Federated learning trains one shared model across many phones or sensors without collecting their data. Each round, devices compute an update on their local data and send it to a server, which averages the updates. Over a real cellular network, some of these uploads fail. The failures are not random. Devices far from a base station, or near busy neighboring cells, lose their uplink far more often than devices close to a tower. When what a device's data looks like also depends on where it is (a regional dialect, local weather, the kinds of activity common in one country), the server keeps receiving updates that over-represent the well-covered regions. The model then settles on the optimum for those regions rather than for the whole population. This proposal argues that the size of this bias can be predicted in closed form. It uses stochastic geometry, the branch of applied probability that models base stations and devices as random point patterns and derives the probability that a link at a given distance succeeds. The same prediction gives a correction that needs only one distance or signal-strength reading per device.
The core observation is simple. If device $i$ gets its update through with probability $p_i$, then at the global optimum the server's expected aggregate is off by the covariance, across devices, between $p_i$ and the device's update. Stochastic geometry supplies $p_i$ as an explicit function of the device's distance to its base station, the density of interferers, and the decoding threshold. The bias therefore becomes a computable quantity that depends on network design parameters. We then correct it by reweighting with the analytic success probability, and test whether this beats the obvious data-driven fix.
Core Claim
Two existing results do not fit together. On the Unreasonable Effectiveness of Federated Averaging with Heterogeneous Data shows that on naturally heterogeneous data such as FEMNIST and StackOverflow, the average of client updates at the global optimum is close to zero. Under this "client consensus" property, heterogeneity does not hurt plain federated averaging. That result assumes every client is equally likely to report. Stochastic-geometry studies of cellular networks, such as Impact of Dual Slope Path Loss on User Association in HetNets and Downlink Multi-Antenna Heterogeneous Cellular Network With Load Balancing, show the opposite: link success varies sharply with position, especially past the distance where path loss steepens. Our claim is that spatially filtered participation can break client consensus even when the full population satisfies it. The resulting error floor equals
$$ b(\mathbf{w}^*) \;=\; \frac{\sum_{i=1}^{N} (p_i - \bar p)\,\Delta_i(\mathbf{w}^*)}{\sum_{j=1}^{N} p_j}, $$
where $\Delta_i(\mathbf{w}^*)$ is client $i$'s local update evaluated at the global optimum, $p_i$ is its upload success probability, and $\bar p$ is the mean success probability. The floor vanishes when success is uncorrelated with data. It grows as outage becomes more position-dependent.
For interferers forming a Poisson point process (points scattered independently at density $\lambda$) with Rayleigh fading, the success probability for a device at distance $r$ takes the standard form
$$ p(r) \;=\; \exp\!\left(-\frac{\theta r^{\alpha}\sigma^2}{P}\right)\,\mathcal{L}_I\!\left(\frac{\theta r^{\alpha}}{P}\right). $$
Here $\theta$ is the decoding threshold, $\alpha$ the path-loss exponent, $\sigma^2$ the noise power, $P$ the transmit power, and $\mathcal{L}_I$ the Laplace transform of the aggregate interference given in the research file. This expression plays two roles. It predicts $b$ from design parameters before training begins, and it supplies inverse-probability weights $1/p(r_i)$ from a single distance or received-power reading. The weights are clipped at a floor $p_{\min}$, because unbounded weights make the variance explode. The analysis will give the bias–variance tradeoff in $p_{\min}$ explicitly.
Relation to Prior Work
Over-the-air federated learning is the setting that motivates this work. In this approach, devices transmit analog signals simultaneously so the channel itself sums their updates. Existing analyses all assume a single isolated cell or treat interference abstractly. Over-the-Air Federated Learning From Heterogeneous Data removes channel noise with time-varying precoding inside one multiple-access channel. Rethinking Federated Learning Over the Air: The Blessing of Scaling Up proves that fading averages out as the number of clients grows. Adaptive Federated Learning Over the Air takes the heavy-tail index of interference as an exogenous parameter. None of the three lets participation depend on geometry. We keep their aggregation schemes and add the one geometric fact they omit. The scaling result in particular is complementary rather than contradictory. Channel hardening averages out small-scale fading across clients. It does not average out a systematic dependence of success on distance, which remains even as $N$ grows.
