PaperlayerView exact revision on arXiv

Source-linked AI summary

From Theory to Practice: Sub-Nyquist Sampling of Sparse Wideband Analog Signals

Moshe Mishali, Yonina C. Eldar

arXiv:0902.4291v3cs.IT

TL;DR

The paper tackles blind sub-Nyquist acquisition of multiband analog signals whose sparse spectral support is unknown, while seeking practical hardware and efficient digital processing. It proposes the modulated wideband converter and a digital architecture for recovery or direct low-rate band processing. The system achieves unique recovery under stated conditions and, in a high-SNR design example, does so with at least 35 channels at less than 18% of the Nyquist rate.

  • Problem

    Blind sub-Nyquist sampling is needed for multiband analog signals with unknown or time-varying frequency support, whose Nyquist rates can exceed available ADC specifications by orders of magnitude.

  • Method

    The modulated wideband converter uses periodic modulation, lowpass filtering, and low-rate uniform sampling, followed by digital processing for reconstruction or low-rate extraction of individual bands.

  • Results

    The proposed system uniquely determines multiband inputs under stated parameter conditions; in a high-SNR example, correct recovery uses at least 35 channels, less than 18% of the Nyquist rate.

  • Takeaways & Limitations

    The architecture supports processing any transmitted band without first interpolating to the high Nyquist rate and can accommodate time-varying spectral support.

Abstract

from arXiv · show

Conventional sub-Nyquist sampling methods for analog signals exploit prior information about the spectral support. In this paper, we consider the challenging problem of blind sub-Nyquist sampling of multiband signals, whose unknown frequency support occupies only a small portion of a wide spectrum. Our primary design goals are efficient hardware implementation and low computational load on the supporting digital processing. We propose a system, named the modulated wideband converter, which first multiplies the analog signal by a bank of periodic waveforms. The product is then lowpass filtered and sampled uniformly at a low rate, which is orders of magnitude smaller than Nyquist. Perfect recovery from the proposed samples is achieved under certain necessary and sufficient conditions. We also develop a digital architecture, which allows either reconstruction of the analog input, or processing of any band of interest at a low rate, that is, without interpolating to the high Nyquist rate. Numerical simulations demonstrate many engineering aspects: robustness to noise and mismodeling, potential hardware simplifications, realtime performance for signals with time-varying support and stability to quantization effects. We compare our system with two previous approaches: periodic nonuniform sampling, which is bandwidth limited by existing hardware devices, and the random demodulator, which is restricted to discrete multitone signals and has a high computational load. In the broader context of Nyquist sampling, our scheme has the potential to break through the bandwidth barrier of state-of-the-art analog conversion technologies such as interleaved converters.

I. INTRODUCTION

The paper addresses blind sub-Nyquist sampling of sparse multiband analog signals with unknown frequency support, targeting practical hardware and efficient low-rate processing. It introduces the modulated wideband converter and a supporting digital architecture for recovery and band-specific processing.

  • Wideband multiband signals can have Nyquist rates exceeding available ADC specifications by orders of magnitude, motivating structure-aware acquisition.
  • The blind-recovery problem is challenging when carrier frequencies are unknown or time-varying, because the receiver lacks prior frequency-support information.
  • The modulated wideband converter multiplies the input by periodic waveforms, lowpass filters the products, and samples them uniformly at a low rate.
  • With appropriate waveform period and sampling-rate parameters, the MWC uniquely determines multiband input signals and can trade channel count against per-branch rate and processing.
  • The digital architecture supports analog reconstruction or low-rate sequences for individual bands, including inputs whose spectral support varies over time.
  • Simulations examine robustness to noise and mismodeling, hardware simplifications, adaptation to time-varying support, and other engineering aspects.

B. Multicoset using practical ADCs

Multicoset sampling lowers the average sampling rate, but practical ADC bandwidth limits distort wideband inputs and undermine its implementation. The MWC replaces time shifts with periodic analog mixing, filtering, and synchronized low-rate sampling using available devices.

  • Multicoset sampling: Multicoset sampling uses m<M cosets on a periodic nonuniform grid, reducing the average rate to m/(MT), below the Nyquist rate.
  • Practical ADC limitations: Practical ADCs attenuate and distort spectral content beyond their analog bandwidth, so multicoset samples no longer equal pointwise samples of wideband signals.The cited example gives front-end bandwidth up to b=780 MHz.
  • Practical ADC limitations: Multicoset implementation requires a specialized wideband ADC and still samples each coset at the nonstandard rate b/M, wasting available conversion resources.
  • Practical ADC limitations: Accurate time shifts on the order of the Nyquist interval are difficult, and compensating timing mismatches adds substantial receiver complexity.
  • Modulated wideband converter: The MWC multiplies the signal by periodic waveforms, lowpass filters each output, and samples synchronously at a low rate, avoiding analog bandwidth issues and nonzero time synchronization.Its channels use different mixtures so that the resulting measurements can recover a relatively sparse multiband signal.
  • Modulated wideband converter: The MWC uses periodic mixing functions, including piecewise-constant ±1 waveforms, with design parameters consisting of channel count, period, sampling rate, and mixing functions.The analog mixer and filter operate on wideband signals, so this preprocessing must remain analog.

