Federated learning keeps raw data on phones, sensors and hospital machines and ships only model updates to a server, which avoids the breach risk of pooling everyone's records in one place. One wireless variant goes further. In over-the-air aggregation, every device transmits its analog update at the same moment, and the radio channel adds the signals together physically. The server receives only the noisy sum, never any single device's contribution, and the time to aggregate stays the same no matter how many devices join. This note makes a simple observation: a Bayesian neural network (one whose weights are Gaussian distributions rather than single numbers) combines client posteriors by multiplying Gaussians, and multiplying Gaussians is exactly adding their "natural parameters", the precision (inverse variance) and the precision-weighted mean. The wireless channel therefore computes the global Bayesian posterior for free, in one transmission.
The catch is where the research lies. Recovering a usable mean requires dividing the received precision-weighted mean by the received precision, and both quantities arrive noisy. At low signal-to-noise ratio (SNR), the received precision can approach zero or turn negative. The ratio then explodes, and the network either diverges or becomes wildly overconfident. This is a new instance of the threshold "cliff" seen in other nonlinear estimators. We propose to predict where this cliff sits, to move it with a noise-aware server estimator and a precision-sharing transmission design, and to schedule transmissions so the whole system stays on the right side of the cliff under a latency and energy budget.
Core Claim
Analog superposition over a multiple-access channel (many transmitters sharing one frequency band at once) is the natural transport for probabilistic federated learning. We claim three things. First, its failure mode at low SNR is a predictable breakdown of the precision ratio, not a gradual loss of accuracy. Second, a server that treats channel noise as a known observation model, rather than something to cancel, can keep both accuracy and calibration several dB below the point where naive over-the-air Bayesian aggregation collapses. Third, because each device's precision measures how much evidence it contributes, precision is a better scheduling signal than channel quality alone when trading training progress against delay.
Concretely, client $k$ holds a Gaussian approximate-likelihood factor with per-weight precision $\Lambda_k$ and precision-weighted mean $\eta_k = \Lambda_k \mu_k$. The global posterior has precision $\Lambda = \Lambda_0 + \sum_k \Lambda_k$ and mean $\mu = (\eta_0 + \sum_k \eta_k)/\Lambda$, where subscript $0$ denotes the prior. After power scaling $\alpha$, the server receives
$$\tilde\Lambda = \Lambda + n_\Lambda, \qquad \tilde\eta = \eta + n_\eta, \qquad n_\Lambda, n_\eta \sim \mathcal{N}(0, N_0/\alpha),$$
where $N_0$ is the channel noise power. The plug-in mean $\tilde\eta/\tilde\Lambda$ fails once the probability that $\tilde\Lambda$ falls near zero becomes non-negligible. That happens roughly when $\sqrt{N_0/\alpha}$ becomes comparable to the smallest per-weight precision $\Lambda$ in the network. This gives a testable prediction: the cliff SNR should shift by about 20 dB per decade of typical precision, and should improve as clients are added, because $\Lambda$ grows with the number of clients $K$ while the noise does not.
Research Program
The work builds on three lines of prior work, each playing a different role. Partitioned Variational Inference supplies the method. It writes the global posterior as a prior times client factors and has each client delete its own old factor, forming a "cavity" distribution, before updating so its evidence is not counted twice. It assumes error-free links. How to Combine Variational Bayesian Networks in Federated Learning motivates the design. It found that aggregation rules which keep the posterior's spread low, including the Gaussian product rule we are effectively computing over the air, beat deterministic averaging, while spread-inflating rules hurt. Channel noise is a new source of exactly that inflation, and that paper never considered it. The over-the-air papers supply the physical-layer tools but only for point-estimate weights. COTAF uses time-varying precoding so that noise fades out over rounds. Adaptive Federated Learning Over the Air filters heavy-tailed interference with server-side adaptive moments. Bayesian Over-the-Air Computation and Bayesian Federated Learning over Wireless Networks use transmitted moments as priors for minimum-mean-squared-error denoising of the signal. None of these makes the transmitted object itself a posterior whose variance must remain meaningful after reception. That distinction is our contribution.
The first component is the noise-aware server estimator. Since $N_0/\alpha$ is known from channel estimation, the server forms a posterior over the true $(\Lambda, \eta)$. It uses a truncated prior $\Lambda \ge \Lambda_0$ and centres the prior on the previous round's global posterior. From this it computes a shrunken mean and an honestly inflated variance. It does not clip negative precisions or divide noisy ratios. At high SNR this reduces to exact product-of-Gaussians aggregation. At low SNR it shrinks toward the last trusted posterior and reports extra uncertainty. The second component is precision sharing. Precision is tied across groups of weights (per output neuron or per layer), so far fewer precision symbols are sent than mean symbols. The transmit power saved on the precision stream is spent on those few symbols, which pushes the cliff to lower SNR. It also halves the parameter-doubling cost that mean-field Bayesian networks usually pay. Precision sharing coarsens the posterior, and ablations will measure what that coarsening costs in calibration. The third component preserves privacy structure. The server only ever sees the superposed sum. Each client keeps its own $\Lambda_k, \eta_k$ locally and computes its cavity from the broadcast global posterior. The PVI deletion step therefore survives without the server ever learning an individual factor.
The fourth component is a delay-aware scheduler built on Lyapunov drift-plus-penalty (Huang et al.). In each round it chooses participating devices and transmit power. The penalty term rewards the expected gain in global log-precision, net of the channel noise the transmission would add. Virtual queues enforce long-term energy budgets, and a parameter $V$ trades training progress against queueing delay. This follows Perazzone et al., whose objective is a loss bound for deterministic models. Deferral has a cost that the Bayesian setting makes explicit: a client that waits for a better channel computes its cavity against a stale global posterior. We measure this staleness rather than assume it away. The failure identified by Hu, where waiting for peak channels costs more than it gains, is the regime we expect to see at small $V$.
