1 Identification of Underspread Linear Systems with...
Transcript of 1 Identification of Underspread Linear Systems with...
arX
iv:1
008.
0851
v1 [
cs.IT
] 4
Aug
201
01
Identification of Underspread Linear Systems
with Application to Super-Resolution RadarWaheed U. Bajwa,Member, IEEE, Kfir Gedalyahu, and Yonina C. Eldar,Senior Member, IEEE
Abstract
Identification of time-varying linear systems, which introduce both time-shifts (delays) and frequency-
shifts (Doppler-shifts) to the input signal, is one of the central tasks in many engineering applications.
This paper studies the problem of identification ofunderspread linear systems, defined as time-varying
linear systems whose responses lie within a unit-area region in the delay–Doppler space, by probing them
with a single known input signal and analyzing the resultingsystem output. One of the main contributions
of the paper is that it characterizes the conditions on the temporal support and the bandwidth of the input
signal that ensure identification of underspread linear systems described by a set of discrete delays and
Doppler-shifts—and referred to as parametric underspreadlinear systems—from single observations. In
particular, the paper establishes that sufficiently underspread parametric linear systems are identifiable
as long as the time–bandwidth product of the input signal is proportional to the square of the total
number of delay–Doppler pairs in the system. In addition, analgorithm is developed in the paper that
enables identification of parametric underspread linear systems in polynomial time. Finally, application
of these results to super-resolution target detection using radar is discussed. Specifically, it is shown
that the procedure developed in the paper allows to distinguish between multiple targets with very close
proximity in the delay–Doppler space and results in resolution that substantially exceeds those of standard
matched-filtering-based techniques and the recently proposed compressed sensing-based methods.
Index Terms
Compressed sensing, delay–Doppler estimation, matrix-pencil algorithm, rotational invariance tech-
niques, subspace methods, super-resolution radar, systemidentification, time-varying linear systems
W. U. Bajwa is with the Program in Applied and Computational Mathematics, Princeton University, Princeton, NJ 08544
USA (Email: [email protected]). K. Gedalyahu and Y. C. Eldar are with the Department of Electrical
Engineering, Technion—Israel Institute of Technology, Haifa 32000, Israel (Phone: +972-4-8293256, Fax: +972-4-8295757, E-
mails:kfirge@techunix,[email protected]. Y. C. Eldar is currently also a Visiting Professor at Stanford
University, USA.
August 5, 2010 DRAFT
2
Fig. 1. Schematic representation of identification of a time-varying linear systemH by probing it with a known input signal.
Characterization of an identification scheme involves specification of the input probe,x(t), and the accompanying sampling and
recovery stages.
I. INTRODUCTION
A. Background
Physical systems arising in a number of application areas can often be best described as time-varying
linear systems [1], [2]. Identification of such systems in many situations either helps to improve the
overall system performance, e.g., the bit-error rate in communications [1], or constitutes an integral part
of the overall system operation, e.g., target detection using radar or active sonar [2].
Mathematically, identification of a given time-varying linear systemH involves probing it with a
single known input signalx(t) and identifyingH by analyzing thesingle system outputH(x(t)) [3],
as illustrated in Fig. 1. Unlike the case oftime-invariant linear systems, however, a single observation
of a time-varying linear system does not lead to a unique solution unless additional constraints on the
response of the system are imposed. This is due to the fact that time-varying linear systems introduce both
time-shifts (delays) and frequency-shifts (Doppler-shifts) to the input signal. It is now a well-established
fact in the literature that a time-varying linear systemH can only be identified from a single observation
if H(δ(t)) is known to lie within a regionR of the delay–Doppler space such thatarea(R) < 1 [3]–[6].
Such identifiable time-varying linear systems are termed asunderspreadlinear systems in the literature,
as opposed to nonidentifiableoverspreadlinear systems, which satisfyarea(R) > 1 [3], [6].1
In this paper, we study the problem of identification of underspread linear systems whose responses
can be described by a set of discrete delays and Doppler-shifts. That is,
H(x(t)) =
K∑
k=1
αkx(t− τk)ej2πνkt (1)
1It is still an open research question as to whethercritically spread linear systems, which correspond toarea(R) = 1, are
identifiable or nonidentifiable [6]; see [3] for a partial answer to this question for the case whenR is a rectangular region.
August 5, 2010 DRAFT
3
where (τk, νk) denotes a delay–Doppler pair andαk ∈ C is the complex attenuation factor associated
with (τk, νk). Unlike most of the existing work in the literature, however, our goal in this paper is to
explicitly characterize the conditions on the temporal support and the bandwidth of the input signal that
ensure identification of such underspread linear systems from single observations. The importance of this
goal can be best put into perspective by realizing that underspread linear systems of the form (1) tend
to arise frequently in many engineering applications. Consider, for example, the case of a single-antenna
transmitter communicating wirelessly with a single-antenna receiver in a mobile environment. Then, over
a small-enough time interval, the wireless channel betweenthe transmitter and the receiver is of the
form (1) with each triplet(τk, νk, αk) corresponding to a distinct physical path taken by the transmitted
signal from the transmitter to the receiver [7]. Identification of H in this case enables one to establish a
relatively error-free communication link between the transmitter and the receiver. But wireless systems
also need to identify channels using signals that have as small time–bandwidth productas possible so
that they can allocate the rest of the temporal degrees of freedom to communicating data [7], [8].
Similarly, in the case of target detection using radar or active sonar, the (noiseless, clutter-free) received
signal is of the form (1) with each triplet(τk, νk, αk) corresponding to an echo of the transmitted signal
from a distinct target in the delay–Doppler space [2]. Identification of H in this case, therefore, enables
one to accurately obtain the radial position and the radial velocity of the targets. But radar systems also
need to operate with signals (waveforms) that have as small temporal support and bandwidth as possible.
This is because the temporal support of the radar waveform isdirectly tied to the time it takes to identify
all the targets while the bandwidth of the waveform—among other technical considerations—is tied to
the sampling rate of the radar receiver [2].
B. Our Contributions and Relation to Previous Work
Given the ubiquity of time-varying linear systems in engineering applications, there exists considerable
amount of existing literature that studies the problem of identification of such systems in an abstract
setting. Kailath was the first to recognize that the identifiability of a time-varying linear systemH from a
single observation is directly tied to the area of the regionR that containsH(δ(t)) [4]. Kailath’s seminal
work in [4] laid the foundations for the groundbreaking future works of Bello [5], Kozek and Pfander
[3], and Pfander and Walnut [6], which establish the nonidentifiability of overspread linear systems
and provide constructive proofs for the identifiability of underspread linear systems. Nevertheless, the
constructive proofs provided in [3]–[6] are for the identification ofarbitrary underspread linear systems.
None of these results therefore shed any light on the temporal support and the bandwidth of the input
August 5, 2010 DRAFT
4
signal needed to ensure identification of underspread linear systems of the form (1). On the contrary, the
constructive proofs provided in [3]–[6] for the identification of underspread linear systems require use
of input signals with infinite temporal support and infinite bandwidth.
In contrast, to the best of our knowledge, this paper is the first attempt to develop a theory for
identification of underspread linear systems of the form (1)—henceforth referred to asparametric un-
derspread linear systems—that parallels that of [3]–[6] for identification of arbitrary time-varying linear
systems. One of the main contributions of this paper is that we establish using a constructive proof that
sufficiently underspread parametric linear systems are identifiable as long as the time–bandwidth product
of the input signal is proportional to square of the total number of delay–Doppler pairs in the system.
Equally importantly, as part of our constructive proof, we also concretely specify the nature of the input
signal and the structure of a corresponding polynomial-time recovery procedure that enable identification
of parametric underspread linear systems. The developments in the paper leverage key results on time-
delay estimation presented in [9] and the literature on direction-of-arrival (DOA) estimation [10]–[12].
Unlike the traditional DOA estimation literature, however, we do not assume that the system output is
observed by an array of antennas. Similarly, unlike the time-delay estimation results presented in [9], our
goal here is not a reduction in the sampling rate; rather, we are interested in characterizing the minimum
temporal degrees of freedom of the input signal needed to ensure identification of parametric underspread
linear systems.
In order to best put the significance of the results of this paper into perspective, it is also instructive to
compare it with some of the corresponding results reported in [13]–[18] for the identification of parametric
underspread linear systems. The key factor that distinguishes this paper from [13]–[18] is that [13]–[18]
and similarly related works only consider specialized versions of parametric underspread linear systems.
