Deborah Cohen and Yonina C.

Eldar

Sub-Nyquist Radar Systems

©istockphoto.com/MrJub

Temporal, spectral and spatial compression

R
adar is an acronym for “radio detection and ranging.” gets. The goal of this article is to review these novel sub-Nyquist
However, the functions of today’s radar systems, both in radars and their potential applications.
civilian and military applications, go beyond simple target Conventional radar systems transmit electromagnetic waves
detection and localization; they extend to tracking, imag- of near-constant power in very short pulses toward the targets
ing, classification, and more and involve different types of radar of interest. Between outgoing pulses, the radar measures the
systems, such as through-the-wall [1], ground-penetration [2], signal reflected from the targets to determine their presence,
automotive [3], and weather [4]. Although radar technology has range, velocity, and other characteristics. Different systems use
been well established for decades, a new line of compressed different radar waveforms and varying transmit strategies. One
radars has recently emerged. These aim at reducing the com- of the most popular methods is pulse-Doppler radar, which
plexity of classic radar systems by exploiting inherent prior periodically transmits identical pulses. In contrast, stepped-
information on the structure of the received signal from the tar- frequency radars (SFRs) [5] vary the carrier frequency of each
pulse. Some systems rely on simple traditional waveforms
such as Gaussian pulses while others adopt more complex sig-
Digital Object Identifier 10.1109/MSP.2018.2868137
Date of publication: 13 November 2018 nals, such as chirps [6], [7]. Each configuration corresponds

1053-5888/18©2018IEEE IEEE Signal Processing Magazine | November 2018 | 35

 to a certain choice in the complexity-performance tradeoff, circular SAR are presented in [33] for both two-dimensional
between complex waveform and system designs and target (2-D) and three-dimensional (3-D) images. In [34], diverse SAR
detection and estimation. applications, such as wide-angle SAR imaging, joint imaging,
State-of-the-art radar systems operate with large bandwidths, and autofocusing from data with phase errors, moving targets,
large coherent processing intervals (CPIs), and high number of analysis, and design of SAR sensing missions, are reviewed.
antennas in multiple-input, multiple-output A survey of statistical sparsity-based tech-
(MIMO) settings [8], [9], to achieve high- Although the natural niques for radar imagery applications is
range velocity and azimuth resolution, respec- presented in [35], including superresolution
tively. This, in turn, generates large data sets
application of CS is imaging, enhanced-target imaging, auto-
to be sampled, stored, and processed, creating typically the reduction focusing, and moving-target imaging. The
a bottleneck in terms of both analog system of the required number review of [22] presents three applications of
complexity, including high-rate analog-to-dig- of samples to perform a CS radars: pulse compression, radar imag-
ital converters (ADCs), and subsequent digital certain signal processing ing, and airspace surveillance with array
processing [10]. task, it was first used antennas. At the time it was written, there
In the past few years, novel approaches was a small number of publications address-
to radar signal processing have emerged that
by the radar community ing the application of CS to radar, as stated
allow radar signal detection and parameter to increase a target’s by the authors.
estimation using a much smaller number of parameter resolution. In this article, we focus on nonradar-
measurements than that required by spatial imaging applications and survey many
and temporal Nyquist sampling. While temporal sampling re­­­­cent works that exploit CS in different radar systems to
refers to taking samples in time intervals determined by the achieve various goals. We consider different transmit wave-
sampling rate, spatial sampling extends this notion to placing forms and processing approaches, while focusing on pulse-
transmit and receive antennas, whose locations are governed Doppler radar—one of the most popular systems—and its
by the signal wavelength. These works capitalize on the fact extension to MIMO configurations. Our goal is to review
that, in most radar applications, the reflectivity scene consists of the main impacts of compressed radar on parameter reso-
a small number of strong targets. That is, the reflected signals lution as well as digital and analog complexity. The survey
by only a few targets have high enough power to be detected includes fast time compression schemes, which reduce the
by the radar receiver. In pulse-Doppler radar, the target scene number of acquired samples per pulse; slow time compres-
is often sparse in the joint time–frequency, or ambiguity, do­­ sion techniques, which decrease the number of pulses; and
main [5]. In synthetic aperture radar (SAR) [11], the scene is spatial compression approaches, in which the number of
often sparse in the Fourier or wavelet domain, or even in the transmit and receive antenna elements is reduced. We show
image domain. that, beyond a substantial rate reduction, compression may
Over the past decade, many works have exploited the inher- also enable communication and radar spectrum sharing [36]–
ent sparsity of the target scene to enhance radar-estimation [38], as elaborated on in [39]. Throughout this article, we
capabilities. These rely on the compressed sensing (CS) [10], consider both theoretical and practical aspects of compressed
[12] framework, brought to the forefront by the works of Candes, radar and present hardware prototype implementations [40]–
Romberg, and Tao [13] and of Donoho [14]. Although the natu- [43] of the theoretical concepts, demonstrating the real-time
ral application of CS is typically the reduction of the required target parameters’ recovery from low-rate samples in pulse-
number of samples to perform a certain signal processing task, Doppler and MIMO radars.
it was first used by the radar community to increase a target’s
parameter resolution [15]–[20]. It was later applied to reduce Radar systems
the number of samples to be processed [21]–[25] and finally Radar systems aim to estimate targets’ parameters to determine
to reduce the sampling rate [26], [27] and number of anten- their location and motion. In its simplest form, the radar trans-
nas [28] required in radar systems, performing time and spatial mits a single pulse toward targets in one direction and recovers
compression and alleviating the burden on both the analog and their range, i.e., distance to the radar, which is proportional to
digital sides. In particular, the recently proposed Xampling, the received pulse delay. More elaborate systems are able to
i.e., compressed sampling concept [10], [29], has been applied provide additional information on the targets. Pulse-Doppler
to radar [30]–[32] to break the link between bandwidth, CPI, radars transmit several pulses, enabling them to resolve both
and the number of antennas on the one hand, and range, Dop- the targets’ ranges and radial velocities, which are proportional
pler, and azimuth resolution, respectively, on the other hand. to the Doppler frequency. Stepped-frequency-based approaches
The reviews of compressive radar [22], [33]–[35] mostly achieve highly effective bandwidths that increase range resolu-
deal with radar imaging. The works in [33] and [34] focus on tion, while allowing for narrow instantaneous bandwidth.
SAR imaging and consider sparsity-based radar imagery using MIMO radars use several elements both at the transmitter and
both greedy algorithms, which iteratively recover the sparse tar- at the receiver to illuminate the entire target scene and recover
get scene, and convex relaxations of sparsity-inducing regular- targets’ azimuths in addition to their ranges and velocities.
ization. The special cases of interferometric, polarimetric, and In this article, we consider the application of compression in

36 IEEE Signal Processing Magazine | November 2018 |

terms of the number of required samples,
pulses, and antennas, as well as its im­­
αl αl e–jvlτ αl e–j 2vlτ
pact on different aspects of the radar
system, including parameter resolution Tx
and system complexity, for several types Rx
of radars.

Pulse-Doppler radar τl t
A standard pulse-Doppler radar trans- τ
ceiver detects targets by transmitting a
periodic stream of pulses and processing
FIGURE 1. The pulse-Doppler radar transmitted and received pulse trains with P = 3 pulses and
its reflections. The transmitted signal L = 4 targets Tx: transmitted; Rx: received.
x T (t) consists of P equally spaced puls-
es h (t) such that

P-1 Based on the three assumptions A1–A3 presented in

x T (t) = / h (t - px), 0 # t # Px. (1)
“Targets’ Assumptions,” the received signal can be writ-
p=0
ten as
The pulse-to-pulse delay x is the pulse-repetition interval
P-1 L-1
(PRI), and its reciprocal 1/ x is the pulse-repetition frequency x R (t) = / / a l h (t - x l - px) e -j o l px
, 0 # t # Px. (2)
(PRF). The entire span of the signal in (1), i.e., Px, is the CPI. p=0 l=0
The pulse time support is denoted by T p, with 0 1 T p 1 x.
The pulse h (t) is typically a known time-limited baseband It will be convenient to express x R (t) as a sum of single frames
function with continuous-time Fourier transform (CTFT)
H ( f ) = 8 -33 h (t) e -j2rft dt that has negligible energy at fre- P-1

quencies beyond B h /2, where B h is referred to as the band-

x R (t) = / x p (t), (3)
p=0
width of h (t) . An example of a transmitted pulse train is
illustrated in Figure 1. where
It is typically assumed that the target scene is composed
of L nonfluctuating point-targets, according to the Swerling-0 L-1

model [5]. This is one of the popular models in the radar signal
x p (t) = / a l h (t - x l - px) e -j o l px
, 0 # t # Px. (4)
l=0
processing literature since, by describing an idealized target,
it allows simplifying the radar equations while constituting a An illustration of a received pulse train is shown in Figure 1
fairly good approximation in many applications [6], [7]. Other with L = 4 targets. In pulse-Doppler radar, the goal is to recover
models, such as Swerling-1, which applies to targets composed the three L parameters {x l, o l, a l} for 0 # l # L - 1 from the
of many independent scatters, or fluctuating target models, are received signal x R (t) . In particular, estimating the time delays x l
beyond the scope of this article. The pulses reflect off the L and Doppler frequencies o l enables an approximation of the targets’
targets and propagate back to the transceiver. Each target l is distances and radial velocities.
defined by three parameters:
1) a time delay x l = 2rl /c, proportional to the target’s dis- Stepped-radar waveforms
tance to the radar or range rl, where c is the speed of light In classic pulse-Doppler radar, high-range resolution re­­
2) a Doppler-radial frequency o l = 2ro l fc /c, proportional to quires a large signal bandwidth. This technology bottleneck
the target’s radial velocity to the radar, i.e., the target’s is partially overcome by stepped-frequency-based wave-
velocity radial component ro l, and the radar’s carrier fre- forms, in which the large bandwidth is obtained se­­quentially
quency fc by stepping the frequency of each pulse, keeping the instan-
3) a complex amplitude a l, proportional to the target’s radar taneous bandwidth low. Two popular examples of such
cross section (RCS), dispersion attenuation, and other waveforms are SFRs and stepped chirps. An SFR [5] system
propagation factors. transmits P-narrowband pulses, in which each pulse p has
The targets are defined in the radar radial coordinate sys- carrier frequency
tem and are typically assumed to lie in the radar unambigu-
ous time–frequency region: delays up to the PRI and Doppler f p = f0 + pD f , (5)
frequencies up to the PRF. When this assumption does not
hold, several processing techniques have been proposed that for 0 # p # P - 1, with f0 the initial frequency and D f the
require the transmission of multiple pulse trains with differ- frequency increment. The pth-transmitted pulse is a rectangu-
ent parameters, e.g., different PRFs. We review this setting in lar pulse modulated by its carrier f p . The corresponding
the “Range-Velocity Ambiguity Resolution” section. received signal is then of the form

IEEE Signal Processing Magazine | November 2018 | 37

 Targets’ Assumptions
To simplify the received signal model, the following constant o l . This condition is satisfied when the velocity
assumptions of the targets’ locations and motions are typi- change induced by acceleration is smaller than the
cally made [5]: velocity resolution:
• A1: Far targets: The target-radar distance is large com-
pared with the distance change during the coherent c & rp % c
rpl Px % l .(S4)
processing intervals (CPIs), which allows for constant 2fc Px 2fc (Px) 2
a l within the CPI:
Although these assumptions may seem hard to comply
rol Px % rl & o l % fc x l .(S1) with, they all rely on slow enough relative motion between
Px
the radar and its targets. Radar systems tracking people,
• A2: Slow targets: The constant-Doppler phase during
ground vehicles, and sea vessels usually comply quite
pulse time,
easily [6].
In multiple-input, multiple-output settings, two additional
o l Tp % 1, (S2)
assumptions are adopted on the array structure and trans-
mitted waveforms:
and low target velocity allows for constant x l during the
• A4: Collocated array: The target radar cross sections
CPI. This condition holds when the baseband Doppler
a l and i l are constant over the array [44].
frequency is smaller than the frequency resolution:
• A5: Narrowband waveform: A small aperture allows
2rol B h % 1 & o % fc .(S3) x l to be constant over the channels:
l
c Px PxB h

• A3: Small acceleration: The target velocity remains 2Zm % 1 .(S5)

approximately constant during the CPI allowing for c Bh

L-1 L- 1
rect ` t - x l j e j2r (f p (t - x l) + 2 (t - x l) ), (9)
c
x p (t) = / a l rect (t - x l) e -j2 f (t -
r p x l)
e jol px . (6) x p (t) = / al e j zp
x
2

l=0 l= 0

To process the received signal, the delay is neglected in the is dechirped with a reference linear-frequency waveform of
signal envelope because of the narrowband assumption. An fixed frequency equal to the first carrier f0:
SFR traditionally obtains one sample from each received pulse
m (t) = rect ` t - x r j e -j2r` f0 t + 2 t j . (10)
c 2
and computes the phase detector output sequence as
x
L-1 The receive window is x r = 2 (rmax + rmin) /c, and the refer-
yp = / a l e j2 f r p xl
e jo l px . (7)
ence delay is t r = (rmax + rmin) /c , with rmax and rmin as the
l=0
maximal and minimal ranges, respectively. The resulting
The phase detector signal y p can be modeled as the product of dechirped received signal can be written as
the received signal (6) and the reference signal, followed by a
low-pass filter (LPF). Conventional processing applies an L-1
x p (t) = / a l e j^ z p - 2rf p x l h
rect ` t - x l + t r j e j2r^ f p - f0 - cx lht . (11)
inverse discrete Fourier transform (DFT) on the output to esti- l=0
x
mate the targets’ time delays x l and Doppler frequencies o l .
The range resolution achieved by SFR is c/2PD f , where PD f Classic processing of the received signal includes a DFT oper-
is the total effective bandwidth of the signal over P pulses. ation to recover the targets’ delays x l .
Another popular stepped waveform is the stepped chirp or
multifrequency chirp signal. The corresponding transmitted MIMO pulse-Doppler radar
signal is given by MIMO radar presents significant potential for advancing state-
of-the-art modern radar in terms of flexibility and perfor-
P-1 mance. This configuration [8] combines several antenna elements
rect ` t j e j2r` f p t + 2 t j, (8)
c 2
x T (t) = / ej zp
x both at the transmitter and receiver. Unlike phased-array sys-
p=0
tems, each transmitter radiates a different waveform, which
where c is the common chirp rate and f p and z p are the fre- offers more degrees of freedom (DoF) [9]. There are two main
quency and complex phase of the pth subcarrier. The returned configurations of MIMO radar, depending on the location of
signal corresponding to the pth pulse, given by the transmitting and receiving elements; collocated MIMO

