A wireless sensor powered by sunlight or radio energy has to decide, slot by slot, how much stored energy to spend on transmitting. When the energy source is "sticky," meaning a sunny or dark spell lasts a characteristic duration called the coherence time, a known result says something counterintuitive: for a fixed battery, longer coherence time lowers throughput. The long dark spells drain the battery and the node goes silent. That result, and the fix it implies (grow the battery in proportion to coherence time), covers only one goal: maximize average bits. Real nodes also care about not going silent. This note asks how the whole curve of best-achievable (throughput, reliability) pairs changes with coherence time. We make two predictions. First, the curve collapses onto a single shape once battery size is measured in units of coherence time, times a reliability-dependent factor. Second, as coherence time grows, the policies along that curve change their behavior very unevenly: nearly all the change is packed into a narrow band of preferences near the high-reliability end.
The second prediction is where the Fisher information metric comes in. The Fisher metric is the standard way to measure how different two probability distributions are, independent of how they are parameterized. We use it to measure how much a policy's actual behavior changes as the user's preference slides along the trade-off. Standard methods for tracing trade-off curves spread their preference samples uniformly, whether by fixed weights or by random directions. If the behavioral change is concentrated, uniform spreading wastes most of the compute on near-identical policies and under-resolves the regime operators care about. We propose to reparameterize the preference coordinate by Fisher arc length, so that neighboring sample points are equally distinguishable in behavior, and to test whether this recovers the reliability end of the curve at a fixed sample budget.
Core Claim
The claim has three parts. First, in a block-correlated energy-harvesting channel, the throughput–outage Pareto front (the set of operating points where neither objective can improve without hurting the other) obeys a scaling collapse in normalized battery size. Second, the Fisher length of the preference-conditioned policy family concentrates as coherence time grows. Third, sampling preferences uniformly in that Fisher length traces the front with less compute than sampling uniformly in weight space or on the preference sphere.
The scaling conjecture comes from a back-of-envelope argument. The block i.i.d. model of Online Power Control for Block i.i.d. Energy Harvesting Channels holds harvested energy constant for $T$ slots, then redraws it independently. Suppose a block is "dry" with probability $p$ and staying above an outage rate costs at least $e_{\min}$ energy per slot. Then avoiding outage through $k$ consecutive dry blocks needs $kTe_{\min}$ stored energy, and such runs occur with probability about $p^k$. Keeping the outage fraction below $\varepsilon$ therefore suggests
$$B^\star(\varepsilon, T) \approx T\, e_{\min}\, \frac{\log(1/\varepsilon)}{\log(1/p)},$$
where $B^\star$ is the smallest battery achieving outage $\varepsilon$ at negligible throughput loss. This keeps the linear-in-$T$ law of Shaviv and Özgür, whose throughput-only large-battery regime is $B \ge \mu + T(E_{\max}-\mu)$ with $\mu$ the mean and $E_{\max}$ the peak arrival. The slope, however, now grows logarithmically in the reliability target, so fronts plotted against $B/T$ should overlay across $T$.
For the geometric part, let $\pi_t$ be the policy selected by preference coordinate $t \in [0,1]$, and let $d_t$ be its long-run state distribution. The Fisher length element is
$$\frac{ds}{dt} = \sqrt{\mathbb{E}_{s \sim d_t,\, a \sim \pi_t}\big[(\partial_t \log \pi_t(a \mid s))^2\big]}.$$
This is the Fisher metric pulled back onto the one-dimensional family of policies. We predict that as $T$ grows at fixed $B/T$, most of the total length $S=\int ds$ accumulates in a small interval of $t$ near the reliability end. There, a small change in preference flips the node from "spend" to "hoard through the drought."
