I.

INTRODUCTION

The goal of array processing is to estimate the

Source Localization and locations and waveforms of sources by combining
the received data from multiple sensors so that the
Sensing: A Nonparametric desired signal is enhanced, while the unwanted
signals, such as interference and noise, are suppressed.
Iterative Adaptive Approach In active sensing applications such as radar/sonar
range-Doppler imaging, the aim is to find targets
Based on Weighted Least present in a region of interest. For channel estimation
in communications, the aim is to estimate the non-zero
Squares channel taps and their Doppler shifts, which are then
fed to the subsequent equalizer for symbol detection.
TARIK YARDIBI, Student Member, IEEE Low numbers of snapshots and low signal-to-noise
JIAN LI, Fellow, IEEE ratios (SNR) are among the many challenges that
University of Florida array processing systems frequently face. Another
PETRE STOICA, Fellow, IEEE challenge is the presence of nearby sources, in terms
Uppsala University of location or Doppler, since closely spaced sources
Sweden are harder to discriminate. Herein both passive array
MING XUE, Student Member, IEEE processing and active sensing are considered.
University of Florida The most basic approach to array processing
ARTHUR B. BAGGEROER, Fellow, IEEE is the classical delay-and-sum (DAS) method,
Massachusetts Institute of Technology in which the received signal from each sensor is
Array processing is widely used in sensing applications weighted and delayed so as to focus on different
for estimating the locations and waveforms of the sources in
a given field. In the absence of a large number of snapshots, points in space. However, this method suffers from
which is the case in numerous practical applications, such low resolution and high sidelobe levels. There is a
as underwater array processing, it becomes challenging to
estimate the source parameters accurately. This paper presents vast amount of literature on methods that provide
a nonparametric and hyperparameter, free-weighted, least superior performance over the DAS approach when
squares-based iterative adaptive approach for amplitude and
phase estimation (IAA-APES) in array processing. IAA-APES can certain assumptions are met [1]. The well-known
work well with few snapshots (even one), uncorrelated, partially standard Capon beamformer (SCB) [2] and multiple
correlated, and coherent sources, and arbitrary array geometries.
IAA-APES is extended to give sparse results via a model-order signal classification (MUSIC) [3, 4] methods provide
selection tool, the Bayesian information criterion (BIC). Moreover, superresolution when the sources are uncorrelated and
it is shown that further improvements in resolution and accuracy the number of snapshots is high. Many extensions
can be achieved by applying the parametric relaxation-based
cyclic approach (RELAX) to refine the IAA-APES&BIC to these methods have been proposed to deal with
estimates if desired. IAA-APES can also be applied to active modelling errors, such as steering vector mismatches.
sensing applications, including single-input single-output
(SISO) radar/sonar range-Doppler imaging and multi-input (See, e.g., [5]—[9].) However, none of these methods
single-output (MISO) channel estimation for communications. is able to cope with very low snapshot numbers,
Simulation results are presented to evaluate the performance of
IAA-APES for all of these applications, and IAA-APES is shown coherent or highly correlated sources, or severe noise.
to outperform a number of existing approaches. Only a few snapshots are available when the
environment being sensed by the array is stationary
Manuscript received December 11, 2007; revised July 21 and
November 4, 2008; released for publication November 16, 2008. for a short duration of time. Moreover, to avoid
IEEE Log No. T-AES/46/1/935952. smearing, i.e., losing resolution because of wide
Refereeing of this contribution was handled by W. Koch. main beamwidths in the array response, averaging
This work was supported in part by the Office of Naval Research
can only be done over a small bandwidth [10, 11].
(ONR) under Grants N00014-07-1-0193, N00014-07-1-0293, Therefore the number of available snapshots, which is
and N00014-01-1-0257, the Army Research Office (ARO) directly related to the time-bandwidth product, can be
under Grant W911NF-07-1-0450, the National Aeronautics and
Space Administration (NASA) under Grant NNX07AO15A, the very small, sometimes as small as 3, for applications
National Science Foundation (NSF) under Grants CCF-0634786, such as underwater array processing. Furthermore,
ECS-0621879, and ECS-0729727, the Swedish Research Council
(VR), and the European Research Council (ERC). as discussed later, the data models for single-input
Opinions, interpretations, conclusions, and recommendations are single-output (SISO) radar/sonar range-Doppler
those of the authors and are not necessarily endorsed by the United imaging and multi-input single-output (MISO) channel
States Government.
estimation in communication problems are similar
Authors’ addresses: T. Yardibi, J. Li, M. Xue, Dept. of Electrical
and Computer Engineering, University of Florida, PO Box 116130, to the model used in array processing with a single
Gainesville, FL 32611, E-mail: ([email protected]); P. Stoica, Dept. snapshot and an arbitrary array geometry (dictated by
of Information Technology, Uppsala University, Uppsala, Sweden;
A. B. Baggeroer, Depts. of Mechanical Engineering & Electrical the probing waveforms).
Engineering and Computer Science, Massachusetts Institute of The array processing problem has been carried into
Technology, Cambridge, MA 02139. the sparse signal representation area by noticing that
c 2010 IEEE
0018-9251/10/$26.00 ° the number of actual sources is usually much smaller

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010 425

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
 TABLE I
Notation used in the Text

k ¢ k0 `0 -norm
k ¢ k1 `1 -norm
k ¢ k2 `2 -norm
k ¢ kF Frobenius norm of a matrix
¯ the Hadamard (elementwise) matrix product
tr(¢) trace of a matrix
(¢)T transpose of a vector or matrix
(¢)H conjugate transpose of a vector or matrix
Fig. 1. Far-field linear array.

than the number of potential source points that can

be considered [12—29]. Sparsity-based techniques used to estimate the desired source characteristics.
have also been used in spectral estimation [30], image This section first introduces the data model for such
processing, and array design [31—33], among many applications and then the IAA-APES algorithm in this
other application areas. Sparse signal representation model.
algorithms can deal with a few snapshots (even one).
However, they may require large computation times A. Data Model
and the fine tuning of one or more hyperparameters. Consider the wavefield generated by K sources
Application areas, ranging from passive source ¢
localization to active radar/sonar range-Doppler located at μ, where μ =[μ1 , μ2 , : : : , μK ] and μk 1 are the
imaging and channel estimation for communications location parameters of the kth signal, k = 1, : : : , K. In
are addressed here. Section II considers passive the narrowband, multi-snapshot case, the M £ 1 array
sensing applications and describes a weighted output vector of an M element array in the presence
least squares-based iterative adaptive approach of additive noise can be represented as [5, 43]
for amplitude and phase estimation (IAA-APES). y(n) = A(μ)s(n) + e(n), n = 1, : : : , N (1)
The algorithm is named IAA-APES herein since
its derivation resembles that of the amplitude and where N is the number of snapshots, A(μ) is the
¢
phase estimation (APES) algorithm [34—36]. (See M £ K steering matrix defined as A(μ) =[a(μ1 ),
Section IIC.) This name also distinguishes it from the ¢
a(μ2 ), : : : , a(μK )], and s(n) =[s1 (n), s2 (n), : : : , sK (n)]T ,
maximum likelihood (ML)-based iterative adaptive
approach (IAA-ML) discussed in the Appendix to n = 1, : : : , N, is the source waveform vector at time n.
further motivate IAA-APES. Using the Bayesian The array steering matrix has different expressions,
information criterion (BIC) [37, 38], IAA-APES can depending on the array geometry and on whether the
be extended to yield point source estimates. (This source is in the near-field or far-field of the array. For
approach is referred to as IAA-APES&BIC.) Next, a instance, the steering vector corresponding to the kth
parametric relaxation-based cyclic approach, namely source for a far-field linear array (where μk represents
RELAX [39, 40], is discussed as a way to further the impinging angle of source k in this case) is given
refine the results of IAA-APES&BIC. (This approach by
is referred to as IAA-APES&RELAX.) Section III a(μk ) = [e¡j(2¼f=c0 )x1 cos(μk ) , : : : , e¡j(2¼f=c0 )xM cos(μk ) ]T
presents the data models for the aforementioned active
sensing applications and emphasizes the similarities (2)
to the passive sensing case. IAA-APES is evaluated where f is the center frequency, c0 is the wave
via comprehensive simulations in Section IV, and the propagation velocity, and xm is the position of the
IAA-APES’ performance is compared with that of a mth sensor, m = 1, : : : , M. (See Fig. 1.) Note that for a
number of existing approaches. planar array or near-field sources, only the expressions
Notation: We denote vectors and matrices by for the steering vectors have to be modified; the
boldface lowercase and boldface uppercase letters, algorithms that are presented can be applied without
respectively. The kth component of a vector x is any modifications, since a(μ) is assumed to be a
written as xk . The kth diagonal element of a matrix known function of μ. The number of sources, K, is
P is written as Pk . See Table I for other symbols and usually unknown; hence, here, K is considered to be
their meanings. the number of scanning points in the region. In other
words every point of a predefined grid that covers
II. PASSIVE SENSING the region of interest is considered as a potential
source whose power is estimated. Consequently,
In passive sensing applications such as
aeroacoustic noise measurements [41] and underwater 1 With a slight abuse of notation, we do not use bold font for μk ,
acoustic measurements [42], an array of sensors is k = 1, : : : , K, which might be multi-dimensional, for simplicity.

