IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO.

9, SEPTEMBER 2016

6449

High-Resolution Direct Position Determination

Using MVDR
Liran Tzafri, Student Member, IEEE, and Anthony J. Weiss, Fellow, IEEE

Abstract The direct position determination (DPD) approach

is a single-step method, which uses the maximum likelihood
estimator to localize sources emitting electromagnetic energy
using combined data from all available sensors. The DPD is
known to outperform the traditional two-step methods under
low signal-to-noise ratio conditions. We propose an improvement
to the DPD approach, using the well-known minimum-variancedistortionless-response (MVDR) approach. Unlike maximum likelihood, the number of sources needs not be known before applying
the method. The combination of both the direct approach and
MVDR yields unprecedented localization accuracy and resolution
for weak sources. We demonstrate this approach on the problem
of multistatic radar, but the method can easily be extended to
general localization problems.
Index Terms Positioning algorithms, adaptive array processing, minimum variance distortionless response (MVDR),
maximum likelihood estimation, direct position determination.

(.)T
(.) H
a
diag (A)

N OMENCLATURE
Transpose.
Conjugate transpose.
Euclidean norm of a vector.
Main diagonal of a matrix.
Kronecker product.
I. I NTRODUCTION

HE localization of sources emitting electromagnetic or

acoustic energy is needed in wild-life tracking, radioastronomy, seismology, medical-diagnosis, communications,
and other engineering applications. Common localization
methods use two estimation steps. First, intermediate parameters are estimated. Intermediate parameters are usually time of
arrival, direction of arrival, Doppler frequency shift or signal
strength. These estimated parameters are then used, in a second
step, to estimate the actual location of the emitter. The Direct
Position Determination (DPD) approach has been recently
proposed [1], [2] as a single-step Maximum Likelihood localization technique. A single-step approach is a technique in
which the estimator uses exactly the same data as used in
two-step methods but estimates the source location directly,
Manuscript received November 11, 2015; revised February 20, 2016 and
June 5, 2016; accepted June 22, 2016. Date of publication June 27, 2016;
date of current version September 8, 2016. The associate editor coordinating
the review of this paper and approving it for publication was R. M. Buehrer.
The authors are with the School of Electrical Engineering, Tel Aviv University, Tel Aviv 6997801, Israel (e-mail: ltzafri@gmail.com; ajw@eng.tau.ac.il).
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TWC.2016.2585116

skipping the intermediate (first) step. This can be viewed as

searching for the emitter location that best explains the collected data. From estimation theory point-of-view the two-step
approach is inferior, since in the first step the parameters
are measured independently, ignoring the constraint that the
measurements relate to the same emitter location. Indeed,
the DPD method has been shown to be superior to the two-step
methods for low SNR. In addition, this method has been
shown to be more robust by inherently selecting reliable
observations without the need for a goodness-of-fit test (such
as the chi-square test). This method was also extended to
radar scenarios in [3], where the Maximum Likelihood target
location estimation was developed as well as the Cramer Rao
lower bound for the estimation error.
When there are multiple sources the DPD is no longer
equivalent to the Maximum Likelihood Estimator (MLE). The
exact MLE can be derived but it requires a multi-dimensional
search which is usually impractical. An alternative for the
Maximum Likelihood parameter estimator is the Minimum
Variance Distortionless Response (MVDR) estimator. It was
originally proposed by Capon [4] for frequency-wavenumber
power spectral density analysis, but has since been used extensively as a high resolution method. The idea is to maximize
a score function given by multiplying the inputs by a set
of weight vectors, where the weight vectors are adaptively
selected in order to fix the response for the parameter value
of interest while minimizing the output power. Unlike the
Maximum Likelihood approach, the MVDR approach does not
need to know a-priori the number of targets (or model order)
and therefore it is a robust approach with good resolution
and immunity to jamming and interference. The MVDR was
proposed previously for Synthetic Aperture Radar (SAR) high
resolution imaging. DeGraff [5] describes many approaches
for improving SAR images including MVDR which he terms
MVM for Minimum Variance Method. According to [5] MVM
is one of the most successful approaches.
Our test case is multistatic radar, thus a brief introduction
is presented here. The idea of bistatic radar has been studied
from the earliest days of radars. By spatially separating the
transmitter from the receiver, these radar systems have been
shown to exhibit some advantages over common monostatic
radar systems. A multistatic radar system is an extension of
the bistatic radar concept, where multiple cooperative receivers
are used for target localization. This could be generalized
further with the addition of multiple transmitters (a scheme
usually termed MIMO radar, e.g. [7][9]), but we limit our

1536-1276 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

6450

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO. 9, SEPTEMBER 2016

demonstration to a single transmitter, thus ignoring difficulties

caused by mutual interference of the transmitted signals.
The focus of this paper is the demonstration of the singlestep (direct) MVDR concept for source localization. As an
example, we use a simplified multi-static radar model, neglecting radar clutter for brevity. Future work includes the full
blown radar model. Note that MVDR is applied to the targets location estimation in a single step, and not for intermediate parameters estimation. Some previous publications
touch upon our proposed approach but in no way cover the
full potential of this method (See for example [10], where
only the signals direction of arrivals are taken into account,
and [11][13] where the receivers are collocated). Some papers
like [14] mention location in their titles but their text discuss
application of MVDR to Angle of Arrival. We show that the
estimation method proposed here can significantly improve
target resolution compared with the single-source MLE. Fine
target resolution can prove very useful for target localization
within many dense decoys.
Our novel contributions are summarized as follows. 1) We
provide a 1-step MVDR based localization algorithm and
demonstrate its unprecedented resolution capabilities. MVDR
is applied to localization and not, as done in previous works,
to intermediate parameters, where actual localization is performed in a second step. 2) We develop a way to predict estimation performance for a single target at high SNR based on
small-error analysis. 3) The localization problem is presented
in a way which easily lends itself to the application of other
single-step Beamformers.
In the sequel the terms DPD and Maximum Likelihood
refer to the Maximum Likelihood Estimator assuming a single
source, also known as conventional Beamforming, and not
to the exact Maximum Likelihood, which needs to know
(or guess) a-priori the number of sources, and which requires
a multi-dimensional search that is often impractical.

transmitter to the target and then to the -th array, which

satisfies,

II. P ROBLEM F ORMULATION

where the dependence of  on pt is suppressed, r  and n 

are vectors of size Mr K 1, A is a matrix of dimensions
Mr K K and s is a vector of size K 1. We can now write (3)
in a vector form

