Low-Power Estimation (LoPE): Resource-Aware Methods

• Low-Power Estimation (LoPE) is a framework that uses compact models, lightweight predictors, and in-sensor processing to operate under stringent resource constraints.
• It employs architectural strategies such as quantization, duty-cycled sensing, and hardware-software co-design to reduce data movement and computational cost.
• Mathematical simplifications and low-order predictor designs in LoPE enable near-state-of-the-art accuracy with significantly reduced energy and memory overhead.

Searching arXiv for relevant papers on Low-Power Estimation and the provided topic scope. Low-Power Estimation (LoPE) denotes estimation methods designed to operate under stringent resource constraints while maintaining useful predictive or inferential accuracy. In the material summarized here, the term spans several domains—indoor LP-IoT channel estimation, edge vision, received-power prediction in Low-Power and Lossy Networks, time-of-flight depth sensing, embedded power modeling, gaze estimation, backscatter channel estimation, low-power radar tracking, edge-TPU satellite pose estimation, wearable EEG arousal estimation, FPGA power characterization, low-power DNN inference, optical-flow estimation on MCU clusters, LPWAN lifetime estimation, and decentralized control in the Witsenhausen problem—but the recurring objective is consistent: reduce energy, memory, compute, data movement, or sensing overhead without materially sacrificing estimation quality (Arif et al., 2024, Bonazzi et al., 2023, Dargie, 25 Jan 2025). Across these settings, LoPE is realized through compact models, lightweight predictors, quantization, on-sensor or in-sensor processing, reduced-duty-cycle sensing, and hardware-software co-design.

1. Conceptual scope and defining characteristics

LoPE in indoor LP-IoT is defined as channel estimation that meets “stringent resource constraints—limited energy, memory, compute cycles—while still producing accurate, robust estimates of the wireless channel to maintain reliable connectivity” (Arif et al., 2024). In the edge-vision setting, it is framed as shrinking “latency and energy from the moment photons hit the sensor to the moment the device emits a prediction, without compromising accuracy,” with in-sensor AI reducing data movement and emitting only compact results (Bonazzi et al., 2023). In Low-Power and Lossy Networks, LoPE appears as a “lightweight, n-step predictor for the received power” that tolerates successive packet losses with minimal computational and memory resources (Dargie, 25 Jan 2025). In time-of-flight imaging, the same principle is expressed by reducing the emitter’s duty cycle and estimating depth from concurrently acquired RGB imagery, only re-enabling the active sensor when reliability is insufficient (Noraky et al., 2018).

These formulations suggest that LoPE is not tied to a single estimator family. Rather, it is a design regime characterized by constrained-state operation, low-overhead observables, and explicit trade-offs between estimation fidelity and system energy. A plausible implication is that LoPE functions as a cross-domain systems concept unifying signal processing, control, machine learning, and hardware architecture under a common resource-aware objective.

A second common feature is the preference for measurements or representations that are already available in the platform. Indoor wireless LoPE relies on RSSI because it is “universally available, simple to collect, and imposes little protocol overhead” (Arif et al., 2024). Software-based embedded power estimation uses “frequency, utilization” traces that low-end meters cannot supply directly but operating systems can log cheaply (Wang et al., 2024). EEG LoPE uses a single Cz channel and the spectral slope of the multitaper power spectrum rather than a large feature stack or deep model (Demirel et al., 2021). These choices reduce acquisition burden as much as inference burden.

A third defining property is architecture-awareness. Some LoPE methods minimize pilots, others compress model size, others reduce memory traffic, and others duty-cycle sensing hardware. TinyTracker is co-designed for Sony’s IMX500 so that computation occurs in the logic die of the image sensor (Bonazzi et al., 2023). Depth estimation for time-of-flight imaging is explicitly designed for the ODROID XU-3 using only Cortex-A7 cores (Noraky et al., 2018). Low-power DNN inference exploits the statistics of quantized tensors to reduce interconnect and memory switching without changing numerical outputs (Bamberg et al., 2023). These examples indicate that LoPE is often inseparable from the execution substrate.

2. Mathematical structure and recurring estimators

Despite the diversity of applications, the mathematical structures used in LoPE are comparatively narrow. Many formulations begin with a standard forward model and then replace expensive or overhead-heavy estimation with an economical surrogate.