426 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
K will be much larger than the actual number of [26]. Fuchs [27, 28] uses a sparsity-constrained
sources present, and only a few components of deconvolution approach that assumes the sources
fs(n)g will be non-zero. This is the main reason why are uncorrelated and that the number of snapshots
sparse algorithms can be used in array processing is large. The sparsity-constrained solution is
applications. obtained with a LASSO or BP type of algorithm.
Reference [29] introduces two hyperparameter
B. Related Work free deconvolution algorithms exploiting sparsity:
a sparsity-based extension to the deconvolution
We focus our attention on array processing approach for the mapping of acoustic sources
algorithms exploiting sparsity, which have (DAMAS) [45] (which is similar to [27] and [28]
gained noticeable interest recently. Sparse signal and widely used in practice) and a sparsity based
representation aims at finding the sparsest s, such covariance-matrix fitting approach. Extensions
that y = As is satisfied, i.e., to minimize ksk0 , such to the correlated source case are also provided.
that y = As, where A is known and y is measured. However, the methods in [29] are based on the
The problem in its original form is a combinatorial sample covariance matrix, and hence, these
problem and is nondeterministic polynomial-time methods do not work well with a limited number of
(NP) hard, making it impractical [12]. Fortunately, snapshots.
when s is sufficiently sparse [12—15], ksk0 can
be replaced by ksk1 , which leads to a convex C. IAA-APES
optimization problem that can be solved much more
easily by using, for instance, the least absolute IAA-APES is a data-dependent, nonparametric
shrinkage and selection operator (LASSO) [16] or algorithm based on a weighted least squares (WLS)
basis pursuit (BP) [17] algorithms. Alternatively, the approach. Let P be a K £ K diagonal matrix, whose
focal underdetermined system solution (FOCUSS) diagonal contains the power at each angle on the
algorithm [18], which is derived using Lagrange scanning grid. Then P can be expressed as
multipliers, can be used to iteratively solve the N
sparse problem. A Bayesian approach, such as 1X
Pk = jsk (n)j2 , k = 1, : : : , K: (3)
sparse Bayesian learning (SBL) [19, 20] or the N
n=1
approach in [21], can also be used to estimate s.
Furthermore, define the interference (signals at angles
The algorithms mentioned up to this point are for
other than the angle of current interest μk ) and noise
the single-snapshot case only. The extensions of
covariance matrix Q(μk ) to be
FOCUSS and SBL to the multiple-snapshot case
are M-FOCUSS [22] and M-SBL [23], respectively. Q(μk ) = R ¡ Pk a(μk )aH (μk ) (4)
Another algorithm, the `1 -SVD (singular value ¢
decomposition) algorithm [24, 25] is similar where R = A(μ)PAH (μ). Then the WLS cost function
to BP or LASSO, but this algorithm can work is given by (see, e.g., [34]—[36] and [43])
with multiple snapshots. (For the single-snapshot N
case, `1 -SVD becomes a LASSO and BP type X
ky(n) ¡ sk (n)a(μk )k2Q¡1 (μk ) (5)
of method.) M-FOCUSS requires the tuning of
n=1
two hyperparameters, which might affect the
¢
performance of the algorithm significantly. `1 -SVD where kxk2Q¡1 (μk ) = xH Q¡1 (μk )x and sk (n) represents the
requires the tuning of a hyperparameter and an
signal waveform at angle μk and at time n. Minimizing
estimate for the number of sources. Moreover,
(5) with respect to sk (n), n = 1, : : : , N, yields
implementing `1 -SVD requires convex optimization
software, such as SeDuMi [44]. M-SBL does aH (μk )Q¡1 (μk )y(n)
ŝk (n) = , n = 1, : : : , N: (6)
not require any hyperparameters. However, aH (μk )Q¡1 (μk )a(μk )
M-SBL converges quite slowly in its original form
[19, 20, 23]. This looks like the result that would be obtained by
Besides the above algorithms, which are the employing APES [34—36], but it is actually different
focus of our attention in the numerical examples, than APES since APES obtains Q(μk ) from the data
there are other sparsity based approaches worth by forming subapertures, while IAA-APES computes
mentioning. Reference [26] adds an additional spatial Q(μk ), as in (4). Moreover, IAA-APES is iterative, but
sparsity regularizing term (an `2 -norm constraint) APES is not, and APES cannot be used with arbitrary
to the `1 -norm constraint, and it minimizes a cost array geometries.
function similar to that of `1 -SVD. However, this Using (4) and the matrix inversion lemma, (6) can
method has two hyperparameters, assumes that be written as
the source waveforms can be represented by a aH (μ )R¡1 y(n)
ŝk (n) = H k ¡1 , n = 1, : : : , N: (7)
sparse basis, and has high computational complexity a (μk )R a(μk )

YARDIBI ET AL.: SOURCE LOCALIZATION AND SENSING: A NONPARAMETRIC ITERATIVE ADAPTIVE APPROACH 427

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
 TABLE II than necessary reduction of the number of degrees
The IAA-APES Algorithm of freedom when some of the interfering sources are
X
N coherent since the cancellation of multiple coherent
1
P̂k = jaH (μk )y(n)j2 , k = 1, : : : , K interfering sources would require only one DOF if
(aH (μk )a(μk ))2 N
n=1 the correct structure of P were known. However, we
repeat do not assume that the true structure of P is known.
R = A(μ)P̂AH (μ) Moreover, it is the diagonal structure of P assumed by
for k = 1, : : : , K
IAA-APES that makes the algorithm work properly
aH (μk )R¡1 y(n)
ŝk (n) = , n = 1, : : : , N even for low number of snapshot cases and coherent
aH (μk )R¡1 a(μk )
sources.
1 X
N
P̂k = jŝk (n)j2
N D. IAA-APES&BIC
n=1
end for
In many applications it is desirable to obtain point
until (convergence)
estimates rather than a continuous spatial estimate.
To achieve this sparsity we incorporate a model-order
TABLE III selection tool, i.e., the BIC [37, 38], into IAA-APES.
The IAA-APES&BIC Algorithm Let P denote a set containing the indices of the peaks
selected from the IAA-APES spatial power spectrum
P: Set of peaks obtained from IAA-APES
I = Ø, ´ = 1, quit = 0, BICold = 1
estimate. Also let I denote the set of the indices of
repeat the peaks selected by the BIC algorithm so far. The
i0 = arg min BICi (´) IAA-APES&BIC algorithm works as follows: first
i2P¡I
the peak, from the set P, giving the minimum BIC is
if BICi0 (´) < BICold
selected. Then the second peak, from the set P ¡ I,
I = fI, i0 g which together with the first peak gives the minimum
BICold = BICi0 (´) BIC, is selected, and so on, until the BIC value
´ = ´+1 does not decrease anymore.2 The IAA-APES&BIC
else quit = 1
algorithm is summarized in Table III. BICi (´) is
until (quit = 1)
calculated as follows (see [38])
0 ° °2 1
XN ° X °
This avoids the computation of Q¡1 (μk ) for each B ° ° C
° a(μj )ŝj (n)°
scanning point, i.e., K times. Moreover, fŝk (n)g can
BICi (´) = 2MN ln @ °y(n) ¡ ° A
n=1 ° j2fI[ig °
be computed in parallel for each scanning point, 2

which makes IAA-APES amenable to implementation + 3´ ln(2MN) (8)

on parallel hardware. IAA-APES is summarized in
Table II. Since IAA-APES requires R, which itself where ´ = jIj + 1, jIj denotes the size of the set I, i is
depends on the unknown signal powers, it has to the index of the current peak under consideration, and
be implemented as an iteration. The initialization is fŝj (n)gN
n=1 is the IAA-APES signal waveform estimate
done by a standard DAS beamformer. Our empirical corresponding to angle μj , j 2 fI [ ig. Note that the
experience is that IAA-APES does not provide second term on the right side of (8) does not matter
significant improvements in performance after to peak selection; it matters only when (8) is used to
about 15 iterations. In IAA-APES, P and, hence select the number of peaks to retain.
R, are obtained from the signal estimates of the
previous iteration and not from the snapshots as in E. IAA-APES&RELAX
conventional adaptive beamforming algorithms, such
RELAX [39, 40] is a parametric cyclic algorithm
as SCB, which fails to work properly with coherent
that requires an estimate of the number of sources in
or highly correlated sources, or few snapshots [43].
the field. The results of IAA-APES&BIC can be used
The Appendix provides an alternative derivation of
to provide a good initial estimate for the last step of
IAA-APES, based on the ML principle. IAA-APES
RELAX. This approach is outlined in Table IV. The
is shown to be an approximation of the IAA-ML
RELAX iterations can be terminated when the norm
algorithm, which is locally convergent due to
of the difference between two consecutive estimates
cyclically maximizing the likelihood function. Hence,
is smaller than a certain threshold (5 £ 10¡4 for the
the analysis in the Appendix provides an approximate
examples considered herein). The maximization
calculation that shows the local convergence of
IAA-APES. 2 Alternatively, the largest ´ peaks can be selected so that when
Because P is assumed to be a diagonal matrix, the (´ + 1)st largest peak is added to I, the BIC value does not
one degree of freedom (DOF) is needed to suppress decrease. This simplified version gives similar results to the one
one interfering source. This may lead to a larger described above in our numerical examples.