38 IEEE Signal Processing Magazine | November 2018 |

[46], in which the elements are close to each other relative to Current challenges
the signal wavelength, and multistatic MIMO [47], where they Standard radar processing samples and processes the received
are widely separated. In this article, we focus on collocated signal at its Nyquist rate B h . For example, the pulse-Doppler
pulse-Doppler MIMO systems. classic radar processing, described in “Classic Pulse-Doppler
Collocated MIMO radar systems exploit waveform diver- and Multiple-Input, Multiple-Output Processing,” first filters
sity, based on mutual orthogonality of the transmitted signals the sampled signal by a matched filter (MF). In modern sys-
[9]. Consequently, the performance of MIMO systems can be tems, the MF operation is performed digitally and therefore
characterized by a virtual array constructed by the convolution requires an ADC capable of sampling at rate B h . Other radar
of the locations of the transmit and receive antenna locations. systems similarly require sampling the received signal at its
In principle, with the same number of antenna elements, this Nyquist rate. The radar bandwidth B h is inversely proportion-
virtual array may be much larger than the array of an equiva- al to the system fast time, or range resolution, and can thus be
lent traditional system [48]. hundreds of megahertz or even up to several gigahertz, requir-
The standard approach to collocated MIMO adopts a ing a high sampling rate and resulting in a large number of
virtual uniform linear array (ULA) structure [49], where R samples per pulse N = xB h to process.
receivers, spaced by m / 2 and T transmitters and spaced by The slow time (Doppler) resolution is inversely proportion-
R ^m / 2 h (or vice versa), form two ULAs. Here, m is the signal al to the CPI Px. The Doppler processing stage can be viewed
wavelength. Coherent processing of the resulting TR chan- as an MF in the pulse dimension, i.e., slow time domain, to a
nels generates a virtual array equivalent to a phased array constant radial velocity target. As such, it increases the signal-
with TR ^m / 2 h - spaced receivers and normalized aperture to-noise ratio (SNR) by P compared to the SNR of a single
Z = TR / 2. Denote by {p m} Tm-=10 and {g q} qR =- 01, the normal- pulse [7]. Since an MF is the linear time-invariant system that
ized transmitters’ and receivers’ locations, respectively. For maximizes SNR, it follows that a factor P increase is optimal
the traditional virtual ULA structure, denote g q = q / 2 and for P pulses. A large number of pulses increases resolution and
p m = Rm / 2. This standard-array structure and the corre- SNR but leads to large time on target and a large total number
sponding virtual array are illustrated in Figure 2 for R = 3 of samples to process, given by PN.
and T = 5. The circles represent the receivers, and the squares The required computational power corresponds to P con-
are the transmitters. volutions of a signal of length N = xB h and N-fast Fourier
Each transmit antenna sends P pulses, such that the mth- transforms (FFTs) of length P (see “Classic Pulse-Doppler and
transmitted signal is given by Multiple-Input, Multiple-Output Processing”). The growing
demands for improved estimation accuracy and target separa-
P-1
tion dictate an ever-growing increase in the signal’s bandwidth
s m (t) = / h m ^t - pxhe j2 f t,
r c
0 # t # Px, (12)
and CPI. This creates bottlenecks in sampling and processing
p=0
rates in the fast time (intrapulse) domain and in time on target
where h m (t), 0 # m # T - 1 are orthogonal pulses with band- in the slow time (interpulse) dimension.
width B h and modulated with carrier frequency fc . For con- In MIMO radar, the additional spatial dimension increas-
venience, it is typically assumed that fc x is an integer, so that es the system’s complexity, as may be seen in “Classic Pulse-
the initial phase for every pulse e -j2rfc xp is canceled in the Doppler and Multiple-Input, Multiple-Output Processing.”
modulation for 0 # p # P - 1 [6]. In such systems, the array aperture determines the azimuth
MIMO radar architectures impose several requirements on resolution. In a traditional virtual array configuration, the
the transmitted waveform family. Besides traditional demands product between the number of transmit and receive antennas
from radar waveforms such as low sidelobes, MIMO transmit scales linearly with the aperture. Consequently, high resolution
antennas rely on orthogonal waveforms. In addition, to avoid
cross talk between the T signals and form TR channels, the
orthogonality condition should be invariant to time shifts, that
- 3 s i ^ t h s j ^ t - x 0 h dt = d ^ i - j h for i, j ! 60, T - 1@ and
* λZ
is 8 3
for all x 0 . The main waveform families typically considered
are time-division multiple access (TDMA), frequency-division λ λ
R
2 2
multiple access (FDMA), and code-division mul­ tiple access
(a)
(CDMA), respectively. Time-invariant orthogonality is achieved
by FDMA and TDMA and approximately achieved by CDMA, λZ
as the latter involves overlapping frequency bands [50].
Besides the traditional assumptions on the targets, MIMO λ
systems present additional requirements on the radar array and 2
waveforms with respect to the targets, as described in “Tar- (b)
gets’ Assumptions.” In the MIMO configuration, the goal is
to recover the targets’ azimuth angles i l in addition to their FIGURE 2. An illustration of MIMO arrays: (a) a standard array and (b) a
delays x l and Doppler shifts o l from the received signals. corresponding receiver virtual array [32].

IEEE Signal Processing Magazine | November 2018 | 39

 Classic Pulse-Doppler and Multiple-Input, Multiple-Output Processing
The classic methods for radar processing typically consist peak is considered to originate from a target so that a
of the following stages [5], [45]: required probability of FA is achieved.
1) Sampling: Sample each incoming frame x p (t ) at its Ny- Classic collocated multiple-input, multiple-output radar pro-
quist rate B h, equal to the double-sided bandwidth cessing traditionally includes the following stages:
of h (t ), creating the samples x p [n], 0 # n # N - 1, 1) Sampling: At each receiver 0 # q # R - 1, where R
where N = xB h . We assume, for simplicity, that N is denotes the number of receivers, the signal x q (t ) is sam-
an integer. pled at its Nyquist rate Btot. In code-division multiple
2) Matched filter (MF): Apply a standard MF on each frame access and time-division multiple access, B tot = B h as all
x p [n] . This results in the outputs y p [n] = x p [n] ) h [- n], waveforms overlap in frequency, whereas in frequency-
where h [n] is the sampled version of the transmit- division multiple access, B tot = TB h, where B h denotes
ted pulse h (t ) at its Nyquist rate and ) is the convolu- the bandwidth of a single waveform in both cases, and
tion operation. The time resolution attained in this step T is the number of transmitters.
is 1/B h . 2) MF: The sampled signal is convolved with a sampled
3) Doppler processing: For each discrete time n, perform version of h m (t ), for 0 # m # T - 1. The time resolution
a P-point discrete Fourier transform along the pulse dimen- attained in this step is 1/B tot .
sion, i.e., z n 6k@ = DFTP " y p [n] , = R Pp -= 10 y p [n] e - j 2rpk/P for 3) Beamforming: The correlations between the observa-
0 # k # P. The Doppler resolution is 1/Px. tion vectors from the previous step and the steering vec-
4) Delay-Doppler map: Stacking the vectors z n and tak- tors corresponding to each azimuth on the grid defined
ing absolute value, we obtain a delay-Doppler map by the array aperture are computed. The spatial resolu-
Z = abs 6z 0, f, z N - 1@ ! R P # N . tion attained in this step is 2 TR .
5) Peak detection: A heuristic detection process, in which 4) Doppler detection: The correlations between the result-
knowledge of the number of targets, targets’ powers, ing vectors and Doppler vectors, with Doppler frequen-
clutter location, and so on, may help in ­discovering cies lying on the grid defined by the number of pulses,
targets’ positions. For example, if we know there are L are computed. The Doppler resolution is 1/Px.
targets, then we can choose the L -strongest points in 5) Peak detection: This is similar to classic radar, but de-
the map. Alternatively, constant false alarm (FA) rate de- tection is performed on the three-dimensional range-
tectors determine the power threshold, above which a azimuth-Doppler map.

requires a large number of antennas, thus increasing the sys- While CS is typically applied to signal processing tasks to
tem’s complexity in terms of hardware and processing. reduce the associated sampling rate [10], earlier papers that
In the following sections, we review fast time-compressed applied CS recovery to pulse-Doppler radar and SFR were
radar systems that allow for low-rate sampling and processing aimed at increasing delay-Doppler resolution [15]–[17], [20]
of radar signals, regardless of their bandwidth, while retaining using Nyquist samples. More recent approaches use CS recov-
the same SNR scaling. We then demonstrate how compres- ery techniques on low-rate, or sub-Nyquist samples, which
sion can be extended to the slow time, thereby reducing time enable sampling and processing rate reduction while achiev-
on target, and to the spatial dimension allowing one to achieve ing the same resolution as traditional Nyquist radars. Later in
resolution similar to a filled array but with significantly fewer this section, we review radar recovery methods that increase
elements. In reality, the received signal x R (t) is further contam- delay-Doppler resolution using CS techniques on Nyquist
inated by additive noise and clutter. We will also demonstrate samples. In the next sections, we consider the application of
the impact of SNR and clutter on compressed radar system pro- CS to reduce the fast time-sampling rate and the number of
totypes [30], [40], [54]. Finally, we show how compression and pulses and antennas, while preserving the resolution achieved
sub-Nyquist sampling may be used to address other challenges, by Nyquist systems.
such as communication and radar spectrum sharing. In the works of [15]–[17] and [20], the signal is still sampled
at its Nyquist rate B h, but the delay and Doppler resolutions are
Increased parameter resolution determined by the CS grid containing N > xB h grid points,
In many radar applications, the reflectivity scene consists of a rather than the signal’s bandwidth and CPI, respectively. The
small number L of strong targets. Therefore, CS techniques key idea in [15], which adopts a pulse-Doppler radar model,
(see “Compressed Sensing Recovery”) are a natural process- is that the received signal x R (t) defined in (2) is generally a
ing tool for radar systems. Shortly after the idea of CS was sparse superposition of time- and frequency-shifted replicas of
brought forward by the works of Candes, Romberg, and Tao the transmitted waveforms. The time–frequency plane is dis-
[13] and of Donoho [14] a decade ago, it was introduced to cretized into an N # N grid in which each point represents a
pulse-Doppler radar [15], [16 ], [55] and SFR [17]. unique time–frequency shift H i, expressed as the product of

40 IEEE Signal Processing Magazine | November 2018 |

 Compressed Sensing Recovery
Compressed sensing (CS) [10], [12] is a framework for subtracting the contribution of a partial estimate x
t , of the
simultaneous sensing and compression of finite-dimen- signal at the ,th iteration, from z, as follows:
sional vectors, which relies on linear dimensionality
reduction. In particular, the field of CS focuses on the r = z - Ax
t , .(S9)
recovery problem
It is initialized by r = z. Once the support set is updated
z = Ax, (S6) by adding the index i, the coefficients of x t , over the sup-
port set are updated, so as to minimize the residual error.
where x is an N # 1 sparse vector, i.e., with few nonzero Other greedy techniques include thresholding algo-
entries, and z is a vector of measurements of size rithms. We focus here on the IHT method proposed in
M 1 N. CS provides recovery conditions and algorithms [53]. Starting from an initial estimate x t 0 = 0, the algo-
to reconstruct x from the low-dimensional vector z. rithm iterates a gradient-descent step with step size n fol-
Two popular CS greedy recovery algorithms, orthogonal lowed by hard thresholding, i.e.,
matching pursuit (OMP) and iterative hard thresholding
(IHT), attempt to solve the optimization problem x t , - 1 + nA H (z - Ax
t , = T (x t , - 1), k), (S10)

x
t = argmin x
x 0 s.t. z = Ax, (S7) until a convergence criterion is met. Here, T (x, k) denotes
a thresholding operator on x that sets all but the k entries
where · 0 denotes the , 0 - norm. OMP [51], [52] iterative- of x with the largest magnitudes to zero, and k is the
ly proceeds by finding the column of A most correlated to sparsity level of x (assumed to be known).
the signal residual r, Alternative approaches to greedy recovery are convex-
relaxation-based methods using , 1 regularization such as
i = argmax | A H r | , (S8) basis pursuit and least absolute shrinkage and selection
operator, better known as LASSO. Further details on CS
where the absolute value is computed element-wise and recovery conditions and techniques can be found in [10]
(·) H is the Hermitian operator. The residual is obtained by and [12].

time-shift and frequency-modulation matrices, denoted by T (.) where the vector f contains the Nyquist rate samples h [n]
and M (.), respectively. In particular, of the transmitted signal h (t) . The latter is chosen so that
the samples correspond to the Alltop sequence h [n] =
H i = M i mod N T6i/N@, (13)
3
` 1 N j e 2rjn /N [56], for some prime N $ 5. This yields a
low-coherence matrix U, i.e., a matrix whose columns have
where small correlation.
The vector s is reconstructed from y using CS techni­
J0 0 J1 N
1N K 0 O ques, as described in “Compressed Sensing Recovery.” The
K O 2r
K1 0 0O K e N
j O time–frequency shifts, determined by the targets’ delays and
T=K O , M=K O. (14)
j j K j O Doppler frequencies, are thus recovered with a resolution
KK OO K0
0 1 0 j 2 r (N - 1 ) O of 1 / N.
L P L e N
P The CS recovery in [15] is performed without an MF, which
Here, 6 · @ and mod denote the floor and modulation functions, reduces performance in low-SNR regimes. Additionally, [15]
respectively. considers only delay recovery. Alternatively, CS techniques
The vector y that concatenates the Nyquist samples of a can be performed after applying an analog MF [16] on the
single pulse x p (t) can then be expressed as pulse-Doppler-received signal (2). The MF output of the pth
pulse, sampled at the Nyquist rate 1/B h, is given by
y = Us, (15)
L-1

where s is the L -sparse vector of size N 2 whose nonzero

w p [k] = / al e j ol xl
e jo l px C h [k - x l /x], (17)
l=0
entries are the targets’ RCS a l with locations determined by
the corresponding time–frequency shift. The ith column, i.e., where C h [k] is the discrete autocorrelation function of the
atom of the N # N 2 matrix U, is given by transmitted waveform. For each sampling time k, the Nyquist
samples have a sparse representation in the frequency
U i = H i f, (16) (Doppler) domain using a Fourier matrix as a dictionary. A

IEEE Signal Processing Magazine | November 2018 | 41

 two-step approach is therefore proposed to apply CS recovery product operator. As in [15], the dictionary U contains N 2
for each k. However, the sidelobes of C h [k] lead to ambigui- atoms. The Nyquist samples can then be expressed in the
ty. To avoid this, pairs of Golay complementary sequences x 1 form (15), and the compressed samples z are given by
and x 2 of length N, whose correlation functions satisfy
z = Ay, (21)
C x 1 [k] + C x 2 [k] = 2Nd [k], (18)
where A is an M # N matrix, with M 1 N constructed by
are transmitted alternatively by phased coding of the baseband randomly selecting M rows of the N # N identity matrix,
waveform. This allows for unambiguous delay-Doppler recov- which corresponds to the M-selected pulses.
ery, provided that all of the Doppler coordinates are within the In these approaches, processing is performed at a low rate;
interval 6- r/2, r/2@ . however, the random discarding of samples is difficult to imple-
CS has also been applied to SFR to increase the range reso- ment in a sampling system for the purpose of effectively reduc-
lution [17]. As in pulse-Doppler radar, the target scene is dis- ing the sampling rate. Furthermore, the large dictionary size
cretized over an N # N delay-Doppler map [17]. The outputs discussed in the previous section remains an issue. Alternative
of the phase detector (7) are then expressed, as in (15), where practical radar systems using CS to reduce the sampling rate have
y is the vector of size P with the pth entry given by y p, and U been proposed and rely on two main techniques: uniform low-rate
is a DFT-based dictionary such that sampling using appropriate waveforms and analog preprocessing.

