When the Signal Powers the Sensor: Endogenous Gains in Energy-Harvesting Compressive Sensing

Picture a field of cheap solar-powered sensors that measure sunlight, or temperature under the same sky, and all transmit at once. Their radio signals add up in the air, so a base station receives a few compressed sums instead of one packet per sensor. This scheme, known as compressive sensing (recovering a signal from far fewer measurements than its length, provided the signal is sparse in some basis), is attractive because it needs only tens of channel uses for hundreds of nodes. Each node, however, scales its transmission by whatever power its battery allows, and the battery is charged by the same weather the node is measuring. The base station does not know these per-node power levels. Theory treats the resulting mismatch as random, independent noise in the measurement matrix. We argue that it is not random. Brighter nodes transmit louder, so the unknown gain is correlated with the very signal being recovered. That correlation is a form of unobserved confounding, and it produces systematic errors that more measurements cannot remove.

The central prediction is concrete. With independent gain errors, a signal that is sparse in frequency is blurred into a diffuse noise floor. When the gains are correlated with the signal, their product with the signal creates coherent spurious frequency components, roughly harmonics and sum frequencies of the true ones. Standard recovery reports these as real, confident support. We propose to show this failure, measure it, and fix it using the temporal memory of energy harvesting. A battery integrates past harvest, so a node's power level is a slowly evolving, predictable function of the signal's own history. That memory turns an unidentifiable per-frame calibration problem into a tractable joint estimation of signal and battery trajectories.

Core Claim

In the over-the-air model of Wireless Compressive Sensing for Energy Harvesting Sensor Nodes, the base station receives $y_t = H_t S_t (\phi_t \odot x_t) + e_t$. Here $x_t \in \mathbb{R}^n$ is the spatial field sampled by the $n$ nodes at frame $t$ and is sparse in the discrete Fourier basis. $H_t$ holds the channel fading gains. $S_t$ holds known pseudo-random spreading codes. $\phi_t$ is the vector of per-node transmit amplitudes set by battery state, and $e_t$ is additive noise. Yang et al. prove that the composite matrix satisfies the restricted isometry property, the standard condition guaranteeing stable sparse recovery. They assume both that the base station knows $\phi_t$ and that $\phi_t$ is independent of $x_t$. We drop both assumptions.

If the base station uses nominal amplitudes $\bar\phi$, it effectively recovers $z_t = (\phi_t/\bar\phi) \odot x_t$ rather than $x_t$. When $\phi_t$ depends on $x_t$, either directly because the sensed quantity is the harvest source, or through a shared latent weather state $U_t$, the vector $z_t$ is not sparse in frequency and its extra components are coherent. Our claim has two parts. First, this endogenous distortion creates an error floor and a false-support rate that do not shrink with the number of measurements. Independent gain errors of the same magnitude do not produce this effect. Second, modeling the battery as a state-space system driven by the estimated signal removes the floor in the regime where the sensed field is the harvest source. It also yields honest worst-case bounds in the regime where a hidden weather state drives both.

Relation to Prior Work

The motivation comes from the gap that Yang et al. leave open. They treat harvest and fading as exogenous and assume a perfectly known sensing matrix. They also name time-correlated power management as unresolved. Two noise-robustness results constrain the design. General Deviants: An Analysis of Perturbations in Compressed Sensing proves that basis pursuit error grows linearly with a bounded matrix perturbation $E$. Their bound still holds here, but it is worst-case and treats $Ex$ as unstructured. It therefore cannot distinguish a diffuse error from a coherent spurious peak, and it gives no route to removing the bias. Price of Quality shows that a decoder blind to per-sample noise levels loses at most a factor of two, because Lasso depends only on the average noise variance. That result assumes noise independent of the signal. Our experiments test whether the benign "agnostic is nearly enough" conclusion survives when the corruption is correlated with the support. We expect that it does not.

