Recent Patents on Signal Processing, 2012, 2, 127-139 127
3:99/8346/12 $100.00+.00 © 2012 Bentham Science Publishers
Compressive Sampling Methods and their Implementation Issues
Nikos Petrellis
*
Computer Science and Telecommunications Dept., TEI of Larisa, TEI Campus, Larisa 41110, Greece
Received: 28 April 2012; Revised: 22 June, 2012; Accepted: 22 June 2012
Abstract: The ultra wideband telecommunication systems require nowadays mixed signal front end modules like Analog-
to-Digital Converters with ever increasing cost in terms of speed, power consumption, die area and complexity if the
Nyquist rate is respected. Since the signals in various applications are inherently sparse or compressible, a great attention
has been recently given to systems that use under-sampling methods without significant information loss. These methods
are covered by a signal processing branch known as Compressed (or Compressive) Sensing or Compressive Sampling. A
Compressive Sampling approach is useful if it leads to a lower cost solution compared to the cost of the architecture that
would be required if Nyquist sampling was adopted or if the signal frequency is so high that no appropriate Nyquist
sampler is available at all. A number of recent Compressive Sampling patents will be reviewed in this article focusing on
the feasibility of their implementation since the mathematical modeling that is often adopted is based on discrete input
values and cannot be directly applied to the real world analog signals. Moreover, computational intensive optimization
problems require to be solved in order to fully reconstruct the original input signal from a small number of samples.
Keywords: Compressive sampling (sensing), undersampling, analog digital conversion, VLSI, DSP.
1. INTRODUCTION
The Compressive Sampling (CS) is formally defined as a
method for finding solutions to underdetermined linear
systems i.e., to linear systems with fewer equations than the
unknown variables. In signal processing this method is used
basically for the reconstruction of original signals with fewer
samples than the ones required by the Nyquist rate. The
Nyquist sampling rate requires a sampling frequency that is
at least twice as high as the highest frequency component of
the input signal. In the underdetermined linear system
mentioned above, the vector with the measured values can be
expressed as a product of an appropriate matrix with the
vector containing the Nyquist input samples. The selection
of the matrix that multiplies the input vector is a key point
for the success of every CS approach.
Moving to higher communication throughputs and carrier
frequencies in ultra wideband systems, the requirements
posed to Analog-to-Digital Converters (ADCs) are ever
getting stricter. More specifically, if an ADC has to operate
in a multi-GS/s sampling rate, expensive architectures like
Flash ADCs have to be used. These architectures require
large die area and consume extremely high power even for
resolutions as low as 7-bits.
The CS is a method that can often relax the sampling
frequency requirements of an ADC since it would be forced
to sample the input only at the time points that are necessary
for the input signal reconstruction. Nevertheless, the
sampling intervals are not always regular (non-uniform
sampling) and the original signal reconstruction is difficult
since it cannot be carried out using low cost hardware.
*Address correspondence to this author at the Computer Science and
Telecommunications Dept., TEI of Larisa, TEI Campus, Larisa 41110,
Greece; Tel: +302410684542; Fax: +302410684573;
E-mail: npetrellis@teilar.gr
The input signal has to be sparse (either in the time or the
frequency domain) or compressible in order to apply CS
techniques. A signal is considered to be K-sparse if its N
samples retrieved at Nyquist rate can be represented by a
vector of size N that contains K non-zero elements (K<<N).
If a signal is not sparse but can be described by less than N
values (e.g., as a sequence of vectors) then it is considered as
compressible. For example, a straight line in an image can be
defined by the coordinates of its endpoints instead of storing
all of its pixels. Similarly, a circle can be defined only by the
coordinates of its center and the length of its radius, a
triangle can be defined only by the coordinates of its
vertices, etc. If a signal has a lot of near-zero values, they
can all be considered as zero without significant information
loss in many CS applications. This process is often called
“sparsification”. Such signals may represent images like the
ones produced by Magnetic Resonance Imaging (MRI),
radar imaging, Optical Character Recognition (OCR), tele-
surgery, multi-player online games, speech recognition,
occasional sensor sampling, etc.
The fundamental theory for Compressive Sampling is
described in [1] by Donoho. A more abstract introduction to
CS is given in [2]. In [3-10] implementable CS methods are
proposed for general applications. In [11] CS techniques are
discussed for the channel estimation of telecommunication
systems that use Orthogonal Frequency Division
Multiplexing (OFDM). The application of CS in sensor
networks like GPS is shown in [12-16] while its use in
imaging applications like MRI is shown in patents [17] and
[18]. As already mentioned the architecture of ADCs that
can support compressive sampling is a very important issue
for the feasibility of CS methods. Appropriate ADC
implementations and architectures are described in [19, 20].
In Section 2 the mathematical modeling of the CS
problem is given as well as a number of optimization targets
that have been proposed in the referenced approaches for the