Consider a transmitter and L widely separated receiving

arrays. Each receiving array consists of Mr elements. The
array aperture is typically a few signal wavelengths. The
transmitted signal consists of several consecutive pulses, where
the waveform of each pulse, s(t), is confined to the time
interval t [0, T ]. It is assumed that the signal s(t) is
perfectly known, which is usually the case when there is line
of sight from the transmitter to the receiving array. In the
following analysis the signal impinges on a single target
whose coordinates vector is denoted by pt , and is reflected
by the target towards the receiving arrays. We assume that the
transmitter, arrays and target are all confined to a plane, and
the transmitter and receiving arrays locations are known while
the target location needs to be estimated.
The -th array output for a single pulse is given by the
Mr 1 vector,
r (t) = a (pt ) s(t  (pt ))ei2 f D, t + n (t)

(1)

where  is the signal attenuation at the -th array,  is

the signal delay associated with the propagation from the

 (pt ) =

pt p  pt pTx 

+
c
c

(2)

where pTx is the transmitters location, p is the -th array

location, and c is the speed of propagation. Further, n (t) is
a Mr 1 wide-sense stationary, white, zero mean, complex
Gaussian noise, a (pt ) is a Mr 1 vector representing the
-th array response to a target at pt , and f D, is the Doppler
frequency shift. We note that without loss of generality we
can impose the constraint a (pt )2 = 1. Since the transmitter
and arrays cooperate the signal transmission time is perfectly
known. This can be accomplished by direct interception of the
transmitted signal or by synchronization of the transmitter and
receiving arrays. Finally, we assume the target is illuminated
by M p consecutive pulses. To simplify the exhibition, it is
assumed that the target speed is small enough to neglect
the Doppler effect. This assumption would be addressed
in Sec. III. The DFT of the received j -th pulse is given by
r ,k ( j ) = a (pt )  ( j ) sk ei2 fk  (pt ) + n ,k ( j )

(3)

where fk = Kk f s is the frequency associated with the

k-th coefficient, K is the number of samples, f s is the sampling
frequency, and r ,k , sk and n ,k are the k-th Fourier coefficients
of r (t), s(t) and n (t), respectively, and where it is assumed
that the observation time is longer than the received signal
interval plus its delays at all sensors.
Define


T
T
T
r  ( j ) = [r,1
( j ) , r ,2
( j ) , . . . , r ,K
( j )]T


T
T
T
n  ( j ) = [n ,1
( j ) , n ,2
( j ) , . . . , n ,K
( j )]T


A (pt ) = diag(e j 2 f 1 , . . . , e j 2 f K  ) a (pt )



s = [s1 , . . . , sK ]T

r  ( j ) =  ( j ) A (pt ) s + n  ( j ) .

(4)

(5)

In the next section we derive the (single-source) Maximum

Likelihood estimator for pt , where { ( j )} are treated as
unknown parameters. As explained in the introduction, it is
possible to derive an exact Maximum Likelihood estimator.
However, such an estimator needs to know a-priori the number
of targets, and it requires a multi-dimensional search. In Sec.
II-B we use the single-source Maximum Likelihood estimator
to obtain the multi-source MVDR estimator, which is our main
goal.
A. Target Localization Using the Maximum
Likelihood Estimator
Using (5) and the fact that {n  ( j )} are statistically independent complex Gaussian vectors, the Maximum Likelihood cost

TZAFRI AND WEISS: HIGH-RESOLUTION DPD USING MVDR

6451

function is
Q(pt ) =

Mp
L 


r ( j )  ( j ) A (pt ) s2 .

(6)

=1 j =1

The signal attenuation { ( j )} is assumed to be independent

from pulse to pulse, as is suggested by the well known
Swerling II and IV target models, which assume the target
radar-cross-section (RCS) is independent from pulse to
pulse (see [15]). This assumption will be revisited in Sec. III.
The attenuation coefficient that minimizes the cost function (6)
is given by
 ( j ) = (s H AH (pt ) A (pt ) s)1 s H AH (pt ) r  ( j )
= (s H s)1 s H AH (pt ) r  ( j )

(7)

where we used
AH (pt ) A (pt ) = I Mr K ||a ||2 = I Mr K .

(8)

Substituting (7) back into (6) we get

Q ML (pt )
Mp
L 



 r  ( j )  1 A (pt ) ss H A H (pt ) r  ( j ) 2
=

=

=1 j =1
Mp 
L 


r ( j ) 
2


2 

H
s A (pt ) r  ( j )

1  H

=1 j =1

Algorithm 1 Target Localization Using Maximum Likelihood

 } according to (12)
Evaluate {R
Define the area of interest and determine a suitable grid of
locations p1 ,p2 pG , where G is the number of grid points
for j = 1 to G do
Set Q(p j ) = 0
for  = 1 to L do
Evaluate  (p j ) according to (2)
to (4)
Evaluate A pj according

Evaluate wML p j according to (15)
  
H
 
 wML
Let Q p j = Q p j + wML R

end for
end for
Find the grid point for which Q is the biggest. This grid
point is the estimated position.

(9)

a single target and no other interference. This expression

can be seen as focusing on the hypothesized
target location

using the weight vectors wML (pt ) , then combining the
responses.
Note that the estimator involves the maximization of a
nonlinear cost function. A possible algorithm, which uses a
simple grid search, is displayed in Algorithm 1. However, it is
possible to use more sophisticated algorithms.

where


 = s H s.

(10)

Since r ( j ) is constant, minimization of this expression is

equivalent to the maximization of
2

Mp
L 
2




ML (pt ) = 1
Q
 1 s H AH (pt ) r  ( j )
Mp
=1 j =1

L

=1

 A (pt ) s
 1 s H AH (pt ) R

(11)

where the sample covariance is a Mr K Mr K matrix,

defined as
Mp
1 

R =
r  ( j )rH ( j ) .
Mp


(12)

j =1

Note that it is always possible, through proper normalization

of the received signals, to set  = s H s = 1.
Finally, we can express the ML estimator as
ML (pt ) = arg max
= arg max Q

pML
t
pt

pt

L

=1

Q ML
 (pt )

where the individual cost functions are given by

H


ML
 wML (pt )
R
Q ML
 (pt ) = w (pt )

(13)

(14)

We now turn to the derivation of the MVDR estimator for

this case. As discussed in the introduction, high resolution
DOA (Direction of Arrival) or frequency-wavenumber estimation can be achieved by replacing the Maximum Likelihood
weight vector with a weight vector designed to minimize the
array output power while maintaining a constant response for
the hypothesized parameter value. The same idea is implemented here for the targets position estimation.
First, note that the Maximum Likelihood cost function (11)
can be expressed as

where

H
ML (pt ) = vML
ML (pt )
Q
(pt ) v

(16)

T  T

ML
ML
, . . . , wL
vML (pt ) = w1

(17)

where the dependence of wML on pt is suppressed, vML (pt )

is a (L Mr K ) 1 vector, and where  is a block diagonal
matrix of dimensions (L Mr K ) (L Mr K ) given by

1 ...
0
R

..
..
 = ...
(18)
.
.
0

and
wML (pt ) = A (pt )s

B. Target Localization Using MVDR

...

