1

Design and Analysis of Compressive Sensing
Radar Detectors
Laura Anitori,

Student Member, IEEE, Arian Maleki, Member, IEEE, Matern Otten,
Richard Baraniuk, Fellow, IEEE, and Peter Hoogeboom
Abstract
We consider the problem of target detection from a set of Compressive Sensing (CS) radar measurements
corrupted by additive white Gaussian noise. We propose two novel architectures and compare their
performance by means of Receiver Operating Characteristic (ROC) curves. Using asymptotic arguments
and the Complex Approximate Message Passing (CAMP) algorithm, we characterize the statistics of the

1
-norm reconstruction error and derive closed form expressions for both the detection and false alarm
probabilities of both schemes. Of the two architectures, we demonstrate that the best one consists of
a reconstruction stage based on CAMP followed by a detector. This architecture, which outperforms
the
1
-based detector in the ideal case of known background noise, can also be made fully adaptive by
combining it with a conventional Constant False Alarm Rate (CFAR) processor. Using the state evolution
framework of CAMP, we also derive Signal to Noise Ratio (SNR) maps that, together with the ROC
curves, can be used to design a CS-based CFAR radar detector. Our theoretical findings are confirmed
by means of both Monte Carlo simulations and experimental results.
Index Terms
Radar, Compressive Sensing, Detection Probability, False Alarm Probability, Constant False Alarm
Rate (CFAR), Complex Approximate Message Passing (CAMP).
EDICS: SAM-RADR, SSP-DETC.
L. Anitori, M. Otten, and P. Hoogeboom are with TNO, Oude Waalsdorperweg 63, 2597 AK, The Hague, The Netherlands.
E-mail: {laura.anitori, matern.otten, peter.hoogeboom}@tno.nl.
L. Anitori, and P. Hoogeboom are also with the Technical University of Delft, Stevinweg 1, 2628 CN Delft, The Netherlands.
E-mail: {l.anitori, p.hoogeboom}@tudelft.nl.
A. Maleki, and R. Baraniuk are with the Dept. of Electrical and Computer Engineering, Rice University, Houston, 77005 TX,
USA. E-mail: {arian.maleki, richb}@rice.edu.
Part of the material presented in this paper in Section VII was presented at the 2012 IEEE Radar Conference.
March 8, 2012 DRAFT
2
I. INTRODUCTION
High resolution radar detection and classification of targets requires transmission of wide bandwidth
during a short observation time. Wide instantaneous bandwidth necessitates wideband receivers and fast
Analog to Digital Converters (ADC) that drive up the cost of a system. In applications using interleaving
of radar modes in time or space (antenna aperture), multi-function operation often leads to conflicting
requirements on sampling rates in both the time and space domains. Measurement time may also be a
constraint in real-time applications.
Recently, Compressive Sensing (CS) has been proposed to alleviate these problems [1]–[12]. CS
exploits the sparsity or compressibility of a signal to reduce both the sampling rate (while keeping
the resolution fixed) and the amount of data generated. In many radar applications, the prerequisite on
the signal sparsity is often met, as the number of targets is typically much smaller than the number of
resolution cells in the illuminated area or volume.
Classical radar architectures (without CS) use well-established signal processing algorithms and detec-
tion schemes, such as Matched Filtering (MF) and Constant False Alarm Rate (CFAR) processors. CS
instead relies on a nonlinear reconstruction scheme such as
1
-minimization for reconstructing the target
scene. Most of the existing CS recovery algorithms have a threshold or a regularization parameter, that
affects the quality of the reconstruction result; for the best performance it is essential to optimally tune
this parameter [13]. A similar problem is encountered in classical radar detectors, where the detection
threshold adapts to the environment using a CFAR scheme. These detectors first estimate the background
noise and clutter surrounding the Cell Under Test (CUT) and then calculate the threshold to preserve a
constant false alarm rate even in changing environments. Several CFAR schemes have been designed for
different types of noise, clutter, and target environments [14], [15].
The design of CFAR-like schemes for CS radar systems has not been addressed so far in the literature,
mostly due to the unknown relations between the false alarm probability/noise statistics and the parameters
of the recovery algorithm. Moreover, the main focus of CS has been on the reconstruction Mean
Square Error (MSE), while in radar, the primary performance parameters are the detection and false
alarm probabilities. Furthermore, most of the theoretical results provide only sufficient conditions with
unknown/large constants that are not useful for applications [16]–[19]. These issues have been recently
addressed in a series of papers [20]–[23]. In these papers a novel CS recovery algorithm has been
proposed, called Complex Approximate Message Passing (CAMP). The statistical properties of CAMP
enable accurate calculations of different performance criteria such as the false alarm and detection
March 8, 2012 DRAFT
3
probabilities.
In this paper, we use the features of the CAMP algorithm to design a fully adaptive CS-CFAR radar
detector. Toward this goal we address the following questions. How can the detection probability be
optimized against false alarm rate? How can the false alarm rate be controlled adaptively against unknown
noise and clutter? How can a CS-radar be designed; in particular, what amount of undersampling is
acceptable, and at what cost in terms of power?
To this end, in Section II we begin with a summary of the main features of the CAMP algorithm,
originally presented in [23] for the CS recovery of complex signals and previously in [20], [21], [24]
for the real signal case. In Section III we propose two novel CS radar detection architectures based on
CAMP. We predict the performance of these two detectors theoretically by characterizing the statistical
properties of the recovered signal. Although the theoretical results on CAMP assume sensing matrices with
independent identically distributed (i.i.d.) random entries, both our simulations and empirical results show
that the general trend of our findings also applies to randomly subsampled (or partial) Fourier sensing
matrices. Such matrices are particularly important for CS-radar range compression, where subsampling
in frequency is a power-efficient way of applying CS [9], as opposed to time undersampling.
From the analytical, simulation, and experimental results presented respectively in Sections III, VI,
and VII, it is clear that the best architecture consists of a reconstruction stage based on CAMP followed
by a detector. For this architecture, in Section IV we propose an adaptive version of CAMP followed
by a CFAR processor, resulting in a fully adaptive detector. In Section VIII we derive Signal-to-Noise
Ratio (SNR) plots in which the reconstruction SNR, which is input to the detection stage, is plotted
against the undersampling factor (δ) and the relative signal sparsity (ρ). Given a maximum number of
expected targets in the scene, such graphs can be used to evaluate how transmitted power can be traded
for undersampling and thus serve as a guide for the CS radar designer. In Section IX we draw some
conclusions from our theory and experiments.
II. COMPLEX APPROXIMATE MESSAGE PASSING (CAMP)
A. Heuristic explanation of the CAMP algorithm
Consider the problem of recovering a sparse vector x
o
∈ C
N
from a noisy undersampled set of linear
measurements y ∈ C
n
y = Ax
o
+n, (1)
where A ∈ C
n×N
is called the sensing matrix and n is i.i.d. complex Gaussian noise with variance σ
2
.
In radar systems, y corresponds to the sampled received signal. Each non-zero element of the vector x
o
March 8, 2012 DRAFT
4
corresponds to the (complex) Radar Cross Section (RCS) of a target and may also include propagation
and other factors normally associated with the radar equation [25]. In the remainder of this paper, [x
o,i
[
2
represents the received power from a target at position i. Hence, x
o,i
is zero except at the target locations.
If there are k ¸N targets, then x
o
is sparse.
It has been proved that, under certain conditions on the measurement matrix and the sparsity of x
o
,
it is possible to recover x
o
accurately [16], [26]. One of the most successful recovery algorithms is

1
-regularized least squares, also known as the LASSO or Basis Pursuit Denoising (BPDN) [27]–[29],
given by
´ x(λ) = arg min
x
1
2
|y −Ax|
2
2
+ λ|x|
1
, (2)
where |x|
1


N
i=1
_
(x
R
i
)
2
+ (x
I
i
)
2
, with x
R
i
and x
I
i
denoting the real and imaginary parts of x
i
,
respectively. λ is called the regularization parameter. The cost function in (2) is convex and can be solved
by standard techniques such as interior point or homotopy methods [30], [31]. However, these approaches
are computationally expensive, and therefore researchers have considered several iterative algorithms
with inexpensive per-iteration computations. For more information on these algorithms, see [23] and the
references therein. These methods rely on the fact that the optimization problem arg min
x
1
2
|u −x|
2
2
+
λ|x|
1
has a closed form solution η(u; λ) ([u[ − λ)e
ju
, where η(; λ) is called the complex soft
thresholding function.
In this paper, we focus on a specific algorithm called Complex Approximate Message Passing (CAMP)
[23]. This algorithm has several appealing properties for radar applications that will become clear as we
proceed. The CAMP algorithm is given in Algorithm I, where ¸¸ denotes the average of a vector, η
I
and η
R
are the imaginary and real parts of the complex soft thresholding function,
∂η
R
∂xR
is the partial
derivative of η
R
with respect to the real part of the input,
∂η
I
∂xI
is the partial derivative of η
I
with respect to
the imaginary part of the input, 1 is the indicator function, and maxiter is the (user specified) maximum
number of iterations. δ =
n
N
is called the compression factor, i.e., the number of CS measurements
divided by the number of samples to be recovered. We first explain each component of CAMP.
(i) ´ x
t
is an estimate of x
o
at iteration t. If the parameter τ is tuned properly, ´ x
t
→ ´ x(λ) as t → ∞.
The tuning of τ in terms of λ is given later in (5) and will be explained in detail in Section II-B.
(ii) ¯ x
t
is a non-sparse, noisy estimate of x
o
. Define the “noise” vector w
t
= ¯ x
t
− x
o
at iteration t of
CAMP. The histogram of w
t
, which is shown in Figure 1, suggests that the empirical distribution
of w
t
is “close” to a zero mean Gaussian probability density function. We will more rigorously
discuss the Gaussianity of w
t
in Section II-B.
March 8, 2012 DRAFT
5
Algorithm I: Ideal CAMP Algorithm
Input: y, A, τ, xo
Initialization x
0
= 0, z
0
= y
for t = 1 : maxiter
x
t
= A

z
t−1
+ x
t−1
σt = std( x
t
−xo)
z
t
= y −A x
t−1
+z
t−1 1

∂η
R
∂x
R
( x
t
; τσt) +
∂η
I
∂x
I
( x
t
; τσt)

x
t
= η( x
t
; τσt)
end
Output: x, x, σ∗
−0.5 0 0.5
0
1
2
3
Data

D
e
n
s
i
t
y
Iteration 2


−0.5 0 0.5
0
2
4
Data
Iteration 5


−0.5 0 0.5
0
2
4
6
8
Data
Iteration 10


(a) Histogram of the real part of w
t
.
−0.5 0 0.5
0
1
2
3
Data

D
e
n
s
i
t
y
Iteration 2


−0.5 0 0.5
0
2
4
Data
Iteration 5


−0.5 0 0.5
0
2
4
6
8
Data
Iteration 10


(b) Histogram of the imaginary part of w
t
.
Fig. 1. Histograms of the real and imaginary parts of the CAMP residual noise signal w
t
at different iterations of Ideal CAMP.
A Gaussian probability density function (pdf) is fitted to the histograms (red line). In these plots, N = 4000, δ = 0.6, and
σ = 10
−3
.
(iii) σ
t
is the standard deviation of w
t
. In the rest of the paper we use the notation σ

lim
t→∞
σ
t
.
CAMP first finds a noisy estimate of the signal called ¯ x
t
. Since this estimate is not sparse, the soft
thresholding function is applied to obtain a sparse estimate ´ x
t
. Here for the clarity of exposition we have
assumed that the algorithm uses x
o
in computing the noise standard deviation σ
t
. Hence, we refer to this
algorithm as Ideal CAMP. In Section IV we propose a more practical scheme where σ
t
is replaced with
an estimate.
March 8, 2012 DRAFT
6
B. State evolution: A framework for the analysis of CAMP
There are three important questions that have not been answered yet. (i) Under what conditions are the
above heuristics accurate? (ii) Can we use the properties of CAMP to predict its performance theoretically?
(iii) What is the formal connection between CAMP and the LASSO problem defined in (2).
The first question is accurately discussed in [22]–[24], [32]. It has been proved that in the asymptotic
setting N →∞, while δ is fixed, the above heuristics are correct. Consider the following definition from
[24].
Definition II.1. A sequence of instances ¦x
o
(N), A(N), n(N)¦ is called a converging sequence if the
following conditions hold:
- The empirical distribution of x
o
(N) ∈ R
N
converges weakly to a probability measure p
X
with
bounded second moment as N →∞.
- The empirical distribution of n ∈ R
n
(n = δN) converges weakly to a probability measure p
n
with
bounded second moment as N →∞.
- The elements of A(N) ∈ R
n×N
are i.i.d. drawn from a Gaussian distribution.
Theorem 1. [24] Let ¦x
o
(N), A(N), n(N)¦ be a converging sequence, and let ¯ x
t
(N) be the estimate
provided by the CAMP algorithm. The empirical law of w
t
(N) = ¯ x
t
(N) −x
o
(N) converges to a zero-
mean Gaussian distribution almost surely as N →∞.
While the above result has been proved for Gaussian measurement matrices in the asymptotic setting,
simulation results confirm that it is still accurate even for medium problem sizes with N ∼ 200 and
different classes of sensing matrices.
1
We explore this claim for the partial Fourier matrices that are of
particular interest in radar applications using Stepped Frequency (SF) waveforms in Section VI-A.
Theorem 2 enables us to answer the second question as well. Using the Gaussianity of the noise in
the asymptotic regime, one can predict the performance of CAMP through what is called the “state
evolution” (SE). SE tracks the evolution of the standard deviation of the noise σ
t
across iterations. Let
the marginal distribution of x
o
converge to p
X
, and let σ
t
denote the standard deviation of w
t
. In the
asymptotic setting, the value of the standard deviation at time t + 1 is calculated from σ
t
according to
the following equation:
σ
2
t+1
= Ψ(σ
t
), (3)
1
Empirical studies have already confirmed that this theoretical prediction holds for other sensing matrices with i.i.d. elements
other than Gaussian, [20], [22], [24].
March 8, 2012 DRAFT
7
where
Ψ(σ
t
) = σ
2
+
1
δ
E
_
[η(X + σ
t
Z; τσ
t
) −X[
2
_
, (4)
Z ∼ (A(0, 1), and the expectation is with respect to the two independent random variables X ∼ p
X
and
Z. In fact, if σ
t
is the standard deviation of the noise at iteration t, then E
_
[η(X + σ
t
Z; τσ
t
) −X[
2
_
equals the MSE of the estimate ´ x
t
after applying the soft thresholding function. It has been proved [23]
that the function Ψ is concave, and therefore the iteration (3), (4) has at most one stable fixed point σ
2

.
Also, CAMP converges to this fixed point exponentially fast (linear convergence according to optimization
literature). Appendix A provides the calculations involved in the Ψ function for a given distribution p
X
.
The following two consequences of (3), (4) are particularly useful for the radar application:
(i) The SE framework establishes the input/output relation for CAMP. Particularly, the output noise
power σ
2

is the sum of the actual system noise (σ
2
) plus a noise-like component caused by the
reconstruction itself (
1
δ
MSE). Consequently, for a given σ
2
, minimizing the reconstruction error
also minimizes σ
2

, therefore maximizing the CAMP reconstruction SNR.
2
(ii) The fixed point σ

depends on τ as well as on δ, p
X
, and σ. Figure 2 exhibits the dependence of
σ

on τ for two distinct values of δ and for a fixed problem instance, i.e., fixed p
X
and σ. As is
clear from the figure, there is a value of τ, say τ
o
, for which σ

is minimized. Moreover, as the
number of measurements decreases, both the optimal threshold τ
o
and the corresponding output
noise standard deviation increase.
Finally, to answer the third question we observe that there is a nice connection between the CAMP and
LASSO algorithms. According to [23], if τ is chosen to satisfy
λ τσ

_
1−
1

E
_
∂η
R
∂x
R
(X + σ

Z; τσ

)+
∂η
I
∂x
I
(X + σ

Z; τσ

)
__
, (5)
then in the asymptotic setting CAMP with threshold τ solves the LASSO problem (2) with parameter λ.
III. CS TARGET DETECTION USING CAMP
We now propose two CS target detection schemes based on CAMP and evaluate their performance
as measured by their Receiver Operating Characteristic (ROC). Let k be the number of targets, i.e., the
number of non-zero coefficients in x
o
. Define ρ =
k
n
, and define G as the distribution of the non-zero
2
For a given target received power a
2
, the CAMP input and output SNR are defined here respectively as SNRin = a
2
/(nσ
2
)
and SNRout = a
2

2

. Note that the output (reconstruction) SNR depends on, amongst other quantities, the CAMP threshold τ
via σ∗. A more rigorous definition of SNR is given in Section V.
March 8, 2012 DRAFT
8
0.5 1 1.5 2 2.5 3
0.25
0.3
0.35
0.4
0.45
0.5
0.55
0.6
τ
σ
*


δ = 0.6
δ = 0.2
Fig. 2. Fixed point σ∗ versus threshold τ for Ideal CAMP with σ = 0.23, compression factors δ = 0.2 (dashed, red line) and
δ = 0.6 (solid, black line). The sensing matrix A has i.i.d. Gaussian entries.
elements of x
o
. As before, σ
2
is the variance of the system noise. In this section we assume that k, G,
and σ are known. In Section IV we will investigate the realistic scenario where these parameters are not
known a priori and describe how to implement the proposed architectures in this case.
The two architectures we consider in this paper are displayed in Figure 3. Since any sparse recovery
algorithm is intrinsically a detection scheme, Architecture 1 seems a natural choice for a CS radar detector.
In this architecture, the measurements y are input to a recovery algorithm (here CAMP or equivalently
LASSO). This algorithm returns a sparse vector ´ x with the non-zero values being detections. Clearly,
the threshold parameter τ in CAMP (or equivalently the regularization parameter λ in LASSO) controls
the false alarm probability α and detection probability (P
d
) of the algorithm.
Proposition III.1. Consider the CAMP iteration with threshold τ
α
=