More specifically, [13], [15]–[18] study the problem of identification of parametric underspread linear
systems whose delays and Doppler-shifts lie on a quantized grid in the delay–Doppler space, while [14]
considers the problem of identification of parametric underspread linear systems in which there are no
more than two Doppler-shifts associated with the same delay. Note that while the insights of [13], [15]–
[18] can be extended to arbitrary parametric underspread linear systems by taking infinitesimally-fine
quantization of the delay–Doppler space, the results reported in [13], [15]–[18] in that case require—
similar to the case of [3]–[6]—input signals with infinite temporal support and infinite bandwidth. On
the other hand, we are unaware of a straightforward generalization of the results of [14] to arbitrary
parametric underspread linear systems. Our ability to avoid quantization of the delay–Doppler space is
due to the fact that we treat the system-identification problem directly in the analog domain. This follows
August 5, 2010 DRAFT
5
the philosophy in much of the recent work in analog compressed sensing, termed Xampling, which
provides a framework for incorporating and exploiting structure in analog signals without the need for
quantization [19]–[24].
Before concluding this discussion, it is also worth mentioning here that responses of arbitrary under-
spread linear systems can always be represented as (1) underthe limitK → ∞. Therefore, the main result
of this paper can also be construed as an alternate constructive proof of the identifiability of sufficiently
underspread linear systems. Nevertheless, just like [3]–[6], this interpretation of the results presented in
this paper again seem to suggest that identification of arbitrary underspread linear systems requires use
of input signals with infinite temporal support and infinite bandwidth.
C. Notation
We make use of the following notational convention throughout this paper. Vectors and matrices are
denoted by bold-faced lowercase and bold-faced uppercase letters, respectively. Thenth element of a
vectora is written asan, and the(i, j)th element of a matrixA is denoted byAij . Superscripts(·)∗, (·)T
and (·)H represent conjugation, transposition, and conjugate transposition, respectively. In addition, the
Fourier transform of a continuous-time signalx (t) ∈ L2(C) is defined byX (ω)def=
∫∞
−∞x (t) e−jωtdt,
while 〈x (t) , y (t)〉 =∫∞
−∞x (t) y∗(t)dt denotes the inner product between two continuous-time signals
in L2(C). Similarly, the discrete-time Fourier transform of a sequence a [n] ∈ ℓ2(C) is defined by
A(ejωT
)=
∑n∈Z a [n] e
−jωnT and is periodic inω with period2π/T . Finally, we useA† to write the
Moore–Penrose pseudoinverse of a matrixA.
D. Organization
The rest of this paper is organized as follows. In Section II,we formalize the problem of identification
of parametric underspread linear systems along with the accompanying assumptions. In Section III, we
propose a polynomial-time recovery procedure used for the identification of parametric underspread linear
systems, while we specify in Section IV the conditions on theinput signal needed to guarantee unique
identification using the proposed procedure. In Section V, we compare the results of this paper to some
of the related literature on identification of parametric underspread linear systems and super-resolution
target detection using radar. Finally, we present some numerical results in Section VI and concluding
remarks in Section VII.
August 5, 2010 DRAFT
6
II. PROBLEM FORMULATION AND MAIN RESULTS
In this section, we formalize the problem of identification of a parametric underspread linear system
H whose response is described by a total ofK arbitrary delay–Doppler-shifts of the input signal. The
task of identification ofH essentially requires specifying two distinct but highly intertwined steps. First,
we need to specify the conditions on the temporal support andthe bandwidth of the input signalx(t)
that ensure identification ofH from a single observation. Second, we need to provide a polynomial-time
recovery procedure that takes as inputH(x(t)) and provides an estimateH of the system response by
taking advantage of the structure of the input signal specified in the first step. Before specifying these
steps, we need to be more specific about our system model and the accompanying assumptions.
We first express the response of a parametric underspread linear systemH comprising ofK delay–
Doppler pairs [cf. (1)] in terms ofKτ ≤ K distinct delays as follows2
H(x(t)) =
Kτ∑
i=1
Kν,i∑
j=1
αijx(t− τi)ej2πνijt (2)
whereνij denotes thejth Doppler-shift associated with theith distinct delayτi, αij ∈ C denotes the
attenuation factor associated with the delay–Doppler pair(τi, νij), andKdef=
∑Kτ
i=1Kν,i. In addition,
throughout the rest of this paper, we useτdef= τi, i = 1, . . . ,Kτ to denote the set ofKτ distinct
delays associated withH. The first main assumption that we make here concerns the footprint of H in
the delay–Doppler space.
[A1] The responseH(δ(t)) of H lies within a rectangular region of the delay–Doppler space; in other
words, (τi, νij) ∈ [0, τmax] × [−νmax/2, νmax/2]. This is indeed the case in many engineering
applications (see, e.g., [1], [2]), and the parametersτmax andνmax are termed in the parlance of
linear systems as thedelay spreadand theDoppler spreadof the system, respectively.
Next, we useT andW to denote the temporal support and the two-sided bandwidth of the known
input signalx(t) used to probeH, respectively. Without loss of generality, we expressx(t) in the form
x(t) =
N−1∑
n=0
xng(t− nT ), 0 ≤ t ≤ T (3)
whereg(t) is a prototype pulse of bandwidthW that is (essentially) temporally supported on[0, T ] and
is assumed to have unit energy(∫|g(t)|2dt = 1), while xn ∈ C is anN -length probing sequence
that is assumed to have unitℓ2-norm (∑
n |xn|2 = 1). Here, the parameterN is proportional to the
2Note that (1) and (2) are equivalent in terms of the mathematical characterization; nevertheless, we choose to expressH(x(t))
as in (2) so as to facilitate the forthcoming analysis.
August 5, 2010 DRAFT
7
time–bandwidth product of the input signalx(t), which roughly defines the number of temporal degrees
of freedom available for estimatingH [8]: N = T /T ∝ T W.3 The final two assumptions that we make
in this paper concern the relationship between the delay spread and the Doppler spread ofH and the
temporal support and bandwidth of the pulseg(t).
[A2] The delay spread of the underspread linear systemH is strictly smaller than the temporal support
of the prototype pulse (in other words,τmax < T ).
[A3] The Doppler spread of the underspread linear systemH is much smaller than the bandwidth of the
prototype pulse (in other words,νmax ≪ W).
Note that, sinceW ∝ 1/T , [A3] equivalently imposes thatνmaxT ≪ 1; in words, this assumption states
that the temporal scale of variations in the system responseis large relative to the temporal scale of
variations inx(t). It is worth pointing out here that linear systems that are sufficiently underspread in
the sense thatτmaxνmax ≪ 1 can always be made to satisfy[A2] and [A3] for any given budget of the
time–bandwidth product.
Remark1. In order to further elaborate on the validity of[A2] and [A3], note that there indeed exist
many communication applications where the underlying linear systems tend to be highly underspread
[1, § 14.2]. On the other hand,[A2] and [A3] in the context of radar target detection simply mean that
the targets are not too far away from the radar and that their velocities are not too high. Consider, for
example, anL-band radar (operating frequency of1.3 GHz) that transmits a pulseg(t) of bandwidth
W = 10 MHz after everyT = 50 µs. Then both[A2] and [A3] are satisfied for the case when the
distance between the radar and any target is at most7.5 km (τmax ≈ 50 µs) and the radial velocity of
any target is at most185 km/h (νmax ≈ 445 Hz) [2].
We are now ready to summarize the key findings of this paper concerning identification of parametric
underspread linear systems in terms of the following theorem.
Theorem 1 (Identification of Parametric Underspread Linear Systems). Suppose thatH is a parametric
underspread linear system that is completely described by atotal ofK =∑Kτ
i=1Kν,i triplets (τi, νij , αij).
Then, irrespective of the distribution of(τi, νij) within the delay–Doppler space,H can be identified in
polynomial-time from a single observationH(x(t)) as long as it satisfies[A1]–[A3], the probing sequence
xn remains bounded away from zero in the sense that|xn| > 0 ∀ n, and the time–bandwidth product
3Recall from elementary signal processing that the temporalsupport and the bandwidth of an arbitrary pulseg(t) are related
to each other asW ∝ 1/T .
August 5, 2010 DRAFT
8
Fig. 2. Schematic representation of the polynomial-time recovery procedure proposed in the paper for identification ofparametric
underspread linear systems from single observations.
of the known input signalx(t) satisfies the condition
T W ≥ 8πKτKν,max (4)
whereKν,max is the maximum number of Doppler-shifts associated with anyone of the distinct delays;
in other words,Kν,maxdef= maxiKν,i. In addition, the time–bandwidth product ofx(t) is guaranteed to
satisfy(4) as long asT W ≥ 2π(K + 1)2.
The rest of this paper is devoted to providing a proof of Theorem 1. Specifically, we will make use of
(2) and (3) in the sequel to describe:
[1] The polynomial-time recovery procedure used for the identification ofH (cf. Section III), and
[2] The accompanying conditions onx(t) needed to guarantee identification ofH (cf. Section IV).
