PaperlayerView exact revision on arXiv

Source-linked AI summary

Sub-Nyquist Sampling for Power Spectrum Sensing in Cognitive Radios: A Unified Approach

Deborah Cohen, Yonina C. Eldar

arXiv:1308.5149v1cs.ITmath.IT

TL;DR

Wideband spectrum sensing is demanding because Nyquist-rate sampling can strain ADC bandwidth and processing. This paper reconstructs power spectra from sub-Nyquist samples across analog and digital models, deriving minimal noiseless rates and recovery methods. Simulations show sub-Nyquist power-spectrum sensing outperforms sub-Nyquist spectrum sensing and can match Nyquist-rate power-spectrum detection up to an SNR threshold.

  • Problem

    Wideband cognitive-radio signals require high Nyquist sampling rates that can exceed ADC front-end bandwidths and increase sample-processing costs.

    Source: §I · p. 1

  • Method

    The paper reconstructs signal power spectra from sub-Nyquist samples across analog and digital models, deriving minimal noiseless sampling rates and recovery techniques for sparse and non-sparse cases.

    Source: §I · pp. 1–2

  • Results

    Sub-Nyquist power-spectrum sensing outperforms sub-Nyquist spectrum sensing and has performance comparable to Nyquist-rate power-spectrum detection up to an SNR threshold.

    Source: §VI · p. 12

  • Takeaways & Limitations

    The analysis treats analog and digital model variations uniformly in the frequency domain, while time-domain analysis is more complex.

    Source: §VI · p. 12

Abstract

from arXiv · show

In light of the ever-increasing demand for new spectral bands and the underutilization of those already allocated, the concept of Cognitive Radio (CR) has emerged. Opportunistic users could exploit temporarily vacant bands after detecting the absence of activity of their owners. One of the crucial tasks in the CR cycle is therefore spectrum sensing and detection which has to be precise and efficient. Yet, CRs typically deal with wideband signals whose Nyquist rates are very high. In this paper, we propose to reconstruct the power spectrum of such signals from sub-Nyquist samples, rather than the signal itself as done in previous work, in order to perform detection. We consider both sparse and non sparse signals as well as blind and non blind detection in the sparse case. For each one of those scenarii, we derive the minimal sampling rate allowing perfect reconstruction of the signal's power spectrum in a noise-free environment and provide power spectrum recovery techniques that achieve those rates. The analysis is performed for two different signal models considered in the literature, which we refer to as the analog and digital models, and shows that both lead to similar results. Simulations demonstrate power spectrum recovery at the minimal rate in noise-free settings and show the impact of several parameters on the detector performance, including signal-to-noise ratio (SNR), sensing time and sampling rate.

I. INTRODUCTION

Cognitive radios need reliable spectrum sensing across wide bands, but Nyquist-rate sampling is costly; this paper reconstructs power spectra from sub-Nyquist samples instead of reconstructing signals. It unifies analog and digital models and analyzes rates and recovery methods across sparse and non-sparse settings.

I. INTRODUCTION

  • Wideband cognitive-radio sensing requires high Nyquist rates, which can exceed ADC front-end bandwidths and increase processing, speed, and power costs.Cognitive radios seek vacant licensed bands by detecting primary-user activity.Source: §I · p. 1
  • Rather than reconstructing the original signal, the paper reconstructs its power spectrum from sub-Nyquist samples for detection.Earlier methods derived rates and recovery techniques for perfect signal reconstruction, although that reconstruction is unnecessary when the goal is sensing.Source: §I · p. 1

A. Analog Model

  • The analog model treats stationary multiband signals whose frequency support lies in continuous bands; the paper’s goal is to recover their power spectrum from sub-Nyquist samples.The spectrum is the Fourier transform of the signal autocorrelation, and its support matches the signal spectrum’s support.Source: §I · p. 1, A. Analog Model · pp. 2–3

I. INTRODUCTION

  • Prior work had not compared the two models or established general sampling-matrix conditions and universal minimal rates across them.The paper positions its framework as bridging the analog and digital models.Source: §I · p. 2
  • The paper frames a unified analysis around analog multiband and digital models, with frequency-domain analysis providing a common route where time-domain analysis is more complex.For the analog model it considers schemes operating on spectrum bins; for the digital model it gives time- and frequency-domain reconstruction approaches.Source: §I · p. 2
  • The paper derives minimal rates for noiseless power-spectrum reconstruction across its scenarios and presents recovery methods that achieve those rates.For each scenario, the required power-spectrum rate is half the rate for perfect signal reconstruction; the analysis concerns generic systems, not optimally designed sampling patterns.Source: §I · p. 2

1) No sparsity assumption:

The paper distinguishes non-sparse, sparse non-blind, and sparse blind cases, then describes a digital multi-tone signal model using Fourier-series support. In the sparse non-blind case, the potentially active frequency support is known; in the blind case, carrier frequencies are not assumed known.