−ln α. If A(N), x
o
(N), w(N)
is a converging sequence and ´ x(N) is the fixed point of CAMP, then
lim
N→∞
1
N −k
N

i=1
1
{ xi(N)=0,xi(N)=0}
= α. (6)
almost surely. Also, τ
α
is the only value of τ for which (6) holds.
Proof: Define z ∼ (A(0, 1). According to [32] we know that
lim
N→∞
1
N −k
N

i=1
1
{ xi(N)=0,xi(N)=0}
= P([σ

z[ > τ
α
σ

) = e
−τ
2
α
.
The uniqueness is also clear from the above equation.
Roughly speaking, the above proposition states that the number of false alarms are on the order of
(N − k)α. Note that the connection between τ and the detection probability is not clear yet. We will
discuss this issue later.
March 8, 2012 DRAFT
9
(a) Architecture 1 (b) Architecture 2
Fig. 3. Block diagrams of the proposed architectures for CS-radar detection.
The second architecture is inspired by the standard, non-CS radar approach. Most commonly, radar
detectors comprise two stages: an estimation stage, where a noisy estimate of the signal is computed,
followed by a detection stage. Usually, a Matched Filter (MF), matched to the transmitted waveform,
is used to obtain a noisy estimate of the signal with optimum SNR. Then this noisy estimate is fed to
a detection block that controls the False Alarm Probability (P
fa
or FAP). Inspired by this philosophy
and by the properties of CAMP, we introduce Architecture 2. We first use CAMP to obtain a noisy,
non-sparse estimate of the signal ¯ x = x
o
+w. Similar to classical estimation procedures, the goal of the
first stage is to minimize σ

by choosing the optimal CAMP threshold τ = τ
o
. Once σ

is minimized
or equivalently the SNR is maximized, the noisy signal ¯ x can be fed to a detection block with fixed
threshold κ = σ

_
−ln(α), which is used to control the false alarm rate. From the Gaussianity of w, it
is clear that in the asymptotic setting this choice of κ results in the false alarm probability α as derived
in Proposition III.1.
As will be clarified later, Architecture 2 is much more appropriate for practical radar applications,
since all of the parameters can be optimized and estimated efficiently even without prior knowledge of k,
G, and σ. Furthermore, from a detection perspective, Architecture 2 outperforms Architecture 1. Suppose
that the Gaussianity of w
t
holds. Let τ
o
be the optimal value of τ that leads to the minimum σ

and that
this optimal value is unique and can be computed using (4). Under these assumptions, which provably
hold under the conditions specified in Theorem 1, the following theorem can be derived.
Theorem 2. Set the probability of false alarm to α for both Architecture 1 that uses τ
α
and Architecture
2 that uses τ
o
in CAMP. If P
d,1
and P
d,2
are the detection probabilities of the two schemes, then
P
d,1
≤ P
d,2
.
Furthermore the equality is satisfied at only one specific value α

= e
−τ
2
o
.
The above theorem is proved in Appendix B by comparing the detection and false alarm probabilities
March 8, 2012 DRAFT
10
10
−6
10
−4
10
−2
10
0
0
0.2
0.4
0.6
0.8
1
P
fa
P
d


Arch. 2, Theoretical
Arch. 1, Theoretical
Arch. 2, MC
Arch. 1, MC
0.2 0.4 0.6
0.9998
1


Fig. 4. Comparison of the ROC curves for Architectures 1 and 2 at δ = 0.6, ρ = 0.1 and σ
2
= 0.05. The distribution
G of the non-zero coefficients of xo is chosen such that all non zero components have the same amplitude (equal to 1) and
phase uniformly distributed between −π and π. The solid and dashed lines represent the theoretical predictions based on the
SE equation. The dots are the results of MC simulations using Ideal CAMP. The sensing matrix for MC simulations had i.i.d.
Gaussian entries.
of the two architectures. However, since in Architecture 2 the CAMP threshold is designed to minimize
the output noise variance, intuitively it is clear that, for the same P
fa
, any other choice of τ will lead to
a higher σ

, i.e., a lower SNR, and therefore a lower P
d
. Example ROC curves for Architectures 1 and
2 are shown in Figure 4. The solid and dashed lines are obtained using the analytical equations derived
in Appendix B. The theoretical ROC curves are verified by Monte Carlo (MC) simulations (dots). In
the simulations we averaged over 10000 realizations of the Ideal CAMP algorithm given in Algorithm
I for several values of P
fa
ranging from 10
−1
to 10
−5
. In Figure 4 we can also observe an interesting
characteristic of Architecture 1 in the region around P
fa
= 0.4 (the zoomed area); the probability of
detection, P
d,1
, decreases as P
fa
increases above this value. To understand why, recall that in Architecture
1 the CAMP threshold τ
α
varies with P
fa
. Therefore σ

(and hence the CAMP reconstruction SNR) also
changes with P
fa
, and it is not constant along the ROC curve. This explains why P
d,1
reaches its
maximum at around P
fa
= 0.4 and then decreases again as the P
fa
goes to 1. Instead, in Architecture
2, σ

is fixed to its minimum along the ROC curve, and therefore P
d,2
increases with increasing FAP.
From this plot it is also possible to observe that there exists only one value of FAP (α

= 0.22) where
P
d,1
= P
d,2
.
3
3
The difference in performance between the two architectures depends significantly on the values of (δ, ρ, σ) and on the
sensing matrix in use, as explained later in Section VI.
March 8, 2012 DRAFT
11
IV. ADAPTIVE CAMP CFAR RADAR DETECTOR
So far, we have assumed we know exactly ρ, σ, and G (to derive the theoretical fixed point) and x
o
(to run Ideal CAMP). However, in practical radar systems such information is not available, and both
the CAMP and detector parameters must be estimated from the CS measurements y. In this section we
demonstrate how these issues can be handled in practice and propose a fully adaptive scheme, i.e., one
that does not require any prior information and that adapts to changes in noise level. There are three
main issues that have to be settled. (i) How to estimate σ
t
without knowing x
o
. (ii) How to compute the
optimal value τ
o
for Architecture 2, efficiently and accurately, without the SE. (iii) How to replace the
fixed threshold κ for the detector in Architecture 2 with an adaptive threshold to maintain a CFAR.
The first question can be answered in several different ways. For instance, in this paper we use the
median to estimate the standard deviation via
´ σ
t
=
_
1
ln 2
median([¯ x
t
[). (7)
This estimator is unbiased if x
o
= 0. However, in the presence of targets, i.e., when x
o
,= 0, (7) is a
biased estimator. The main advantage of this scheme is its robustness to high SNRs. To see this, we
consider the performance of the median estimation in the asymptotic setting. As mentioned above, in the
asymptotic setting at each iteration of CAMP we have ¯ x
t
= x
o
+ w
t
. Assume that the elements of x
o
are distributed as x
o,i
i.i.d.
∼ (1 −)G(x) + (1 −)δ
0
(x) with = δρ ¸1 and w
t
i
∼ (A(0, σ
2

). The goal
is to estimate the median µ

of [w
t
[. However, since we only have access to ¯ x, we estimate the median
of [w
t
[ as the ´ µ that satisfies P([¯ x
t
i
[ > ´ µ) =
1
2
. The following theorem provides an upper bound on the
deviation of the estimated median from the true median.
Proposition IV.1. The error of the estimated median is bounded above by
[´ µ −µ

[
σ


[ ln(1 −)[
2

ln 2
. (8)
This proposition is proved in Appendix C. It is important to note several interesting properties of the
upper bound (8). First, the distribution G of the non zero components of x
o
does not play any role. Second,
for small values of , i.e., in very sparse situations, [ ln(1−)[ ≈ and, hence, the error is proportional to
the sparsity level. So, when the noise level and x
o
are unknown, a practical implementation of Algorithm
I is obtained by replacing σ
t
with the estimate in (7). We will refer to this algorithm simply as CAMP
or Median CAMP.
The estimate of σ
t
provides an approach to answer the second above question as well, namely how to
estimate, in Architecture 2, the optimum CAMP threshold ´ τ
o
that minimizes ´ σ
2

. Suppose that we know
March 8, 2012 DRAFT
12
or can estimate τ
max
such that τ
o
< τ
max
. Given a step size δ
τ
, we define a sequence of thresholds
τ = ¦τ

¦
L
=1
such that τ
1
= τ
max
and τ

= τ
−1
−δ
τ
. Starting from τ
max
, at each new iteration CAMP
is initialized with ´ x
0

= ´ x
−1
and z
0

= z
−1
. Using the solution of CAMP at the previous iteration −1
as initial value for τ = τ

, CAMP needs only a few iterations to converge to the solution, and therefore
the entire process is very fast. After L iterations, we have a matrix of solutions
´
X = [´ x
1
, ´ x
2
, , ´ x
L
] of
size N L, where each column contains the CAMP solution for a given τ

. Also, we have L estimates
¦´ σ


¦
L
=1
. The optimum estimated threshold ´ τ
o
is chosen as the one that minimizes the estimated CAMP
output noise variance ´ σ
2

. It is clear that δ
τ
specifies the trade-off between computational complexity and
the accuracy of the algorithm in estimating ´ τ
o
. Decreasing δ
τ
increases the number of points L needed
to span the same τ search region, but it also results in a more accurate estimate of ´ τ
o
.
We now explain how to set τ
max
. At the first iteration ( = 1 and t = 1) the CAMP algorithm is
initialized with ´ x
0
= 0 and z
0
= y, and we have ¯ x = A
H
y, where A
H
is the Hermitian of the matrix A.
Suppose that ´ σ
0
is estimated from this vector. Consider now the LASSO problem in (2). It is well known
that for λ > λ
max
= |A
H
y|

the only solution is the zero solution. Using the calibration equation (5)
with λ = λ
max
and σ

= ´ σ
0
, we can compute an estimate of ´ τ
max
. We will refer to this algorithm as
Adaptive CAMP, since both the noise variance ´ σ
t
the threshold ´ τ
o
are adaptively estimated inside the
algorithm itself, and the only input variables are y and A.
It remains now to establish how to replace the fixed threshold κ with an adaptive one for Architecture
2. In Appendix B we show how to set the fixed threshold κ to achieve the desired FAP when the noise
variance σ
2
is known. In practice, however, the noise statistics are not known in advance. Hence, in
classical radar detectors a CFAR processor is employed. In CFAR schemes the Cell Under Test (CUT),
is tested for the presence of a target against a threshold that is derived from an estimated clutter plus
noise power. The 2N
w
cells (CFAR window) surrounding the CUT are used to derive an estimate of
the local background and they are assumed to be target free. Commonly, 2N
G
guard cells immediately
adjacent to the CUT are excluded from the CFAR window. The great advantage of CFAR schemes is
that they are able to maintain a constant false alarm rate via adaptation of the threshold to a changing
environment.
The general form of a CFAR test is
X
H1

H0
βY, (9)
where the random variable X represents some function (generally envelope or power) of the CUT, say
x
o,i
, β is a threshold multiplier, and Y is also a random variable function of the cells in the CFAR
March 8, 2012 DRAFT
13
Fig. 5. Block diagram of the Adaptive CAMP CFAR detector.
window [x
o,i−Nw−NG
, , x
o,i−NG−1
, x
o,i+NG+1
, , x
o,i+Nw+NG
]. In the well known Cell Averaging
CFAR (CA-CFAR) detector, Y is the average of the cells in the CFAR window [14]. Clearly, one has to
know the relation between the CFAR threshold multiplier and P
fa
so that β can be adjusted to maintain
P
fa
constant during the observation time. Hence, the noise distribution should be known to design a
CFAR scheme. If the CAMP estimate ´ x were input to a CFAR processor, then estimation of the noise
characteristics would be very difficult, since many samples in ´ x are identically zero. Instead, the signal
¯ x contains all non-zero samples, and it is modeled as the sum of targets plus Gaussian noise; this can
be directly input to a conventional CFAR processor. A block diagram of the Adaptive CAMP CFAR
detector based on Architecture 2 is shown in Figure 5. For the properties of CAMP, in Architecture
2 replacing the fixed threshold detector with a CFAR detector should provide comparable results as in
classical CFAR without CS.
V. DEFINING THE SNR FOR CS CAMP
Before proceeding to the simulations and experiments, we provide a definition of SNR for the CAMP-
based CS radar system. This definition will be useful not only for understanding the performance of
the proposed detectors, but also to compare the novel CS based architectures to more classical ones for
which the performance is known. We observed in (4) that the variance σ
2

of the noise present in the
CAMP estimate depends on the 5-tuple (δ, ρ, p
X
, τ, σ). In fact, even for fixed p
X
, σ and for a specific
Architecture (that fixes τ), σ
2

, and therefore the (P
d
, P
fa
) curves, can vary significantly with δ and ρ.
This is uncommon in classic radar systems, where the SNR after the MF is uniquely determined for a
given input SNR and integration time. Therefore it is important to relate the performance of a given CS
scheme to that of MF-based classical systems.
Let A
+
∈ C
N×N
be the measurement matrix for the case δ = 1, i.e., no undersampling. A conventional
radar processor feeds the measurement vector y = A
+
x
o
+σz, z ∼ (A
N
(0, I), to a MF to obtain a noisy
March 8, 2012 DRAFT
14
estimate ´ x
MF
= x
o
+σz of the target received signal with optimum SNR.
4
CAMP enables us to define
SNR in a similar way. Once CAMP has converged we have access to both the sparse estimate ´ x and
its noisy version ¯ x. As stated before, even in medium size problems ¯ x can be accurately approximated
by the sum of the true target vector plus white Gaussian noise with variance σ
2

, i.e. ¯ x = x
o
+ σ

z.
Indicating with a
2
the target received power at bin i, we define the SNR at the output of MF or CAMP
respectively as
SNR
MF
=
a
2
σ
2
, SNR
CS
=
a
2
σ
2

. (10)
In the remainder of the paper we assume that the total transmitted (and received) power is independent of
δ.
5
In the case of partial Fourier sensing matrix, described later in Sections VI and VII, this is achieved
in practice by dividing the total available transmit power over the subset of transmitted frequencies. With
this assumption, dividing both sides of (4) by a
2
, we can derive the CAMP output SNR as a function of
the equivalent MF output SNR, i.e., the SNR that we would obtain without compression using a MF. The
constraint of keeping the received target power equal for different amounts of undersampling enables us
to evaluate the changes in SNR
CS
due to reconstruction and not due to a reduction in total signal power.
In other words, if we normalize (4) by a
2
, then we keep the first term on the right hand side fixed (which
is equal to SNR
MF
) and observe the change in SNR
CS
produced by the reconstruction error term (1/δ
MSE). Note that, with the above assumptions made, SNR
MF
represents an upper bound on the highest
SNR that can be obtained from CAMP.
VI. SIMULATION RESULTS
In this section we investigate the performance of median and Adaptive CAMP using MC simulations
and compare it with the theoretical results obtained from the SE. Moreover, we replace the Gaussian
sensing matrices, for which SE provably applies, with partial Fourier matrices, which are of particular
interest in radar applications transmitting Stepped Frequency (SF) waveforms [9], [11], [33]. An n N
partial Fourier matrix is obtained from an N N discrete Fourier transform matrix by preserving only
a random subset n of the original N matrix rows. The elements of a partial Fourier matrix are given by
a
i,j
= e
−j2π(i−1)(j−1)/N
, for i = 1, , n and j = 1, , N.
4
In a multiple target scenario, the MF SNR is optimum and independent of the number of targets as long as each target is
exactly on a grid point and the matrix A+ is orthogonal. We assume these (ideal) conditions are satisfied when computing the
MF SNR.
5
This is achieved by using sensing matrices with unit column norms in all simulations.
March 8, 2012 DRAFT
15
−0.2 0 0.2
0
2
4
6
δ = 0.2, ρ = 0.1


Data
−2 0 2
0
0.5
1
Data
δ = 0.2, ρ = 0.7


−0.5 0 0.5
0
1
2
Data
δ = 0.9, ρ = 0.5


−0.4 −0.2 0 0.2 0.4
0
2
4

D
e
n
s
i
t
y
δ = 0.7, ρ = 0.2


Data
(a) Histogram of the real part of w.
−0.5 0 0.5
0
1
2
Data
δ = 0.9, ρ = 0.5


−1 0 1
0
0.5
1
Data
δ = 0.2, ρ = 0.7


−0.2 0 0.2
0
2
4

D
e
n
s
i
t
y
δ = 0.7, ρ = 0.2


Data
−0.2 0 0.2
0
2
4
6
δ = 0.2, ρ = 0.1


Data
(b) Histogram of the imaginary part of w.
Fig. 6. Histograms (bars) of the real and imaginary parts of the the noise signal w for different combinations (δ, ρ) using CAMP
with threshold τ = 1.8. A Gaussian pdf is fitted to the histograms (solid, red line). In these plots σ
2
= 0.1 and N = 4000.
The sensing matrix is partial Fourier. See Appendix D for the p-values of Kolmogorov-Smirnov test.
A. Gaussianity of w using partial Fourier matrices
In this section we investigate the Gaussianity of the reconstruction noise vector w
t
for a partial Fourier
sensing matrix using MC simulations. In Figure 6 we show a few examples of the empirical distribution
of w at convergence for different combinations of δ and ρ. To obtain the histograms we used CAMP with
a fixed (not necessarily optimal) threshold τ and we fixed σ = 0.1. Simulation results confirm that the
Gaussianity of the noise vector is preserved for partial Fourier matrices as well. We further investigate
its Gaussianity using the Kolmogorov-Smirnov test in Appendix D.
B. Accuracy of the SE
We investigate the accuracy of the SE by comparing the theoretical results obtained from (4) with
simulations results obtained using the Ideal CAMP algorithm. Moreover, we study the behavior of σ

for
the case of a partial Fourier sensing matrix and investigate how it deviates from the theoretical case of
a Gaussian sensing matrix, for which the SE applies. Figure 7 compares σ

obtained from Ideal CAMP
for the case of complex Gaussian and partial Fourier sensing matrices with the theoretical one from the
SE. The following two remarks can be made:
(i) SE does not predict the performance of CAMP accurately for partial Fourier matrices. However, for
τ = τ
o
, the value of σ

for Fourier is not very different from the value for Gaussian.
March 8, 2012 DRAFT
16
2 2.5 3
0.26
0.27
0.28
0.29
0.3
τ
σ
*
δ = 0.05