III. POLYNOMIAL -TIME IDENTIFICATION OF UNDERSPREADL INEAR SYSTEMS
In this section, we characterize the polynomial-time recovery procedure used in the paper for identifi-
cation of underspread linear systems of the form (2). In order to facilitate understanding of the proposed
algorithm, shown in Fig. 2, we conceptually partition the procedure into two stages: the sampling stage
and the recovery stage. The rest of this section is devoted todescribing these two steps of the recovery
in detail. Before proceeding further, however, it is instructive to first make use of (2) and (3) and rewrite
August 5, 2010 DRAFT
9
the output ofH as
H(x(t)) =
Kτ∑
i=1
Kν,i∑
j=1
N−1∑
n=0
αijxnej2πνijtg(t− τi − nT )
(a)≈
Kτ∑
i=1
Kν,i∑
j=1
N−1∑
n=0
αijxnej2πνijnT g(t− τi − nT )
=
Kτ∑
i=1
N−1∑
n=0
ai[n]g(t − τi − nT ) (5)
where(a) follows from the assumption thatνmaxT ≪ 1, which implies thatej2πνijt ≈ ej2πνijnT for all
t ∈ [(n− 1)T, nT ), and the sequencesai[n], i = 1, . . . ,Kτ , are defined as
ai[n] =
Kν,i∑
j=1
αijxnej2πνijnT , n = 0, . . . , N − 1. (6)
A. The Sampling Stage
We leverage some of the results reported in [9] on time-delayestimation in order to describe the
sampling stage of our proposed recovery procedure. Note that the primary objective in [9] is time-delay
estimation using low-rate samples, whereas the development in here is carried out with an eye towards
identification of parametric underspread linear systems regardless of the distribution of system parameters
within the delay–Doppler space—the so-calledsuper-resolution identification. Specifically, [9] introduced
a general multi-channel sampling scheme for the purpose of recovering a set of unknown delays from
signals of the form (5). In this paper, we focus on one specialcase of that sampling scheme, which
consists of a low-pass filter (LPF) followed by a uniform sampler. This instance of the sampling scheme
proposed in [9] is preferable from an implementation viewpoint since it requires only one sampling
channel, thereby simplifying the analog front-end of the sampling hardware. Note that the LPF, besides
being required by the sampling stage, also serves as the front-end of the system-identification hardware
and rejects noise and interference outside the working spectral band in practical settings.
To be concrete, the sampling stage of our recovery method first passes the system outputy(t)def=
H(x(t)) through a LPF whose impulse response is given bys∗(−t) and then uniformly samples the
output of the LPF at timest = nT/p
. Here, we assume that the frequency response,S∗(ω), of this
LPF is contained in the spectral bandF , defined as
F =[−π
Tp,π
Tp], (7)
while S∗(ω) is zero for frequenciesω /∈ F . The only assumption that we make here is that the parameter
p is even and satisfies the conditionp ≥ 2Kτ ; exact conditions thatp eventually needs to satisfy to ensure
August 5, 2010 DRAFT
10
identification ofH will be given in Section IV. Now in order to relate the sampledoutput of the LPF
with the multi-channel sampling formulation of [9], we firstdefinep sampling (sub)sequencescℓ[n]
as follows
cℓ[n] = 〈y(t), s(t− nT − (ℓ− 1)T/p)〉 , ℓ = 1, . . . , p. (8)
Note that these sampling (sub)sequences correspond to periodically splitting the sampled sequence at the
output of the LPF, which is generated at a rate ofp/T , into p slower sequences at a rate of1/T each
using a serial-to-parallel converter; see Fig. 2 for a schematic representation of this splitting.
Next, we define the vectorc(ejωT
)as thep-length vector whoseℓth element isCℓ
(ejωT
), which
denotes thediscrete-time Fourier transform(DTFT) of cℓ[n]. In a similar fashion, we definea(ejωT
)as
theKτ -length vector whoseith element is given byAi
(ejωT
), the DTFT ofai[n]. Then it can be argued
along the lines of the developments carried out in [9] that these two vectors are related to each other as
follows
c(ejωT
)= W
(ejωT
)N (τ )D
(ejωT , τ
)a(ejωT
). (9)
Here,W(ejωT
)is a p× p matrix whose(ℓ,m)th element is given by
Wℓm
(ejωT
)= ejω(ℓ−1)T/p 1
TS∗
(ω +
2π
Tm′
)G
(ω +
2π
Tm′
)ej
2π
p(ℓ−1)m′
, (10)
wherem′ def= m− p/2− 1, while N (τ ) is a p×Kτ matrix whose(m, i)th element is given by
Nmi (τ ) = e−j 2π
Tm′τi , (11)
andD(ejωT , τ
)is aKτ ×Kτ diagonal matrix whoseith diagonal element is given bye−jωτi . Further,
if we assume for the time being thatW(ejωT
)is a stably-invertible matrix then we can obtain a new
vectord(ejωT
)from c
(ejωT
)as follows
d(ejωT
)= W
−1(ejωT
)c(ejωT
). (12)
It is also easy to verify that this vector, which we refer to asthe modified measurement vector, satisfies
d(ejωT
)= N (τ )b
(ejωT
), (13)
where the vectorb(ejωT
)is defined as
b(ejωT
)= D
(ejωT , τ
)a(ejωT
). (14)
August 5, 2010 DRAFT
11
Finally, note that sinceN (τ ) is not a function ofω, (13) can be expressed in the discrete-time domain
using the linearity of the DTFT as follows
d[n] = N (τ )b[n], n ∈ Z. (15)
Here, the elements of the vectorsd[n] andb[n] are discrete-time sequences that are given by the inverse
DTFT of the elements of the vectorsd(ejωT
)andb
(ejωT
), respectively.
The key insight to be drawn here is that (13), and its time-domain equivalent (15), can be viewed as an
infinite ensemble of modified measurement vectors in which each element corresponds to a distinct matrix
N (τ ) that, in turn, depends on the set of (distinct) delaysτ . Linear measurement models of the form
(15)—in which the measurement matrix is completely determined by a set of (unknown) parameters—
have been studied extensively in a number of research areas such as system identification [25] and
direction-of-arrival and spectrum estimation [26], [27].One specific class of methods that has proven
to be quite useful in these areas in efficiently recovering the parameters that describe the measurement
matrix are the so-calledsubspace methods[26]. Consequently, our approach in the recovery stage of the
proposed procedure will be to first use subspace methods in order to recover the setτ from d[n] and
afterwards recover the vectora(ejωT
)from the measurement vectord[n] using linear filtering operations
as follows [cf. (14), (13)]
a(ejωT
)= D
−1(ejωT , τ
)N
† (τ )d(ejωT
). (16)
Finally, we recover the Doppler-shifts and the attenuationfactors associated with the linear systemHfrom the vectora
(ejωT
)by making another use of the subspace methods.
Before proceeding to the recovery stage, however, recall that we first need to justify the assumption
that the matrixW(ejωT
)can be stably inverted. To this end, we first observe from (10)thatW
(ejωT
)
can be decomposed as
W(ejωT
)= Φ
(ejωT
)FHΨ
(ejωT
), (17)
whereΦ(ejωT
)is a p× p diagonal matrix withℓth diagonal element
Φℓℓ
(ejωT
)=
√p(−1)ℓ−1ejω(ℓ−1)T/p, (18)
F is a p-point discrete Fourier transform (DFT) matrix with(ℓ,m)th element equal to
Fℓm =1√pe−j 2π
p(ℓ−1)(m−1), (19)
August 5, 2010 DRAFT
12
andΨ(ejωT
)is a p× p diagonal matrix whosemth diagonal element is given by
Ψmm
(ejωT
)=
1
TS∗
(ω +
2π
T(m− p/2− 1)
)G
(ω +
2π
T(m− p/2− 1)
). (20)
It can now be easily seen from the decomposition in (17) that,in order for W(ejωT
)to be stably
invertible, each of the above three matrices has to be stablyinvertible. By construction, however, it
is easy to see that bothΦ(ejωT
)and F
H are stably invertible. On the other hand, the invertibility
requirement on the diagonal matrixΨ(ejωT
)leads to the following conditions on the pulseg(t) and the
kernels∗(−t) of the LPF.
Condition 1. In order for Ψ(ejωT
)to be stably invertible, the continuous-time Fourier transform of the
pulseg (t) has to satisfy
a ≤ |G (ω)| ≤ b a.e.ω ∈ F (21)
for some positive constantsa > 0 and b <∞.
Condition 2. In order for Ψ(ejωT
)to be stably invertible, the continuous-time Fourier transform of the
LPF s∗ (−t) has to satisfy
c ≤ |S (ω)| ≤ d a.e.ω ∈ F (22)
for some positive constantsc > 0 and d <∞.
Note that Condition 1 requires that the bandwidthW of the prototype pulseg(t) has to satisfy
W ≥ 2π
Tp. (23)
In Section IV, we will also derive conditions on the parameter p and combine them with (23) to obtain
equivalent conditions on the time–bandwidth product of theinput signalx(t) that will ensure invertibility
of the matrixW(ejωT
).
