{"id":"e9c91f00-4aa8-4ded-97ea-f6f5c17e963c","arxiv_id":"2412.16779","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Fed-ZOE compresses over-the-air federated learning updates to a few thousand random projections, cutting communication overhead to about 0.07% of baseline while keeping similar accuracy.","lead":"This paper proposes Fed-ZOE, a way to compress the information that edge devices send to a central server during over-the-air federated learning. Instead of transmitting millions of model weights, each device sends a few thousand scalar projections, and the server reconstructs a useful gradient estimate.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 1 and Theorem 1 rest on an incorrect fourth-moment identity (Eq. 35), so the stated convergence guarantee is not proven as written.","rationale":"The reader's weakest_assumption correctly identifies Eq. (35) as the load-bearing flaw in the theoretical section. I verified the fourth-moment computation: for independent standard Gaussian entries, E[u_i u_s u_s u_j] summed over s equals (S+2) delta_ij, not 3S delta_ij. This invalidates Proposition 1's variance formula and the learning-rate condition in Theorem 1 as displayed. Because the paper explicitly claims theoretical guarantees as a contribution, an incorrect moment identity is a genuine correctness risk. However, the corrected constants are smaller, so the convergence rate and the communication-efficiency conclusions are not overturned; the theorem can likely be repaired with straightforward modifications. I also noticed a second, independent scaling error in the channel-noise term of Appendix B (a missing factor of L) and a suspicious extra K in the compensation term of Eq. (15), but neither changes the qualitative assessment. The empirical results are extensive, though they lack repeated runs and released code, which the reader already noted. Given that the qualitative claims appear robust and the identified proof errors are fixable, the conditional-accept verdict remains appropriate; no change is needed.","tokens_in":22182,"tokens_out":13059,"duration_ms":102905,"concrete_test":"Independently re-derive Eq. (35) for u ~ N(0, I_S): compute E[(u u^T)^2]_{ij} via Isserlis and confirm the result is (S+2) delta_ij. Then re-run Appendices B and C with this corrected fourth moment to obtain the compression-variance term (S+1)/(K^2 L) sum ||Delta w||^2 and the adjusted learning-rate condition eta_t <= 2/[beta(1+(S+1)I/(K^2 L))]. Also recompute part A with the noise-corrupted y_G having the same scalar r^H n in all L entries, yielding an added variance S N0 ||r||^2 / L rather than /L^2. If the resulting Theorem 1 bound still exhibits O(1/sqrt(T)) decay and the required L scales as before, the central claim stands after a proof revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The theoretical guarantee is load-bearing for the paper's central claim of providing convergence guarantees alongside communication reduction. Proposition 1 and Theorem 1 use Eq. (35), which asserts E[(u u^T)^2] = 3S I for 'spherically symmetric, unit-variance' u in R^S. For u ~ N(0, I_S), Isserlis' theorem gives E[(u u^T)^2]_{ij} = sum_s E[u_i u_s u_s u_j] = sum_s (2 delta_is delta_js + delta_ij delta_ss) = (S+2) delta_ij, not 3S delta_ij. A similar calculation for Rademacher or uniform-sphere vectors gives S I, again not 3S I. This error propagates into Proposition 1's compression-variance term (3S-1) in Eq. (22) and into the learning-rate condition and final bound in Theorem 1. As written, the proof of the stated convergence result is invalid. The corrected constant is smaller (S+1 for Gaussian, S-1 for Rademacher), so the qualitative O(1/sqrt(T)) rate and the empirical claims likely survive a revised proof. A separate issue: in Appendix B, part A treats the channel noise as entering through a single u_l, giving S N0 ||r||^2 / L^2, but the noise is added to all L entries of y_G, so the variance is S N0 ||r||^2 / L; this understates the noise term by a factor of L. These corrections change constants but not the algorithm's effectiveness, so the concern is with proof rigor, not the method's core idea.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Fed-ZOE, an over-the-air federated learning (OtA-FL) framework in which each edge device compresses its local model update vector by transmitting L scalar inner products with shared random vectors, rather than the full S-dimensional update. The parameter server reconstructs the aggregated update via randomized gradient estimation, preserving the superposition property required for over-the-air computation. The authors provide a convergence analysis (Proposition 1 and Theorem 1), a beamforming optimization for unbiased aggregation, and extensive ResNet-18 experiments on CIFAR-10, CIFAR-100, SVHN, Tiny-ImageNet, and Brain-CT. The empirical results show that Fed-ZOE approaches the accuracy of uncompressed OtA-FL while using roughly 8,192 scalars per round for a model with about 1.1×10^7 parameters, corresponding to a per-round compression