428 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
 TABLE IV
The IAA-APES&RELAX Algorithm

μ0 : Locations of the peaks obtained from IAA-APES&BIC

K 0 : Number of peaks obtained from IAA-APES&BIC
fŝk (n)g: Corresponding waveforms obtained from
IAA-APES&BIC
repeat
for k = 1, : : : , K 0
0
X
K
yk (n) = y(n) ¡ a(μi0 )ŝi (n), n = 1, : : : , N
i=1
i6=k

X
N
μ̂k0 = arg maxμ0 2R jaH (μ 0 )yk (n)j2
n=1
1 H 0
ŝk (n) = a (μ̂k )yk (n), n = 1, : : : , N
M
end for Fig. 2. Pulse compression for radar/sonar range-Doppler imaging.
until (convergence)

the Doppler interval of interest is divided into L bins,

step of the algorithm can be implemented without and the Doppler frequency for the lth Doppler bin is
much computational effort by means of a fine search denoted as !l . Then the M samples of the received
only around the peak values of the IAA-APES&BIC signal that is temporally aligned with the return from
result.3 RELAX can be useful in estimating off-grid the range bin of current interest r (see Fig. 2) can be
sources accurately and for further improving the represented by
IAA-APES waveform and angle estimates. L M¡1 L
X X X
yr = ®r,l s(!l ) + ®r+m,l Jm s(!l ) + er
III. ACTIVE SENSING l=1 m=¡M+1
m6=0
l=1

In active sensing applications, besides the (12)

receiver, there are also one or more transmitters. for r = 1, : : : , R, where ®r,l (which is proportional to
The radar/sonar range-Doppler imaging problem for the complex “voltage” radar-cross section (RCS)
a SISO system is first investigated in this section. of the corresponding target) denotes the complex
Then the channel estimation problem for MISO amplitude of the returned signal from the range bin
communications is discussed. of current interest r and the lth Doppler bin f®m+r,l g
denote the complex amplitudes of the returned signals
A. Range-Doppler Imaging from the adjacent range bins, and er denotes the noise.
The reflections from nearby range bins are considered
Pulse compression refers to the process of
to be clutter. The M £ M shift matrix Jm takes into
transmitting a modulated pulse and then matched
account the fact that the clutter returns from adjacent
filtering the returned signal, which arrives at the
range bins need different propagation times to reach
antenna altered by complex coefficients that bear
the radar/sonar receiver:
target information [48]. 2 3
1) Data Model: Consider a range-Doppler á m ¡!
imaging radar/sonar with a transmitted pulse 6 7
6 0 ¢¢¢ 1 ¢¢¢ 07
6 7
s̃ = [s̃(1), s̃(2), : : : , s̃(M)]T (9) 6 .. .. .. .7
6 . . . .. 7
6 7
where M is the pulse length. Let Jm = 6 .. 7 = JT¡m (13)
6 0 . 17
6 7
s(!l ) = s̃ ¯ d(!l ) (10) 6 . .. .. .7
6 . . . .. 7
4 . 5
be the Doppler shifted signal, where
0 ¢¢¢ 0 ¢¢¢ 0
d(!l ) = [1, ej!l , : : : , ej(M¡1)!l ]T , l = 1, : : : , L
for m = 0, : : : , M ¡ 1. Equation (12) can be written as
(11)
yr = S®r + ẽr (14)
3 The fine search can be implemented efficiently using the fast where
Fourier transform (FFT) [39, 40] or a derivative-free uphill search
method, such as the Nelder-Mead algorithm [46, 47]. Note that the S = [s(!1 ), s(!2 ), : : : , s(!L )] (15)
latter method is available in the MATLAB optimization toolbox
with the name of “fminsearch.” ®r = [®r,1 , ®r,2 , : : : , ®r,L ]T (16)

YARDIBI ET AL.: SOURCE LOCALIZATION AND SENSING: A NONPARAMETRIC ITERATIVE ADAPTIVE APPROACH 429

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
and can be computed off-line, whereas APC updates
M¡1
X L
X the filter coefficients iteratively and adaptively. In
ẽr = ®r+m,l Jm s(!l ) + er : (17) the nonnegligible Doppler case, however, adaptive
m=¡M+1 l=1
m6=0 approaches can perform much better. DC-APC is
Note that (12) is similar to the data model for passive purported to work with at most one target per range
sensing arrays, see (1), but with a single snapshot. The bin, and the method used in [55] to estimate the
IAA-APES estimate at iteration number i 2 f1, 2, : : :g Doppler values is not appealing from a performance
becomes viewpoint. On the other hand, IAA-APES can work
sH (!l )R¡1 with multiple Doppler targets located at the same
(i) (i¡1) (r)yr
®ˆ r,l = ¡1
(18) range bin, and IAA-APES estimates the Doppler
H
s (!l )R(i¡1) (r)s(!l )
values in a robust manner. Moreover, APC and
for l = 1, : : : , L, r = 1, : : : , R, where DC-APC require the tuning of hyperparameters,
whereas IAA-APES is hyperparameter free. Some
M¡1
X L
X (i¡1) 2 other differences are discussed below.
R(i¡1) (r) = j®ˆ r+m,l j Jm s(!l )sH (!l )JTm :
DC-APC assumes that there is at most one
m=¡M+1 l=1
target per range bin. Let f®r¡M+1 , : : : , ®r+M¡1 g and
(19) f!˜ r¡M+1 , : : : , !˜ r+M¡1 g denote the complex amplitudes
Here IAA-APES is applied in a slightly different and the Doppler frequencies, respectively, of the
manner than in the passive array processing case. targets in the corresponding range bin r. DC-APC
When ®r is to be estimated, the previous values of estimates the target parameters iteratively as follows
f®r+m gm=M¡1
m=¡M+1 are used to estimate R(r). However,
only ®r is updated using R(r) and yr . When all ®r(i) = j®r(i¡1) j2 sH (!˜ r )R¡1
(i¡1) (r)yr (20)
f®r gRr=1 are updated in this way, we advance to the
for r = 1, : : : , R, where
next IAA-APES iteration. IAA-APES is initialized by
matched filtering, matched to each range and Doppler M¡1
X
(i¡1) 2
bin, to obtain the initial f®r,l g. R(i¡1) (r) = j®r+m j Jm s(!˜ r+m )sH (!˜ r+m )JTm + Qr :
2) Related Work: Matched filtering has been m=¡M+1
widely used for pulse compression because it gives (21)
optimal signal-power-to-output-power performance
in the presence of a single target and white noise. Qr is the true noise covariance matrix of er , which
However, in practical radar/sonar systems, matched is assumed to be known, and i 2 f1, 2, : : :g represents
filter performance is far from desirable since the pulse the current iteration number. The initial estimates
compression problem is usually clutter limited rather of f®r+m gM¡1
m=¡M+1 are obtained by using a standard
than noise limited (see, e.g., [49] and [50]). Many data matched filter that neglects the Doppler effect [55].
independent (see, e.g., [49]—[54]) and data-adaptive DC-APC also requires the estimates f!˜ r+m gM¡1 m=¡M+1 of
(see, e.g., [48] and [55]) approaches have been the target Doppler frequencies at each iteration. The
proposed to achieve improved pulse compression. Doppler frequency estimation approaches suggested
Data independent approaches can be designed in [55] are ad-hoc and not very accurate, especially
off-line, and hence, these approaches are convenient for large Doppler shifts. This limitation of DC-APC,
for real-time implementations in practical systems. though, can be easily corrected by replacing !˜ r in
Although receive filters based on data-independent (20) with all possible Doppler frequencies f!l g to
instrumental variables (IV) can be used to achieve deal with multiple targets per range bin and to form
excellent pulse compression for the negligible range-Doppler images.
Doppler case, their performance is unsatisfactory The DC-APC iterations defined in (20)—(21) result
in the nonnegligible Doppler case because of high in the numerical ill-conditioning of R(i¡1) (r). To
sidelobe level problems [49]. On the other hand mitigate this problem, [48] suggests using j®r(i¡1) j±
data-adaptive approaches result in better performance in (20) instead of j®r(i¡1) j2 , and ¾± instead of ¾ 2 ,
but at the cost of implementation complexity. where 0 · ± · 2 and where the noise er is assumed
Two important data-adaptive methods for pulse to be white with a known variance ¾ 2 . This approach
compression are the adaptive pulse compression requires the delicate tuning of ± (at each iteration),
(APC) [48] and the Doppler-compensated APC but [48] and [55] do not provide a clear guideline on
(DC-APC) [55] algorithms. APC is an iterative how to do this. We remark that if the noise covariance
minimum mean-squared error (MMSE) based data matrix Qr in (21) is set to zero, then APC/DC-APC
adaptive approach, and DC-APC is the extension of becomes identical to FOCUSS, with the sparsity
APC to the nonnegligible Doppler case. However, parameter p = 0 and the regularization parameter
an IV approach [49] may be preferred to APC since ¸ = 0. (See, e.g., (16) in [22]). In addition, with the
the sidelobe level of the former method can be made introduction of ±, APC/DC-APC is still identical to
arbitrarily low and because the filter coefficients FOCUSS, now with 1 ¡ p=2 = ±=2 and ¸ = ¾± . We