One tempting claim should be avoided, and we avoid it. Both objectives are long-run averages of per-slot quantities (log-rate and an outage indicator). The set of values reachable by randomized stationary policies is therefore convex, because it is a linear image of the set of achievable state–action frequencies. The zero-duality-gap result of Constrained Reinforcement Learning Has Zero Duality Gap is consistent with this. So the claim that weighted sums miss concave regions, which Optimal Scalarizations for Sublinear Hypervolume Regret proves in general, need not bite here for randomized policies. We claim non-convexity only for the small deterministic policy classes that fit on a microcontroller, and we test that claim separately.
Research Program
The first component is analysis. We will derive the throughput–outage front for the block i.i.d. channel with a Bernoulli-dry arrival law. We will make the heuristic scaling above rigorous as an upper and lower bound in $B/T$, extending the constant-gap machinery Shaviv and Özgür use for their Fixed Fraction policy. Their Policy 1 spends a fixed fraction $q$ of the battery per block and is provably within about 0.72 bits per channel use of optimal throughput. The open question is whether some one-parameter extension of it is also near-Pareto-optimal for outage. If it is, that is a clean theoretical result. It also changes what the learned solver is for: it becomes a verifier rather than the product.
The second component is the solver. We train a single policy network conditioned on the preference, in the style of Basaklar et al.. Inputs are battery level, current arrival, and position within the block. Updates are Fisher-preconditioned natural policy gradient steps with KL anchoring, taken from Anchor-Changing Regularized Natural Policy Gradient for Multi-Objective Reinforcement Learning. That paper supplies the optimizer and its $\tilde O(1/T)$ convergence, but it traces only one policy per budget and never models batteries or correlated arrivals. Scalarization uses the hypervolume scalarization of Zhang's paper (a max–min scalarization with provably optimal coverage rate).
Our change is in how preferences are sampled. Every few iterations we estimate $ds/dt$ on a grid from rollouts, which requires only first derivatives of log-probabilities, and resample training preferences uniformly in $s$. This contrasts with Multi-Objective Reinforcement Learning with Continuous Pareto Frontier Approximation, whose policy-manifold method uses a Euclidean parameterization fixed in advance and needs noisy sampled Hessians (111–390 iterations even on 2D LQG). We keep a learned conditional policy and use Fisher geometry only to decide where to spend samples.
The third component is realism. Ku et al. fit Gaussian-mixture hidden Markov weather models to the Elizabeth City State University solar data. We use those models to generate arrivals whose correlation is not block-shaped, and we define an effective coherence time as the mean hidden-state sojourn. This tests whether the $B/T$ collapse survives outside the idealized model, which is exactly the limitation the file flags for block i.i.d. analysis.
Baselines
The naive reference is charge-then-immediate-discharge: spend everything as soon as it arrives. It is the default behavior that SmartON improves on by 1–7× in captured events on hardware. We run it on identical arrival traces. It yields a single point on the front, and any serious method must dominate it.
The strongest structured baseline is the Fixed Fraction family of Shaviv and Özgür, swept densely over $q$, optionally with a reserve floor. It costs almost nothing, it is provably near-optimal for throughput (gap at most about 0.72 bits per channel use in that paper), and it is what an engineer would actually deploy. We give it the same simulator and evaluate the same number of policies the learned method produces. Beating it means strictly higher hypervolume, specifically in the outage region below 1%.
The strongest learned baseline is ARNPG-EPD from Zhou et al., run once per outage budget on an evenly spaced budget grid, with total environment steps matched to our single conditional run. On synthetic tabular constrained problems, Zhou et al. report log-log convergence slopes between −0.9 and −1. On Acrobot they report higher reward than NPG-PD and CRPO within cost limits, but the file gives no exact figure. For throughput–outage fronts on block i.i.d. channels we found no reported result on Lacuna, so all numbers here will be our own. Beating ARNPG-EPD means matching its per-budget points while also covering budgets it was not trained for, at equal total samples.
