Unlimited Sampling of Multiband Signals: Single-Channel Acquisition and Recovery

Gal Shtendel, Ayush Bhandari · 2025

Read Unlimited Sampling of Multiband Signals: Single-Channel Acquisition and Recovery by Gal Shtendel, Ayush Bhandari (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-27110.txt.

Introduction

Gal Shtendel and Ayush Bhandari tackle recording signals that span several frequency bands without clipping loud parts. Standard sampling folds when signals exceed sensor range, losing the peaks. The unlimited sensing framework instead records modulo-folded samples, meaning values wrap around a fixed range, and recovers the original signal afterward by unwrapping.

Working with a low-complexity single-channel setup, the pair prove recovery guarantees showing sampling below the classic Nyquist rate is achievable in this setting. They also tighten a prior sampling theorem for bandpass signals. Hardware experiments with up to six spectral bands show up to a thirteenfold gain in dynamic range. The results enable high-range multiband capture where oversampling and clipping once blocked practical use.

Catalog record

CREATORGal Shtendel, Ayush Bhandari
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE30 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-27110

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Unlimited Sampling of Multiband Signals: Single-Channel Acquisition and Recovery Thanks: This work was supported in part by the the UKRI’s HASC Program under Grant EP/X040569/1, in part by European Research Council’s Starting Grant for “CoSI-Fold” under Grant 101166158, and in part by UK Research and Innovation council’s FLF Program “Sensing Beyond Barriers via Non-Linearities” MRC Fellowship under Award MR/Y003926/1. The authors are with the Department of Electrical and Electronic Engineering, Imperial College London, South Kensington, London SW7 2AZ, U.K. (Email: {g.shtendel21, a.bhandari}@imperial.ac.uk or [email protected]). Gal Shtendel, Ayush Bhandari Abstract In this paper, we address the problem of reconstructing multiband signals from modulo-folded, pointwise samples within the Unlimited Sensing Framework (USF). Focusing on a low-complexity, single-channel acquisition setup, we establish recovery guarantees demonstrating that sub-Nyquist sampling is achievable under the USF paradigm. In doing so, we also tighten the previous sampling theorem for bandpass signals. Our recovery algorithm demonstrates up to a 13x dynamic range improvement in hardware experiments with up to 6 spectral bands. These results enable practical high-dynamic-range multiband acquisition in scenarios previously limited by dynamic range and excessive oversampling. I Introduction Multiband (MB) signals are characterized by spectra consisting of multiple disjoint frequency bands, as shown in Fig. 1 . Such signals arise in a wide range of applications, including communications [ 1 ] , radar [ 2 ] , and cognitive radio [ 3 ] . When sampling MB signals, applying the Shannon–Nyquist theorem directly requires treating them as bandlimited to their total spectral span. This is often impractical, as it leads to prohibitively high sampling rates for analog-to-digital converters (ADCs). Mathematically, this excessive sampling rate represents the cost of ignoring the spectral gaps inherent in the Fourier structure of MB signals. When these gaps are carefully exploited, they enable sub-Nyquist sampling strategies [ 4 , 5 , 6 , 7 , 8 ] . Fig. 1: The spectrum of a multiband signal x(t) (orange), and its corresponding modulo folded signal \mathscr{M}_{\lambda}({x(t)}) (blue), which is not bandlimited. However, beyond sampling rate limitations, ADC quantization imposes a fundamental trade-off between high dynamic range (HDR) and high digital resolution (HDRes), a constraint that becomes particularly critical in MB scenarios where both strong and weak signal components coexist. In such cases, conventional ADCs may either saturate on strong signals or fail to capture weaker ones, giving rise to the well-known near–far problem [ 9 , 1 ] . Under a fixed bit budget [ 10 ] , one can optimize for either HDR or HDRes, but not both simultaneously, creating a foundational bottleneck at the interface between the analog and digital domains. The Unlimited Sensing Framework (USF) [ 11 , 12 , 13 ] introduces a novel digitization paradigm that breaks through the fundamental HDR–HDRes trade-off imposed by conventional ADCs, achieving both simultaneously under a fixed bit budget. At the core of USF is the modulo ADC ( \mathscr{M} – ADC ) [ 13 , 14 ] , which folds the analog input signal into a low dynamic range before digitization. This folding prevents saturation and effectively enhances resolution within the given bit constraints. Signal reconstruction is then performed algorithmically to recover the original input. In practice, USF has demonstrated the ability to recover signals that are 60x larger [ 14 ] than the folding threshold \lambda . Its advantages over traditional approaches have been validated across several domains, including computed tomography [ 15 ] , radar [ 16 ] , and