Abstract
Following the Unlimited Sampling strategy to alleviate the omnipresent dynamic range barrier, we study the problem of recovering a bandlimited signal from point-wise modulo samples, aiming to connect theoretical guarantees with hardware implementation considerations. Our starting point is a class of non-idealities that we observe in prototyping an unlimited sampling based analog-to-digital converter. To address these non-idealities, we provide a new Fourier domain recovery algorithm. Our approach is validated both in theory and via extensive experiments on our prototype analog-to-digital converter, providing the first demonstration of unlimited sampling for data arising from real hardware, both for the current and previous approaches. Advantages of our algorithm include that it is agnostic to the modulo threshold and it can handle arbitrary folding times. We expect that the end-to-end realization studied in this paper will pave the path for exploring the unlimited sampling methodology in a number of real world applications.
TL;DR
This paper bridges the gap between the theory of "Unlimited Sampling"—which uses modulo arithmetic to bypass the dynamic range limits of standard digitizers—and real-world hardware. The authors introduce a robust Fourier-Prony recovery algorithm that handles hardware non-idealities and demonstrate its efficacy using a custom-built prototype ADC capable of recovering signals 24 times larger than its native voltage threshold.
What Problem Does This Paper Address?
Standard Analog-to-Digital Converters (ADCs) suffer from a "dynamic range barrier": if an input signal exceeds a certain voltage threshold ($\lambda$), the output clips, destroying information. While the "Unlimited Sensing Framework" (USF) theoretically solves this by using a modulo operation to "fold" high-amplitude signals back into a measurable range, translating this to hardware is difficult.
Real-world hardware is messy. Modulo circuits often exhibit non-idealities like delayed folding times, inaccurate thresholds, or spurious jumps. Previous recovery algorithms relied on high-order finite differences, which are extremely sensitive to these hardware artifacts. This paper addresses the need for a recovery method that is robust to hardware imperfections and can operate at lower, more practical sampling rates.
Methods
The authors propose a Fourier-Prony Recovery Algorithm. Instead of working purely in the time domain with finite differences, they move the problem into the Fourier domain.
- Signal Modeling: The signal is treated as a periodic, bandlimited function. The "residue"—the difference between the original signal and its modulo-folded version—is a piecewise constant function.
- Spectral Estimation: In the Fourier domain, the discontinuities caused by folding (even non-ideal ones) manifest as a sum of complex exponentials. This transforms the reconstruction task into a classic spectral estimation problem.
- Prony’s Method / Matrix Pencil: High-resolution spectral estimation techniques are used to identify the exact "folding instants" and amplitudes.
- Threshold Agnosticism: Unlike previous methods, this approach doesn't need to know the exact modulo threshold ($\lambda$) used by the hardware, making it much more resilient to component tolerances and thermal drift.
Datasets and Evaluation
As an experimental hardware study, the evaluation uses real-world signals captured by a custom US-ADC prototype:
- Synthetic Signals: Randomly generated trigonometric polynomials and Dirichlet kernels used to test backwards compatibility.
- Burst Signals: Signals with clustered folds to test "super-resolution" capabilities.
- Real-World HDR Signal: The 50 Hz AC voltage from a UK mains power socket (~48V peak-to-peak).
- Metrics: Performance was measured using Mean Squared Error (MSE) compared against ground truth from a high-end 8-bit oscilloscope (DSO-X 3024A).
Key Insights
- Hardware Non-idealities as "Spikes": Hardware errors, such as delayed folds, contribute to the same "out-of-band" spectral components as ideal folds in the Fourier domain.
- Exploiting Out-of-Band Space: By analyzing Fourier coefficients where the bandlimited signal should be zero, the algorithm isolates folding-induced "noise" to reconstruct the original signal.
- Lowering Sampling Requirements: By using spectral estimation, accurate recovery is possible at sampling rates closer to the Nyquist limit, unlike previous USF theory which required high oversampling for finite differences.
Key Results
The prototype hardware successfully recovered signals with amplitudes up to 24 times the ADC's native threshold. A critical finding was the algorithm's robustness: in comparative tests, the Fourier-Prony method achieved an MSE of $0.343 \times 10^{-3}$, while previous state-of-the-art algorithms failed (MSE of 3.7) due to hardware non-idealities.
The system also successfully digitized UK mains voltage using a circuit with a much smaller native range, demonstrating a practical path for HDR sensing in power electronics.
Limitations and Future Work
The current algorithm assumes the number of folds ($M$) is known or can be estimated; future work will focus on robustly estimating $M$ from data. Additionally, while robust to folding errors, a formal analysis of performance under additive Gaussian noise is needed. The authors also suggest hybrid methods combining Fourier domain correction with standard USF for computational efficiency.
Summary updated