the authors can likely fix it. I would not cite it in its current form.","headline":"The epsilon-constrained Chebyshev fairness idea is a genuine and clearly stated contribution, but the over-the-air half rests on a false covariance assumption and the central Lemma 2 does not hold as written.","tokens_in":9907,"tokens_out":4157,"would_cite":false,"duration_ms":38565,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T05","90C29"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes OTA-FFL, a fair federated learning algorithm that uses over-the-air computation, and derives closed-form transmit scalars that make the server's weighted gradient estimate unbiased with minimum variance.","keywords":["over-the-air computation","federated learning","fairness","multi-objective optimization","Chebyshev method","weighted gradient aggregation","wireless multiple-access channel","adaptive weighting"],"falsifier":"On a heterogeneous data split, compute the actual per-symbol power of the transmitted signal $\\mathbb{E}|b_{t,k}s_{t,k}[i]|^2$. If $\\mathbb{E}|s_{t,k}[i]|^2 = (v_{t,k}+(m_{t,k}-m_t)^2)/v_t \\neq 1$, then the choice $b_{t,k}=\\lambda^\\star_{t,k}c_t/h_{t,k}$ can exceed the stated power constraint, and the variance formula $\\mathbb{E}^\\star = d v_t \\sigma^2/P_0 \\max_k (\\lambda^\\star_{t,k})^2/|h_{t,k}|^2$ would not match measurements.","tokens_in":8788,"feed_emoji":"📡","tokens_out":7247,"duration_ms":67331,"temperature":0.7,"pith_summary":"The paper proposes OTA-FFL, a federated learning algorithm designed to train a fair shared model when clients communicate over a wireless multiple-access channel using over-the-air computation. It treats fairness as a multi-objective optimization problem and resolves it with a modified Chebyshev method that computes adaptive aggregation weights each round, interpolating between FedAvg and agnostic FL. The central technical claim is that this is the first fair FL scheme compatible with over-the-air aggregation: the paper derives closed-form transmit scalars at the clients and a denoising scalar at the server that make the weighted gradient estimate unbiased with minimum variance. If the derivation holds, fairness can be achieved without giving up the communication efficiency of over-the-air computation.","feed_headline":"Fair federated learning can run over-the-air with adaptive weights","feed_subtitle":"Adaptive Chebyshev weights and optimized transmit scalars shrink accuracy gaps across clients while keeping OTA aggregation.","key_machinery":"The load-bearing mechanism is the combination of a modified Chebyshev objective and an optimal OTA encoder-decoder pair. The modified Chebyshev maximization constrains the weight vector to stay within an $\\epsilon$-ball of the FedAvg weights, so the aggregation adapts to underperforming clients while retaining average-performance behavior. Lemma 2 then supplies the transmit scalars $b_{t,k}$ and receive scalar $c_t$ that make the server's estimate of the weighted gradient sum unbiased and minimize its variance under the per-symbol power constraint.","core_discovery":"The paper claims that fairness in federated learning can be combined with over-the-air aggregation in a single algorithm. OTA-FFL treats each client's local loss as one objective in a multi-objective minimization and, each round, solves a constrained maximization to obtain weights $\\lambda^\\star_t$ close to the data-proportional FedAvg weights but free to shift toward the worst-off client. These weights enter a gradient aggregation whose transmit scalars are chosen as $b_{t,k} = \\lambda^\\star_{t,k} c_t / h_{t,k}$ and whose server denoising scalar is $c_t = \\min_{k\\in S_t} \\sqrt{P_0}|h_{t,k}|/\\lambda^\\star_{t,k}$, giving an estimation variance of $d v_t \\sigma^2 / P_0 \\cdot \\max_{k\\in S_t} (\\lambda^\\star_{t,k})^2/|h_{t,k}|^2$. In experiments on four datasets, the paper reports that OTA-FFL reduces the spread of per-client accuracies compared with OTA-FedAvg, OTA-TERM, and OTA-q-FFL.","pith_inferences":["The derivation of the optimal scalars assumes that after normalization each client's gradient entries have the same mean and variance; with heterogeneous data, $\\mathbb{E}|s_{t,k}[i]|^2 = (v_{t,k}+(m_{t,k}-m_t)^2)/v_t$ is generally not 1, so the power constraint used in Lemma 2 may not hold exactly as stated. A power-aware re-derivation or client-specific normalization would close this gap.","Because the fair weights depend only on scalar losses, the same modified Chebyshev mechanism could be coupled with compressed or one-bit over-the-air schemes, not only the linear analog encoders used here.","If the fairness metric were changed from the standard deviation of accuracies to a worst-case quantile, the inner optimization would change, but Lemma 2's encoder-decoder design would carry over unchanged since it depends only on the resulting weights."],"forward_implications":["Fairness no longer forces the server to collect each client's gradient vector separately; the weighted sum can be recovered directly from the superimposed wireless signal.","The $\\epsilon$ parameter in the modified Chebyshev