Learning Who Feeds a Wireless Harvester: Transfer Entropy for Pilot-Free Interference Scheduling

Tiny wireless sensors can power themselves by rectifying radio waves into direct current, and some of the most useful waves are other people's transmissions: interference that would normally be cancelled. Every scheme that deliberately exploits interference for energy, however, assumes the network already knows its channels, which radio links reach which receiver and how strongly. Learning that is expensive. A survey of the field notes that a node which spends its harvested energy on channel estimation has nothing left for data (Wireless Networks with RF Energy Harvesting: A Contemporary Survey). This note proposes a cheaper route. The node passively logs two things: its own harvested-power trace, and coarse on/off activity of nearby transmitters, obtained by energy detection or packet timestamps. From these logs we infer a small directed graph of which transmitters, and which upstream triggers of those transmitters, predict the node's future energy intake. The graph then tells the node two things: which few channels are worth monitoring, and when to switch from decoding data to harvesting.

The statistical tool is transfer entropy, a measure of how much the past of one signal reduces uncertainty about the future of another beyond what the target's own past already predicts. Its multivariate, graph-building form is the directed information graph (Directed Information Graphs). The central bet is that the rectifier makes this structure nonlinear and conditional. Diodes barely conduct at low input power: conversion efficiency is above 50% at -5 dBm and about 0.4% at -40 dBm (Information and Energy Transfer with Experimentally-Sampled Harvesting Functions). A weak interferer therefore contributes energy mainly when a stronger one is already active. Linear or pairwise methods misrank such interferers. A conditional, model-free measure should rank them correctly and should also find upstream "triggers", such as a source whose packet causes a relay to retransmit. Triggers give the node lead time to switch modes before the energy arrives.

Core Claim

Under realistic nonlinear rectification and protocol-driven transmission cascades, a sparse set of $K$ causal parents can be inferred from cheap activity logs. Choosing which neighbors to monitor from that parent set, and scheduling the switch between decoding and harvesting from it, should harvest more net energy at a fixed sensing budget than three alternatives: monitoring chosen by correlation, monitoring chosen by linear Granger causality, or a fixed policy that ignores the network. The relevant quantity is the conditional transfer entropy from a candidate source $X$ (a neighbor's activity indicator) to the node's harvested-power trace $E$, given the other already-selected sources $Z$:

$$TE_{X \to E \mid Z} = I\left(E_{t+1} ; X_{t-\ell:t} \mid E_{t-\ell:t}, Z_{t-\ell:t}\right),$$

where $I(\cdot;\cdot\mid\cdot)$ is conditional mutual information and $\ell$ is the history length. Greedy parent selection with this quantity, under the bounded-in-degree result of Quinn et al., needs only $(K+1)$-dimensional statistics rather than the joint distribution of the whole neighborhood.

What would be new is narrow but concrete. Directed information graphs have so far been applied to abstract time series such as Twitter timestamps, never to radio interference or to systems with energy constraints. Interference-harvesting designs, meanwhile, assume perfect CSI (channel state information: the measured gains of each radio link) and a linear harvesting model.

Relation to Prior Work

Three papers motivate the gap from the harvesting side. Wireless Information and Energy Transfer in Multi-Antenna Interference Channel shows that time-splitting receivers, which alternate between decoding and harvesting, beat the idealized simultaneous receiver when cross-link interference is strong. Their schedules are designed with perfect CSI and linear harvesting; we keep the time-splitting receiver but learn the schedule from causal structure. Joint Optimization of Power and Data Transfer in Multiuser MIMO Systems shows that letting interference reach harvesting users, instead of nulling it, raises harvested power by about 50%. That result needs full transmitter-side CSI, so we use it only as an oracle reference, not as a competitor. On the Performance of Interference-Based Energy-Harvesting-Enabled Wireless AF Relaying Communication Systems shows that interference powers relays. Relays also create the cascade structure we expect to exploit: a source's transmission predicts the relay's retransmission.

One paper constrains the design. Information and Energy Transfer with Experimentally-Sampled Harvesting Functions proves that energy delivery tolerates errors in the estimated harvesting curve much better than information rate does. That is a warning: our inference errors will hurt decoding more than harvesting, so we must report both.

Two sources supply the method. From TRENTOOL we take the nearest-neighbor estimator and the shift test. The shift test matters here because radio coupling is effectively instantaneous at our time scales, and we want only lagged, predictive influence. From TENDE, a diffusion-model estimator of conditional mutual information that stays robust to redundant noise dimensions at a higher training cost, we take the approach for long histories.

One paper is the policy competitor. Dynamic spectrum access in cognitive radio networks with RF energy harvesting uses a Markov decision process over battery and queue states, with primary-user activity modeled as a fixed Markov chain. That baseline is strongest exactly where its stationarity assumption holds.

Research Program

The first stage builds the testbed. We extend the ns-3 energy framework (An energy framework for the network simulator 3 (NS-3)), which models nonlinear battery depletion on a 7x7 grid with standard routing, by adding a harvesting source. That source maps received RF power to DC output through the measured 1-stage and 9-stage rectifier curves from Seo and Varshney. The physical layer reuses the fading assumptions of the multi-antenna downlink and amplify-and-forward relay setups in the file. Traffic comes from real protocol behavior, namely routing, relaying and contention backoff, so that cascades arise from the protocols rather than being planted.

