Distance-Indexed Nested Compression for Wireless-Powered Over-the-Air Learning

Picture a base station that powers its sensors by radio and then collects their model updates over the air. A distant sensor loses twice. It harvests little energy because the power beam has faded by the time it arrives, and its reply fades by the same factor on the way back. The usable signal budget therefore falls roughly as $d^{-2\alpha}$, where $d$ is distance and $\alpha$ is the path-loss exponent. This is the "doubly near-far problem." Standard network designs respond by starving or dropping far devices. That is harmful in federated learning, because those devices may hold data no one else has. Meanwhile, model-compression research has shown that combining several compression methods (fewer dimensions, fewer nonzeros, fewer bits per value) beats pushing any single one to an extreme. This proposal turns the doubly near-far gap into a compression-allocation problem. Each device sends a prefix of a shared, nested low-dimensional code of its update. The prefix length is chosen from its distance and harvested energy. The device spends its scarce energy on fewer, stronger symbols rather than many weak ones.

The central observation is that in over-the-air computation, the analog analogue of bit precision is energy per transmitted symbol. Over-the-air computation (AirComp) means all devices transmit at once and the radio channel adds their signals for free. A far device that sends fewer coordinates can put more energy into each one. Because the coordinate basis is shared and nested, everyone still sums coherently on the coordinates they share. The weakest device then limits accuracy only on the leading coordinates it participates in, not on the whole update. Two things would be new: deriving how a device's optimal code length should scale with its distance, and showing that this distance-indexed allocation beats both dropping far devices and handing everyone the same compression.

Core Claim

The closest existing work is Optimized Joint Beamforming for Wireless Powered Over-the-Air Computation. It jointly optimizes energy beamforming, receive beamforming and device power for wireless-powered AirComp. It finds that the optimal power control is truncated channel inversion: energy-rich devices invert their channel, while far devices simply transmit at their harvested maximum. The data, however, are uncompressed scalars, and every device sends the same full-length signal. Its reported benchmark scheme is "limited by the worst-case device," and the joint design mitigates this only through power and beams.

We add a second control: the length $r_k$ of the code device $k$ transmits. Let $g_k \in \mathbb{R}^D$ be device $k$'s model update. Let $U$ be an orthonormal basis shared by the server and all devices, regenerated each round from a common seed or from the previous global update. Device $k$ sends the first $r_k$ coefficients $U_{r_k}^\top g_k$, using harvested energy $E_k$ spread over $r_k$ channel uses. To first order, its contribution to aggregation error is

$$ \mathcal{E}_k(r_k) \approx \big\|(I - U_{r_k}U_{r_k}^\top)\, g_k\big\|^2 + c\, \frac{r_k\, \sigma^2}{E_k\, \beta_k}, $$

The symbols are defined as follows. The first term is the update energy lost by truncating the code. In the second term, $\sigma^2$ is receiver noise, $\beta_k \propto d_k^{-\alpha}$ is the uplink channel gain, $E_k \propto d_k^{-\alpha}$ in the linear regime of the energy harvester, and $c$ absorbs alignment constants. Suppose the truncation tail decays like $r^{-\gamma}$, which is a power-law singular spectrum. Balancing the two terms then gives

$$ r_k^\star \propto d_k^{-2\alpha/(1+\gamma)}, $$

so the code length should shrink polynomially with distance, at a rate set jointly by the propagation environment and by how compressible the updates are.

Below the harvester's turn-on threshold, $E_k = 0$ and $r_k^\star = 0$. At that point the device is "pruned" from the round, and the base station must decide whether to steer energy toward it to lift it over the threshold. The proposal therefore integrates three levers, one per device: dimension (rank), precision (energy per symbol), and participation (zero rank). All three are tied to physical position.

Research Program

The first component is the aggregation protocol. Coordinate $j$ receives a superposition from exactly the devices with $r_k \ge j$. The server normalizes each coordinate by its own participant count and by its own receive scaling factor. As a result, the scaling on high-index coordinates is no longer limited by far devices, which are absent there. The scaling on leading coordinates benefits from far devices concentrating their energy.

Two biases need handling. Leading coordinates are averaged over more devices than trailing ones. And far devices' trailing information is systematically missing. We will analyze both effects and test a correction in the spirit of the cooldown reweighting of Over-the-Air Federated Learning with Energy Harvesting Devices. That paper scales each device's update by the time since it last transmitted, under ambient rather than radio-powered energy, so it never faces the $d^{-2\alpha}$ coupling.

The second component is the allocation algorithm. We will extend the alternating optimization of the Zhang et al. paper, which handles energy beams, receive beams and power, with a fourth block that sets integer code lengths $r_k$ by one-dimensional search on the error model above. The spectrum exponent $\gamma$ is estimated online from the server's aggregate.

Because the harvester is nonlinear, the energy-beam step must trade spreading power, which pushes edge devices below threshold, against concentrating it on a few edge devices so that they clear threshold with a short code. This addresses the open problem in the research file that linear harvesting models hide edge blackouts. We model the harvester with a threshold-and-saturation curve, with parameters taken from the Southampton measurements.

The third component is the choice of basis. Several compression methods supply candidates, each with a different role. SVD-LLM V2 shows that truncating singular vectors weighted by activation statistics, with layer-wise heterogeneous ratios, beats uniform truncation, and that SVD composes well with quantization. That result motivates allocating rank non-uniformly across layers as well as across devices. Bayesian Bits supplies the nested idea in the precision domain: a 2-bit base plus gated residual refinements, with a 0-bit gate acting as pruning. Our nested prefix is the dimension-domain counterpart, and its gates are set by physics rather than learned. 3BASiL shows that sparse-plus-low-rank beats either alone for weights. Following it, we test a variant in which near devices append a sparse residual on top of their prefix. Far devices cannot afford one.

