{"id":"23d06da8-b1b3-4245-b8d8-5fac097d3f75","arxiv_id":"2506.06090","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A Bayesian federated learning framework uses over-the-air superposition to aggregate local posterior distributions, with convergence analysis and power control, improving accuracy and calibration under scarce non-i.i.d. wireless data.","lead":"This paper proposes a wireless federated-learning scheme in which each device transmits a distribution over model weights, and the radio channel itself combines the distributions into a global one. The authors report faster, better-calibrated learning than standard methods when edge devices have very few and unevenly distributed samples, with a convergence proof and a transmit-power rule.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The two-phase encoding does not transmit sufficient statistics of a single local posterior, so the exact reconstruction of the global conflation in Eqs. (18)-(19) is not established as written.","rationale":"The reader's weakest_assumption identifies exactly this gap, and my reading agrees. The concern is load-bearing because the abstract and Section IV claim distribution-level aggregation by AirComp; if ρ and ν are not derived from a single local posterior, the channel superposition does not compute product-of-Gaussians conflation. The convergence analysis and power-control results do not repair this, because Theorem 1 analyzes a two-phase objective with local gradients, not the conflation identity (18)-(19). The paper may still contain a useful approximate scheme with plausible convergence behavior, so the appropriate response is not to reject the entire contribution but to require a derivation of the missing equivalence or an explicit reframing as approximate Bayesian aggregation. Since the reader already reached CONDITIONAL and identified the same assumption, my verdict is UNCHANGED.","tokens_in":19222,"tokens_out":4398,"duration_ms":42302,"concrete_test":"Independently re-derive one round of the protocol for K=2 Gaussian clients with known local posteriors N(μ_1, Σ_1) and N(μ_2, Σ_2) with μ_1, μ_2 distinct from μ_t, implementing Eqs. (23)-(37) exactly, and compare the server-reconstructed ρ_{t+1}, ν_{t+1} with the true conflation in Eqs. (18)-(19). If the reconstructed mean differs from Σ_{t+1} Σ_k π_k Σ^{-1}_{t,k} μ_{t,k}, the exact-reconstruction claim is refuted as written.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that the pair (ρ_{t,k}, ν_{t,k}) defined in Eqs. (21)-(22) are sufficient statistics of the same local posterior q_{θ_{t,k}} = N(μ_{t,k}, Σ_{t,k}) from Eq. (16), so that superposition in the channel computes the conflation (18)-(19). The paper does not establish this. Phase 1 (Eqs. 23-25) updates ρ_{t,k} (precision) while holding the sampling mean at μ_t, the previous global mean, not at μ_{t,k}. Phase 2 (Eqs. 33-35) initializes ν_{t,k} = Σ^{-1}_{t,k} Σ_{t+1} μ_t and then samples from N(μ_{t,k}, Σ_{t+1}) with the covariance fixed to the newly aggregated global covariance; it never optimizes the mean of the local posterior N(μ_{t,k}, Σ_{t,k}). Consequently, the reconstructed ν_{t+1} in Eq. (37) is, at best, a weighted sum of statistics from two different distributions, and it does not follow that ν_{t+1} = μ_{t+1} as claimed after Eq. (37). The exact Bayesian conflation claim therefore fails unless a missing derivation supplies the missing link; without it, the method is a two-phase approximate heuristic.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a wireless federated learning framework that combines Bayesian variational inference with over-the-air computation (AirComp). Each client maintains a Gaussian variational posterior, and the server aggregates the local posteriors by product-of-Gaussians conflation. To do this over the wireless channel, the authors design a two-phase analog transmission scheme: phase 1 transmits a precision vector (diagonal of the inverse covariance), and phase 2 transmits a precision-weighted mean vector. The superposition property of the channel is claimed to compute the conflation in constant time per round. The paper also provides a convergence analysis under smoothness assumptions, a closed-form transmit power-control solution, and MNIST simulations showing accuracy and calibration gains over FedAvg, FedProx, and SCAFFOLD.","tokens_in":19594,"tokens_out":12271,"duration_ms":111920,"significance":"If the central reconstruction claim is correct, the paper would be