Phase Synchronization For Coherent MIMO Radar Algo

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 21

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/224248836

Phase Synchronization for Coherent MIMO Radar: Algorithms and Their


Analysis

Article  in  IEEE Transactions on Signal Processing · December 2011


DOI: 10.1109/TSP.2011.2162509 · Source: IEEE Xplore

CITATIONS READS

45 674

2 authors:

Yang Yang Rick S. Blum


Wuhan University Lehigh University
199 PUBLICATIONS   4,242 CITATIONS    403 PUBLICATIONS   13,782 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Information Theory of MIMO RADAR View project

Photocatalysis View project

All content following this page was uploaded by Rick S. Blum on 14 July 2014.

The user has requested enhancement of the downloaded file.


5538 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 11, NOVEMBER 2011

Phase Synchronization for Coherent MIMO Radar:


Algorithms and Their Analysis
Yang Yang and Rick S. Blum, Fellow, IEEE

Abstract—Multiple-input multiple-output (MIMO) radar can fit within a given target beamwidth. For the case of coherent
achieve improved localization performance by employing a co- MIMO radar, the antennas are within a given target beamwidth,
herent processing approach with proper antenna positioning. and this leads to identical reflection coefficients. The optimum
Coherent processing, however, entails the challenge of ensuring
phase coherence of the carrier signals from different distributed processing of the coherent MIMO radar returns has been
radar elements. In this work, we aim to address such a challenge shown to give excellent performance in some very important
by providing a systematic treatment of the phase synchronization cases, such target localization [2], [3], but it requires phase
problem in coherent MIMO radar systems. We propose and study synchronization between the separated antenna elements1. In
three different approaches for reaching a common notion of phase fact, phase synchronization embodies a major difference be-
in coherent MIMO radar, namely, the master-slave closed-loop
algorithm, the round-trip algorithm and the broadcast consensus tween the operations of noncoherent MIMO radar and coherent
based algorithm. These algorithms range from centralized to MIMO radar, as will be mathematically demonstrated later
distributed types, and include both noniterative and iterative in this work. Unfortunately, in reality, the radar elements are
approaches. They do not require a priori establishment of the time usually operated with physically different local oscillators, and
synchronization, and thus are all time asynchronous in nature. each of them suffers from statistically independent phase offset,
Under a similar analytical framework, we mathematically charac-
terize each of these algorithms, and further derive and study the indicating that the phase of the carrier signal transmitted gets
statistical properties of a few relevant figures of merit including rotated by an unknown amount. As a consequence, practical
the resulting phase synchronization error. Simulation results are implementation of coherent MIMO radar entails the challenge
presented to validate our theoretical analysis. of ensuring phase coherence of the carrier signals from different
Index Terms—Coherent processing, multiple-input mul- distributed radar elements.
tiple-output (MIMO) radar, phase synchronization. The imperfection in phase synchronization that would in-
evitably occur in coherent MIMO radar, on the other hand,
has recently attracted the attention of some researchers. For
I. INTRODUCTION
example, some theoretical analysis conducted on coherent