1.5 2 2.5 3
0.26
0.27
0.28
0.29
0.3
τ
σ
*
δ = 0.1


1.5 2 2.5 3
0.26
0.27
0.28
0.29
0.3
τ
σ
*
δ = 0.2


1 1.5 2 2.5 3
0.25
0.26
0.27
0.28
0.29
0.3
τ
σ
*
δ = 0.5


CAMP Ideal, Gaussian
CAMP Ideal, Fourier
Theoretical SE
Fig. 7. σ∗ versus τ from Ideal CAMP for both complex Gaussian and partial Fourier sensing matrices. The empirical curves
are obtained by averaging over 100 MC samples for σ
2
= 0.05, ρ = 0.05, N = 4000, and different δ. The solid, blue line
shows the analytical σ∗ computed from (13) in Appendix A.
(ii) As δ → 0 the predictions of SE become more accurate for Fourier. However, as the number of
measurements increases, i.e., δ →1, the columns of the matrix become more correlated, and hence
the true behavior deviates from the SE that holds for matrices with i.i.d. entries.
C. Effects of the median estimator in CAMP
We investigate the effect of replacing the true σ
t
with the median based estimate ´ σ
t
from (7) when
x
o
,= 0. In Figure 8 the estimated output noise standard deviation is shown for both Ideal and Median
CAMP. We observe that the estimate ´ σ

deviates from the Ideal CAMP case because of the bias introduced
by the estimator. Furthermore, as is clear from this figure and confirmed by the upper bound provided
in Appendix C, the deviation diminishes as = δρ decreases.
In Architecture 1, overestimating ´ σ

will result in a loss in performance compared to the case of using
Ideal CAMP. This is because the soft thresholding function in CAMP uses the parameter τ
α
´ σ

. For a fixed
FAP (that fixes τ
α
), the bias will have the effect of increasing the overall threshold, therefore resulting in
losing both detection performance and the CFAR property (since the output P
fa
is underestimated). This
is shown in Figure 9, where we plot the ROC for Architecture 1 using both Ideal and Median CAMP.
From this figure we also observe that, for the value of δ used here, Architecture 1 performs much better
when using the partial Fourier sensing matrix as compared to the Gaussian one. This is related to the
March 8, 2012 DRAFT
17
2 2.5 3
0.26
0.27
0.28
0.29
0.3
τ
σ
*
δ = 0.05


1.5 2 2.5 3
0.26
0.27
0.28
0.29
0.3
τ
σ
*
δ = 0.1


1.5 2 2.5 3
0.26
0.27
0.28
0.29
0.3
τ
σ
*
δ = 0.2


1 1.5 2 2.5 3
0.25
0.26
0.27
0.28
0.29
0.3
τ
σ
*
δ = 0.5


CAMP Adaptive, Gaussian
CAMP Adaptive, Fourier
CAMP Ideal, Gaussian
CAMP Ideal, Fourier
Fig. 8. Output noise standard deviation versus τ for both Ideal (σ∗) and Median ( σ∗) CAMP for complex Gaussian and partial
Fourier sensing matrices. The curves are obtained by averaging over 100 MC realizations for σ
2
= 0.05, ρ = 0.05, N = 4000,
and various δ.
10
−4
10
−3
10
−2
10
−1
0.4
0.6
0.8
1
P
fa
P
d


Median CAMP
Ideal CAMP
(a) Gaussian sensing matrix
10
−4
10
−3
10
−2
10
−1
0.75
0.8
0.85
0.9
0.95
1
P
fa
P
d


Median CAMP
Ideal CAMP
(b) Partial Fourier sensing matrix
Fig. 9. ROC curves for Architecture 1 using Ideal and Median CAMP. Here N = 1000, δ = 0.6, ρ = 0.1, a
2
= 1, and
σ
2
= 0.05 (corresponding to a MF SNR = 13dB).
variation of the noise variance σ
2

, and therefore the SNR, with the threshold τ
α
along the ROC curve.
As can be seen from Figure 8, in the partial Fourier case as δ increases the variance curve becomes
flatter than in the Gaussian case. This implies that, for decreasing FAPs, in the partial Fourier case the
SNR in Architecture 1 deviates much less from the optimum SNR that is achieved for τ = τ
o
.
March 8, 2012 DRAFT
18
10
−4
10
−3
10
−2
10
−1
0.7
0.75
0.8
0.85
0.9
0.95
1
P
fa
P
d


Ideal CAMP CFAR
Adaptive CAMP FT
Adaptive CAMP CFAR
Ideal CAMP FT
Theoretical CA−CFAR
(a) Gaussian sensing matrix
10
−4
10
−3
10
−2
10
−1
0.75
0.8
0.85
0.9
0.95
P
fa
P
d


Ideal CAMP CFAR
Adaptive CAMP FT
Ideal CAMP FT
Adaptive CAMP CFAR
Theoretical CA−CFAR
(b) Partial Fourier sensing matrix
Fig. 10. ROC curves for Architecture 2 for different levels of adaptivity. Here N = 1000, δ = 0.6, ρ = 0.1, a
2
= 1, and
σ
2
= 0.05 (corresponding to a MF SNR = 13dB). FT denotes the use of an (ideal) fixed threshold detector.
D. Adaptive CAMP CFAR detector performance
We investigate now the (P
d
, P
fa
) performance of the Adaptive CAMP CFAR detector. In the simula-
tions for Architecture 2, Adaptive CAMP is followed by a CA-CFAR processor preceded by a Square
Law (SL) detector (recall Figure 5). For the CFAR processor, we use a CFAR window of length 20 with
4 guard cells.
It is a well known fact that, if one or more targets are present in the CFAR window, then they cause
a rise in the adaptive threshold, thus possibly masking the target in the CUT that has yet to be detected.
Therefore ,in the basic analysis of CFAR schemes, it is assumed that there are no targets present in the
CFAR window. This scenario is referred to as the non-interfering targets scenario. Following a similar
approach, we consider the case of multiple but non-interfering targets. For the case of interfering targets,
to reduce the interference losses encountered in CA-CFAR, several dedicated CFAR schemes have been
proposed in literature, such as Order Statistic (OS) CFAR [15]. To keep the discussions concise, we do
not pursue these directions and leave them for future research.
Adaptivity imposes extra losses on the system. One loss is due to the use of Adaptive- instead of Ideal
CAMP. This means that there is an error in estimating τ
o
. A second loss is caused by the CFAR processor
and its estimate of the noise standard deviation. In Figure 10 we show the ROC curves for Architecture
2 obtained using: (a) Ideal CAMP with ideal (fixed threshold) detector, (b) Ideal CAMP in combination
with the CA-CFAR detector, (c) Adaptive CAMP with ideal (fixed threshold) detector, and (d) a fully
adaptive scheme consisting of Adaptive CAMP followed by CA-CFAR processor. We also show the
theoretical curve of a CA-CFAR processor with the same window length and SNR = 11.8dB for the
Gaussian sensing matrix and 12.4dB for the partial Fourier sensing matrix. The SNR can be estimated
March 8, 2012 DRAFT
19
during simulations, or it can be derived using the procedure described later in Section VIII. For the
Gaussian sensing matrix the optimal threshold for Architecture 2 (using Ideal CAMP) is computed using
the SE. For the partial Fourier sensing matrix, the threshold in Ideal CAMP is set to τ
o
= 1.85, which is
derived from a plot like the ones shown in Figure 7 for the case δ = 0.6. It is clear from this figure that
Adaptive CAMP introduces almost no loss in the detection performance of Architecture 2. This is due
to the fact that, although ´ σ

is biased, the value of ´ τ
o
at which the minimum ´ σ

occurs is very close to
the true optimal τ
o
(see Figure 8). The main loss instead is introduced by the adaptive CFAR detector.
This is the well-known CFAR loss [14], which can be controlled by changing the CFAR window length.
In general, the window length depends on the specific application (e.g., the expected types of target and
environment). Furthermore, by comparing the curve of Adaptive CAMP plus CFAR with the theoretical
one of a CA-CFAR processor (without CS) with equal parameters, we observe that the CA-CFAR detector
performance seems independent of the fact that the input to the detector is obtained by running CAMP
instead of a conventional MF.
By comparing Figures 9 and 10 it can be seen that, in the fixed threshold case, Architecture 2 always
outperforms Architecture 1, as predicted by Theorem 2.
VII. EXPERIMENTAL RESULTS
To verify our theoretical and simulated results, we performed several experiments at Fraunhofer FHR
using the software defined LabRadOr experimental radar system. The digital SF waveforms were designed
to acquire a set of CS measurements for δ = 0.5 and 0.25. A more detailed description of the radar
system and the setup is given in [33]. In the experiments, we use five stationary corner reflectors of
different sizes (RCS) as targets. The corner reflectors are about 5m apart in range. For each transmitted
(TX) waveform 300 measurements (with the same set-up) were performed.
A. Transmitted waveform
We use a SF waveform, meaning that the TX signal consists of a number of discrete frequencies. In the
Nyquist case (that represents unambiguous mapping of ranges to phases over the whole bandwidth) we
transmit N = 200 frequencies over a bandwidth of 800MHz. The achievable range resolution is therefore
δ
R
= 18.75cm. Each frequency is transmitted during 0.512µs, thus implying a bandwidth of B
f
=
1.95MHz, and sequential frequencies are separated by ∆f = 4MHz, resulting in 37.5m unambiguous
range (∆R).
March 8, 2012 DRAFT
20
0 5 10 15 20 25 30 35
10
1
10
2
10
3
10
4
Range [m]
l
o
g
(
a
m
p
l
i
t
u
d
e
)



CAMP Arch. 1
CAMP Arch. 2
MF
T1
T2
T3
T4
T5
Fig. 11. Reconstructed range profile using: CAMP Architectures 1 and 2, and MF. For the MF, N = 200; for CS, δ = 0.5.
In the CS case, the number of TX frequencies is reduced from N to n (n < N). The subset of
transmitted frequencies is chosen uniformly at random within the total transmitted bandwidth. In our
experiments, we fix N = 200, and n = 100, 50, corresponding to δ = 0.25, 0.5 and ρ = 0.1, 0.05.
After reception and demodulation, each range bin, r
i
= r
0
+i ∆R/N, i = 1, . . . , N, maps to n phases
proportional to the n transmitted frequencies f
m
= f
0
+ m∆f, m = 1, . . . , n. Therefore, the sensing
matrix A can be represented as a partial Fourier matrix, and the n samples y
m
of the compressed
measurement vector y are given by
y
m
=
1

n
N

i=1
e
−j2kmri
x
o,i
= a
(m)
x
o
, (11)
where a
(m)
is the m
th
row of the partial Fourier matrix, k
m
= 2πf
m
/c is the wave number, and c is
the speed of light. We assume that the same total power is transmitted, irrespective of the number of
transmitted frequencies. This means that when the number of measurements is reduced by δ, the power
per transmitted frequency P
T
is 1/δ times higher than in the Nyquist waveform case, so that the total
transmitted energy (P
T
n/B
f
) is the same in all cases, irrespective of n. This enables us to analyze
the effects of undersampling and transmitted power separately.
B. CFAR CAMP adaptive detector performance
We show now the ROC curves obtained using CAMP-based detection on the measured data. Figure
11 exhibits the signals reconstructed using Architectures 1 and 2 in addition to the MF. For Architecture
1, τ
α
was set using a FAP = 10
−4
. In this figure the five corner reflectors are indicated as T1–T5. In
the experimental measurements, due to the fact that targets are not exactly on Fourier grid points, there
is a leakage of target power into neighboring range bins both for MF and CS. Also, since the SNR
is very high for all targets (in all cases above 20dB), to evaluate the performance of the detectors at
March 8, 2012 DRAFT
21
10
−3
10
−2
10
−1
10
0
0.2
0.4
0.6
0.8
1
P
fa

P
d


Arch. 2 CA−CFAR
Arch. 2 FT
Arch. 1
T1
T3
T4
(a) δ = 0.5
10
−3
10
−2
10
−1
10
0
0
0.2
0.4
0.6
0.8
1
P
fa

P
d


Arch. 2 CA−CFAR
Arch. 2 FT
Arch. 1
T3
T1
T4
(b) δ = 0.25
Fig. 12. ROC curves for Architecture 1 and Architecture 2 using both fixed threshold (FT) and CA-CFAR detectors. The
curves correspond to targets T1, T3 and T4 in Figure 11.
medium SNR values, we added additional white Gaussian noise (with σ = 500) to the raw frequency data
samples, resulting in a MF output SNR of 17.2, 16.6, 14, 10.2 and 26dB, respectively, from the closest
corner (T1) to the farthest one (T5). Figure 12 shows the ROC curves for three of the five targets with
different SNRs, indicated as T1, T3, and T4 in Figure 11, for δ = 0.5 and 0.25. In these plots, the FAP
was estimated by excluding the range bins corresponding to the target locations plus four guard cells.
For P
d
estimation, we used the detection at the location of the highest target peak. For each target the
ROC curve is shown using both Architecture 1 and Architecture 2. In Architecture 1, the highest FAP
that can be estimated is equal to δ, since the sparse estimated signal ´ x cannot have more than n out of
N non-zero coefficients. In Architecture 2, CAMP recovery is followed by either a fixed threshold (FT)
detector or a CA-CFAR processor that uses four guard cells and a CFAR window of length 20.
Recall that in CAMP the reconstruction SNR is the ratio of the estimated target power to the system
plus reconstruction noise power (σ
2

). Therefore, while the total transmitted power remains fixed for
δ = 0.5 and 0.25, the reconstruction SNR for each target depends on the architecture used in addition to
the compression factor δ. As observed in the simulated results in Figures 7 and 8, the reconstruction noise
variance σ

increases, and thus the CAMP SNR decreases, as the number of measurements is reduced.
Since a loss in SNR translates directly into a loss in detection probability, for a given FAP, CAMP will
perform better for larger δ.
From Figure 12 we observe that, in agreement with our theoretical findings, the detection probability
of Architecture 2 with the fixed threshold detector is always higher than that of Architecture 1. When
Architecture 2 is followed by a CA-CFAR processor, as expected, there is a loss in detection performance
compared to the fixed threshold case.
March 8, 2012 DRAFT
22
At very low and very high SNRs, the two proposed architectures perform very similarly. However,
at low FAPs and high P
d
, which is the most relevant case in practical situations, Architecture 2 always
outperforms Architecture 1. Furthermore, we observe that Architecture 1 is comparable to an OS-CFAR
detector where the CFAR window is the entire signal, i.e., 2N
w
= N, including the CUT. However, a
serious disadvantage of Architecture 1 is that, since the entire signal is used in the noise estimation and
the threshold τ
α
is fixed, it can not adapt to local variation of noise level. This makes Architecture 1
unsuitable for many radar applications. Architecture 2, in contrast, provides the flexibility to choose both
the most appropriate CFAR processor and CFAR window length depending on the specific scenario.
VIII. DESIGN METHODOLOGY
Using the tools developed in the previous sections, we propose here a methodology for designing CS
radar detectors based on CAMP. Given the detection range, target RCS and system noise, the first step
is to compute the transmitted power necessary to reach a given CAMP output SNR. For example, if we
consider Architecture 2 and we assume that the received power is equal to a
2
for all targets,
6
then using
equation (4) for a given σ
2
we obtain a value of σ
2

for each couple (δ, ρ). Computing the ratio a
2

2

we
obtain the CAMP output SNR map, an example of which is shown in Figure 13 for a Gaussian sensing
matrix.
7
It is clear that CS radar performance depends, besides on the SNR, on δ and ρ. Therefore, in
the system design phase, an estimate of the expected number of targets should be made. To design the
system for the worst-case scenario, k could be set based on the maximum expected number of targets,
i.e., k = k
max
. Then, for a given number of range (or Doppler) bins N and a given σ, we can vary the
number of CS measurements n to obtain several values of δ = n/N and corresponding ρ = k
max
/n. By
evaluating the SNR at these points, we obtain a curve that shows how power (SNR
CS
) and undersampling
(δ) can be traded against one another.
In Figure 14 we show an example of such curves for several values of k
max
and N = 200. In Figure
14(a) the sensing matrix has i.i.d. Gaussian entries and the curves are obtained using the theoretical SE. In
Figure 14(b) the sensing matrix is partial Fourier and the curves are obtained using MC simulations with
Adaptive CAMP. Observe that in both cases the curves are equal for the same k
max
up to approximately
6
This choice of target amplitudes distribution provides a lower bound on the SNR performance, being this the least favorable
distribution for the non-zero entries in xo [23].
7
For sensing matrices other than Gaussian, the SNR map of CAMP can be obtained via simulations using the Ideal or Adaptive
CAMP algorithms, as shown in Figure 14(b) for a few sets of points in the map. For Figure 13 we used the analytical equations
from Appendix A.
March 8, 2012 DRAFT
23
δ
ρ