B. Frequency domain analysis

The frequency-domain analysis expresses each sampled channel as a mixture of shifted copies of the unknown spectrum and organizes those copies into a sparse vector. Parameter choices and mixing functions determine the slice structure, recovery conditions, implementation burden, and noise sensitivity.

  • Signal-to-sample relation: Periodic mixing produces a finite combination of fp-shifted copies of X(f), which the lowpass filter restricts to the retained frequency interval.
  • Signal-to-sample relation: The DTFTs of the sampled sequences are related linearly to the unknown vector z(f), whose entries are shifted slices of X(f).Determining z(f) over the fundamental interval is sufficient to recover x(t).
  • Parameter roles: Choosing fp≥B ensures that each band contributes at most one nonzero element to z(f), making z(f) at most N-sparse for a given frequency.In practice fp is chosen slightly above B to avoid edge effects.
  • Parameter roles: The number of channels m sets the overall sampling rate mfs, while fp and fs determine L, the number of spectrum slices that may contain signal energy.The simple choice fs=fp≈B controls the sampling rate at a resolution of fp.
  • Mixing functions: Different periodic mixing functions generate rows of the measurement matrix, and their differing Fourier coefficients are intended to produce linearly independent mixtures of the spectrum slices.
  • Implementation and robustness: For sign-alternating mixers, M≥L is a blind-recovery condition, while M also determines the shift-register length and clock rate.
  • Implementation and robustness: Periodicity permits calibration of repeated waveform imperfections, but sign-alternating mixers weight spectrum regions asymmetrically and make noise sensitivity depend on band location.

C. Choice of parameters

The sampling parameters must satisfy necessary and sufficient conditions for unique blind recovery, while allowing trade-offs between minimum rate and hardware complexity.

  • Blind and non-blind modes: The architecture can use half as many channels when band locations are known, enabling non-blind reconstruction with m ≥ N.The same parameter selection supports both blind and non-blind modes, with non-blind reconstruction requiring m ≥ N.
  • Necessary conditions: Necessary blind-recovery conditions include fs ≥ fp, m ≥ 2N, and M ≥ Mmin for sign-waveform mixing.With fp = B, these conditions are required for exact spectrum-blind recovery of arbitrary multiband signals.
  • Sufficient conditions: Sufficient recovery requires fs ≥ fp ≥ B, a bounded fs/fp ratio, M ≥ Mmin, m ≥ 2N, and linear independence of every 2N columns of SF.Under these conditions, z(f) is the unique N-sparse solution for every f ∈ Fs.
  • Minimum-rate design: Choosing fs = fp = B and m = 2N yields the minimum average sampling rate 2NB for blind recovery.For the representative Option A, this corresponds to 615 MHz versus the minimal 2NB = 600 MHz.
  • Parameter trade-offs: Hardware and rate constraints can be traded: increasing channel sampling rate can resolve cases where channel-count requirements conflict with M ≤ 2m−1.An alternative uses narrower conceptual bands and more channels, while higher fs can reduce the required number of channels.

D. Trading channels for sampling rate

The design trades physical channel count against per-channel sampling rate by digitally expanding each sampled channel into multiple lower-rate sequences.

  • Digital expansion: With fs = qfp, each physical channel provides q equations over Fp, making m hardware branches equivalent to mq channels sampled at fp.The expansion uses q sequences per channel and preserves the effective equation count needed for recovery.
  • Digital processing: Truncating the expanded sequences to Fp is beneficial because z(f) is 2N-jointly sparse there, while its joint support may be larger over Fs.This truncation reduces the digital processing burden before reconstruction.
  • Option B: Option B guarantees uniqueness with 3 physical channels instead of the 12-channel setting, at the cost of a higher sampling rate and 15 digitally expanded channels.The example expands 3 channels to 3q = 15 channels.
  • Trade-offs: Theoretical channel collapse to one channel requires fs = mfp, but q digital filters per channel increase computation and larger q requires longer filters.The filter approximation becomes more demanding as the cutoff π/q decreases.