T HE last few years have seen a great surge of interest and


progress in understanding the concept of multiple-input
multiple-output (MIMO) radar and developing paradigms
MIMO radar has taken into account the phase synchronization
error or mismatch, such as [4]–[7]. However, the technical
feasibility of phase coherence in coherent MIMO radar as well
to assess its potential and actual effectiveness. For all cases as approaches for achieving such phase coherence has received
where the MIMO radar antennas are not colocated, the delays very little attention. Moreover, although the synchronization
from a given transmit antenna to a given receive antenna problem has been intensely studied in several closely related
change significantly as we change the transmit and/or receive contexts, for example, in traditional radar systems (see, e.g.,
antenna involved. Among all these noncolocated cases are two [8]–[14]) and in wireless networks including sensor networks
cases that have received special attention: one called coherent (see, e.g., [15]–[22]), we have not seen much work that is
MIMO radar and one called noncoerent MIMO radar [1]. The specifically targeted at attaining phase coherence, or phase
distinguishing feature is whether the antennas are spaced to synchronization in coherent MIMO radar. Without doubt, this
line of work, which addresses the challenge that is inherent
Manuscript received January 21, 2011; revised May 24, 2011; accepted July to coherent MIMO radar, is of both practical and theoretical
11, 2011. Date of publication July 22, 2011; date of current version October 12,
importance, and merits interest on its own terms.
2011. The associate editor coordinating the review of this manuscript and ap-
proving it for publication was Prof. Visa Koivunen. This work was supported As a result, a systemic treatment of the phase synchroniza-
by the U.S. Army Space and Missile Defense Command under contract number tion problem in coherent MIMO radar becomes the primary
W9113M-08-C-0221 to ANDRO Computational Solutions, LLC, who in turn,
objective of our study, which, is also a major contribution of
sponsored a subcontract to Syracuse University. The U.S. Government is autho-
rized to reproduce and distribute reprints for Governmental purposes notwith- our work. As far as the novelty of our work goes, the very
standing any copyright notation thereon. The views and conclusions contained few efforts to date to this problem has made ours the first of
herein are those of the authors and should not be interpreted as necessarily rep-
its kind, according to our best knowledge. More specifically,
resenting the official polices or endorsements, either expressed or implied, of
the US Army Space and Missile Defense Command, ANDRO Computational in our work, we propose and discuss three different yet ef-
Solutions, LLC, Syracuse University or the U.S. Government. ficient approaches to achieve phase coherence in coherent
The authors are with the Department of Electrical and Computer Engineering,
MIMO radar, namely, the master-slave closed-loop algorithm,
Lehigh University, Bethlehem, PA 18015 USA (e-mail: yay204@lehigh.edu;
rblum@ece.lehigh.edu). the round-trip algorithm and the broadcast consensus based
Color versions of one or more of the figures in this paper are available online algorithm. These phase synchronization algorithms range from
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TSP.2011.2162509 1In this paper, we use the terms sensor, node, and element interchangely.