0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0
2
4
6
8
10
12
Fig. 13. CAMP output SNR for Architecture 2 for a = 1 and σ
2
= 0.05, corresponding to SNRMF = 13dB. The sensing
matrix is i.i.d. Gaussian. The colormap is the output SNR of CAMP in a dB scale.
0 0.2 0.4 0.6 0.8 1
2
4
6
8
10
12
14
δ
S
N
R
C
S


k
max
= 1
k
max
= 6
k
max
= 12
k
max
= 24
k
max
= 48
(a) i.i.d. Gaussian sensing matrix
0 0.2 0.4 0.6 0.8 1
2
4
6
8
10
12
14
δ
S
N
R
C
S


k
max
= 1
k
max
= 6
k
max
= 12
k
max
= 24
k
max
= 48
(b) partial Fourier sensing matrix
Fig. 14. CAMP output SNR versus δ for Architecture 2 for different numbers of targets kmax. Here N = 200, a = 1, and
σ
2
= 0.05, corresponding to SNRMF = 13dB.
δ = 0.8. However, as δ →1 the partial Fourier matrix approaches a full Fourier matrix, i.e., it becomes
orthogonal, and SNR
CS
→ 13 dB. On the contrary, even in the limit n = N (i.e., δ = 1) the Gaussian
sensing matrix is not orthogonal. Hence, we estimate that if k
max
> 1, just as in conventional MF, there
will be losses due to target interference. In this case the upper bound (SNR = 13dB) cannot be achieved.
Once we have computed the SNR maps or curves, we can use them to predict the performance of our
CAMP CFAR detector. Assume, for example, that we would like to choose δ = 0.6 and k
max
= 12.
For this combination (δ, ρ) we derive a value for the CAMP output SNR equal to 11.7dB that can be
plugged in the selected CFAR P
d
and P
fa
equations. If we use a CA-CFAR processor and we assume
March 8, 2012 DRAFT
24
that the targets are not interfering, then for P
fa
= 10
−4
we obtain P
d
= 0.7. This means that, if we want
to increase the detection probability for the desired FAP, since k
max
is fixed, we either have to increase
the number of CS measurements (δ) or the received power. The latter can be achieved by increasing the
transmitted power or any term in the radar equation that increases the received power, such as antenna
gain.
IX. CONCLUSIONS
In this paper we have achieved two goals. Firstly, we have presented the first architecture for adaptive
CS radar detection with CFAR properties. Secondly, we have provided a methodology to predict the
performance of the proposed detector, which makes it possible to design practical CS based radar systems.
These goals have been achieved by exploiting CAMP, which features closed-form expressions for the
detection and false alarm probabilities in the ideal case of known system parameters. Based on the SE
theoretical results, we have demonstrated that, out of two proposed architectures, the combination of a
recovery stage based on CAMP with a separate detector achieves the best performance. With a simple
modification to CAMP, we have combined conventional CFAR processing with
1
-norm minimization to
provide a fully adaptive detection scheme. Our theoretical findings have been supported by both numerical
and experimental evidence.
Furthermore, by comparing theoretical and experimental results we have been able to understand the
behavior of CAMP for sensing matrices other than i.i.d. Gaussian. In fact, our experimental data has shown
that our conclusions still hold for the case of partial Fourier sensing matrices, for which, unfortunately,
no theoretical claims can yet be made.
We have derived closed form expressions for the CAMP output SNR as a function of system and
target parameters. These relations can be used to obtain CS link budget plots that allow the system
designer to evaluate the trade off between power and undersampling. Such charts play an important role
in determining when and how CS can be applied and at what cost.
We believe this work paves the way for the use of CS in radar detection.
APPENDIX A
RISK OF THE SOFT THRESHOLDING FUNCTION
We derive an analytical expression for (4) for the case when the non-zero coefficients in x
o
all have
equal amplitude. At the end of the appendix, we also explain how the calculations can be easily generalized
to the more general setting where the non-zero elements have different amplitude distributions. For the
March 8, 2012 DRAFT
25
complex signal case, it has been proved in [23] that the function Ψ(σ

) is independent of the phase
of the non-zero coefficients, and therefore, without loss of generality, we will assume in the following
calculations that the phase of the non-zero coefficients is equal to zero. Therefore, we have
f
x
(x) = (1 −δρ)δ(0) + δρδ(a), (12)
where δρ =
k
N
and a is the amplitude of all non-zero coefficients in x
o
.
Using f
x
(x) in (4), we obtain
σ
2

= σ
2
+
1
δ
E
Z,X
¦[(η(X + σ

Z; τσ

) −X)[
2
¦
= σ
2
+
1
δ
_
E
Z
¦[(η(σ

Z; τσ

))[
2
¦P(x = 0) +E
Z
¦[(η(a + σ

Z; τσ

) −a)[
2
¦P(x = a)
_
= σ
2
+
σ
2

δ
_
(1 −δρ)E
Z
¦[(η(Z; τ))[
2
¦ + δρE
Z
¦[(η(µ + Z; τ) −µ)[
2
¦
_
, (13)
where Z ∼ (A(0, 1) and can be decomposed as Z = (Z
r
+ jZ
i
), where Z
r
, Z
i
∼ A(0, 1/2), and
µ = a/σ

. Define the two independent random variables w = [z[ ∼ Rayleigh(
1
2
) and θ = (z) ∼ that is
uniformly distributed between 0 and 2π. The first expectation in (13) can be computed as
E
Z
¦[η(Z; τ)[
2
¦ =
_
θ
_
w>τ
(w −τ)
2
f
w
(w)f
θ
(θ)dwdθ = 2
_
w>τ
w(w −τ)
2
e
−w
2
dw
= 2

π
_
_
_
w>τ
w
3
e
−w
2

π
dw −2τ
_
w>τ
w
2
e
−w
2

π
dw + τ
2
_
w>τ
w
e
−w
2

π
dw
_
_
, (14)
where each of the integrals in the last line is the incomplete moment of a Gaussian random variable with
parameters (0,
1

2
) of order 3, 2, and 1, respectively. These integrals can be computed numerically.
After some changes of variable and coordinate transformations, the second expectation in (13) can be
written as
E([η(µ + Z
r
+ jZ
c
; τ) −µ[
2
) =

_
φ=0
_
r≤τ
µ
2
1
π
e
−r
2
−µ
2
+2rµcos(φ)
rdrdφ
+

_
φ=0
_
r>τ
[(r −τ)
2
+ µ
2
−2µ(r −τ) cos(φ)]e
−r
2
−µ
2
+2rµcos(φ)
rdrdφ
= µ
2
+ e
−µ
2
_

τ
r(r −τ)e
−r
2
_
−4µI
1
(2rµ) + 2(r −τ)I
0
(2rµ)
_
dr, (15)
March 8, 2012 DRAFT
26
where I
0
and I
1
are modified Bessel functions of the first kind of order 0 and 1, respectively, and J
1
is
the Bessel function of the first kind of order 1 [34]. For a given τ, the fixed point solution σ

can be
iteratively computed using the integrals above.
Finally, note that if the amplitudes of the non-zero coefficients are drawn from an arbitrary distribution
G, then we can also calculate the expected value of the expressions in (15) with respect to µ ∼ G.
APPENDIX B
PROOF OF THEOREM 2
To prove Theorem 2, we derive the detection and false alarm probabilities of Architectures 1 and 2.
Since the thresholds τ
α
and τ
o
are different for Architectures 1 and 2, in what follows, σ
∗,α
and σ
∗,o
denote the fixed point solutions of SE in Architectures 1 and 2, respectively.
Architecture 1
Recall from Figure 3(a) that in Architecture 1 the non-zero coefficients in ´ x represent the final detections
and the threshold τ
α
is selected so as to achieve the desired FAP α. In this architecture, the output of
CAMP is given by
´ x = η(x
o
+ σ
∗,α
z; τ
α
σ
∗,α
), (16)
where z ∼ (A(0, I). The Gaussianity of the noise is due to the assumption of the Theorem, and it holds
in the asymptotic settings according to Theorem 1. The test statistic at bin i, i = 1, , N, is given by
[´ x
i
[
H1

H0
0. (17)
Therefore, we have
P
fa
1
= P([´ x
i
[ , = 0[H
0
) = P([σ
∗,α
z[ > τ
α
σ
∗,α
) = P([z[ > τ
α
) = e
−τ
2
α
,
where z ∼ (A(0, 1). This leads us to the following straightforward parameter tuning: τ
α
=

−ln α.
For evaluating the detection probability, define a as the square root of the power received from a target
at location i. We then have
P
d1
= P([´ x
i
[ , = 0[H
1
) = P([η(˜ x
i
; τ
α
σ
∗,α
)[ , = 0[H
1
) = P([x
o,i
+ σ
∗,α
z[ > τ
α
σ
∗,α
) (18)
= P([a + σ
∗,α
z[ > τ
α
σ
∗,α
) = Q
1
(
_
2SNR
CS,1
,

−2 ln α),
where SNR
CS,1
= a
2

2
∗,α
, σ
2
∗,α
evaluated using (4) with threshold parameter τ
α
, and Q
1
(., .) is the
Marcum-Q function.
March 8, 2012 DRAFT
27
Architecture 2
Suppose that after the estimation block we obtain the following noisy estimate of x
o
¯ x = x
o
+ σ
∗,o
z.
For Architecture 2, the decision statistic at bin i is given by
[¯ x
i
[
H1

H0
κ.
Therefore,
P
fa2
= P([¯ x
i
[ > κ[H
0
) = P([σ
∗,o
z[ > κ) = P([z[ > κ/σ
∗,o
) = e
−κ
2

2
∗,o
.
The last equality comes from the fact that [z[ ∼ Rayleigh(1/2). Thus, for a desired FAP α, the detector
threshold can be set as κ = σ
∗,o

−ln α, where as before σ
∗,o
can be computed from (4) with threshold
parameter τ
o
. The detection probability is given by
P
d2
= P([¯ x
i
[ > κ[H
1
) = P([x
o,i
+ σ
∗,o
z[ > κ) = P([a + σ
∗,o
z[ > κ) (19)
= Q
1
(
_
2SNR
CS,2
,

−2 ln α),
where SNR
CS,2
= a
2

2
∗,o
.
The proof of Theorem 2 is now straightforward. Since σ
∗,o
≤ σ
∗,α
, for the same target received power
a
2
, SNR
CS,2
≥SNR
CS,1
and, therefore, for the same FAP α, P
d2
≥ P
d1
, with equality if and only if
α = e
−τ
2
o
. Note that the above arguments are independent of the distribution G of the non-zero elements
of x
o
.
APPENDIX C
MEDIAN ESTIMATOR ERROR BOUND
To quantify the error of the median estimator as a function of δ and ρ, consider the random variable
x with pdf as in (12) and z ∼ (A(0, σ
2
). Define µ

as the median of the absolute value of the random
variable z, i.e., P([z[ ≤ µ

) =
1
2
. Since [z[ ∼ Rayleigh(σ
2
/2), we obtain µ

= σ
_
log(2).
Define = δρ and µ as the median of the absolute value of the random variable x + z, i.e.,
P([x +z[ ≤ µ) = P([a +z[ ≤ µ) + (1 −)P([z[ ≤ µ) = P([a +z[ ≤ µ) + (1 −)(1 −e
−µ
2

2
) =
1
2
.
Since 0 ≤ P([a + z[ ≤ µ) ≤ 1, using the upper bound we obtain
(1 −)(1 −e
−µ
2

2
) ≤
1
2

March 8, 2012 DRAFT
28
and therefore
µ ≤ σ
_
log
_
2(1 −)
_
.
Therefore the normalized bias introduced by the median estimator in the presence of a target is bounded
by
[µ −µ

[
σ

¸
¸
¸
_
log
_
2(1 −)
_

_
log(2)
¸
¸
¸.
Furthermore, using the inequality log
_
2(1 −)
_
< log(2), the bound in (8) follows.
APPENDIX D
GAUSSIANITY OF w FOR PARTIAL FOURIER MATRIX
We report simulation results that demonstrate the approximate Gaussianity of the noise vector w for the
partial Fourier sensing matrix. In Table I we report the p-values from the Kolmogorov-Smirnov (KS) test
[35] for N = 200. The KS test compares the empirical distribution of the input samples to the Gaussian
TABLE I
p-VALUES OF THE REAL (R) AND IMAGINARY (I) PARTS OF w FOR PARTIAL FOURIER SENSING MATRIX, N = 200.
δ 0.1 0.18 0.2 0.2 0.3 0.4 0.4 0.5 0.6 0.65 0.7 0.8 0.8 0.9 0.9
ρ 0.01 0.5 0.05 0.75 0.6 0.2 0.3 0.15 0.3 0.8 0.6 0.3 0.65 0.1 0.45
p-value (R) 0.84 0.99 0.73 0.86 0.93 0.7 0.7 0.73 0.67 0.89 0.87 0.86 0.93 0.96 0.78
p-value (I) 0.72 0.7 0.98 0.94 0.86 0.84 0.9 0.96 0.82 0.94 0.83 0.97 0.78 0.89 0.9
distribution, and the p-value measures the similarity of the input samples to the reference distribution. In
the null hypothesis it is assumed that the samples belong to the reference probability density function.
In our simulations, in all cases the null hypothesis was accepted.
We performed the KS test also for different combinations of N, δ, and ρ and obtained similar results.
ACKNOWLEDGMENTS
Thanks to Prof. J. Ender and T. Mathy from Fraunhofer FHR, Wachtberg, Germany, for making the
LabRadOr radar system available available and for technical support during the experiments. Many thanks
also to W. van Rossum from TNO, The Netherlands, for frequent discussions and constructive comments.
March 8, 2012 DRAFT
29
REFERENCES
[1] R. G. Baraniuk and T. P. H. Steeghs. Compressive radar imaging. In Proc. IEEE Radar Conf., 2007.
[2] A. C. Gurbuz, J. H. McClellan, and W. R. Scott. Compressive sensing for subsurface imaging using ground penetrating
radar. J. Signal Process., 89(10):1959–1972, Oct. 2009.
[3] I. Stojanovic, W. C. Karl, and M. Cetin. Compressed sensing of mono-static and multi-static SAR. In Proc. SPIE Algorithms
for Synthetic Aperture Radar Imagery XVI, 2009.
[4] C. Y. Chen and P. P. Vaidyanathan. Compressed sensing in MIMO radar. In Proc. Asilomar Conf. Signals, Systems, and
Comput., 2008.
[5] J. J. Fuchs. The generalized likelihood ratio test and the sparse representations approach. In Proc. Int. Conf. Image and
Signal Process. (ICISP), 2010.
[6] M. A. Herman and T. Strohmer. High-resolution radar via compressed sensing. IEEE Trans. Signal Process., 57(6):2275–
2284, Jun. 2009.
[7] Y. Wang, G. Leus, and A. Pandharipande. Direction estimation using compressive sampling array processing. In Proc.
IEEE Work. Stat. Signal Process., 2009.
[8] Y. Yu, A. P. Petropulu, and H. V. Poor. MIMO radar using compressive sampling. IEEE J. Sel. Topics Sig. Proc.,
4(1):146–163, Feb. 2010.
[9] J. H. G. Ender. On compressive sensing applied to radar. J. Signal Process., 90(5):1402–1414, May 2010.
[10] L. C. Potter, E. Ertin, J. T. Parker, and M. Cetin. Sparsity and compressed sensing in radar imaging. Proc. IEEE,
98(6):1006–1020, Jun. 2010.
[11] S. Shah, Y. Yu, and A. Petropulu. Step-frequency radar with compressive sampling SFR-CS. In Proc. IEEE Int. Conf.
Acoust., Speech, and Signal Process. (ICASSP), 2010.
[12] L. Anitori, M. Otten, and P. Hoogeboom. Compressive sensing for high resolution radar imaging. In Proc. IEEE Asia-Pacific
Microwave Conf. (APMC), 2010.
[13] A. Maleki and D. L. Donoho. Optimally tuned iterative reconstruction algorithm for compressed sensing. IEEE J. Sel.
Topics Sig. Proc., 4(2):330–341, Apr. 2010.
[14] P. P. Gandhi and S.A. Kassam. Analysis of CFAR processors in homogeneous background. IEEE Trans. Aerosp. Electron.
Syst., 24(4):427–445, Jul. 1988.
[15] H. Rohling. Radar CFAR thresholding in clutter and multiple target situations. IEEE Trans. Aerosp. Electron. Syst.,
19(4):608–621, Jul. 1983.
[16] J. Romberg E. Candes and T. Tao. Stable signal recovery from incomplete and inaccurate measurements. Comm. Pure
Appl. Math., 59(8):1207–1223, Aug. 2006.
[17] M. Elad D. L. Donoho and V. Temlyakov. Stable recovery of sparse overcomplete representations in the presence of noise.
IEEE Trans. Inf. Theory, 52(1):6–18, Jan. 2006.
[18] E. Candes and T. Tao. The dantzig selector: statistical estimation when p is much larger than n. Ann. Stat., 35:2392–2404,
2007.
[19] B. Bubacarr and J. Tanner. Improved bounds for restricted isometry constants for Gaussian matrices. SIAM J. Matrix
Anal., 31(5):2882–2898, 2010.
[20] D. L. Donoho, A. Maleki, and A. Montanari. Message passing algorithms for compressed sensing. Proc. Natl. Acad. Sci.,
106(45):18914–18919, 2009.
March 8, 2012 DRAFT
30
[21] A. Maleki and A. Montanari. Analysis of approximate message passing algorithm. In Proc. IEEE Conf. Inform. Science
and Systems (CISS), 2010.
[22] D. L. Donoho, A. Maleki, and A. Montanari. The noise-sensitivity phase transition in compressed sensing. IEEE Trans.
Inf. Theory, 57(10):6920–6941, Oct. 2011.
[23] A. Maleki, L. Anitori, Y. Zai, and R. G. Baraniuk. Asymptotic analysis of complex LASSO via complex approximate
message passing (CAMP). Submitted to IEEE Trans. Inf. Theory, 2011.
[24] M. Bayati and A. Montanari. The dynamics of message passing on dense graphs, with applications to compressed sensing.
IEEE Trans. Inf. Theory, 57(2):764 – 785, Feb. 2011.
[25] M. Skolnik. Introduction to Radar Systems. McGraw Hill, Third edition, 2002.
[26] D. L. Donoho and J. Tanner. Precise undersampling theorems. Proc. IEEE, 98(6):913 – 924, Jun. 2010.
[27] S. S. Chen and D. L. Donoho. Examples of basis pursuit. In Proc. SPIE Ann. Meeting: Wavelet Apps. Sig. Imag. Proc.,
1995.
[28] R. Tibshirani. Regression shrinkage and selection via the LASSO. J. Roy. Stat. Soc., Series B, 58(1):267–288, 1996.
[29] S. S. Chen, D. L. Donoho, and M. A. Saunders. Atomic decomposition by basis pursuit. SIAM J. on Sci. Computing,
20:33–61, 1998.
[30] B. Efron, T. Hastie, I. Johnstone, and R. Tibshirani. Least angle regression. Ann. Stat., 32(2):407–499, 2004.
[31] S. J. Kim, K. Koh, M. Lustig, S. Boyd, and D. Gorinevsky. A method for large-scale 1-regularized least squares. IEEE
J. Sel. Topics Sig. Proc., 1(4):606–617, Dec. 2007.
[32] M. Bayati and A. Montanari. The LASSO risk for Gaussian matrices. To appear in IEEE Trans. Inf. Theory, 2011.
[33] L. Anitori, A. Maleki, W. van Rossum, R. G. Baraniuk, and M. Otten. Compressive CFAR radar detection. In To appear
in Proc. IEEE Radar Conf., 2012.
[34] M. Abramowitz and I. A. Stegun. Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables.
National Bureau of Standards, U.S. Dept. of Commerce, 1972.
[35] F. J. Massey. The Kolmogorov-Smirnov test for goodness of fit. J. Amer. Statist. Assoc., 46(253):68–78, 1951.
March 8, 2012 DRAFT