ratio of about 0.07%.","tokens_in":22476,"tokens_out":12076,"duration_ms":101066,"significance":"If the theoretical guarantee is made rigorous, the paper offers a simple and practical communication-reduction technique for OtA-FL. The core idea of applying a randomized gradient estimator to compress the local update vector rather than using it to replace backpropagation is a useful twist, and the method is explicitly designed to preserve the analog superposition property. The empirical evaluation is broad and includes several baselines (OtA-FL, LoRA-OtA-FL, ZO-OtA-FL) under non-IID data and varying numbers of devices. A notable strength is that the predicted dependence on the number of samples L (more samples improve convergence, and higher learning rates require larger L) is tested out-of-sample in Figures 3 and 5, giving a non-circular check of the core mechanism. However, several technical errors in the proof of Proposition 1 and Theorem 1 must be corrected before the convergence guarantee can be accepted as stated.","major_comments":[{"comment":"The claimed fourth-moment identity is incorrect. For u ~ N(0, I_S), Isserlis' theorem gives E[(u u^T)^2]_{ij} = sum_s E[u_i u_s u_s u_j] = (S+2) δ_{ij}, not 3S δ_{ij}. For a unit-variance spherically symmetric vector with ||u||^2 = S (e.g., uniform on the scaled sphere), the value is S δ_{ij}. The equality in Eq. (35) is therefore false, so the variance expression (3S−1) in Proposition 1, Eq. (22), and the learning-rate condition involving (3S−1) in Theorem 1 are not proven as written. The result can likely be repaired because (S+2) ≤ 3S for S ≥ 1, so the compression-loss term remains an upper bound if the derivation is rewritten as an inequality, but the proof must be corrected and the theorem restated accordingly.","section":"Appendix B, Eq. (35)"},{"comment":"The channel-noise contribution is undercounted by a factor of L. The receiver noise r^H n is a scalar added to all L entries of y_G, so the reconstruction noise is (1/L) U (1_L r^H n), whose squared norm has expectation S N_0 ||r||^2 / L, not S N_0 ||r||^2 / L^2. Consequently, the channel-noise term in Proposition 1 (Eq. 22) and the corresponding term in Theorem 1 (Eq. 23) are too small by a factor of L, and the final bound does not follow as written. This is a load-bearing issue for the stated convergence guarantee, although the qualitative O(1/√T) rate would likely survive a corrected derivation.","section":"Appendix B, Part A (Eqs. 30–33)"},{"comment":"The bound σ_{k,t}^2 ≤ (1/L)||U_t Δw_{k,t}||^2 ≤ ||Δw_{k,t}||^2 is not valid for the random matrices used in the paper. For Gaussian or unit-variance spherical u_l, E[U U^T/L] = I, but U U^T/L is not ≤ I almost surely; for example, with L=1, the operator norm of u u^T is ||u||^2 = S, which is much larger than 1. Thus the pointwise inequality (a) in Eq. (32) is false, and the subsequent noise-variance bound in Eq. (33) does not follow. A correct derivation must handle the maximum over k more carefully, e.g., through an expectation bound on the maximum eigenvalue or by a different concentration argument; this affects the same noise term in Proposition 1 and Theorem 1.","section":"Appendix B, Eq. (32), inequality (a)"}],"minor_comments":[{"comment":"The noise term is written as n 1_S^T, but y_G is L-dimensional and the additive noise during the L-symbol transmission should be an N×L noise matrix, not the N-dimensional vector n times a length-S vector. The dimensions are inconsistent and should be corrected throughout.","section":"Section III-C, Eq. (15) and Eq. (19)"},{"comment":"The algorithm body says 'Send σ_k to the PS,' while the derivation in Section III-C assumes both μ_{k,t} and σ_{k,t} are known at the PS. Please clarify that both scalars are transmitted or explain how μ is obtained.","section":"Algorithm 1, line 16"},{"comment":"The variance expression is written as an equality, but both the intended derivation and the corrected version give an upper bound rather than an exact equality. Please state the result as a bound to avoid overclaiming.","section":"Proposition 1, Eq. (22)"},{"comment":"The sentence 'as training a classifier for the given datasets is getting harder' should be rephrased, e.g., 'as the classification task becomes harder,' for clarity.","section":"Section V-C, last paragraph"},{"comment":"The abstract states that Fed-ZOE requires 'at most 30% of the communication overhead of traditional approaches,' but the main compression claim is given as 0.07% per round. The origin of the 30% figure is not clearly tied to a specific comparison in the main text; please reconcile these numbers.","section":"Abstract and Section I"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable contribution to over-the-air federated learning, and the empirical study is extensive. The main concern is the correctness of the theoretical results: the errors in Appendix B are localized and likely repairable, but they currently invalidate the stated convergence guarantee. I see no sign of citation manipulation or scope mismatch. The novelty is moderate — random projection for compression in FL is not new in general, but its specific combination with OtA-FL and the RGE-based reconstruction is a worthwhile practical contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: Fed-ZOE is a genuinely useful compression idea for over-the-air FL, and the experiments back it. The convergence proof has a concrete error in the fourth-moment computation, but the method's core claim survives a corrected proof.