Steering Wireless Power Toward the Nodes That Hold a Learning Network Together

Picture a network of small battery-free sensors that train a shared model with no central server: each sensor takes a gradient step on its own data, then averages its parameters with its radio neighbors. The sensors run on radio-frequency energy that a few multi-antenna transmitters beam at them. Today those transmitters aim for physical fairness, splitting power so that every sensor harvests about the same amount. This proposal argues that fairness is the wrong target. How fast the network agrees on a good model depends very unevenly on which sensors are awake. A sensor that bridges two clusters, or one that holds unusual data, matters far more than a redundant sensor in a dense patch. We propose to aim the energy beams so that they directly speed up the learning, using a convergence bound for decentralized learning under random node participation as the objective that the beamformers optimize.

In two sentences: energy beamforming decides which nodes of a decentralized learner can afford to participate, so it effectively reshapes the network's communication graph. We optimize the beams for that graph's mixing speed and for coverage of heterogeneous data, not for equal harvested power, and the optimization runs with only neighbor-to-neighbor messages.

Where the two literatures stand

Work on spatial energy focusing treats the beam as a controllable actuator. Multi-antenna transmitters concentrate power on receivers (Wireless Networks with RF Energy Harvesting: A Contemporary Survey). Cluster-based schemes beam power to relay nodes to maximize the worst user's throughput. Multi-Antenna Enabled Cluster-Based Cooperation in Wireless Powered Communication Networks reports about 40% higher max-min throughput than cooperation without energy beamforming, using a centralized solver. Distributed variants exist, but each stops short of what this proposal needs. Distributed energy beamforming with one-bit feedback and its companion Distributed Wireless Power Transfer with Energy Feedback align transmitter phases toward a single receiver from one bit of harvested-power feedback. Optimal Distributed Beamforming for MISO Interference Channels passes beamformers around a logical ring. In every case the objective is a physical-layer quantity, either received power or rate, and the receivers do nothing with the energy except harvest it.

On the optimization side, On the Convergence of Decentralized Gradient Descent analyzes decentralized gradient descent, the "step locally, then average with neighbors" rule. It shows that convergence speed is governed by the spectral gap of the mixing matrix, a measure of how quickly averaging spreads information across the graph, and it implicitly assumes every node is always powered. The closest work at the intersection is Decentralized Federated Learning With Energy Harvesting Devices. It motivates this proposal by deriving a convergence bound with a persistent error term caused by nodes that drop out for lack of energy, and it learns participation policies that recover accuracy. Its energy, however, comes from solar irradiance traces, an exogenous process no one can steer. The open question is what happens when energy becomes a decision variable. This proposal builds on that bound's structure, but the decision variable is where the power goes, not only when to spend it.

Core Claim

Let node $i$ participate in a round with probability $p_i$, the probability that its harvested energy covers one gradient step plus one exchange with its neighbors. Harvested energy is $E_i = \psi\big(\sum_k h_{ki}^{H} Q_k h_{ki}\big)$. Here $Q_k \succeq 0$ is the transmit covariance of energy transmitter $k$, with trace bounded by its regulatory power cap, $h_{ki}$ is the channel from $k$ to $i$, and $\psi$ is a nonlinear rectifier curve with a turn-on threshold and saturation. Under random activation, the expected graph Laplacian is $\bar L(p) = \sum_{(i,j)\in\mathcal E} w_{ij}\, p_i p_j\, (e_i - e_j)(e_i - e_j)^{\top}$, where $w_{ij}$ is the mixing weight on the link between nodes $i$ and $j$, and $e_i$ is the $i$-th standard basis vector. We choose the beams to solve

$$\max_{\{Q_k\}} \; \lambda_2\big(\bar L(p(Q))\big) \;-\; \mu \sum_i (1 - p_i)\, \zeta_i^2 ,$$

where $\lambda_2$ is the algebraic connectivity (the spectral-gap quantity that controls the averaging speed), $\zeta_i^2$ measures how far node $i$'s local gradient departs from the network average (its data heterogeneity), and $\mu$ trades the two terms. The second term mirrors the persistent-error term in the energy-harvesting bound of Decentralized Federated Learning With Energy Harvesting Devices. Absent nodes with atypical data bias the consensus point.

Two properties make the problem decentralizable. First, the gradient of $\lambda_2$ with respect to $p_i$ equals $\sum_j w_{ij} p_j (v_i - v_j)^2$, where $v$ is the Fiedler vector (the eigenvector associated with $\lambda_2$). Each node needs only its own entry of $v$ and its neighbors' entries. Nodes can estimate those entries by a few rounds of power iteration piggybacked on the parameter-averaging messages they already send. Second, each transmitter's constraint (positive semidefinite, bounded trace) is local, so its projection is a small eigenvalue clip. The heavier approximate-projection machinery of Approximate projections for decentralized optimization with SDP constraints is therefore unnecessary. We note this explicitly because the coupling between transmitters enters only through the objective.

The falsifiable prediction is this: on graphs with bottleneck nodes and on non-IID data, beams that optimize this objective will reach a target accuracy in fewer rounds than max-min-fair beams with the same transmit power. On homogeneous random graphs with IID data, the two should roughly tie.

Research Program

The work has three parts. The first is analysis. We extend the random-activation convergence bound so that $p$ is an explicit function of the beams through the nonlinear rectifier $\psi$. This addresses the linear-harvesting simplification flagged in Joint Offloading and Computing Optimization in Wireless Powered Mobile-Edge Computing Systems. We then characterize when the bound-optimal allocation differs from max-min fairness. Our conjecture is that the gap grows with the variance of node betweenness and of $\zeta_i$.