430 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
also note that both APC/DC-APC and FOCUSS are Consequently, IAA-APES can be applied directly to
related to the approaches in [21], [30], and [56]. It is (23) to estimate f®r,l (i)g. Like in radar/sonar range
interesting to note that many different ideas lead to the compression problems, to determine f®r,l (i)g with
same result. high accuracy, the transmitted pulses need to have
IAA-APES assumes that multiple targets can exist both good auto and cross-correlation properties
within the same range bin and that f®r,l g is calculated [59, 60].
as in (18), where the term in the denominator can be 2) Related Work: Most of the demodulation
viewed as the current estimate of j®r(i¡1) j2 used in (20). algorithms used in practice rely heavily on the
This would be best estimate for j®r(i¡1) j2 , obtained accurate estimation of the channel impulse response.
by applying the minimum variance distortionless As in many other applications, sparse signal
criterion, had the true covariance matrix been known estimation approaches have also been proposed in
[5]. One more advantage of IAA-APES is that the this context. The main motivation for this is that
parameter estimates are inherently unbiased for the underwater communications [61—63] and wireless
signal of interest, whereas this is not the case in channels are appropriately modelled as sparse
DC-APC (or FOCUSS). Furthermore, in IAA-APES, channels consisting of only a few non-zero taps
the statistical properties of er are assumed unknown [64]. The existing approaches we evaluate for sparse
and are taken into account implicitly as false targets in channel estimation, besides IAA-APES, include the
the current and adjacent range and Doppler bins; see matching pursuit (MP), orthogonal matching pursuit
(19). No matrix inversion problems were encountered (OMP) [65—67], and least squares matching pursuit
with IAA-APES during our numerical simulations (LSMP) [68] algorithms, which have been used for
because of the unbiasedness property of IAA-APES. sparse channel estimation and equalization in many
applications [69—72]. It is difficult to determine
B. Channel Estimation the stopping criterion when using matching pursuit
The purpose of channel estimation in algorithms, and user intervention is needed.
communications is to provide the subsequent
equalizer with an accurate channel estimate so that the IV. NUMERICAL EXAMPLES
transmitted signals can be recovered successfully at
the receiver side [57, 58]. SISO channel estimation We evaluate the performance of IAA-APES
for communications problem is similar to SISO and compare it with various alternative methods
radar/sonar range-Doppler imaging, with the only in this section. We first focus on passive sensing
difference being that the delay for the former is due to applications and then shift our attention to active
single-trip propagation, while the delay for the latter sensing applications.
results from round-trip propagation.
1) Data Model: For the MISO channel estimation A. Passive Sensing Examples
problem with I transmitters and a single receiver, let
This subsection investigates the performance
s̃i = [s̃i (1), s̃i (2), : : : , s̃i (M)]T , i = 1, : : : , I (22) of IAA-APES, M-FOCUSS, M-SBL and `1 -SVD
for various passive sensing scenarios. Unless noted
denote the ith transmitted pulse, and let L denote the
otherwise, M-FOCUSS is implemented by setting
total number of Doppler bins. Then similar to (12),
the sparsity parameter p = 0:8 and fine tuning the
the received signal that is temporally aligned with the
regularization parameter ¸ to get the best results.4
return from the rth tap, r = 1, : : : , R, with R denoting
`1 -SVD is implemented by assuming that the number
the total number of taps of the channel, can be written
of sources is known a priori and by fine tuning
as ( L
I the hyperparameter. Since the fine tuning of the
X X
yr = ®r,l (i)si (!l ) hyperparameters assume knowledge of the true source
i=1 l=1
parameters, the so obtained results of M-FOCUSS and
9 `1 -SVD are impractical. M-SBL is implemented by
M¡1
X L
X >
= using the alternative update method for the parameters
+ ®r+m,l (i)Jm si (!l ) + er (23) as described in [19], [20], and [23]. For all approaches
>
;
m=¡M+1
m6=0
l=1 considered, the scanning grid is uniform in the range
from 1± to 180± , with 1± increment between adjacent
where Jm was defined in (13), grid points, unless noted otherwise.
si (!l ) = s̃i ¯ d(!l ) (24) We consider a uniform linear array with M = 12
sensors and half-wavelength interelement spacing.
and d(!l ) was defined in (11). When determining
the parameters of the ith channel f®r,l (i)g, the term 4 See Section VE in [22]. Also note that the M-FOCUSS algorithm
PL
l=1 ®r,l (i)si (!l ) is the signal term, and all other used here is referred to as regularized M-FOCUSS in [22] since
terms in (23) are considered as clutter and noise. ¸ 6= 0. We refer to the approach as M-FOCUSS for simplicity.

YARDIBI ET AL.: SOURCE LOCALIZATION AND SENSING: A NONPARAMETRIC ITERATIVE ADAPTIVE APPROACH 431

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
The far-field narrowband signal waveforms and the as the estimated signals for FOCUSS and `1 -SVD.
additive noise signals are assumed to be circularly We run the sparse algorithms, with a fine grid of size
symmetric, independent identically distributed (i.i.d.) 0:01± , around the true source locations. The sparse
complex Gaussian random processes with zero mean algorithms are run with K = 950 scanning points,
and variance ¾2 , which is varied to obtain various whereas IAA-APES is run with the same resolution
SNR values. Furthermore, each Psignal waveform as before, i.e., K = 180. IAA-APES&RELAX
is normalized such that (1=N) N 2
n=1 jsk (n)j = Pk , is used to estimate the off-grid sources. M-SBL
k = 1, : : : , K0 , for a given Pk value, where K0 denotes results are not shown because of the excessive
the true number of sources. SNR is defined as computation time needed. Each point in Fig. 5 is the
10 log10 (Pk =¾ 2 ), k = 1, : : : , K0 , in decibels (dB), where average of 100 Monte-Carlo trials. We observe that
¾2 is the noise variance. According to the comments IAA-APES&RELAX has both better variance and bias
made in the Introduction, we consider very low characteristics than the other methods for lower SNR.
snapshot cases, viz. N = 3 and N = 1. Note that FOCUSS and `1 -SVD both have two
First we consider three uncorrelated sources at 60± , hyperparameters, and their selection affects the
82 , and 90± , with 5 dB, 10 dB, and 10 dB powers,
±
performances of these two algorithms significantly.
respectively, and with N = 3. The noise power is Moreover, a different parameter setting should be used
0 dB, which results in a minimum SNR of 5 dB. depending on the SNR, N, and the source structure,
Fig. 3 shows the power and location estimates of i.e., the number of sources, source spacing, source
the algorithms. The circles and the vertical dotted power levels, and correlation levels. We were able to
lines that align with these circles represent the true tune the parameters relatively easily in our simulation
source locations and powers, and the results of 10 scenarios, but when the problem dimensions are large
Monte-Carlo trials are shown in each plot. DAS and when there is no prior knowledge of the scenario,
clearly suffers from smearing and leakage. IAA-APES it becomes difficult to find good hyperparameters.
provides a much better result than DAS, with low M-SBL, on the other hand, does not require any
sidelobes and peaks at the true source locations. hyperparameters, but it takes the longest time to
Moreover, IAA-APES&RELAX indicates the number converge. These are the main reasons why BP or
of sources and their locations and powers accurately. LASSO, which are the single-snapshot counterpart of
As observed in Fig. 3(d)—(f), the sparse algorithms `1 -SVD, and SBL and FOCUSS, which are the single
encounter source splitting, location bias, and power snapshot counterparts of M-FOCUSS and M-SBL, are
underestimation problems. Also it is hard to tell how not considered in the active sensing examples below.
many sources are present by solely using the spatial Assuming K À M, the complexity of each
estimates of the sparse algorithms. IAA-APES iteration is O(M 2 K). The complexity
Next Fig. 4 considers three coherent sources of the BIC extension is negligible compared to
at 60± , 80± , and 90± , with 10 dB power each and that of IAA-APES, and the complexity of the
with N = 3. The source waveforms are assumed RELAX extension depends on how many sources
to be identical for all three coherent sources. The IAA-APES&BIC determines, the termination
noise power is 0 dB, resulting in a 10 dB SNR. condition, and the method chosen for the
Similar to Fig. 3, the circles and the vertical dotted maximization step. In our experiments RELAX
lines represent the true source locations and the usually converged faster than the time needed
powers, and the results of 10 Monte-Carlo trials are for IAA-APES to converge. The complexities of
shown in each plot. We observe that IAA-APES M-FOCUSS and M-SBL are also O(M 2 K) per
is able to resolve the sources successfully and iteration [23]. The complexity of `1 -SVD, on the
that IAA-APES&RELAX provides accurate point other hand, is O(K 3 Ksvd
3
) [25], where Ksvd is the
estimates. On the other hand DAS fails to resolve the estimate of the number of sources. In our simulations
two closely spaced sources. The performances of the IAA-APES was always faster than M-FOCUSS,
sparse algorithms are similar to those of the previous M-SBL, and `1 -SVD, especially when the SNR was
example. low. Note that the actual number of iterations required
Finally Fig. 5(a) compares the total mean-squared for convergence depends heavily on the specific
error (MSE) (sum of each individual MSE) of scenario. For example for the scenario considered in
the angle estimates of each algorithm with the Fig. 5 at SNR = 2 dB, the times required by FOCUSS
Cramer-Rao bound (CRB) [5], and Fig. 5(b) compares and `1 -SVD are 3.9 and 5.6 s, respectively, whereas
the total angle estimation bias (sum of each individual IAA-APES&RELAX requires 0.1 s on average.5
bias’ modulus) of each algorithm for varying SNR.
Two uncorrelated sources are placed at 77:51± and 5 The timing values in seconds are given as an example. The
90:51± , and N is set to 1. (The angle values are convergence times may vary depending on many factors such as
picked so that they are not on the size 1± grid used how the algorithms are implemented, the specific hardware, etc.
by IAA-APES.) To calculate the MSE and bias, we `1 -SVD has been implemented using the software described in [73]
consider only the signals with the two largest powers and [44].