1053-587X/$26.00 © 2011 IEEE


YANG AND BLUM: PHASE SYNCHRONIZATION FOR COHERENT MIMO RADAR 5539

centralized to distributed types, and include both noniterative nections, such as coaxial cable or fiber optic links. For example,
and iterative methods. They are versatile and practically imple- the broadcast consensus based phase synchronization algorithm
mentable, and should cater to different needs arising as a result given in this paper is applicable to other types of media than the
of different number of MIMO radar elements, different sensor wireless environment, as long as broadcast is enabled. Addition-
placements, as well as different operational preferences. They ally, we assume sensors operate in the half-duplex mode, which
are all time asynchronous algorithms, and do not require the means that they cannot receive while they are transmitting.
establishment of time synchronization in advance. In particular, The remainder of this paper is organized as follows. In
the master-slave closed-loop algorithm and the round-trip algo- Section II, we give a brief introduction to the coherent MIMO
rithm also require no a priori frequency synchronization. Under radar model, demonstrate the difference between coherent and
a similar analytical framework, for each of these algorithms, we noncoherent processing for target localization, and most im-
derive and also provide an analysis of the statistical properties portantly, elucidate the motivation of our work. In Section III,
of some interesting figures of merit, such as the aggregate we describe a simple master-slave closed-loop phase synchro-
phase error after synchronization and the phase deviation from nization approach, and examine the statistical properties of the
consensus. Our analysis clearly reveals the impact of the phase cumulative phase error. In Section IV, we propose a round-trip
and frequency estimation errors on the phase synchronization phase synchronization algorithm, and analyze the phase syn-
accuracy. chronization error. In Section V, we propose a broadcast
Of particular note is that while each of these phase syn- consensus based phase synchronization algorithm, and study
chronization algorithms may resemble some algorithm(s) in some convergence properties of this algorithm. In Section VI,
existing works that are designed for or applied to situations we summarize and compare these three phase synchronization
other than phase synchronization in coherent MIMO radar, algorithms, and discuss some issues that may arise in practice.
they do contain distinct features and contributions when in In Section VII, we provide some simulation results on the
comparison. For example, [11] proposes a broadcast reference broadcast consensus based algorithm. In Section VIII, we end
self-cohering technique for calibrating a large, sparse, flexing this paper with some concluding remarks.
array that has a similar physical setup as the coherent MIMO
radar system. This technique directly broadcasts phase refer- II. COHERENT MIMO RADAR MODEL
ence signals from a reference element to other array elements,
In this section, we provide a concise description of the co-
which resembles our master-slave closed-loop algorithm. It,
herent MIMO radar model, and illustrate the difference between
however, requires several auxiliary beacons with either accu-
coherent and noncoherent processing in the context of target
rately known or loosely known locations to provide additional
localization. Note that to make the discussion very clear, we
phase measurements. Note that none of our proposed methods
consider here some simplified assumptions. Let us consider a
requires auxiliary beacons or alike. Another approach, similar
MIMO radar that is equipped with transmitters and re-
to our master-slave closed-loop algorithm, has been used in
ceivers. Thus, the total number of radar elements is ,
[18] to synchronize the carrier phase of a group of nodes,
and for notational convenience we denote . The
but [18] implicitly assumes the a priori establishment of time
positions of the th, transmitter and the th,
synchronization. In contrast, our work does not assume either
receiver are and , respectively,
time nor frequency synchronization, and features an asyn-
in a two-dimensional Cartesian coordinate system. The lowpass
chronous time model and timing analysis. The round-trip type
equivalent of the signal transmitted from the th transmitter is
method proposed in [19]–[22] for global clock synchronization
in sensor networks or carrier synchronization for distributed , where denotes the total transmitted energy, and
beamforming share the same essential features as our round-trip the waveform is normalized such that
algorithm, but in general they cannot be applied to our problem
in a straightforward manner. Besides, in our work, we provide
a thorough analysis of the resulting phase synchronization error
at each radar element, which is not available in these works.
Finally, the application of broadcast-based consensus algorithm Assume the target’s location and velocity are and ,
to phase synchronization and the analysis of its convergence respectively. For notational convenience, we define a parameter
behavior in the presence of nonzero-mean phase perturbations, vector as follows:
is an extension of our work in [38], and is new on its own.
Therefore, each algorithm proposed in our work represents a
unique addition to their respective category of methods. Overall,
our work, providing a practical orientation along with a theoret-
ical depth, is a first attempt to explicitly address practical issues In the coherent mode, the antennas are assumed to be all
related to phase synchronization in coherent MIMO radar. It is within the effective beamwidth of the target, if the target is
also worth noting that in our work, we only consider space as modeled as an antenna [1]. Thus, the reflection coefficient is
the transmission medium through which phase synchronization assumed to be the same for each th path, and is denoted as
is implemented, but we note that these synchronization algo- . We assume is unknown but deterministic2. Then, the signal
rithms can remain effective when phase synchronization is to 2Considering random leads to very similar results but the deterministic case
be achieved across different radar elements through wired con- simplifies the analysis.
5540 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 11, NOVEMBER 2011

emanated from transmitter and captured by receiver can be As the noise components are i.i.d. Gaussian, the log-likelihood
expressed as function of the received signal vector, after simplification, can
be expressed as
(1)

where is the carrier frequency of the probing signal; and


represent the time delay and Doppler shift, respectively, cor-
responding to the th path. In particular, the time delay
(3)
is a function of the unknown target location , i.e.
where and are constants independent of the parameters to
be estimated. Thus, the ML estimate of the deterministic but
unknown parameter vector using coherent
processing is given by [2]

where denotes the speed of light, the distance between the


target and the th transmitter, and the distance between the
target and the th receiver. The Doppler shift is a function of
the unknown target location and velocity , i.e.,
(4)