L
R

Note that
(15)

is a Mr K 1 weight vector. This ends the derivation for the

Maximum Likelihood estimator of the target position assuming

H
vML
(pt ) vML (pt ) =

H

wML (pt ) wML (pt ) = L
=1

(19)

6452

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO. 9, SEPTEMBER 2016

where we used

H
wML (pt ) wML (pt ) = (A (pt ) s) H A (pt ) s
= s H AH (pt ) A (pt ) s = s H s = 1.
(20)
The proposed MVDR weight vectors should satisfy

vMVDR = arg min v H v,

(21)

v H vML (pt ) = L.

(22)

subject to

This is a minimization of a quadratic function under a linear

constraint. This problem was solved in [16] using the complex
gradient operator. The solution is given by
vMVDR (pt ) = L

H
vML

 1 vML (pt )
.
(pt )  1 vML (pt )

(23)

The MVDR location estimator is therefore given by



H

v

pMVDR
=
arg
max
v
(p
)
(p
)
t
MVDR
t
t
MVDR
pt

= arg max
pt

H (p ) 
1 
 1 vML (pt )
L 2 vML
t

2
H (p ) 
1 v
vML
t
ML (pt )

L2
H (p ) 
1 v
pt
vML
t
ML (pt )

= arg min Q MVDR (pt )

= arg max

(24)

pt

H

=  1 and where
where we used  1
 H
Q MVDR (pt ) = vML
(pt )  1 vML (pt ) .

...

(25)

1
R
L

Substituting (26) and (17) back in (25) we get

Q MVDR (pt ) =

L


Q  (pt )

(27)

=1

where

1 A (pt ) s .
Q  (pt ) = s H AH (pt ) R


easy task. Further, a mistake in the number of targets might

cause severe degradation in its performance. Finally, the search
dimension is Nt times that of the search dimension of the
MVDR estimator developed here, where Nt is the number of
targets. A possible algorithm is displayed in Algorithm 2.
C. Robustness of the Beamformer and Diagonal Loading
In practical scenarios the number of pulses M p is usually
 is singular and cannot be
small. In such cases the matrix R
inverted. We note that
 


R = E R


= E  ( j ) A (pt ) ss H AH (pt ) H ( j ) + n2 I
= 2 A (pt ) ss H AH (pt ) + n2 I

This expression can be further simplified, using the fact that

the inverse of  is a block diagonal matrix as well, given by

...
0
R
1

..
..
(26)
 1 = ...
.
.
0

Algorithm 2 Target Localization Using MVDR

 } according to (12)
Evaluate {R
Define the area of interest and determine a suitable grid of
locations p1 ,p2 pG , where G is the number of grid points
for j = 1 to G do
Set Q(p j ) = 0
for  = 1 to L do
Evaluate  (p j ) according to (2)
Evaluate A p j according to (4)
 
 
1 A s
Let Q p j = Q p j + s H AH R

end for
end for
Find the grid point for which Q is the smallest. This grid
point is the estimated position.

(28)

This ends the derivation of the MVDR target location

estimator. To summarize, the proposed estimator uses the raw
data gathered at the sensors to evaluate the exact MVDR
cost function of the targets position in a single step. Thus,
MVDR is used for direct localization and not the estimation
of the Direction of Arrival (DOA). It is important to note that
no assumption was made on the number of targets during the
derivation of the MVDR estimator. In comparison, an exact
multi-target Maximum Likelihood estimator would need to
know a-priori the number of targets, which is not always an

(29)

where n2 and 2 are, respectively, the noise and signal attenuation variances. Diagonal loading is a well known method
(e.g. [17][19]) for inverting singular matrices. By replacing
 with
R
0 + I
 = R
R

(30)

0 is the original sample covariance matrix given

where R

by (12), the matrix can be inverted. We note that in radar
systems, the noise variance n2 can be estimated by examining
range cells distant from the cell-under-test, similar to the
concept employed in constant-false-alarm-rate (CFAR) filters
(see [20]). The effects of the level of diagonal loading were
shown in [21] for the case of adaptive arrays. It was shown that
diagonal loading reduces the effects of noise by increasing the
noise-associated eigenvalues to the level of . While this has
negligible effect on the nulling of interference represented by
large eigenvalues, it may reduce the nulling capability of weak
interference with small eigenvalues. Using the estimated noise
variance, it is possible to choose the diagonal loading level
according to the interference rejection needs. A good choice
for is between the estimated noise level and the expected
signal level.
III. N UMERICAL E XAMPLES
In the following section we demonstrate the estimation
and resolution capabilities of the proposed algorithm by

TZAFRI AND WEISS: HIGH-RESOLUTION DPD USING MVDR

6453

generated as complex-Gaussian
random vectors as well, thus

n  ( j ) CN 0, n2 I . The SNR is defined as


2
 E  ( j ) A (pt ) s 
SNR =


E n  ( j )2


E  ( j )2 A (pt ) s2
1
=
=
(32)
K Mr n2
K Mr n2
where we used

"
!
E  ( j )2 = 1

s H AH (pt ) A (pt ) s = s H s = 1.