432 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
 Fig. 3. Three uncorrelated sources at 60± (5 dB power), 82± (10 dB power), and 90± (10 dB power), as represented by circles and
vertical dotted lines in each plot. N = 3, the noise power is 0 dB (which results in minimum SNR of 5 dB), and 10 Monte-Carlo
trials are shown. (a) DAS spatial estimate. (b) IAA-APES spatial estimate. (c) IAA-APES&RELAX point source estimates.
(d) M-FOCUSS spatial estimate. (e) M-SBL spatial estimate. (f) `1 -SVD spatial estimate.

B. Active Sensing Examples ©l = !l M(180± =¼), l = 1, : : : , L, which is the total

phase shift (in degrees) for the transmitted pulse
This section considers examples for the SISO
duration.
radar/sonar and the MISO channel estimation We consider three moving targets with 5 dB
problems. power and with Doppler shifts of ¡30± , ¡25± , and
1) SISO Range-Doppler Imaging Example: We 15± and six moving targets with 25 dB power and
now evaluate the performance of IAA-APES for Doppler shifts of ¡70± , ¡55± , ¡10± , 10± , 20± , and
SISO range-Doppler imaging and compare it with 60± , as shown in Fig. 6(a). The background noise is
that of the matched filtering, MP, OMP, and LSMP. assumed to be a circularly symmetric i.i.d. complex
We use a 30-element P3 code for the transmitted Gaussian random process, with mean zero and a
2
pulse, i.e., s̃(m) = ej(m¡1) ¼=M , and m = 1, : : : , M, where variance of 0 dB. The number of range bins is set
M = 30 [55, 74]. The Doppler shift is expressed as to 100, and the number of Doppler bins is set to

YARDIBI ET AL.: SOURCE LOCALIZATION AND SENSING: A NONPARAMETRIC ITERATIVE ADAPTIVE APPROACH 433

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
Fig. 4. Three coherent sources at 60± , 80± , and 90± , each with 10 dB power, as represented by circles and vertical dotted lines in each
plot. N = 3, noise power is 0 dB (SNR = 10 dB), and 10 Monte-Carlo trials are shown. (a) DAS spatial estimate. (b) IAA-APES
spatial estimate. (c) IAA-APES&RELAX point source estimates. (d) M-FOCUSS spatial estimate. (e) M-SBL spatial estimate.
(f) `1 -SVD spatial estimate.

37. The resulting minimum SNR is 5 dB. Fig. 6 target locations and Doppler frequencies being easily
shows that the matched filter smears the targets in observable) than the matched filter. IAA-APES&BIC
both the Doppler and range domains significantly. can estimate the strong targets accurately and the
Figs. 6(c)—(e) show that MP, OMP, and LSMP results weak targets reasonably well, as shown in Fig. 6(g).
are not satisfactory in this example. The MP, OMP, IAA-APES&RELAX helps improve the results of
and LSMP algorithms are all terminated manually IAA-APES&BIC even further, as shown in Fig. 6(h).
to give the best performance.6 Fig. 6(f) shows that In Figs. 6(c)—(e) and Figs. 6(g)—(h), the cross marks
IAA-APES provides a much more useful result (the represent the targets selected by the algorithms, and
the circles represent the ground truth. The numbers
6 See [70] for a complexity analysis of MP, OMP, and LSMP. In our shown in Figs. 6(g)—(h) are the power estimates
examples MP and OMP took less time than IAA-APES, whereas obtained by the corresponding algorithms. Recall
LSMP took longer than IAA-APES to converge. that BIC selects only the dominant components of the

434 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
 Fig. 5. Two uncorrelated sources at 77:51± and 90:51± with N = 1. (a) Total angle estimation MSE in dB together with CRB.
(b) Total angle estimation bias versus SNR. Each point is average of 100 Monte-Carlo trials. Recall that FOCUSS is single snapshot
version of M-FOCUSS.

Fig. 6. SISO range-Doppler imaging with three 5 dB and six 25 dB targets, indicated by circles. Noise power is 0 dB, resulting in
minimum SNR of 5 dB. (a) Ground truth with power levels. (b) Matched filter. (c) MP and ground truth. (d) OMP and ground truth.
(e) LSMP and ground truth. (f) IAA-APES. (g) IAA-APES&BIC and ground truth. (h) IAA-APES&RELAX and ground truth.
All power levels are in dB.

YARDIBI ET AL.: SOURCE LOCALIZATION AND SENSING: A NONPARAMETRIC ITERATIVE ADAPTIVE APPROACH 435

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
Fig. 7. MISO channel estimation with six transmitters and one receiver. Channel taps are indicated by circles. Channel tap power levels
are in range of ¡30 dB to 0 dB, and noise level is ¡20 dB, which results in ¡10 dB minimum SNR. Estimates for first channel
via (a) MP, (b) OMP, (c) LSMP, and (d) IAA-APES&BIC. All power levels are in dB.

IAA-APES spatial estimate, but BIC does not modify with mean zero and a ¡20 dB variance. The locations
either the range-Doppler values or the power levels of the non-zero taps of the channels are also simulated
estimated by IAA-APES, whereas RELAX improves randomly and by trying to mimic practical channel
upon the IAA-APES estimates. functions encountered in underwater communications,
2) Channel Estimation: The performance of MP, see, e.g., [69] and [70]. The power levels of the
OMP, LSMP, and IAA-APES are also investigated channel taps were in the range from ¡30 dB to 0 dB,
for the channel estimation problem encountered which results in a minimum SNR of ¡10 dB. Fig. 7
in communications. We consider six transmitters, illustrates the performances of the algorithms for
viz. I = 6, and a single receiver. The probing pulses the first channel; IAA-APES&BIC provides the best
used in the simulations are obtained by using a result. The results for the other channels, which are
cyclic algorithm to ensure both good auto and not shown for conciseness, are similar to that for the
cross-correlation properties; see [59] and [60] for first channel.
details. The sampling frequency is assumed to be
24 kHz. M is chosen to be 1024 (which yields a V. CONCLUSIONS
42.7 ms pulse duration). The number of delay taps
are set at 80 to yield a 3.3 ms of delay span, and This paper has presented IAA-APES in
13 Doppler bins are used to cover the range from array processing applications. IAA-APES is a
¡30 Hz to 30 Hz with 5± resolution. The six channels nonparametric, hyperparameter free algorithm that is
are assumed to be independent of each other, and they designed to work under severe snapshot limitations
are simulated as circularly symmetric i.i.d. complex and for uncorrelated, partially correlated, and
Gaussian random variables, with mean zero and a coherent sources, as well as for arrays with arbitrary
variance decreasing exponentially with increasing geometries. Because of the similarities between
delay, specifically ¾k2 = e¡0:03(k¡1) , where ¾k2 is the many active sensing applications and passive array
variance of the channel tap with delay index k, processing, IAA-APES can be applied to these cases
k = 1, : : : , 80. The noise is simulated as a circularly as well, without any essential modifications. The
symmetric i.i.d. complex Gaussian random process, BIC can be used in conjunction with the IAA-APES

436 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
algorithm to yield sparse solutions, which are and hence f 00 (P̃k ) is
desirable in many applications. Furthermore, the
¡1 2
application of the parametric RELAX algorithm to the (aH
k Qk ak )
f 00 (P̃k ) = (28)
¡1
IAA-APES&BIC results provides further performance (1 + P̂k aH
k Qk ak )
2

improvement. Simulations showed that IAA-APES

and its variations outperform the most prominent which is strictly positive. This means that P̃k is the
methods in the literature in the corresponding unique minimizer of f(Pk ). In principle, P̃k may be
application areas. IAA-APES is believed to be a negative. However, since Pk represents power, it should
viable candidate for practical applications since it be nonnegative. The minimizer of f(Pk ) subject to the
does not require any hyperparameters, has a simple constraint Pk ¸ 0 is
formulation, provides superresolution, facilitates
parallel processing, and shows good performance. P̂k = max(0, P̃k ) (29)

since P̃k is the unique minimizer of f(Pk ) (when

APPENDIX. AN APPROXIMATE ML INTERPRETATION P̃k < 0, f(0) · f(Pk ) for 8Pk ¸ 0).
OF IAA-APES
By using the matrix inversion lemma in (27) to
We derive a locally convergent ML-based iterative replace Qk by R, we get
adaptive approach called IAA-ML. IAA-ML is similar μ ¡1 ¡1
¶
aH
k R (¡ ¡ R)R ak
to IAA-APES in nature, and we show that IAA-APES P̂k = max 0, Pk + : (30)
can be obtained as a simplified version of IAA-ML. (aH ¡1
k R ak )
2