1) No sparsity assumption:

  • In the non-sparse scenario, the signal is not assumed sparse, and NsigB may be on the order of fNyq.No a priori signal knowledge is assumed.Source: 1) No sparsity assumption · p. 3

2) Sparsity assumption and non blind detection:

  • In the sparse non-blind scenario, NsigB ≪ fNyq and the potentially active transmissions’ support is known from the licensed-user standard.The detection algorithm is intended to remain independent of a specific known signal support as primary-user activity varies.Source: 2) Sparsity assumption and non blind… · p. 3

3) Sparsity assumption and blind detection:

  • In the sparse blind scenario, sparsity is assumed but carrier-frequency knowledge is not; only maximum transmission count and bandwidth are assumed known.The signal’s Fourier frequencies are not assumed known a priori.Source: 3) Sparsity assumption and blind… · p. 3

B. Digital Model

  • The digital model represents a continuous-time multi-tone signal over [0, T) using a finite Fourier-series support.Each transmission has a finite set of Fourier coefficients, and the signal is wide-sense stationary.Source: B. Digital Model · p. 3
  • For the digital model, the signal is sampled on a Nyquist grid and its autocorrelation has a Fourier representation whose coefficients share the signal’s support.The model defines N = Q + 1 from the sampling interval and signal duration.Source: B. Digital Model · p. 3

1) No sparsity assumption:

The paper frames detection as identifying active transmissions from sub-Nyquist samples by recovering a full or partial power spectrum. Its analog sampling discussion covers multicoset and MWC schemes and relates the signal spectrum to sample statistics.

1) No sparsity assumption:

C. Problem Formulation

  • Detection tests whether each of the Nsig transmissions is absent or active, using sub-Nyquist samples of x(t).These are formulated as the hypotheses Hi,0 and Hi,1.Source: C. Problem Formulation · p. 3
  • The method reconstructs a full spectrum in the non-sparse and blind sparse cases, but only potentially occupied locations when carrier frequencies are known.It then performs detection on the fully or partially reconstructed spectrum, without sampling or computing Nyquist-rate samples.Source: C. Problem Formulation · p. 3
  • Energy detection on the reconstructed spectrum can identify unoccupied bands from sub-Nyquist samples.The paper notes that other detection forms are also possible after spectrum recovery.Source: C. Problem Formulation · p. 3

A. Analog Model: Sampling and the Analog Spectrum

  • The analog model considers multicoset sampling and the Modulated Wideband Converter, which yield identical expressions relating the signal power spectrum to that of the samples.The section treats whole-spectrum reconstruction, with a later case recovering only potentially occupied locations when carrier frequencies are known.Source: A. Analog Model: Sampling and the… · p. 4

1) Multicoset sampling:

  • Multicoset sampling retains M samples from each block of N consecutive points on the uniform grid.The selected samples are determined by the coset offsets.Source: 1) Multicoset sampling · p. 4
  • The analog reconstruction relation uses the sample DTFTs and a matrix built from the sampling matrix to recover the signal’s power spectrum.The cited passage states that recovering rx(f) over the frequency interval recovers the entire power spectrum.Source: 1) Multicoset sampling · p. 4

2) MWC sampling:

The MWC samples through parallel mixing channels, and the analog power spectrum can be recovered from correlations among the sub-Nyquist samples. The same recovery relation applies to both sampling schemes, with frequency-domain sparsity enabling the minimal rate.

2) MWC sampling:

  • The MWC uses M parallel channels with mixing and lowpass filtering, giving an overall sampling rate of Mfs.The per-channel sampling rate fs is chosen equal to the mixing-function period frequency fp.Source: 2) MWC sampling · p. 4

3) Analog Power Spectrum Reconstruction:

  • Under the full-spark assumption, both sampling schemes share a system that recovers the analog power spectrum from sub-Nyquist-sample correlations.The autocorrelation matrix is diagonal, with entries corresponding to frequency-shifted values of the power spectrum.Source: 3) Analog Power Spectrum Reconstruction · p. 4
  • At each frequency rx(f) has at most Kf nonzeros, while rx[n] is 2Kf-sparse, so the minimal sampling rate is achieved only in the frequency domain.The frequency-domain relation also holds in the time domain, but the minimal rate is not achieved there.Source: 3) Analog Power Spectrum Reconstruction · p. 4

B. Discrete Model: Reconstruction of the Digital Spectrum

For the digital model, the target is the Fourier coefficients of the signal autocorrelation. The paper develops a frequency-domain approach because analyzing recovery conditions for the time-domain formulation is difficult.