We conclude this discussion of the sampling stage of our proposed recovery procedure by pointing
out that, besides the invertibility condition, the decomposition in (17) also provides an efficient way to
implement thedigital-correction filter bankW−1(ejωT
). Specifically, we have from (17) that this filter
bank satisfies
W−1
(ejωT
)= Ψ
−1(ejωT
)FΦ
−1(ejωT
). (24)
August 5, 2010 DRAFT
13
It is now easy to see from this representation that the implementation ofW−1(ejωT
)can be done in three
stages, where each stage corresponds to one of the three matrices in the above expression. Specifically,
define the set of digital filtersφℓ[n] andψℓ[n] as
φℓ[n] = IDTFTΦ
−1ℓℓ
(ejωT
)[n], 1 ≤ ℓ ≤ p (25)
and
ψℓ[n] = IDTFTΨ
−1ℓℓ
(ejωT
)[n], 1 ≤ ℓ ≤ p, (26)
where IDTFT denotes the inverse DTFT operation. Then the first correction stage involves filtering the
sampling (sub)sequencescℓ[n] using the set of filtersφℓ[n]. Next, multiplication with the DFT matrix
F can be efficiently implemented by applying the fast Fourier transform (FFT) to the outputs of the filters
in the first stage. Finally, the third correction stage involves filtering the resulting sequences using the
set of filtersψℓ[n] to get the desired sequencesdℓ[n]. Note that this last correction stage can also be
interpreted as an equalization stage that compensates for the non-flatness of the frequency responses of
the prototype pulse and the kernel of the LPF. A detailed schematic representation of the sampling stage
of the proposed recovery procedure, which is based on the preceding interpretation ofW−1(ejωT
), is
provided in Fig. 2.
B. The Recovery Stage
We conclude this section by describing in detail the recovery stage of the proposed procedure, which—
as noted earlier—consists of two steps. In the first step, we rely on subspace methods to recover the set
of delaysτ from d[n] [cf. (15)]. In the second step, we make use of the recovered delays to obtain the
Doppler-shifts and the attenuation factors associated with each of the delays.
1) Recovery of the Delays:In order to recover the set of delaysτ by processing the output of the
sampling stage, we rely on the approach advocated in [9] and make use of the well-known ESPRIT
algorithm [28] together with an additional smoothing stage[29]. Specifically, we propose to recoverτ
from d[n] using the following method; see [9], [28] for further details concerning this algorithm.
(i) Construct the matrix
Rdd =1
M
M∑
m=1
∑
n∈Z
dm[n]dHm[n], (27)
wheredm is theM = p/2 length subvector which is given by
dm [n] =[dm [n] dm+1 [n] . . . dm+M [n]
]T. (28)
August 5, 2010 DRAFT
14
(ii) RecoverKτ as the rank ofRdd.
(iii) Perform a singular value decomposition(SVD) of Rdd and construct the matrixEs consisting of
theKτ singular vectors corresponding to theKτ nonzero singular values ofRdd as its columns.
(iv) Compute the matrixΦ = E†s↓Es↑, whereEs↑ andEs↓ denote the submatrices extracted fromEs
by removing its first row and its last row, respectively.
(v) Compute the eigenvalues ofΦ, λi, i = 1, 2, . . . ,Kτ .
(vi) Recover the unknown delays as follows:τi = − T2πarg(λi).
2) Recovery of the Doppler-Shifts and the Attenuation Factors: The second step of the recovery stage
consists of determining the Doppler-shifts and the attenuation factors by making use of the recovered
delays. Once the unknown delays are found, we can recover thevectorsa[n] through the frequency
relation (16). Next, define for each delayτi, the set of corresponding Doppler-shiftsνi as
νidef= νij , j = 1, . . . ,Kν,i (29)
and recall that theith element ofa[n] is given by (6). We can therefore write theN -length sequence
ai[n] for each indexi in the following matrix–vector form
ai = XR(ν i)αi, (30)
whereai is a length-N vector whosenth element isai[n], X is anN ×N diagonal matrix whosenth
diagonal element is given byxn, R(ν i) is anN ×Kν,i matrix with njth elementej2πνijnT , andαi is
length-Kν,i vector with jth elementαij . Now since the sequencexn is completely determined by the
input signalx(t), the matrixX in (30) can be inverted under the assumption that the probingsequence
xn satisfies|xn| > 0 ∀ n and we therefore obtain
ai = R(νi)αi, (31)
where we have thataidef= X
−1ai. It now follows from a simple inspection of the elements ofai that
ai[n] =
Kν,i∑
j=1
αijej2πνijnT , 0 ≤ n ≤ N − 1. (32)
It is now easy to see from this representation that the recovery of the Doppler-shifts from the sequences
ai[n] is equivalent to the problem of recovering distinct frequencies from a (weighted) sum of complex
exponentials. In the context of our problem, for each fixed index i, the frequency of thejth exponential
is given byωij = 2πνijnT and its amplitude is given byαij.
Fortunately, the problem of recovering frequencies from a sum of complex exponentials has been
studied extensively in the literature and various strategies exist for solving this problem (see [27] for a
August 5, 2010 DRAFT
15
review). One of these techniques that has gained interest recently, especially in the literature on finite
rate of innovation [30]–[33], is theannihilating-filtermethod. The annihilating-filter approach, in contrast
to some of the other techniques, allows the recovery of the frequencies associated with theith index
even at the critical value ofN = 2Kν,i. On the other hand, subspace methods such as ESPRIT [10],
matrix-pencil algorithm [11], and the Tufts and Kumaresan approach [12] are generally more robust to
noise in the system but also require more samples than2Kν,i. In summary, we conclude that there are
a number of methods in the existing literature that can be used for recovery of the Doppler-shifts from
(32) depending upon the temporal degrees of freedomN available for identification ofH. In particular,
if one is faced with the condition thatN = 2Kν,i for any one of the indices then the annihilating filter
should be used.
Finally, under the assumption that the Doppler-shifts for each indexi have been recovered using any
one of the subspace methods, the attenuation factorsαij associated with each of the delaysτi can
simply be determined as
αi = R†(νi)ai, i = 1, . . . ,Kτ . (33)
IV. SUFFICIENT CONDITIONS FORIDENTIFICATION OF UNDERSPREADL INEAR SYSTEMS
Our focus in Section III was on developing a recovery algorithm for the identification of underspread
linear systems. We now turn to specify conditions that guarantee that the proposed recovery procedure
recovers the set of triplets,(τi, νij , αij)
, that describe the parametric underspread linear systemH.
As noted earlier, we represent these conditions in terms of equivalent conditions on the time–bandwidth
product T W of the input signalx(t). This is a natural way to describe the performance of system
identification schemes sinceT W roughly defines the number of temporal degrees of freedom available
for estimatingH [8].
To begin with, recall that the recovery stage involves first determining the unknown delaysτ from the
set of equations given by (15) (cf. Section III-B). Therefore, to ensure that our algorithm successfully
returns the parameters ofH, we first need to provide conditions that guarantee unique solutions to (15).
To this end, we first introduce some new notation to facilitate the forthcoming analysis. Specifically,
let d [Λ] denote the set of all modified measurement vectorsd [n], d [Λ]def= d [n] , n ∈ Z, and let
b [Λ]def= b [n] , n ∈ Z denote the set of all unknown vectorsb [n]. Using this notation, we can then
rewrite (15) as
d [Λ] = N (τ )b [Λ] . (34)
August 5, 2010 DRAFT
16
We now make use of this new notation to leverage the analysis carried out in [9] and provide sufficient
conditions for a unique solution to (34) in terms of the following proposition; see [9] for a formal proof.
Proposition 1. If(τ , b [Λ] 6= 0
)solves(34) and if
p > 2Kτ − dim(span
(b [Λ]
))(35)
then(τ , b [Λ]
)is the unique solution of(34). Here, span
(b [Λ]
)is used to denote the subspace of
minimal dimensions that containsb [Λ].
Note that Proposition 1 suggests that a unique solution to (34)—and, by extension, unique recovery
of the set of delaysτ—is guaranteed through a proper selection of the parameterp. In particular, since
dim(span
(b [Λ]
))is a positive number in general, we have from Proposition 1 that p ≥ 2Kτ is a
sufficient condition for unique recovery of the set of delaysτ and the corresponding vectorsb [Λ]. In
addition, since we have from Condition 1 in Section III that the parameterp effectively determines the
minimum bandwidthW of the prototype pulse [cf. (23)], Proposition 1 effectively results in the following
sufficient condition on the bandwidth of the input signal forunique recovery ofτ andb [Λ] by combining
the conditionp ≥ 2Kτ and the condition obtained earlier in (23):
W ≥ 4πKτ
T. (36)
The second step in the recovery stage involves recovering the Doppler-shifts and the attenuation factors
(cf. Section III-B). We now investigate the conditions required for unique recovery of the Doppler-shifts.