The causal side supplies the framing and one tool. Long-term Causal Inference via Modeling Sequential Latent Confounding models confounding bias as an autoregressive process and reduces PEHE (precision in estimating heterogeneous effects) from 36.12 to 3.11 against a static-bias baseline. It depends on unconfounded experimental checkpoints, which physical sensor deployments do not have. We replace those checkpoints with a known physical law, battery bookkeeping. Sharp Bounds for Generalized Causal Sensitivity Analysis bounds hidden confounding by a density-ratio parameter $\Gamma$ on treatment assignment. We borrow that idea of a single interpretable strength parameter, but apply it to the dependence between sensing gains and signal rather than to treatment propensity.

Temporal-energy models supply the dynamics. Data-Driven Stochastic Models and Policies for Energy Harvesting Sensor Communications fits a four-state hidden Markov weather model to solar data. Online power control for block i.i.d. energy harvesting channels shows that long coherence times force prolonged droughts. Both results imply that battery-driven gains are persistent and predictable, which our method relies on.

The closest sparse-plus-confounding work, Sparse Causal Effect Estimation using Two-Sample Summary Statistics in the Presence of Unmeasured Confounding, uses sparsity to identify effects despite confounding in cross-sectional genetic data. It has no sensing operator and no dynamics, and its confounder acts on outcomes rather than on the measurement matrix.

Research Program

The first component is analysis. For a stylized model with amplitudes $\phi_j = g(\alpha x_j + u_j)$, where $u_j$ is independent noise and $\alpha$ controls endogeneity, we will derive how much energy of $z$ falls outside the true frequency support. We will show that it scales with $\alpha$ and the curvature of the amplitude policy $g$, and that it is independent of the measurement count $m$. The asymptotic floor in $m$ follows because recovery converges to $z$, not $x$. We will contrast this with the $\alpha = 0$ case, where the same gain variance acts as noise that is spread across frequencies and that averages out.

The second component is the estimator. It alternates two steps. The first step recovers the frequency coefficients of $x_t$ given current amplitude estimates, using the square-root Lasso of Belloni et al.. This is Lasso on the square root of the residual sum of squares, whose penalty does not depend on the unknown noise level. The residual gain error makes the effective noise level unknown and time-varying, so a solver that needs this level would have to be re-tuned every frame. The second step runs a forward-backward smoother over each node's battery:

$$b_j(t+1) = \min\big(B,\ b_j(t) + \eta_j\, h(\hat x_j(t), \hat w(t)) - c(\phi_j(t))\big), \qquad \phi_j(t) = g(b_j(t)).$$

Here $b_j$ is node $j$'s battery level and $B$ its capacity. $\eta_j$ is an unknown per-node harvesting efficiency. $h$ is the harvest as a function of the estimated field and, in the shared-cause regime, of a hidden weather state $\hat w$ tracked with a hidden Markov model in the style of Ku et al. The function $c$ is the energy cost of transmitting at amplitude $\phi_j$, and $g$ is the nodes' known firmware power policy. The unknowns drop from $n$ amplitudes per frame to $n$ efficiencies plus initial charges for the whole run. That is the identification gain supplied by temporal correlation. We will state identifiability conditions: a known policy $g$, enough temporal variation of $x$ within the battery coherence time, and at least one weakly harvesting or externally referenced node to fix the overall scale.

The third component covers the shared-cause regime, where the sensed field is, say, temperature and the harvest is solar. There the gains are only partially identified. We will define a strength parameter $\Gamma$ bounding how far the conditional distribution of gains given the signal can deviate from the model's prediction. We will then compute the worst-case NMSE (normalized mean-squared error) and false-support set over that ball, yielding the smallest $\Gamma$ at which a detected frequency could be spurious.

Baselines

The naive reference point is standard basis pursuit or Lasso with nominal amplitudes $\bar\phi$. This is what a deployment of the Yang et al. scheme would do if nodes adapt power without reporting it, and it shows the size of the problem.

The strongest existing method is the Yang et al. recovery under its own assumptions. They report reliable recovery of a $k = 5$ sparse signal over 500 nodes with about 68 measurements at 25 dB average SNR, even with a 25 dB SNR spread across nodes. We will first reproduce that operating point with exogenous random gains, then introduce endogeneity at matched gain variance. Beating it means recovering the 68-measurement performance level under endogenous gains, where we expect the baseline to plateau.