In indoor wireless channel estimation, the baseband relation is

$y = h\,x + n,$

with multicarrier form

$\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{n}.$

The target is the channel-gain magnitude inferred from RSSI, with idealized mapping

$|h|^2 \propto \frac{P_r}{P_t},$

and path-loss model

$$P_r[\mathrm{dBm}] = P_t[\mathrm{dBm}] + G_t + G_r - PL(d),$$

$$PL(d) = PL(d_0) + 10 n \log_{10}\!\left(\frac{d}{d_0}\right) + X_\sigma$$

(Arif et al., 2024). Classical LS and MMSE estimators are provided as baselines:

$\hat{h}_{\mathrm{LS}} = y/x,$

and

$$\hat{\mathbf{H}}_{\mathrm{MMSE}} = \mathbf{R}_h \mathbf{X}^H \left(\mathbf{X}\mathbf{R}_h \mathbf{X}^H + \sigma^2 \mathbf{I}\right)^{-1}\mathbf{Y},$$

but the LoPE contribution is to avoid their pilot and statistical burdens through compact FCNNs (Arif et al., 2024).

In rough-environment LLNs, the estimator is deliberately simpler. Received power is modeled as a wide-sense stationary process, and the general MMSE predictor is

$\hat{r}(t+\tau) = \rho_r\,r(t) + \rho_{r'}\,r'(t).$

Under small-horizon linearization, this reduces to the lightweight predictor

$\hat{r}(t+\tau) = r(t) + \tau\,r'(t),$

with finite-difference slope

$r'(t) \approx \frac{r(t)-r(t-\Delta)}{\Delta}.$

For multiple losses,

$\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{n}.$0

This is an archetypal LoPE construction: a principled MMSE model simplified to an $\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{n}.$1 forecaster that preserves enough predictive structure for control (Dargie, 25 Jan 2025).

In software-based power estimation for embedded devices, the methodology again reduces a richer physical process to a low-order predictive form. The chosen per-frequency polynomial is

$\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{n}.$2

with the general energy relation

$\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{n}.$3

The model is trained from long-duration measurements but deployed online at segment granularity using cpufreq transitions and utilization traces (Wang et al., 2024).

In wearable EEG, the estimation problem is spectral rather than geometric or control-oriented. The multitaper PSD is

$\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{n}.$4

$\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{n}.$5

and the feature is the spectral slope obtained by fitting

$\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{n}.$6

over 30–45 Hz (Demirel et al., 2021). Here LoPE relies on feature parsimony rather than architectural compression alone.

A plausible synthesis is that LoPE repeatedly transforms high-dimensional or expensive inference into low-order structure: compact regressors, low-dimensional sufficient statistics, sparse pilot structures, or cheap state recursions.

3. Architectural strategies for reducing energy and overhead

One major branch of LoPE reduces sensing or communication overhead directly. Indoor LP-IoT LoPE replaces “pilot-heavy, statistic-dependent estimators with compact FCNNs that learn directly from low-overhead RSSI” (Arif et al., 2024). In LLNs deployed over water or in rough environments, LoPE closes the ATPC feedback gap during ACK loss by forecasting received power locally rather than waiting for explicit feedback (Dargie, 25 Jan 2025). In LPWAN lifetime estimation, total energy is decomposed into Tx, Rx, computation, standby, and self-discharge terms so that operational and protocol overhead can be budgeted explicitly (Galin-Pons, 2019).

A second branch reduces data movement. TinyTracker on the IMX500 exemplifies this most directly: images are processed “at the source, in the sensor’s logic die,” and the measured timing decomposition is approximately $\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{n}.$7 ms for readout and on-sensor processing, $\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{n}.$8 ms for inference, and $\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{n}.$9 ms for result retrieval, yielding a total latency of approximately $|h|^2 \propto \frac{P_r}{P_t},$0 ms and throughput of approximately $|h|^2 \propto \frac{P_r}{P_t},$1 frames per second (Bonazzi et al., 2023). The same work reports total energy per frame of approximately $|h|^2 \propto \frac{P_r}{P_t},$2 mJ, of which inference contributes approximately $|h|^2 \propto \frac{P_r}{P_t},$3 mJ, implying that the dominant savings arise from architectural collapse of I/O and host processing rather than from kernel-level optimization alone (Bonazzi et al., 2023).

