The Sample Allocation Problem and Non-Uniform Compressive Sampling

TL;DR

This paper investigates optimal sample allocation in frequency-domain compressive sampling, proposing non-uniform sampling strategies based on spectral support and magnitude, leading to improved reconstruction especially for multi-band and complex spectral signals.

Contribution

It introduces a non-uniform sampling scheme that allocates samples based on spectral support and magnitude, enhancing CS performance for multi-band and ambiguous spectral signals.

Findings

Sampling within signal bands improves reconstruction accuracy.

Knowledge of spectral peaks allows exact sampling at peak locations.

Horizontal spectral slicing reduces sample requirements and enables sample reuse.

Abstract

This paper discusses sample allocation problem (SAP) in frequency-domain Compressive Sampling (CS) of time-domain signals. An analysis that is relied on two fundamental CS principles; the Uniform Random Sampling (URS) and the Uncertainty Principle (UP), is presented. We show that CS on a single- and multi-band signals performs better if the URS is done only within the band and suppress the out-band parts, compared to ordinary URS that ignore the band limits. It means that sampling should only be done at the signal support, while the non-support should be masked and suppressed in the reconstruction process. We also show that for an N-length discrete time signal with K-number of frequency components (Fourier coefficients), given the knowledge of the spectrum, URS leads to exact sampling on the location of the K-spectral peaks. These results are used to formulate a sampling scheme when the boundaries of the bands are not sharply distinguishable, such as in a triangular- or a stacked-band- spectral signals. When analyzing these cases, CS will face a paradox; in which narrowing the band leads to a more number of required samples, whereas widening it leads to lessen the number. Accordingly; instead of signal analysis by dividing the signal's spectrum vertically into bands of frequencies, slicing horizontally magnitude-wise yields less number of required sample and better reconstruction results. Moreover, it enables sample reuse that reduces the sample number even further. The horizontal slicing and sample reuse methods imply non-uniform random sampling, where larger-magnitude part of the spectrum should be allocated more sample than the lower ones.