2

I. I NTRODUCTION High resolution radar detection and classification of targets requires transmission of wide bandwidth during a short observation time. Wide instantaneous bandwidth necessitates wideband receivers and fast Analog to Digital Converters (ADC) that drive up the cost of a system. In applications using interleaving of radar modes in time or space (antenna aperture), multi-function operation often leads to conflicting requirements on sampling rates in both the time and space domains. Measurement time may also be a constraint in real-time applications. Recently, Compressive Sensing (CS) has been proposed to alleviate these problems [1]–[12]. CS exploits the sparsity or compressibility of a signal to reduce both the sampling rate (while keeping the resolution fixed) and the amount of data generated. In many radar applications, the prerequisite on the signal sparsity is often met, as the number of targets is typically much smaller than the number of resolution cells in the illuminated area or volume. Classical radar architectures (without CS) use well-established signal processing algorithms and detection schemes, such as Matched Filtering (MF) and Constant False Alarm Rate (CFAR) processors. CS instead relies on a nonlinear reconstruction scheme such as
1 -minimization

for reconstructing the target

scene. Most of the existing CS recovery algorithms have a threshold or a regularization parameter, that affects the quality of the reconstruction result; for the best performance it is essential to optimally tune this parameter [13]. A similar problem is encountered in classical radar detectors, where the detection threshold adapts to the environment using a CFAR scheme. These detectors first estimate the background noise and clutter surrounding the Cell Under Test (CUT) and then calculate the threshold to preserve a constant false alarm rate even in changing environments. Several CFAR schemes have been designed for different types of noise, clutter, and target environments [14], [15]. The design of CFAR-like schemes for CS radar systems has not been addressed so far in the literature, mostly due to the unknown relations between the false alarm probability/noise statistics and the parameters of the recovery algorithm. Moreover, the main focus of CS has been on the reconstruction Mean Square Error (MSE), while in radar, the primary performance parameters are the detection and false alarm probabilities. Furthermore, most of the theoretical results provide only sufficient conditions with unknown/large constants that are not useful for applications [16]–[19]. These issues have been recently addressed in a series of papers [20]–[23]. In these papers a novel CS recovery algorithm has been proposed, called Complex Approximate Message Passing (CAMP). The statistical properties of CAMP enable accurate calculations of different performance criteria such as the false alarm and detection

March 8, 2012

DRAFT

3

probabilities. In this paper, we use the features of the CAMP algorithm to design a fully adaptive CS-CFAR radar detector. Toward this goal we address the following questions. How can the detection probability be optimized against false alarm rate? How can the false alarm rate be controlled adaptively against unknown noise and clutter? How can a CS-radar be designed; in particular, what amount of undersampling is acceptable, and at what cost in terms of power? To this end, in Section II we begin with a summary of the main features of the CAMP algorithm, originally presented in [23] for the CS recovery of complex signals and previously in [20], [21], [24] for the real signal case. In Section III we propose two novel CS radar detection architectures based on CAMP. We predict the performance of these two detectors theoretically by characterizing the statistical properties of the recovered signal. Although the theoretical results on CAMP assume sensing matrices with independent identically distributed (i.i.d.) random entries, both our simulations and empirical results show that the general trend of our findings also applies to randomly subsampled (or partial) Fourier sensing matrices. Such matrices are particularly important for CS-radar range compression, where subsampling in frequency is a power-efficient way of applying CS [9], as opposed to time undersampling. From the analytical, simulation, and experimental results presented respectively in Sections III, VI, and VII, it is clear that the best architecture consists of a reconstruction stage based on CAMP followed by a detector. For this architecture, in Section IV we propose an adaptive version of CAMP followed by a CFAR processor, resulting in a fully adaptive detector. In Section VIII we derive Signal-to-Noise Ratio (SNR) plots in which the reconstruction SNR, which is input to the detection stage, is plotted against the undersampling factor (δ ) and the relative signal sparsity (ρ). Given a maximum number of expected targets in the scene, such graphs can be used to evaluate how transmitted power can be traded for undersampling and thus serve as a guide for the CS radar designer. In Section IX we draw some conclusions from our theory and experiments. II. C OMPLEX A PPROXIMATE M ESSAGE PASSING (CAMP) A. Heuristic explanation of the CAMP algorithm Consider the problem of recovering a sparse vector xo ∈ CN from a noisy undersampled set of linear measurements y ∈ Cn
y = Axo + n,

(1)

where A ∈ Cn×N is called the sensing matrix and n is i.i.d. complex Gaussian noise with variance σ 2 . In radar systems, y corresponds to the sampled received signal. Each non-zero element of the vector xo
March 8, 2012 DRAFT

(2) where x 1 N i=1 (xR )2 + (xI )2 . We first explain each component of CAMP. In the remainder of this paper. and therefore researchers have considered several iterative algorithms with inexpensive per-iteration computations.4 corresponds to the (complex) Radar Cross Section (RCS) of a target and may also include propagation and other factors normally associated with the radar equation [25]. The histogram of wt . One of the most successful recovery algorithms is 1 -regularized least squares. Hence. see [23] and the references therein. it is possible to recover xo accurately [16]. noisy estimate of xo . (i) xt is an estimate of xo at iteration t. i. where η(·. It has been proved that. For more information on these algorithms. xo. If there are k N targets. 1 y − Ax 2 given by x(λ) = arg min x 2 2 + λ x 1. i i i i respectively. [26]..i |2 represents the received power from a target at position i. [31]. we focus on a specific algorithm called Complex Approximate Message Passing (CAMP) [23]. these approaches are computationally expensive. However. also known as the LASSO or Basis Pursuit Denoising (BPDN) [27]–[29]. λ) is called the complex soft thresholding function. The tuning of τ in terms of λ is given later in (5) and will be explained in detail in Section II-B. xt → x(λ) as t → ∞. 1 is the indicator function. under certain conditions on the measurement matrix and the sparsity of xo . (ii) xt is a non-sparse. λ is called the regularization parameter.i is zero except at the target locations. derivative of η R with respect to the real part of the input. ∂η I ∂xI ∂η R ∂xR is the partial is the partial derivative of η I with respect to the imaginary part of the input. suggests that the empirical distribution of wt is “close” to a zero mean Gaussian probability density function. These methods rely on the fact that the optimization problem arg minx λ x 1 1 2 u−x 2 2 + has a closed form solution η(u. η I and η R are the imaginary and real parts of the complex soft thresholding function. This algorithm has several appealing properties for radar applications that will become clear as we proceed. and maxiter is the (user specified) maximum number of iterations. with xR and xI denoting the real and imaginary parts of xi .e. Define the “noise” vector wt = xt − xo at iteration t of CAMP. 2012 DRAFT . March 8. |xo. where · denotes the average of a vector. the number of CS measurements divided by the number of samples to be recovered. If the parameter τ is tuned properly. which is shown in Figure 1. δ = n N is called the compression factor. In this paper. The cost function in (2) is convex and can be solved by standard techniques such as interior point or homotopy methods [30]. We will more rigorously discuss the Gaussianity of wt in Section II-B. The CAMP algorithm is given in Algorithm I. λ) (|u| − λ)ej u. then xo is sparse.

Hence. Since this estimate is not sparse. A. Here for the clarity of exposition we have assumed that the algorithm uses xo in computing the noise standard deviation σt . x. σ∗ Iteration 2 3 Density 2 2 1 0 −0. Histograms of the real and imaginary parts of the CAMP residual noise signal wt at different iterations of Ideal CAMP.5 0 Data 0. In these plots. and σ = 10−3 . 1. In the rest of the paper we use the notation σ∗ limt→∞ σt .5 0 Data 0. CAMP first finds a noisy estimate of the signal called xt .5 4 Iteration 5 8 6 4 2 0 −0. In Section IV we propose a more practical scheme where σt is replaced with an estimate. τ σt ) end Output: x.5 0 −0. 2012 DRAFT . xo Initialization x0 = 0. the soft thresholding function is applied to obtain a sparse estimate xt . τ σt ) xt = η(xt . (iii) σt is the standard deviation of wt .5 Algorithm I: Ideal CAMP Algorithm Input: y. δ = 0. March 8. N = 4000.5 Iteration 10 (b) Histogram of the imaginary part of wt .5 Iteration 10 0 Data 0. Iteration 2 3 Density 2 2 1 0 −0.6. τ σt ) ∂xR ∂η I ∂xI + (xt . τ. z0 = y for t = 1 : maxiter xt = A† zt−1 + xt−1 σt = std(xt − xo ) 1 zt = y − Axt−1 + zt−1 2δ ∂η R (xt .5 0 Data 0.5 0 −0.5 0 Data 0.5 4 Iteration 5 8 6 4 2 0 Data 0.5 (a) Histogram of the real part of wt . we refer to this algorithm as Ideal CAMP. A Gaussian probability density function (pdf) is fitted to the histograms (red line).5 0 −0. Fig.

A(N ).d. Consider the following definition from [24]. While the above result has been proved for Gaussian measurement matrices in the asymptotic setting. 1 (3) Empirical studies have already confirmed that this theoretical prediction holds for other sensing matrices with i. [22]. It has been proved that in the asymptotic setting N → ∞.1 We explore this claim for the partial Fourier matrices that are of particular interest in radar applications using Stepped Frequency (SF) waveforms in Section VI-A. Using the Gaussianity of the noise in the asymptotic regime. elements other than Gaussian.The empirical distribution of n ∈ Rn (n = δN ) converges weakly to a probability measure pn with bounded second moment as N → ∞. (i) Under what conditions are the above heuristics accurate? (ii) Can we use the properties of CAMP to predict its performance theoretically? (iii) What is the formal connection between CAMP and the LASSO problem defined in (2).i. Definition II. 2012 DRAFT . drawn from a Gaussian distribution.i. [24]. Let the marginal distribution of xo converge to pX . March 8.The elements of A(N ) ∈ Rn×N are i. simulation results confirm that it is still accurate even for medium problem sizes with N ∼ 200 and different classes of sensing matrices. In the asymptotic setting. [32]. Theorem 2 enables us to answer the second question as well.1. A sequence of instances {xo (N ). and let xt (N ) be the estimate provided by the CAMP algorithm. SE tracks the evolution of the standard deviation of the noise σt across iterations. . The first question is accurately discussed in [22]–[24]. State evolution: A framework for the analysis of CAMP There are three important questions that have not been answered yet. n(N )} be a converging sequence. [24] Let {xo (N ). while δ is fixed.6 B. A(N ). the above heuristics are correct. The empirical law of wt (N ) = xt (N ) − xo (N ) converges to a zeromean Gaussian distribution almost surely as N → ∞. .d. the value of the standard deviation at time t + 1 is calculated from σt according to the following equation: 2 σt+1 = Ψ(σt ). n(N )} is called a converging sequence if the following conditions hold: .The empirical distribution of xo (N ) ∈ RN converges weakly to a probability measure pX with bounded second moment as N → ∞. Theorem 1. one can predict the performance of CAMP through what is called the “state evolution” (SE). [20]. and let σt denote the standard deviation of wt .

therefore maximizing the CAMP reconstruction SNR. and the expectation is with respect to the two independent random variables X ∼ pX and Z . the number of non-zero coefficients in xo . to answer the third question we observe that there is a nice connection between the CAMP and LASSO algorithms. τ σ∗ )+ (X + σ∗ Z. Also. the CAMP input and output SNR are defined here respectively as SNRin = a2 /(nσ 2 ) 2 and SNRout = a2 /σ∗ . τ σt ) − X|2 equals the MSE of the estimate xt after applying the soft thresholding function.. τ σ∗ ) 2δ ∂xR ∂xI . if τ is chosen to satisfy λ τ σ∗ 1− ∂η R ∂η I 1 E (X + σ∗ Z. τ σt ) − X|2 . Let k be the number of targets. as the number of measurements decreases. It has been proved [23] 2 that the function Ψ is concave. for which σ∗ is minimized. for a given σ 2 . Consequently. The following two consequences of (3). 2012 DRAFT . (4) are particularly useful for the radar application: (i) The SE framework establishes the input/output relation for CAMP. say τo . δ (4) Z ∼ CN (0. III. amongst other quantities. fixed pX and σ .. In fact. the CAMP threshold τ via σ∗ . (4) has at most one stable fixed point σ∗ . i. 1). Moreover. the output noise 2 power σ∗ is the sum of the actual system noise (σ 2 ) plus a noise-like component caused by the reconstruction itself ( 1 MSE). pX .e. As is clear from the figure.2 (ii) The fixed point σ∗ depends on τ as well as on δ .7 where 1 Ψ(σt ) = σ 2 + E |η(X + σt Z. minimizing the reconstruction error δ 2 also minimizes σ∗ . CS TARGET DETECTION USING CAMP We now propose two CS target detection schemes based on CAMP and evaluate their performance as measured by their Receiver Operating Characteristic (ROC). and σ . Appendix A provides the calculations involved in the Ψ function for a given distribution pX . i. if σt is the standard deviation of the noise at iteration t. CAMP converges to this fixed point exponentially fast (linear convergence according to optimization literature). Note that the output (reconstruction) SNR depends on. and therefore the iteration (3). then E |η(X + σt Z. According to [23]. Define ρ = 2 k n. Particularly. March 8. (5) then in the asymptotic setting CAMP with threshold τ solves the LASSO problem (2) with parameter λ. there is a value of τ . and define G as the distribution of the non-zero For a given target received power a2 . both the optimal threshold τo and the corresponding output noise standard deviation increase. A more rigorous definition of SNR is given in Section V. Finally. Figure 2 exhibits the dependence of σ∗ on τ for two distinct values of δ and for a fixed problem instance.e.

Gaussian entries. 2012 DRAFT . 1). and σ are known. xo (N ). In this section we assume that k .1. The two architectures we consider in this paper are displayed in Figure 3.35 0.i. Clearly. black line). τα is the only value of τ for which (6) holds.2 (dashed. According to [32] we know that 1 lim N →∞ N − k N 1{xi (N )=0.2 σ* 0.5 δ = 0. the measurements y are input to a recovery algorithm (here CAMP or equivalently LASSO). Note that the connection between τ and the detection probability is not clear yet. Architecture 1 seems a natural choice for a CS radar detector. i=1 (6) almost surely. In Section IV we will investigate the realistic scenario where these parameters are not known a priori and describe how to implement the proposed architectures in this case. w(N ) 1{xi (N )=0. then 1 lim N →∞ N − k N √ − ln α.23.xi (N )=0} = α. Roughly speaking. the threshold parameter τ in CAMP (or equivalently the regularization parameter λ in LASSO) controls the false alarm probability α and detection probability (Pd ) of the algorithm.6 0.6 δ = 0.3 0. i=1 2 The uniqueness is also clear from the above equation.8 0. If A(N ).d. In this architecture.45 0. compression factors δ = 0.5 1 1. Proof: Define z ∼ CN (0. Consider the CAMP iteration with threshold τα = is a converging sequence and x(N ) is the fixed point of CAMP.5 τ 2 2.5 3 Fig. elements of xo . σ 2 is the variance of the system noise. G. Fixed point σ∗ versus threshold τ for Ideal CAMP with σ = 0.6 (solid.xi (N )=0} = P(|σ∗ z| > τα σ∗ ) = e−τα . Proposition III. As before. red line) and δ = 0.55 0. Also. the above proposition states that the number of false alarms are on the order of (N − k)α. The sensing matrix A has i. We will discuss this issue later.25 0. This algorithm returns a sparse vector x with the non-zero values being detections. 2. March 8. Since any sparse recovery algorithm is intrinsically a detection scheme.4 0.