U ( p,(i - 1) N + k) = e j2r f p x i e jo k px . (19) Uniform low-rate sampling

In [26], the authors consider SFR using multifrequency chirps, as
The vector s is then recovered from y using CS techniques. described in (8). Low-rate samples are uniformly taken from the
The approaches mentioned in this section may increase reso- received signal (11) at rate 2cx r, with x r = 2 (rmax - rmin) / c,
lution by taking a large grid size N. However, bounds on N are with c being the common chirp rate. This results in the aliasing
not discussed, and it is not clear how large it can be. Denser grids of the multiple sinusoids to baseband with random complex coef-
reduce the sensitivity of the reconstruction to off-grid targets but ficients. Upon discretization of the target range, as denoted by s,
increase the computational complexity by a square factor since the low-rate samples may be modeled as
the dictionaries contain N 2 atoms. More importantly, higher
grid dimensions cause a significant increase to the coherence of y = As. (22)
the CS dictionary, which may degrade recovery performance.
The parameter space discretization, typically used in CS Here, the kth column of the sensing matrix A is the FFT of the
recovery techniques, assumes the targets’ delays and Dop- samples of (11) for a singular target at range bin k correspond-
plers lie on the predefined grid. Several approaches have been ing to a delay of x l = 2 (rmin - kD) / c, where D is the range-
proposed to solve off-grid issues, including grid refinement, discretization step. The targets’ delays are therefore recovered
which adjusts the detected delay-Doppler peak [32], parameter, from the low-rate uniform sampling of the chirp waveforms.
perturbation-based, adaptive-sparse reconstruction techniques
[21], and sensing matrix perturbation [57]. More references Random demodulation
may be found in [58]. Many analog-to-information-conversion systems have been
proposed to sample wideband signals at sub-Nyquist rates.
Fast time compression Among them, the random demodulator (RD) [59], random-
In the works reviewed thus far, sampling and digital process- modulation preintegrator (RMPI) [60], and Xampling-based
ing are still performed at the Nyquist rate. We next consider [29] systems have been used for radar applications. All three
compressed radar that reduces sampling and processing rates. approaches consider pulse-Doppler radar.
The RD modulates the input signal using a high-rate se­­quence
Random sampling p (t) created by a pseudorandom number generator, aliasing its
Random sampling has been considered in SFR systems by frequency content. The random sequence used for demodulation
selecting random measurements out of the Nyquist samples is a square wave, which alternates between the levels ! 1 with
[21], [22]. The SFR approach of (1) is adopted in [22], with a equal probability. The mixed output is filtered by a bandpass fil-
random selection of M out of P pulses with different carriers. ter h bp (t), with center frequency fc and bandwidth B CS % B h,
The sparse representation of the received signal used is a and sampled at a low rate, as shown in Figure 3(a).
delay-Doppler shifted dictionary [21] similar to [15]. Consider The RD is adopted in [27] as the analog-mixing front end
the matrix U whose ith column is given by of a proposed quadrature-compressive-sampling (referred to as
QuadCS by Liu et al.) system. The mixed and filtered output
U i = h (t - x i) % e j2ro i t, (20) y (t), shown in Figure 3, is given by

where t is the N # 1 vector containing the sampling instants

at the Nyquist rate, i.e., t i = i / B h, and c is the Hadamard
y (t) = #-33 h bp ( t) p (t - t) x R (t - t) dt, (23)

42 IEEE Signal Processing Magazine | November 2018 |

where x R (t) is defined as the real part of (7). The RD samples method is independent of the waveform, and MF is performed
y (t) at rate fs = 1/Ts = fc /k, with k an integer satisfying directly on the low-rate samples before parameter recovery.
k # 6 fc /2B CS@ . The samples are fed to the quadrature-process-
ing system [61], which extracts the baseband in-phase and Fast time Xampling
quadrature (I and Q) components, respectively, of the radar An alternative sub-Nyquist radar method is the Xampling-
echoes. As shown in [27], the complex samples of the RD out- based system proposed in [30] and [40]. This approach, which
put can be written as may be used with any transmitted pulse shape, achieves the
minimal sampling rate required for target detection, while pro-
y = Ax. (24) viding optimal SNR.
The sub-Nyquist analog front end is composed of an ADC
Here, x is a sparse vector that contains the complex ampli- that filters the received pulse-Doppler signal (2) to predeter-
tudes a l at the corresponding delays x l, and the (m, p) ele- mined frequencies before taking pointwise samples. These
ment of the matrix A is given by compressed samples, or “Xamples,” contain the information
needed to recover the desired signal parameters, i.e., the target’
A m, p = # h bp (t) e - j2rfc t p (mTs - t) h (mTs - px - t) dt.
3
delay-Doppler map. To see this, note that the Fourier-series
-3
(25) coefficients of the aligned frames x p (t + p x) are given by

L-1
The samples of P pulses are concatenated in a matrix Y such
c p [k] = 1 H [k] / a l e -j2rkx l /x e -jol px, 0 # k # N - 1, (26)
that each column corresponds to a pulse. The subsequent pro­­ x l=0
cessing of the quadrature-compressive sampling, referred to as
compressive-sampling pulse Doppler, is composed of a DFT where H [k] are the Fourier coefficients of the known trans-
step on the rows of Y that acts as an MF in slow time followed mitted pulse h (t), and N = B h x is the number of Fourier sam-
by a MF in each column, corresponding to the fast time. ples. From (26), we see that the unknown parameters
The RMPI is a variant of the RD composed of a parallel {a l, x l, o l} lL=-01 are contained in the Fourier coefficients c p [k] .
set of RD channels driven by a common input, in which each We now show how the Fourier coefficients c p [k] may be
RD uses a distinct pseudorandom binary sequence. A hard- obtained from low-rate samples of x p (t) and how the targets’
ware RMPI-based prototype has been implemented in [43] that parameters can then be recovered from c p [k] [30].
recovers radar pulses and estimates their amplitude, phase, and The received signals x p (t) exist in the time domain; thus,
carrier frequency. In the next section, we discuss an alternative there is no direct access to c p [k] . To obtain any arbitrary set
prototype with a different analog front end, which also recov- of Fourier-series coefficients, the direct multichannel sam-
ers the targets’ parameters from low-rate samples. pling scheme [63] illustrated in Figure 4 can be used. The ana-
Note that the considerations behind waveform design for log input x p (t) is split into k = l channels, where, in each
CS recovery in the approaches [15]–[17] presented in the previ- channel k i with i ! [0, K - 1], it is mixed with the harmonic
ous section are similar to traditional radar requirements. The signal e - j2rk i t/x, integrated over the PRI duration, and then sam-
well-known ambiguity function (AF) impacts CS radar in a pled. Xampling thus allows one to obtain an arbitrary set l out of
way that is similar to traditional radar systems. Indeed, the N = xB h frequency components from K pointwise samples of the
mutual coherence of the dictionary is linearly related to the received signal after appropriate analog preprocessing. An alter-
highest sidelobe value of the AF [58], [62]. In contrast, we will native Xampling method uses the sum-of-sincs filter described in
see in the next section that the CS dictionary of the Xampling [64]. This class of filters, which consists of a sum-of-sinc ­function

Sub-Nyquist Sampling Subsystem Quadrature Demodulation Subsystem

Ics [m ]
↓2
y (t )
(–1)m
r (t )
h bp(t ) A /D
–2sin(kπ /2)
p (t )
t = k /fs
Pseudorandom hlp(t ) ↓2
±1 Generator Qcs [m ]

(a) (b)

FIGURE 3. A quadrature-compressive-sampling implementation with (a) RD sampling followed by (b) quadrature demodulation [27].

IEEE Signal Processing Magazine | November 2018 | 43

 The resulting equation (27) is a standard delay-estimation
τ
problem for each o and may be solved using multiple tech-
1
X (.)dt c [k0] niques [10]. However, improved performance can be obtained
τ 0
by jointly processing the sequences {U o [k]} for different val-
e–j 2πk0t /τ ues of o. Thus, instead of searching separately for each of the
x (t ) delays x l, l ! K (o), the L delays are estimated by jointly pro-
cessing overall Doppler frequencies.
1 τ A particularly convenient method in this case is to employ a
X (.)dt c k|k |–1
τ 0 matching pursuit-type approach in which the strongest overall
peak o, assuming a single delay, is first found:
e–j 2πk|k |–1t /τ

FIGURE 4. A multichannel direct sampling of the Fourier coefficients [63].

(xt l, ot l) = argmax
x l, o l
/W ol [k] e j2rkx l /x . (29)
k!l

in the frequency domain, is a general sampling scheme for arbi- Once the optimal values xt l and ot l are determined, their influ-
trary pulse shapes. ence is subtracted from the focused sub-Nyquist samples as
A less expensive and more practical approach for the Fou-
P-1
rier-series coefficients acquisition proposed in [40] is based
Wlo [k] = W o [k] - 1 at l e -j2rkx l /x / e j( - t ) p , (30)
t o ol x

on multiple bandpass filters and is adopted in the Xampling x p=0

hardware radar prototype described in the next section. This
system is composed of a few channels, with each sampling where
the content of a narrow frequency band of the received sig-
nal. Each channel thus yields a group of several consecutive
at l =
x
/ W t [k] e j2 k t /
P | l | k!l ol
r xl x
. (31)
Fourier coefficients. The multiple bandpass constellation has
the advantage of acquiring the measurements over a wider fre-
quency aperture. At the same time, it still allows for practical The same operations are performed iteratively to find all of the
hardware implementation, as detailed in the next section. By desired L peaks. This approach does not require discretization
widening the frequency aperture, a finer resolution grid may of the targets’ parameters, and these are recovered over the
be employed during the recovery process. Moreover, empiri- continuous domain from a minimal number of samples.
cal results show that highly distributed frequency samples In practice, the search for peaks can be limited to a grid,
provide better noise robustness [40]. However, widening the which enables all of the computations to be carried out using
frequency aperture eventually requires increasing the number simple FFT operations. Suppose we limit ourselves to the
of samples K; otherwise, recovery performance may degrade. Nyquist grid, i.e., the grid defined by the Nyquist resolu-
This ­tradeoff is observed in the experiments presented in [40]. tion so that x l /x = s l /N, where s l is an integer satisfying
Once a set of Fourier coefficients c p [k] has been acquired, 0 # s l # N - 1. Then, (26) is approximately written in vector
the delays and Doppler frequencies can be recovered using dif- form as
ferent techniques. Doppler focusing [30], summarized in “Dop-
pler Focusing,” is one approach that has several advantages, as W o = PHF KN a o, (32)
detailed next. This method uses target echoes from all of the pulses
to generate a focused pulse at a specific Doppler frequency. It then where W o = 6W o [k 0] fW o [k K - 1]@, k i ! l for 0 # i # K - 1,
jointly recovers the delay-Doppler map by reducing the detection H is a diagonal matrix that contains the Fourier coefficients
problem to a one-dimensional, delay-only estimation. Performing H [k] of the transmitted waveforms, and F KN is the partial-
the Doppler focusing operation in frequency results in computing Fourier matrix that contains the K rows of the N # N Fourier
the DFT of the coefficients c p [k] in the slow time domain: matrix indexed by k. The entries of the L -sparse vector a o
are the values a l at the indices s l for the Doppler frequ­
P- 1
W o [k] = / c p [k] e j p o x encies o l in the “focus zone,” i.e., | o - o l | 1 r/Px. The P
p= 0 equations (32) are simultaneously solved using CS-based algo-
L- 1 P- 1
rithms, which, during each iteration, the maximal projection
= 1 H [k] / a l e - j2 k r x l /x
/ e j( - o o l ) px
. (27)
x l= 0 p= 0 of the observation vectors onto the measurement matrix is
retained [30].
Note that W o [k] is the Fourier series of U (t, o), defined in Some results comparing different configurations of low-rate
(S11), with respect to t. Following the same argument as in sampling and processing are shown in Figure 5 [30]. The recov-
(S12), we have ery performance of the classic processing applied to Nyquist
samples is presented as a baseline. Sub-Nyquist ap­­proaches,
W o [k] . P H [k] / al e
-j2rkx l /x
. (28) performed at 1/10 of the Nyquist rate, include the same classic
x l ! K (o) processing applied to sub-Nyquist samples, a t­wo-stage CS