B. Discrete Model: Reconstruction of the Digital Spectrum

  • The digital model seeks the Fourier coefficients of the signal autocorrelation, rather than reconstructing the signal itself.The proposed frequency-domain approach exploits equivalent signal properties and permits the same analysis as for the analog model.Source: B. Discrete Model: Reconstruction of… · p. 5

1) Time domain:

  • The time-domain formulation relates the recovery variables to sub-Nyquist-sample covariance through a linear system involving C, but its rank conditions are difficult to analyze.C is expressed using the sampling matrix and a selection matrix.Source: 1) Time domain:  · p. 5

2) Frequency domain:

The frequency-domain digital approach relates sub-Nyquist samples to Nyquist samples through a sampling matrix, then uses sample covariance to recover the digital spectrum. The resulting recovery system can exploit sparsity in the spectrum coefficients.

2) Frequency domain:

  • For a sensing frame, the digital model represents sub-Nyquist samples z as a linear transform of Nyquist samples x using a sampling matrix A.The frame has M=fsT<N samples, and the paper assumes the associated matrix is full spark.Source: 2) Frequency domain · p. 5
  • The covariance relation yields a recovery system for rc, whose nonzero entries correspond to the Kf nonzero Fourier coefficients in the sparse scenarios.The paper derives conditions for this system to have a unique solution.Source: 2) Frequency domain · p. 5
  • The frequency-domain approach shares the analog model's analysis, although the analog model has one equation per frequency and the digital model has a single equation.The paper summarizes the notation used for the two models in Table I.Source: 2) Frequency domain · p. 5, Table I · p. 6

IV. MINIMAL SAMPLING RATE AND RECONSTRUCTION

For full-spark sampling matrices, the analog and digital recovery systems have a unique solution under rank and sampling-dimension conditions. In the non-sparse case, the paper gives a rate range and attributes spectrum recovery to signal stationarity.

1) Minimal Rate for Perfect Reconstruction:

  • For full-spark sampling matrices, the analog and digital systems have unique solutions when M^2≥N and 2M>N.The corresponding full-column-rank proposition gives the same conditions for C.Source: 1) Minimal Rate for Perfect… · p. 6
  • For M≥2, the non-sparse unique-solution range is N/2<M<N.The passage states that the minimal sampling rate is then given, but its expression is not included in the supplied text.Source: 1) Minimal Rate for Perfect… · p. 6
  • Even without signal sparsity, the power spectrum can be retrieved by exploiting stationarity, although the measurement vector z generally has no stationarity constraints.The stated conditions are expressed in terms of the sampling matrix A, which depends directly on the sampling pattern.Source: 1) Minimal Rate for Perfect… · p. 6

2) Power Spectrum Reconstruction:

For sparse signals with known frequency support, the paper reduces spectrum recovery to the potentially nonzero components. Under full-spark and rank conditions, it derives a minimal sampling rate that is half the Landau rate for non-blind spectrum recovery.

2) Power Spectrum Reconstruction:

  • At rates f ≥ f(1), the analog and digital power spectra are recovered using Moore–Penrose pseudoinverses.The corresponding systems are overdetermined.Source: 2) Power Spectrum Reconstruction · p. 6

1) Minimal Rate for Perfect Reconstruction:

  • Known frequency support reduces each model’s recovery problem to its Kf potentially nonzero spectrum components.The analog model’s support can vary with frequency; in the digital model, it equals Sx.Source: 1) Minimal Rate for Perfect… · pp. 6–7
  • For a full-spark T, the reduced matrix D has full column rank when M^2 ≥ Kf and 2M > Kf.This condition supports unique solutions for the analog and digital recovery systems.Source: 1) Minimal Rate for Perfect… · p. 7
  • The minimal sampling rate for non-blind spectrum recovery is half the Landau rate, 2NsigB, for perfect signal reconstruction.The uniqueness conditions require full-spark sampling matrices and M^2 ≥ Kf and 2M > Kf.Source: 1) Minimal Rate for Perfect… · p. 7

2) Power Spectrum Reconstruction:

The paper gives recovery procedures for reduced spectrum systems and considers blind recovery when the sparse spectrum’s support is unknown. Blind recovery requires twice the minimal rate of the known-support setting.