Recall that the vectorsb[n] anda[n] are related to each other by the invertible frequency relation (14).
Therefore, since the diagonal matrixD(ejωT , τ
)is invertible and completely specified by the set of delays
τ , the condition given in (36) also guarantees unique recovery of the vectorsa[n] from the vectorsb[n].
Further, it can be easily verified that the matrixR(ν i) in (31) has the same parametric structure as that
required by Proposition 1. We can therefore once again appeal to the results of Proposition 1 in providing
conditions for unique recovery of the sets of Doppler-shifts νi from the vectorsai [cf. (31)]. To that
end, the only thing that needs to be done is to interchangep with N andKτ with Kν,i in Proposition 1
and use the fact that dim(span(ai)) = 1 (since it is a nonzero vector). Therefore, by making use of
Proposition 1, we get that a sufficient condition for unique recovery ofνi from (31) is
N ≥ 2Kν,i. (37)
Note that the above condition is intuitive in the sense that there are2Kν,i unknowns in (31) (Kν,i
unknown Doppler-shifts andKν,i unknown attenuation factors) and therefore at least2Kν,i equations are
August 5, 2010 DRAFT
17
required to solve for these unknown parameters. Finally, since we need to ensure unique recovery of the
Doppler-shifts and the attenuation factors for each distinct delayτi, we finally have the condition
N ≥ 2maxiKν,i (38)
which trivially ensures that (37) holds for everyi = 1, . . . ,Kτ . We summarize these results for unique
recovery of the set of triplets(τi, νij , αij) in terms of the following theorem.
Theorem 2 (Sufficient Conditions for System Identification). Suppose thatH is a parametric underspread
linear system that is completely described by a total ofK =∑Kτ
i=1Kν,i triplets (τi, νij , αij). Then,
irrespective of the distribution of(τi, νij) within the delay–Doppler space, the recovery procedure
specified in Section III with the samples taken att = 2nπ/W uniquely identifiesH from a single
observationH(x(t)) as long as the system satisfies[A1]–[A3], the probing sequencexn remains
bounded away from zero in the sense that|xn| > 0 ∀ n, and the time–bandwidth product of the (known)
input signalx(t) satisfies the condition
T W ≥ 8πKτKν,max (39)
whereKν,max is the maximum number of Doppler-shifts associated with anyone of the distinct delays;
in other words,Kν,maxdef= maxiKν,i.
Proof: Recall from the previous discussion that the delays, Doppler-shifts, and the attenuation factors
associated withH can be uniquely recovered as long asN ≥ 2Kν,max, W ≥ 4πKτ
T , andp ≥ 2Kτ . Now
takeN = T W4πKτ
and note that under the assumptionT W ≥ 8πKτKν,max, we trivially haveN ≥ 2Kν,max.
Further, sinceT = NT and since sampling rate of2π/W implies p = 2π/(WT ), we also have that
W = 4πKτ
T ⇒ p = 2Kτ . This completes the proof of the theorem.
Note that Theorem 2 implicitly assumes thatKτ (or an upper bound onKτ ) andKν,max (or an upper
bound onKν,max) are known at the transmitter side. We explore this point in further detail in Section V
and numerically study the effects of “model-order mismatch” on the robustness of the proposed recovery
procedure. It is also instructive (especially for comparison purposes with related work such as [16],
[17]) to present a weaker version of Theorem 2 that only requires knowledge of the total number of
delay–Doppler pairsK.
Corollary 1 (Weaker Sufficient Conditions for System Identification). Suppose that the assumptions
of Theorem 2 hold. Then the recovery procedure specified in Section III with the samples taken at
August 5, 2010 DRAFT
18
t = 2nπ/W uniquely identifiesH from a single observationH(x(t)) as long as the time–bandwidth
product of the known input signalx(t) satisfies the conditionT W ≥ 2π(K + 1)2.
Proof: This corollary is simply a consequence of Theorem 2 and the fact thatKτKν,max ≤ (K+1)2
4 .
As to the validity of the fact thatKτKν,max ≤ (K+1)2
4 , first note that for any fixedK andKτ , we always
haveKν,max ≤ K − (Kτ − 1). To see this, observe that ifKν,max were greater thanK − (Kτ − 1)
then either∑Kτ
i=1Kν,i > K or there exists ani such thatKν,i = 0, both of which are contradictions.
Consequently, for any fixedK, we have that
KτKν,max ≤ −K2τ + (K + 1)Kτ (40)
and since the maximum of−K2τ + (K + 1)Kτ occurs atKτ = K+1
2 , we getKτKν,max ≤ (K+1)2
4 .
V. D ISCUSSION
In Section III and Section IV of this paper, we have proposed and analyzed a polynomial-time recovery
procedure that ensures identification of parametric underspread linear systems under certain conditions.
In particular, one of the key contributions of the precedinganalysis is that it parlays the low-rate sampling
results for time-delay estimation presented in [9] and the direction-of-arrival (DOA) estimation results
[10]–[12] into conditions on the time–bandwidth product,T W, of the input signalx(t) that guarantee
that the proposed recovery procedure can identify arbitrary linear systems as long as they are sufficiently
underspread. Specifically, in the parlance of system identification, Corollary 1 states that the recovery
procedure of Section III achievesinfinitesimally-fine resolutionin the delay–Doppler space as long as the
temporal degrees of freedom available to excite an underspread linear system are on the order ofΩ(K2).
In addition, we carry out extensive simulations in the latter part of the paper, which confirm that—as long
as the conditionT W ≥ 2π(K + 1)2 is satisfied—the ability of the proposed procedure to distinguish
between (resolve) closely spaced delay–Doppler pairs is primarily a function of the signal-to-noise ratio
(SNR) and its performance degrades gracefully in the presence of noise. Below, we compare these results
to some of the related ones in the literature on identification of parametric underspread linear systems
and also discuss an application of these results to super-resolution target detection using radar.
A. Comparisons with Previous Work
In order to best put the significance of the preceding resultsinto perspective, it is instructive to
compare them with some of the corresponding results reported in recent literature. There exists a large
body of existing work—especially in the communications andradar literature—treating identification of
August 5, 2010 DRAFT
19
parametric time-varying linear systems; see, e.g., [2], [13]–[18]. One of the approaches that is commonly
taken in many of these works, such as in [13], [15]–[18], is toquantize the delay–Doppler space(τ, ν)
by assuming that bothτi andνij lie on a grid. The following theorem is representative of some of the
known results in this case in the literature.4
Theorem 3 ([16], [17]). Suppose thatH is a parametric underspread linear system that is completely
described by a total ofK =∑Kτ
i=1Kν,i triplets (τi, νij , αij). Further, let the delays and the Doppler-shifts
of the system be such thatτi = riW−1 and νij = ℓijT −1 for ri ∈ Z+ and ℓij ∈ Z. ThenH can be
identified in polynomial-time from a single observationH(x(t)) as long as the system satisfies[A1]–[A3]
and the time–bandwidth product of the input signalx(t) satisfiesT W = Ω(K2/ log T W).
There are two conclusions that can be immediately drawn fromTheorem 3. First, both [16], [17]
require about the same scaling of the temporal degrees of freedom as that required by Corollary 1:
T W ≈ Ω(K2).5 Second, the resolution of the recovery procedures proposedin [16], [17] is limited to
W−1 in the delay space andT −1 in the Doppler space because of the assumption thatτi = riW−1 and
νij = ℓijT −1. Similarly, in another related recent paper [18], two recovery procedures are proposed that
have been numerically shown to uniquely identifyH as long asT W ≫ 1 and each(τi, νij) corresponds
to one of the points in the quantized delay–Doppler space with resolution proportional toW−1 andT −1
in the delay space and the Doppler space, respectively.
Finally, we conclude here by discussing the connections between the results reported in this paper and
somewhat related results obtained in [14]. In particular, [14] also leverages some of the results in DOA
estimation to propose a scheme for the identification of linear systems of the form (2) without requiring
that τi = riW−1 andνij = ℓijT −1. Nevertheless, the results presented in this paper differ from those in
[14] in three important respects. First, we explicitly state the relationship between the time–bandwidth
product T W of the input signalx(t) and the number of delay–Doppler pairsK =∑Kτ
i=1Kν,i that
guarantees recovery of the system response by explicitly studying the sampling stage and the recovery
stage of our proposed recovery procedure. On the other hand,the method proposed in [14] assumes
the sampling stage to be given and, as such, fails to make explicit the connection between the time–
4It is worth mentioning here that a somewhat similar result was also obtained independently in [34] in an abstract setting.
