Quantum federated learning trains a shared quantum model across many devices. Each device runs a small parameterized quantum circuit, which is a sequence of quantum gates whose rotation angles are the trainable weights. It estimates gradients by repeatedly running and measuring that circuit, and it sends only the classical angles to a server for averaging. Each repeated run is called a "shot". Every shot costs energy and time, and fewer shots mean noisier gradients.
Now suppose the devices are powered by radio-frequency energy that a base station beams to them. Real rectifiers, the diode circuits that turn radio waves into direct current, behave very non-linearly. They harvest almost nothing below a turn-on threshold, and conversion efficiency falls to about 0.4% at −40 dBm against more than 50% at −5 dBm (Lu et al.). Above a ceiling they saturate and harvest nothing extra.
This note proposes treating the harvested energy budget as a measurement budget. The base station then decides two things together: where to beam power, and how many shots each client spends. The curve's shape matters for that decision. Because the curve is convex near its threshold, focusing energy on one client at a time can beat spreading it thinly. A simple linear energy model cannot see this effect.
Core Claim
In quantum federated learning with wirelessly powered clients, the server faces a choice between two sources of error. Variance falls when shots are concentrated on clients that are cheap to power. Bias falls when every client's data eventually contributes, which matters when data differ across clients. The non-linear harvesting curve moves this trade-off in a specific way.
Saturation caps how much a nearby client can usefully absorb. The dead zone means a distant client cannot slowly accumulate energy from weak, spread-out beams. It gets nothing unless power is concentrated on it for a while.
We claim that a scheduler which models the curve explicitly will reach a target accuracy using fewer transmitted joules and waste less energy than the best existing composition of methods. It does so by time-sharing concentrated energy beams and assigning shots with inverse-variance weights. The existing composition we compare against is energy-queue scheduling plus post-hoc noise filtering.
We further claim that this advantage appears mainly when client data differ across clients. If a simpler baseline matches us there, the claim is false.
Relation to Prior Work
The file contains no work that joins these two directions, so each piece of prior work plays a separate role.
On Federated Learning with Energy Harvesting Clients motivates the proposal and supplies the scheduling baseline. It models each device's battery, sets the learning rate in proportion to the square root of the number of active clients, and shows that a "myopic" policy, which picks the clients with the fullest batteries, beats greedy and round-robin selection. However, it uses idealized random energy arrivals, classical networks, and data drawn from the same distribution on every client. We keep its battery-memory model but replace the energy arrivals with a physical rectifier curve that the base station can steer. We also make the quantity being scheduled shots rather than participation alone.
Sporadic Federated Learning Approach in Quantum Environment to Tackle Quantum Noise (SpoQFL) supplies the strongest quantum-side baseline. It also constrains our design. It estimates each client's noise-induced gradient deviation, shrinks the update by an exponential factor, and drops it entirely above a threshold. It assumes unlimited power. On a harvesting client, dropping an update after the shots have been paid for throws that energy away. Cutting shots to save energy raises the deviation and makes a drop more likely. Our design moves that decision to before the shots are spent.
Mixed-Timescale Beamforming and Power Splitting for Massive MIMO Aided SWIPT IoT Network supplies the physical-layer method. It jointly optimizes base-station beamforming and device power splitting under a saturating harvesting model, working on two timescales to limit signaling overhead. We reuse its non-linear constraint and its two-timescale structure. It optimizes communication utility, not learning.
Streaming IoT Data and the Quantum Edge supplies a regression workload and a centralized reference point. It runs variational quantum regression on the Seoul Bike Sharing data, with principal-component analysis (PCA) reducing the features to 7, and reports an MSE of 0.0209. It assumes tethered power and no federation.
We stay with piecewise and sigmoidal harvesting models (Do et al.). We do not use the higher-order waveform models of Zuhra et al., because our question is about scheduling rather than waveform design.
We must address one physical objection directly. Present-day quantum processors are not powered by microwatts. We therefore do not claim that a harvested budget runs a cryogenic processor. Instead, the energy cost of a shot is a parameter that stands in for the client-side cost of preparing, running, and reading out one circuit. We sweep that parameter over several orders of magnitude. The result must hold across the sweep, or be reported as holding only in a stated range.
Research Program
The model works as follows. Client $k$ receives RF power $p_k$ from the base station and harvests $\Phi(p_k)$. Here $\Phi$ is a logistic curve with a turn-on threshold and a saturation ceiling, fitted to the efficiency figures in Lu et al..
To compute a gradient, the client spends $2 P s_k$ circuit runs, where $P$ is the number of circuit parameters and $s_k$ is the number of shots per evaluation. The factor of two comes from the parameter-shift rule, which evaluates the circuit twice per parameter. Each circuit run costs $c$ joules, and sending the angles costs a further $E_{\text{tx}}$. The battery carries charge over from round to round, as in Shen et al..
Shot noise makes each client's gradient variance roughly $\sigma^2/s_k$. This is the standard behavior of an average of $s_k$ independent measurement outcomes. The server aggregates with weights $w_k \propto s_k$, which gives an aggregate variance of $\sigma^2 / \sum_k s_k$. Under data that differ across clients, it also accrues a bias term, which grows with the time since each client last contributed.
The per-round control problem is to choose base-station power allocations $\{p_k\}$ and shot counts $\{s_k\}$ that minimize
$$ \frac{\sigma^2}{\sum_k s_k} + \lambda \sum_k \pi_k \, a_k , $$
subject to battery dynamics and a total transmit-power budget. In this expression, $\pi_k$ is client $k$'s share of the data, $a_k$ is the number of rounds since client $k$ last contributed, and $\lambda$ trades bias against variance.