A second strong competitor is perturbation-aware basis pursuit following Herman and Strohmer, with its tolerance inflated to cover the gain error. It receives the same measurement budget, and its tolerance is tuned on a held-out grid with the same number of settings we allow our method.

The upper reference is an oracle that knows $\phi_t$ exactly. A pilot-calibration baseline is discussed below. All methods see identical measurement sequences, noise realizations and node counts, and every hyperparameter gets the same grid size.

Three ablations isolate the key ingredients. The first is the most important. We randomly permute the true amplitudes across nodes, which keeps their distribution but breaks their correlation with the signal. If the naive method's error and false-support rate collapse under permutation, the harm is attributable to endogeneity rather than gain variance. The second removes the battery model and estimates amplitudes independently per frame, which tests whether temporal memory provides the identification. The third swaps square-root Lasso for ordinary Lasso with cross-validated penalty, which checks whether pivotal tuning matters or is incidental.

Why not just…?

Why not just have each node report its power level? This is the most serious alternative. Per-node calibration in an over-the-air sum needs on the order of $n$ separated pilot or header slots, about 500 at the Yang et al. scale, against about 68 measurement slots. That overhead is paid again whenever amplitudes change. If amplitudes change only once per coherence block of $T$ frames, the overhead amortizes to roughly $n/T$ per frame, and for long coherence times this could beat our method. We will not argue this away. The experiments sweep $T$ and count total channel uses and transmit energy including headers, and we will report the crossover coherence time below which calibration loses. The shared-cause regime also leaves a residual role for our method even with reports, because reported power levels quantize the battery state.

Why not just treat the gain mismatch as extra noise and rely on robust recovery? The Herman–Strohmer bound and the Price-of-Quality result both suggest that unstructured corruption only costs a constant factor. Their premise is independence between corruption and signal. The permutation ablation tests this alternative directly. If robust recovery performs equally well on permuted and unpermuted gains, our central claim is false and we will say so.

Why not just transmit at fixed power whenever the battery allows? This removes amplitude uncertainty, but it turns the gain into an on/off switch that is still correlated with the signal: sunny nodes transmit and shaded ones stay silent. It also gives up the throughput advantage of spreading power evenly that Shaviv and Ozgur exploit. We include fixed-power on/off transmission as a baseline, reporting both recovery error and delivered energy efficiency.

Proposed Experiments

We will use three testbeds. The first is the Yang et al. synthetic setting (500 nodes, $k = 5$, Rayleigh fading, 25 dB SNR with a 25 dB spread), extended with batteries and an endogeneity knob $\alpha$. The second places nodes on pixels of the EUMETSAT satellite irradiance fields from the BSRN/EUMETSAT dataset used in Improving day-ahead Solar Irradiance Time Series Forecasting by Leveraging Spatio-Temporal Context. Harvest follows the imagery, and the sensed field is either the irradiance itself (self-powered regime) or a correlated synthetic field (shared-cause regime). That paper reports forecasting MAE of 50.35 at the CAB station, but no reported result found on Lacuna for compressive recovery on this data, so all recovery baselines there are run by us. The third reuses the Elizabeth City solar traces and four-state hidden Markov model from Ku et al. to drive harvest time series. No reported recovery number is found on Lacuna for that dataset either.

Metrics are NMSE, false-support rate (fraction of reported frequencies absent from the true support), the measurement count needed to reach a target NMSE, and total channel uses and energy including any calibration.

Three results would be decisive. The first would show that naive and perturbation-aware recovery reach an NMSE and false-support floor that stays flat as $m$ grows from 68 to several hundred under endogenous gains, while the same methods improve under permuted gains. The second would show that the battery-tracking estimator approaches oracle NMSE in the self-powered regime, with the per-frame ablation failing, which establishes that temporal memory does the identifying work. The third would show that in the shared-cause regime the $\Gamma$-bounds cover the true error on the satellite fields and flag the spurious frequencies the naive method reports. If the permutation test shows no endogeneity-specific harm, the proposal reduces to ordinary gain calibration and should be abandoned in favor of the pilot baseline.