Baselines
The naive reference is plug-in over-the-air Bayesian aggregation. It sends $\Lambda_k$ and $\eta_k$ at equal power, divides the received quantities, and clips negative precisions to the prior. This baseline is also the ablation that isolates our key ingredient: it is identical to our method except for the noise-aware estimator and precision sharing. We will also ablate each of those two ingredients separately. A second reference is error-free PVI, the ceiling that no channel method should exceed.
The strongest point-estimate competitor is Adam-OTA from Adaptive Federated Learning Over the Air. It reports reaching 70% CIFAR-10 accuracy in approximately 200 rounds, while FedAvgM-OTA failed to reach 40% after 600 rounds. We will also run COTAF (paper), for which no reported CIFAR-10 accuracy number was found on Lacuna. The strongest Bayesian-aggregation competitor for noisy channels is SBFL (paper), which reports 139 rounds to 90% MNIST accuracy versus 638 for signSGD, and about 95% final accuracy. For scheduling, the baseline is the Lyapunov channel-and-energy scheduler of Perazzone et al., which reports up to 8.5x faster wall-clock convergence than uniform selection on CIFAR-10.
All methods share the same simulated fading channel realizations, client partitions, total energy per round, and network backbone. Bayesian variants use the same architecture with mean-field Gaussian weights. Every method receives the same hyperparameter search budget, tuned at a single mid-range SNR and then frozen across the sweep. Beating a baseline means two things: matching or exceeding Adam-OTA and COTAF accuracy at equal energy and wall-clock time, and achieving clearly lower expected calibration error and negative log-likelihood at every SNR. The central requirement is that our method keeps working at least 5 dB below the SNR where the plug-in baseline diverges.
Why not just…?
Why not just send the variational parameters digitally, with error-correcting codes and one device at a time, and run ordinary PVI? This is the most serious threat. At high SNR with few clients it will likely win on accuracy. The argument for analog transmission is latency and privacy. Orthogonal digital access takes time linear in $K$, and its rate collapses at low SNR. Superposition takes constant time and never exposes an individual factor. We will settle this empirically by comparing wall-clock time to target accuracy and NLL, at matched energy, while varying $K$ and SNR. If digital PVI dominates everywhere, the proposal reduces to a negative result about analog Bayesian aggregation, which is still worth reporting.
Why not just train a deterministic model with Adam-OTA and fix calibration afterwards with temperature scaling or a local ensemble? Post-hoc temperature scaling needs a held-out set, and it assigns one global confidence correction. It cannot tell the model's own lack of knowledge apart from input noise, the distinction Cooperative Variance Estimation shows Bayesian networks can make. We will include Adam-OTA plus temperature scaling as a baseline and test out-of-distribution detection, where post-hoc scaling should help least. If it matches our method on ECE and out-of-distribution AUROC, the Bayesian machinery is not justified.
Why not just use SBFL-style MMSE denoising on gradients and skip weight distributions? SBFL denoises point gradients using transmitted moments. It yields a better point estimate, not a posterior, so it offers no uncertainty for downstream use. It also assumes its side-information moments arrive cleanly, which is exactly what fails near the cliff. Our SNR sweep will show whether SBFL's own moment channel breaks down at a similar SNR.
Why not just clip negative received precisions? Clipping is the plug-in baseline. Our prediction is that clipping stops divergence but produces overconfident posteriors near the cliff, because clipped precisions are biased upward. The calibration curve against SNR will show whether this prediction holds.
Proposed Experiments
We will simulate a Rayleigh-fading multiple-access channel with $K \in \{10, 100\}$ clients and Dirichlet non-IID partitions of MNIST, Fashion-MNIST and CIFAR-10, sweeping SNR from −10 to 20 dB. SVHN serves as the out-of-distribution set for CIFAR-10. We will also include MIMIC-III mortality prediction, following PVI's clinical setting, since it matches the privacy-sensitive regime that motivates federated learning. The metrics are accuracy, ECE, NLL, out-of-distribution AUROC, rounds and wall-clock time to target, and energy.
The decisive check is the cliff prediction. We will measure the SNR at which each method's NLL diverges and plot it against median per-weight precision and against $K$. Our analysis predicts roughly 20 dB shift per decade of precision, with the cliff moving to lower SNR as $K$ grows. If the measured cliff does not track precision this way, our account of the failure is wrong. A second check sweeps the precision-sharing granularity to trace the tradeoff between cliff location and calibration. A third sweeps the scheduler's $V$ parameter and records delay, staleness of cavity distributions, and NLL. This tests whether precision-aware scheduling reaches a given NLL with lower delay than channel-only scheduling, and whether deferral-induced staleness erases that gain at small $V$.
References
- Partitioned Variational Inference: A Framework for Probabilistic Federated Learning
- How to Combine Variational Bayesian Networks in Federated Learning
- Over-the-Air Federated Learning From Heterogeneous Data
- Adaptive Federated Learning Over the Air
- Bayesian Over-The-Air Computation
- Bayesian Federated Learning over Wireless Networks
- Utility optimal scheduling in processing networks
- Communication-Efficient Device Scheduling for Federated Learning Using Lyapunov Optimization
- Opportunistic Spectrum Access: Does Maximizing Throughput Minimize File Transfer Time?
- Cooperative Variance Estimation and Bayesian Neural Networks for Disentangling Aleatoric and Epistemic Uncertainties