IV. RECONSTRUCTION

Reconstruction recovers the Nyquist-rate sequence or analog signal from the sampled sequences and can also produce low-rate outputs for individual bands.

  • Inputs: The reconstruction stage takes the sample sequences yi[n] or their decimated sequences ỹi,k[ñ] as input.These sequences are the digital inputs to the recovery architecture.
  • Reconstruction: It recovers the Nyquist-rate sequence x(nT) or its analog version x(t).Thus the architecture supports either digital or analog reconstruction of the input.
  • Low-rate processing: The same digital architecture outputs low-rate sequences capturing the information in each band without interpolation to the high Nyquist rate.This enables low-rate processing of a selected transmitted band.

A. IMV model

The IMV model represents infinitely many linear systems whose solution vectors share a sparse support, enabling support recovery before reconstructing the full vector collection.

  • Model: The model uses an m × M matrix A with m < M and a parameterized family indexed by a possibly infinite set Λ.The vectors u(λ) solve the associated linear systems for every λ ∈ Λ.
  • Joint sparsity: The collection u(Λ) is jointly K-sparse when the union of its nonzero indices contains at most K entries.Every vector’s nonzero entries lie within one shared set of at most K indices.
  • Known support: If the support S is known and AS has full column rank, the vectors can be recovered from the corresponding measurements using the pseudoinverse of AS.AS contains the columns of A indexed by S.
  • Unknown support: For unknown support, exact invertibility requires K to be known and every set of 2K columns of A to be linearly independent.Finding the support is generally NP-hard, motivating sparse-recovery algorithms.
  • Recovery strategy: The IMV approach constructs a finite frame, recovers the common support from a finite-dimensional system, and then reconstructs u(Λ).This isolates the infinite-cardinality aspect in frame construction and reduces recovery to one finite-dimensional compressive-sensing problem.

B. Multiband reconstruction

The reconstruction stage converts recovered spectral slices into either the original analog signal or time-domain outputs, with a finite frame replacing an otherwise noncausal construction.

  • Support recovery: The CTF block constructs a frame from the sample sequences and recovers the active support S of the spectral representation.The frame can be formed from finitely many linearly independent samples because sparsity bounds its rank.
  • Slice recovery: After support recovery, the sequences z_i[n] provide the inverse-DTFT representations of the spectral slices at the input rate f_s.These sequences form the basis for subsequent analog or digital reconstruction.
  • Digital reconstruction: Digital reconstruction zero-pads and interpolates the slice sequences to the Nyquist rate, then shifts and sums them to generate x(nT).An analog lowpass can recover x(t) from the resulting Nyquist-rate sequence when that rate is manageable.
  • Analog reconstruction: An alternative uses analog lowpass filters on each slice sequence and combines the resulting real and imaginary components to recover x(t) entirely in the time domain.The paper emphasizes that the recovery flow is time-domain based despite frequency-domain analysis.

C. Architecture and advantages

The architecture separates support estimation from realtime slice processing, enabling low-rate outputs while accommodating changing spectral support and parameter inaccuracies.

  • Architecture: The architecture expands input sequences when needed, triggers CTF support estimation on initialization or support changes, and outputs a low-rate sequence for each active spectral slice.Support changes can be signaled by an application or detected by monitoring inactive slice sequences.
  • Computational trade-offs: Support recovery is computationally difficult because the CTF requires solving an MMV system that is NP-hard, so practical polynomial-time methods trade tractability for higher sampling rates.This is an explicit computational limitation of the recovery block.
  • Realtime processing: Once support is known, closed-form recovery can run in realtime, avoiding a CS problem whose dimensions would otherwise be dictated by the ambient Nyquist dimension.The CTF runs only when support changes, and about 2N input vectors are stored to bridge its update delay.
  • Robustness: The recovery flow depends mainly on clock ratios rather than exact f_s and f_p values, and numerical results show stability in noise.The design uses f_s ≥ f_p ≥ B and guard regions to tolerate hardware inaccuracies or signal mismodeling.
  • Low-rate processing: Compared with multicoset processing, the architecture can operate at the input rate f_s instead of requiring interpolation to the Nyquist rate.The paper identifies digital processing at f_s as preferable to a Nyquist-rate processor.

D. Choosing the sign patterns

Sign patterns determine the sampling matrix conditions needed for sparse support recovery, while practical recovery uses polynomial-time relaxations and simulation-based assessment.