Federated Learning Over Wireless Fading Channels compares digital and analog uplinks in one cell with 25 devices. We use both uplink types, but in a multi-cell layout. Over-the-Air Federated Edge Learning With Hierarchical Clustering and Federated Learning in UAV-Enhanced Networks: Joint Coverage and Convergence Time Optimization optimize fixed placements, the first deterministically and the second with bandits. Neither studies the learning bias that random placement induces.
Two lines of work constrain the design. Towards Understanding Biased Client Selection in Federated Learning shows that deliberately biasing selection toward high-loss clients speeds training, at the cost of an error floor. Our analysis treats channel-induced bias as an unintended version of the same phenomenon. That work also warns that some bias can help early in training, so our correction must be compared against loss-biased selection and not only against plain averaging. The consensus paper warns that on artificially partitioned data, drift dominates regardless of channel. So the benchmarks must separate channel-induced bias from ordinary drift.
SGFusion: Stochastic Geographic Gradient Fusion in Federated Learning groups geographic zones by how similar their label distributions are, and reports gains on a 33-country heart-rate dataset. That dataset supplies the location-correlated data we need, but SGFusion models no radio link at all.
The stochastic-geometry machinery comes from the cited network papers. We reuse their Laplace-transform coverage expressions, their dual-slope path loss model, and their multi-tier point processes. We do not re-derive them.
Research Program
The analytic part bounds the stationary error of federated averaging under independent, position-dependent upload success. The bound decomposes into the covariance term $b$ and an effective-participation variance term, both expressed through $p(r)$ and the distribution of client distances under the point-process model. Two regimes are treated. The first is a digital uplink, where success means $\mathrm{SINR}>\theta$. The second is over-the-air aggregation with truncated channel inversion, where the truncation threshold plays the role of $\theta$; this is the dropout problem listed as open in the file. For clustered users, modeled as a Thomas cluster process (users scattered around random hotspot centers), the distance distribution changes but the covariance structure does not.
The algorithmic part is geometry-weighted aggregation: each received update is scaled by $1/\max(p(r_i), p_{\min})$, with $p(r)$ evaluated from the closed form using the network's density and threshold. We also study the design question the bound makes answerable. Given a target bias, which pairs of base-station density and threshold $(\lambda, \theta)$ minimize energy or rounds?
Baselines
The first reference point is federated averaging over an ideal channel where every upload succeeds. It sets the accuracy ceiling. The second is naive federated averaging that simply averages whatever arrives, which is the realistic default. Beating naive averaging means closing a measurable fraction of the gap between these two.
The strongest direct competitor is empirical inverse-propensity weighting. The server tracks each client's historical success rate and weights by its inverse. It needs no geometry at all. We will run it with the same rounds, the same clients, and a tuned smoothing window and clip level, using the same tuning budget as our $p_{\min}$. No reported result found on Lacuna for this baseline in wireless federated learning. Our method must match it in static settings and beat it in final accuracy and error-floor size when clients move or join fresh.
For selection, the competitor is Power-of-Choice. The cited paper reports reaching target loss up to 3 times faster and up to 10% higher test accuracy than random selection on FMNIST, CIFAR-10, and Sent140 (Towards Understanding Biased Client Selection in Federated Learning). We run it under the same outage model, with the candidate-pool size tuned, since in a real network its chosen high-loss clients may fail to upload. For drift correction, the competitor is SCAFFOLD: Stochastic Controlled Averaging for Federated Learning under identical outage. For over-the-air aggregation, the competitor is COTAF precoding with truncated inversion (Over-the-Air Federated Learning From Heterogeneous Data).
