IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 57, NO. 3, MARCH 2009 993
Blind Multiband Signal Reconstruction: Compressed
Sensing for Analog Signals
Moshe Mishali, Student Member, IEEE, and Yonina C. Eldar, Senior Member, IEEE
Abstract—We address the problem of reconstructing a multi-
band signal from its sub-Nyquist pointwise samples, when the
band locations are unknown. Our approach assumes an existing
multi-coset sampling. To date, recovery methods for this sampling
strategy ensure perfect reconstruction either when the band
locations are known, or under strict restrictions on the possible
spectral supports. In this paper, only the number of bands and
their widths are assumed without any other limitations on the sup-
port. We describe how to choose the parameters of the multi-coset
sampling so that a unique multiband signal matches the given
samples. To recover the signal, the continuous reconstruction
is replaced by a single finite-dimensional problem without the
need for discretization. The resulting problem is studied within
the framework of compressed sensing, and thus can be solved
efficiently using known tractable algorithms from this emerging
area. We also develop a theoretical lower bound on the average
sampling rate required for blind signal reconstruction, which is
twice the minimal rate of known-spectrum recovery. Our method
ensures perfect reconstruction for a wide class of signals sampled
at the minimal rate, and provides a first systematic study of com-
pressed sensing in a truly analog setting. Numerical experiments
are presented demonstrating blind sampling and reconstruction
with minimal sampling rate.
Index Terms—Landau–Nyquist rate, multiband, multiple mea-
surement vectors (MMV), nonuniform periodic sampling, sparsity.
I. INTRODUCTION
T
HE well-known Whittaker, Koteln
´
ikov, and Shannon
(WKS) theorem links analog signals with discrete repre-
sentations, allowing signal processing in a digital framework.
The theorem states that a real-valued signal bandlimited to
Hertz can be perfectly reconstructed from its uniform samples if
the sampling rate is at least
samples per second. This state-
ment highlights the basic ingredients of a sampling theorem.
First, a signal model which appears in practical applications
and relates to physical properties. In this case, the maximal fre-
quency
parameterizes a bandlimited signal model. Second,
a minimal rate requirement to allow perfect reconstruction,
referred to as the Nyquist rate in the WKS setting. Finally, a
design of sampling and reconstruction stages which is based on
the model parameters.
Manuscript received September 11, 2007; revised November 24, 2008. First
published January 13, 2009; current version published February 13, 2009.
The associate editor coordinating the review of this manuscript was Dr. Zoran
Cvetkovic.
The authors are with the Technion—Israel Institute of Technology, Haifa
32000, Israel (e-mail: moshiko@tx.technion.ac.il; yonina@ee.technion.ac.il).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TSP.2009.2012791
In this paper, we consider the class of multiband signals,
whose frequency support resides within several continuous
intervals, spread over a wide spectrum. This model naturally
arises when observing the sum of a few narrowband trans-
missions which are modulated by relatively high carriers,
with different frequencies. When the band locations and their
widths are known prior to sampling, the signal model defines a
subspace of possible inputs. Landau [1] developed a minimal
rate requirement for an arbitrary sampling method that allows
perfect reconstruction in this setting. The Landau rate equals
the sum of the band widths, which is below the corresponding
Nyquist rate. Several works proposed sampling and reconstruc-
tion stages, designed according to the specific spectral support
of the input signal. Low-rate uniform sampling was studied in
[2] for a real bandpass signal, whereas [3] suggested periodic
nonuniform sampling at the same average sampling rate. Lin
and Vaidyanathan [4] extended this approach to multiband sig-
nals. These methods allow exact recovery at rates approaching
that derived by Landau.
A much more challenging problem is to design a
blind sam-
pling and reconstruction system, that does not rely on knowl-
edge of the band locations. Blindness is important whenever
detecting the spectral support, prior to sampling, is impossible
or too expensive to implement. Reconstruction under partial
knowledge of the support was addressed in a series of confer-
ence papers [5]–[7]. These works do not assume the exact sup-
port but require that the band locations obey a certain math-
ematical condition. Sampling is carried out by a multi-coset
strategy, independent of the band locations. In order to recover
the signal, [5], [6] identified similarities with direction of ar-
rival problems, implying potential use of techniques from this
field, such as MUSIC [8]. An alternative method was proposed
in [7] under the assumption that the support obeys a strict math-
ematical condition. Similar results were derived in [9]. This ap-
proach, however, suffers from two main drawbacks. First, as de-
tailed in Section V-B, if the support violates the condition of [7],
then the samples match many possible input signals, thus hin-
dering any chance for recovery. Since this condition depends on
the band locations which are unknown, it cannot be validated in
practice. The difficulty arises form the fact that this requirement
is not directly related to simple physical properties of a multi-
band model (i.e., the number of bands and their widths). Second,
even when the condition is met, exact recovery is not guaran-
teed. This is a result of the fact that the proposed sampling rate
approaches that dictated by Landau. As we prove in Section III,
perfect blind reconstruction requires a higher sampling rate.
In subsequent publications, Herley and Wong [10] suggested
a half-blind system. Similar ideas were later suggested in [11].
1053-587X/$25.00 © 2009 IEEE