A batteryless sensor runs on harvested sunlight or radio energy and loses power many times a minute. Such a sensor can see hundreds of thousands of readings over its lifetime but can store only a few hundred. We propose to let it report a statistically valid confidence interval for a quantity like a median, a quantile, or a regression slope, computed on the device itself. The tool is the bag of little bootstraps (BLB), a resampling method that estimates full-dataset uncertainty from a small stored subset. It does this by giving each stored point a random integer weight so that the weights sum to the full count $N$ (The Big Data Bootstrap, A Scalable Bootstrap for Massive Data). BLB was built to fit large datasets on clusters. Its key property is that memory scales with the stored subset, not with $N$, and that is exactly the constraint of a microcontroller with a few kilobytes of non-volatile memory.
The difficulty is not running BLB on a small chip. It is that power failures quietly corrupt the statistics. Under intermittent power, some resampling jobs finish and others are cut off. The ones that get cut off are not random. Resamples whose weighted problems are harder to solve take longer, so they brown out more often, and the surviving bootstrap distribution loses its tails. Harvested energy is also often correlated with what is being measured: sunny hours bring both more solar power and more of the phenomenon being sensed. Our central claim is that energy-adaptive BLB, done naively, gives intervals that are too narrow. We further claim that a simple restructuring restores nominal coverage at negligible overhead. In that restructuring, each resample is addressed by a stored random seed and executed as a power-failure-safe task, and the energy-aware scheduler is only allowed to change when work runs, never which resamples count.
Core Claim
We make two claims, one statistical and one systems.
The statistical claim is that best-effort BLB under stochastic harvesting produces cost-correlated survivorship. Here best-effort means using whichever resamples complete, or shrinking the resample count when energy is low. Resamples drawn from the same multinomial law are kept with a probability that depends on their own solver cost and on the energy trace. The resulting quantile estimates are biased toward the center. This is a different failure from the one documented for federated learning. In that setting, Energy-Harvesting Distributed Machine Learning shows energy-agnostic participation skews a point estimate toward energy-rich clients, reaching 64% versus roughly 80% CIFAR-10 accuracy. Our concern is the shape of a resampling distribution, not the location of a point estimate.
The systems claim is that the fix is cheap because BLB's inner loop has an unusual property. Resample $k$ of bag $j$ can be regenerated exactly from a seed instead of stored:
$$ (w^{(j,k)}_1,\dots,w^{(j,k)}_b) = \mathrm{Multinomial}\big(N,\ (1/b,\dots,1/b);\ \text{seed} = h(j,k)\big), $$
where $b$ is the number of stored points in the bag, $w_i$ is the integer weight on point $i$, and $h$ is a fixed hash of the bag and resample indices. Every resample task is therefore idempotent: rerunning it after a brownout gives the same answer. The only state needing an atomic commit is a completion counter and the resulting statistic. The device pre-commits to a fixed list of resample indices and finishes every one eventually, so completion probability is one by construction. The energy scheduler can then do what power-neutral systems do, slowing down or deferring work as energy varies, without touching the sampling distribution.
Research Program
The work has three parts.
On-device BLB runtime. We will build a BLB runtime for MSP430FR-class microcontrollers in the style of task-based intermittent execution. Alpaca, which splits programs into atomic tasks and privatizes data written after being read (Alpaca: Intermittent Execution without Checkpoints), supplies the execution model. Seed addressing means almost nothing in the inner loop needs privatizing. Solvers are written with a fixed, data-independent iteration count per task chunk. This gives tasks bounded energy that can be checked against capacitor size, addressing the livelock problem in task sizing that Alpaca leaves to the programmer. Chinchilla's adaptive checkpointing (Adaptive Dynamic Checkpointing for Safe Efficient Intermittent Computing) is the generic alternative we compare against, not a component.
Streaming data collection. As data streams in, the device maintains its $s$ bags in ferroelectric non-volatile memory (FRAM) by reservoir sampling, and it tracks $N$. The device itself may sample only when powered, so the stream it sees can be energy-selected. For that case we test one extension, borrowed in spirit from the $1/\beta_i$ gradient scaling of Energy-Harvesting Distributed Machine Learning. The extension draws multinomial weights with probabilities proportional to the inverse estimated sampling propensity rather than uniformly. This is the only place energy enters the statistics, and it enters deliberately.
Energy-aware but distribution-neutral scheduling. Frequency and duty cycle follow capacitor voltage, as in Graceful Performance Modulation for Power-Neutral Transient Computing Systems. Energy may also decide when to stop adding resamples, but only through a rule fixed in advance and applied at bag boundaries. One example is BLB's own convergence check on interval width, which never discards a started resample.
We will characterize analytically when best-effort truncation biases interval endpoints. The expected condition is a nonzero correlation between a resample's solver cost and its statistic. We will then measure the size of that bias empirically.
Target estimators. We deliberately choose estimators where resampling earns its keep: medians and upper quantiles of a sensed quantity, and least-absolute-deviation slopes, as well as a weighted logistic slope as a smooth control. For quantiles and least-absolute-deviation, closed-form standard errors require estimating a density at the quantile, which is fragile on small samples.
Baselines
Naive reference. The do-nothing reference is the point estimate with no interval. The cheapest interval is the analytic asymptotic one: a sandwich standard error for the logistic slope, and a kernel-density-based standard error for quantiles. Both are computed on the stored $b$ points and rescaled by $\sqrt{b/N}$. Both run on the same device, under the same energy traces, with bandwidths chosen by standard rules. Beating it means coverage within two points of 95% where it falls short, at no more than a few times its energy.
m-out-of-n bootstrap. This bootstraps the stored $b$ points at size $b$ and rescales, with the rate assumed known. The Big Data Bootstrap reports it is sensitive to the choice of $m$. We give it the same resample budget and tune $m$ over the same grid we use for BLB's $b$.