The negative log-likelihood function of fy(n)gN

n=1 Computing P̂k requires knowledge of Pk and R (recall
can be represented as
that R = A(μ)PAH (μ), where P is a diagonal matrix
ln jRj + tr(R¡1 ¡ ) (25) with fPk g on its diagonal). Therefore the algorithm
PN must be implemented iteratively; the initialization is
where ¡ = (1=N) n=1 y(n)yH (n) is the so-called done with DAS. IAA-ML will be locally convergent
sample covariance matrix. It is assumed that the if R is recalculated after each Pk is updated because of
received signal is a multivariate, complex Gaussian the cyclical maximization of the likelihood function.
random vector with zero-mean and covariance matrix Note that (30) can be written as (assuming P̃k ¸ 0)
R and that the snapshots are i.i.d. [5]. Minimizing
¡1 ¡1
(25) is equivalent to minimizing the Stein’s loss, aH
k R ¡ R ak 1
P̂k = + Pk ¡ H ¡1 : (31)
¡ ln jR¡1 ¡ j + tr(R¡1 ¡ ) [75, 76], which stresses the H ¡1
(ak R ak ) 2 (ak R ak )
covariance fitting character of the cost function, but
¡1
without the i.i.d. Gaussian assumptions needed by the By the properties of SCB, Pk ¼ 1=(aH
k R ak ) [6, 43].
maximum likelihood loss function. Hence an approximate solution of (31) is
Assume that Q(μk ) is known and that the signal ¡1 ¡1
aH
k R ¡ R ak
power at μk is to be estimated. Using the fact that P̂k = , k = 1, : : : , K: (32)
jI + ABj = jI + BAj and the matrix inversion lemma (aH ¡1
k R ak )
2

together with (4), it can be shown that minimizing Iterating this equation by building R from the
(25), with respect to Pk , is equivalent to minimizing latest estimate of fPk g, we get IAA-APES. This
Pk aH ¡1 ¡1 approximation has two advantages. First the P̂k in (32)
¢ ¡1 k Qk ¡ Qk ak
f(Pk ) = ln(1 + Pk aH
k Qk ak ) ¡ ¡1 is guaranteed to be nonnegative, and it alleviates the
1 + Pk aH
k Qk ak need for the procedure in (29). Secondly when fPk g
(26) is accurate, the difference Pk ¡ 1=(aH ¡1
k R ak ) in (31) is
where Q(μk ) and a(μk ) have been replaced by Qk and small, as discussed above. When fPk g is inaccurate,
ak , respectively, for notational simplicity. Setting the however, this difference may not be small, and forcing
first derivative of (26), with respect to Pk , to zero, i.e., the difference to zero may provide a better estimate of
f 0 (P̃k ) = 0, gives Pk . (IAA-ML tends to work well when the snapshot
number is large. It is not considered in the numerical
¡1 ¡1
aH
k Qk (¡ ¡ Qk )Qk ak examples because of our focus on cases with few
P̃k = ¡1
: (27)
(aH
k Qk ak )
2 snapshots.) Since IAA-APES can be obtained as a
close approximation to IAA-ML, which is locally
The second derivative of (26), with respect to Pk , is convergent, IAA-APES is expected to also enjoy
aH ¡1 ¡1 local convergence. We have never come across an
¡1 2 k Qk ¡ Qk ak H ¡1
¡(aH
k Qk ak ) + 2 ¡1
ak Qk ak example where IAA-APES did not converge; however,
(1 + Pk aH
k Qk ak )
f 00 (Pk ) = ¡1
the search for a convergence proof could be a useful
(1 + Pk aH
k Qk ak )
2
direction for future work.

YARDIBI ET AL.: SOURCE LOCALIZATION AND SENSING: A NONPARAMETRIC ITERATIVE ADAPTIVE APPROACH 437

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
ACKNOWLEDGMENT [15] Fuchs, J. J.
Recovery of exact sparse representations in the presence
We thank Professor Aaron Lanterman for his of bounded noise.
IEEE Transactions on Information Theory, 51, 10 (2005),
useful comments on the revisions of a former version
3601—3608.
of this paper. [16] Tibshirani, R.
Regression shrinkage and selection via the lasso.
REFERENCES Journal of the Royal Statistical Society, 58, 1 (1996),
267—288.
[1] Li, J., and Stoica P. (Eds.)
[17] Chen, S. S., Donoho, D. L., and Saunders, M. A.
Robust Adaptive Beamforming.
Atomic decomposition by basis pursuit.
New York: Wiley, 2005.
SIAM Journal on Scientific Computing, 20, 1 (1998),
[2] Capon, J.
33—61.
High resolution frequency-wavenumber spectrum
[18] Gorodnitsky, I. F., and Rao, B. D.
analysis.
Sparse signal reconstruction from limited data using
Proceedings of the IEEE, 57 (Aug. 1969), 1408—1418.
FOCUSS: A re-weighted minimum norm algorithm.
[3] Schmidt, R. O.
IEEE Transactions on Signal Processing, 45, 3 (1997),
Multiple emitter location and signal parameter estimation.
600—616.
IEEE Transactions on Antennas and Propagation, AP-34, 3
[19] Tipping, M. E.
(Mar. 1986), 276—280.
Sparse Bayesian learning and the relevance vector
[4] Stoica, P., and Nehorai, A.
machine.
MUSIC, maximum likelihood, and Cramer-Rao bound.
Journal of Machine Learning Research, 1 (2001), 211—244.
IEEE Transactions on Acoustics, Speech, and Signal
[20] Wipf, D. P., and Rao, B. D.
Processing, 5 (May 1989), 720—741.
Sparse Bayesian learning for basis selection.
[5] Van Trees, H. L.
IEEE Transactions on Signal Processing, 52, 8 (2004),
Optimum Array Processing: Part IV of Detection,
2153—2164.
Estimation, and Modulation Theory.
[21] Figueiredo, M. A. T.
New York: Wiley, 2002.
Adaptive sparseness for supervised learning.
[6] Li, J., Stoica, P., and Wang, Z.
IEEE Transactions on Pattern Analysis and Machine
On robust Capon beamforming and diagonal loading.
Intelligence, 25, 9 (2003), 1150—1159.
IEEE Transactions on Signal Processing, 51, 7 (July
2003), 1702—1715. [22] Cotter, S. F., Rao, B. D., Kjersti, E., and Kreutz-Delgado, K.
[7] Swindlehurst, A. L., and Kailath, T. Sparse solutions to linear inverse problems with multiple
A performance analysis of subspace-based methods in the measurement vectors.
presence of model errors, Part I: The MUSIC algorithm. IEEE Transactions on Signal Processing, 53, 7 (2005),
IEEE Transactions on Signal Processing, 40, 7 (July 2477—2488.
1992), 1758—1773. [23] Wipf, D. P., and Rao, B. D.
[8] Stoica, P., Wang, Z., and Li, J. An empirical Bayesian strategy for solving the
Extended derivations of MUSIC in the presence of simultaneous sparse approximation problem.
steering vector errors. IEEE Transactions on Signal Processing, 55, 7 (2007),
IEEE Transactions on Signal Processing, 53, 3 (Mar. 3704—3716.
2005), 1209—1211. [24] Malioutov, D. M.
[9] Li, J., Stoica, P., and Wang, Z. A sparse signal reconstruction perspective for source
Doubly constrained robust Capon beamformer. localization with sensor arrays.
IEEE Transactions on Signal Processing, 52, 9 (Sept. M.S. thesis, Massachusetts Institute of Technology,
2004), 2407—2423. Cambridge, MA, July 2003.
[10] Baggeroer, A. B., and Cox, H. [25] Malioutov, D. M., Çetin, M., and Willsky, A. S.
Passive sonar limits upon nulling multiple moving ships A sparse signal reconstruction perspective for source
with large aperture arrays. localization with sensor arrays.
In Proceedings of the 33th Asilomar Conference on Signals, IEEE Transactions on Signal Processing, 53, 8 (2005),
Systems and Computers, vol. 1, 1999, 103—108. 3010—3022.
[11] Kraay, A. L., and Baggeroer, A. B. [26] Model, D., and Zibulevsky, M.
A physically constrained maximum-likelihood method for Signal reconstruction in sensor arrays using sparse
snapshot-deficient adaptive array processing. representations.
IEEE Transactions on Signal Processing, 55, 8 (Aug. Signal Processing, 86, 3 (2006), 624—638.
2007), 4048—4063. [27] Fuchs, J. J.
[12] Donoho, D. L., and Elad, M. Linear programming in spectral estimation. Application to
Optimally sparse representation in general array processing.
(nonorthogonal) dictionaries via l 1 minimization. In Proceedings of the IEEE Conference on Acoustics,
Proceedings of the National Academy of Sciences, 100, 5 Speech, and Signal Processing, 1996, 3161—3164.
(Mar. 2000), 2197—2202. [28] Fuchs, J. J.
[13] Tropp, J. A. On the application of the global matched filter to DOA
Just relax: Convex programming methods for identifying estimation with uniform circular arrays.
sparse signals. IEEE Transactions on Signal Processing, 49, 4 (2001),
IEEE Transactions on Information Theory, 51, 3 (Mar. 702—709.
2006), 1030—1051. [29] Yardibi, T., Li, J., Stoica, P., and Cattafesta, L. N.
[14] Fuchs, J. J. Sparsity constrained deconvolution approaches for
On sparse representations in arbitrary redundant bases. acoustic source mapping.
IEEE Transactions on Information Theory, 50, 6 (2004), Journal of the Acoustical Society of America, 123, 5 (May
1341—1344. 2008), 2631—2642.