For the purpose of comparison, we also briefly describe the


noncoherent processing approach. In this mode, the antennas are
where denotes the wavelength of the carrier. In (1), de- assumed to be separated sufficiently so that each path observes
notes the noise-plus-clutter for the th path. For simplicity an independent aspect of the target. If we denote the reflection
of discussion, we assume is a temporally white, zero- coefficient corresponding to the th path as , then is
mean complex Gaussian random process. More specifically, we independent for different and/or . Further, we model as
have a zero-mean complex Gaussian random variable with variance
, i.e., . This is usually referred to as the
Swerling I model [24]. Then the signal emanated from trans-
mitter and captured by receiver can be expressed as
where is a constant and is a unit impulse function. Ad-
ditionally, to explain the concept of coherent processing in a
(5)
simple way, we assume elements of the noise-plus-clutter are
spatially white, i.e.,
The ML estimate of using noncoherent processing is given by
[23] [see (6) at the bottom of the page].
Comparing (4) and (6), it is clearly observed that when es-
Let be an vector which contains all the signals timating the unknown target parameters, the relative phase in-
received at the receive elements, i.e., formation at different antennas is utilized in the coherent pro-
cessing, which is in contrast to the noncoherent case. As a result,
implementation of coherent MIMO radar systems needs to cope
with the fundamental challenge of ensuring a common notion of
The scatter reflectivity can be estimated using the maximum phase among all the antennas, and calls for efficient algorithms
likelihood (ML) method, and we provide this result directly as which can lead to good performance in terms of phase coher-
follows ence. In the sequel, we propose and discuss three schemes of
this kind, and further provide some analysis on their accuracy.

III. MASTER-SLAVE CLOSED-LOOP PHASE SYNCHRONIZATION


We present in this section an intuitive yet very simple phase
synchronization scheme. This scheme employs a master-slave
(2) architecture as illustrated in Fig. 1, where we denote sensor

