{"id":"336f85d6-7e6c-4375-8920-03aeb06ce9d6","arxiv_id":"2501.17874","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Multi-task over-the-air federated learning in cell-free massive MIMO performs best with centralized or partially centralized combining, outperforming cellular mMIMO, especially when devices are spread across cells.","lead":"Wireless devices training separate AI models can share the same cell-free massive MIMO network by sending their model updates as analog signals that add up in the air. The paper designs transmit and receive coefficients for this multi-task over-the-air federated learning setup and shows that centralized or partly centralized processing clearly beats conventional cellular base stations when devices are scattered.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The optimized designs assume perfect knowledge and error-free feedback of per-device model statistics; under realistic control-channel errors the multi-group gains over Cellular could shrink or reverse.","rationale":"The paper's core mathematics is sound: the channel estimation model, the MSE derivations, and the alternating optimization are internally consistent, and the convergence analysis in Theorem 1 provides a principled link between aggregation error and FL loss. The simulations are extensive, and the conclusion that centralized or semi-centralized cell-free processing outperforms the presented cellular baseline is plausible. However, the strongest load-bearing risk is the error-free side-information assumption. The optimized transmit coefficients in (35) and the combining vectors in (28) both depend directly on the per-device standard deviation nu_t_k, and the recovered model in (21) requires the per-device mean theta-bar_t_k. If these control-channel values are noisy or delayed, the target alignment in (22) is biased and the claimed optimality of the closed forms collapses. The multi-group comparison in Fig. 4 indicates that TCO provides a substantial gain for Level 2/3; because TCO is built on nu, imperfect nu would directly erode that gain. The paper does not analyze this robustness, so the headline claim that Cell-free mMIMO 'significantly outperforms' Cellular is only established under the idealized control channel. This is the same weakest assumption the reader identified, and the proposed SNR/quantization test would settle whether the concern actually lands. I concur with the CONDITIONAL verdict; no verdict change is needed.","tokens_in":20597,"tokens_out":17679,"duration_ms":181885,"concrete_test":"Re-run the Section V experiments for G=3, M=64, L=64, N=4, Pmax=20 dBm, with nu and theta-bar corrupted at the CPU by i.i.d. Gaussian noise at SNRs of 10, 20, and 30 dB (or quantized to 8/12 bits). Compare the weighted sum-MSE and FL test accuracy curves for Level 2/3 vs Cellular under both device distribution modes. If the Level 2/3 advantage shrinks by more than half its current margin at 20 dB SNR, the error-free assumption is load-bearing; if the qualitative ranking is unchanged, the concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The MSE expressions (22), (28), and (35) treat the per-device mean theta-bar_t_k and standard deviation nu_t_k as perfectly known to the CPU, and the optimized transmit coefficients b_t_k,d are assumed to be delivered to devices over error-free channels (Section III, after eq. (13)). In a practical system these quantities are themselves wireless transmissions: theta-bar and nu must be sent from devices to the CPU, and the coefficients must be broadcast back. If nu is estimated with error, the target gamma_j nu_j in (22) is biased and the closed-form solutions (28) and (35) are no longer optimal; the alternating optimization then minimizes the wrong objective. The numerical comparison in Fig. 4 shows that transmit-coefficient optimization (TCO) is essential for the multi-group gain of Level 2/3 over Cellular. Since TCO directly uses nu, the magnitude of the claimed advantage likely degrades under imperfect side information. The paper provides no robustness analysis, so the central claim that Level 2/3 'significantly outperforms' Cellular is conditional on this idealization.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper considers over-the-air federated learning (OtA FL) with multiple FL groups/tasks in a cell-free massive MIMO network. Devices normalize their local model updates and transmit them over shared time-frequency resources; APs and the CPU use linear receive combining to recover each group's weighted model average. Three levels of AP cooperation are studied: fully centralized processing (Level 3), local processing with centrally designed combiners (Level 2), and fully local processing with simple averaging (Level 1). The paper derives closed-form MSE expressions under channel estimation error, spatial correlation, and inter-group interference; formulates a weighted sum-MSE minimization over