Then this noisy estimate is fed to a detection block that controls the False Alarm Probability (Pf a or FAP). Theorem 2. it is clear that in the asymptotic setting this choice of κ results in the false alarm probability α as derived in Proposition III. Furthermore. the noisy signal x can be fed to a detection block with fixed threshold κ = σ∗ − ln(α). As will be clarified later. 3. then Pd. non-CS radar approach. where a noisy estimate of the signal is computed.9 (a) Architecture 1 Fig. Similar to classical estimation procedures. Under these assumptions. we introduce Architecture 2. We first use CAMP to obtain a noisy. which provably hold under the conditions specified in Theorem 1. If Pd. is used to obtain a noisy estimate of the signal with optimum SNR. 2012 DRAFT . Furthermore the equality is satisfied at only one specific value α = e−τo . non-sparse estimate of the signal x = xo + w. (b) Architecture 2 The second architecture is inspired by the standard.2 . which is used to control the false alarm rate. From the Gaussianity of w. radar detectors comprise two stages: an estimation stage.1. Set the probability of false alarm to α for both Architecture 1 that uses τα and Architecture 2 that uses τo in CAMP. the following theorem can be derived. from a detection perspective.1 and Pd. Suppose that the Gaussianity of wt holds. and σ . Block diagrams of the proposed architectures for CS-radar detection. the goal of the first stage is to minimize σ∗ by choosing the optimal CAMP threshold τ = τo .1 ≤ Pd. a Matched Filter (MF). Let τo be the optimal value of τ that leads to the minimum σ∗ and that this optimal value is unique and can be computed using (4). The above theorem is proved in Appendix B by comparing the detection and false alarm probabilities 2 March 8. G. Architecture 2 is much more appropriate for practical radar applications. Architecture 2 outperforms Architecture 1. matched to the transmitted waveform. Once σ∗ is minimized or equivalently the SNR is maximized. Inspired by this philosophy and by the properties of CAMP. Usually. followed by a detection stage. since all of the parameters can be optimized and estimated efficiently even without prior knowledge of k . Most commonly.2 are the detection probabilities of the two schemes.

and therefore Pd. a lower SNR. Therefore σ∗ (and hence the CAMP reconstruction SNR) also changes with Pf a . as explained later in Section VI.3 3 The difference in performance between the two architectures depends significantly on the values of (δ. σ∗ is fixed to its minimum along the ROC curve. MC 10 10 Pfa −2 10 0 Fig.2 increases with increasing FAP. The dots are the results of MC simulations using Ideal CAMP.i. σ) and on the sensing matrix in use.6 Arch. 2.. the probability of detection.e. The theoretical ROC curves are verified by Monte Carlo (MC) simulations (dots). This explains why Pd. In Figure 4 we can also observe an interesting characteristic of Architecture 1 in the region around Pf a = 0.8 0. any other choice of τ will lead to a higher σ∗ . 4. in Architecture 2.4 and then decreases again as the Pf a goes to 1. MC Arch. Example ROC curves for Architectures 1 and 2 are shown in Figure 4.9998 0. decreases as Pf a increases above this value.2 . However. i.1 and σ 2 = 0. ρ. Instead. March 8.1 . 1. Theoretical Arch. Comparison of the ROC curves for Architectures 1 and 2 at δ = 0. From this plot it is also possible to observe that there exists only one value of FAP (α = 0.2 0 −6 10 −4 1 0.4 0. since in Architecture 2 the CAMP threshold is designed to minimize the output noise variance. and it is not constant along the ROC curve.1 = Pd. Theoretical Arch. In the simulations we averaged over 10000 realizations of the Ideal CAMP algorithm given in Algorithm I for several values of Pf a ranging from 10−1 to 10−5 . The solid and dashed lines are obtained using the analytical equations derived in Appendix B.05. Pd.4 0.6. Gaussian entries. To understand why.6 Pd 0. intuitively it is clear that.2 0. and therefore a lower Pd . for the same Pf a . The distribution G of the non-zero coefficients of xo is chosen such that all non zero components have the same amplitude (equal to 1) and phase uniformly distributed between −π and π. of the two architectures.22) where Pd. The sensing matrix for MC simulations had i. 2. recall that in Architecture 1 the CAMP threshold τα varies with Pf a .d.4 (the zoomed area). 2012 DRAFT . ρ = 0.10 1 0.1 reaches its maximum at around Pf a = 0. The solid and dashed lines represent the theoretical predictions based on the SE equation. 1.

and G (to derive the theoretical fixed point) and xo (to run Ideal CAMP). i. in Architecture 2. namely how to 2 estimate. ln 2 (7) This estimator is unbiased if xo = 0. σ∗ ). σ∗ 2 ln 2 (8) This proposition is proved in Appendix C. for small values of . in very sparse situations.11 IV. However. Proposition IV.e. and both the CAMP and detector parameters must be estimated from the CS measurements y. efficiently and accurately.e.d.. So. Second. in the asymptotic setting at each iteration of CAMP we have xt = xo + wt . since we only have access to x. when xo = 0. First. the distribution G of the non zero components of xo does not play any role. σ . (i) How to estimate σt without knowing xo . (iii) How to replace the fixed threshold κ for the detector in Architecture 2 with an adaptive threshold to maintain a CFAR. There are three main issues that have to be settled. i. The first question can be answered in several different ways. The main advantage of this scheme is its robustness to high SNRs. It is important to note several interesting properties of the upper bound (8). one that does not require any prior information and that adapts to changes in noise level.i ∼ (1 − )G(x) + (1 − )δ0 (x) with i. the error is proportional to the sparsity level.1. we consider the performance of the median estimation in the asymptotic setting. To see this. in the presence of targets. The estimate of σt provides an approach to answer the second above question as well. However. (ii) How to compute the optimal value τo for Architecture 2.i. For instance.. the optimum CAMP threshold τo that minimizes σ∗ . We will refer to this algorithm simply as CAMP or Median CAMP. Assume that the elements of xo are distributed as xo. we estimate the median of |wt | as the µ that satisfies P(|xt | > µ) = 1 . in practical radar systems such information is not available. without the SE. a practical implementation of Algorithm I is obtained by replacing σt with the estimate in (7). 2012 DRAFT . = δρ t 2 1 and wi ∼ CN (0. The following theorem provides an upper bound on the i 2 deviation of the estimated median from the true median. hence. The goal is to estimate the median µ∗ of |wt |. in this paper we use the median to estimate the standard deviation via σt = 1 median(|xt |). when the noise level and xo are unknown. In this section we demonstrate how these issues can be handled in practice and propose a fully adaptive scheme. (7) is a biased estimator.. However. we have assumed we know exactly ρ. A DAPTIVE CAMP CFAR RADAR DETECTOR So far. The error of the estimated median is bounded above by |µ − µ∗ | | ln(1 − )| √ ≤ . i. As mentioned above. | ln(1 − )| ≈ and. Suppose that we know March 8.e.

We now explain how to set τmax . 2NG guard cells immediately adjacent to the CUT are excluded from the CFAR window. say xo. Given a step size δτ . In CFAR schemes the Cell Under Test (CUT). The great advantage of CFAR schemes is that they are able to maintain a constant false alarm rate via adaptation of the threshold to a changing environment. At the first iteration ( = 1 and t = 1) the CAMP algorithm is initialized with x0 = 0 and z0 = y. Hence. The general form of a CFAR test is H1 X H0 βY.i . Starting from τmax . and therefore the entire process is very fast. and the only input variables are y and A. the noise statistics are not known in advance. 2012 DRAFT . we have L estimates {σ∗ }L . Using the calibration equation (5) with λ = λmax and σ∗ = σ0 . The 2Nw cells (CFAR window) surrounding the CUT are used to derive an estimate of the local background and they are assumed to be target free. β is a threshold multiplier. The optimum estimated threshold τo is chosen as the one that minimizes the estimated CAMP =1 2 output noise variance σ∗ . and we have x = AH y. Using the solution of CAMP at the previous iteration − 1 as initial value for τ = τ . we can compute an estimate of τmax .12 or can estimate τmax such that τo < τmax . where AH is the Hermitian of the matrix A. xL ] of size N × L. since both the noise variance σt the threshold τo are adaptively estimated inside the algorithm itself. We will refer to this algorithm as Adaptive CAMP. we define a sequence of thresholds τ = {τ }L such that τ1 = τmax and τ = τ =1 −1 − δτ . Consider now the LASSO problem in (2). · · · . After L iterations. where each column contains the CAMP solution for a given τ . but it also results in a more accurate estimate of τo . x2 . at each new iteration CAMP is initialized with x0 = x −1 and z0 = z −1 . In practice. we have a matrix of solutions X = [x1 . It is clear that δτ specifies the trade-off between computational complexity and the accuracy of the algorithm in estimating τo . and Y is also a random variable function of the cells in the CFAR March 8. Suppose that σ0 is estimated from this vector. (9) where the random variable X represents some function (generally envelope or power) of the CUT. It is well known that for λ > λmax = AH y ∞ the only solution is the zero solution. Decreasing δτ increases the number of points L needed to span the same τ search region. Commonly. in classical radar detectors a CFAR processor is employed. Also. In Appendix B we show how to set the fixed threshold κ to achieve the desired FAP when the noise variance σ 2 is known. CAMP needs only a few iterations to converge to the solution. however. is tested for the presence of a target against a threshold that is derived from an estimated clutter plus noise power. It remains now to establish how to replace the fixed threshold κ with an adaptive one for Architecture 2.

we provide a definition of SNR for the CAMPbased CS radar system. V. in Architecture 2 replacing the fixed threshold detector with a CFAR detector should provide comparable results as in classical CFAR without CS. Hence. Instead. Block diagram of the Adaptive CAMP CFAR detector. We observed in (4) that the variance σ∗ of the noise present in the CAMP estimate depends on the 5-tuple (δ. pX . Pf a ) curves. ρ. the signal x contains all non-zero samples. z ∼ CN N (0. xo. In fact. window [xo. the noise distribution should be known to design a CFAR scheme. xo. For the properties of CAMP. Let A+ ∈ CN ×N be the measurement matrix for the case δ = 1.i−Nw −NG . Y is the average of the cells in the CFAR window [14]. A conventional radar processor feeds the measurement vector y = A+ xo +σz. σ and for a specific 2 Architecture (that fixes τ ). then estimation of the noise characteristics would be very difficult. Clearly. but also to compare the novel CS based architectures to more classical ones for 2 which the performance is known. and it is modeled as the sum of targets plus Gaussian noise. · · · .i+Nw +NG ]. can vary significantly with δ and ρ. τ. If the CAMP estimate x were input to a CFAR processor. · · · .i+NG +1 . since many samples in x are identically zero. σ∗ . Therefore it is important to relate the performance of a given CS scheme to that of MF-based classical systems. This definition will be useful not only for understanding the performance of the proposed detectors. A block diagram of the Adaptive CAMP CFAR detector based on Architecture 2 is shown in Figure 5. In the well known Cell Averaging CFAR (CA-CFAR) detector. i. one has to know the relation between the CFAR threshold multiplier and Pf a so that β can be adjusted to maintain Pf a constant during the observation time.i−NG −1 . where the SNR after the MF is uniquely determined for a given input SNR and integration time. even for fixed pX . this can be directly input to a conventional CFAR processor. D EFINING THE SNR FOR CS CAMP Before proceeding to the simulations and experiments. to a MF to obtain a noisy March 8. and therefore the (Pd . no undersampling. 2012 DRAFT . I).. 5.e. σ ). xo.13 Fig. This is uncommon in classic radar systems.

with the above assumptions made.e. With this assumption. 2012 DRAFT .5 In the case of partial Fourier sensing matrix. We assume these (ideal) conditions are satisfied when computing the MF SNR. if we normalize (4) by a2 .e. Once CAMP has converged we have access to both the sparse estimate x and its noisy version x.14 estimate xM F = xo + σz of the target received signal with optimum SNR. The constraint of keeping the received target power equal for different amounts of undersampling enables us to evaluate the changes in SNRCS due to reconstruction and not due to a reduction in total signal power. then we keep the first term on the right hand side fixed (which is equal to SNRM F ) and observe the change in SNRCS produced by the reconstruction error term (1/δ MSE). x = xo + σ∗ z. The elements of a partial Fourier matrix are given by ai. March 8. we define the SNR at the output of MF or CAMP respectively as SNRM F = a2 . As stated before. In other words. 4 In a multiple target scenario. we can derive the CAMP output SNR as a function of the equivalent MF output SNR. Indicating with a2 the target received power at bin i. Note that. σ2 a2 .4 CAMP enables us to define SNR in a similar way.. · · · . the MF SNR is optimum and independent of the number of targets as long as each target is exactly on a grid point and the matrix A+ is orthogonal. n and j = 1. described later in Sections VI and VII. N . the SNR that we would obtain without compression using a MF. with partial Fourier matrices. dividing both sides of (4) by a2 . 5 This is achieved by using sensing matrices with unit column norms in all simulations. i. for which SE provably applies. this is achieved in practice by dividing the total available transmit power over the subset of transmitted frequencies. [11]. An n × N partial Fourier matrix is obtained from an N × N discrete Fourier transform matrix by preserving only a random subset n of the original N matrix rows. · · · . we replace the Gaussian sensing matrices.j = e−j2π(i−1)(j−1)/N . SNRM F represents an upper bound on the highest SNR that can be obtained from CAMP. [33]. for i = 1. S IMULATION RESULTS In this section we investigate the performance of median and Adaptive CAMP using MC simulations and compare it with the theoretical results obtained from the SE. 2 σ∗ SNRCS = (10) In the remainder of the paper we assume that the total transmitted (and received) power is independent of δ . i. VI. which are of particular interest in radar applications transmitting Stepped Frequency (SF) waveforms [9]. Moreover. even in medium size problems x can be accurately approximated 2 by the sum of the true target vector plus white Gaussian noise with variance σ∗ .

In these plots σ 2 = 0. A. The sensing matrix is partial Fourier.7 (b) Histogram of the imaginary part of w. for τ = τo . ρ = 0. ρ = 0.1 2 δ = 0. Histograms (bars) of the real and imaginary parts of the the noise signal w for different combinations (δ.5 1 δ = 0.7. However. To obtain the histograms we used CAMP with a fixed (not necessarily optimal) threshold τ and we fixed σ = 0.2 0 −0.2 0 Data 0. ρ = 0. ρ = 0. March 8.1 2 δ = 0. Accuracy of the SE We investigate the accuracy of the SE by comparing the theoretical results obtained from (4) with simulations results obtained using the Ideal CAMP algorithm. The following two remarks can be made: (i) SE does not predict the performance of CAMP accurately for partial Fourier matrices.2 4 Density 6 4 2 2 0 −0.15 δ = 0. In Figure 6 we show a few examples of the empirical distribution of w at convergence for different combinations of δ and ρ. 6. 2012 DRAFT .2 0 Data 0. red line).2. Fig.2 4 Density 6 4 2 2 0 Data 0.7 0 −0.2 0 −0.5 0 −2 0 Data 2 1 0.2 0. δ = 0.5 δ = 0.2 0 Data 0.2 0 −0.2 (a) Histogram of the real part of w.1. ρ = 0.8. ρ = 0.5 0 Data 0. B.5 1 δ = 0. ρ = 0.4 −0. See Appendix D for the p-values of Kolmogorov-Smirnov test.2. We further investigate its Gaussianity using the Kolmogorov-Smirnov test in Appendix D.2. the value of σ∗ for Fourier is not very different from the value for Gaussian.1 and N = 4000. we study the behavior of σ∗ for the case of a partial Fourier sensing matrix and investigate how it deviates from the theoretical case of a Gaussian sensing matrix.5 0 Data 0.9. ρ) using CAMP with threshold τ = 1. A Gaussian pdf is fitted to the histograms (solid.5 δ = 0.2.9. Simulation results confirm that the Gaussianity of the noise vector is preserved for partial Fourier matrices as well.4 0 −0. ρ = 0. for which the SE applies. Gaussianity of w using partial Fourier matrices In this section we investigate the Gaussianity of the reconstruction noise vector wt for a partial Fourier sensing matrix using MC simulations.7. Moreover. Figure 7 compares σ∗ obtained from Ideal CAMP for the case of complex Gaussian and partial Fourier sensing matrices with the theoretical one from the SE.5 0 −1 0 Data 1 1 0.

the columns of the matrix become more correlated.3 0. This is shown in Figure 9. blue line shows the analytical σ∗ computed from (13) in Appendix A.5 3 0. i. δ → 1. as the number of measurements increases.5 2 τ 2. However. as is clear from this figure and confirmed by the upper bound provided in Appendix C. and different δ.5 2 τ δ = 0. σ∗ versus τ from Ideal CAMP for both complex Gaussian and partial Fourier sensing matrices. where we plot the ROC for Architecture 1 using both Ideal and Median CAMP.27 0.26 0.28 0. The solid.1 0. N = 4000. This is related to the March 8. for the value of δ used here.27 0.25 1. Furthermore. Effects of the median estimator in CAMP We investigate the effect of replacing the true σt with the median based estimate σt from (7) when xo = 0. From this figure we also observe that.29 0.27 0.3 0.5 3 Fig.e.3 0.5 τ 2 2. (ii) As δ → 0 the predictions of SE become more accurate for Fourier. Gaussian CAMP Ideal. For a fixed FAP (that fixes τα ). Architecture 1 performs much better when using the partial Fourier sensing matrix as compared to the Gaussian one.28 0. therefore resulting in losing both detection performance and the CFAR property (since the output Pf a is underestimated).i.2 2.29 σ* 0.d. In Figure 8 the estimated output noise standard deviation is shown for both Ideal and Median CAMP. We observe that the estimate σ∗ deviates from the Ideal CAMP case because of the bias introduced by the estimator.3 0.26 0.29 0.05. the deviation diminishes as = δρ decreases. In Architecture 1. This is because the soft thresholding function in CAMP uses the parameter τα σ∗ . the bias will have the effect of increasing the overall threshold.5 2. 2012 DRAFT .28 σ* 0.5 3 1 1. overestimating σ∗ will result in a loss in performance compared to the case of using Ideal CAMP. C. The empirical curves are obtained by averaging over 100 MC samples for σ 2 = 0. Fourier Theoretical SE σ* δ = 0.26 0. entries.28 σ* 0.16 δ = 0. 7.29 0. and hence the true behavior deviates from the SE that holds for matrices with i.26 2 τ δ = 0. ρ = 0.05..27 0.5 3 1.05 CAMP Ideal.