(6)
YANG AND BLUM: PHASE SYNCHRONIZATION FOR COHERENT MIMO RADAR 5541

the channel phase information can be measured in a certain way


and then fed back explicitly to each slave node for appropriate
phase compensation, the accuracy of phase synchronization can
be much improved. This is the basic idea behind the closed-loop
approach, another type of master-slave synchronization. How-
ever, a fundamental premise that lies underneath this approach is
the channel reciprocity, which is assumed in our work, meaning
that the channel between a pair of sensors remains the same for
both forward and reverse directions.
In the following, we describe a specific time-slotted closed-
loop phase synchronization algorithm. In particular, we assume
each slave element is assigned a specific time slot, during which
Fig. 1. Master-slave architecture. they can exchange information with the master element . We
further assume each slave element knows the number (or order)
of the time slot assigned to it. To this end, this time-slotted
as the master node, and , as slave nodes. All closed-loop approach is summarized as follows.
the slave sensors will then synchronize their local phase with 1) The master element broadcasts an unmodulated beacon
that of the master node. Note that a similar master-slave archi- signal to all the slave elements.
tecture has been employed in [17] and [18] for the distributed 2) Upon receiving the beacon signal, each slave sensor per-
beamforming application. In particular, in [18], the authors also forms the following steps concurrently:
used a master-slave based approach to synchronize the carriers a) It estimates the frequency and phase of the received
of a group of source nodes, but they implicitly assumed time signal.
synchronization as given a priori. In contrast, here we assume b) It begins to count the elapsed time using its own local
no a priori establishment of time synchronization across all the clock, and waits for its time slot to come.
radar elements, at least not one with sufficient accuracy to en- 3) When the time slot for slave element starts according
able phase synchronization. Each radar element maintains its to its local clock, the following steps will take place in a
own time using its own independent local oscillator, and is able sequential order:
to track the elapsed time using its local clock with relative pre- a) passes the received beacon signal back to , using
cision. As a result of the absence of a common time reference, the local phase and frequency estimates generated
the instantaneous phase of each radar element can no longer be from its observation.
precisely determined by the local time at that element. For sim- b) Upon receiving the returned beacon signal from ,
plicity, we assume all the radar elements have a similar notion estimates the received phase, and computes the
of the length of a specified time interval, e.g., for a time slot.3 phase difference relative to the originally transmitted
In practice, this can be achieved by using identical stable clocks reference signal. It divides the phase difference by
(e.g., atomic clocks) across all the radar elements. Moreover, we two, precompensates the phase offset, and then trans-
also assume no a priori frequency synchronization among these mits the beacon signal back to .
nodes. But we assume all the radar elements are equipped with c) After receiving the reference signal, estimates
identical frequency and phase estimators. the phase information and adjusts its carrier phase
Given this master-slave architecture, there exist two basic accordingly.
variations, depending on how the master element interacts with As is evident from the above procedures, it is essential to
each of these slave elements. The first one is the open-loop ap- estimate the pre-compensation phase offset for each slave
proach, in which the master node broadcasts an unmodulated node when performing this closed-loop phase synchronization
sinusoidal reference signal to all the slave nodes, and each of scheme. But since the channel between the master node and
these slave nodes will then use such a reference signal to es- each slave node may vary from time to time, though at a
timate and correct its phase offset. In this way, interaction be- slow rate, the value of the pre-compensation offset has to be
tween the master node and slave nodes is minimal. However, the estimated and updated in a periodic manner.
absence of a common time reference, as we pointed out earlier,
makes it difficult for each sensor to generate absolute phase with A. Protocol Timing Analysis
respect to the true time. As a result, phase shift induced by the Since we assume no a priori time synchronization and as-
channel between the master node and each slave node can not sume each slave element keeps its own local time using its own
be disambiguated at each slave node, and hence cannot be com- independent oscillator, it is important to coordinate the trans-
pensated either. So, this open-loop method may lead to inferior missions of these slave sensors in order to minimize or avoid any
accuracy in phase synchronization. On the other hand, suppose possible collisions. To facilitate the ensuing analysis, we model
3Note that, nevertheless, this assumption can be relaxed, for discrepancies in the channel from the master element to each slave element
the actual lengths of the time interval at these radar elements will not affect as a linear time-invariant (LTI) system with impulse response
the final accuracy of our algorithms (as can be seen later from the analysis of . For simplicity, we focus on the single-path and time-in-
the phase synchronization error), though in some cases they can cause possible
transmission collisions during the phase synchronization process if is not well variant channels, i.e., , where
chosen. is a real number, and denotes the propagation delay for the
5542 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 11, NOVEMBER 2011

channel between and . Due to the channel reciprocity, we Combing (8) and (9), we have
have and thus .
Then as far as timing is concerned, we make the following (10)
proposition: (11)
Proposition 1: Each slave element , upon the conclusion
of receiving the beacon signal from the master element , (12)
needs to defer a period of time before it initiates its time slot
(13)
counting. The defer time, denoted as , should satisfy the
following condition:
where in both (11) and (13), we apply the fact that and
.
Remark 1: Of note is that this algorithm is also applicable
Note that Proposition 1 can make sure that each slave node to cases where the communication channels are multipath LTI
will not confuse the initial beacon signal sent by with the channels. The key is to modify and accordingly to accom-
bounced signal from any other slave node. Further, the length of modate the duration of the delay spread of the multipath chan-
the time slot needs to be judiciously designed, so that slave ele- nels, and to estimate the phase and frequency of the received
ments will not interfere with each other when they interact with signals when such signals are composed of replicas of the trans-
the master element. We assume that before the defer time ex- mitted signals corresponding to the full range of delays. We omit
pires at each slave node, each slave node has successfully gener- further discussion for the sake of brevity.
ated estimates of the frequency and phase of its received signal.
B. Phase Synchronization Error Analysis
We also assume the processing time at the master element is
. Then, we make the following proposition with regard to the In the following analysis, for simplicity we assume the overall
length of the time slot. time for phase synchronization is sufficiently short, such that the
Proposition 2: The duration of a time slot, , should satisfy oscillator phase noise [25], [26] is negligible and does not cause
the following condition: appreciable changes to both the phase and frequency informa-
tion. Let the unit-amplitude sinusoidal reference signal gener-
(7) ated at the master element using its local oscillator be given
by
Proof: Suppose slave sensor and are scheduled to
transmit in two adjacent time slots. Let sensor be the first one (14)
to transmit and its time slot number be , where
where and are the carrier frequency and phase of the
. Let us use the instant when the master element broadcasts
reference signal, respectively. broadcasts this signal to all
the beacon signal as the time reference point. Then the ending
the slave elements. The signal received by slave element ,
time of ’s time slot is
, can be written as