438 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
[30] Sacchi, M. D., Ulrych, T. J., and Walker, C. J. [45] Brooks, T. F., and Humphreys, Jr., W. M.
Interpolation and extrapolation using a high-resolution A deconvolution approach for the mapping of acoustic
discrete Fourier transform. sources (DAMAS) determined from phased microphone
IEEE Transactions on Signal Processing, 46, 1 (Jan.1998), arrays.
31—38. Journal of Sound and Vibration, 294 (July 2006), 856—879.
[31] Jeffs, B. [46] Nelder, J. A., and Mead, R.
Sparse inverse solution methods for signal and image A simplex method for function minimization.
processing applications. Computer Journal, 7 (1965), 308—313.
In Proceedings of the IEEE International Conference on [47] Lagarias, J. C., Reeds, J. A., Wright, M. H., and Wright,
Acoustics, Speech and Signal Processing, 3 (May 1998), P. E.
1885—1888. Convergence properties of the Nelder-Mead simplex
[32] Ting, M., Raich, R., and Hero, A. method in low dimensions.
Sparse image reconstruction using sparse priors. SIAM Journal of Optimization, 9, 1 (1998), 112—147.
Presented at the IEEE International Conference on Image
[48] Blunt, S. D., and Gerlach, K.
Processing, Atlanta, GA, Oct. 2006.
Adaptive pulse compression via MMSE estimation.
[33] Raich, R., and Hero, A. IEEE Transactions on Aerospace and Electronic Systems,
Sparse image reconstruction for partially known blur 42, 2 (Apr. 2006), 572—584.
functions.
[49] Stoica, P., Li, J., and Xue, M.
Presented at the IEEE International Conference on Image
Transmit codes and receive filters for pulse compression
Processing, Atlanta, GA, Oct. 2006.
radar systems.
[34] Li, J., and Stoica, P. IEEE Signal Processing Magazine, 25, 6 (Nov. 2008).
An adaptive filtering approach to spectral estimation and
SAR imaging. [50] Stoica, P., Li, J., and Xue, M.
IEEE Transactions on Signal Processing, 44, 6 (June On sequences with good correlation properties: A new
1996), 1469—1484. perspective.
Presented at the 2007 IEEE Information Theory
[35] Stoica, P., Li, H., and Li, J.
Workshop on Information Theory for Wireless Networks,
A new derivation of the APES filter.
Bergen, Norway, July 2007.
IEEE Signal Processing Letters, 6, 8 (Aug. 1999),
205—206. [51] Stutt, C., and Spafford, L.
A “best” mismatched filter response for radar clutter
[36] Stoica, P., Jakobsson, A., and Li, J.
discrimination.
Capon, APES and matched-filterbank spectral estimation.
IEEE Transactions on Information Theory, 14, 2 (Mar.
Signal Processing, 66, 1 (Apr. 1998), 45—59.
1968), 280—287.
[37] Schwarz, G.
Estimating the dimension of a model. [52] Ackroyd, M. H., and Ghani, F.
The Annals of Statistics, 6, 2 (1978), 461—464. Optimum mismatched filters for sidelobe suppression.
IEEE Transactions on Aerospace and Electronic Systems, 9,
[38] Stoica, P., and Selén, Y.
2 (Mar. 1973), 214—218.
Model-order selection: A review of information criterion
rules. [53] Nunn, C.
IEEE Signal Processing Magazine, 21, 4 (July 2004), Constrained optimization applied to pulse compression
36—47. codes, and filters.
In Proceedings of the IEEE International Radar
[39] Li, J., and Stoica, P.
Conference, Arlington, VA, May 9—12, 2005, 190—194.
Efficient mixed-spectrum estimation with applications to
target feature extraction. [54] Levanon, N.
IEEE Transactions on Signal Processing, 44, 2 (Feb. Cross-correlation of long binary signals with longer
1996), 281—295. mismatched filters.
[40] Li, J., Stoica, P., and Zheng, D. IEE Proceedings–Radar, Sonar, and Navigation, 152, 6
Angle and waveform estimation via RELAX. (Dec. 2005), 377—382.
IEEE Transactions on Aerospace and Electronic Systems, [55] Blunt, S. D., Smith, K. J., and Gerlach, K.
33 (July 1997), 1077—1087. Doppler-compensated adaptive pulse compression.
[41] Humphreys, Jr., W. M., Brooks, T. F., Hunter, Jr., W. W., In Proceedings of the IEEE Conference on Radar, Verona,
and Meadows, K. R. NY, Apr. 24—27, 2006, 114—119.
Design and use of microphone directional arrays for [56] Fuchs, J. J.
aeroacoustic measurements. Convergence of a sparse representations algorithm
Presented at the 36th Aerospace Sciences Meeting and applicable to real or complex data.
Exhibit, Reno, NV, Jan. 1998; AIAA Paper 98-0471. IEEE Journal of Selected Topics in Signal Processing, 1, 4
[42] Baggeroer, A. B. (2007), 598—605.
Sonar arrays and array processing. [57] Preisig, J. C., and Deane, G. B.
In Proceedings of AIP Conference, vol. 1, Apr. 2005, Surface wave focusing and acoustic communications in
3—24. the surf zone.
[43] Stoica, P., and Moses, R. L. Journal of the Acoustical Society of America, 116, 4 (Oct.
Spectral Analysis of Signals. 2004), 2067—2080.
Upper Saddle River, NJ: Prentice-Hall, 2005. [58] Preisig, J. C.
[44] Labit, Y., Peaucelle, D., and Henrion, D. Performance analysis of adaptive equalization for
SEDUMI interface 1.02–A tool for solving LMI coherent acoustic communications in the time-varying
problems with SEDUMI. ocean environment.
In Proceedings of the IEEE International Symposium on Journal of the Acoustical Society of America, 118, 1 (July
Computer, Glasgow, UK, Sept. 2002, 272—277. 2005), 263—278.

YARDIBI ET AL.: SOURCE LOCALIZATION AND SENSING: A NONPARAMETRIC ITERATIVE ADAPTIVE APPROACH 439

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
[59] Li, J., Zheng, X., and Stoica, P. [68] Li, W.
MIMO SAR imaging: Signal synthesis and receiver Estimation and tracking of rapidly time-varying
design. broadband acoustic communication channels.
Presented at the 2nd International Workshop on Ph.D. dissertation, Massachusetts Institute of Technology,
Computational Advances in Multi-Sensor Adaptive Cambridge, MA, 2005.
Processing, St. Thomas, US Virgin Islands, 2007. [69] Li, W., and Preisig, J. C.
[60] Li, J., Stoica, P., and Zheng, X. Estimation and equalization of rapidly varying sparse
Signal synthesis and receiver design for MIMO radar acoustic communication channels.
imaging. In Proceedings of the IEEE/MTS Oceans Conference, Sept.
IEEE Transactions on Signal Processing, 56, 8 (Aug. 2006, 1—6.
2008), 3959—3968. [70] Li, W., and Preisig, J. C.
[61] Kocic, M., Brady, D., and Stojanovic, M. Estimation of rapidly time-varying sparse channels.
Sparse equalization for real-time digital underwater IEEE Journal of Oceanic Engineering, 32, 4 (2007),
acoustic communications. 927—939.
In Proceedings of the IEEE/MTS Oceans Conference, [71] Cotter, S. F., and Rao, B. D.
vol. 3, Oct. 1995, 1417—1422. The adaptive matching pursuit algorithm for estimation
[62] Stojanovic, M. and equalization of sparse time-varying channels.
Recent advances in high-speed underwater acoustic In Proceedings of the 25th Asilomar Conference on Signals,
communications. Systems and Computers, vol. 2, Pacific Grove, CA, Oct.
IEEE Journal of Oceanic Engineering, 21, 2 (Apr. 1996), 2000, 1772—1776.
125—136. [72] Cotter, S. F., and Rao, B. D.
[63] Stojanovic, M., Freitag, L., and Johnson, M. Sparse channel estimation via matching pursuit with
Channel-estimation-based adaptive equalization of application to equalization.
underwater acoustic signals. IEEE Transactions on Communications, 50, 3 (Mar. 2002),
In Proceedings of the IEEE/MTS Oceans Conference, 374—377.
vol. 2, Sept. 1999, 985—990. [73] Löfberg, J.
[64] Carbonelli, C., Vedantam, S., and Mitra, U. YALMIP: A toolbox for modeling and optimization in
Sparse channel estimation with zero tap detection. MATLAB.
IEEE Transactions on Wireless Communications, 6, 5 In Proceedings of the 2004 IEEE International Symposium
(2007), 1743—1753. on Computer Aided Control Systems Design, Sept. 2004,
[65] Mallat, S., and Zhang, Z. 284—289.
Matching pursuits with time-frequency dictionaries. [74] Proakis, J. G.
IEEE Transactions on Signal Processing, 41, 12 (1993), Digital Communications (4th ed.).
3397—3415. Columbus, OH: McGraw-Hill, 2001.
[66] Natarajan, B. K. [75] James, W., and Stein, C.
Sparse approximation solutions to linear systems. Estimation with quadratic loss.
SIAM Journal of Computing, 24 (1995), 227—234. In Proceedings of the Fourth Berkeley Symposium on
[67] Cotter, S. F., Adler, R., Rao, R. D., and Kreutz-Delgado, K. Mathematical Statistics Probability, vol. 1, 1961, 361—380.
Forward sequential algorithms for best basis selection. [76] Dey, D. K., and Srinivasan, C.
IEE Proceedings of Vision, Image and Signal Processing, Estimation of a covariance matrix under Stein’s loss.
146, 5 (Oct. 1999), 235—244. The Annals of Statistics, 13, 4 (Dec. 1985), 1581—1591.