We solve the problem on two timescales. On the slow timescale, the server plans power time-sharing over a window of rounds. It uses successive convex approximation on the sigmoid, following the approach of Chen et al. (MJBP). On the fast timescale, each client picks $s_k$ from its battery level before it runs any circuit.
On the analysis side, we will prove a simple structural result. When the sigmoid is convex below its inflection point, allocating the power budget to one client at a time dominates spreading it evenly. When all clients operate on the concave, saturating part of the curve, spreading is optimal. The optimal policy therefore switches depending on where the channel geometry places the clients relative to the dead zone. This switch is the effect that a linear model cannot see.
Baselines
The naive reference point is quantum FedAvg in the style of Chen et al., with a fixed shot count. Energy is spread uniformly, and each client participates whenever its battery can pay for a full round. This shows what happens if energy is ignored except as an on/off gate.
The strongest existing competitor is a composition of two methods. Myopic battery-first scheduling from Shen et al. is combined with SpoQFL filtering from Rahman et al., using a fixed shot count and uniform power. It is the strongest competitor because it already handles both battery memory and noisy updates, the two mechanisms our method claims to unify. For reference, SpoQFL reports 91.92% test accuracy on CIFAR-10, against 87.05% for the wpQFL baseline, in the same paper. Shen et al. report only curves, so no reported result was found on Lacuna for a numerical accuracy of the myopic policy. We will tune the composition's shot count, its filtering threshold and decay rate, and its learning rate. It gets the same grid-search budget we use for $\lambda$ and the planning window.
To keep the comparison fair, every method uses the same simulator, the same client partitions, the same channel draws, and the same total base-station energy. Success means reaching the composition's final accuracy or MSE with at least 25% fewer transmitted joules. It also means a lower fraction of harvested energy spent on updates that are later discarded. Both must hold on at least one partition where client data differ, and across at least two decades of the per-shot cost $c$.
Two ablations isolate our key ingredients. The first is the same planner built on a linear harvesting model and then evaluated in the sigmoid world. It isolates the value of modeling the curve. The second is our shot allocation with uniform power spreading. It isolates the value of time-sharing concentrated beams.
Why not just…?
Why not just stack myopic scheduling and SpoQFL filtering? This is the real threat, and it is exactly the composite baseline above. We expect it to fall short for two reasons. It pays for shots before deciding whether to keep an update. And uniform power leaves dead-zone clients permanently silent, which only shows up as error when data differ across clients. If the composite matches us on those partitions, the proposal's main claim fails.
Why not just use a linear harvesting model, which is convex and easy? A linear model is indifferent between spreading and concentrating power. It also overestimates harvest near saturation (Chen et al. (MJBP)) and misses the dead zone (Miridakis et al.). The linear-planner ablation measures directly how much accuracy and energy this simplification costs. If the loss is negligible for realistic cell geometries, we will report that.
Why not just run classical federated learning on these devices? SpoQFL's comparison of 91.92% against 76.89% for a classical baseline on CIFAR-10 (Rahman et al.) comes from simulation and does not show that quantum models are better at the edge. We make no such claim. Our contribution concerns how to spend energy on shot-based gradient estimates. We include a classical small-model federated learning run under the same energy budget so readers can see where the quantum model stands.
Why not just offload raw data to a cloud quantum processor? Uploading raw data at microwatt budgets, about 5.5 µW at 15 m (Lu et al.), costs more airtime than uploading a few dozen angles. It also breaks the data-locality premise (Uddin et al.). We will report the uplink energy under both options.
Proposed Experiments
We simulate 10 to 50 clients placed in a 100 m cell, using the geometric channel from Chen et al. (MJBP). The quantum circuits are simulated and have 4 to 8 qubits, with depolarizing gate noise, so that we can compare against SpoQFL.
The workloads are the Seoul Bike Sharing regression with 7 PCA features and a centralized reference MSE of 0.0209 (Herbst et al.). We also use the Iris, Wine, Breast Cancer, and Digits classification sets used by Xu et al.. Finally, we run PCA-reduced CIFAR-10, stating clearly that these results are not directly comparable to SpoQFL's 91.92%. Each workload is split in two ways: evenly across clients, and by label-skew or by region.
We track four metrics: accuracy or MSE against total base-station joules, rounds to reach a target, the fraction of energy wasted on discarded updates, and the worst-client participation rate.
Three checks decide the question. First, the structural switch must appear in experiments: concentrated beams should win when clients sit near the dead zone, and uniform spreading should win when they sit near saturation. Second, the linear-model planner should lose measurably when evaluated in the sigmoid world. Third, with data drawn evenly across clients, the composite baseline given extra shots should roughly match us, since coverage matters less there. With data that differ across clients, it should not. If the third check fails in either direction, the bias-versus-variance story is wrong, and we will say so.
References
- Wireless Networks With RF Energy Harvesting: A Contemporary Survey
- On Federated Learning with Energy Harvesting Clients
- Sporadic Federated Learning Approach in Quantum Environment to Tackle Quantum Noise
- Mixed-Timescale Beamforming and Power Splitting for Massive MIMO Aided SWIPT IoT Network
- Streaming IoT Data and the Quantum Edge: A Classic/Quantum Machine Learning Use Case
- Enabling Full-Duplex and Energy Harvesting in Uplink and Downlink of Small-Cell Network Relying on Power Domain Based Multiple Access
- Simultaneous Information and Energy Transmission With Short Packets and Finite Constellations
- Federated Quantum Machine Learning
- Green Cognitive Relaying: Opportunistically Switching Between Data Transmission and Energy Harvesting
- Quantum Federated Learning: A Comprehensive Survey
- Privacy-preserving Intelligent Resource Allocation for Federated Edge Learning in Quantum Internet