44 IEEE Signal Processing Magazine | November 2018 |

recovery method that performs delay and The minimal number of dictionary depends on samples of the wave-
Doppler estimation in parallel, separate­­ samples required for the form h (t), such that the mutual coherence
ly [30], and Doppler focusing. It is clearly of the dictionary is linearly related to the
seen that Doppler focusing applied to ran-
perfect recovery of {al , highest sidelobe value of the AF [58]. In
dom Fourier coefficients, which are wide- xl , ol } with L targets in contrast, the CS dictionary of the Xampling
ly distributed with high probability leading a noiseless environment method is independent of the waveform,
2
to a wide aperture, outperforms other sub- is 4L , with at least K ≥ 2L as shown in (32). Third, in the presence
Nyquist approaches. The use of consec- samples per pulse and at of additive white noise, Doppler focusing
utive coefficients yields small aperture and least P ≥ 2L pulses. achieves an increase in SNR by a factor
poor resolution. of P (a detailed analysis may be found in
The Xampling approach has several ad­­­ [30]). In addition, this approach can oper-
vantages. First, it recovers the targets’ parameters directly ate at the smallest possible sampling rate for recovering the
from the low-rate samples without requiring sampling at the targets’ parameters, as derived in [30]. The minimal number
Nyquist rate. Second, previous CS-based methods typically of samples required for the perfect recovery of {a l, x l, o l}
impose constraints on the radar transmitter, which are not with L targets in a noiseless environment is 4L2, with at least
needed here. Indeed, as may be seen in (20) and (25), the CS K $ 2L samples per pulse and at least P $ 2L pulses. The

Doppler Focusing
Doppler focusing is a processing technique, suggested in
[30], which uses target echoes from different pulses to cre- 200
ate a superimposed pulse focused at a particular Doppler
frequency. This method allows for joint delay-Doppler
recovery of all targets present in the illuminated scene. It 150
results in an optimal signal-to-noise ratio (SNR) boost and
Uniform g (v |v1)

may be carried out in the frequency domain, thus enabl­­

ing sub-Nyquist sampling and processing with the same 100
SNR increase as a matched filter.
The output of Doppler processing can be viewed as a
discrete equivalent of the following time shift and modula- 50
tion operation on the received signal:

P-1
U (t, o) = / x p (t + px) e j p o x
0
p=0 –4 –2 0 2 4
L-1 P-1
Normalized v
= / a l h (t - x l ) / e j (o - o l ) px
. (S11)
l=0 p=0
FIGURE S1. The sum of exponents | g (o | o l ) | for P = 200, x = 1s,
Consider the sum g (o ; o l ) = R Pp -= 10 e j (o - o l ) px . For any given and o l = 0.
o, targets with Doppler frequencies o l in a bandwidth of
2r Px around o will achieve a coherent integration and
an SNR increase of approximatively P. On the other hand, where K (o) = " l : | o - o l | 1 r/Px , . In other words, the
since the sum of P equally spaced points covering the unit sum is only over the targets whose Doppler shifts are in
circle is generally close to zero, targets with o l not “in the interval | o - o l | 1 r/Px.
focus” will roughly cancel out. In summary, we have that For each Doppler frequency o, U (t, o) represents a stan-
dard pulse-stream model in which the problem is to esti-
P-1
P | o - o l | 1 r/Px mate the unknown delays. Thus, using Doppler focusing,
g (o ; o l ) = / ej( -o o l ) px
.)
0 | o - o l | $ r/Px,
(S12)
the two-dimensional delay-Doppler recovery problem is
p=0
reduced to delay-only estimation for a small range of
as shown in Figure S1. Doppler frequencies, with increased SNR by a factor of P
We may therefore estimate the sum of exponents in [10]. The Xampling radar of [30] performs Doppler focus-
(S11) as ing directly on the low-rate samples in the frequency
U (t, o) . P / a l h (t - x l ), (S13) domain, allowing for joint Doppler-delay recovery from
l ! K (o) the “Xamples.”

IEEE Signal Processing Magazine | November 2018 | 45

 Doppler focusing approach achieves this minimal number of
samples. Finally, Doppler focusing is able to deal with certain
Hit Rate
models of clutter and target dynamic range by adding a simple
1
windowing operation in the sum (27) and by prewhitening in
0.9 frequency [54].
0.8 The Xampling radar was implemented in hardware, as
0.7 described in the next section, demonstrating real compressed
radar capabilities. The hardware prototype is built from off-
0.6
the-shelf components, which are bandpass filters and low-rate
0.5 samplers, leading to low hardware complexity.
0.4
Hardware prototype
0.3
Xampling is used in combination with Doppler focusing in
0.2 the sub-Nyquist prototype of [30], [40], which demonstrates
0.1 radar reception at sub-Nyquist rates. The input signal simu-
–30 –25 –20 –15 lates reflections from arbitrary targets and is corrupted by
SNR (dB)
additive noise and clutter. The radar receiver implements
the multichannel topology described in the previous section
Classic Processing (Nyquist)
Classic Processing (Sub-Nyquist) and samples a signal with Nyquist rate of 30 MHz with a com-
Two-Stage CS (Sub-Nyquist) pression factor of 30. Hardware experiments demonstrate
Doppler Focusing-Consecutive Coefficient (Sub-Nyquist) the feasibility of detecting targets from the low-rate sam-
Doppler Focusing-Random Coefficient (Sub-Nyquist)
ples of an analog radar signal using standard radio-frequen-
cy (RF) hardware [30], [40]. Typical experiment results
FIGURE 5. The hit rate of classic processing, two-stage CS recovery, are shown in Figure 6, which depicts the input signal, the
and Doppler focusing for a fixed false alarm rate. A hit is defined as low-rate samples, and the original and recovered delay-
a delay-Doppler estimate circumscribed by an ellipse around the true Doppler maps, including close targets, both in terms of range
target position in the time–frequency plane, with the axes equivalent
and velocity.
to ! 3 times the time and frequency Nyquist bins. The two-stage CS
recovery separates the delay and Doppler estimation, performing At the heart of the receiver lies the Xampling-based ADC,
them in parallel [30]. The sub-Nyquist sampling rate was 1/10 of the which performs analog prefiltering of the signal before tak-
Nyquist rate [30]. ing pointwise samples. A multiple bandpass-sampling approach

FIGURE 6. The Xampling radar LabView experimental interface. From left to right: at top is the received signal from targets only, then, the received signal
from clutter, noise, and overall received signal x p (t). At the bottom are the sub-Nyquist samples of the four channels at 1/30 of the Nyquist rate, then, the
true and recovered delay-Doppler maps. All of the targets (including close targets both in range and in velocity) are correctly detected.

46 IEEE Signal Processing Magazine | November 2018 |

with four channels is adopted. Each channel is composed of Slow time compression
a crystal filter with a bandwidth of 80 KHz and extremely Most works on CS radar focus on compression in the fast
narrow transition bands and then sampled at a rate of 250 kHz. time domain, reducing the number of samples per pulse
The front-end samples four distinct bands below the Nyquist rate. As we have seen,
of radar–signal spectral content, yield- At the heart of the receiver using appropriate CS techniques allows
ing 320 Fourier coefficients after digital lies the Xampling-based for preserving the range resolution while
processing with a total sampling rate of operating in low-rate regimes by break-
1 MHz. The samples are fed into the chas-
ADC, which performs ing the link between bandwidth and sam-
sis controller, and a MATLAB function is analog prefiltering of pling rate. This is illustrated in Figure 5,
launched that computes the 320 Fourier the signal before taking in which Doppler focusing is shown to
coefficients via FFT, composed of four pointwise samples. achieve the same hit rate as classic pro-
groups of 80 consecutive Fourier coef- cessing above a certain SNR, and in Fig­­
ficients. These are then used for digital recovery of the ure 6, where close targets are seen to be correctly recovered
delay-Doppler map using the Doppler focusing recon- despite sampling at 3.3% of the Nyquist rate. We will now
struction algorithm. see that compression may be similarly performed in the
The experimental setup is based on National Instrument slow time domain, as demonstrated in [65], where the num-
(NI) PXI-series equipment that is used to synthesize a radar ber of transmitted pulses is reduced without decreasing
environment and ensure system synchronization. The entire Dopp­­ler resolution.
component ensemble wrapped in the NI chassis as well as
the analog receiver board are depicted in Figure 7. Additional Nonuniform pulse Doppler
information regarding the system’s configuration and synchro- The resolution in Doppler frequency in standard processing
nization can be found in [40]. is governed by the number of transmitted pulses P. More
To demonstrate target detection from low-rate samples, the precisely, it is equal to 2r / Px. However, a large P leads to
Applied Wave Research (AWR) software is used to simulate large CPI and long time on target. Slow time compression
the radar scenario, including pulse transmission and accurate breaks the relation between CPI and time on target. To that
power loss due to wave propagation in a realistic medium. end, M < P pulses are sent nonuniformly over the entire
AWR software provides a computer-based environment for CPI Px, implementing nonuniform time steps between the
designing hardware for use with wireless and high-speed digi- pulses [65]. This way, the same CPI is kept, but a smaller
tal products. It is used for RF, microwave, and high-frequency number of pulses is transmitted, thereby reducing power
analog circuits and system design. A large variety of scenarios, consumption. In addition, the periods of time in which no
consisting of different targets’ parameters, i.e., delays, Doppler pulse is transmitted in a certain direction can be exploited to
frequencies, and amplitudes, are examined in [30] and [40]. send pulses in other directions. This allows the radar to scan
An arbitrary waveform generator module produces an analog several directions at the same time and obtain the corre-
signal that is amplified and routed to the radar receiver board. sponding delay-Doppler maps in a single CPI. However, at
The received radar waveform is contaminated with noise and the same time, this reduces SNR because fewer pulses are
clutter, showing the capabilities of the Xampling receiver to transmitted in every direction.
deal with these [30], [40], [54]. The Nyquist rate of the signal is Consider a nonuniform pulse-Doppler radar such that the
-1
30 MHz, so that sampling at 1 MHz corresponds to a fast time pth pulse is sent at time m p x, where {m p} M
p = 0 is an ordered set
compression factor of 30. of integers satisfying m p $ p. In this case, (1) becomes

Analog Preprocessor User-Control Interface

Radar Display

Auditory Waveform Generator Digital Receiver

FIGURE 7. The Xampling radar prototype including an arbitrary waveform generator, receiver board, NI chassis, and display [40].

IEEE Signal Processing Magazine | November 2018 | 47

 M-1 natively, instead of solving the matrix problem of (35), we can
x T (t) = / h (t - m p x), 0 # t # Px, (33) apply the Doppler focusing operation [30] described in “Dop-
p=0
pler Focusing.” As illustrated in Figure 8, the approximation
and the received frames (4) are written as from (S12) may still be applied in the nonuniform case, where
m p $ p. Therefore, we can rewrite the Fourier coefficients
L-1
from (27) by replacing p with m p for the nonuniform case.
x p (t) = / a l h (t - x l - m p x) e -j ol m p x
, 0 # t # Px. (34)
These may then be approximately expressed in vector form as
l=0
in (32) and recovered as previously described. It is shown in
The same Xampling-based method in [30] is used to obtain [65] that the minimal number of nonuniform pulses required
the Fourier coefficients c p 6k@ of the received pulses. Suppose to recover the Doppler frequencies of L targets is identical to
we limit ourselves to the Nyquist grid, as previously men- the uniform case, that is, two L pulses.
tioned, so that x l /x = s l /N, where s l is an integer satisfying
0 # s l # N - 1, and o l x = 2rrl /M, where rl is an integer in Hardware simulation
the range 0 # rl # M - 1. Similar to the derivations in the pre- The transmission of nonuniform pulses has been implemented in
vious section, we can write the Fourier coefficients c p 6k@ in the Xampling prototype [40]. Recall that the received signal has a
matrix form [65] as bandwidth of 30 MHz and is sampled at the rate of 1 MHz. To
this fast time compression, we now add compression in the slow
T
X = HF KN A ^F PM h , (35) time domain. In the hardware simulation, P = 50 pulses over a
CPI of MPx = 2.5 s are considered. Half of the pulses, i.e.,
where H is a diagonal matrix that contains the Fourier coef- M = 25, chosen at random, are sent in one direction, while the
ficients H 6k@ . The partial-Fourier matrix F PM contains M other half are sent in a second direction. Two delay-Doppler maps
rows from the P # P Fourier matrix, indexed by the values are then simultaneously recovered during a single CPI, as shown
of the transmitted pulses m p, 1 # p # M; when sampling at in Figure 9. Both of the maps are fully recovered, as previously
the Nyquist rate, K = N and F KN become the standard mentioned, from compressed samples in both the fast and slow
N # N Fourier matrix. Similarly, when considering uniform- time domains.
ly spaced pulses, M = P and F PM are the standard P # P
matrix. The goal is to recover the sparse matrix A that con- Range-velocity ambiguity resolution
tains the values a l at the L indices {s l, rl} from the Fourier As presented thus far, targets are traditionally assumed to lie
coefficients matrix X. in the radar-unambiguous range-velocity region. For a given
CS matrix-recovery algorithms are directly applicable to PRI x, the maximum unambiguous range is rmax = cx/2, and
(35) by extending CS techniques presented in vector form, such the maximum unambiguous velocity is ro max = m/ (4x), where
as orthogonal matching pursuit or the fast iterative-shrinkage- m is the radar wavelength. When the target range and velocity
thresholding algorithm [10], [12] to matrix settings [66]. Alter- intervals of interest are large, traditional pulse-Doppler radar

200 100
180 90
160 80
140 70
Nonuniform g (v |vl)
Uniform g (v |vl)

120 60
100 50
80 40
60 30
40 20
20 10
0 0
–0.5 –0.4 –0.3 –0.2 –0.1 0 0.1 0.2 0.3 0.4 0.5 –0.5 –0.4 –0.3 –0.2 –0.1 0 0.1 0.2 0.3 0.4 0.5
Normalized v Normalized v
(a) (b)

FIGURE 8. The sum of exponents | g (o | o l ) | for M = 200, x = 1s, and o l = 0 in the (a) uniform and (b) nonuniform cases. In the nonuniform case,
P = 100 pulses are chosen uniformly at random [65].