the first distribution-level AirComp scheme for Bayesian FL, with constant-time uplink aggregation that scales independently of the number of clients and with explicit convergence guarantees under fading and noise. The paper is clearly written, the simulation setup is described in detail with concrete hyperparameters (Tables I and II), and the power-control solution in Eq. (43) is a useful closed-form contribution. However, the load-bearing claim that the two transmitted vectors are sufficient statistics of a single local posterior is not established, which puts the exactness of the distribution-level conflation in question. The convergence analysis is a standard smoothness-descent argument and is presented with a proof, but it is not connected to the simulated setting.","major_comments":[{"comment":"The two-phase parameterization does not transmit sufficient statistics of a single local posterior. Phase 1 (Eqs. 23-25) updates the precision vector rho_{t,k} while keeping the sampling mean fixed at the old global mean mu_t. Phase 2 (Eqs. 33-35) updates the vector nu_{t,k} while fixing the sampling covariance to the newly aggregated global covariance Sigma_{t+1}, not to the local covariance Sigma_{t,k} from phase 1. The conflation formulas (18)-(19) require the mean and precision of the same local posterior q_{theta_{t,k}} = N(mu_{t,k}, Sigma_{t,k}) of Eq. (16). The paper does not derive that the pair (rho_{t,k}, nu_{t,k}) are sufficient statistics of any single local posterior, and therefore the assertion after Eq. (37) that nu_{t+1} is identical to mu_{t+1} does not follow. This is a load-bearing gap: without a missing derivation, the algorithm is an approximate two-phase heuristic rather than the claimed exact distribution-level conflation. The authors should either supply the missing derivation or explicitly reframe the contribution as an approximate scheme with a quantified approximation error.","section":"Section IV-B, Eqs. (20)-(22), (23)-(25), (33)-(37)"},{"comment":"The convergence analysis assumes E=1 and d=F, giving N1=N2=1, while the simulations use E=3, d=62,346, F=1,024, and N1=N2=61. As written, Theorem 1 does not cover the setting in which all simulation results are obtained. The paper should provide an analysis for multiple local epochs and for d>F, or state clearly that the theoretical result applies only to the simplified E=1, d=F case and that the extension is left for future work.","section":"Section V-A and Tables I-II"},{"comment":"The corollary drops the channel-noise term that appears in Theorem 1 (the last term in Eq. (40)) without a formal argument. In the proof in Appendix A, that term behaves as O(1/T) only if the local gradient norms are uniformly bounded; this assumption is not stated. The claim that the impact of additive noise on convergence becomes negligible needs either a bounded-gradient assumption or a separate argument that the term vanishes as T grows.","section":"Corollary 1, Eq. (41)"},{"comment":"The power-control optimization in Eq. (42) minimizes the per-device squared distance between the update magnitude and the transmitted magnitude, but the convergence bound (40) depends on the squared norm of the aggregate error xi_{rho,t} = sum_k pi_k sgn(Delta_{rho,t,k}) odot e_{rho,t,k}. Minimizing the sum of per-device squared errors does not necessarily minimize the norm of the aggregate error, since errors from different devices can add or cancel. If this is intended as a heuristic proxy, the text should say so explicitly; as written, the claim that the power control is 'grounded in' the convergence analysis overstates the connection.","section":"Section V-B, Eqs. (42)-(43)"}],"minor_comments":[{"comment":"Lines 14 and 26 update rho_{t+1} and nu_{t+1} with an extra factor of bar{delta}_{rho,t} and bar{delta}_{nu,t}, respectively, while Eqs. (32) and (37) do not include this factor. Please align the algorithm with the equations.","section":"Algorithm 1, lines 14 and 26"},{"comment":"The text says each device generates MC samples from N(mu_{t,k}, Sigma_{t+1}), but the displayed formula gives the sampling mean as Sigma_{t+1}^{-1} Sigma_{t,k} nu_{t,k}, which is not mu_{t,k} in general. Please correct either the text or the formula.","section":"Eq. (34)"},{"comment":"In Eq. (47), the argument of nabla_rho L is written as (nu_t, rho_t) at the first occurrence and as (mu_t, rho_t) later in the same display; Theorem 1 and Assumption 1 use (mu_t, rho_t) consistently. This notational inconsistency