The curves are obtained by averaging over 100 MC realizations for σ 2 = 0.95 0. This implies that.5 2 τ 2.29 0.1 0. 2 variation of the noise variance σ∗ .05.27 0. Fourier σ* δ = 0.28 0.5 2 τ δ = 0.85 0.5 3 Fig.29 0.28 σ* 0.1. 2012 DRAFT . for decreasing FAPs. 9.5 3 1. Here N = 1000. 8.5 τ 2 2. Fourier CAMP Ideal.6 0. March 8.3 0. Gaussian CAMP Adaptive.25 1.3 0. with the threshold τα along the ROC curve.3 0. δ = 0.5 2. (b) Partial Fourier sensing matrix ROC curves for Architecture 1 using Ideal and Median CAMP.28 σ* 0. N = 4000. and various δ. in the partial Fourier case the SNR in Architecture 1 deviates much less from the optimum SNR that is achieved for τ = τo . and therefore the SNR.27 0.4 Pd Median CAMP Ideal CAMP −4 0. As can be seen from Figure 8.05 CAMP Adaptive. in the partial Fourier case as δ increases the variance curve becomes flatter than in the Gaussian case.75 10 −4 Median CAMP Ideal CAMP 10 −3 10 10 −3 10 −2 10 −1 Pfa 10 Pfa −2 10 −1 (a) Gaussian sensing matrix Fig.5 3 0. a2 = 1.29 0.9 Pd 0.26 0.29 σ* 0.05.2 2.6.8 1 0.27 0.27 0. Gaussian CAMP Ideal.26 2 τ δ = 0. 1 0.28 0.26 0.26 0.5 3 1 1.05 (corresponding to a MF SNR = 13dB). Output noise standard deviation versus τ for both Ideal (σ∗ ) and Median (σ∗ ) CAMP for complex Gaussian and partial Fourier sensing matrices.17 δ = 0.8 0. ρ = 0. and σ 2 = 0.3 0. ρ = 0.

8dB for the Gaussian sensing matrix and 12.18 1 0.9 Ideal CAMP CFAR Adaptive CAMP FT Adaptive CAMP CFAR Ideal CAMP FT Theoretical CA−CFAR −4 Pd Pd 0. and (d) a fully adaptive scheme consisting of Adaptive CAMP followed by CA-CFAR processor.8 0. Therefore .in the basic analysis of CFAR schemes.6. a2 = 1.95 0.7 10 10 −3 0. δ = 0. (b) Ideal CAMP in combination with the CA-CFAR detector. D.05 (corresponding to a MF SNR = 13dB). then they cause a rise in the adaptive threshold. 10. One loss is due to the use of Adaptive. We also show the theoretical curve of a CA-CFAR processor with the same window length and SNR = 11. For the CFAR processor. The SNR can be estimated March 8. Here N = 1000.95 0.85 0.9 0. to reduce the interference losses encountered in CA-CFAR. Adaptive CAMP is followed by a CA-CFAR processor preceded by a Square Law (SL) detector (recall Figure 5). (c) Adaptive CAMP with ideal (fixed threshold) detector. To keep the discussions concise.1. Adaptivity imposes extra losses on the system. Pf a ) performance of the Adaptive CAMP CFAR detector. A second loss is caused by the CFAR processor and its estimate of the noise standard deviation. it is assumed that there are no targets present in the CFAR window. such as Order Statistic (OS) CFAR [15]. This means that there is an error in estimating τo .75 0. In Figure 10 we show the ROC curves for Architecture 2 obtained using: (a) Ideal CAMP with ideal (fixed threshold) detector. ρ = 0. Following a similar approach. (b) Partial Fourier sensing matrix ROC curves for Architecture 2 for different levels of adaptivity. several dedicated CFAR schemes have been proposed in literature.8 0. and σ 2 = 0. Adaptive CAMP CFAR detector performance We investigate now the (Pd .4dB for the partial Fourier sensing matrix. we consider the case of multiple but non-interfering targets. It is a well known fact that.75 −4 10 Ideal CAMP CFAR Adaptive CAMP FT Ideal CAMP FT Adaptive CAMP CFAR Theoretical CA−CFAR 10 −3 10 −2 10 −1 Pfa 10 Pfa −2 10 −1 (a) Gaussian sensing matrix Fig. if one or more targets are present in the CFAR window. we use a CFAR window of length 20 with 4 guard cells.85 0. This scenario is referred to as the non-interfering targets scenario.instead of Ideal CAMP. thus possibly masking the target in the CUT that has yet to be detected. 2012 DRAFT . FT denotes the use of an (ideal) fixed threshold detector. For the case of interfering targets. In the simulations for Architecture 2. we do not pursue these directions and leave them for future research.

which is derived from a plot like the ones shown in Figure 7 for the case δ = 0. In the experiments. Each frequency is transmitted during 0. resulting in 37. A more detailed description of the radar system and the setup is given in [33]. 2012 DRAFT . as predicted by Theorem 2. For the partial Fourier sensing matrix. the expected types of target and environment).95MHz. In the Nyquist case (that represents unambiguous mapping of ranges to phases over the whole bandwidth) we transmit N = 200 frequencies over a bandwidth of 800MHz.25. by comparing the curve of Adaptive CAMP plus CFAR with the theoretical one of a CA-CFAR processor (without CS) with equal parameters. E XPERIMENTAL R ESULTS To verify our theoretical and simulated results. A. By comparing Figures 9 and 10 it can be seen that. March 8. thus implying a bandwidth of Bf = 1.5 and 0.5m unambiguous range (∆R). meaning that the TX signal consists of a number of discrete frequencies. This is due to the fact that. The achievable range resolution is therefore δR = 18. Architecture 2 always outperforms Architecture 1. the window length depends on the specific application (e. This is the well-known CFAR loss [14]. Transmitted waveform We use a SF waveform.85. in the fixed threshold case. The corner reflectors are about 5m apart in range. VII. we use five stationary corner reflectors of different sizes (RCS) as targets. For the Gaussian sensing matrix the optimal threshold for Architecture 2 (using Ideal CAMP) is computed using the SE.512µs. Furthermore. which can be controlled by changing the CFAR window length. the value of τo at which the minimum σ∗ occurs is very close to the true optimal τo (see Figure 8).19 during simulations. and sequential frequencies are separated by ∆f = 4MHz.. although σ∗ is biased. we performed several experiments at Fraunhofer FHR using the software defined LabRadOr experimental radar system. The digital SF waveforms were designed to acquire a set of CS measurements for δ = 0. In general. It is clear from this figure that Adaptive CAMP introduces almost no loss in the detection performance of Architecture 2. For each transmitted (TX) waveform 300 measurements (with the same set-up) were performed. the threshold in Ideal CAMP is set to τo = 1. or it can be derived using the procedure described later in Section VIII.6.g.75cm. The main loss instead is introduced by the adaptive CFAR detector. we observe that the CA-CFAR detector performance seems independent of the fact that the input to the detector is obtained by running CAMP instead of a conventional MF.

In the CS case. .05. . N . Figure 11 exhibits the signals reconstructed using Architectures 1 and 2 in addition to the MF. there is a leakage of target power into neighboring range bins both for MF and CS. Therefore. In this figure the five corner reflectors are indicated as T1–T5. For the MF. 1 CAMP Arch.1. τα was set using a FAP = 10−4 . The subset of transmitted frequencies is chosen uniformly at random within the total transmitted bandwidth. Also. ri = r0 + i ∆R/N . n. . irrespective of n. each range bin. i = 1.20 10 log(amplitude) 4 CAMP Arch. 2012 DRAFT . This enables us to analyze the effects of undersampling and transmitted power separately. In the experimental measurements. . irrespective of the number of transmitted frequencies. 50. and MF. Reconstructed range profile using: CAMP Architectures 1 and 2. After reception and demodulation. to evaluate the performance of the detectors at March 8. km = 2πfm /c is the wave number. the number of TX frequencies is reduced from N to n (n < N ). and the n samples ym of the compressed measurement vector y are given by 1 ym = √ n N e−j2km ri xo. We assume that the same total power is transmitted. for CS. since the SNR is very high for all targets (in all cases above 20dB). so that the total transmitted energy (PT × n/Bf ) is the same in all cases. CFAR CAMP adaptive detector performance We show now the ROC curves obtained using CAMP-based detection on the measured data. due to the fact that targets are not exactly on Fourier grid points. This means that when the number of measurements is reduced by δ . and n = 100. . . 0. corresponding to δ = 0. N = 200. δ = 0. we fix N = 200. i=1 (11) where a(m) is the mth row of the partial Fourier matrix. 11.i = a(m) xo . In our experiments. m = 1.5 and ρ = 0. For Architecture 1. maps to n phases proportional to the n transmitted frequencies fm = f0 + m∆f . the power per transmitted frequency PT is 1/δ times higher than in the Nyquist waveform case. and c is the speed of light.5. .25. . 2 MF T1 T2 T5 T3 T4 10 3 10 2 10 1 0 5 10 15 20 Range [m] 25 30 35 Fig. 0. B. the sensing matrix A can be represented as a partial Fourier matrix.

CAMP will perform better for larger δ .25 ROC curves for Architecture 1 and Architecture 2 using both fixed threshold (FT) and CA-CFAR detectors. while the total transmitted power remains fixed for δ = 0.5 Fig. 2 CA−CFAR Arch.5 and 0. T3.8 T1 1 0.5 and 0. Therefore.6 0. Since a loss in SNR translates directly into a loss in detection probability.25. March 8. When Architecture 2 is followed by a CA-CFAR processor. In Architecture 1. In these plots. 1 10 Pfa −1 0. 2 CA−CFAR Arch. for δ = 0. indicated as T1.2 and 26dB. 10. 12. CAMP recovery is followed by either a fixed threshold (FT) detector or a CA-CFAR processor that uses four guard cells and a CFAR window of length 20. the FAP was estimated by excluding the range bins corresponding to the target locations plus four guard cells. For Pd estimation. As observed in the simulated results in Figures 7 and 8. we used the detection at the location of the highest target peak. in agreement with our theoretical findings. 2012 DRAFT . The curves correspond to targets T1. 16. the highest FAP that can be estimated is equal to δ . respectively. as expected. the reconstruction noise variance σ∗ increases.2 0 T4 10 −2 0 10 10 −3 10 −2 10 Pfa −1 10 0 (a) δ = 0.25.21 1 0. we added additional white Gaussian noise (with σ = 500) to the raw frequency data samples. and thus the CAMP SNR decreases. From Figure 12 we observe that. for a given FAP.4 0. 2 FT Arch. medium SNR values.6.6 0. since the sparse estimated signal x cannot have more than n out of N non-zero coefficients. T3 and T4 in Figure 11. (b) δ = 0.2 10 −3 0.4 Arch. resulting in a MF output SNR of 17. Figure 12 shows the ROC curves for three of the five targets with different SNRs. 2 FT Arch. 1 T1 T3 T3 P Pd T4 d 0. from the closest corner (T1) to the farthest one (T5). the reconstruction SNR for each target depends on the architecture used in addition to the compression factor δ . For each target the ROC curve is shown using both Architecture 1 and Architecture 2. there is a loss in detection performance compared to the fixed threshold case. Recall that in CAMP the reconstruction SNR is the ratio of the estimated target power to the system 2 plus reconstruction noise power (σ∗ ).8 Arch. the detection probability of Architecture 2 with the fixed threshold detector is always higher than that of Architecture 1. as the number of measurements is reduced. 14. In Architecture 2. and T4 in Figure 11.2.

as shown in Figure 14(b) for a few sets of points in the map.e. 7 For sensing matrices other than Gaussian. since the entire signal is used in the noise estimation and the threshold τα is fixed.e. Given the detection range. the SNR map of CAMP can be obtained via simulations using the Ideal or Adaptive CAMP algorithms. an estimate of the expected number of targets should be made. if we consider Architecture 2 and we assume that the received power is equal to a2 for all targets. in contrast. Observe that in both cases the curves are equal for the same kmax up to approximately 6 This choice of target amplitudes distribution provides a lower bound on the SNR performance. i.i. the two proposed architectures perform very similarly. 2012 DRAFT . target RCS and system noise. i. it can not adapt to local variation of noise level. Gaussian entries and the curves are obtained using the theoretical SE. For Figure 13 we used the analytical equations from Appendix A.. we can vary the number of CS measurements n to obtain several values of δ = n/N and corresponding ρ = kmax /n. k could be set based on the maximum expected number of targets.7 It is clear that CS radar performance depends.22 At very low and very high SNRs. we obtain a curve that shows how power (SNRCS ) and undersampling (δ ) can be traded against one another. By evaluating the SNR at these points. Architecture 2 always outperforms Architecture 1. Computing the ratio a2 /σ∗ we obtain the CAMP output SNR map. To design the system for the worst-case scenario.. an example of which is shown in Figure 13 for a Gaussian sensing matrix.d. we propose here a methodology for designing CS radar detectors based on CAMP. Therefore. we observe that Architecture 1 is comparable to an OS-CFAR detector where the CFAR window is the entire signal. on δ and ρ. Architecture 2. ρ). VIII. Then. However. provides the flexibility to choose both the most appropriate CFAR processor and CFAR window length depending on the specific scenario. a serious disadvantage of Architecture 1 is that. the first step is to compute the transmitted power necessary to reach a given CAMP output SNR. March 8. including the CUT. for a given number of range (or Doppler) bins N and a given σ .6 then using 2 2 equation (4) for a given σ 2 we obtain a value of σ∗ for each couple (δ. In Figure 14(b) the sensing matrix is partial Fourier and the curves are obtained using MC simulations with Adaptive CAMP. This makes Architecture 1 unsuitable for many radar applications. in the system design phase. However. k = kmax . D ESIGN M ETHODOLOGY Using the tools developed in the previous sections. Furthermore. at low FAPs and high Pd . For example. In Figure 14(a) the sensing matrix has i. In Figure 14 we show an example of such curves for several values of kmax and N = 200. besides on the SNR. which is the most relevant case in practical situations. 2Nw = N . being this the least favorable distribution for the non-zero entries in xo [23].

a = 1. and σ 2 = 0.3 0.d.8 1 10 8 6 4 2 0 0.i. δ = 1) the Gaussian sensing matrix is not orthogonal. as δ → 1 the partial Fourier matrix approaches a full Fourier matrix. that we would like to choose δ = 0. it becomes orthogonal.2 0. (b) partial Fourier sensing matrix CAMP output SNR versus δ for Architecture 2 for different numbers of targets kmax .05. δ = 0. ρ) we derive a value for the CAMP output SNR equal to 11. and SNRCS → 13 dB.. even in the limit n = N (i.d.8 0.9 0.5 0.6 kmax = 1 kmax = 6 kmax = 12 kmax = 24 kmax = 48 0. we can use them to predict the performance of our CAMP CFAR detector.4 0.23 12 0.4 δ 0. On the contrary.5 0. Assume. In this case the upper bound (SN R = 13dB) cannot be achieved.2 0.6 0. For this combination (δ.05. CAMP output SNR for Architecture 2 for a = 1 and σ 2 = 0.7 0. 13.2 0. we estimate that if kmax > 1. However. corresponding to SNRM F = 13dB.1 0 4 8 10 ρ 6 0. Gaussian.e. Hence. there will be losses due to target interference.8 0. 14. Here N = 200.8 1 8 6 4 2 0 (a) i..7dB that can be plugged in the selected CFAR Pd and Pf a equations.2 2 0. 14 12 10 SNRCS SNRCS 14 12 kmax = 1 kmax = 6 kmax = 12 kmax = 24 kmax = 48 0. corresponding to SNRM F = 13dB. 2012 DRAFT .e. for example.6 0.6 0.9 δ Fig.i.7 0. just as in conventional MF. Gaussian sensing matrix Fig. Once we have computed the SNR maps or curves. If we use a CA-CFAR processor and we assume March 8. i.4 0.8.1 0. The colormap is the output SNR of CAMP in a dB scale.6 and kmax = 12.4 δ 0. The sensing matrix is i.3 0.