48 IEEE Signal Processing Magazine | November 2018 |

systems suffer from the so-called Doppler dilemma [67], a [70], relies on the Chinese remainder theorem [5] and uses two
tradeoff between range and velocity ambiguity, whose product PRFs, such that the numerator and denominator of the ratio
is limited to rmax ro max = cm/8. between them are prime numbers. The ambiguous velocities
Several techniques have been proposed over the years to are computed for each train i as
mitigate the range-velocity ambiguity by increasing either of
these parameters. Two main PRF variation-based methods
rto i, k = rto i, 0 + k m , k ! Z, (36)
are staggered PRFs and multiple PRFs (MPRFs). Staggered 2x
PRFs are used to raise the first blind speed ro max significantly
without degrading the unambiguous range [7]. Pulse-to-pulse where rto i, 0 is the velocity estimate within the unambiguous
stagger varies the PRF from one pulse to the next, achieving velocity interval (- ro max, ro max] . Congruence between these
increased Doppler coverage [68]. The main are found by an exhaustive search, so that
disadvantage of this approach is that the all rto i, k fall within a small, interval, or
data correspond to a nonuniformly sampled Ambiguity resolution is correlation bin. The resulting velocity esti-
sequence, making it more difficult to apply typically achieved by mate is computed by averaging overall rto i, k .
coherent Doppler filtering [7]. In addition, searching for coincidence Assuming T = 2 pulse trains with PRFs and
clutter cancellation becomes more challeng- between either unfolded ratio x 1 /x 2 = m/n, where m and n are
ing, and the sensitivity to noise increases Doppler or delay estimates relative prime numbers, the expanded
[69]; therefore, MPRF techniques are typi- velocity interval is of size mm/2x 1 = nm/2x 2 .
cally preferred. We now review some of the
for each PRF. However, in this approach, a small range
MPRF-based methods and then present a error on a single PRF can cause a large
Xampling approach that solves the delay-Doppler ambiguity error in the resolved range with no indication that this has
using phased-coded-transmit pulses. happened [71].
A clustering algorithm proposed in [71] implements the
MPRF search for a matching interval by computing average distances
The MPRF approach transmits several pulse trains, each with to cluster centers. The average squared error is defined as
a different PRF. Ambiguity resolution is typically achieved by
T
searching for coincidence between either unfolded Doppler or C (k) = / | rti, k - rk | 2, k = 0, f, ramb /rmax, (37)
delay estimates for each PRF. A popular approach, adopted in i= 1

(a) (b)

FIGURE 9. The experimental interface of the Xampling radar, with both fast and slow time compression. (a) The true targets’ ranges in two directions (top)
and superposed low-rate samples from both directions (bottom). (b) The range-velocity map of true and recovered targets in both directions [65].

IEEE Signal Processing Magazine | November 2018 | 49

 where rk is the median value of the T ranges with index k The Fourier series of the received signal (40) can be writ-
and ramb is the maximal ambiguous range. The best cluster ten in matrix form, similarly to (35), and recovered using
occurs at the value of k, where C (k) is minimized. This matrix CS recovery techniques [72]. The minimal number of
happens when all of the ambiguous rang- samples per pulse allowing recovery of X
es are unfolded correctly, and, hence, all with high probability is found to be K > 2L,
of the range estimates have nearly the same
Random PC has been and the minimal number of pulses P is
range. This technique still requires an adopted in polarimetric 2L + Q + 2. This method resolves a maxi-
exhaustive search over clusters and does weather radars, which mum unambiguous range rmax = cQx/2,
not process the samples jointly, thereby exploit the inherent while preserving the maximum unambigu-
decreasing the SNR. random phase between ous velocity ro max = m/ (4x), thereby increas-
ing their product rmax ro max by a factor of Q,
pulses of the popular
Phased-coded pulses under the prior mentioned conditions on the
A random-pulse, phase-coding (PC) ap­­
magnetron transmitters. number of samples and pulses.
proach is adopted in [72] to increase the This approach has three main advantag-
range-unambiguous region, while preserving that of the es. First, it improves the delay estimation with respect to the
Doppler frequency and using a single PRF. A similar tech- MPRF methods since it preserves the resolution of traditional
nique may be used to increase the Doppler-frequency- pulse-Doppler radar, i.e., 1/B h, while increasing the unambigu-
unambiguous region. Random PC has been adopted in ous delay region to Qx. Second, it increases the SNR by jointly
polarimetric weather radars, which exploit the inherent processing the samples from all of the pulse trains, rather than
random phase between pulses of the popular magnetron matching the estimated parameters from each pulse processed
transmitters. In this context, PC mitigates out-of-trip echoes separately. Finally, it provides a systematic delay-Doppler-
[73]. The approach of [72] introduces a random phase, which recovery method that does not involve an exhaustive search.
differs from pulse to pulse. The joint processing of recei­­ From a practical point of view, this approach does not require
ved signals from all of the trains is the key to range ambigui- the use of several pulse trains with different PRFs, thus sim-
ty resolution. plifying hardware implementation.
The pulse-Doppler radar transceiver se­­quentially transmits
one modulated pulse train, consisting of P equally spaced Cognitive radar and spectrum sharing
pulses. For 0 # t # Px, the transmitted signal is given by Recently, the concept of cognitive radar (CR) [74], inspired by
the echo-location system of a bat, has been presented as a nat-
P-1 ural next step for traditional radar. The cognition property
x T (t) = / h (t - px) e jc [p], (38) requires adaptive transmission and reception capabilities, i.e.,
p=0
both the transmitter and receiver are able to dynamically
adjust to the environment conditions. Many interpretations of
where 6 @ is uniformly distributed in the interval
c p [0 , 2r) and this idea have been proposed. We focus on one aspect of cog-
represents the phase shift of the pth pulse. nition, the dynamic and flexible adaptation to the spectral
As opposed to the common assumption in traditional radar, environment, which allows for spectrum sharing between
the targets’ time delays xu l are not assumed to lie in the unam- communication and radar systems [36]–[38]. The interest in
biguous time region, i.e., less than the PRI x, but rather in the these spectrum sharing radars is largely due to electromagnet-
ambiguous range xu l ! [0, Qx), where Q < P is the ambiguous ic spectrums being a scarce resource, with most services hav-
factor defined by the targets’ maximal range. For convenience, ing a need for a greater access to it.
the delay xu l is decomposed into its integer part (the ambiguity The spectrum sharing solution proposed in [39] capitalizes
order) q l x and fractional part (the folded or reduced delay) on the cognitive abilities of the radar system. It is shown how
x l such that compressed radars may be adapted to allow for spectral coex-
istence between communication and radar signals and flexibil-
xu l = x l + q l x, (39) ity of the radar transmission. This demonstrates that, beyond
increasing resolution and realizing compression in the time,
where 0 # q l # Q - 1 is an integer and 0 # x l < x. frequency, and spatial domains, compressed radars have the
The received signal is then potential to enable otherwise challenging technologies.

P-1 L-1
Spectral adaptive transmission
x R (t) = / / a l h (t - xu l - px) e -j2 ro l (p + q l) x
e jc [p], (40)
In previous works that implement fast time compression, e.g.,
p=0 l=0
Xampling radar [30], [40], the transmitter broadcasts a wide-
for 0 # t < (P + Q) x. The main difference with traditional band signal, which reflects off the targets and propagates back
pulse-Doppler radar, aside from the coded phase, is that the to the receiver. The received signal is then filtered before sam-
PRI index in the Doppler shift term is p + q l, rather than the pling so that only the content of a few narrow bands is sam-
pulse index p. pled and processed. These works only deal with the reception

50 IEEE Signal Processing Magazine | November 2018 |

side of the radar, providing sampling and processing tech- tion between the two systems. The CR configuration is key
niques that can be used with any traditional radar transmitter. to spectrum sharing since the radar transceiver can adapt its
However, for broadband frequency occupation and power sav- transmission to available bands, thus achieving coexistence
ings, only the narrow frequency bands with communication signals. Suppose the set
that are to be sampled may be transmit- of all frequencies of the available common
To increase the efficiency
ted [31], [39]. This will not affect any system spectrum is given by F. The commu-
aspect of the processing since the re­­ of spectrum management, nication and radar systems occupy the subsets
ceived signal is preserved in the bands of a dynamic opportunistic FC and FR of F, respectively. The goal is
interest. In fact, since all the signal exploitation of temporarily to design the radar waveform and its support
power is concentrated in the processed vacant spectral bands FR, conditional on the fact that the commu-
bands, the SNR increases, and the detec- by secondary users has nication occupies frequencies Fc, which are
tion performance improves [75]. unknown to the radar transceiver [39]. To
u ( f ) be the CTFT of the new been considered, called
Let H de­­tect the bands left vacant by the commu-
transmitted radar pulse, cognitive radio. nication signals, spectrum sensing needs to
be performed over a large bandwidth. Such
u ^ f h = ) H ^ f h f ! 8 f r - B r 2, f r + B r 2B for 1 # i # N b
i i i i a task has recently received tremendous interest in the com-
H munication community, which faces a bottleneck in terms of
0 else,
(41) spectrum availability. To increase the efficiency of spectrum
managment, a dynamic opportunistic exploitation of tempo-
where N b is the number of filtered bands and B ir and f ri are rarily vacant spectral bands by secondary users has been con-
the bandwidth and center frequency of the ith band, respec- sidered, called cognitive radio (CRo) [80], [81].
tively. Obviously, the computation of the relevant Fourier A spectrum sharing paradigm using Xampling techni­
coefficients c p 6k@ from (25) will not change. Therefore, the ques, the spectral coexistence via Xampling (SpeCX) sys-
recovery methods presented in the “Fast Time Xampling” sec- tem [39] is composed of a sub-Nyquist CRo receiver [81] to
tion are applicable here as well. detect the occupied communication bands so that the radar
The concept of transmitting only a few subbands that the transmitter may subsequently exploit the spectral holes. In
receiver processes is one way to formulate a frequency-agile this setting, the received signal at the communication receiv-
CR in terms of its ability to adapt to spectral demands. Com- er is given by
plying with CR requirements, the support of the subbands
varies with time to allow for dynamic and flexible adaptation. x (t) = x C (t) + x R (t), (42)
Such a system also enables the radar to disguise the transmit-
ted signal as an electronic countermeasure or to cope with where x R (t) = rTX (t) + rR X (t) is the radar signal sensed by the
crowded spectrums by using a smaller, interference-free por- communication receiver, composed of the transmitted and
tion, as further discussed in the following section. received radar signals. The goal is therefore to recover the
support of x C (t), given the known support of x R (t), which is
Application to spectrum sharing shared by the radar transmitter with the communication
The unhindered operation of a radar that shares its spectrum with receiver. This can be formulated as a sparse recovery with par-
communication systems has captured a great deal of attention tial-support knowledge, studied under the framework of a
within the operational radar community in recent years [36]– modified CS [82].
[39]. Recent research programs in spectrum sharing radars Once FC is identified, the communication receiver provides
include the Enhancing Access to the Radio Spectrum project a spectral map of occupied bands to the radar. Equipped with
by the National Science Foundation [38] and the Shared the detected spectral map and known radio environmental map,
Spectrum Access for Radar and Communication (SSPARC) pro- the objective of the radar is to identify an appropriate transmit-
gram [37], [76], initiated by the Defense Advanced Research frequency set FR 1 F = FC such that the radar’s probability
Projects Agency. of ­detection Pd is maximized. For a fixed probability of false
A variety of system architectures have been proposed for alarm Pfa, the Pd increases with a higher signal-to-interference
spectrum sharing radars, and most place an emphasis on opti- plus-noise ratio (SINR) [83]. Hence, the frequency selection pro-
mizing the performance of either radar [77] or communication cess can, alternatively, choose to maximize the SINR or mini-
[78] while ignoring the performance of the other. In nearly all mize the spectral power in the undesired parts of the spectrum.
cases, the real-time exchange of information between radar To find available bands with the least amount of interference, a
and communication hardware has not yet been integrated into structured sparsity framework [84] is adopted in [39]. Additional
the system architectures. Exceptions to this are automotive requirements of transmit-power constraints, range-sidelobe lev-
solutions in which the same waveform is used for both target els, and minimum separation between the bands can also be
detection and communication [79]. imposed. At the receiver of this spectrum-sharing radar, the sub-
In a similar vein, the sub-Nyquist, CR-based approach Nyquist processing method of [30] recovers the delay-Doppler
from [39] incorporates the handshaking of spectral informa- map from the subset of Fourier coefficients defined by FR .

IEEE Signal Processing Magazine | November 2018 | 51

 This CR system leads to three main advantages. First, the CS performed at the low rate of 120 MHz, very low computational
reconstruction, as presented in [30] on the transmitted fragment- load is required to achieve real-time recovery. The prototype
ed bands, achieves the same resolution as traditional Nyquist is fed with RF signals composed of up to N sig = 5 real com-
processing over a significantly smaller bandwidth. Second, by munication transmissions, i.e., ten spectral bands with a total
concentrating all of the available power in the transmitted nar- bandwidth occupancy of up to 200 MHz and varying support,
row bands rather than over a wide bandwidth, the CR increases with a Nyquist rate of 6 GHz.
the SNR. Finally, this technique allows for a dynamic form of The input transmissions then go through an RF combiner,
the transmitted signal spectrum, in which only a small portion resulting in a dynamic multiband input signal that enables fast
of the whole bandwidth is used at each transmission, thereby carrier switching for each of the bands. This input is specially
enabling spectrum sharing with communication signals, as designed to allow the testing of the system’s ability to rapidly
illustrated in Figure 10(d). There, the coexistence between radar- sense the input spectrum and adapt to changes, as required
transmitted bands in red and existing communication bands in by modern CRo and shared spectrum standards, e.g., in the
white is shown. SSPARC program. The system’s effective sampling rate, equal
to 480 MHz, is only 8% of the Nyquist rate. Support recovery
SpeCX prototype is digitally performed on the low-rate samples. The prototype
The SpeCX prototype, presented in Figure 10, demonstrates successfully recovers the support of the communication trans-
radar and communication spectrum sharing. It is composed of mitted bands, as demonstrated in Figure 10(b) and (c). Once
a CRo receiver and a CR transceiver. At the heart of the CRo the support is recovered, the signal itself can be reconstructed
system lies the proprietary modulated wideband converter from the sub-Nyquist samples. This step is performed in real
board [29] that implements a sub-Nyquist analog front-end time, reconstructing the signal bands one sample at a time.
receiver, which processes signals with Nyquist rates up to The CR receiver system is identical to the sub-Nyquist sam-
6 GHz. The card first splits the wideband signal into M = 4 pling prototype of [30], [31] and [40]. In the cognitive case, the
hardware channels with an expansion factor of q = 5, yield- transmitter only transmits over N b = 4 bands, which constitute
ing M q = 20 virtual channels after digital expansion (see 3.2% of the original wideband signal bandwidth after the spec-
[85]). In each channel, the signal is mixed with a periodic trum-sensing process has been completed by the communica-
sequence p i (t), which are truncated versions of Gold codes tion receiver. Figure 10(d) illustrates the coexistence between
[86], generated on a dedicated field-programmable gate array, the radar-transmitted bands in red and the existing communi-
with a periodic frequency f p = 20 MHz. cation bands in white. The gain in power is demonstrated in
Next, the modulated signal passes through an analog anti- Figure 10(e); the wideband radar spectrum is shown in blue,
aliasing LPF. Finally, the low-rate analog signal is sampled by the CR in red, and the noise in yellow on a logarithmic scale.
an NI ADC operating at fs = (q + 1) f p = 120 MHz (with in­­ The true and recovered range-velocity maps are presented in
tended oversampling), leading to a total sampling rate of 480 MHz. Figure 10(f). All of the L = 10 targets are perfectly recovered,
The digital receiver is implemented on an NI PXIe-1065 com- and the clutter, depicted in blue, is discarded. Below the map,
puter with a dc-coupled ADC. Since the digital processing is the range-recovery accuracy is shown for three scenarios: from

(a) (c)

(b)

(d)

Communication Communication Communication

Analog Rx Digital Rx Display

(e)

Signal Generator Radar Analog Rx Radar Digital Rx Radar Display

FIGURE 10. The SpeCX prototype. The system consists of a signal generator, a CRo communication analog receiver, including the modulated wideband
converter (MWC) analog front-end board, a communication digital receiver, a CR analog, and a receiver. The SpeCX communication system display
shows (a) low-rate samples acquired from one MWC channel at a rate of 120 MHz and (b) a digital reconstruction of the entire spectrum from sub-
Nyquist samples. The SpeCX radar display shows (c) the coexisting communication and CR, (d) the CR spectrum compared with the full-band radar, and
(e) the range-velocity display of both the detected and true locations of the targets [39].