The DNN-inference literature generalizes this principle to internal accelerator datapaths. “Exploiting Neural-Network Statistics for Low-Power DNN Inference” uses lossless fixed-width transforms such as XOR-MSB, sign-magnitude encoding, XOR-ZP, and stream-wise XOR decorrelation to decrease bit probabilities and switching activity across SRAMs, interconnects, and MAC blocks (Bamberg et al., 2023). The reported gains are “up to 80% memory/interconnect energy reduction” and “up to 39% compute energy reduction” with “no loss of accuracy” (Bamberg et al., 2023). This does not reduce model semantics or arithmetic precision at the algorithmic level; it reduces physical switching under the same inference computation.

A third branch duty-cycles expensive sensors. In time-of-flight imaging, LoPE replaces many active TOF measurements with RGB-based depth propagation from the previous depth map and only turns the emitter back on when a RANSAC-based residual test fails (Noraky et al., 2018). The reported outcome is 640×480 depth maps at 30 fps, overall median MRE of $|h|^2 \propto \frac{P_r}{P_t},$4, and an 85% reduction in TOF usage, with estimated total power reduction up to 73% depending on emitter power (Noraky et al., 2018). This suggests that LoPE can be interpreted not only as low-power inference, but also as low-duty-cycle sensing with confidence-triggered refresh.

A fourth branch exploits parallelism to lower clock frequency. On the GAP8 eight-core cluster, optical-flow estimation reaches a speedup factor of $|h|^2 \propto \frac{P_r}{P_t},$5, enabling 529 FPS at 50 MHz in the preferred local-optimization implementation, with approximate energy per frame of $|h|^2 \propto \frac{P_r}{P_t},$6 given a 25 mW platform power (Kühne et al., 2023). The core LoPE idea here is not to shrink the estimator mathematically, but to map a lightweight estimator onto a cluster such that throughput rises while operating frequency falls.

4. Representative application domains

The indoor LP-IoT channel-estimation study provides one of the clearest domain-specific instantiations of LoPE. Two FCNNs are proposed: Model A for time-series prediction at fixed distance and Model B for spatial prediction across receiver locations (Arif et al., 2024). Model A uses a 10-sample sliding RSSI window with a 4-layer topology of input-10, hidden-10, hidden-10, output-1, Leaky ReLU with slope 0.01, NAdam optimization, and 231 trainable parameters. Model B uses 2 input neurons, the same two hidden layers of 10 neurons, 1 output neuron, and 151 trainable parameters (Arif et al., 2024). Reported gains include a 99.02% MSE reduction for Model A and 90.03% for Model B relative to cited prior indoor RSSI prediction benchmarks; detailed testing MSEs are 0.11 dBm and 0.23 dBm in Scenario 1 LoS and NLoS for Model A, and 2.63 dBm and 4.82 dBm in Scenario 2 LoS and NLoS for Model B (Arif et al., 2024). The work also reports lower or comparable test MSE/RMSE than RNN and LSTM on the same dataset, with far smaller parameter count and training time (Arif et al., 2024).

Edge gaze estimation supplies a complementary vision-oriented LoPE case. TinyTracker is derived from iTracker but removes eye crops and face grid input complexity, uses a MobileNetV3 backbone with one extra convolutional layer and two fully connected layers, and is quantized to int8 except for outputs (Bonazzi et al., 2023). The model is approximately 0.6 MB, with 455k parameters and 11.8M MACs, versus iTracker’s approximately 24.6 MB, 6,287k parameters, and 2651M MACs (Bonazzi et al., 2023). The fully quantized model incurs a maximum accuracy loss of 0.16 cm while enabling IMX500 deployment at around 19 ms end-to-end latency (Bonazzi et al., 2023).

Received-power prediction in rough LLNs shows LoPE in a control loop rather than a feedforward inference pipeline. The method is designed for CC2538 and CC1200 radios, with update intervals of approximately 100 ms and 500 ms respectively, and evaluated under more than 30% packet loss in aquatic and coastal deployments (Dargie, 25 Jan 2025). Reported RMSE values for CC2538 range from 6.48% to 15.48% depending on site and horizon, with average prediction accuracy at or above 90% for small horizons in most environments; CC1200 achieves approximately 85% average prediction accuracy for LAG 1 (Dargie, 25 Jan 2025). The practical importance is not merely predictive accuracy but maintaining ATPC performance when acknowledgments are absent.