5Note that there is also a Bayesian variant of Theorem 3 in [17]that requiresT W ≈ Ω(K) under the assumption thatH
has a uniform statistical prior over the quantized delay–Doppler space. A somewhat similar Bayesian variant of Corollary 1 can
also be obtained by trivially extending the results of this paper to the case whenH is assumed to have a uniform statistical
prior over thenon-quantizeddelay–Doppler space.
August 5, 2010 DRAFT
20
bandwidth product ofx(t) and the number of delay-Doppler pairs. Second, the algorithm of [14] requires
searching over aK-dimensional continuous parameter space, which can be computationally prohibitive
for large-enough values ofK. Last, but not the least, this procedure is guaranteed to work as long as
there are no more than two delay–Doppler pairs having the same delay:maxiKν,i ≤ 2. In contrast, the
recovery algorithm proposed in this paper does not impose any such restrictions on the distribution of
(τi, νij) within the delay-Doppler space.
B. Application: Super-Resolution Radar
We have established in Section IV that the polynomial-time recovery procedure of Section III achieves
infinitesimally-fine resolution in the delay–Doppler spaceunder mild assumptions on the temporal degrees
of freedom of the input signal. This makes the proposed recovery procedure extremely useful for
application areas in which the system performance depends critically on the ability to resolve closely
spaced delay–Doppler pairs. In particular, our method can be used for super-resolution target detection
using radar. This is because the noiseless, clutter-free received signal in the case of monostatic radars
is exactly of the form (1) with each triplet(τk, νk, αk) corresponding to an echo of the radar waveform
x(t) from a distinct target [2].6 The fact that our recovery procedure allows us to identify arbitrary
parametric underspread linear systems, therefore, enables us to distinguish between multiple targets even
if their radial positions are quite close to each other and/or their radial velocities are quite similar to one
another—the so-called super-resolution detection of targets.
On the other hand, note that apart from the fact that none of the methods based on the assumption of a
quantized delay–Doppler space can ever carry out super-resolution target detection, a major drawback of
the radar target detection approach in works such as [17], [18] is that targets in the real-world do not in fact
correspond to points in the quantized delay–Doppler space,which causesleakageof their energies in the
quantized space. In order to elaborate further on this point, defineLdef= ⌈Wτmax⌉ andM
def= ⌈Tνmax/2⌉
and note that (canonical) quantization corresponds to transforming theC = [0, τmax]×[−νmax/2, νmax/2]
continuous delay–Doppler space into aQ def= 0, . . . , L×−M/2, . . . ,M/2 two-dimensional quantized
grid, which in turn transforms the received signalH(x(t)) at the radar into [35], [36, Chapter 4]
H(x(t)) ≈L∑
ℓ=0
M∑
m=−M
αℓmx(t− τℓ)ej2πνmt (41)
6In the radar literature, the term “monostatic” refers to thecommon scenario of the radar transmitter and the radar receiver
being collocated.
August 5, 2010 DRAFT
21
Fig. 3. Quantized representation of four targets (represented by ∗) in the delay–Doppler space withτmax = 10 µs and
νmax = 10 kHz. The quantized delay–Doppler approximation of the targets corresponds toW = 1.6 MHz andT = 0.32 ms.
where we have thatαℓmdef=
∑Kτ
i=1
∑Kν,i
j=1 αijejπ(m−T νij)sinc(m − T νij)sinc(ℓ − TWτi) and that the
quantized delay–Doppler pairs(τℓ, νm) ∈ Q. It is now easy to conclude from (41) that, unless the
original targets (delay–Doppler pairs) happen to lie inQ, most of the attenuation factorsαℓm would
be nonzero because of the sinc kernels—the so-called “leakage effect.” This has catastrophic implications
for target detection using radar since leakage makes it impossible to reliably identify the original set of
delays and Doppler-shifts. This limitation of target-detection methods that are based on the assumption
of a quantized delay–Doppler space is also depicted in Fig. 3for the case of four hypothetical targets.
The figure illustrates that each of the four non-quantized targets not only contributes energy to its own
(τℓ, νm) in Q but also leaks its energy to the nearby points in the quantized space.
Owing to the fact that leakage can cause missed detections and false alarms, conventional radar literature
in fact tends to focus only on recovery procedures that do no impose any structure on the distribution
of (τi, νij) within the delay–Doppler space. The most commonly used approach in the radar signal
processing literature corresponds to matched-filtering (MF) the received signal with the input signalx(t)
in the delay–Doppler space [2]. Specifically, the MF outputχ(τ, ν) takes the form
χ(τ, ν)def=
∫
R
H(x(t))x∗(t− τ) exp(−j2πνt)dt =Kτ∑
i=1
Kν,i∑
j=1
αijA(τ − τi, ν − νij) (42)
whereA(τ, ν)def=
∫Rx(t)x∗(t − τ) exp(−j2πνt)dt is termed as the Woodward’sambiguity functionof
x(t). It can be easily deduced from (42) that the resolution of theMF-based recovery procedure is tied
August 5, 2010 DRAFT
22
(a)
(b)
Fig. 4. Comparison between the target-detection performance of matched-filtering and our proposed recovery procedurefor
the case of four targets (represented by∗) in the delay–Doppler space withτmax = 10 µs, νmax = 10 kHz, W = 1.6 MHz,
andT = 0.32 ms. The probing sequencexn corresponds to a random binary(±1) sequence withN = 32, the pulseg(t)
is designed to have a nearly-flat frequency response in the working spectral bandF , and the pulse repetition interval is taken
to beT = 10 µs. (a) Target detection by matched-filtering the received signal H(x(t)) with the input signalx(t). (b) Target
detection using the proposed recovery procedure withp = 16.
to the supportof the ambiguity function in the delay–Doppler space. Ideally, one would like to have
A(τ, ν) = δ(τ)δ(ν) for super-resolution detection of targets but two fundamental properties of ambiguity
functions, namely,|A(0, 0)|2 =∫|x(t)|2dt and
∫∫|A(τ, ν)|2dτdν =
∫|x(t)|2dt, dictate that no real-
world signalx(t) can yield infinitesimally-fine resolution in this case either [2]. In fact, the resolution
August 5, 2010 DRAFT
23
of MF-based recovery techniques also tends to be on the orderof W−1 andT −1 in the delay space and
the Doppler space, respectively, which severely limits their ability to distinguish between two closely-
spaced targets in the delay–Doppler space. This inability of MF-based methods to resolve closely-spaced
delay–Doppler pairs is depicted in Fig. 4. Specifically, Fig. 4 compares the target-detection performance
of MF and the recovery procedure proposed in this paper for the case of four closely-spaced targets. It
is easy to see from Fig. 4(a) that matched-filtering the received signalH(x(t)) with the input signalx(t)
only gives rise to two peaks that are not even centered at the true targets. On the other hand, Fig. 4(b)
illustrates that our proposed recovery procedure perfectly identifies the locations of all four of the targets
in the delay–Doppler space.
VI. N UMERICAL EXPERIMENTS
In this section, we explore various issues using numerical experiments that were not treated theoretically
earlier in the paper. These include robustness of the proposed recovery procedure in the presence of noise
and the effects of truncated digital filters, use of finite number of samples, and model-order mismatch on
the recovery performance. Throughout this section, the numerical experiments correspond to a parametric
underspread linear systemH that is described by a total ofK = 4 delay–Doppler pairs withKτ = 2 and
Kν,1 = Kν,2 = 2. The locations of these pairs in the delay–Doppler space aregiven by Fig. 5, while the
attenuation factors associated with each of the four delay–Doppler pairs are taken to have unit amplitudes
and random phases.
In order to identifyH, we design the prototype pulseg(t) to have a constant frequency response over
the working spectral bandF =[− π
T p,πT p
]with p = 4 andT = 10 µs, that is,
G(ω) ≈
1, ω ∈ F
0, ω /∈ F .(43)
In other words, the input signalx(t) is chosen to have bandwidthW = 8πT . In addition, we use a
probing sequencexn corresponding to a random binary(±1) sequence withN = 16, which leads to
a time–bandwidth product ofT W ≈ 128π. Note that the chosen time–bandwidth product here is more
than the lower bound of Theorem 2 by a factor of4 so as to increase the robustness to noise. Also,
unless otherwise stated, all experiments in the following use an ideal (flat) LPF as the sampling filter
(cf. Fig. 2). Further, note that we make use of the ESPRIT method described in Section III for recovery
of the delays, while we make use of the matrix-pencil method [11] for recovery of the corresponding
Doppler-shifts. Finally, the performance metrics that we use in this section are the mean squared error
August 5, 2010 DRAFT
24
Fig. 5. Delay–Doppler representation of the parametric underspread linear systemH corresponding toK = 4 delay–Doppler
pairs withτmax = 10 µs andνmax = 10 kHz.