52 IEEE Signal Processing Magazine | November 2018 |

left to right, the CR in blue (2.5 m), the four adjacent bands with the columns of the dictionary A are given by vec (Z) for all
the same bandwidth (12.5 m), and the wideband (4 m). The sec- possible combinations of i, o, and x on a predefined grid.
ond configuration selects four adjacent frequency bands with The targets’ parameters are recovered by matching the
the same bandwidth as the CR (with nonadjacent bands) for received signal with dictionary atoms. To achieve measure-
transmission. Its poor resolution stems from its small aperture. ment diversity, random waveforms may be used, while the
The CR system with nonadjacent bands yields better resolution antenna locations are deterministic. ULAs are considered in
than traditional wideband transmission, sampling, and pro- [18] for both the transmit and receive arrays that do not benefit
cessing at the Nyquist rate, due to the increased SNR. from the virtual array configuration. Alternatively, determinis-
tic waveforms can be used, e.g., Kerdock codes [19], while the
Compressed MIMO radar antenna locations are selected uniformly at random over the
Compressed radar methods have recently been extended to aperture Z = TR/2. Bounds on N, with respect to the number
MIMO settings, in which their impact may be even greater of antennas T and R and the number of samples that ensure
than for single-antenna configurations. MIMO radar systems targets’ parameters recovery, are provided in [18] and [19].
belong to the family of array radars, which allow for the si­­ A similar approach extends the framework of [15] to the
multaneous recovery of the targets’ ranges, Dopplers, and MIMO setting by adding an azimuth matrix to the time-shift
azimuths. This 3-D recovery results in high digital processing and frequency-modulation matrices T and M, respectively, as
complexity. One of the main challenges of MIMO radar is defined in (14). In this case, each target lying on the grid is rep-
therefore coping with complicated systems in terms of cost, resented by a time shift, a frequency modulation, and an angle
high computational load, and hardware implementation. CS A q, m = e ji (p m + g q) [87].
has been naturally applied to MIMO to reduce the processing In both works, assuming N grid points in each dimension,
complexity on the digital side as well as allow for spatial com- the number of columns of A is N 3 . The processing efficiency
pression, in addition to the time compression achieved in sin- is thus penalized by a very large dictionary that contains every
gle-antenna systems. In MIMO radars, the array aperture, parameter combination. Note that the previously mentioned
which depends on the number of antennas, dictates the azimuth works focus on increased parameter resolution and do not deal
resolution. Since the aperture is determined by the number of with reduced time/spatial sampling and processing rates.
antennas in traditional virtual ULAs, high-azimuth resolution
requires a large number of antennas. Reduced processing
Fast time compression is performed in [23]–[25], in which the
Increased resolution Nyquist rate samples are compressed in each antenna before
As in single-antenna radar systems, CS has first been exploit- being forwarded to the central unit. A circular array is adopted in
ed to increase the parameter resolution. Here, the MIMO array [23], with transmit and receive nodes uniformly distributed on a
is composed of T transmitters and R receivers so as to disk with a small radius. At each receive antenna, linear projec-
achieve the desired aperture Z = TR/2, as shown in Figure 2. tions of the measurement vector are retained so that the resulting
The transmitted signal at the mth transmit antenna is given by samples are compressed in both the slow and fast time domains.
(12), and each receiver samples the received signal at the Both individual reconstruction at each receiver and joint process-
Nyquist rate, as in a traditional MIMO. Assuming a sparse tar- ing at a fusion center are proposed, using CS recovery methods.
get scene, in which the ranges, Dopplers, and azimuths lie on The actual sampling is still performed at the Nyquist rate.
a predefined grid, the work of [15] is extended to MIMO The MIMO matrix completion (MIMO-MC) radar [24],
architectures in [18] and [19]. The transmit and receive array [25] employs MC techniques to avoid parameter discretization,
manifolds respectively, are given by which is typically used in CS methods. Two configurations are
proposed for azimuth-Doppler recovery in a range bin of inter-
a T (i) = 6e j2rp 1 i, e j2rp 2 i, f, e j2rp T i@ ,
T
(43) est. In the first scenario, each receiver performs an MF and
forwards the maximum of each MF output to the fusion center.
and The samples from the pth pulse transmitted to the fusion center
can then be written in matrix form as
a R (i) = [e j2rg 1 i, e j2rg 2 i, f, e j2rg R i] T , (44)
X p = A R RD p A TT , (46)
where p m and g q are the relative mth transmit and qth receive
antenna spacings. The R # N received signal matrix from a where X p is the R # T matrix that contains the maximum of
unit strength target at direction i, with delay x and Doppler o the MF output for each transmitter and each receiver. For
is defined as ULA configurations, the lth column of the T # L transmitter-
steering matrix A T is given by (A T ) l = 61, e j d T sin (i l), f,
r 2
m

j 2 (T - 1) d T sin (i l)@T
r
Z= a R (i) a TT (i) S T (x, o) . (45) e m , where d T is the interelement spacing. The
steering matrix A R at the receiver is similarly defined. The
Here, S T (x, o) i, m = s m (t i - x) e j2rot i, where t i are the sampling diagonal matrix R contains the targets’ RCS a l, and the diag-
times and m indexes the transmitted waveforms. In this case, onal matrix D contains the targets’ Dopplers such that

IEEE Signal Processing Magazine | November 2018 | 53

 bin is investigated in [28]. Spatial compressive sampling is
λZ performed, in which the number of antennas is reduced while
preserving the azimuth resolution. The classic MIMO virtual
λ λ array configuration requires receivers with maximum spacing
R
2 2 m/2 and transmitters with spacing Rm/2 (or vice versa). The
(a) product RT thus scales linearly with aperture, which sets
the azimuth resolution. Spatial compression is achieved by
λZ
using a sparse random-array architecture [28], in which a low
number of transmit and receive elements are placed at ran-
(b) dom over the same aperture Z, achieving similar resolu-
tion as a filled array, but with significantly fewer elements.
FIGURE 11. An illustration of MIMO arrays: (a) the standard array and (b)
The random-array configuration is illustrated in Figure 11.
the random thinned array [32].
Beamforming is applied on the time-domain samples ob­­
tained from the thinned array at the Nyquist rate, and the
r 2
D p (l, l) = e j 2o l (p - 1) x . In this scheme, each receiver transmits
m
azimuths are recovered using CS techniques. Recovery guar-
the output of a few randomly chosen MFs to the fusion center antees and guidelines concerning the choice of the product
so that X p is only partially known. RT and the antenna locations are provid-
In the second scenario, the receivers for- ed. The methods for choosing the antenna
ward Nyquist samples to the fusion center Spatial compression locations using deep networks are investi-
without performing the MF. In this case, the is achieved by using a gated in [89].
samples are written as sparse random-array
architecture, in which a Time and spatial compression
X p = A R RD p A TR S, (47) low number of transmit In all of the previously discussed works,
recovery is performed in the time domain on
and receive elements are
where S is the T # N matrix that contains acquired or reconstructed Nyquist rate sam-
the Nyquist rate samples of each transmit- placed at random over the ples for each antenna. The sub-Nyquist
ted waveform s m (t) . In this scheme, each same aperture Z, achieving MIMO radar (SUMMeR) system, presented
receive antenna randomly acquires a subset similar resolution as in [32], extends the Xampling concept to
of the Nyquist samples and transmits these a filled array, but with MIMO configurations and breaks the link
to the fusion center. In both cases, the significantly fewer between the aperture and the number of
fusion center performs MC before para- antennas, similar to [28]. The concept of
elements.
metric estimation methods are applied to Xampling is applied both in space (antenna
extract i l and o l, such as multiple signal deployment) and in time (sampling scheme)
classification, also known as MUSIC [88]. In these works, sam- to simultaneously reduce the required number of an­­tennas
pling and processing rate reduction are not addressed and samples per receiver, while preserving time and spatial res-
since compression is performed in the digital domain after olution. In particular, targets’ azimuths, ranges, and Dopplers
sampling, and the missing samples are reconstructed before are recovered from compressed samples in both space and time,
recovering the targets’ parameters. Instead, these approaches while keeping the same resolution induced by Nyquist rate sam-
are aimed at reducing the communication overhead between ples obtained from a full virtual array with low computation-
the receivers and the fusion center. al cost.
The SUMMeR system implements a collocated MIMO
Spatial compression radar system with M < T transmit and Q < R receive anten-
Several recent works have considered applying CS to MIMO nas, whose locations are chosen uniformly at random within
radar to reduce the number of antennas or the number of sam- the aperture of the virtual array described previously in this
section, i.e., {p m} mM=- 01 ~ U 60, Z@ and {g q} q = 0 ~ U 60, Z@,
Q-1
ples per receiver without degrading resolution. The problem
of azimuth recovery of targets all in the same range-Doppler respectively, as shown in Figure 11. Note that, in principle, the
antenna locations may be chosen on the ULAs’ grid; however,
this configuration is less robust than range-azimuth ambiguity
TBh TBh and leads to coupling between these parameters in the pres-
ence of noise [32]. An FDMA framework is adopted so that
Bh f Bh f spatial compression, which, in particular reduces the number
of transmit antennas, removes the corresponding transmitting
frequency bands as well. The transmitted signals are illus-
(a) (b)
trated in Figure 12 in the frequency domain. Figure 12(a) and
FIGURE 12. The frequency division multiple access transmissions: (a) (b) shows a standard FDMA transmission for T = 5 and the
standard and (b) spatial compression [32]. resulting signal after spatial compression for M = 3.

54 IEEE Signal Processing Magazine | November 2018 |

 The transmitted pulses, defined in (12), are reflected by the Hardware prototype
targets and collected at the receive antennas. Under the assump- The cognitive SUMMeR prototype [41], [42] extends the
tions described in “Targets’ Assumptions,” the received signal Doppler focusing, Xampling-based prototype [40] to the
xu q (t) at the qth antenna is the sum of time-delayed, scaled rep- MIMO configuration. It simultaneously recovers the targets’
licas of the transmitted signals: delays, Dopplers, and azimuths from sub-Nyquist samples.
T- 1 L More specifically, it implements a receiver with a maximum
R l, mq
xu q (t) = / / a l s m c cc +- vv l c t - c + vl
mm, (48) of eight transmit and ten receive antenna elements. The same
m= 0 l= 1 l
hardware is used for each receive element and serially feeds
where R l, mq is the sum of the distances from the mth transmit- the signals of all R = 10 receivers to the same prototype.
ter and qth receiver to the lth target, which account for the To avoid the use of an overwhelmingly large number of
array geometry. After demodulation to the baseband, the ADCs and bandpass filters for an 8 # 10 array, a cognitive
received signal can be further simplified to transmission is adopted wherein each transmit signal lies in
N b = 8 disjoint, narrow slices over a 15-MHz band. Each sub-
P-1 M-1 L
band is the width of 375 kHz, leading to a total signal band-
/ / / a l h m ^ t - p x - x l h e j2
D
x q (t) = rb mq j l
e j2rf l px, (49)
p=0 m=0 l=1 width of 3 MHz. The transmit subbands, locations were chosen
so that all can be subsampled using a single low-rate ADC
where b mq = ^g q + p mh^ fm mc + 1h, with fm the mth-transmis- without aliasing between them [41]. This allows the reduction
sion-carrier frequency and m the signal wavelength. The goal of the number of samplers. The signal is subsampled at 7.5 MHz,
is to estimate the targets’ ranges, azimuths, and velocities, i.e., whereas a noncognitive signal would have occupied the entire
to estimate x l, j l, and f lD from low-rate samples of x q (t), and 15-MHz spectrum requiring a Nyquist sampling rate of 30 MHz.
small numbers m and Q of the antennas. Therefore, the use of cognitive transmission enables spec-
Similar to the Xampling processing in [30], SUMMeR con- tral sampling reduction by a factor of four for each channel.
siders the Fourier coefficients of the received signal x qp (t) The effective signal bandwidth is reduced by a factor of five
at the qth antenna. To jointly recover the targets’ ranges, azimuths, (= 15/3 MHz), respectively, for each channel.
and Doppler frequencies, the concept of Doppler focusing from The system may be configured to operate in various array
[30] (see “Doppler Focusing”) is applied to the MIMO setting, configurations simulating different numbers and locations of
and the CS algorithms are extended to simultaneous matrix the antennas. The hardware switches off the inactive channels
recovery [32]. The minimal number of channels required for and does not sample any data over the corresponding ADCs.
the recovery of L targets’ parameters in noiseless settings is This governs the spatial compression by reducing the number
MQ $ 2L, with a minimal number of MK $ 2L samples per of receivers and transmitters. In its baseline configuration, the sys-
receiver and P $ 2L pulses per transmitter [32]. The SUM- tem uses only half of the antennas with respect to the full vir-
MeR system has been implemented in hardware, as described tual array, i.e., M = 4 transmitters and Q = 5 receivers. Fig­­ure 13
in the following section. shows the sub-Nyquist MIMO prototype, user interface, and

Analog Preprocessing Digital Rx

Radar Display Waveform Generator Ir , 1 (t )
I-Channel ADC
Ir [t ] App Ir , 2 (t )
Laptop/PC DAC ADC
:×8 :×8
Ir [n ] Ir , 1 . . . 8 [n ] Laptop/PC
User LAN (Serialized)
FPGA Ir , 8 (t )
Interface ADC
Qr [n ] Qr , 1 (t )
ADC LAN Radar
sr [n ] = Ir [n ] + jQr [n ] Qr [t ] Qr , 1 . . . 8 [n ] FPGA
Qr , 2 (t )
(Serialized) DAC ADC (Serialized) Display
Q-Channel
App :×8 :×8
Qr , 8 (t )
ADC

User-Control Interface Waveform Generator Analog Preprocessing Digital Rx

FIGURE 13. The sub-Nyquist MIMO prototype and user interface. The analog preprocessor module consists of two cards mounted on opposite sides of a common
chassis. The inset shows the simplified block diagram of the system. The subscript r represents the received signal samples for the r th receiver. Wherever ap-
plicable, the second subscript corresponds to a particular transmitter. The square brackets (parentheses) are used for digital (analog) signals [41], [42].