Other domains broaden the scope of the concept. Low-power radar tracking on the Infineon BGT60TR13D reports distance-estimation error up to $|h|^2 \propto \frac{P_r}{P_t},$7 m and velocity-estimation error $|h|^2 \propto \frac{P_r}{P_t},$8 m/s at power “in the range of 10s of milliwatts,” using simple FFT processing and peak detection (Ronco et al., 2023). Satellite pose estimation on a Coral Dev Board Mini reaches 7.7 frames per second at approximately 2.2 W using MobileDet plus EfficientNet-Lite and an EPnP+RANSAC+LM solver, with quantized pipelines preserving much of the accuracy of larger float models (Lotti et al., 2022). Single-channel EEG arousal estimation achieves more than 80% accuracy and can run on devices with minimum RAM of 512 KB and 55 mJ average energy consumption using multitaper PSD and a thresholded spectral slope (Demirel et al., 2021). These examples show that LoPE is neither confined to learning-based methods nor to wireless communications.

5. Complexity, metrics, and evaluation criteria

LoPE studies tend to report a mixture of task-level accuracy and systems-level efficiency, and the latter is often decisive. In the indoor LP-IoT study, Model A requires approximately 210 MACs for linear layers and total well under 1k operations per inference, while Model B requires approximately 130 MACs plus activation overhead and total under approximately 500 operations (Arif et al., 2024). Raw parameter memory in float32 is approximately 0.9 KB for Model A and approximately 0.6 KB for Model B (Arif et al., 2024). The paper states that reported testing times are on the order of milliseconds for the FCNN, suggesting suitability for real-time microcontroller deployment (Arif et al., 2024).

TinyTracker’s evaluation is explicitly end-to-end rather than kernel-only. The comparison below summarizes the platform-level measurements already present in the source material.

Platform	End-to-end latency	End-to-end energy
IMX500	$|h|^2 \propto \frac{P_r}{P_t},$9 ms	$$P_r[\mathrm{dBm}] = P_t[\mathrm{dBm}] + G_t + G_r - PL(d),$$0 mJ
Coral Dev Micro	$$P_r[\mathrm{dBm}] = P_t[\mathrm{dBm}] + G_t + G_r - PL(d),$$1 ms	$$P_r[\mathrm{dBm}] = P_t[\mathrm{dBm}] + G_t + G_r - PL(d),$$2 mJ
Sony Spresense	$$P_r[\mathrm{dBm}] = P_t[\mathrm{dBm}] + G_t + G_r - PL(d),$$3 ms	$$P_r[\mathrm{dBm}] = P_t[\mathrm{dBm}] + G_t + G_r - PL(d),$$4 mJ

These measurements are paired with normalized power-efficiency and MAC/cycle figures, but the broader significance lies in the decomposition: inference energy is only approximately $$P_r[\mathrm{dBm}] = P_t[\mathrm{dBm}] + G_t + G_r - PL(d),$$5 mJ on IMX500, whereas the remainder is dominated by readout, on-sensor processing, and result retrieval (Bonazzi et al., 2023). This reinforces the LoPE thesis that minimizing movement and staging can outweigh reducing arithmetic alone.

Software-based embedded power estimation uses conventional regression metrics—MSE, MAE, and $$P_r[\mathrm{dBm}] = P_t[\mathrm{dBm}] + G_t + G_r - PL(d),$$6—but the design objective is instantaneous power observability from low-cost instrumentation (Wang et al., 2024). The best-performing per-frequency polynomial reports MSE 0.0182, MAE 0.1040, and $$P_r[\mathrm{dBm}] = P_t[\mathrm{dBm}] + G_t + G_r - PL(d),$$7 0.9221 on Jetson Nano benchmarks, corresponding to approximately 92% accuracy relative to long-duration meter-based measurements (Wang et al., 2024). Here the metric of interest is not only predictive error, but portability and instrumentation cost.