should be fixed.","section":"Appendix A, Eq. (47)"},{"comment":"The phrase 'Hadarmard product' should be 'Hadamard product'.","section":"Section III-A"}],"recommendation":"major_revision","confidential_remarks":"The key issue in major comment 1 is not a stylistic concern but a matter of whether the central claimed result is true as stated. The two-phase scheme as written does not appear to transmit sufficient statistics of a single local posterior, so the exact conflation claim is not established. If the authors cannot provide the missing derivation, the manuscript should be reframed as an approximate Bayesian AirComp method, with the approximation error analyzed or at least clearly acknowledged. Given the promising empirical results and the well-structured presentation, I do not think rejection is warranted, but substantial revision is needed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this paper has a genuinely new idea, and a load-bearing gap in how it is justified. The idea is to transmit two precision-weighted statistics—diag(Σ^{-1}_{t,k}) and Σ_{t+1}Σ^{-1}_{t,k}μ_t—over the air so that superposition directly computes the product-of-Gaussians conflation. That is a real design step, and the closed-form power control from a KKT argument, plus a convergence bound that explicitly includes channel noise and power-constrained distortion, goes beyond prior Bayesian AirComp work [19]–[21]. The empirical results, particularly the ECE improvement on a very data-scarce, label-skewed MNIST setup, are plausible and worth seeing.\n\nThe soft spot is the two-phase encoding. Equations (21)–(22) are claimed to be sufficient statistics of the local posterior q_{θ_{t,k}} = N(μ_{t,k}, Σ_{t,k}) of Eq. (16), but the algorithm never produces that posterior. In phase 1, ρ is optimized while the sampling mean is fixed to the previous global μ_t (Eq. 24), not to the local mean. In phase 2, ν is initialized with Σ^{-1}_{t,k}Σ_{t+1}μ_t and then optimized while the covariance is fixed to the newly aggregated global Σ_{t+1} (Eq. 34); the local mean is never optimized. The pair therefore comes from two different optimization problems, and there is no derivation that their conflation equals the global posterior of (18)–(19). The statement after (37) that ν_{t+1} = μ_{t+1} does not follow. This is not a cosmetic flaw; it means the 'distribution-level aggregation' is, as written, an approximate two-phase heuristic rather than the exact Bayesian conflation advertised. The reader's stress-test note pins this down correctly.\n\nTwo smaller issues track the same pattern. Theorem 1 is stated and proved only for E=1 and d=F, while the simulations run E=3, d=62,346, and N=61 symbols; Corollary 1 drops the channel-noise term without a formal argument. No error bars or code are provided, so I could not check the empirical claims or the apparent subscript inconsistencies.\n\nI still think this paper deserves a serious referee. The design is clever, the power-control analysis is useful, and the calibration result is interesting. The right fix is to either supply the missing derivation or honestly reframe the method as approximate Bayesian aggregation, then extend or qualify the theorem and release code. I would bring it to a reading group as an example of a promising idea that overreaches its own equations.\n\nRecommendation: send it to peer review, but tell the authors the conflation claim and the theorem/simulation mismatch have to be addressed.","headline":"Clever AirComp design for Bayesian FL, but the exact conflation claim doesn't follow from the algorithm as written; worth a rigorous revision.","tokens_in":20025,"tokens_out":4133,"would_cite":false,"duration_ms":37081,"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":"The paper claims that encoding each client's Gaussian posterior as two precision-weighted statistics lets wireless channel superposition perform Bayesian posterior conflation exactly, making Bayesian federated learning over the air…","keywords":["wireless federated learning","over-the-air computation","Bayesian inference","posterior conflation","Gaussian variational inference","power control","data scarcity","non-i.i.d. data"],"falsifier":"Run the two-phase protocol in a noiseless, infinite-power, perfect-channel setting and compare the server's reconstructed global Gaussian to the exact product of the local posteriors in Eqs. (17)-(19); any mismatch in the recovered mean or covariance would falsify the exact-reconstruction