Two ablations isolate the key ingredient. The first is the identical conditional natural-gradient solver with uniform preference sampling, both uniform weights and hypervolume-scalarization directions. The second replaces the Fisher length with Euclidean parameter-space arc length, $\|\partial_t\theta\|$. If Fisher resampling does not beat both at equal compute, the geometric claim fails. We also compare against the LSTM-DQN approach of Chu et al. swept over its loss weight, as a representative weighted-sum learner. That paper reports an 11.93% average-reward gain over a myopic policy under memoryless Poisson arrivals, not correlated ones.
Why not just…?
Why not just use a weighted sum of throughput and outage? For randomized policies on long-run averages, the reachable region is convex, so a weighted sum with mixing can in principle reach every front point. We concede this. The weakness is resolution: if behavior change is concentrated, uniform weights put most samples where nothing happens. The ablation with uniform weights against Fisher resampling settles whether this matters. A separate test on deterministic threshold policies checks whether Zhang's concavity argument applies to the policy class actually deployed.
Why not just sweep the Fixed Fraction fraction $q$? This is the most serious threat. If a $q$-sweep plus reserve floor reaches within noise of the learned front, the right conclusion is that the problem is solved analytically and the solver is unnecessary. We would report that as the main result, alongside the scaling law. Our reason to expect a gap is that fixed-fraction spending decays geometrically through a drought and ignores how far into the block the node is. Survival-oriented policies need spending that depends on remaining drought risk. The head-to-head hypervolume comparison below 1% outage decides this.
Why not just solve one constrained problem per operating budget? Paternain et al. guarantee this is well posed, and ARNPG-EPD solves each instance fast. The cost is retraining for every budget and for every shift in coherence time. The matched-sample comparison measures whether one conditional run with Fisher resampling matches a grid of constrained runs.
Why not just refine where the front has gaps in objective space? Placing new preferences between points that are far apart in (throughput, outage) is a real competitor. It may track the same concentration without any Fisher computation. We include it as an ablation. The Fisher version should win when behavior changes sharply but objectives move little, such as switching between two hoarding strategies with similar averages but different drought tails. If it does not win, the Fisher component should be dropped.
Proposed Experiments
We sweep coherence time $T$ from 1 to 200 slots and normalized battery $B/T$ over a decade on the block i.i.d. simulator. At each setting we compute throughput (bits per channel use), outage fraction, and hypervolume against a fixed reference point.
The first decisive check is the scaling collapse. Fronts plotted against $B/(T\log(1/\varepsilon))$ should overlay across $T$, and they should fail to overlay against $B/T$ alone at small $\varepsilon$.
The second is concentration. For each $T$, we measure the fraction of total Fisher length $S$ lying in the top 10% of the preference interval, and predict it rises with $T$.
The third is efficiency. We plot hypervolume against environment steps for Fisher resampling, uniform-weight sampling, hypervolume-scalarization directions, objective-gap refinement, Euclidean arc length, the ARNPG-EPD grid, and the Fixed Fraction sweep, using five seeds and matched tuning budgets. We then repeat the collapse and efficiency tests on the solar-derived arrival models to see whether the idealized findings transfer.
The proposal fails cleanly if any of three things happens: the collapse does not appear, Fisher length does not concentrate, or the $q$-sweep matches the learned front. The first two outcomes would each be informative about how correlated energy scarcity shapes reliability.
References
- Online Power Control for Block i.i.d. Energy Harvesting Channels
- Constrained Reinforcement Learning Has Zero Duality Gap
- Optimal Scalarizations for Sublinear Hypervolume Regret
- PD-MORL: Preference-Driven Multi-Objective Reinforcement Learning Algorithm
- Anchor-Changing Regularized Natural Policy Gradient for Multi-Objective Reinforcement Learning
- Multi-Objective Reinforcement Learning with Continuous Pareto Frontier Approximation
- Data-Driven Stochastic Models and Policies for Energy Harvesting Sensor Communications
- SmartON: Just-in-Time Active Event Detection on Energy Harvesting Systems
- Reinforcement Learning-Based Multiaccess Control and Battery Prediction With Energy Harvesting in IoT Systems