communications [ 17 ] . USF naturally requires redundancy for signal recovery due to its non-linear folding mechanism. This appears fundamentally at odds with the sub-Nyquist sampling philosophy, which seeks to minimize redundancy by operating below the Nyquist rate. At first glance, the two approaches seem contradictory. However, a deeper insight—explored in recent work—is that when signals possess additional structure, such as bandpass structure [ 18 ] or complex exponentials [ 19 ] , USF can indeed enable sub-Nyquist recovery despite its many-to-one mapping. An often overlooked aspect is that, even in the absence of noise, sub-Nyquist sampling typically demands high digital resolution. Without sufficient bit depth, algorithmic recovery is easily compromised by noise folding. USF offers a distinct advantage in this setting: for a fixed bit budget, it enhances effective quantization resolution [ 14 ] , reducing the reliance on excessive bit depth. This makes USF-enabled sub-Nyquist schemes particularly attractive for real-world deployment [ 19 ] . Contributions. In this paper, we address the problem of modulo sampling and recovery of MB signals. By leveraging their Fourier structure, we show that modulo unfolding enables HDR and HDRes acquisition beyond the reach of conventional methods. These findings are validated through both numerical simulations and hardware experiments. To reduce hardware complexity and mitigate challenges related to calibration and synchronization in multi-channel systems, we adopt a single-channel design well-suited for compact, resource-constrained systems. The proposed architecture is backwards compatible with previous frameworks, such as hardware-demodulated radar systems, single-channel MB sampling [ 20 , 21 ] , and multi-channel periodic nonuniform sampling methods [ 7 , 5 ] . Our key contributions are: 1. Theoretical guarantee for MB signal recovery from modulo samples, based on effective bandwidth. Moreover, the result tightens the existing USF-bandpass sampling theorem [ 18 ] . 2. Recovery algorithm for USF-based sub-Nyquist sampling. 3. Hardware validation showing up to 13 x dynamic range improvement using 7 -bit quantized measurements, achieving clear gains over existing methods. Notation. The sets of real, complex-valued, and integer numbers are denoted by \mathbb{R},\mathbb{C} , and \mathbb{Z} , respectively, and \mathbbm{I}_{N}=\left\{0,\ldots,N-1\right\} is the set of N contiguous integers. Set cardinality is \# . Continuous functions are denoted by f(t) with a Fourier transform (when it exists) denoted \widehat{f}\left(\omega\right) . We define {||f||}_{\infty}\stackrel{{\scriptstyle\rm{def}}}{{=}}\mathrm{sup}_{t\in\mathbb{R}}\absolutevalue{f(t)} for functions, and {||f||}_{\infty}\stackrel{{\scriptstyle\rm{def}}}{{=}}\max_{k}{\absolutevalue{f\sqb{k}}} for sequences. The norm \norm{f}_{1} refers to the L^{1} norm if f\in L^{1} , or \ell^{1} norm if f\in\ell^{1} . II Towards Recovery Signal Model. We consider complex-valued MB signals x\left(t\right)=\sum\limits_{p=0}^{P-1}\varphi_{p}\left(t\right)e^{-\jmath\omega_{p}t},\quad\varphi_{p}\in\mathcal{B}_{\Omega_{\mathsf{B}}}, (1) comprising P baseband signals \left\{\varphi_{p}\right\}_{p=0}^{P-1} of bandwidth 2\Omega_{\mathsf{B}} , which are modulated to frequencies \left\{\omega_{p}\right\}_{p=0}^{P-1} . The support of the Fourier transform \supp(\widehat{x}) is assumed to be disjoint , that is \absolutevalue{\oc- \omega_{q}}>\Omega_{\mathsf{B}},\ \forall p\neq q,\ p,q\in\mathbb{I}_{P} , as illustrated in Fig. 1 . Problem Formulation. The \mathscr{M} – ADC produces samples y\left[k\right]=\mathscr{M}_{\lambda}({x})\big|_{t=kT_{\mathsf{S}}},\quad k\in\mathbb{Z}, (2) where \mathscr{M}_{\lambda}({\cdot}) is the centered modulo operator defined by \mathscr{M}_{\lambda}({f}):f\left(t\right)\mapsto\left({\frac{\lambda}{\pi}}\right)\angle\exp\left({\jmath\frac{\pi}{\lambda}f\left(t\right)}\right) with threshold \lambda>0 . For a complex-valued input z\in\mathbb{C} , the folded signal is defined as y=\mathscr{M}_{\lambda}({\Re{z}})+\jmath\mathscr{M}_{\lambda}({\Im{z}}) as in [ 22 ] . The corresponding inverse problem amounts to recovering x\left[k\right]=x\left(t\right)\big|_{t=kT_{\mathsf{S}}} from the discrete modulo samples y\left[k\right] . Solution Approach. Our recovery leverages the separation of modulo samples into their smooth and non-smooth components based on the modulo decomposition property [ 12 ] , y\left[k\right]=x\left[k\right]-r\left[k\right],\quad r\left[k\right]\in 2\lambda\mathbb{Z}, (3) where x and r denote the MB and residual parts, respectively. Though smoothness can be measured via a digital derivative filter 1 1 1 The filter \Delta is used in the US – Alg unfolding method presented in [ 12 ] , to shrink the amplitudes of \Omega -bandlimited signals for sufficiently small T_{\mathsf{S}} . \Delta=\begin{bmatrix}-1&1\end{bmatrix} , this would be non-ideal for modulated signals . Therefore, we redefine the filter to accommodate local carrier frequencies \left\{\omega_{p}\right\}_{p=0}^{P-1} , which achieves more precise shrinkage.

Read the complete text