(MSE) of the estimated delays and the estimated Doppler-shifts (averaged over100 noise realizations),
defined as
e2delay =1
2
2∑
i=1
[(τi − τi)/τmax
]2, (44)
and
e2Doppler =1
4
2∑
i=1
2∑
j=1
[(νij − νij)/νmax
]2, (45)
whereτi and νij denote the estimated delays and the estimated Doppler-shifts, respectively.
1) Robustness to Noise:The first numerical experiment that we carry out examines therobustness
of the proposed recovery procedure when the received signalH(x(t)) is corrupted by additive noise.
The results of this experiment are shown in Fig. 6, which plots the MSE of the estimated delays and
the estimated Doppler-shifts as a function of the signal-to-noise ratio (SNR). It can be easily seen from
the figure that the ability to resolve delay–Doppler pairs isprimarily a function of the SNR and its
performance degrades gracefully in the presence of noise.
2) Effects of Truncated Digital-Correction Filter Banks:Recall from Section III that our recovery
method is composed of various digital-correction stages (see also Fig. 2). The filters used in these
stages, which includeφℓ[n] andψℓ[n], have infinite impulse responses in general so that practical
implementations of these filters require truncation of their impulse responses. Note further that the
August 5, 2010 DRAFT
25
Fig. 6. Mean squared error of the estimated delays and the estimated Doppler-shifts as a function of the signal-to-noiseratio.
(a) (b)
Fig. 7. Mean squared error of the estimated delays and the estimated Doppler-shifts as a function of the signal-to-noiseratio
for various lengths of the impulse responses of the filters.
truncated lengths of these filters also determine the (computational) delay and the computational load of
the proposed procedure. The next numerical experiment we carry out examines the effects of impulse
response truncation on the performance.
The results of this numerical experiment are reported in Fig. 7, which plots the MSE of the estimated
delays (Fig. 7(a)) and the MSE of the estimated Doppler-shifts (Fig. 7(b)) as a function of the SNR for
various lengths of the impulse responses of the filters. There are two important remarks that can be made
from looking at the results of this experiment. First, for a fixed length of the impulse responses, there is
always some SNR beyond which the estimation error caused by the truncation of the impulse responses
August 5, 2010 DRAFT
26
(a) (b)
Fig. 8. Mean squared error of the estimated delays and the estimated Doppler-shifts as a function of the signal-to-noiseratio
for different numbers of samples collected at the output of the sampling filter (corresponding to an ideal low-pass filter).
becomes more dominant than the error caused by the additive noise (as evident by the error floors
in Fig. 7). Second, and perhaps most importantly, filters with 49 taps seem to provide good estimation
accuracy up to an SNR of60 dB, whereas filters with even just25 taps yield good estimation performance
at SNRs below40 dB.
3) Effects of Finite Number of Samples:It can be seen from Fig. 2 that the sampling filter used at
the front-end of the proposed recovery procedure is bandlimited in nature and, therefore, it has infinite
support in the time domain. Consequently, our proposed procedure theoretically requires collecting an
infinite number of samples at the back-end of the sampling filter. The next numerical experiment that we
carry out examines the effect of collecting instead a finite number of samples on the performance.
The results of this numerical experiment are reported in Fig. 8, which plots the MSE of the estimated
delays (Fig. 8(a)) and the MSE of the estimated Doppler-shifts (Fig. 8(b)) as a function of the SNR for
different numbers of samples collected at the output of the sampling filter (corresponding to an ideal
LPF). As in the case of truncation of digital-correction filter banks, it can also be seen from Fig. 8 that
there is always some SNR for every fixed number of samples beyond which the estimation error caused
by the finite number of samples becomes more dominant than theerror caused by the additive noise.
Equally importantly, however, note that the signalx(t) in these experiments is comprised ofN = 16
prototype pulses. Therefore, under the assumption ofp = 4 samples per pulse periodT , it is clear that
we require at leastN · p = 64 samples in total to represent just the input signalx(t). On the other hand,
Fig. 8 shows that collecting128 samples, which is twice the minimum number of samples required,
provides good (delay and Doppler) estimation accuracy for SNRs up to70 dB.
August 5, 2010 DRAFT
27
Fig. 9. Frequency response of the raised-cosine filter with roll-off factor equal to1.
Finally, it is worth noting here that making use of an ideal LPF as the sampling filter requires collecting
relatively more samples at the filter back-end due to the slowly-decaying nature of the sinc kernel.
Therefore, in order to reduce the number of samples requiredat the back-end of the sampling filter for
reasonable estimation accuracy, we can instead make use of sampling filters whose (time-domain) kernels
decay faster than the sinc kernel. One such possible choice for the sampling filter is the raised-cosine
filter with roll-off factor equal to1, whose frequency response is given by
S(ω) =
p2T
(1 + cos(Tp ω)
), ω ∈ F
0, otherwise.(46)
It is a well-known fact (and can be easily checked) that this filter decays faster in the time domain than
the sinc kernel. However, the main issue here is that the raised-cosine filter does not satisfy Condition 2
in Section III, since its frequency response is not bounded away from zero at the ends of the spectral
bandF (see, e.g., Fig. 9).
However, we now show that this problem with the raised-cosine filter can be overcome by slightly
increasing the sampling rate and the bandwidth requirementstated in Section IV. Specifically, note
that Proposition 1 requires that the parameterp, which controls the minimal bandwidth ofx(t) and
the sampling rate of our proposed procedure, satisfiesp ≥ 4 under the current simulation setup (since
Kτ = 2). We now instead choosep = 6 and argue that the raised-cosine filter can be successfully used
under this choice ofp. To this end, recall from Section III that the function of thedigital-correction
August 5, 2010 DRAFT
28
(a) (b)
Fig. 10. Mean squared error of the estimated delays and the estimated Doppler-shifts as a function of the signal-to-noise
ratio for different numbers of samples collected at the output of the sampling filter (corresponding to the raise-cosinefilter with
roll-off factor equal to1).
filters ψ1[n] andψ6[n] is to invert the frequency response of the sampling kernel corresponding to the
frequency bands denoted by1 and 6 on Fig. 9, respectively (under the assumption that the pulseg(t)
has a flat frequency response). In the case of the raised-cosine filter, however, we cannot compensate
for the non-flat nature of these two bands due to the fact that they are not bounded away from zero.
Nevertheless, because of the fact that we are usingp = 6, we can simply disregard the channels1 and
6 after the first digital-correction stage and work with the rest of the four channels (2-4) only. We make
use of this insight to repeat the last numerical experiment using the raised-cosine filter (instead of an
ideal LPF) and report the results in Fig. 10. It is easy to see from Fig. 10 that, despite increasingp to
6, the raised-cosine filter performs better than an ideal LPF using fewer samples.
4) Effects of Model-Order Mismatch:The final numerical experiment that we carry out studies the
situation where the conditions of Theorem 2 do not exactly hold. To this end, we simulate identification
of a parametric underspread linear systemH with Kτ = 4 distinct delays. Further, for the first3 delays
we takeKν,i = 2, i = 1, 2, 3, whereas we chooseKν,4 = 8 for the last delay. Finally, we take the
prototype pulseg(t) as at the start of this section (with bandwidthW = 8πT ), but we use a probing
sequencexn corresponding to a random binary(±1) sequence withN = 8. Clearly, this does not
satisfy the conditions of Theorem 2 because of the large number of Doppler-shifts associated with the
last delay(Kν,4 = 8).
The results of this numerical experiment are reported in Fig. 11. It can be easily seen from the figure
that, despite the fact that the input signal does not satisfythe conditions of Theorem 2, our algorithm
August 5, 2010 DRAFT
29
Fig. 11. Effects of model-order mismatch on the performanceof the proposed recovery procedure corresponding to a parametric
underspread linear systemH with K = 14 delay–Doppler pairs.
successfully recovers the first three delays and the corresponding Doppler-shifts. In addition, the procedure
also recovers the fourth delay correctly but (as expected) fails to recover all the Doppler-shifts associated
with the last delay. Note that in addition to demonstrating the robustness of our procedure in the presence
of model-order mismatch, this experiment also highlights the advantage of the sequential nature of our
approach where we first recover the delays and then estimate the Doppler-shifts and the attenuation factors
associated with the recovered delays. The main advantage ofthis being that if the input signal does not
satisfyN ≥ 2Kνifor somei then the recovery of the Doppler-shifts fails only for the ones associated
with the ith delay. Moreover, the recovery of theith delay itself does not suffer from the mismodeling
and it will be recovered correctly as long as the bandwidth ofthe input signal is not too small.