transmit coefficients and receive combining solved by alternating optimization; and proves a standard FL convergence bound (Theorem 1) that justifies minimizing the aggregation error. A cellular mMIMO baseline is optimized under the same framework. Numerical results show that Level 2 and Level 3 outperform the cellular baseline in MSE and test accuracy, especially for multiple FL groups and distributed device placement, while Level 1 does not.","tokens_in":20779,"tokens_out":10168,"duration_ms":105066,"significance":"If the reported results hold, the paper is a useful extension of OtA FL to multi-task scenarios and provides a systematic comparison of AP cooperation levels that includes fronthaul signaling. The analytical work is a clear strength: the MSE expressions in (22), (28), (35), and (40) are derived under standard massive MIMO assumptions, and Theorem 1 is a clean smooth-and-strongly-convex convergence bound. The fair total-antenna comparison between cell-free and cellular systems and the fronthaul accounting in Table II and Section V-C are also valuable. No fitted constants or circular reasoning are apparent; the convergence analysis supports the MSE-minimization objective rather than assuming it. The main limitation is that the claimed gains are conditional on perfect side information about per-device statistics and transmit coefficients; no robustness analysis is provided for that assumption.","major_comments":[{"comment":"The central performance comparison assumes that the per-device statistics theta-bar and nu are conveyed to the CPU over error-free channels and that the optimized transmit coefficients b_{k,d} are broadcast to the devices error-free. These quantities are inputs to the optimization: nu_j is the target in (22), and the closed-form solutions (28) and (35) use it as an input. If nu_j is estimated with error, the target gamma_{jg} nu_j in (22) is biased and the alternating algorithm minimizes an incorrect objective; if the b-coefficient broadcast is corrupted or delayed, the signal model in (14) no longer matches the assumed transmit coefficients. Since Fig. 4 shows that transmit-coefficient optimization is important for the Level 2/3 advantage over the cellular baseline in the multi-group case, the headline claim that cell-free 'significantly outperforms' cellular is conditional on this idealization. The paper should add a robustness study (e.g., perturbed nu and b, or a limited/delayed feedback link) or explicitly qualify the conclusions, and it should account for the control overhead needed to deliver D coefficients per device per training round when discussing communication-resource scalability.","section":"Section III, after (13) and (35); MSE expressions (22), (28), (35)"},{"comment":"The manuscript repeatedly describes the proposed solutions as 'optimal' (Abstract, Section I contributions, Section III-A). However, problem (24) is non-convex because of the coupling between transmit coefficients and receive combining, and the proposed alternating optimization only yields a stationary point; its convergence is demonstrated numerically in Fig. 2 but not proved, and no global optimality certificate is given. The wording should be changed to 'optimized' or the local-stationarity nature of the solution should be stated explicitly wherever 'optimal' appears, so that the claims match the mathematics.","section":"Section III-A, problem (24); contribution list in Section I"}],"minor_comments":[{"comment":"In the second sum of (43), the transmitted signal is written as s_{i,d} but should be s_{j,d} to match the device index in that term.","section":"Eq. (43)"},{"comment":"The text 'only AL model parameters are exchanged' appears to contain a typo; this should read 'ML model parameters' or 'AI model parameters'.","section":"Section I, paragraph 2"},{"comment":"In the Level 3 row, the expression '(tau_p + tau_u) N L - K L N^2 / 2' uses a dash that could be misread as a minus sign; using a separator or parentheses would improve clarity.","section":"Table II"},{"comment":"The simulations do not specify the values of the group-priority weights omega_g used in the weighted sum-MSE; the authors should state whether equal weights are assumed and whether the conclusions are sensitive to this choice.","section":"Section V-A"},{"comment":"The convergence of the alternating optimization is shown for one random channel realization; reporting the average over several realizations would make the claim that the algorithm converges 'after a few iterations' more robust.","section":"Section V-A, Fig. 