claim.","tokens_in":18999,"feed_emoji":"📶","tokens_out":15307,"duration_ms":123605,"temperature":0.7,"pith_summary":"This paper tries to establish that Bayesian federated learning can be carried out over a shared wireless channel: each client transmits two precision-weighted statistics of its Gaussian variational posterior, and the physical superposition of the signals sums them into the global posterior conflation. If true, uplink cost for Bayesian aggregation becomes constant in the number of clients, while retaining the uncertainty-aware benefits that make Bayesian FL attractive when local data are tiny and skewed. The authors back the scheme with a convergence analysis that includes Rayleigh fading, receiver noise, and power constraints, and with a closed-form power-control rule that minimizes power-limited update distortion. Their simulations indicate higher converged test accuracy and roughly half the expected calibration error of conventional FL baselines in data-scarce, non-i.i.d. settings.","feed_headline":"Two-phase AirComp makes Bayesian FL aggregation a wireless sum","feed_subtitle":"Clients send two statistics each; channel superposition performs posterior aggregation, cutting cost and calibration.","key_machinery":"The central object is the product-of-Gaussians conflation of local variational posteriors, $q_{\\theta_{t+1}}(w)\\propto\\prod_{k\\in\\mathcal{K}}N(w|\\mu_{t,k},\\Sigma_{t,k})^{\\pi_k}$, whose sufficient statistics are the precision-weighted sums in Eqs. (18)-(19). The mechanism carrying the argument is the two-phase sufficient-statistics encoding: phase 1 transmits $\\rho_{t,k}$, the diagonal of the local precision matrix, and phase 2 transmits $\\nu_{t,k}$, a precision-scaled mean vector tied to the local mean under the freshly aggregated global covariance. These quantities enter the global formulas as sums, so the waveform superposition of the shared channel computes the aggregation in constant time, and the power-control vectors of Eq. (28) align received amplitudes under each device's per-symbol power constraint.","core_discovery":"The paper's central claim is that distribution-level Bayesian aggregation can be offloaded to the wireless channel. Each client's mean-field Gaussian posterior $N(\\mu_{t,k},\\Sigma_{t,k})$ is encoded as the diagonal precision $\\rho_{t,k}=\\mathrm{diag}(\\Sigma^{-1}_{t,k})$ and the precision-scaled mean $\\nu_{t,k}=\\Sigma_{t+1}\\Sigma^{-1}_{t,k}\\mu_{t,k}$ (Eqs. 21-22). Because the global conflation updates (Eqs. 18-19) are sums of these quantities, the superposition of analog OFDM transmissions yields the product-of-Gaussians aggregation $q_{\\theta_{t+1}}(w)=N(\\mu_{t+1},\\Sigma_{t+1})$ (Eq. 17) after one post-processing step at the base station. The protocol runs in two phases: phase 1 aggregates the precisions into the global covariance, phase 2 aggregates the precision-scaled means into the global mean. Theorem 1 (Eq. 40) bounds the expected loss decrease under fading, noise, and power limits, Corollary 1 identifies power-limited update distortion as the dominant asymptotic obstruction, and the closed-form power allocation (Eq. 43) is designed to minimize that distortion. The paper presents this as the first convergence-guaranteed AirComp scheme for Bayesian FL.","pith_inferences":["The exact-reconstruction claim is an idealization: the paper does not derive that the phase-1 precision and phase-2 mean come from the same local posterior, so in practice the scheme is best read as a two-phase approximate Bayesian aggregation whose bias against the exact conflation should be measured.","Because phase 2 uses the freshly aggregated global covariance, the two phases are coupled across the aggregation round; analyzing this as a delayed or two-timescale update could reveal convergence effects the current single-round bound omits.","A natural testable variant would transmit both statistics in a single concurrent phase, which would halve aggregation latency if the sufficient-statistics issue is resolved.","The power-control analysis assumes devices know their own CSI while the base station does not; quantifying how the closed-form allocation degrades under imperfect or delayed local CSI is a direct extension within the same framework."],"forward_implications":["Uplink communication for Bayesian FL aggregation becomes independent of the number of clients per round, because the channel itself performs the summation.","Under the convergence analysis, fading and additive noise