VII. C ONCLUSION
In this paper, we have revisited the problem of identification of underspread linear systems that are
completely described by a set of discrete delays and Doppler-shifts and are termed parametric underspread
linear systems. We have established that sufficiently underspread parametric linear systems are identifiable
as long as the time–bandwidth product of the input signal is proportional to the square of the total number
of delay–Doppler pairs. In addition, we have also concretely specified the nature of the input signal and the
structure of a corresponding polynomial-time recovery procedure that enable identification of parametric
underspread linear systems. Extensive simulation resultsconfirm that—as long as the time–bandwidth
August 5, 2010 DRAFT
30
product of the input signal satisfies the requisite conditions—the proposed recovery procedure is quite
robust to noise and other implementation issues. This makesour algorithm extremely useful for application
areas in which the system performance depends critically onthe ability to resolve closely spaced delay–
Doppler pairs. In particular, our method for identifying parametric underspread linear systems can be
used for super-resolution target detection using radar.
REFERENCES
[1] J. G. Proakis,Digital Communications, 4th ed. New York, NY: McGraw-Hill, 2001.
[2] M. I. Skolnik, Introduction to Radar Systems, 3rd ed. New York, NY: McGraw-Hill, 2001.
[3] W. Kozek and G. E. Pfander, “Identification of operators with bandlimited symbols,”SIAM J. Math. Anal., vol. 37, no. 3,
pp. 867–888, 2005.
[4] T. Kailath, “Measurements on time-variant communication channels,”IRE Trans. Inform. Theory, vol. 8, no. 5, pp. 229–236,
Sep. 1962.
[5] P. Bello, “Measurement of random time-variant linear channels,”IEEE Trans. Inform. Theory, vol. 15, no. 4, pp. 469–475,
Jul. 1969.
[6] G. E. Pfander and D. F. Walnut, “Measurement of time-variant linear channels,”IEEE Trans. Inform. Theory, vol. 52,
no. 11, pp. 4808–4820, Nov. 2006.
[7] W. U. Bajwa, J. Haupt, A. M. Sayeed, and R. Nowak, “Compressed channel sensing: A new approach to estimating sparse
multipath channels,”Proc. IEEE, vol. 98, no. 6, pp. 1058–1076, Jun. 2010.
[8] D. Slepian, “On bandwidth,”Proc. IEEE, vol. 64, no. 3, pp. 292–300, Mar. 1976.
[9] K. Gedalyahu and Y. C. Eldar, “Time-delay estimation from low-rate samples: A union of subspaces approach,”IEEE
Trans. Signal Processing, vol. 58, no. 6, pp. 3017–3031, Jun. 2010.
[10] R. Roy, A. Paulraj, and T. Kailath, “ESPRIT—A subspace rotation approach to estimation of parameters of cisoids in
noise,” IEEE Trans. Acoust., Speech, Signal Processing, vol. 34, no. 5, pp. 1340–1342, Oct. 1986.
[11] Y. Hua and T. K. Sarkar, “Matrix pencil method for estimating parameters of exponentially damped/undamped sinusoids
in noise,” IEEE Trans. Acoust., Speech, Signal Processing, vol. 38, no. 5, pp. 814–824, May 1990.
[12] D. W. Tufts and R. Kumaresan, “Estimation of frequencies of multiple sinusoids: Making linear prediction perform like
maximum likelihood,”Proc. IEEE, vol. 70, no. 9, pp. 975–989, Sep. 1982.
[13] A. W. Habboosh, R. J. Vaccaro, and S. Kay, “An algorithm for detecting closely spaced delay/Doppler components,” in
Proc. IEEE Int. Conf. Acoustics, Speech and Signal Processing (ICASSP’97), Munich, Germany, Apr. 1997, pp. 535–538.
[14] A. Jakobsson, A. L. Swindlehurst, and P. Stoica, “Subspace-based estimation of time delays and Doppler shifts,”IEEE
Trans. Signal Processing, vol. 46, no. 9, pp. 2472–2483, Sep. 1998.
[15] I. C. Wong and B. L. Evans, “Low-complexity adaptive high-resolution channel prediction for OFDM systems,” inProc.
IEEE Global Telecommunications Conf. (GLOBECOM’06), San Francisco, CA, Dec. 2006, pp. 1–5.
[16] W. U. Bajwa, A. M. Sayeed, and R. Nowak, “Learning sparsedoubly-selective channels,” inProc. 45th Annu. Allerton
Conf. Communication, Control, and Computing, Monticello, IL, Sep. 2008, pp. 575–582.
[17] M. A. Herman and T. Strohmer, “High-resolution radar via compressed sensing,”IEEE Trans. Signal Processing, vol. 57,
no. 6, pp. 2275–2284, Jun. 2009.
August 5, 2010 DRAFT
31
[18] X. Tan, W. Roberts, J. Li, and P. Stoica, “Range-Dopplerimaging via a train of probing pulses,”IEEE Trans. Signal
Processing, vol. 57, no. 3, pp. 1084–1097, Mar. 2009.
[19] M. Mishali, Y. C. Eldar, O. Dounaevsky, and E. Shoshan, “Xampling: Analog to digital at sub-Nyquist rates,” CCIT
Report # 751, EE Dept., Technion—Israel Institute of Technology, Dec. 2009. [Online]. Available: arXiv:0912.2495
[20] M. Mishali, Y. C. Eldar, and A. Elron, “Xampling: Signalacquisition and processing in union of subspaces,” CCIT
Report # 747, EE Dept., Technion—Israel Institute of Technology, Oct. 2009. [Online]. Available: arXiv:0911.0519
[21] M. Mishali and Y. C. Eldar, “From theory to practice: Sub-Nyquist sampling of sparse wideband analog signals,”IEEE J.
Select. Topics Signal Processing, vol. 4, no. 2, pp. 375–391, Apr. 2010.
[22] ——, “Blind multi-band signal reconstruction: Compressed sensing for analog signals,”IEEE Trans. Signal Processing,
vol. 57, no. 3, pp. 993–1009, Mar. 2009.
[23] Y. C. Eldar, “Compressed sensing of analog signals in shift-invariant spaces,”IEEE Trans. Signal Processing, vol. 57,
no. 8, pp. 2986–2997, Aug. 2009.
[24] Y. C. Eldar and M. Mishali, “Robust recovery of signals from a structured union of subspaces,”IEEE Trans. Inform.
Theory, vol. 55, no. 11, pp. 5302–5316, Nov. 2009.
[25] E. Moulines, P. Duhamel, J. Cardoso, S. Mayrargue, and T. Paris, “Subspace methods for the blind identification of
multichannel FIR filters,”IEEE Trans. Signal Processing, vol. 43, no. 2, pp. 516–25, Feb. 1995.
[26] H. Krim and M. Viberg, “Two decades of array signal processing research: The parametric approach,”IEEE Signal
Processing Mag., vol. 13, no. 4, pp. 67–94, Jul. 1996.
[27] P. Stoica and R. L. Moses,Introduction to Spectral Analysis. Englewood Cliffs, NJ: Prentice-Hall, 1997.
[28] R. Roy and T. Kailath, “ESPRIT—Estimation of signal parameters via rotational invariance techniques,”IEEE Trans.
Acoust., Speech, Signal Processing, vol. 37, no. 7, pp. 984–995, Jul. 1989.
[29] T.-J. Shan, M. Wax, and T. Kailath, “On spatial smoothing for direction-of-arrival estimation of coherent signals,” IEEE
Trans. Acoust., Speech, Signal Processing, vol. 33, no. 4, pp. 806–811, Aug. 1985.
[30] P. L. Dragotti, M. Vetterli, and T. Blu, “Sampling moments and reconstructing signals of finite rate of innovation: Shannon
meets Strang-Fix,”IEEE Trans. Signal Processing, vol. 55, no. 5, pp. 1741–1757, May 2007.
[31] M. Vetterli, P. Marziliano, and T. Blu, “Sampling signals with finite rate of innovation,”IEEE Trans. Signal Processing,
vol. 50, no. 6, pp. 1417–1428, Jun. 2002.
[32] R. Tur, K. Gedalyahu, and Y. C. Eldar, “Multichannel sampling of pulse streams at the rate of innovation,” submittedto
IEEE Trans. Signal Processing. [Online]. Available: arXiv:1004.5070
[33] R. Tur, Y. C. Eldar, and Z. Friedman, “Low rate sampling of pulse streams with application to ultrasound imaging,”
submitted toIEEE Trans. Signal Processing. [Online]. Available: arXiv:1003.2822
[34] G. E. Pfander and H. Rauhut, “Sparsity in time-frequency representations,”J. Fourier Anal. Appl., Aug. 2009.
[35] P. Bello, “Characterization of randomly time-variantlinear channels,”IEEE Trans. Commun., vol. 11, no. 4, pp. 360–393,
Dec. 1963.
[36] W. U. Bajwa, “New information processing theory and methods for exploiting sparsity in wireless systems,” Ph.D.
dissertation, University of Wisconsin-Madison, Madison,WI, 2009.
August 5, 2010 DRAFT