(15)
and the starting time of ’s time slot is
where . Given the noisy observation
in (15), slave element generates estimates of the received fre-
quency and phase, denoted as and , respectively. Using
Obviously when , or equivalently , there will its estimates, transmits the following signal back to in its
be no time overlap at all. But when , or equivalently designated time slot
, will start its time slot before the time slot as-
signed to ends.
To avoid collisions resulting from the overlap of time slots, The signal captured by can be written as
the duration of should include a guard time interval or cushion
duration so that even if starts its time slot before ’s ends,
no collision will happen as a result. The guard time should be at
least as long as the maximum possible length of time overlap, where . then generates an estimate
i.e. of the received phase, which is denoted as . Therefore, the
phase offset that is induced by the propagation channel can be
computed through4
(8)
Note that should include the duration of the round-trip com- (16)
munication between each slave sensor and the master sensor,
4Ideally the phase offset should be converted into a value in the radian range
and the processing time at the master sensor. This leads to
or . In this paper, for the ease of presentation, we choose not to
make such an explicit conversion, but note this radian conversion can be made
(9) after the phase synchronization process completes.
YANG AND BLUM: PHASE SYNCHRONIZATION FOR COHERENT MIMO RADAR 5543

This phase offset is used for the purpose of phase pre-com- where . Specific to this case, we define the
pensation, and is incorporated into the signal to be transmitted SNR as . Then, under the assumption of reasonably high
back to . Thus, the transmit signal is expressed as SNR, ML estimates and of and can be com-
puted, respectively, through [27] [see (20) and (21), shown at the
bottom of the page, where denotes the measure-
and the received phase at is given by ment phase of ]. From (20) and (21), we can see that im-
plementation of the above ML estimators makes use of both the
(17) measurement phase and the measurement mag-
nitudes of the received signal samples
An estimate of is generated at , and is denoted as . . It, however, requires neither the knowledge of the
is the final phase at , as a result of the phase synchronization signal amplitude , nor that of the noise power . Addition-
process. ally, since the measurement phase is obtained from
In the following, we will compute the aggregate phase error the principal argument of the complex phasor , phase un-
resulting from the implementation of this synchronization wrapping [28] is necessary here, which is a process to recover
algorithm, and will also characterize its statistical properties. the absolute value of the phase angle of a continuous function
But firstly, let us examine the frequency and phase estimates at that extends over a range of more than (relative to a pre-
each radar element. In general, these estimates can be obtained defined starting point). A phase unwrapping algorithm which
through a phase-locked loop (PLL). However, as indicated in complements the frequency and phase estimation in (20) and
[18], there is significant sub-optimality in using an analog PLL, (21) can be found in [27]. Note that due to the application of
and digital processing is more preferred when better accuracy this phase unwrapping procedure, the mod- phase ambiguity
is desired. Therefore, we apply the ML estimator instead to ob- issue5, which was discussed in [11] and [18], will not be a con-
tain estimates of these unknown but deterministic parameters. cern here.
While, as it is well known that nonlinear estimation is generally Let the complex AWGN in (19) be decomposed
plagued by threshold effects, which are closely related to the into two orthogonal components, and ,
signal-to-noise ratio (SNR), we hereby assume the channel with being parallel to, and perpendicular
between the master element and each slave element has a rea-
to . Suppose and are
sonably high SNR such that the variance of the ML estimator
statistically independent with identical distribution, i.e.,
of the received frequency (phase) equals or closely approaches
its corresponding Cramer-Rao lower bound (CRLB). and . Let the esti-
Without loss of generality, let the signal received at sensor mation errors for the received frequency and phase be defined
and sent by be given by as