(33)

Thus, the noise variance is given by

1
.
(34)
K Mr SNR
Unless stated otherwise, the number of pulses in all the
simulations is M p = 2. As discussed in Sec. II-C, in order to
 invertible we use diagonal loading, where
make the matrix R
the diagonal loading level is chosen to be n2 , motivated by
the expected sample covariance matrix given in (29).
The simulations compare three positioning algorithms:
the proposed MVDR method, the (single-target) Maximum
Likelihood method, and a 2-step TOA-based localization
method. We are interested in high resolution, so the TOA
estimates for the 2-step method are obtained using MVDR,
instead of traditional Maximum Likelihood based Matched
Filters. This can be achieved by the same derivation
in Sec. II-B and setting the number of receivers to one. The
targets position is then given by minimizing the TOA cost
function
 H 1 

 
(35)
m
m (pt )
m (pt ) TOA
Q TOA (pt ) = m
n2 =

Fig. 1. Problem geometry for single element arrays layout. There are two
targets present, but they are hard to discern due to the figures scale.

numerical examples. We examine the algorithm for the case

of multiple close targets. These targets may represent several
separate targets, one target with several jamming signals,
mimicking legitimate targets, or a single target with multiple resolvable scatterers, given that the systems parameters,
such as the signals bandwidth, provide such high resolution.
We examine two possible layouts, the first consists of single
element arrays (Mr = 1), and the second consists of arrays
composed of several elements (Mr > 1). We emphasize
that in both cases, the sensors themselves do not estimate
TOA (Time of Arrival) or DOA (Direction of Arrival) parameters, but rather record the intercepted signals.
A. Single Element Arrays
The layout of the radar system used during simulations
consists of a transmitter at {300, 0}[km] and four single
element arrays distributed evenly from {100, 150}[km] to
{100, 150}[km]. The problem geometry is shown in Fig. 1.
This geometry is used in all the single-element arrays
simulations.
The simulated signal sampling frequency is f s = 488[kHz],
the number of samples per pulse is K = 128 and the signal
spectrum is flat with bandwidth of 150[kHz]. In order to
gather enough statistics, the simulation results were based
on 1000 Monte-Carlo runs. In each run the channel coefficients were drawn from a normal distribution given by
 T
 
and
 CN (0,  ) where  =  (1) , . . . ,  M p

1
. . . M p 1

..
..
..
.
.
.

 =
(31)
.

..
..
..
.
.

M p 1 . . .
1
is a M p M p matrix, and is the correlation between
the channel coefficients of adjacent pulses, equal to zero
unless specifically stated otherwise. The noise samples were

where m
is a vector of L TOA measurements, L is the number
of receivers, and where
T
 
m (pt ) = 1 (pt ), 2 (pt ), . . . , L (pt )
"
!

(36)
TOA = diag 12 , 22 , . . . , L2
where  (pt ) is given by (2) and where 2 denotes the variance
of the -th TOA measurement, which is equal for all receivers
in the simulated layout. Note that when there are multiple
targets each receiver measures several TOAs. Finding the
association of different TOAs at different receivers is generally
not an easy task. In our simulations we avoid this issue by
associating each measurement with the correct target. It should
be emphasized that this is only an issue for 2-step algorithms.
Fig. 2 and Fig. 3 show the contour plots for the
(single-target) Maximum Likelihood cost function defined
in (11) and the inverse of the MVDR cost function defined
in (27) for the case of two targets located at {0.5, 50}[km]
and {0.5, 50}[km] and no noise. It is clearly seen that the
MVDR method succeeds in separating the targets whereas
the ML method fails. In order to gain some insight into the
difference between single-step and two-step methods, compare
Fig. 2 and Fig. 3 to Fig. 4, which shows the contour plot
for the inverse of the 2-step MVDR cost function defined
in (35). Note that this shows the score of the second step only,

6454

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO. 9, SEPTEMBER 2016

Fig. 2. Single element arrays layout: ML target position cost function in dB.

Fig. 3.
in dB.

Fig. 5. Single element arrays layout: LRMSE as a function of SNR for

two targets.

Single element arrays layout: MVDR target position cost function

Fig. 4. Single element arrays layout: 2-step MVDR target position cost
function in dB.

where the input to the second step is the TOA measurements

which correspond to the target at {0.5, 50}[km].
Next, we examine the effects of SNR on the estimation accuracy. Fig. 5 shows the location-root-mean-squareerror (LRMSE) as a function of SNR for the three

Fig. 6.
Single element arrays layout: LRMSE as a function of for
two targets.

estimation techniques. The estimated targets positions are

the peaks closest to the actual targets position of the corresponding cost functions. Thus, we ignore the problem of
target detection, which could be addressed in future work.
The LRMSE is evaluated separately for each target. However,
performance for both targets are equivalent so only the
LRMSE of one target is shown. Clearly, the single-target
Maximum Likelihood does not improve with SNR while
the MVDR shows excellent performance. The 2-step MVDR
method converges to the 1-step approach, but is inferior at
intermediate SNR.
It is interesting to examine the effect of the value of
presented in (31) on the LRMSE. In the previous section
it was assumed that the signal attenuation is independent
in each pulse, corresponding to = 0. This assumption
may not hold in some cases, e.g. for a slow moving,
non-maneuvering target, or a target with relatively constant
radar-cross-section (RCS). It is desirable to examine the
impact of this dependency on the LRMSE. Fig. 6 shows the
LRMSE for the previous case of two targets, as a function
of the channel attenuation correlation , for SNR = 30[dB].

TZAFRI AND WEISS: HIGH-RESOLUTION DPD USING MVDR

Fig. 7.

Single element arrays layout: LRMSE as a function of SNR.

It is shown that the LRMSE gradually increases with increasing . However, this effect becomes critical only for a very
high correlation of the channel attenuation, leading to the
conclusion that the method proposed holds for practical radar
scenarios. In comparison, the LRMSE for the 2-step method
is less robust. We note that for = 1, corresponding to
a set of linearly dependent signals, the LRMSE using the
proposed method is equivalent to the LRMSE using the
Maximum Likelihood method. Note that the tolerability to
signal attenuation correlation, coupled with the use of M p = 2,
enables the use of a small number of pulses with low pulserepetition-interval (PRI) for processing, resulting in a short
target dwell time, which in turn justifies the assumption that
target speed can be considered negligible for the processing
duration.
It is also of interest to examine the difference in performance
of the estimation methods when only a single target is present.
It is expected that in such a scenario the performance of the
MVDR estimator would be inferior to the optimal single-target
Maximum Likelihood estimator.
Fig. 7 shows the LRMSE as a function of SNR for the
case of a single target located at {0, 50}[km]. The performance of the 2-step method are not shown in order to
reduce clutter, since the superiority of the 1-step approach over
the 2-step method for single target localization was already
shown in [3]. The simulated performance and a theoretical
small-error analysis for the MVDR estimator are all shown
in the figure. The derivation of the small error performance
expression is explained in the Appendix. Also shown in Fig. 7
is the performance for the previous case of two close targets.
We note that both estimation techniques converge to the
theoretical analysis and to one another at high SNR for a
single target. This is a pleasing result, since the DPD method
was shown in [3] to converge to the Cramer-Rao-LowerBound for this estimation problem. We also note that even for
moderate SNR, both estimation methods yield similar results.
Finally, note that the LRMSE for two close targets does not
converge to the LRMSE for one target. This is because the two
targets interfere with each other, causing an estimation bias.
In order to provide a lower bound for the performance of the
estimation, also shown in Fig. 7 is the LRMSE for the exact

6455

Fig. 8. Single element arrays layout: LRMSE as a function of the distance

between targets.

