REVIEW 1 major objections 6 minor 2 cited by
Over-the-Air Fair Federated Learning via Multi-Objective Optimization
T0 review · 1 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read 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.
desk verdict 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. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
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.
What would settle it
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.
Extended reading notes
Core claim
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.
Load-bearing premise
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.
Editorial extensions
If this is right
- 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.
Reading between the lines
- 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.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
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.
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 (1)
- [Section V.B, Eq. (12a)-(14), and Appendix (Lemma 2)] 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.
minor comments (6)
- [Section IV.A] There is a typo: 'can leas to' should be 'can lead to'.
- [Section V and Appendix] The phrase "proof is differed to the Appendix" should read "proof is deferred to the Appendix".
- [Table I] 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 VI.A] The sentence "We further set batch-size is set to 64" contains a redundant verb and should be rephrased.
- [Definition 3 and Table I] 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 V and VI] 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.
Circularity Check
No circular derivation: OTA-FFL's transceiver design is derived from stated distortion and power constraints; only a minor non-load-bearing self-citation appears.
full rationale
The derivation chain of OTA-FFL is not circular. The adaptive Chebyshev weights are computed from current local losses and then used in the same round to define the weighted gradient aggregation target; this is a standard adaptive optimization loop, not a parameter fitted to data and then renamed as a prediction. Lemma 2 is derived in the appendix from the unbiasedness condition (16), the variance expression (17), and the transmit power constraint (13); although the power constraint itself rests on the unproven claim E[s_t,k s_t,k^H]=I_d under heterogeneous client statistics, that is a correctness/soundness gap, not a circular reduction. The per-round weight computation does not presuppose the fairness outcomes that are later reported. The only self-citation that could be viewed as load-bearing is the statement that the FL Pareto front is non-convex, attributed to [13] by the same first author, but the modified Chebyshev objective and the OTA encoder/decoder scalars do not reduce to that citation's content: the scalars are optimized from first-order conditions over the paper's own model. The experimental benchmarks are external (OTA-FedAvg, OTA-TERM, OTA-q-FFL), so the reported superiority is not implied by construction. Thus the paper contains no circular step; the score reflects one minor, non-load-bearing self-citation.
Assumptions & free parameters
free parameters (2)
- epsilon (Chebyshev constraint radius) =
Not reported
- zeta (reference loss lower-bound vector) =
Not reported
assumptions (4)
- ad hoc to paper Normalized gradients have zero mean and identity covariance
- domain assumption Local objectives are differentiable and gradient statistics exist
- standard math Mukai's lemma connecting Pareto-stationarity and Pareto-optimality
- domain assumption Channel is constant over a FL iteration and noise is complex Gaussian with variance sigma^2
Cite this review
Pith. "Pith review of Over-the-Air Fair Federated Learning via Multi-Objective Optimization." pith.science (2026). https://pith.science/paper/NGTW23ZK
@misc{pith2026250103392,
author = {Pith},
title = {Pith review of: Over-the-Air Fair Federated Learning via Multi-Objective Optimization},
year = {2026},
howpublished = {\url{https://pith.science/paper/NGTW23ZK}},
note = {Machine review of arXiv:2501.03392}
}
read the original abstract
In federated learning (FL), heterogeneity among the local dataset distributions of clients can result in unsatisfactory performance for some, leading to an unfair model. To address this challenge, we propose an over-the-air fair federated learning algorithm (OTA-FFL), which leverages over-the-air computation to train fair FL models. By formulating FL as a multi-objective minimization problem, we introduce a modified Chebyshev approach to compute adaptive weighting coefficients for gradient aggregation in each communication round. To enable efficient aggregation over the multiple access channel, we derive analytical solutions for the optimal transmit scalars at the clients and the de-noising scalar at the parameter server. Extensive experiments demonstrate the superiority of OTA-FFL in achieving fairness and robust performance compared to existing methods.
Figures
Forward citations
Cited by 2 Pith papers
-
FAIRVAR: Fair Federated Learning via Variance Regularization
FairGrad and FairGrad* add gradient variance penalties to federated learning, and the authors report these reduce accuracy variance across clients while keeping or improving mean accuracy, though the exact version's g...
-
Distributed Quasi-Newton Method for Fair and Fast Federated Learning
DQN-Fed updates a global model in a direction that makes every client's loss decrease at a rate tied to its local quasi-Newton step, with claimed linear-quadratic convergence.