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid but incremental extension of the authors' prior work [1] to multi-task FL with a cellular baseline. The main risk to the headline conclusion is the perfect side-information assumption; a robustness experiment or a clear qualification would address it. The reliance on the authors' prior paper is acknowledged and is not a concern. The fit with the journal's scope is appropriate. I would support acceptance after the major revision is completed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Conditional is the right call. This paper is a genuine extension of the single-task OtA computation work to multi-task FL in cell-free massive MIMO, with three cooperation levels, closed-form MSE expressions, and a clean convergence bound (Theorem 1) that justifies minimizing aggregation error. The alternating optimization with closed-form updates is internally consistent, converges in a handful of iterations in their simulations, and the comparison against Cellular mMIMO is fair: same total antenna count, same propagation model. Credit where due: Level 1 being unable to beat Cellular is reported honestly, which makes the Level 2/3 gains more credible.\n\nThe soft spots are what the reader flagged. The abstract and conclusions say 'optimal designs,' but problem (24) is non-convex and the algorithm finds stationary points of the alternating problem, not a global optimum. That's a wording fix, not a flaw in the math.\n\nThe larger issue is the error-free side information. The design assumes devices send their local mean and standard deviation (theta-bar and nu) to the CPU over error-free channels, and the optimized transmit coefficients are broadcast back error-free. The MSE expressions (22), (28), (35) treat nu as perfectly known. The transmit-coefficient optimization, which their own Fig. 4 shows is essential for the multi-group gain, directly uses nu. If nu arrives with estimation error, the target gamma_j nu_j is biased, and the closed-form solutions are no longer optimal. No robustness analysis is offered. So the 'significantly outperforms' claim is conditional on this idealization. It is not a fatal flaw—the side information is low-dimensional scalar statistics, and similar assumptions are common in the OtA FL literature—but it is the right thing to ask about in review.\n\nMinor: the group-priority weights omega_g are never specified in the simulation, and there are no error bars or code. Also, the convergence proof of the alternating optimization is only numerical (Fig. 2), not analytical; that's fine for a system paper.\n\nWho is this for? Researchers in communication-efficient FL and cell-free mMIMO. It gives a practical design rule: use Level 2 or 3, skip Level 1, and expect big gains over cellular when devices are spread across cells. A serious referee should engage with it. I'd send it to review, asking for a softened 'optimal' claim, a robustness section on side-information errors, and specification of the simulation weights.","headline":"Solid multi-task OtA FL extension with a fair cell-free vs cellular comparison; the 'optimal' wording and error-free side-information assumption are the things to push on in review.","tokens_in":21301,"tokens_out":2572,"would_cite":true,"duration_ms":25218,"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":"This paper claims that, with optimized transmit coefficients and receive combining, cell-free Massive MIMO using Level 2 or Level 3 access-point cooperation supports multiple simultaneous over-the-air federated learning tasks with much…","keywords":["over-the-air computation","federated learning","cell-free massive MIMO","multi-task learning","model aggregation error","transmit coefficient optimization","receive combining","6G networks"],"falsifier":"In the same $500\\times500$ m simulation setup, replace the assumed perfect control channel with a finite-rate link carrying quantized values of $\\bar{\\theta}_k^t$, $\\nu_k^t$, and $b_{k,d}^t$, and rerun the Level 2 and Level 3 experiments. If the resulting test accuracy no longer exceeds the cellular baseline, the central claim collapses; otherwise the design is robust to the control-channel idealization.","tokens_in":20394,"feed_emoji":"📡","tokens_out":7036,"duration_ms":66684,"temperature":0.7,"pith_summary":"The paper asks whether cell-free Massive MIMO can make over-the-air federated learning work when several groups of devices train different models at once. It argues that the answer is yes, provided the access points cooperate centrally or semi-centrally: the proposed transmit-coefficient and receive-combining designs minimize the weighted sum of model-aggregation mean-squared errors across all groups, and in simulation the resulting Level 2 and Level 3 schemes clearly outperform a cellular Massive MIMO baseline with the same antenna count. The paper also shows that the lowest-cooperation scheme, Level 1, cannot beat that baseline, so the benefit comes specifically from coordinated combining. A convergence bound connects the aggregation error to the gap between training loss and optimal loss, which is why a reader should care: aggregation error, not channel noise per se, is the quantity a system designer can attack.","feed_headline":"Cell-free MIMO beats cellular in multi-task