(multi-source) Maximum Likelihood. We emphasize again that

the exact Maximum Likelihood has information which the
other estimators do not, namely, it has to know a-priori the
number of targets. In addition, it requires a multi-dimensional
search, two major drawbacks when compared to the proposed
method. In order to find the exact Maximum Likelihood we
revisit (5). The received signal when several targets are present
is given by
r  ( j ) =

Nt


,m ( j ) A,m (pm ) s + n  ( j )

(37)

m=1

where Nt is the number of targets and pm is the location of the

m-th target. The exact Maximum Likelihood cost function is
Q({pm }m=1,...,Nt )
2
Mp 
Nt
L 






=
,m ( j ) A,m (pm ) s
r  ( j )


=1 j =1

(38)

m=1

Since {,m ( j )} are unknown deterministic constants, we can

write
Q({pm }m=1,...,Nt )
=

Mp
L 

=1 j =1


2
Nt





min r  ( j )
,m ( j ) A,m (pm ) s

{,m ( j )}m 
m=1

(39)
The exact Maximum Likelihood estimate of the set of targets
locations is given by minimizing (39) over all possible combinations of {pm }. The exact Maximum Likelihood performs
better than the proposed method at intermediate SNR, which
could be attributed to the extra information it possesses.
However, at high SNR the MVDR converges to the exact ML
which is the optimal estimator for this problem.
Next, it is interesting to examine what happens when the
distance between the two targets is increased. It is expected
that estimation performance will converge to the performance
of a single target. Fig. 8 shows the LRMSE as a function of
the targets separation d p for SNR = 30[dB], where the targets

6456

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO. 9, SEPTEMBER 2016

Fig. 9. Single element arrays layout: LRMSE as a function of the number

of targets.

Fig. 10. Single element arrays layout: LRMSE as a function of the number
of pulses.

are at { 2p , 50}[km] and { 2p , 50}[km]. Also shown is

the theoretical small-error analysis for a single target at
d
{ 2p , 50}[km]. For small target distance, all estimators fail
to separate the targets, generally estimating a single position
which is somewhere between the true targets locations. Thus,
as the distance is increased the LRMSE increases as well. This
is true up to a point where the distance is high enough so that
the targets are successfully separated. Further increase of the
distance decreases the LRMSE, which finally converges to the
small-error analysis for a single target. It is clearly seen that
the MVDR method achieves target separation for closer targets
than all other methods.
So far our layout consisted of at most two targets. Fig. 9
shows the LRMSE as a function of the number of targets
present. The targets were placed along the x-axis around
{0, 50}[km], spaced 1km from each other. The number of
pulses in this simulation is M p = 32, in order to suppress
the effect of the linear dependence of the signals returned
from different targets. It is evident that the single-step MVDR
performs better than the other methods, especially for a large
number of targets.
In order to complete our analysis of the performance for
more than two targets, Fig. 10 shows the LRMSE as a
function of the number of pulses for 3, 5 and 7 targets for the
single-step and 2-step MVDR methods. As expected, when
M p is below the number of targets separation is not possible
since the rank of the sample covariance matrix is lower than
the number of targets. However, it is evident that the LRMSE
improves rapidly when the number of pulses is increased
further. It can be seen that the single-step MVDR performs
better than the 2-step method for a smaller number of pulses.
So far we have assumed perfect synchronization between the
receivers. We examine now the impact of receivers synchronization errors on the estimators performance. Fig. 11 shows
the LRMSE of the estimators as a function of the synchronization error standard deviation t for SNR = 30[dB] for a single
target at {0, 50}[km] and for two targets at {0.5, 50}[km]
and {0.5, 50}[km]. In order to reduce clutter, the LRMSE of
the 2-step method for a single target is not shown, since it

Fig. 11. Single element arrays layout: LRMSE as a function of the receivers
synchronization error standard deviation.

is practically the same as the LRMSE of both other methods.

Also not shown is the LRMSE of the ML method for two
targets, since it was shown in Fig. 5 that it fails to separate
the targets. It is seen that performance for all estimators for
a single target are mostly the same, and that MVDR achieves
target separation even for high synchronization errors. Note
that synchronization errors smaller than 100nsec can easily be
achieved using, for example, GPS (e.g. [22]). Finally, we note
that MVDR is more robust compared with the 2-step method.
B. Multi Element Receiving Arrays
We now turn to examine the case of receiving arrays
(Mr > 1). Such arrays are usually harder to implement,
as they require calibration of the array response. However,
the additional data improves the LRMSE, possibly allowing
the use of less arrays for target localization, depending on
the problem geometry. To demonstrate this, the layout of
this section consists of a transmitter at {300, 0}[km], and two
receiving arrays, located at {50, 20}[km] and {50, 20}[km].
See Fig. 12. Five elements uniform linear arrays are used,

TZAFRI AND WEISS: HIGH-RESOLUTION DPD USING MVDR

Fig. 12.

6457

Problem geometry for five elements arrays layout.

with half wavelength element spacing. The array baseline is

parallel to the y axis. All other simulation parameters were
unchanged.
Similarly to Sec. III-A, a 2-step based algorithm is used
in the comparison. Here, in addition to measuring TOA,
the sensors are also capable of DOA measurement, which
was calculated using MVDR. We emphasize again that these
measurements of intermediate parameters are only used for
the 2-step approach. The targets position using this approach
is given by minimizing the cost function


Q 2step (pt ) = Q TOA (pt ) + Q DOA (pt )

where Q TOA is given in (35) and


H


1
Q DOA (pt ) = h h (pt ) DOA
h h (pt )

Fig. 13. Five elements arrays layout: ML target position cost function in dB.

(40)

(41)

where h is a vector of L DOA measurements, L is the number

of receivers, and where
T
 
h (pt ) = 1 (pt ), 2 (pt ), . . . , L (pt )
"
!