\n\nWhat's actually new is the application of RGE-style compression to the local update vector before transmission, preserving superposition and cutting uplink to L scalars, with shared random seeds and downlink of only L scalars. That combination is not in the cited prior work. The mathematical core is standard E[uu^T]=I, but the system design is the contribution.\n\nThe paper does well on evidence. The experiments are extensive: five datasets, K values from 20 to 160, several baselines, and a learning-rate/L tradeoff that matches the qualitative prediction of the theorem. They also acknowledge the limitation about random direction selection, which is honest.\n\nThe soft spots are in the proof. In Appendix B, Eq. (35), they compute E[(u u^T)^2] = 3S I. For Gaussian u ~ N(0,I), the correct value is (S+2) I; for Rademacher, S I. The 3S-1 factor in Proposition 1 and the learning-rate condition in Theorem 1 depend on this. Separately, the channel-noise term in Part A treats the noise as entering through a single u_l, but it enters all L entries of y_G, so the variance is S N0 ||r||^2 / L, not /L^2. These change constants and the stated condition, but not the O(1/sqrt(T)) rate. So the theorem as written is not rigorously established; the qualitative behavior is likely right. Lack of code and repeated runs is a minor reproducibility issue.\n\nWho is this for? People working on wireless FL and communication-efficient training. It deserves a serious referee. The proof error is fixable and the contribution is worth engaging. I'd send it to review, with a request to correct the proof and release code.","headline":"A solid, practical compression idea for over-the-air FL with an honest but fixable error in the convergence proof.","tokens_in":23076,"tokens_out":2353,"would_cite":true,"duration_ms":19519,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Over-the-air federated learning can cut uplink traffic to ~0.07% by sending random projections of local updates, while keeping accuracy close to uncompressed aggregation.","keywords":["over-the-air federated learning","randomized gradient estimation","zeroth-order optimization","communication efficiency","gradient compression","superposition property","wireless edge learning","6G"],"falsifier":"Run the reported ResNet-18/CIFAR-10 experiment at $L=8192$; if the final test accuracy is more than a couple of points below the full-gradient OtA-FL baseline, the paper's headline claim fails.","tokens_in":21941,"feed_emoji":"📡","tokens_out":9773,"duration_ms":72761,"temperature":0.7,"pith_summary":"The paper proposes Fed-ZOE, a method that makes over-the-air federated learning (OtA-FL) drastically more communication-efficient. Instead of transmitting the full gradient vector from every edge device, each device computes its local weight update and then transmits only a few thousand scalar inner products of that update with shared random vectors. The parameter server reconstructs a good approximation of the aggregated gradient from those scalars, and the linearity of the reconstruction preserves the superposition property that makes over-the-air aggregation possible. The paper argues that with L = 8,192 projections for a ResNet-18 model with 11 million parameters, a compression ratio of about 0.07%, the trained model reaches accuracy close to uncompressed OtA-FL while using far less uplink and downlink bandwidth. This matters because model sizes are growing rapidly and OtA-FL's main bottleneck is the number of symbols that must be exchanged each round.","feed_headline":"8,192 scalars replace 11 million gradients in over-the-air learning","feed_subtitle":"Fed-ZOE sends random projections of local updates, approaching full-gradient accuracy at 0.07% of the uplink cost.","key_machinery":"The central object is the randomized gradient estimator (RGE), a standard tool in zeroth-order optimization. Its key identity is $\\mathbb{E}[u u^T] = I$, which lets $\\frac{1}{L}\\sum_{l=1}^L (u_l^T v) u_l$ approximate any vector $v$. Fed-ZOE applies this identity to the local model update vector $\\Delta w_{k,t}$, producing compressed scalars $\\phi_{t,k,l} = u_{t,l}^T \\Delta w_{k,t}$. The same random vectors are regenerated at the server from a shared seed, so the server reconstructs the average update as a linear combination of random vectors weighted by the aggregated scalars. This linearity is what preserves the superposition property of over-the-air computation, making the method compatible with OtA-FL.","core_discovery":"The central discovery is that the randomized