IEEE Signal Processing Magazine | November 2018 | 55

 ful detection (green circle) occurs when the estimated target is
True Locations within one range cell, one azimuth bin, and one Doppler bin
Correct Detections of the ground truth (blue circle). More experiments in [41] and
[42] demonstrate that the prototype performance is robust, with

Velocity (m/s)
40 SNRs dropping to as low as –10 dB, and the time and spatial
0 resolution are preserved by simulating couples of close targets
–40 in range, Doppler, and azimuth.
10
)
0 3 6 9 12 15 10 0 –10 0 (km Conclusions and future challenges
Range (km) y (km) x
In this article, we reviewed several compressed radar systems
(a) (b) that aim to reduce complexity while preserving parameter
­resolution. Throughout this article, we considered different
FIGURE 14. The SUMMeR prototype recovery performance: (a) the plan popular radar systems, including pulse-Doppler and step-fre-
position indicator (PPI) display. The origin is the location of the radar. The quency radars as well as MIMO configurations. In particular,
red dot indicates the north direction relative to the radar. The positive/
negative distances along the horizontal axis correspond to the east/west
we showed that temporal, spectral, and spatial compression
direction of the radar. Similarly, the positive/negative distances along the can be implemented without decreasing Doppler, range, and
vertical axis correspond to the north/south direction of the radar. The es- azimuth resolution. To recover these parameters for L targets,
timated targets are plotted over the ground truth. (b) The range-azimuth- the minimum number of required samples per pulse, the mini-
Doppler map for the same targets. The lower axes represent the Cartesian mum number of pulses, and the minimum number of channels
coordinates of the polar representation of the PPI plots from (a). The
vertical axis represents the Doppler spectrum [32].
are each equal to 2L. These are determined by the actual num-
ber of DoF of the parameter estimation problem, governed by
L, rather than a function of design parameters, such as signal
radar display. The inset graph depicts the signal flow through a bandwidth, CPI, or aperture. This is essential since the latter
simplified block diagram. determine range, Doppler, and azimuth resolution and are
The experimental process consists of the following steps. increased for higher performance. By breaking the traditional
The simulated radar scenario is stored in a custom-designed links between the sampling rate, number of pulses, and anten-
waveform generator. The scenario includes pulse-transmis- nas on the one hand and parameter estimation on the other
sion modeling, accurate power loss due to wave propagation hand, increased performance may be achieved without increasing
in a realistic medium, and in­­teraction of a transmit signal sampling and processing rates.
with the target. A large variety of scenarios, consisting of An advantage of the Xampling system is that traditional
different targets’ parameters, i.e., delays, Doppler frequen- radar-processing algorithms can be easily adapted and applied
cies, and amplitudes, and array configu- directly to the sub-Nyquist samples. For exam-
rations, i.e., the number of transmitters and ple, clutter-cancellation techniques have
receivers and antenna locations, may be The transmit subbands, been implemented on the Xampling radar
ex­­amined using the prototype. The waveform locations were chosen so prototypes. These significantly enhance
ge­­nerator board then produces an analog that all can be subsampled the performance of compressed radars with-
signal corresponding to the synthesized using a single low-rate out requiring the reconstruction of Nyquist
radar environment, which is amplified and ADC without aliasing rate samples. In addition, while CS-based
routed to the MIMO radar-receiver board. methods traditionally do not perform well
The prototype samples and processes the
between them. in the presence of large noise, since they
signal in real time. The physical array aper- inherently reduce SNR, Doppler focusing,
ture and simulated target response correspond to an X-band applied to samples obtained using Xampling, enjoys an SNR
^ fc = 10 GHz h radar. improvement that scales linearly with the number of pulses,
Figure 14 presents some recovery results from the proto- obtaining good detection at low SNRs.
type. In the experiment, P = 10 pulses were transmitted at a An essential part of the approach adopted in this survey is
uniform PRF of 100 nHz. The received signal corresponding the relation between the theoretical algorithms and p­ ractical
to the echoes from L = 10 targets, placed at arbitrary ranges hardware implementation, demonstrating real-time target
with azimuths and with arbitrary velocities, was injected into detection from compressed samples in the fast and slow time
the transmit waveform generator. In the experiment, when the domains, as well as in space. The prototypes presented here
angular spacing (in terms of the sine of azimuth) between any were built from off-the-shelf components, paving the way to
two targets was greater than 0.025 and the signal SNR = -8 dB, enable commercial, compressed radar systems. To this end,
the recovery performance of the compressed configuration in such hardware prototypes should be further extended to imple-
time and in space was equivalent to that of a full array, i.e., ment radar transmit and receive systems and deploy them to
with eight transmitters and ten receivers. The figure shows the be tested on real data. This would permit assessing their per-
obtained plan position indicator plot and range-azimuth-Dop- formance in real-world conditions, including different types of
pler maps for both true and recovered targets. Here, a success- noise, clutter, and interference.

56 IEEE Signal Processing Magazine | November 2018 |

Authors [11] J. C. Curlander and R. N. McDonough, Synthetic Aperture Radar. New York:
Wiley, 1991.
Deborah Cohen ([email protected]) received her
[12] Y. C. Eldar and G. Kutyniok, Compressed Sensing: Theory and Applications.
B.S. degree in electrical engineering (summa cum laude) in Cambridge, U.K.: Cambridge Univ. Press, 2012.
2010 and her Ph.D. degree in electrical engineering in 2016, [13] E. J. Candès, J. Romberg, and T. Tao, “Robust uncertainty principles: Exact
both from the Technion–Israel Institute of Technology, signal reconstruction from highly incomplete frequency information,” IEEE Trans.
Inf. Theory, vol. 52, no. 2, pp. 489–509, 2006.
Haifa. She has been granted several awards, including the [14] D. L. Donoho, “Compressed sensing,” IEEE Trans. Inf. Theory, vol. 52, no. 4,
Meyer Foundation Excellence Award, the David and Tova pp. 1289–1306, 2006.
Freud and Ruth Freud-Brendel Memorial Scholarship, the [15] M. A. Herman and T. Strohmer, “High-resolution radar via compressed sens-
ing,” IEEE Trans. Signal Process., vol. 57, no. 6, pp. 2275–2284, 2009.
Sandor Szego Award, the Vivian Konigsberg Award, and the
[16] Y. Chi, R. Calderbank, and A. Pezeshki, “Golay complementary waveforms
Muriel and David Jacknow Award for Excellence in for sparse delay-Doppler radar imaging,” in Proc. IEEE Int. Workshop Computational
Teaching. Since 2014, she has been an Azrieli fellow. She is Advances Multi-Sensor Adaptive Processing, 2009, pp. 177–180.
currently a research scientist with Google Israel. Her research [17] S. Shah, Y. Yu, and A. Petropulu, “Step-frequency radar with compressive
sampling (SFR-CS),” in Proc. IEEE Int. Conf. Acoustics, Speech and Signal
interests include theoretical aspects of signal processing, Processing, 2010, pp. 1686–1689.
compressed sensing, reinforcement learning, and machine [18] T. Strohmer and H. Wang, “Sparse MIMO radar with random sensor arrays and
learning for dialogues. Kerdock codes,” in Proc. IEEE Int. Conf. Sampling Theory and Applications,
2013, pp. 517–520.
Yonina C. Eldar ([email protected]) is a profes-
[19] T. Strohmer and B. Friedlander, “Analysis of sparse MIMO radar,” Appl.
sor in the Department of Electrical Engineering at the Comput. Harmon. Anal., vol. 37, no. 3, pp. 361–388, 2014.
Technion–Israel Institute of Technology, Haifa, where she [20] W. U. Bajwa, K. Gedalyahu, and Y. C. Eldar, “Identification of parametric
holds the Edwards chair in engineering. She is also an underspread linear systems and super-resolution radar,” IEEE Trans. Signal
Process., vol. 59, no. 6, pp. 2548–2561, 2011.
adjunct professor at Duke University, Durham, North Ca­­
[21] O. Teke, A. C. Gurbuz, and O. Arikan, “A robust compressive sensing-based
rolina, and a research affiliate with the Research Laboratory technique for reconstruction of sparse radar scenes,” Digit. Signal Process., vol. 27,
of Electronics at the Massachusetts Institute of Technology, pp. 23–32, Apr. 2014.

Cambridge, and she was a visiting professor at Stanford [22] J. H. Ender, “On compressive sensing applied to radar,” Signal Process., vol.
90, no. 5, pp. 1402–1414, 2010.
University, California. She is a member of the Israel Ac­­ [23] Y. Yu, A. P. Petropulu, and H. V. Poor, “MIMO radar using compressive sam-
ademy of Sciences and Humanities and a Fellow of the IEEE pling,” IEEE J. Sel. Topics Signal Process., vol. 4, no. 1, pp. 146–163, 2010.
and the European Association for Signal Processing. She has [24] D. S. Kalogerias and A. P. Petropulu, “Matrix completion in colocated MIMO
radar: Recoverability, bounds & theoretical guarantees,” IEEE Trans. Signal
received many awards for excellence in research and teach- Process., vol. 62, no. 2, pp. 309–321, 2014.
ing, including the IEEE Signal Processing Society Technical [25] S. Sun, W. U. Bajwa, and A. P. Petropulu, “MIMO-MC radar: A MIMO radar
Achievement Award, the IEEE/Aerospace and Electronic approach based on matrix completion,” IEEE Trans. Aerosp. Electron. Syst., vol.
51, no. 3, pp. 1839–1852, 2015.
Systems Society Fred Nathanson Memorial Radar Award,
[26] E. Ertin, L. C. Potter, and R. L. Moses, “Sparse target recovery performance of
the IEEE Kiyo Tomiyasu Award, the Michael Bruno multi-frequency chirp waveforms,” in Proc. IEEE European Signal Processing
Memorial Award from the Rothschild Foundation, the Conf., 2011, pp. 446–450.
Weizmann Prize for Exact Sciences, and the Wolf Foundation [27] C. Liu, F. Xi, S. Chen, Y. D. Zhang, and Z. Liu, “Pulse-Doppler signal process-
ing with quadrature compressive sampling,” IEEE Trans. Aerosp. Electron. Syst.,
Krill Prize for Excellence in Scientific Research. She is the vol. 51, no. 2, pp. 1217–1230, 2015.
editor-in-chief of Foundations and Trends in Signal [28] M. Rossi, A. M. Haimovich, and Y. C. Eldar, “Spatial compressive sensing for
Processing and serves the IEEE on several technical and MIMO radar,” IEEE Trans. Signal Process., vol. 62, no. 2, pp. 419–430, 2014.