(42)
DOA = diag 12 , 22 , . . . , L2
where  is the angle between the x axis and the line
connecting the target at the hypothesized location pt and
the -th receiver and where 2 denotes the variance of the
-th DOA measurement. The variances of the DOA and TOA
measurements were calculated numerically.
Fig. 13 and Fig. 14 show the contour plots for the Maximum
Likelihood cost function defined in (11) and the inverse of the
MVDR cost function defined in (27) for the case of two targets
located at {0.5, 20}[km] and {0.5, 20}[km] and no noise.
The number of pulses is M p = 2. Again, the MVDR method
succeeds in separating the targets whereas the single target
Maximum Likelihood method fails. For completeness, Fig. 15
shows the contour plot for the inverse of the 2-step MVDR
cost function defined in (40). Remember that this is the second
step of the localization, where the input for this function is the
DOA and TOA measurements corresponding to the target at
{0.5, 20}[km].
Next, we show that the effects of SNR on the estimation
accuracy are similar to the previous layout. Fig. 16 shows the
LRMSE as a function of SNR for each target using MVDR,
ML, and the 2-step MVDR method. We note that the addition

Fig. 14.
in dB.

Five elements arrays layout: MVDR target position cost function

Fig. 15. Five elements arrays layout: 2-step MVDR target position cost
function in dB.

of array elements allows for as few as two arrays to achieve

high resolution target positioning. The single-step MVDR is
again superior to the 2-step method at intermediate SNR.

6458

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO. 9, SEPTEMBER 2016

superior resolution with respect to the Maximum Likelihood

and to current 2-step MVDR based localization methods by
analyzing multiple near targets localization. The targets used
in the analysis can represent a single target with multiple,
resolvable scatterers or represent separate, small targets. The
fine resolution can be employed to resolving real targets from
many decoys. We have demonstrated localization using a small
number of pulses, and have shown that returns of consecutive
pulses can be highly correlative, allowing the use of this
method in most practical scenarios. Finally, weve developed
a way to predict estimation performance for a single target at
high SNR based on small-error analysis.
A PPENDIX
S MALL E RROR A NALYSIS
Fig. 16. Five elements arrays layout: LRMSE as a function of SNR for two
targets.

We derive the covariance matrix of the MVDR position

error, assuming a single target and small errors. The covariance can be used to predict estimation performance for high
SNR or many pulses. The following could easily be derived
for the ML estimator as well. However, the Cramer-RaoLower-Bound for this estimation problem was derived in [3]
and it was shown that the DPD method converges to it.
The framework and formulas used for the complex calculus
performed here were presented in [23]. For the purpose of a
compact notation, throughout the appendix we use w instead
of wML defined in (15), and the dependence on pt will be
suppressed.
We begin our derivation by defining


pt = p t pt


r 0, ( j ) =  ( j ) A s


r ( j ) = r  ( j ) r 0, ( j ) = n  ( j )
Fig. 17. Multi elements arrays layout: LRMSE as a function of SNR for
two targets.

In order to gain some intuitive understanding of the

difference between the performance of the single-step and
2-step MVDR for multiple targets, Fig. 17 shows the LRMSE
for the three following layouts. 1) The previous 5-element
receiving arrays layout. 2) Two 9-element receiving arrays at
{50, 20}[km] and {50, 20}[km]. 3) Four 5-element receiving arrays at {50, 40}[km], {50, 20}[km], {50, 20}[km] and
{50, 40}[km]. All other parameters are unchanged. It is
evident that the single-step MVDR is consistently superior to
the 2-step method.

(43)

Using the Woodbury identity we get



1


H H
1 
R

A
s

A
=
I
+




{r 0, ( j )}
$
#
A ss H AH
1
=
I


s H AH A s + / 
$
#
A ss H AH
1
=
I


1 + / 

(44)

where is the sample covariance matrix loading factor defined

in (30), and


 =

Mp
1 
| ( j )|2
Mp

(45)

j =1

IV. C ONCLUSION
In this paper we proposed a single-step direct position
determination (DPD) using the MVDR approach rather than
the single target Maximum Likelihood approach. The proposed
method is an adaptive method, using the returns from several
consecutive pulses for computation of the targets location
directly without first estimation of direction of arrival and
delay. We presented DPD based on single target Maximum
Likelihood and DPD based on MVDR for multistatic radar,
but our approach may easily be extended to other localization
problems. We demonstrated that the MVDR approach achieves

We note that

MVDR (pt )
Q
{r

0, ( j ) , j

L

=1


H R
1 
s H A


{r 0, ( j )}

 s
A

$
#
H AH
s

A
1


H
 s
=
s H A
A
I


1
+
/


=1

2


 H H
L
s
A
A
s






1

=
(46)
1

1 + / 
=1
L


TZAFRI AND WEISS: HIGH-RESOLUTION DPD USING MVDR

6459

where we used (44) for the second transition and (8) for the
third transition. It is obvious that minimization of (46) is
achieved by choosing p t = pt . This can be understood in
the following manner: the attenuation coefficient does not
affect the estimated position when there is no noise present.
Since this choice also maximizes the individual cost functions
MVDR , we get
comprising Q


Q  
Q  
=
=0
(47)
pt p t ,{r( j )}
pt pt ,{r0 ( j )}

We now turn to find each of the partial derivatives constituting the above expression.
In the next section, the subscript () is omitted whenever
the discussion is focused on a single receiver-transmitter pair.
 w
. In order to
We begin with an expression for wpT = p
t
achieve compact expressions we assume each receiver consists
of a single element (Mr = 1). It is possible, however,
to develop expressions for the case of a general array response.
Define the quantities
pt
p
pTx
d
dTx

where we used the more compact notation



()|{r0 ( j )} = ()|{r 0, ( j )}

, j

Q
pt

The first order Taylor series of

is given by



Q  
Q  
2 Q  
=
+
pt
pt p t ,{r( j )}
pt pt ,{r0 ( j )}
p2t pt ,{r0 ( j )}
%
&

Mp

2 Q  
+2
Re
r ( j )
pt r  ( j ) pt ,{r0 ( j )}
j =1

(48)
Q
is a real function. Using
where we used the fact that p
t


L

Q MVDR 
Q  
=
(49)

pt 
pt p t ,{r( j )}
p t ,{r ( j )}

=1

and (47) we get

# L
$1

 2 Q 

pt = 2
2 
p
t
pt ,{r0 ( j )}
=1
&
%

M
p
L 

2 Q  

Re
r ( j ) .
pt r  ( j ) pt ,{r0 ( j )}

(50)

E r1 ( j1) rT2 ( j2 )



E r1 ( j1) rH2 ( j2 )


E r ( j ) rT ( j )


E r ( j ) rH ( j )