Depth-estimation LoPE evaluates quality and duty cycle jointly: overall median MRE $$P_r[\mathrm{dBm}] = P_t[\mathrm{dBm}] + G_t + G_r - PL(d),$$8, MAE 2.11 cm, RMSE 7.63 cm, and duty cycle 15% across several RGB-D datasets (Noraky et al., 2018). This paired reporting is central to the concept, because a high-quality estimator that requires full emitter duty cycle would not satisfy the low-power criterion.

A plausible generalization is that LoPE should be judged on at least two axes simultaneously: a task metric such as MSE, RMSE, BER, EVM, angular error, or reprojection error, and a systems metric such as parameter count, SRAM footprint, MACs, runtime, energy per estimate, duty cycle, or feedback overhead.

6. Design patterns, limitations, and unresolved issues

Several recurrent design patterns emerge from the surveyed material. One is to start from the smallest viable estimator. The indoor LP-IoT work explicitly advises starting with compact FCNNs with two hidden layers of 10 units and expanding only if accuracy justifies extra cost (Arif et al., 2024). The embedded gaze-estimation study similarly emphasizes low-resolution eye crops, dual-eye channel fusion, and stacked 3×3 convolutions rather than larger face-based inputs or multi-branch pipelines (Lemley et al., 2018). These choices reflect a LoPE bias toward minimal sufficient representations.

A second pattern is quantization or fixed-point compatibility. TinyTracker relies on 8-bit post-training quantization (Bonazzi et al., 2023). Satellite pose estimation uses fully quantized TFLite models with Edge TPU execution (Lotti et al., 2022). EEG LoPE recommends Q15/Q31 tapers and fixed-point-friendly FFT processing where possible (Demirel et al., 2021). The DNN-inference coding framework is compatible with int8 quantized accelerators and exploits their bit statistics directly (Bamberg et al., 2023). This suggests that LoPE commonly treats bit-width as a first-class systems variable.

A third pattern is explicit adaptation to nonstationarity. Indoor RSSI models are trained in LoS and NLoS conditions and can be periodically fine-tuned or domain-adapted when moving across rooms or after furniture rearrangements (Arif et al., 2024). LLN received-power prediction re-estimates slope when new measurements arrive and advises shorter horizons under more dynamic conditions (Dargie, 25 Jan 2025). TOF depth propagation re-enables the emitter whenever confidence collapses (Noraky et al., 2018). Thus, LoPE is often dynamic rather than static compression.

The literature also states its limitations clearly. The indoor LP-IoT study is confined to a single indoor lab environment, a single device pair, fixed distance for time-series prediction, and modest location diversity; BER and throughput impact are not explicitly reported (Arif et al., 2024). TinyTracker is a case study in 2D gaze estimation on specific platforms, and its performance claims center on IMX500, Coral, and Spresense rather than broader deployment classes (Bonazzi et al., 2023). The rough-environment LLN predictor assumes strong short-lag autocorrelation and degrades at longer horizons, harsher dynamics, and lower sampling rates (Dargie, 25 Jan 2025). Software-based power estimation requires board-specific profiling even if the methodology is portable (Wang et al., 2024). TOF depth propagation assumes rigidity and loses efficiency under non-rigid motion, low texture, fast motion, or heavy disocclusion (Noraky et al., 2018).

Some common misconceptions can therefore be rejected. LoPE does not imply deep learning; several core examples are linearized predictors, geometric solvers, multitaper spectral estimators, or control-theoretic constructions (Dargie, 25 Jan 2025, Noraky et al., 2018, Demirel et al., 2021, Zhao et al., 2 Sep 2025). LoPE also does not imply sacrificing accuracy categorically; several studies report near-baseline or state-of-the-art performance at much lower computational or sensing cost (Bonazzi et al., 2023, Lotti et al., 2022, Noraky et al., 2018). Conversely, LoPE does not mean that low arithmetic count alone is sufficient; in-sensor processing and memory/interconnect reduction demonstrate that data movement often dominates the energy budget (Bonazzi et al., 2023, Bamberg et al., 2023).

A plausible implication is that future work in LoPE will continue moving from isolated algorithmic compression toward full-stack estimation design: measurement choice, model topology, memory representation, execution substrate, control policy, and sensor scheduling considered jointly rather than separately.