On the heart-rate data, the relevant reference is that SGFusion reports a 3.23% average utility improvement over its D-ZGD baselines (SGFusion: Stochastic Geographic Gradient Fusion in Federated Learning). That number was obtained without any channel model, so it serves as context rather than a head-to-head target.
The key ablation replaces analytic $p(r)$ with a constant, recovering naive averaging, and with a mis-specified $p(r)$ using the wrong density. This isolates whether gains come from geometry-accurate weights or from reweighting of any kind.
Why not just…?
Why not just estimate each client's success rate empirically and reweight? This is the real threat. For static clients observed over many rounds, it should converge to the same weights as our analytic ones, and we expect to tie there. The geometric formula has an advantage only when history is uninformative: under mobility, where a client's distance changes faster than its success count accumulates; for newly joining clients; and for design-time prediction before deployment. The mobility and cold-start experiments settle this. If empirical weighting matches us even under realistic user speeds, then the proposal's contribution narrows to the predictive bound and design rule, and we will say so.
Why not just raise power or retransmit until everyone succeeds? Retransmission removes the bias but increases latency and energy, and edge devices are exactly the ones with the worst links. The UAV study documents how slow links stall synchronous rounds (Federated Learning in UAV-Enhanced Networks: Joint Coverage and Convergence Time Optimization). Raising power in a multi-cell setting raises everyone's interference, which the Laplace-transform term makes explicit. We compare accuracy per unit of wall-clock time and energy against a retransmit-to-success scheme.
Why not just use a drift-correcting method like SCAFFOLD? Drift correction targets the mismatch between local and global optima under full participation. Its control variates for absent clients go stale, and it has no mechanism to reweight systematically missing regions. It may still partly help, so it appears as a baseline and in combination with our weights.
Why not rely on the blessing of scaling up? Channel hardening averages out small-scale fading across clients. It does not average out a systematic dependence of success on distance, which persists at any $N$. We test this directly by sweeping $N$ from 10 to 1000 at fixed geometry.
Proposed Experiments
Base stations are drawn from a Poisson process; a repulsive Ginibre-like placement is used as a robustness check on the Poisson assumption. Users are drawn from a Thomas cluster process, with single-slope and dual-slope path loss.
Three data regimes separate the effects. First, the heart-rate prediction dataset (956 users, 33 countries) with an LSTM, where countries map to clusters so that data is genuinely location-correlated. Second, CIFAR-10 with ResNet-18, where Dirichlet(0.1) label mixtures are assigned by cluster rather than at random. Third, FEMNIST with writers placed uniformly at random, as a null control in which the covariance term should vanish and no method should beat naive averaging.
Metrics are final test accuracy or error, worst-region accuracy, measured distance from the ideal-channel solution, rounds to target, and uplink energy.
Three decisive checks follow. First, the predicted floor $b$ should track the measured gap between naive averaging and the ideal channel across a sweep of $\lambda$ and $\theta$; a poor correlation falsifies the core claim. Second, on the FEMNIST null, the gap should be near zero. Third, empirical inverse-propensity weighting should tie with our method for static users and lose under mobility and cold start. If that final separation fails to appear, the method is not what the work contributes, and the remaining result is the analytic bias bound.
References
- On the Unreasonable Effectiveness of Federated Averaging with Heterogeneous Data
- Impact of Dual Slope Path Loss on User Association in HetNets
- Downlink Multi-Antenna Heterogeneous Cellular Network With Load Balancing
- Over-the-Air Federated Learning From Heterogeneous Data
- Rethinking Federated Learning Over the Air: The Blessing of Scaling Up
- Adaptive Federated Learning Over the Air
- Federated Learning over Wireless Fading Channels
- Over-the-Air Federated Edge Learning With Hierarchical Clustering
- Federated Learning in UAV-Enhanced Networks: Joint Coverage and Convergence Time Optimization
- Towards Understanding Biased Client Selection in Federated Learning
- SGFusion: Stochastic Geographic Gradient Fusion in Federated Learning
- SCAFFOLD: Stochastic Controlled Averaging for Federated Learning