over-the-air FL","feed_subtitle":"Central or semi-central access-point cooperation keeps multi-task FL accurate even for scattered devices.","key_machinery":"The central object is the over-the-air model aggregation error, measured by the mean squared error between the desired weighted sum of local model vectors and the recovered global model vector. Each device first normalizes its local model vector to zero mean and unit variance per entry, so that transmit coefficients can be designed under a power constraint, and the local mean and standard deviation are assumed to be sent to the CPU over error-free channels. The argument is carried by an alternating optimization: the receive combining vectors are updated in closed form as regularized MMSE combiners built from the channel estimates and their error covariances, and the transmit coefficients are updated from first-order KKT conditions, repeated until the weighted sum-MSE stops decreasing. A convergence bound shows that the gap between the federated training loss and the optimal loss is controlled by the accumulated aggregation error, which justifies minimizing that error rather than treating communication and learning separately.","core_discovery":"On the authors' own terms, the paper establishes that multi-task over-the-air federated learning is practically feasible in cell-free Massive MIMO when access points cooperate: at Level 3 the CPU centrally estimates all channels and computes global combining vectors; at Level 2 each access point applies a slice of the same centrally designed vector locally; at Level 1 each access point estimates channels itself, devices transmit at full power, and the CPU only averages local estimates. Given MMSE channel estimates, spatially correlated fading, and inter-group interference, the closed-form combining rule and transmit-coefficient rule minimize the weighted sum of $\\mathrm{MSE}_{g,d,(3)}$ terms, and the Level 2 replica reaches the same global estimate as Level 3. Numerically, Level 2 and Level 3 achieve much lower weighted sum-MSE and markedly higher federated learning test accuracy than a cellular Massive MIMO baseline with the same total antenna count, for both device distribution modes and for all three learning tasks; Level 1 does not. The conclusion is that appropriate cooperation levels make cell-free Massive MIMO the architecture that can support multi-task over-the-air federated learning, whereas cellular Massive MIMO fails when devices are spread across cells or inter-group interference is strong.","pith_inferences":["Editorial inference: the error-free side-information assumption is the most likely place the comparison could reverse; an end-to-end test with finite-rate, quantized, or delayed control channels would show whether Level 2 and Level 3 retain their margin over the cellular baseline.","Editorial inference: the same weighted sum-MSE formulation could be used as a dynamic scheduler, letting the network shift accuracy among tasks by adjusting the group weights $\\omega_g$, a degree of freedom the paper does not explore.","Editorial inference: the convergence theorem assumes convex loss functions while the experiments use feedforward networks, so the practical claim depends on the error-to-accuracy link persisting for non-convex training; repeating the comparison with convolutional or transformer models would test that extension.","Editorial inference: replacing the analog over-the-air transmission with a digital or quantized aggregation scheme, which the paper names as future work, would change the side-information burden and could make Level 2 or Level 3 attractive even in settings where the error-free control channel is not realistic."],"forward_implications":["If the claim holds, an operator can run multiple simultaneous federated learning tasks over one cell-free Massive MIMO network without devoting separate time-frequency resources to each group, because the transmit-coefficient design handles inter-group interference on the same slots.","Level 1 should be avoided: with no transmit-coefficient optimization and no centralized combining, it yields higher aggregation MSE than the cellular baseline in the paper's comparisons, regardless of device distribution.","Level 2 and Level 3 produce the same global estimate, so the choice between them is a fronthaul tradeoff: Level 2 is more fronthaul-friendly when $N>G$ and a coherence block contains few training rounds, namely $C < \\tau_u(N-G)/(NG)$.","The nonzero MSE floor at high transmit power, caused by channel estimation error, sets a ceiling on convergence: increasing device transmit power alone cannot make over-the-air aggregation arbitrarily accurate."],"supporting_citations":[{"why":"Supplies the cell-free Massive MIMO over-the-air computation model and the analytical MSE floor that this paper extends to multi-task federated learning.","marker":"[1]"},{"why":"Provides the cell-free Massive