Reference graph
Works this paper leans on
-
[1]
Communication-efficient learning of networks from decentralized data,
B. McMahan, E. Moore et al. , “Communication-efficient learning of networks from decentralized data,” in Proc. 20th Int. Conf. Artificial Intelligence and Statistics . PMLR, 2017, pp. 1273–1282
work page 2017
-
[2]
Training neural networks on remote edge devices for unseen class classification,
S. M. Hamidi, “Training neural networks on remote edge devices for unseen class classification,” IEEE Signal Processing Letters , vol. 31, pp. 1004–1008, 2024
work page 2024
-
[3]
Machine learning at the wireless edge: Distributed stochastic gradient descent over-the-air,
M. M. Amiri and D. G ¨und¨uz, “Machine learning at the wireless edge: Distributed stochastic gradient descent over-the-air,” IEEE Trans. Signal Process., vol. 68, pp. 2155–2169, 2020
2020
-
[4]
Federated learning via over-the-air computation,
K. Yang, T. Jianget al., “Federated learning via over-the-air computation,” IEEE Trans. Wirel. Commun. , vol. 19, no. 3, pp. 2022–2035, 2020
work page 2022
-
[5]
Over-the-air federated learning via second-order optimization,
P. Yang, Y . Jianget al., “Over-the-air federated learning via second-order optimization,” IEEE Trans. Wirel. Commun. , vol. 21, no. 12, pp. 60–75, 2022
work page 2022
-
[6]
Over-the- air federated learning exploiting channel perturbation,
S. M. Hamidi, M. Mehrabi, A. K. Khandani, and D. G ¨und¨uz, “Over-the- air federated learning exploiting channel perturbation,” in 2022 IEEE 23rd International Workshop on Signal Processing Advances in Wireless Communication (SPAWC), 2022, pp. 1–5
work page 2022
-
[7]
Fed-it: Addressing class imbalance in federated learning through an information- theoretic lens,
S. M. Hamidi, R. Tan, L. Ye, and E.-H. Yang, “Fed-it: Addressing class imbalance in federated learning through an information- theoretic lens,” in 2024 IEEE International Symposium on Information Theory (ISIT) , 2024, pp. 1848–1853
work page 2024
-
[8]
Fair resource allocation in federated learning,
T. Li, M. Sanjabi et al., “Fair resource allocation in federated learning,” in Proc. Int. Conf. Learning Representations , 2019
work page 2019
Show all 20 references
-
[9]
Tilted empirical risk minimization,
T. Li, A. Beirami et al., “Tilted empirical risk minimization,” in Proc. Int. Conf. Learning Representations , 2020
2020
-
[10]
Distributed quasi-newton method for fair and fast federated learning,
L. Y . Shayan Mohajer Hamidi, “Distributed quasi-newton method for fair and fast federated learning,” Transactions on Machine Learning Research,
-
[11]
Fair wireless federated learning through the identification of a common descent direction,
S. Mohajer Hamidi and O. Damen, “Fair wireless federated learning through the identification of a common descent direction,” IEEE Commun. Lett., vol. 28, no. 3, pp. 567–571, 2024
2024
-
[12]
Federated learning meets multi-objective optimization,
Z. Hu, K. Shaloudegi et al., “Federated learning meets multi-objective optimization,” IEEE Trans. Netw. Sci. Eng. , 2022
2022
-
[13]
Adafed: Fair federated learning via adaptive common descent direction,
S. M. Hamidi and E.-H. YANG, “Adafed: Fair federated learning via adaptive common descent direction,” Trans. Mach. Learn. Res. , 2024
2024
-
[14]
Collette and P
Y . Collette and P. Siarry, Multiobjective optimization: Principles and Case Studies. Springer Science & Business Media, 2013
2013
-
[15]
Algorithms for multicriterion optimization,
H. Mukai, “Algorithms for multicriterion optimization,” IEEE Trans. Autom. Control, vol. 25, no. 2, pp. 177–186, 1980
1980
-
[16]
S. P. Boyd and L. Vandenberghe, Convex Optimization. Cambridge university press, 2004
2004
-
[17]
Agnostic federated learning,
M. Mohri, G. Sivek et al., “Agnostic federated learning,” in Proc. Int. conf. machine learning (ICML) . PMLR, 2019, pp. 4615–4625
2019
-
[18]
Multiple-gradient descent algorithm (MGDA),
J.-A. D ´esid´eri, “Multiple-gradient descent algorithm (MGDA),” Ph.D. dissertation, INRIA, 2009
2009
-
[19]
Device scheduling in over-the-air federated learning via matching pursuit,
A. Bereyhi, A. Vagollari et al. , “Device scheduling in over-the-air federated learning via matching pursuit,” IEEE Trans. Signal Process. , vol. 71, pp. 2188–2203, 2023
2023
-
[2024]
Available: https://openreview.net/forum?id=KbteA50cni
[Online]. Available: https://openreview.net/forum?id=KbteA50cni
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.