2) Power Spectrum Reconstruction:

  • If Proposition 2’s conditions are satisfied—namely, sampling rate f ≥ f(2)—the signal’s power spectrum can be reconstructed by first reducing the systems as shown in (40) and (41).Source: 2) Power Spectrum Reconstruction · p. 7

1) Minimal Rate for Perfect Reconstruction:

  • Blind sparse spectrum recovery requires twice the minimal sampling rate of the known-support setting.With Kf-sparse spectra and Kf < M, the spectrum is the unique sparsest solution to the measurement system.Source: 1) Minimal Rate for Perfect… · p. 7
  • The paper distinguishes its analog sparse-spectrum system from, which recovers average bin power using a finite-vector compressed-sensing problem.The analog system comprises infinitely many equations because frequency is continuous.Source: 1) Minimal Rate for Perfect… · p. 7

2) Power Spectrum Reconstruction:

The analog model’s continuous-frequency recovery is reduced to a finite support-recovery problem, while the digital model uses classical compressed sensing. The simulations use the analog model and MWC front end to examine reconstruction and detection.

2) Power Spectrum Reconstruction:

  • The digital model recovers the sparse vector rc from rz using classical compressed-sensing techniques.Unlike the analog model, it requires solving a single equation.Source: 2) Power Spectrum Reconstruction · p. 7
  • The analog model converts infinitely many frequency-indexed equations into a finite MMV system for support recovery, then reconstructs the spectrum.The CTF block performs this reduction, and the recovered support matches that of the original equations.Source: 2) Power Spectrum Reconstruction · pp. 7–8
  • SBR2 reaches the blind minimal rate f(3), whereas SBR4 requires twice that rate; SBR2 adds complexity and operates only in the frequency domain.SBR2 alternates continuous-to-finite reduction with bisection until the sparsity level is at most Kf.Source: 2) Power Spectrum Reconstruction · p. 8

V. SIMULATION RESULTS

  • The simulations use the analog model with an MWC front end and assess detector ROC against spectrum-reconstruction baselines.They examine SNR, sensing time, averaging, and sampling rate; the number of samples also sets frequency-domain resolution.Source: §V · p. 8, A. Reconstruction in time and… · p. 8

A. Reconstruction in time and frequency domains and detection

  • The analog-model detector sums the reconstructed spectrum over frequency or time, and other statistics can also be used.The paper estimates the sampled autocorrelation using averaging across frames and reconstructs the spectrum from it.Source: A. Reconstruction in time and… · p. 8

B. Spectrum reconstruction

The experiments demonstrate spectrum reconstruction below the Nyquist rate and examine how SNR, averaging, sensing-time allocation, and sampling rate affect detection. Detection performance can plateau as frames or channels increase, while balanced frame and sample counts perform best at fixed total samples.

B. Spectrum reconstruction

C. Practical parameters

  • Below an SNR between 5dB and 0dB in this setting, detection performance decreases rapidly as SNR falls.ROC curves become linear after support recovery failures, when measured band energy is zero under both hypotheses; this transition occurs earlier at lower SNR.Source: C. Practical parameters · p. 10, Figure 8 · p. 10
  • Above a threshold, increasing the number of frames almost does not affect detection performance.This experiment fixes K = 20 samples per frame and sets SNR to 0dB.Source: C. Practical parameters · p. 10, Figure 10 · p. 11
  • At fixed PK = 100 samples, P = 10 frames and K = 10 samples per frame attain the best detection performance.The experiment compares different frame/sample divisions under the same total number of samples per channel.Source: C. Practical parameters · p. 10
  • Above 9 channels, detection performance increases very little, while below the 7-channel minimum it decreases drastically.With noise present, good detection requires sampling above the minimum threshold.Source: C. Practical parameters · p. 11

D. Performance Comparaisons

The paper compares power-spectrum sensing with signal reconstruction and Nyquist-rate sensing, then summarizes its reconstruction framework. Power-spectrum sensing outperforms sub-Nyquist signal-spectrum sensing, while its detection performance is comparable to Nyquist-rate power-spectrum detection down to an SNR threshold.

1) Power Spectrum versus Spectrum Reconstruction:

2) Nyquist versus Sub-Nyquist Sampling:

VI. CONCLUSION

  • Frequency-domain analysis treats the considered analog and digital model variations uniformly, whereas time-domain analysis is more complex.The conclusion covers two sampling schemes for the analog model and two reconstruction schemes for the digital model.Source: §VI · p. 12
  • The paper derives minimal noise-free sampling rates and recovery techniques for sparse and non-sparse signals, including blind and non-blind sparse detection.Its conclusion reports simulations of sub-Nyquist reconstruction and practical effects on detector ROC.Source: §VI · p. 12

Related papers

Loading 1308.5149v1…