Best-effort intermittent BLB. This is the strongest naturally occurring competitor and the one our statistical claim targets. Resamples that brown out are dropped, and the resample count shrinks with stored energy, mirroring the energy-adaptive granularity of Intermittent Inference with Nonuniformly Compressed Multi-Exit Neural Network for Energy Harvesting Powered Devices. It uses identical code minus the seed pre-commitment.
BLB under a generic intermittent runtime. This is the strongest existing systems method: unmodified BLB with weights stored in FRAM, run under Alpaca-style privatization and under Chinchilla. For context, Alpaca reports 4.0–5.2x average speedups over Chain, DINO and Ratchet and up to 17.6x smaller non-volatile footprint than DINO. Chinchilla reports checkpoint and restore overhead under 3.5% of run time and a 2% average speedup over Alpaca. These runtimes should match our coverage, since they are correct. Beating them means lower energy per finished interval and smaller FRAM footprint at equal coverage.
Offloading. The stored bag is sent by radio to a mains-powered server that runs standard BLB. Radio energy is charged at measured per-byte cost.
Ablations. Two ablations isolate the key ingredient. The first stores weights instead of seeds, isolating the logging savings. The second keeps seeds but lets energy drop started resamples, isolating the statistical effect.
For on-device bootstrap coverage under intermittent power: no reported result found on Lacuna. The only reported coverage figure is the nominal 95% achieved on reliable clusters by Simultaneous Inference for Massive Data: Distributed Bootstrap.
Why not just…?
Why not just use the analytic standard error and skip resampling? For the smooth logistic slope this may well be as good, and we expect it to be. That is why the logistic slope is in the study as a control, not as the selling point. For quantiles and least-absolute-deviation slopes, the analytic interval needs a density estimate at the quantile from a few hundred points. The experiments measure directly whether its coverage degrades there. If it does not, the proposal narrows to the systems contribution.
Why not just send the bag to a server? Energy-Harvesting Distributed Machine Learning notes that transmission energy often exceeds computation energy at the edge, but gives no figure for this setting. So we measure it ourselves: millijoules to ship $b$ points versus millijoules to compute the interval locally, across $b$. Offloading might win for very small bags, and we will report the crossover honestly.
Why not just run plain BLB under Chinchilla or Alpaca? These runtimes would give correct intervals. The question is cost. Storing $b$-length weight vectors per resample and logging them consumes FRAM and write energy that seed regeneration avoids. If the measured savings are within Chinchilla's reported under-3.5% overhead, our systems claim fails, and the contribution reduces to the warning about best-effort scheduling.
Why not just wait until the capacitor is full and then run BLB? Waiting also gives correct intervals but stalls progress. Energy-Harvesting Distributed Machine Learning found that wait-for-all reached 52% accuracy against roughly 80% for its best-effort scaling in the same iterations. We report time-to-interval for a full-charge policy under the same traces.
Proposed Experiments
Data. We stream the UCI Bike Sharing data (17,389 hourly records, used by Variable Selection with Scalable Bootstrapping in Generalized Linear Model for Massive Data). The target is the median and 90th percentile of hourly rentals, plus a least-absolute-deviation slope on temperature. This dataset gives a natural energy–data correlation. We synthesize a solar harvest trace from its hour and weather fields, so sunny hours deliver both more energy and more rentals. A second stream uses a few covariates from Lending Club (791,332 loans) for the logistic control. The Magic and Shuttle sets from A Scalable Bootstrap for Massive Data test robustness. In each case $b = N^{\gamma}$ with $\gamma \in \{0.5, 0.6, 0.7\}$.
Hardware. We use WISP5 / MSP430FR5969 under radio harvesting and MSP430FR5739 under replayed photovoltaic traces. Thousands of replications run in a cycle-calibrated simulator, and a subset is validated on hardware.
Ground truth. Ground truth for each parameter comes from the full dataset. Coverage is assessed by repeatedly resampling streams from it.
Metrics. We measure empirical coverage of nominal 95% intervals, width relative to a full mains-powered bootstrap, energy per finished interval, time-to-interval, and peak FRAM use.
Decisive check. On the energy-correlated Bike Sharing stream, best-effort BLB should show coverage well below 95% for the 90th percentile. The amount of under-coverage should grow as the harvest becomes weaker. Seed-committed BLB should hold within Monte Carlo error of 95% at comparable energy. If best-effort BLB already covers correctly, the survivorship claim is false. A secondary check sets the energy–data correlation to zero. This should leave only the cost-driven truncation, so the two sources of bias can be separated.
References
- The Big Data Bootstrap
- A Scalable Bootstrap for Massive Data
- Energy-Harvesting Distributed Machine Learning
- Alpaca: intermittent execution without checkpoints
- Adaptive Dynamic Checkpointing for Safe Efficient Intermittent Computing
- Graceful Performance Modulation for Power-Neutral Transient Computing Systems
- Intermittent Inference with Nonuniformly Compressed Multi-Exit Neural Network for Energy Harvesting Powered Devices
- Simultaneous Inference for Massive Data: Distributed Bootstrap
- Variable Selection with Scalable Bootstrapping in Generalized Linear Model for Massive Data