=0
=0


and thus we get




Cp,p = E pt ptT
# L
$1
 2 Q  


= 2n2
2 
p
t
pt ,{r0 ( j )}

=1
M

 H 
p
L 


2 Q
2 Q

Re


pt r  ( j ) pt r  ( j )

#

=1 j =1

L

=1


2 Q  
p2 
t

$T
.

pt ,{r0 ( j )}

(53)

(x x  )2 + (y y )2
c

(x x Tx )2 + (y yTx )2
c

(54)
The partial derivative with respect to the target coordinate x
is given by
x x

= x
c (x x  )2 + (y y )2
x x Tx
+ (55)
c (x x Tx )2 + (y yTx )2
Define the K K matrix


F = diag ( j 2 f 1 , . . . , j 2 f K )

(56)

Using the definition in (15) we get



A
x x
x x Tx
w
=
s =
(57)
+
FA s
x
x
c d 
c dTx 
which can be expressed in vector form, combining it with the
partial derivative with respect to the target coordinate y, by the
2 K matrix wp , given by

=0

= n = n2 I K Mr

(x, y)T
(x  , y )T
(x Tx , yTx )T
[x x  , y y ]T
[x x Tx , y yTx ]T

where pTx and p are, respectively, the transmitters and

-th receivers locations. The signal delay is given by
pt p  pt pTx 
+
 =
c
- c
-

=1 j =1

Note that

=

=

=

=

=

(51)

wpT =

1
Fw d T
c

(58)

where
dTx
 d
d  =
+
= [cos  + cos Tx , sin  + sin Tx ]T
d  dTx 
(59)

pt ,{r0 ( j )}

(52)

 is the angle between the x axis and the line connecting the
target and the -th receiver. Tx is the angle between the x
axis and the line connecting the target and the transmitter.
q
q
Next, we derive the expressions for ppt and r ( pj ) , where


Q
qpT = p
. Note that the individual cost functions defined
t
in (28) can be expressed as

1 w
Q  = wH R


(60)

6460

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO. 9, SEPTEMBER 2016

Using the chain rule, we get

Q
Q w
Q w
=
+
pt
w pt
w pt
1 wpT + wT R
T wpH
= wH R
!
"
2
1 Fw d T
= Re w H R

c


qpT =

(61)

where we used (58), and where zf is defined as the partial

derivative of f with respect to z when z is treated as a
constant, where z is the complex conjugate of z. Similarly,
f
z is defined as the partial derivative of f with respect
to z when z is treated as a constant. (61) can also be
expressed as

1 H 1
T F w d 
w R Fw + wT R
(62)
qp =
c
The partial derivatives of qp with respect to w and w are
given by the 2 K matrices

qp
1
1 + w H R
1 F
= d  w H F H R
(63)
w
c
and

qp
1 T T
T T
w
d
R
=
F
R
+
w
F

w
c

(64)

Using the chain rule and some simple algebraic manipulations,

we get
qp
qp T
qp H
wp +
=
w
pt
w
w p
!
"
2
1 Fw + w H R
1 FFw d T
= 2 d  Re w H F R
c

(65)

(66)

r g ( j )

= conj

r g ( j )



1 H
1 Fww H + ww H F R
1
d r ( j ) R
c Mp

(70)

where r ( pj ) is a 2 K matrix. To summarize, using (52),

(65) and (70) one can calculate Cp,p and the LRMSE for the
proposed method for high SNR.
In order to obtain some insight we examine this analysis
when the signal attenuation is constant for all pulses. This
condition can be expressed, without loss of generality,
as  ( j ) = 1 for all  and j . Note that under this condition
we get

(71)

Substituting (71) and (44) in (65) we get

0 H
1
1
qp
2
w F Fw 1+
w H F ww H Fw T
= 2 d  Re
d

+ 1+
w H FFw
pt
c

(72)
= 2 d  2 2 d T
c
where we used 0. and where


K


2 f k |w (k)|2 =

k=1

K


2 f k |sk |2

k=1

2
2
3 K
3 K
3
3

 4
2
2
=
(2 fk ) |w (k)| = 4
(2 f k )2 |sk |2

(73)

k=1

and w (k) is the k-th element of w , and where we used

|w (k)|2 = |sk e j 2 f k  |2 = |sk |2
(67)

Define

2
3 K
3
 4
=
(2 f k )2 |sk |2

(74)

(75)

k=1

which can be expressed as

1 T T T T
=
e R F w w R r ( j )
Mp g

k=1

1 Fw
wH R

pt ,{r0 ( j )}

= ww H
R

where
is a matrix of dimensions 2 2.
Next, we derive the partial derivative of qp with respect to
the received signal r ( j ). The partial derivative of (62) with
respect to the g-th element of the received signal, denoted
by r g ( j ), is given by

HR
TR
1 Fw
T F w

w
qp
1
d 
=
+
r g ( j )
c
r g ( j )
r g ( j )

and

T F w
wT R

where we used the fact that the terms in the brackets are
scalars. We can now write (69) in vector form

qp 
r ( j ) 

r ( j ) = w

qp
pt

The partial derivatives in (66) are given by

1 Fw
wH R

1 R R
1 Fw
= w H R
r g ( j )
r g ( j )
1 H 1 H
1 Fw
=
w R eg r ( j ) R
Mp

where eg is a column vector where each element is 0 except

for the g-th element, equal to 1. Substituting (67) and (68)
back in (66) we get

qp 
r g ( j ) pt ,{r0 ( j )}
/
.
1 eg
1 Fww H R
1 r H ( j ) R
d
=
T F w wT R
T r ( j )
+egT R
c Mp


1 Fww H R
1
1 r H ( j ) R
=
d
(69)
1 ww H F R
1 eg
c Mp
+r H ( j ) R

(68)

2 = 2 2

K

k=1

2 f k |sk |2 + 2 s H s = 2 2

(76)

TZAFRI AND WEISS: HIGH-RESOLUTION DPD USING MVDR

6461

Note that and can be understood as the central frequency

and effective bandwidth of s, respectively. Next, substituting (71) and (44) in (70) we get
qp
1
=
d  w H
r ( j )
cM p (1 + )


Fww H + ww H F


1
Fww H + ww H F ww H
1+

1 H

d w F w H F ww H
=
cM p

(77)

Note that



qp
qp H
r ( j ) r ( j )

1
H
H
H
T

= 
2 d  w F Fw w F ww Fw d 
cM p

1
2
2
T
= 
2 d  d 
cM p

Finally, substituting (72) and (78) in (52) we get

. L
/1
n2 c2  T

d d
Cp,p =
2 Mp 2

(78)

(79)