The second part is the algorithm. It has two timescales. On the fast timescale, nodes run ordinary decentralized gradient descent rounds. On the slow timescale, receivers report local estimates of $p_i$, $v_i$ and $\zeta_i$ to the energy transmitters within radio range, and each transmitter takes a projected gradient step on its $Q_k$.

Channel knowledge is the known weak point of distributed beamforming. We study two regimes. In the first, channels are estimated from the sensors' harvested-power feedback in the style of the one-bit protocol. In the second, estimates are deliberately corrupted, to measure robustness. Both nodes and transmitters are simulated using the physical parameters from Dataset for "Energy-neutral Wireless-powered Networks": 4 W EIRP cap, 6 dBi receive gain, and path-loss exponents of 2.04 for line of sight and 2.94 for non-line of sight.

Baselines

The naive reference point is isotropic energy broadcast. Each transmitter radiates the same total power with no spatial focusing, and nodes run standard decentralized gradient descent (On the Convergence of Decentralized Gradient Descent) whenever their batteries allow. This baseline measures what any focusing buys. A second simple reference is beamforming that maximizes total harvested power. It tends to pour energy onto nodes near the transmitters, and it shows what happens when physical efficiency is optimized with no regard for the graph.

The strongest physical-layer competitor is max-min-fair energy beamforming in the spirit of Multi-Antenna Enabled Cluster-Based Cooperation in Wireless Powered Communication Networks. We solve it centrally with full channel knowledge, which deliberately gives it an advantage over our decentralized method. It is the strongest competitor because fairness is the standard design goal and plausibly already keeps everyone awake. Beating it means fewer rounds to target accuracy, or higher accuracy at a fixed round budget, at equal total transmit energy, with the margin exceeding seed-to-seed variation over at least five seeds.

The strongest learning-side competitor is the two-hop participation policy of Decentralized Federated Learning With Energy Harvesting Devices. Its reported Fashion-MNIST accuracy is 84.2% under IID data and about 80% under non-IID data. The paper's one-hop policy reaches 73.3% and about 62%, and no communication reaches 75.9% (IID). Those figures were obtained with solar energy, so they are not directly comparable. We will reimplement the policy and feed it energy from max-min-fair beams. We will also stack it on top of our beams, because the two approaches control different things: when to spend energy versus where energy lands.

Two further comparisons are an oracle and a cheap heuristic. The oracle is a centralized solver of our objective with exact channels, which serves as an upper bound. The heuristic allocates power in proportion to static betweenness centrality or degree.

The key ablations remove the heterogeneity term ($\mu = 0$), remove the connectivity term, and replace the nonlinear rectifier with a linear one at design time while keeping the true nonlinear curve at test time. All methods share the same data partitions, graphs, energy budget, round budget and learning-rate tuning grid.

Why not just…?

Why not just use max-min-fair energy beamforming and run ordinary decentralized learning on top? This is the real threat. If every node harvests enough to participate almost always, then $p_i \approx 1$ everywhere and our objective has nothing left to improve. The proposal therefore matters only in the energy-scarce regime, where the power cap and path loss leave some nodes underpowered. The measured non-line-of-sight exponent of 2.94 in the Southampton dataset suggests that regime is common at room scale. The experiment settles the question by sweeping transmitter power from scarce to abundant. We expect the advantage to vanish at the abundant end, and we will report the crossover point honestly.

Why not just a centrality heuristic, giving more power to bridge nodes? It may capture much of the gain on static graphs. That is why it appears as a baseline and not a straw man. It cannot account for data heterogeneity, rectifier saturation (doubling power to a saturated node is wasted), or graphs whose bottlenecks shift as nodes drop out. If the heuristic matches our method on non-IID data with nonlinear harvesting, the contribution shrinks to explaining why the heuristic works, and we would say so.

Why not just tune the transmitters' phases by random perturbation, keeping any change that lowers the training loss? This is a parallel, zeroth-order cousin of the one-bit feedback protocols. It needs no model of $p$ or $\lambda_2$. Its feedback signal, however, is learning progress, which is noisy and arrives slowly compared with harvested power. The one-bit work itself found its random-perturbation variant slower than structured adaptation even with the cleaner signal of harvested power. We include this method as a baseline with an equal number of perturbation rounds and measure its convergence directly.

Proposed Experiments

The first testbed is the synthetic regression setting of On the Convergence of Decentralized Gradient Descent: 100 agents, connectivity 0.3, Gaussian data. Here the predicted rate can be checked against the measured rate and the bound's tightness assessed. We then add two-cluster graphs joined by a few bridge nodes. The learning benchmarks are Fashion-MNIST and CIFAR-10, each under IID and Dirichlet non-IID partitions, as in Decentralized Federated Learning With Energy Harvesting Devices. No reported result found on Lacuna for CIFAR-10 accuracy in that setting. Room-scale deployments place two to six energy transmitters with four to sixteen antennas each, using the dataset's path-loss and EIRP parameters.

The metrics are rounds to target accuracy, final accuracy at a fixed round budget, total radiated energy, measured $\lambda_2$ of the realized activation graph, and the fraction of nodes that never participate.

The decisive checks are three. First, on bridged graphs with non-IID data in the energy-scarce regime, our beams must beat both centralized max-min fairness and the centrality heuristic. Second, on homogeneous graphs with IID data, the gap should close; if it does not, our explanation is wrong. Third, removing the heterogeneity term should hurt on non-IID data but not on IID data, which isolates the term borrowed from the energy-harvesting bound. A further check is that the decentralized Fiedler-vector estimate, under five power-iteration steps per slow round, recovers most of the oracle's gain. If it does not, the method is only centrally implementable, and the claim must be weakened accordingly.