MIMO processing and simulation baseline used for the numerical comparisons.","marker":"[29]"},{"why":"Prior over-the-air federated learning in cell-free Massive MIMO with suboptimal maximum-ratio combining and full-power transmission, which this paper improves with optimized coefficients.","marker":"[30]"},{"why":"Multi-cell multi-task federated learning baseline whose single-antenna base stations cannot manage inter-group interference.","marker":"[31]"},{"why":"Supports the assumption that the global model is broadcast to devices over error-free channels.","marker":"[32]"},{"why":"Supplies the strong-convexity and Lipschitz-gradient machinery used to derive the convergence bound in Theorem 1.","marker":"[34]"},{"why":"Provides the 3GPP Urban Microcell path-loss and shadow-fading model used in all numerical experiments.","marker":"[36]"},{"why":"Provides the Fashion-MNIST dataset for the first federated learning task.","marker":"[37]"},{"why":"Provides the MNIST dataset for the second federated learning task.","marker":"[38]"},{"why":"Provides the EMNIST dataset for the third federated learning task.","marker":"[39]"}],"fun_headline_variants":["Cell-free MIMO with cooperation dominates multi-task over-the-air FL","Central or semi-central cooperation key for cell-free MIMO FL","Over-the-air federated learning thrives with cell-free MIMO cooperation","Multi-task FL: cell-free MIMO + cooperation beats cellular MIMO","Cooperative cell-free MIMO makes multi-task over-the-air FL feasible"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything depends on the CPU and devices exchanging the local statistics ($\\bar{\\theta}_k^t$, $\\nu_k^t$) and the optimized transmit coefficients over error-free channels; if that control information is delayed, quantized, or lost, the aggregation-error expressions and the claimed superiority of Levels 2 and 3 over the cellular baseline no longer hold.","fun_headline_variants_meta":{"raw":{"variants":["Cell-free MIMO with cooperation dominates multi-task over-the-air FL","Central or semi-central cooperation key for cell-free MIMO FL","Over-the-air federated learning thrives with cell-free MIMO cooperation","Multi-task FL: cell-free MIMO + cooperation beats cellular MIMO","Cooperative cell-free MIMO makes multi-task over-the-air FL feasible"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000201,"raw_usage":{"total_tokens":1432,"prompt_tokens":1051,"completion_tokens":381,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":667,"completion_tokens_details":{"reasoning_tokens":289}},"tokens_in":667,"tokens_out":381,"duration_ms":4018,"temperature":1.0,"reasoning_tokens":289,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:29:53.435400+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the same $500\\times500$ m simulation setup, replace the assumed perfect control channel with a finite-rate link carrying quantized values of $\\bar{\\theta}_k^t$, $\\nu_k^t$, and $b_{k,d}^t$, and rerun the Level 2 and Level 3 experiments. If the resulting test accuracy no longer exceeds the cellular baseline, the central claim collapses; otherwise the design is robust to the control-channel idealization.","supporting_citations":[{"cited_title":"Making cell-free massive MIMO competitive with MMSE processing and centralized implementation,","cited_arxiv_id":null,"evidence_quote":"Provides the cell-free Massive MIMO processing and simulation baseline used for the numerical comparisons."},{"cited_title":"Over-the-air federated learning over scalable cell-free massive MIMO,","cited_arxiv_id":null,"evidence_quote":"Prior over-the-air federated learning in cell-free Massive MIMO with suboptimal maximum-ratio combining and full-power transmission, which this paper improves with optimized coefficients."},{"cited_title":"Interference management for over-the-air federated learning in multi-cell wireless networks,","cited_arxiv_id":null,"evidence_quote":"Multi-cell multi-task federated learning baseline whose single-antenna base stations cannot manage inter-group interference."},{"cited_title":"Deploying federated learning in large-scale cellular networks: Spatial convergence analysis,","cited_arxiv_id":null,"evidence_quote":"Supports the assumption that the global model is broadcast to devices over error-free channels."},{"cited_title":"Hybrid deterministic-stochastic methods for data fitting,","cited_arxiv_id":null,"evidence_quote":"Supplies the strong-convexity and Lipschitz-gradient machinery used to derive the convergence bound in Theorem 1."},{"cited_title":"Further advancements for E-UTRA physical layer aspects,","cited_arxiv_id":null,"evidence_quote":"Provides the 3GPP Urban Microcell path-loss and shadow-fading model used in all numerical experiments."},{"cited_title":"EMNIST: Ex- tending MNIST to handwritten letters,","cited_arxiv_id":null,"evidence_quote":"Provides the EMNIST dataset for the third federated learning task."}],"review_version":1}