award committees. [29] M. Mishali, Y. C. Eldar, O. Dounaevsky, and E. Shoshan, “Xampling: Analog to
digital at sub-Nyquist rates,” IET Circuits, Devices Syst., vol. 5, no. 1, pp. 8–20, 2011.
[30] O. Bar-Ilan and Y. C. Eldar, “Sub-Nyquist radar via Doppler focusing,” IEEE
Trans. Signal Process., vol. 62, no. 7, pp. 1796–1811, 2014.
References
[1] X. P. Masbernat, M. G. Amin, F. Ahmad, and C. Ioana, “An MIMO-MTI [31] D. Cohen, A. Dikopoltsev, R. Ifraimov, and Y. C. Eldar, “Towards sub-Nyquist
approach for through-the-wall radar imaging applications,” in Proc. IEEE Int. cognitive radar,” in Proc. IEEE Radar Conf., 2016. doi: 10.1109/RADAR
Waveform Diversity and Design Conf., 2010, pp. 188–192. .2016.7485122.
[2] D. J. Daniels, Ground Penetrating Radar. London, U.K.: Institution of [32] D. Cohen, D. Cohen, Y. C. Eldar, and A. M. Haimovich, “SUMMeR: Sub-Nyquist
Engineering and Technology, 2004. MIMO radar,” IEEE Trans. Signal Process., vol. 66, no. 16, pp. 4315–4330, 2018.
[3] K. Schuler, M. Younis, R. Lenz, and W. Wiesbeck, “Array design for automo- [33] L. C. Potter, E. Ertin, J. T. Parker, and M. Cetin, “Sparsity and compressed
tive digital beamforming radar system,” in Proc. IEEE Int. Radar Conf., 2005, pp. sensing in radar imaging,” Proc. IEEE, vol. 98, no. 6, pp. 1006–1020, 2010.
435–440. [34] M. Cetin, I. Stojanovic, N. O. Onhon, K. R. Varshney, S. Samadi, W. C. Karl,
[4] V. Bringi and V. Chandrasekar, Polarimetric Doppler Weather Radar: and A. S. Willsky, “Sparsity-driven synthetic aperture radar imaging: Reconstruction,
Principles and Applications. Cambridge, U.K.: Cambridge Univ. Press, 2001. autofocusing, moving targets, and compressed sensing,” IEEE Signal Process.
Mag., vol. 31, no. 4, pp. 27–40, 2014.
[5] M. Skolnik, Radar Handbook. New York: McGraw-Hill, 1970.
[35] L. Zhao, L. Wang, L. Yang, A. M. Zoubir, and G. Bi, “The race to improve
[6] P. Z. Peebles, Radar Principles. Hoboken, NJ: Wiley, 2007. radar imagery: An overview of recent progress in statistical sparsity-based tech-
[7] M. A. Richards, Fundamentals of Radar Signal Processing. New York: niques,” IEEE Signal Process. Mag., vol. 33, no. 6, pp. 85–102, 2016.
McGraw-Hill, 2014. [36] H. Griffiths, L. Cohen, S. Watts, E. Mokole, C. Baker, M. Wicks, and S.
[8] E. Fishler, A. Haimovich, R. Blum, D. Chizhik, L. Cimini, and R. Valenzuela, Blunt, “Radar spectrum engineering and management: Technical and regulatory
“MIMO radar: An idea whose time has come,” in Proc. IEEE Radar Conf., 2004, issues,” Proc. IEEE, vol. 103, no. 1, pp. 85–102, 2015.
pp. 71–78. [37] M. P. Fitz, T. R. Halford, I. Hossain, and S. W. Enserink, “Towards simultane-
[9] J. Li and P. Stoica, MIMO Radar Signal Processing. Piscataway, NJ: IEEE ous radar and spectral sensing,” in Proc. IEEE Int. Symp. Dynamic Spectrum
Press, 2009. Access Networks, 2014, pp. 15–19.
[10] Y. C. Eldar, Sampling Theory: Beyond Bandlimited Systems. Cambridge, [38] J. Bernhard, J. Reed, J. M. Park, A. Clegg, A. Weisshaar, and A. Abouzeid.
U.K.: Cambridge Univ. Press, 2015. (2010). Final report of the National Science Foundation workshop on enhancing

IEEE Signal Processing Magazine | November 2018 | 57

 access to the radio spectrum (EARS). Nat. Sci. Foundation. Arlington. [Online]. [63] K. Gedalyahu, R. Tur, and Y. C. Eldar, “Multichannel sampling of pulse
Available: https://www.nsf.gov/mps/ast/nsf_ears_workshop_2010_final_ streams at the rate of innovation,” IEEE Trans. Signal Process., vol. 59, no. 4, pp.
report.pdf 1491–1504, 2011.
[39] D. Cohen, K. V. Mishra, and Y. C. Eldar, “Spectrum sharing radar: Co­­ [64] R. Tur, Y. C. Eldar, and Z. Friedman, “Innovation rate sampling of pulse
existence via Xampling,” IEEE Trans. Aerosp. Electron. Syst., vol. 54, pp. 1279– streams with application to ultrasound imaging,” IEEE Trans. Signal Process., vol.
1296, 2018. 59, no. 4, pp. 1827–1842, 2011.
[40] E. Baransky, G. Itzhak, I. Shmuel, N. Wagner, E. Shoshan, and Y. C. Eldar, [65] D. Cohen and Y. C. Eldar, “Reduced time-on-target in pulse Doppler radar:
“Sub-Nyquist radar prototype: Hardware and algorithms,” IEEE Trans. Aerosp. Slow time domain compressed sensing,” in Proc. IEEE Radar Conf., 2016. doi:
Electron. Syst., vol. 50, no. 1, pp. 809–822, 2014. 10.1109/RADAR.2016.7485243.
[41] K. V. Mishra, E. Shoshan, M. Namer, M. Meltsin, D. Cohen, R. Madmoni, S. [66] T. Wimalajeewa, Y. C. Eldar, and P. K. Varshney. (2013, 11 Nov.). Recovery of
Dror, R. Ifraimov, et al., “Cognitive sub-Nyquist hardware prototype of a collocated sparse matrices via matrix sketching. arXiv. [Online]. Available: http://arxiv.org/
MIMO radar,” in Proc. Compressed Sensing Theory and its Applications to Radar, abs/1311.2448
Sonar and Remote Sensing (CoSeRa), 2016, pp. 56–60. [67] R. J. Doviak and D. S. Zrnic, Doppler Radar & Weather Observations. New
[42] D. Cohen, K. V. Mishra, D. Cohen, E. Ronen, Y. Grimovich, M. Namer, M. York: Academic, 2014.
Meltsin, and Y. C. Eldar, “Cognitive sub-Nyquist MIMO radar prototype with [68] A. Ferrari, G. Alengrin, and C. Theys, “Doppler ambiguity resolution using
Doppler processing,” in IEEE Radar Conf., 2016, pp. 1179–1184. staggered PRF with a new chirp sweep-rate estimation algorithm,” IEEE Proc.
[43] J. Yoo, C. Turnes, E. B. Nakamura, C. K. Le, S. Becker, E. A. Sovero, M. B. Radar, Sonar Navigation, vol. 142, no. 4, pp. 191–194, 1995.
Wakin, M. C. Grant, et al., “A compressed sensing parameter extraction platform [69] V. Venkatesh, L. Li, M. McLinden, G. Heymsfield, and M. Coon, “A frequen-
for radar pulse signal acquisition,” IEEE J. Emerg. Sel. Topics Circuits Syst., vol. 2, cy diversity pulse-pair algorithm for extending Doppler radar velocity Nyquist
no. 3, pp. 626–638, 2012. range,” in Proc. IEEE Radar Conf., 2016, pp. 1–6.
[44] E. Fishler, A. Haimovich, R. S. Blum, and L. J. Cimini, “Spatial diversity in [70] A. Ludloff and M. Minker, “Reliability of velocity measurement by MTD
radars-models and detection performance,” IEEE Trans. Signal Process., vol. 54, radar,” IEEE Trans. Aerosp. Electron. Syst., vol. AES-21, no. 4, pp. 522–528, 1985.
pp. 823–838, Mar. 2006.
[71] G. Trunk and S. Brockett, “Range and velocity ambiguity resolution,” in Proc.
[45] C. Cook, Radar Signals: An Introduction to Theory and Application. IEEE Nat. Radar Conf., 1993, pp. 146–149.
Amsterdam, The Netherlands: Elsevier, 2012.
[72] X. Liu, D. Cohen, T. Huang, Y. Liu, G. Winerich, L. Shani, and Y. C. Eldar,
[46] J. Li and P. Stoica, “MIMO radar with collocated antennas,” IEEE Signal “Unambiguous delay-Doppler recovery from phased coded pulses,” 2016, to
Process. Mag., vol. 24, no. 5, pp. 106–114, 2007. be published.
[47] A. M. Haimovich, R. S. Blum, and L. J. Cimini, “MIMO radar with widely [73] Q. Cao, G. Zhang, R. D. Palmer, and L. Lei, “Detection and mitigation of sec-
separated antennas,” IEEE Signal Process. Mag., vol. 25, no. 1, pp. 116–129, ond-trip echo in polarimetric weather radar employing random phase coding,” IEEE
2008. Trans. Geosci. Remote Sens., vol. 50, no. 4, pp. 1240–1253, 2012.
[48] D. Bliss and K. Forsythe, “Multiple-input multiple-output (MIMO) radar and [74] S. Haykin, “Cognitive radar: A way of the future,” IEEE Signal Process. Mag.,
imaging: Degrees of freedom and resolution,” in Proc. IEEE Asilomar Conf. vol. 23, no. 1, pp. 30–40, 2006.
Signals, Systems and Computers, 2003, pp. 54–59.
[75] K. V. Mishra, D. Cohen, S. Tsiper, S. Stein, E. Shoshan, M. Namer, M.
[49] C.-Y. Chen, “Signal processing algorithms for MIMO radar,” Ph.D. disserta- Meltsin, R. Madmoni, E. Ronen, Y. Grimovich, and Y. C. Eldar, “Xampling-
tion, California Inst. of Technol., Pasadena, 2009. enabled coexistence in spectrally crowded environments,” in Proc. IEEE Int. Conf.
[50] P. Vaidyanathan, P. Pal, and C.-Y. Chen, “MIMO radar with broadband wave- Acoustics, Speech and Signal Processing, 2017, pp. 6580–6581.
forms: Smearing filter banks and 2D virtual arrays,” in Proc. IEEE Asilomar Conf. [76] G. M. Jacyna, B. Fell, and D. McLemore, “A high-level overview of fundamen-
Signals, Systems and Computers, , pp. 188–192. tal limits studies for the DARPA SSPARC program,” in Proc. IEEE Radar Conf.,
[51] Y. C. Pati, R. Rezaiifar, and P. S. Krishnaprasad, “Orthogonal matching pur- 2016. doi: 10.1109/RADAR.2016.7485100.
suit: Recursive function approximation with applications to wavelet decompo-
[77] P. Stinco, M. S. Greco, and F. Gini, “Spectrum sensing and sharing for cogni-
sition,” in Proc. IEEE Asilomar Conf. Signals, Systems and Computers, 1993,
tive radars,” IET Radar, Sonar Navigation, vol. 10, no. 3, pp. 595–602, 2016.
pp. 40–44.
[78] N. Nartasilpa, D. Tuninetti, N. Devroye, and D. Erricolo, “Let’s share
[52] G. M. Davis, S. G. Mallat, and Z. Zhang, “Adaptive time-frequency decompo- CommRad: Effect of radar interference on an uncoded data communication system,”
sitions with matching pursuit,” SPIE, vol. 2242, pp. 402–414, Mar. 1994. doi: in Proc. IEEE Radar Conf., 2016. doi: 10.1109/RADAR.2016.7485064.
10.1117/12.170041.
[79] P. Kumari, N. Gonzalez-Prelcic, and R. W. Heath, “Investigating the IEEE
[53] T. Blumensath and M. Davies, “Gradient pursuits,” IEEE Trans. Signal 802.11ad standard for millimeter wave automotive radar,” in Proc. Vehicular
Process., vol. 56, pp. 2370–2382, Jun. 2008. Technol. Conf., 2015. doi: 10.1109/VTCFall.2015.7390996.
[54] Y. C. Eldar, R. Levi, and A. Cohen, “Clutter removal in sub-Nyquist radar,” [80] J. Mitola and C. Q. Maguire, Jr., “Cognitive radio: Making software radios
IEEE Signal Process. Lett., vol. 22, no. 2, pp. 177–181, 2015. more personal,” IEEE Personal Commun., vol. 6, no. 4, pp. 13–18, 1999.
[55] R. Baraniuk and P. Steeghs, “Compressive radar imaging,” in Proc. IEEE [81] D. Cohen, S. Tsiper, and Y. C. Eldar, “Analog-to-digital cognitive radio: Sampling,
Radar Conf., 2007, pp. 128–133. detection, and hardware,” IEEE Signal Process. Mag., vol. 35, no. 1, pp. 137–166, 2018.
[56] W. O. Alltop, “Complex sequences with low periodic correlations,” IEEE [82] N. Vaswani and W. Lu, “Modified-CS: Modifying compressive sensing for
Trans. Inf. Theory, vol. 26, no. 3, pp. 350–354, 1980. problems with partially known support,” IEEE Trans. Signal Process., vol. 58, no.
[57] Z. Yang, C. Zhang, and L. Xie, “Robustly stable signal recovery in compressed 9, pp. 4595–4607, 2010.
sensing with structured matrix perturbation,” IEEE Trans. Signal Process., vol. 60, [83] S. M. Kay, Fundamentals of Statistical Signal Processing, Volume II:
no. 9, pp. 4658–4671, 2012. Detection Theory. Englewood Cliffs, NJ: Prentice Hall, 1998.
[58] M. A. Hadi, S. Alshebeili, K. Jamil, and F. E. A. El-Samie, “Compressive [84] J. Huang, T. Zhang, and D. Metaxas, “Learning with structured sparsity,” J.
sensing applied to radar systems: An overview,” Signal, Image Video Process., vol. Mach. Learn. Res., vol. 12, pp. 3371–3412, Nov. 2011.
9, no. 1, pp. 25–39, 2015.
[85] M. Mishali and Y. C. Eldar, “From theory to practice: Sub-Nyquist sampling of
[59] J. A. Tropp, J. N. Laska, M. F. Duarte, J. K. Romberg, and R. G. Baraniuk, sparse wideband analog signals,” IEEE J. Sel. Topics Signal Process., vol. 4, no. 2,
“Beyond Nyquist: Efficient sampling of sparse bandlimited signals,” IEEE Trans. pp. 375–391, 2010.
Inf. Theory, vol. 56, pp. 520–544, Jan. 2010.
[86] R. Gold, “Optimal binary sequences for spread spectrum multiplexing
[60] J. Yoo, S. Becker, M. Monge, M. Loh, E. Candes, and A. Emami-Neyestanak, (Corresp.),” IEEE Trans. Inf. Theory, vol. 13, no. 4, pp. 619–621, 1967.
“Design and implementation of a fully integrated compressed-sensing signal
­acquisition system,” in Proc. IEEE Int. Conf. Acoustics, Speech and Signal [87] C.-Y. Chen and P. Vaidyanathan, “Compressed sensing in MIMO radar,” in
Processing, 2012, pp. 5325–5328. Proc. IEEE Asilomar Conf. Signals, Systems and Computers, 2008, pp. 41–44.
[61] H. Liu, A. Ghafoor, and P. H. Stockmann, “A new quadrature sampling and [88] R. O. Schmidt, “Multiple emitter location and signal parameter estimation,”
processing approach,” IEEE Trans. Aerosp. Electron. Syst., vol. 25, no. 5, pp. 733– IEEE Trans. Antennas Propag., vol. 34, no. 4, pp. 276–280, 1986.
748, 1989. [89] A. M. Elbir, K. V. Mishra, and Y. C. Eldar. (2018). Cognitive radar antenna selec-
[62] X. Song, S. Zhou, and P. Willett, “The role of the ambiguity function in com- tion via deep learning. arXiv. [Online]. Available: https://arxiv.org/abs/1802.09736
pressed sensing radar,” in Proc. IEEE Int. Conf. Acoustics, Speech and Signal
Processing, 2010, pp. 2758–2761.  SP

58 IEEE Signal Processing Magazine | November 2018 |