=1

Furthermore, using (59) we get

 2

d d T = u  u  v2 
u v v

(80)

where

[10] J. J. Handfield, R. M. Rao, and S. A. Dianat, Near-field MVDR source

localization, Proc. SPIE, vol. 6980, p. 69800O, Apr. 2008.
[11] D. W. Rieken and D. R. Fuhrmann, Generalizing MUSIC and MVDR
for multiple noncoherent arrays, IEEE Trans. Signal Process., vol. 52,
no. 9, pp. 23962406, Sep. 2004.
[12] L. Xu, J. Li, and P. Stoica, Target detection and parameter estimation
for MIMO radar systems, IEEE Trans. Aerosp. Electron. Syst., vol. 44,
no. 3, pp. 927939, Jul. 2008.
[13] K. Luo and A. Manikas, Superresolution multitarget parameter estimation in MIMO radar, IEEE Trans. Geosci. Remote Sens., vol. 51, no. 6,
pp. 36833693, Jun. 2013.
[14] A. Hassanien, S. Shahbazpanahi, and A. B. Gershman, A generalized
capon estimator for localization of multiple spread sources, IEEE Trans.
Signal Process., vol. 52, no. 1, pp. 280283, Jan. 2004.
[15] P. Swerling, Probability of detection for fluctuating targets, IRE Trans.
Inf. Theory, vol. 6, no. 2, pp. 269308, 1960.
[16] D. H. Brandwood, A complex gradient operator and its application in
adaptive array theory, IEE Proc. F Commun., Radar Signal Process.,
vol. 130, no. 1, pp. 1116, Feb. 1983.
[17] J. Li, P. Stoica, and Z. Wang, On robust capon beamforming and diagonal loading, IEEE Trans. Signal Process., vol. 51, no. 7, pp. 17021715,
Jul. 2003.
[18] W. Wang, R. Wu, and J. Liang, A novel diagonal loading method
for robust adaptive beamforming, Prog. Electromagn. Res. C, vol. 18,
pp. 245255, Jan. 2011.
[19] T. J. Laseetha and R. Sukanesh, Robust adaptive beamformers using
diagonal loading, Cyber J., Multidiscipl. J. Sci. Technol.-J. Sel. Areas
Telecommun., vol. 2, no. 3, pp. 7379, 2011.
[20] M. I. Skolnik, Introduction to radar, Radar Handbook, vol. 2. New
York, NY, USA: McGraw-Hill, 1962.
[21] B. D. Carlson, Covariance matrix estimation errors and diagonal
loading in adaptive arrays, IEEE Trans. Aerosp. Electron. Syst., vol. 24,
no. 4, pp. 397401, Jul. 1988.
[22] W. Lewandowski, J. Azoubib, and W. J. Klepczynski, GPS: Primary
tool for time transfer, Proc. IEEE, vol. 87, no. 1, pp. 163172,
Jan. 1999.
[23] K. Kreutz-Delgado. (2009). The complex gradient operator and the
CR-calculus. [Online]. Available: http://arxiv.org/abs/0906.4835

u  = cos  + cos Tx


v  = sin  + sin Tx

(81)

Delving into (79), we note that the estimation error is inversely

proportional to the number of pulses and to the effective
bandwidth of the transmitted signal. Also note that the error is
independent of transmitter-target range and the target-receivers
ranges.

Liran Tzafri (S16) was born in Israel in 1986.

He received the B.Sc. degree in electrical engineering from the TechnionIsrael Institute of Technology, Haifa, in 2009, and the M.Sc. degree in
electrical engineering from Tel Aviv University,
Tel Aviv, Israel, in 2014. His main research interests
are statistical and array signal processing for localization, communications, and estimation theory.

R EFERENCES
[1] A. Amar and A. J. Weiss, Localization of narrowband radio emitters
based on doppler frequency shifts, IEEE Trans. Signal Process., vol. 56,
no. 11, pp. 55005508, Nov. 2008.
[2] A. J. Weiss, Direct geolocation of wideband emitters based on delay and
Doppler, IEEE Trans. Signal Process., vol. 59, no. 6, pp. 25132521,
Jun. 2011.
[3] O. Bar-Shalom and A. J. Weiss, Direct positioning of stationary targets
using MIMO radar, Signal Process., vol. 91, no. 10, pp. 23452358,
2011.
[4] J. Capon, High-resolution frequency-wavenumber spectrum analysis,
Proc. IEEE, vol. 57, no. 8, pp. 14081418, Aug. 1969.
[5] S. R. DeGraaf, SAR imaging via modern 2-D spectral estimation
methods, IEEE Trans. Image Process., vol. 7, no. 5, pp. 729761,
May 1998.
[6] N. J. Willis, Bistatic Radar. Raleigh, NC, USA: SciTech Publishing,
2005.
[7] E. Fishler, A. Haimovich, R. Blum, D. Chizhik, L. Cimini, and
R. Valenzuela, MIMO radar: An idea whose time has come, in Proc.
IEEE Radar Conf., Apr. 2004, pp. 7178.
[8] A. M. Haimovich, R. S. Blum, and L. J. Cimini, MIMO radar with
widely separated antennas, IEEE Signal Process. Mag., vol. 25, no. 1,
pp. 116129, Jan. 2008.
[9] J. Li and P. Stoica, MIMO radar with colocated antennas, IEEE Signal
Process. Mag., vol. 24, no. 5, pp. 106114, Sep. 2007.

Anthony J. Weiss (S84M85SM86F97)

received the B.Sc. degree from the TechnionIsrael
Institute of Technology in 1973, and the M.Sc.
and Ph.D. degrees from Tel Aviv University, Tel
Aviv, Israel, in 1982 and 1985, respectively, all in
electrical engineering. From 1973 to 1983, he was
involved in research and development of numerous
projects in the fields of communications, command
and control, and emitter localization. In 1985, he
joined the Department of Electrical EngineeringSystems, Tel Aviv University. From 2006 to 2011,
he was the Head of the EE School with Tel Aviv University. He authored
nearly 200 papers in professional magazines and conferences. He also held
leading scientific positions with Signal Processing Technology Ltd., Wireless
on Line Ltd., and SigmaOne Ltd. His research interests include detection and
estimation theory, signal processing, sensor array processing, and wireless
networks. He was a recipient of the IEEE 1983 Acoustics, Speech, and Signal
Processing Societys Senior Award and the IEEE Third Millennium Medal.
He holds nine U.S. patents. From 1996 to 1999, he served as the Department
Chairman and the Chairman of the IEEE Israel Section.