A PPENDIX A R ISK OF THE SOFT THRESHOLDING FUNCTION We derive an analytical expression for (4) for the case when the non-zero coefficients in xo all have equal amplitude.d. We have derived closed form expressions for the CAMP output SNR as a function of system and target parameters. In fact. We believe this work paves the way for the use of CS in radar detection. At the end of the appendix. since kmax is fixed. This means that. we also explain how the calculations can be easily generalized to the more general setting where the non-zero elements have different amplitude distributions. which makes it possible to design practical CS based radar systems. if we want to increase the detection probability for the desired FAP. unfortunately. Based on the SE theoretical results. Such charts play an important role in determining when and how CS can be applied and at what cost. For the March 8. then for Pf a = 10−4 we obtain Pd = 0. such as antenna gain. we either have to increase the number of CS measurements (δ ) or the received power. we have demonstrated that. 2012 DRAFT . Firstly. the combination of a recovery stage based on CAMP with a separate detector achieves the best performance. no theoretical claims can yet be made.7. IX. C ONCLUSIONS In this paper we have achieved two goals. we have provided a methodology to predict the performance of the proposed detector.i. out of two proposed architectures.24 that the targets are not interfering. we have presented the first architecture for adaptive CS radar detection with CFAR properties. Gaussian. which features closed-form expressions for the detection and false alarm probabilities in the ideal case of known system parameters. for which. we have combined conventional CFAR processing with 1 -norm minimization to provide a fully adaptive detection scheme. These goals have been achieved by exploiting CAMP. The latter can be achieved by increasing the transmitted power or any term in the radar equation that increases the received power. by comparing theoretical and experimental results we have been able to understand the behavior of CAMP for sensing matrices other than i. These relations can be used to obtain CS link budget plots that allow the system designer to evaluate the trade off between power and undersampling. our experimental data has shown that our conclusions still hold for the case of partial Fourier sensing matrices. Secondly. Our theoretical findings have been supported by both numerical and experimental evidence. Furthermore. With a simple modification to CAMP.

Zi ∼ N (0. and therefore. δ (13) where Z ∼ CN (0. The first expectation in (13) can be computed as EZ {|η(Z. After some changes of variable and coordinate transformations. τ )|2 } = θ w>τ (w − τ )2 fw (w)fθ (θ)dwdθ = 2 w>τ 2 e−w 2 e−w w(w − τ )2 e−w dw 2 √ = 2 π w>τ  w3 √ π dw − 2τ w>τ w2 √ π dw + τ 2 w>τ w √ 2 e−w  dw . 1) and can be decomposed as Z = (Zr + jZi ). the second expectation in (13) can be written as 2π E(|η(µ + Zr + jZc . τ σ∗ ))|2 }P (x = 0) + EZ {|(η(a + σ∗ Z. 2012 DRAFT . it has been proved in [23] that the function Ψ(σ∗ ) is independent of the phase of the non-zero coefficients. we obtain 1 2 σ∗ = σ 2 + EZ. we will assume in the following calculations that the phase of the non-zero coefficients is equal to zero.X {|(η(X + σ∗ Z. Therefore. where Zr . we have fx (x) = (1 − δρ)δ(0) + δρδ(a). without loss of generality. τ ) − µ)|2 } . respectively.25 complex signal case. π (14) where each of the integrals in the last line is the incomplete moment of a Gaussian random variable with 1 parameters (0. (12) where δρ = k N and a is the amplitude of all non-zero coefficients in xo . √2 ) of order 3. τ σ∗ ) − X)|2 } δ = σ2 + = σ2 + 1 EZ {|(η(σ∗ Z. Define the two independent random variables w = |z| ∼ Rayleigh( 2 ) and θ = (z) ∼ that is uniformly distributed between 0 and 2π . Using fx (x) in (4). and 1 µ = a/σ∗ . 1/2). (15) March 8. τ ))|2 } + δρEZ {|(η(µ + Z. 2. τ σ∗ ) − a)|2 }P (x = a) δ 2 σ∗ (1 − δρ)EZ {|(η(Z. These integrals can be computed numerically. and 1. τ ) − µ|2 ) = φ=0 r≤τ 2π 1 2 2 µ2 e−r −µ +2rµ cos(φ) rdrdφ π + φ=0 r>τ [(r − τ )2 + µ2 − 2µ(r − τ ) cos(φ)]e−r ∞ τ 2 −µ2 +2rµ cos(φ) rdrdφ = µ2 + e−µ 2 r(r − τ )e−r 2 − 4µI1 (2rµ) + 2(r − τ )I0 (2rµ) dr.

σ∗. N .o denote the fixed point solutions of SE in Architectures 1 and 2.α ) x = P (|a + σ∗. 2012 DRAFT . we have Pf a1 = P (|xi | = 0|H0 ) = P (|σ∗. For evaluating the detection probability. . we derive the detection and false alarm probabilities of Architectures 1 and 2. −2 ln α).i + σ∗.α and σ∗.) is the Marcum-Q function. the fixed point solution σ∗ can be iteratively computed using the integrals above. A PPENDIX B P ROOF OF T HEOREM 2 To prove Theorem 2. and Q1 (. (17) Therefore. define a as the square root of the power received from a target at location i. The test statistic at bin i.1 = a2 /σ∗. Architecture 1 Recall from Figure 3(a) that in Architecture 1 the non-zero coefficients in x represent the final detections and the threshold τα is selected so as to achieve the desired FAP α. in what follows. In this architecture.α z| > τα σ∗.α z. is given by H1 |xi | H0 0. We then have Pd1 = P (|xi | = 0|H1 ) = P (|η(˜i . 1).. The Gaussianity of the noise is due to the assumption of the Theorem. then we can also calculate the expected value of the expressions in (15) with respect to µ ∼ G. (16) where z ∼ CN (0. · · · .α )| = 0|H1 ) = P (|xo. τα σ∗. and J1 is the Bessel function of the first kind of order 1 [34]. Since the thresholds τα and τo are different for Architectures 1 and 2. the output of CAMP is given by x = η(xo + σ∗. For a given τ . 2 where z ∼ CN (0. Finally. σ∗. i = 1. respectively. τα σ∗.α evaluated using (4) with threshold parameter τα .α ) = Q1 ( √ 2SNRCS. and it holds in the asymptotic settings according to Theorem 1. respectively. I).α z| > τα σ∗.α .1 .α ). (18) 2 2 where SNRCS.α ) = P (|z| > τα ) = e−τα . note that if the amplitudes of the non-zero coefficients are drawn from an arbitrary distribution G.26 where I0 and I1 are modified Bessel functions of the first kind of order 0 and 1.α z| > τα σ∗. March 8. This leads us to the following straightforward parameter tuning: τα = √ − ln α.

2 ≥SNRCS.i + σ∗.. for the same FAP α.o .2 .o can be computed from (4) with threshold parameter τo .α . Since σ∗. for the same target received power a2 . 2 Since 0 ≤ P(|a + z| ≤ µ) ≤ 1.o . (19) 2SNRCS.o z| > κ) = P (|z| > κ/σ∗. For Architecture 2. The proof of Theorem 2 is now straightforward. √ −2 ln α).2 = a2 /σ∗. Pd2 ≥ Pd1 . Define µ∗ as the median of the absolute value of the random 1 variable z . for a desired FAP α. SNRCS. Thus.. i. Since |z| ∼ Rayleigh(σ 2 /2). with equality if and only if α = e−τo .o z| > κ) = Q1 ( 2 where SNRCS. A PPENDIX C M EDIAN E STIMATOR E RROR B OUND To quantify the error of the median estimator as a function of δ and ρ. 2012 . The last equality comes from the fact that |z| ∼ Rayleigh(1/2). σ 2 ). H0 Therefore.o ) = e−κ 2 2 /σ∗. therefore. P (|z| ≤ µ∗ ) = 2 . Note that the above arguments are independent of the distribution G of the non-zero elements 2 of xo . Pf a2 = P (|xi | > κ|H0 ) = P (|σ∗.o − ln α.27 Architecture 2 Suppose that after the estimation block we obtain the following noisy estimate of xo x = xo + σ∗. i. the detector √ threshold can be set as κ = σ∗.e. consider the random variable x with pdf as in (12) and z ∼ CN (0. using the upper bound we obtain (1 − )(1 − e−µ 2 /σ 2 )≤ 1 − 2 DRAFT March 8. 2 P (|x + z| ≤ µ) = P (|a + z| ≤ µ) + (1 − )P (|z| ≤ µ) = P (|a + z| ≤ µ) + (1 − )(1 − e−µ /σ 2 1 )= . Define = δρ and µ as the median of the absolute value of the random variable x + z .o ≤ σ∗.e.o z| > κ) = P (|a + σ∗.o z. the decision statistic at bin i is given by |xi | H1 κ. The detection probability is given by Pd2 = P (|xi | > κ|H1 ) = P (|xo. where as before σ∗.1 and. we obtain µ∗ = σ log(2).

9 distribution.45 0.86 0. for frequent discussions and constructive comments.7 0.78 0.82 0.8 0.93 0.83 0. for making the LabRadOr radar system available available and for technical support during the experiments.94 0.94 0.7 0.67 0.8 0.1 0.6 0.96 0.15 0.78 0.18 0.65 0.9 0.98 0.89 0. Ender and T. Many thanks also to W.84 0. ACKNOWLEDGMENTS Thanks to Prof.7 0.99 0. J. In Table I we report the p-values from the Kolmogorov-Smirnov (KS) test [35] for N = 200. Furthermore. the bound in (8) follows.87 0.6 0. Mathy from Fraunhofer FHR.9 0. and the p-value measures the similarity of the input samples to the reference distribution.89 0.9 0. A PPENDIX D G AUSSIANITY OF w FOR PARTIAL F OURIER MATRIX We report simulation results that demonstrate the approximate Gaussianity of the noise vector w for the partial Fourier sensing matrix. In the null hypothesis it is assumed that the samples belong to the reference probability density function.7 0.3 0. The Netherlands.73 0. Therefore the normalized bias introduced by the median estimator in the presence of a target is bounded by |µ − µ∗ | ≤ σ log 2(1 − ) − log(2) . using the inequality log 2(1 − ) < log(2).73 0. δ ρ p-value (R) p-value (I) 0.05 0.1 0.2 0.3 0.2 0. 2012 DRAFT . N = 200.5 0.4 0.3 0.01 0. Germany. We performed the KS test also for different combinations of N .VALUES OF THE REAL (R) AND IMAGINARY (I) PARTS OF w FOR PARTIAL F OURIER SENSING MATRIX .4 0. δ .8 0. March 8.28 and therefore µ≤σ log 2(1 − ) . Wachtberg.86 0.3 0.65 0. The KS test compares the empirical distribution of the input samples to the Gaussian TABLE I p.72 0.2 0. In our simulations.6 0.93 0. van Rossum from TNO.75 0. in all cases the null hypothesis was accepted. and ρ and obtained similar results.97 0.84 0.86 0.96 0.5 0.

SIAM J. Pure Appl. Strohmer. Proc. 35:2392–2404. IEEE Trans. 2010. Compressive sensing for subsurface imaging using ground penetrating radar.. Direction estimation using compressive sampling array processing. Improved bounds for restricted isometry constants for Gaussian matrices. [10] L. Karl.. 2007. Sparsity and compressed sensing in radar imaging. Signal Process. Compressed sensing of mono-static and multi-static SAR. R. 2008. P. Apr. H. Oct. Syst. and H. [16] J. Kassam. [19] B. Step-frequency radar with compressive sampling SFR-CS. Ann. [12] L. [17] M. 106(45):18914–18919. 2009. In Proc. A. In Proc. On compressive sensing applied to radar. Temlyakov. 2010. Wang. and A. IEEE J. Acoust.. L. Pandharipande. Hoogeboom. Conf. 24(4):427–445. 31(5):2882–2898. H. J. [5] J. Proc. Stable recovery of sparse overcomplete representations in the presence of noise. G. Analysis of CFAR processors in homogeneous background. A. and M.. [11] S. 89(10):1959–1972. Speech. and M. M. Aerosp. Signals.A. Jun. [13] A. W. 2006. Acad. Sci. Topics Sig. Compressive sensing for high resolution radar imaging. Tao. G. 1983. Jul. Maleki and D. Donoho and V. Steeghs. IEEE Radar Conf. May 2010. IEEE Trans. [20] D. 2010. Jan. Theory. 59(8):1207–1223. A. Proc. 52(1):6–18. J. Asilomar Conf. Stable signal recovery from incomplete and inaccurate measurements. and Comput. Compressed sensing in MIMO radar. C. [2] A.. McClellan. L. [8] Y. IEEE Int. 4(1):146–163. 2010. Stojanovic. Vaidyanathan. IEEE Asia-Pacific Microwave Conf. Aerosp. (ICISP).. Comm. Cetin. (APMC). IEEE Trans.. Poor. P. [18] E. Yu. March 8. J.29 R EFERENCES [1] R. G.. Potter. [15] H. Int. [4] C. [3] I. IEEE. Electron. SPIE Algorithms for Synthetic Aperture Radar Imagery XVI. Rohling. [9] J. In Proc. 90(5):1402–1414. Jul. IEEE Work. C. Tao. MIMO radar using compressive sampling. Topics Sig. 4(2):330–341. Ertin. 1988. IEEE Trans. 2009. Shah. Petropulu. 98(6):1006–1020. Anitori. 2012 DRAFT . Signal Process. The dantzig selector: statistical estimation when p is much larger than n. and Signal Process. Parker. Donoho. and W. Maleki. Natl. and P. 2010. L. [14] P. Conf. Matrix Anal. Cetin. Tanner... 2006. Math. Fuchs. Sel. 2010. Signal Process. Bubacarr and J. In Proc. Y. Message passing algorithms for compressed sensing. Ender. T. Sel.. 2010. IEEE J. and A. Systems. Syst. Electron. Compressive radar imaging. Gandhi and S. Montanari. 57(6):2275– 2284. Leus. Yu.. C. and A. Aug. The generalized likelihood ratio test and the sparse representations approach. P. In Proc. 2007. Stat.. In Proc. 2009. Chen and P. V. Signal Process. Optimally tuned iterative reconstruction algorithm for compressed sensing. [7] Y. Baraniuk and T. Donoho. Stat. [6] M. Image and Signal Process. E. Jun. J. Inf. Feb. Y. 2009.. In Proc. 2009. Radar CFAR thresholding in clutter and multiple target situations. Romberg E. Proc. P. Candes and T. Elad D. High-resolution radar via compressed sensing. (ICASSP). Candes and T. H. J. Herman and T. 19(4):608–621. Scott.. Gurbuz. Otten. Petropulu.

[35] F. Saunders. Theory. In To appear in Proc. Koh. Maleki and A. National Bureau of Standards. IEEE Trans. Montanari. 2011. on Sci. [33] L. Introduction to Radar Systems. Dept. Statist. U. and M. Sel. [28] R. Montanari. I. Otten. Soc. A. Gorinevsky. 1(4):606–617. and D. 1972. G.30 [21] A. Proc. J. 2012 DRAFT . 1998. Roy. [23] A. 2004. with applications to compressed sensing. 32(2):407–499. Theory. Tibshirani. L. Montanari.S. Stat. SPIE Ann. A. Atomic decomposition by basis pursuit. A method for large-scale J. Stat. S. Maleki. J. [34] M. Donoho. Lustig. Meeting: Wavelet Apps. SIAM J. Submitted to IEEE Trans. Abramowitz and I. Inform. M. Oct. The noise-sensitivity phase transition in compressed sensing. Maleki. Maleki. [26] D. Johnstone. Donoho. In Proc. Topics Sig. Montanari. [32] M. The dynamics of message passing on dense graphs. L. van Rossum. S. Baraniuk. Bayati and A. Y. [27] S. Computing. Handbook of Mathematical Functions with Formulas. R. Tanner.. 1 -regularized least squares. Chen. and R. J. Anitori. Anitori. J. IEEE Radar Conf. Amer. Analysis of approximate message passing algorithm. Ann. Skolnik. Theory. IEEE. 2011. [22] D.. of Commerce. [29] S. IEEE March 8. Stegun. Donoho. G. The Kolmogorov-Smirnov test for goodness of fit. The LASSO risk for Gaussian matrices. IEEE Conf. and Mathematical Tables. 98(6):913 – 924. Inf. Baraniuk. Asymptotic analysis of complex LASSO via complex approximate message passing (CAMP). and A. 2010. To appear in IEEE Trans.. Kim. Massey. Proc. Graphs.. 2011. D. Proc. T. [24] M. Assoc. 58(1):267–288. 57(2):764 – 785. Inf. Regression shrinkage and selection via the LASSO. 2007. Zai. Hastie. Inf. Tibshirani.. Dec. Third edition. 2002. Bayati and A. 46(253):68–78. Donoho and J. K. 57(10):6920–6941. IEEE Trans. L. 1996. Least angle regression. 2010. Science and Systems (CISS). Efron. McGraw Hill. Inf. [25] M. Feb. 20:33–61.. [30] B. A. S. Sig. and R. Chen and D. Precise undersampling theorems. Examples of basis pursuit. 2011. L. Imag. Compressive CFAR radar detection. Series B. Jun. A. In Proc. L. 2012. W. Theory. Boyd. 1995. [31] S. 1951. and M.

Master your semester with Scribd & The New York Times

Special offer for students: Only $4.99/month.

Master your semester with Scribd & The New York Times

Cancel anytime.