gradient estimator can be repurposed as a communication code. A device replaces its S-dimensional update vector with L dot products against random vectors drawn from a shared seed; the server reconstructs the update as $\\frac{1}{L}\\sum_{l=1}^L (u_{t,l}^T \\Delta w_{k,t})u_{t,l}$. Because $\\mathbb{E}[u u^T] = I$ for the spherically symmetric unit-variance random vectors, the reconstruction is unbiased in expectation, and because it is a linear function of the transmitted scalars, it preserves the superposition property needed for over-the-air aggregation. With enough projections L, the compression loss falls below the channel noise floor, so the server aggregates almost as well as if it had received full gradients, at a small fraction of the symbol cost.","pith_inferences":["If the reconstruction error at $L=8192$ is already below the channel noise floor, the same projection count may suffice for billion-parameter models, making per-round communication nearly independent of model scale.","The variance decomposition suggests an optimal $L$: increase projections until compression loss is comparable to channel noise; beyond that, extra symbols buy little accuracy, which could inform adaptive resource allocation.","Because the reconstruction is linear and unbiased, combining Fed-ZOE with gradient sparsity in the projection domain or with error feedback across rounds might further improve accuracy per symbol."],"forward_implications":["Uplink communication cost per round becomes proportional to $L$, the number of random projections, rather than the model size $S$, so OtA-FL can scale to much larger models without a linear increase in bandwidth.","The downlink broadcast can also be compressed: the server sends only the aggregated scalars, and devices reconstruct the global update locally from the shared random seed.","The convergence rate remains $O(1/\\sqrt{T})$ in the number of communication rounds, matching standard FedAvg, with the compression variance controlled by $L$.","Unlike sparsification or quantization schemes, Fed-ZOE preserves the superposition property, so it can run entirely over the air without extra handshakes or iterative decoding.","Because the method works from random initialization, it is not limited to fine-tuning (as LoRA is) and can train full models from scratch."],"supporting_citations":[{"why":"Supplies the randomized gradient estimator identity that Fed-ZOE uses to compress and reconstruct updates.","marker":"[33]"},{"why":"Defines over-the-air federated learning and the superposition property that Fed-ZOE must preserve.","marker":"[6]"},{"why":"Provides the FedAvg local-update and aggregation structure that Fed-ZOE builds on.","marker":"[8]"},{"why":"Gives the uniform-forcing beamforming solution used to keep the aggregated estimate unbiased.","marker":"[29]"},{"why":"Used to compute the fourth moment of the Gaussian projection vectors in the variance bound.","marker":"[48]"}],"fun_headline_variants":["Over-the-air learning shrinks gradients to a few random dot products","Fed-ZOE: send scalars, not huge gradients, keep accuracy","Random projections cut over-the-air FL traffic to 0.07%","Gradient-free trick compresses federated learning over the air","Zeroth-order estimation slashes communication in wireless FL"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The convergence guarantee rests on the fourth-moment identity $\\mathbb{E}[(u u^T)^2] = 3S I$ for the Gaussian random projection vectors, which is not the value the stated distribution actually gives.","fun_headline_variants_meta":{"raw":{"variants":["Over-the-air learning shrinks gradients to a few random dot products","Fed-ZOE: send scalars, not huge gradients, keep accuracy","Random projections cut over-the-air FL traffic to 0.07%","Gradient-free trick compresses federated learning over the air","Zeroth-order estimation slashes communication in wireless FL"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000914,"raw_usage":{"total_tokens":3981,"prompt_tokens":1060,"completion_tokens":2921,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":676,"completion_tokens_details":{"reasoning_tokens":2831}},"tokens_in":676,"tokens_out":2921,"duration_ms":17550,"temperature":1.0,"reasoning_tokens":2831,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:15:58.468031+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the reported ResNet-18/CIFAR-10 experiment at $L=8192$; if the final test accuracy is more than a couple of points below the full-gradient OtA-FL baseline, the paper's headline claim fails.","supporting_citations":[{"cited_title":"Communication-efficient learning of deep networks from decentralized data,","cited_arxiv_id":null,"evidence_quote":"Provides the FedAvg local-update and aggregation structure that Fed-ZOE builds on."},{"cited_title":"A uniform-forcing transceiver design for over-the-air function computation,","cited_arxiv_id":null,"evidence_quote":"Gives the uniform-forcing beamforming solution used to keep the aggregated estimate unbiased."}],"review_version":1}