Basis computation must run on the server or come from a seed. A battery-less device cannot afford an on-device SVD. We will account for device-side projection energy, roughly $D r_k$ multiply-accumulates, inside $E_k$.

Baselines

The do-nothing reference is uncompressed full-dimension AirComp federated averaging with truncated channel inversion and fixed, evenly split energy beams. We also report an error-free channel as the ceiling.

The strongest physical-layer competitor is the joint beamforming and power-control design of Optimized Joint Beamforming for Wireless Powered Over-the-Air Computation, applied to full-length updates. It is the strongest because it already attacks the doubly near-far gap with the same wireless-powered model; we differ only by adding code length. We run it with the same antennas, power cap, channel draws and alternating-optimization iteration budget. That paper gives its MSE only in figures, so there is no reported numeric result on Lacuna for it. We reimplement it, and beating it means lower aggregation MSE and higher final accuracy at equal energy.

The strongest learning-side competitor is Channel-Energy-Data client selection with semidefinite-relaxation beamforming from Joint Client Selection and Receive Beamforming for Over-the-Air Federated Learning With Energy Harvesting. It reports "nearly 80% test accuracy" on unbalanced Fashion-MNIST, against below 60% for simpler energy-aware schemes. It is the natural "just drop the weak devices" answer. We port it to radio-powered energy and use its unbalanced Fashion-MNIST protocol. Beating it means exceeding its accuracy under our radio-powered geometry, especially on classes held mostly by far devices.

The digital competitor is time-division uplink with per-device bit-width chosen from the energy budget, scheduled with the proportional-fairness Lyapunov method of Online Optimization of Wireless Powered Mobile-Edge Computing for Heterogeneous Industrial Internet of Things. That paper reports a Jain's fairness index near 1.0, but it moves abstract bits, not model updates. It receives the same total energy and frame time as our method.

Two ablations isolate the key ingredient. In the first, all devices share one tuned code length; this tests whether heterogeneity by distance matters, or only compression itself. In the second, code lengths follow the distance rule but energy per symbol is not reallocated; this tests whether the gain comes from buying precision with saved channel uses. A further ablation removes threshold-aware energy beaming.

Why not just…?

Why not just drop far devices, as client selection does? Dropping far devices is the most serious threat. If data are independent and identically distributed, losing far devices may cost little. The 80% versus below-60% result in the client-selection paper also shows that careful selection already recovers much of the loss. We therefore expect selection to match us in the IID split. The proposal stands or falls on the geography-correlated non-IID split, where far devices hold distinct classes. There, dropping them should collapse accuracy on those classes, while short codes should preserve it. If selection still matches us there, the nested code adds nothing, and we will report that.

Why not just adapt quantization bit-width per device digitally? The required budget spans a large range. With $\alpha = 2.94$, as in the Southampton non-line-of-sight measurements, doubling distance cuts the budget by about $2^{5.9} \approx 60\times$. Bit-width alone covers at most about $16\times$ between 32 and 2 bits before the low-bit accuracy cliffs that SVD-LLM V2 documents. Time-division channel use also grows with the number of devices, whereas AirComp's does not. The digital baseline above measures directly whether this reasoning holds at realistic device counts.

Why not just add antennas or better energy beams? The 4 W EIRP cap in the Southampton dataset bounds the total radiated power, regardless of how the beam is shaped. The full-length joint-beamforming baseline is exactly this alternative at its best. We will sweep antenna count to find where the gap from code-length allocation closes.

Why not give everyone the same low-rank code? If update spectra are steep, a single short code may suffice for all devices. The uniform-rank ablation tests this. The scaling law also predicts when it should fail: as $\gamma$ falls or $\alpha$ rises, the optimal lengths spread further apart.

Proposed Experiments

The simulation uses a single multi-antenna base station with 40 single-antenna devices placed between 1 and 10 m. Path-loss exponents are 2.04 and 2.94, the transmit cap is 4 W EIRP, receive antennas have 6 dBi gain, and fading is Rician. The harvester has a threshold and saturation. Tasks are CIFAR-10 with 40 devices, following Over-the-Air Federated Learning with Energy Harvesting Devices, which reports a 5–10% accuracy loss from noisy channels relative to error-free training. We also run Fashion-MNIST in IID, unbalanced, and distance-correlated non-IID splits.

Metrics are final accuracy, rounds and energy to reach a target accuracy, per-coordinate aggregation MSE, accuracy on far-held classes, and participation rate by distance ring.

Three checks are decisive. First, the fitted exponent of the chosen $r_k^\star$ against $d_k$ should match $2\alpha/(1+\gamma)$ when $\alpha$ is swept between the two measured values, and should change slope as $\gamma$ is varied across architectures. Second, on the distance-correlated split, accuracy on far-held classes should exceed both the client-selection and the full-length beamforming baselines at equal energy. Third, as the device count grows under the fixed power cap, the threshold-aware variant should avoid the edge blackout that the research file identifies. In that regime, adding devices raises error for full-length AirComp, contrary to the trend reported by Zhang et al. under linear harvesting. If the first check fails, the analysis is wrong. If the second fails, the method is not worth building.