  • Design conditions: Uniqueness requires every 2N columns of the sampling matrix to be linearly independent, while CTF application strengthens this to every 4N columns.Checking all such rank conditions is computationally difficult.
  • Random designs: A random sign matrix satisfies the RIP of order K with high probability when m ≥ CK log(M/K), providing a practical theoretical design guideline.The paper notes that the constant C is positive and independent of the other parameters.
  • Recovery algorithms: Basis pursuit replaces combinatorial sparsity search with a convex program solvable by polynomial-time methods, with recovery and noise-error guarantees under suitable RIP conditions.The cited condition includes δ2K ≤ 2−1.
  • Practical validation: In implemented systems, the sampling matrix is fixed and its RIP constant cannot be calculated efficiently, so simulations are used to assess practical stability without guaranteeing a target RIP value.Binary feedback implementations may also provide far fewer possible sign patterns than 2^M.

V. NUMERICAL SIMULATIONS

Numerical experiments evaluate noisy multiband recovery, hardware simplifications, channel reduction, time-varying support, and implementation-related effects using synthesized wideband signals.

  • Experimental scope: The experiments cover wideband noise, shared shift-register hardware, channel collapsing, fast adaptation to time-varying support, and quantization effects.The study uses numerical simulations to examine several engineering aspects of the proposed system.
  • Design example: The test set contains 500 noisy multiband signals with 3 band pairs, total N = 6, bandwidth B = 50 MHz, and carriers drawn across a 10 GHz Nyquist bandwidth.White Gaussian noise contaminates the synthesized signals.
  • Design example: The baseline simulation uses f_s = f_p = f_NYQ/195 ≃ 51.3 MHz and m = 100 randomly generated channels.Each mixing function alternates sign at most M = 195 times.
  • Recovery procedure: Support recovery uses dominant eigenvectors, frame construction, and MMV solving, with correct recovery also allowing a few extra linearly independent support entries.The simulations threshold negligible eigenvalues at 10^-9.
  • Recovery results: 35 channels achieve correct recovery in the high-SNR regime, corresponding to less than 18% of the Nyquist rate.The observed threshold is consistent with a prediction of approximately 30 channels for stable recovery.

C. Collapsing analog channels

The experiments evaluate hardware-channel collapse, time-varying support, and quantized samples. They show reduced-channel recovery, short support-estimation delays, and successful support recovery with few quantization bits.

  • Channel collapse: 7 channels achieve an acceptable recovery rate after collapsing q=5 sampling channels, implying substantial hardware savings.The configuration uses 35/q=7 channels, with performance trends remaining similar across SNR levels and channel counts.
  • Time-varying support: The time-varying-support experiment uses a 20-vector memory for a 50-sample CTF frame, causing temporary invalid low-rate sequences after support changes.Sequences remain valid for 20 cycles and become invalid for 30 cycles while the CTF produces a new support estimate; normal operation should use NMEM ≥ NCTF.
  • Time-varying support: Once the new support estimate is ready, the baseband error and reconstruction become correct.The experiment intentionally uses NMEM < NCTF to expose the transient error; NCTF can be lower under other SNR and channel settings.
  • Quantization: Support recovery functions properly from quantized samples represented with only a few bits.The experiment isolates quantization by removing additive wideband noise and uses uniformly spaced quantization across the samples’ dynamic range.

VI. DISCUSSION

The discussion contrasts the MWC with the random demodulator and emphasizes the paper’s practical bridge from sub-Nyquist theory to engineering implementation. The MWC targets analog multiband signals while supporting low-rate processing and practical hardware choices.

  • Comparison with the random demodulator: The random demodulator mixes with a high-rate pseudorandom sign waveform, integrates, and samples at a constant low rate.Its analysis recovers signals represented as finite-grid multitone functions through a linear system.
  • Comparison with the random demodulator: For analog signals, the random demodulator requires about Nyquist-rate tones and a huge sensing matrix, creating high computational demands.The comparison states that approximating analog signals leads to millions of rows and tens of millions of columns in Φ.
  • Hardware comparison: The MWC accommodates arbitrary periodic waveforms through recalculated Fourier coefficients, whereas the random demodulator is more tailored to sign waveforms.The hardware comparison also contrasts accurate integration with flexible analog filter design.
  • Digital architecture: The MWC uses compressed sensing only for support recovery, reducing recovery complexity and enabling low-rate processing.The discussion contrasts this with the random demodulator’s Nyquist-rate recovery objective.
  • Concluding remarks: The paper presents the MWC as a practical system for spectrum-blind reconstruction and processing any information band at a low rate.Its digital support-recovery stage needs few observations and introduces a short delay before realtime computations become possible.
  • Concluding remarks: Future work will sharpen the theoretical understanding and report on circuit-level implementation.The paper identifies engineering aspects as its prime focus, including hardware simplifications, low-rate and realtime processing, time-varying spectrum, and quantization.
Loading 0902.4291v3…