The second stage is inference. Each harvesting node logs one-bit activity per neighbor channel along with its own DC trace. Offline, or at an access point fed by low-rate logs, we run greedy multivariate parent selection with $TE_{X \to E \mid Z}$, apply the shift test, and test significance by permutation. Nearest-neighbor estimation is used for short histories and TENDE for long ones. The output is a parent set of size at most $K$, together with lead-time estimates for each parent.

The third stage is control. The node monitors only its $K$ parents and uses a simple predictor built on them to decide, slot by slot, whether to harvest or decode. Keeping the controller deliberately simple ensures that any gain comes from parent selection rather than from a clever policy. All energy accounting is net of sensing costs: each monitored channel has an explicit listening cost, and the CSI-based alternatives pay for pilots.

Baselines

The naive reference is a node that monitors nothing and switches between harvesting and decoding on a fixed duty cycle tuned on held-out traffic. Every method must beat it on net harvested energy, or the whole exercise is pointless.

The two direct competitors for parent selection are Pearson correlation and linear Granger causality, a lag-regression test fitted by least squares. Each picks the top $K$ neighbors from the same logs, with the same history length, the same $K$, and the same downstream controller. The Pearson baseline reflects the point of Benchmarking methods for mapping functional connectivity in the brain that symmetric co-movement measures ignore direction.

The strongest competitor overall is a learned energy forecaster: gradient-boosted trees or a small recurrent network given all neighbor channels, with group-lasso sparsity, whose top $K$ important features define the monitored set. It receives the same training logs and the same hyperparameter-search budget as our estimator's tuning of history length and nearest-neighbor count.

The policy baseline is the Markov decision process of Lu et al., rebuilt on our state space. For this method the file gives no numeric throughput result; Lu et al. report only a throughput "sweet spot". The oracle reference is MM precoding with full CSI, where Rubio et al. report a maximum sum rate of 4.5 bit/s/Hz versus 2.75 bit/s/Hz for block diagonalization, plus about 50% more harvested power, on i.i.d. Rayleigh channels with 4–6 transmit antennas. We will reproduce their configuration to calibrate our simulator, then run MM with the energy cost of pilots charged against it.

The key ablation replaces conditional TE with bivariate TE, which drops the conditioning on $Z$. This isolates whether conditioning on other interferers, our hypothesized response to rectifier thresholds, is what matters. A second ablation removes the shift test.

Beating a baseline means a statistically significant gain in net harvested energy at equal decoding throughput, with confidence intervals across random topologies and seeds, at the same $K$.

Why not just…?

Why not just use linear Granger causality, which is closed-form and cheap? If interferers are strong enough that every rectifier operates in its near-linear high-power region, linear Granger causality should match TE, and we expect it to. The proposal predicts a gap only when received powers straddle the diode turn-on region, where contributions stop being additive. We will sweep transmit power and node density across the measured efficiency curve and report where the two methods diverge. If they never diverge, the conclusion is that linear methods suffice for this problem.

Why not just train a forecaster on all neighbors and read off feature importances? This is the real threat. A sparse forecaster may select the same parents. Our case rests on two testable differences. First, conditional TE is designed to separate a trigger from the transmitter it triggers and to avoid picking redundant correlated neighbors. Second, permutation tests give a principled stopping rule for $K$. We will compare the two at matched sensing budgets, both on stationary traffic and after mid-run topology changes. If the forecaster ties, we will report that the contribution reduces to an interpretable diagnostic, not a better controller.

Why not just estimate CSI and run MM precoding? Precoding needs transmitter coordination and pilots, while our setting is a microwatt receiver, about 5.5 µW at 15 m per Seo and Varshney, that cannot coordinate anyone. The net-energy accounting in the experiments settles whether pilot cost actually dominates at these power levels.

Why not just use the Markov decision process of Lu et al.? On stationary traffic it may win, since it optimizes the policy directly. We test it under non-stationary traffic as well, and against network sizes at which its state space becomes impractical.

Proposed Experiments and Decisive Checks

We first validate the estimator pipeline before applying it to wireless data. On the coupled AR(10) benchmark, it must match TRENTOOL's reported 100% direction accuracy with false positives typically below 5%. On the Twitter crisis data, the parent-selection code must approach the roughly 95% influence accuracy reported for directed information graphs. For TE applied to wireless or harvesting traces, no reported result was found on Lacuna, so the simulator results will be the first.

The main experiment has two parts. The first measures parent-recovery precision and recall against the simulator's ground-truth interference and triggering graph, compared across all selection methods as a function of rectifier operating point. The second measures net harvested energy and decoding throughput under each monitoring policy, plus node lifetime against the ns-3 framework's reported baseline relay-failure time of 1164.8 seconds.

The decisive check is the interaction between the selection method and the operating point. The proposal is supported if conditional TE outperforms bivariate TE and linear Granger causality specifically near the diode threshold. It is undercut if the sparse forecaster matches it everywhere.