eventually become secondary; the power-limited update distortion is what determines how close the training can get to a stationary point.","The closed-form power control (Eq. 43) gives each device a cheap rule, from local CSI and one scalar feedback, for clipping its update to fit its power budget with minimal distortion.","In data-scarce, label-skewed settings, distribution-level aggregation is claimed to yield higher converged accuracy and better-calibrated confidence scores than conventional frequentist FL aggregation.","The same power-control rule applies to both phases, so the mean and covariance updates are handled symmetrically by the optimization machinery."],"supporting_citations":[{"why":"Defines the global posterior as the product of local posteriors, the exact distribution-level aggregation the AirComp scheme must reproduce.","marker":"[10]"},{"why":"Establishes that the normalized product of distributions is the conflation minimizing information loss, justifying the product-of-Gaussians form.","marker":"[25]"},{"why":"Introduces AirComp for FL over fading channels, giving the superposition-based aggregation concept that the proposed two-phase scheme adapts to posteriors.","marker":"[11]"},{"why":"Supplies the optimized power-control design for AirComp FL whose distortion-minimization approach is extended to the Bayesian setting.","marker":"[15]"},{"why":"Defines the conventional averaging baseline over which the proposed framework's gains are demonstrated.","marker":"[1]"}],"fun_headline_variants":["Two-phase AirComp: Bayesian FL posterior aggregation as a wireless sum","AirComp turns Bayesian FL aggregation into channel superpositions","Wireless sum enables distribution-level Bayesian FL aggregation","Bayesian FL aggregation: two-phase AirComp as a wireless sum"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole construction rests on the assumption that the two quantities a client transmits—its precision in the first phase and its mean in the second—describe one and the same local posterior, an assumption the paper uses without deriving.","fun_headline_variants_meta":{"raw":{"variants":["Two-phase AirComp: Bayesian FL posterior aggregation as a wireless sum","AirComp turns Bayesian FL aggregation into channel superpositions","Wireless sum enables distribution-level Bayesian FL aggregation","Bayesian FL aggregation: two-phase AirComp as a wireless sum"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000736,"raw_usage":{"total_tokens":3350,"prompt_tokens":1070,"completion_tokens":2280,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":686,"completion_tokens_details":{"reasoning_tokens":2212}},"tokens_in":686,"tokens_out":2280,"duration_ms":19491,"temperature":1.0,"reasoning_tokens":2212,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T06:02:15.770727+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the two-phase protocol in a noiseless, infinite-power, perfect-channel setting and compare the server's reconstructed global Gaussian to the exact product of the local posteriors in Eqs. (17)-(19); any mismatch in the recovered mean or covariance would falsify the exact-reconstruction claim.","supporting_citations":[{"cited_title":"A Bayesian federated learning framework with online laplace approximation,","cited_arxiv_id":null,"evidence_quote":"Defines the global posterior as the product of local posteriors, the exact distribution-level aggregation the AirComp scheme must reproduce."},{"cited_title":"Conflations of probability distributions,","cited_arxiv_id":null,"evidence_quote":"Establishes that the normalized product of distributions is the conflation minimizing information loss, justifying the product-of-Gaussians form."},{"cited_title":"Federated learning over wireless fading channels,","cited_arxiv_id":null,"evidence_quote":"Introduces AirComp for FL over fading channels, giving the superposition-based aggregation concept that the proposed two-phase scheme adapts to posteriors."},{"cited_title":"Optimized power control design for over-the-air federated edge learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the optimized power-control design for AirComp FL whose distortion-minimization approach is extended to the Bayesian setting."},{"cited_title":"Communication-efficient learning of deep networks from decentralized data,","cited_arxiv_id":null,"evidence_quote":"Defines the conventional averaging baseline over which the proposed framework's gains are demonstrated."}],"review_version":1}