440 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
 Tarik Yardibi (S’08) received his B.S. degree from Hacettepe University, Ankara,
Turkey in 2004 and his M.S. degree from Bilkent University, Ankara, Turkey in
2006, both in electrical engineering.
He is currently pursuing a Ph.D. degree at the Spectral Analysis Laboratory
at the University of Florida, Gainesville. His research interests include statistical
signal processing, sensor array processing, aeroacoustic measurements, and sparse
signal representations.

Jian Li (S’87–M’91–SM’97–F’05) received her M.Sc. and Ph.D. degrees in

electrical engineering from The Ohio State University, Columbus, in 1987 and
1991, respectively.
From April 1991 to June 1991, she was an adjunct assistant professor with
the Department of Electrical Engineering, The Ohio State University, Columbus.
From July 1991 to June 1993, she was an assistant professor with the Department
of Electrical Engineering, University of Kentucky, Lexington. Since August
1993, she has been with the Department of Electrical and Computer Engineering,
University of Florida, Gainesville, where she is currently a professor. In Fall
2007, she was on sabbatical leave at MIT, Cambridge, Massachusetts. Her
current research interests include spectral estimation, statistical and array signal
processing, and their applications.
Dr. Li is a Fellow of IET. She is a member of Sigma Xi and Phi Kappa Phi.
She received the 1994 National Science Foundation Young Investigator Award
and the 1996 Office of Naval Research Young Investigator Award. She was an
executive committee member of the 2002 International Conference on Acoustics,
Speech, and Signal Processing, Orlando, FL, May 2002. She was an Associate
Editor of the IEEE Transactions on Signal Processing from 1999 to 2005, an
Associate Editor of the IEEE Signal Processing Magazine from 2003 to 2005,
and a member of the Editorial Board of Signal Processing, a publication of the
European Association for Signal Processing (EURASIP), from 2005 to 2007. She
has been a member of the Editorial Board of Digital Signal Processing–A Review
Journal, a publication of Elsevier, since 2006. She is presently a member of the
Sensor Array and Multichannel (SAM) Technical Committee of IEEE Signal
Processing Society. She is a coauthor of papers that have received the First and
Second Place Best Student Paper Awards, respectively, at the 2005 and 2007
Annual Asilomar Conferences on Signals, Systems, and Computers in Pacific
Grove, CA. She is also a coauthor of the paper that has received the M. Barry
Carlton Award for the best paper published in IEEE Transactions on Aerospace
and Electronic Systems in 2005.

YARDIBI ET AL.: SOURCE LOCALIZATION AND SENSING: A NONPARAMETRIC ITERATIVE ADAPTIVE APPROACH 441

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
 Petre Stoica (SM’91–F’94) received the D.Sc. degree in automatic control from
the Polytechnic Institute of Bucharest (BPI), Bucharest, Romania, in 1979 and
an honorary doctorate degree in science from Uppsala University (UU), Uppsala,
Sweden, in 1993.
He is a Professor of Systems Modeling with the Division of Systems and
Control, the Department of Information Technology at UU. Previously, he was
a Professor of System Identification and Signal Processing with the Faculty
of Automatic Control and Computers at BPI. He held longer visiting positions
with Eindhoven University of Technology, Eindhoven, The Netherlands;
Chalmers University of Technology, Gothenburg, Sweden (where he held a
Jubilee Visiting Professorship); UU; University of Florida, Gainesville, FL; and
Stanford University, Stanford, CA. His main scientific interests are in the areas
of system identification, time series analysis and prediction, statistical signal and
array processing, spectral analysis, wireless communications, and radar signal
processing.
Dr. Stoica has published nine books, ten book chapters, and some 500 papers
in archival journals and conference records. The most recent book he coauthored,
with R. Moses, is Spectral Analysis of Signals (Prentice-Hall, 2005). He is on the
editorial boards of six journals: Journal of Forecasting, Signal Processing, Circuits,
Signals, and Signal Processing, Digital Signal Processing, CA Review Journal,
Signal Processing Magazine, and Multidimensional Systems and Signal Processing.
He was a coguest editor for several special issues on system identification, signal
processing, spectral analysis, and radar for some of the aforementioned journals,
as well as for IEE Proceedings. He was corecipient of the IEEE ASSP Senior
Award for a paper on statistical aspects of array signal processing. He was also
recipient of the Technical Achievement Award of the IEEE Signal Processing
Society. In 1998, he was the recipient of a Senior Individual Grant Award of
the Swedish Foundation for Strategic Research. He was also corecipient of the
1998 EURASIP Best Paper Award for Signal Processing for a work on parameter
estimation of exponential signals with time-varying amplitude, a 1999 IEEE
Signal Processing Society Best Paper Award for a paper on parameter and rank
estimation of reduced-rank regression, a 2000 IEEE Third Millennium Medal,
and the 2000 W. R. G. Baker Prize Paper Award for a paper on maximum
likelihood methods for radar. He has been a member of the international program
committees of many topical conferences. From 1981 to 1986, he was a Director
of the International Time-Series Analysis and Forecasting Society, and he was
also a member of the IFAC Technical Committee on Modeling, Identification,
and Signal Processing. He is also a member of the Royal Swedish Academy of
Engineering Sciences, an honorary member of the Romanian Academy, and a
fellow of the Royal Statistical Society.

Ming Xue (S’08) received his B.S. degree in electrical engineering from the
University of Science and Technology of China, Hefei, China in 2006.
He is working towards a Ph.D. degree in the Department of Electrical and
Computer Engineering, University of Florida, Gainesville. His research interests
include statistical signal processing and its applications.

442 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
 Arthur B. Baggeroer (S’62–M’68–SM’87–F’89) received the degrees of
B.S.E.E. from Purdue University in 1963 and Sc.D. from MIT in 1968. He is a
Ford Professor of Engineering in the Departments of Mechanical Engineering
and Electrical Engineering & Computer Science at the Massachusetts Institute
of Technology. He was been a consultant to the Chief of Naval Research at
the NATO SACLANT Center (now NURC) in 1977 and a Cecil and Ida Green
Scholar at the Scripps Institution of Oceanography in 1990 while on sabbatical
leaves.
Dr. Baggeroer is a Fellow of the Acoustical Society of America. He received
the IEEE Oceanic Engineering Society Distinguished Technical Achievement
Award in 1991, was an elected member of the Executive Council of the
Acoustical Society from 1994—1997, and was awarded the Rayleigh-Helmholtz
Medal from the Acoustical Society in 2003. He was elected to the National
Academy of Engineering in 1995 and awarded a Secretary of the Navy/Chief
of Naval Operations Chair in Oceanographic Science in 1998. He serves as a
senior advisor to the Navy on numerous committees and panels. He recently
chaired the NSB panel on Distributed Remote Surveillance (DRS). Professor
Baggeroer was awarded the “Distinguished Alumni Award” of the Department of
Electrical and Computer Engineering from his alma mater, Purdue University. He
was recently awarded the ADM Charles Martel—David Bushnell Award by NDIA
for “outstanding technical contributions to the defense of the US in the field of
Undersea Warfare.” He has been chief scientist on fifteen oceanographic cruises
with seven in the Arctic Ocean. His research has concerned signal and array
processing for sonar, radar and seismic systems, ocean acoustic telemetry, global
acoustics for ocean thermometry and matched field array processing. He also has
had long affiliations with the Woods Hole Oceanographic Institution where he
was Director of the MIT—Woods Hole Joint Program from 1983—1988 and the
MIT—Lincoln Laboratory. Finally, he was elected four times to be a member of
the School Committee for the Town of Westwood, MA (1978—1990) and was
elected to be chairman four times.
Some of the Navy Committees he has been involved in are: 1) the Naval
Studies Board, 2) the Ocean Studies Board (both for six years), 3) the Submarine
Superiority Technical Advisory Group (he was a member of the original
committee for ADM Demars which led to APB/ARCI.), 4) the Fixed Surveillance
Systems Technical Advisory Group, 5) the SSIPT for N84 (twice), 6) the “Red
Team” special programs component for the Way Ahead for ASW, 7) an advisory
panel member for several programs for the Navy and DARPA, and 8) the Naval
Research Advisory Committee.

YARDIBI ET AL.: SOURCE LOCALIZATION AND SENSING: A NONPARAMETRIC ITERATIVE ADAPTIVE APPROACH 443

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.