problem creates a tunable trade-off between average accuracy and fairness, recovering FedAvg at $\\epsilon=0$ and agnostic FL at $\\epsilon=1$.","Per round, each client only needs to report its scalar loss value and its gradient mean and variance, so the fairness mechanism adds negligible communication overhead to OTA-FL.","The closed-form variance expression shows how channel gains and fairness weights jointly determine aggregation error, which can guide client scheduling decisions.","If the derivation is correct, existing OTA-FL systems can be upgraded to fair training by replacing the fixed aggregation weights with the adaptively computed Chebyshev weights and the optimized scalars."],"supporting_citations":[{"why":"defines FedAvg and the data-proportional weights that OTA-FFL modifies as its baseline.","marker":"[1]"},{"why":"sets up the over-the-air computation model of transmit scalars and denoising scalar over the MAC.","marker":"[4]"},{"why":"supplies the Gibbs-sampling client scheduler used in the OTA-FFL implementation.","marker":"[5]"},{"why":"introduces the q-FFL baseline and the standard-deviation fairness metric used in evaluation.","marker":"[8]"},{"why":"introduces the TERM baseline whose exponential losses are adapted for OTA comparison.","marker":"[9]"},{"why":"is the prior wireless fair-FL scheme whose need for individual gradients motivates OTA compatibility.","marker":"[11]"},{"why":"frames federated learning as multi-objective minimization, the starting point of the paper's formulation.","marker":"[12]"},{"why":"introduces agnostic FL and the Chebyshev minimax method that the modified Chebyshev scheme adapts.","marker":"[17]"}],"fun_headline_variants":["OTA-FFL: Fair federated learning via multi-objective optimization","Over-the-air fairness: adaptive weights for federated learning","Fair FL meets OTA: multi-objective optimization for adaptive aggregation","Shrinking accuracy gaps: OTA-FFL's adaptive Chebyshev weights"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The optimal transmit-scalar derivation assumes that after normalization every client's gradient entries have the same mean and variance, even though heterogeneous data means those statistics differ from client to client.","fun_headline_variants_meta":{"raw":{"variants":["OTA-FFL: Fair federated learning via multi-objective optimization","Over-the-air fairness: adaptive weights for federated learning","Fair FL meets OTA: multi-objective optimization for adaptive aggregation","Shrinking accuracy gaps: OTA-FFL's adaptive Chebyshev weights"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000634,"raw_usage":{"total_tokens":2906,"prompt_tokens":909,"completion_tokens":1997,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":1919}},"tokens_in":525,"tokens_out":1997,"duration_ms":14601,"temperature":1.0,"reasoning_tokens":1919,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:52:46.071167+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a heterogeneous data split, compute the actual per-symbol power of the transmitted signal $\\mathbb{E}|b_{t,k}s_{t,k}[i]|^2$. If $\\mathbb{E}|s_{t,k}[i]|^2 = (v_{t,k}+(m_{t,k}-m_t)^2)/v_t \\neq 1$, then the choice $b_{t,k}=\\lambda^\\star_{t,k}c_t/h_{t,k}$ can exceed the stated power constraint, and the variance formula $\\mathbb{E}^\\star = d v_t \\sigma^2/P_0 \\max_k (\\lambda^\\star_{t,k})^2/|h_{t,k}|^2$ would not match measurements.","supporting_citations":[{"cited_title":"Communication-efficient learning of networks from decentralized data,","cited_arxiv_id":null,"evidence_quote":"defines FedAvg and the data-proportional weights that OTA-FFL modifies as its baseline."},{"cited_title":"Federated learning via over-the-air computation,","cited_arxiv_id":null,"evidence_quote":"sets up the over-the-air computation model of transmit scalars and denoising scalar over the MAC."},{"cited_title":"Over-the-air federated learning via second-order optimization,","cited_arxiv_id":null,"evidence_quote":"supplies the Gibbs-sampling client scheduler used in the OTA-FFL implementation."},{"cited_title":"Fair resource allocation in federated learning,","cited_arxiv_id":null,"evidence_quote":"introduces the q-FFL baseline and the standard-deviation fairness metric used in evaluation."},{"cited_title":"Tilted empirical risk minimization,","cited_arxiv_id":null,"evidence_quote":"introduces the TERM baseline whose exponential losses are adapted for OTA comparison."},{"cited_title":"Fair wireless federated learning through the identification of a common descent direction,","cited_arxiv_id":null,"evidence_quote":"is the prior wireless fair-FL scheme whose need for individual gradients motivates OTA compatibility."},{"cited_title":"Federated learning meets multi-objective optimization,","cited_arxiv_id":null,"evidence_quote":"frames federated learning as multi-objective minimization, the starting point of the paper's formulation."},{"cited_title":"Agnostic federated learning,","cited_arxiv_id":null,"evidence_quote":"introduces agnostic FL and the Chebyshev minimax method that the modified Chebyshev scheme adapts."}],"review_version":1}