(18) (22)

and let the sampling interval be . We assume is com- and


plex additive white Gaussian noise (AWGN) with zero mean (23)
and variance . We assume samples of noisy discrete-time
5The authors of [11] also assumed that the mod- phase ambiguity in each
observations are available, and are given by
phase measurement can be removed. They assumed that the number of multi-
ples of removed by the phase measurement process can be determined sep-
arately, either by prior knowledge of the approximate element positions or by
(19) other means.

(20)

(21)
5544 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 11, NOVEMBER 2011

respectively. Then, and can be explicitly calculated we denote the aggregate phase error after implementing the syn-
through [27] [see (24) and (25) at the bottom of the page]. Thus, chronization algorithm as
it can be shown that conditioning on the magnitudes
of the received signal6, the estimation errors and are
both Gaussian distributed with zero mean. Moreover, using the
approximation , it is easy to verify that In the following theorem, we formally state some synchroniza-
the ML estimators given in (20) and (21) are both unbiased at tion related results.
high SNR, i.e., and . The quality Theorem 1: After running this master-slave closed-loop
of the frequency and phase estimates can be further evaluated in phase synchronization algorithm, all the slave elements will
terms of their variances through the CRLB bound. The results synchronize their phase, in expectation, to the master node’s
are given as follows [27]: phase , i.e.

(27)

The aggregate phase error is given by

(26)

where is Gaussian distributed with zero mean.


Equation (26) indicates that and are statistically corre-
Proof: See Appendix I.
lated. This is because they are generated at the same radar ele-
Due to the statistical independence among , , , and
ment and from the same noisy observation (18). Note that,
, except between and , the variance of can be
however, observations obtained in different time slots at a par-
calculated through
ticular sensor, or at different sensors, will involve independent
noise realizations; we term these observations loosely as inde-
pendent observations. It is then reasonable to conjecture that es-
timates (frequency and/or phase) computed from these indepen-
dent observations are statistically independent from each other. where values of these different variances, including
Additionally, it is also observed that the CRLB bounds given , can be obtained from (26), by setting the
subscripts and other parameters including and to ap-
in(26) depend on the SNR (i.e., ) and the number of inde-
propriate values. To give an example, let us consider a special
pendent samples (i.e., ).
case where and are the same for the ML estimators at
Next, we define the error in estimating of (17) as
all these radar elements, and the SNR value, denoted by ,
, where can be computed through (25), provided that
is identical for both forward and reverse channels between the
samples of the received signal are given. For notational con-
master sensor and slave element , i.e., ,
formity, we define , thus . Finally,
. For this case, we can write explicitly as
6Note that the magnitudes are available at . shown in the equation at the bottom of the page.

(24)

(25)
YANG AND BLUM: PHASE SYNCHRONIZATION FOR COHERENT MIMO RADAR 5545

IV. ROUND-TRIP PHASE SYNCHRONIZATION


The master-slave closed-loop method described in the pre-
vious section requires a centralized entity to act as a coordinator
or controller. This method, however, does not seem very suitable
to cases where such a master entity does not exist. In this section,
we propose a round-trip phase synchronization algorithm. This
algorithm is inspired by the all-node-based method proposed
in [19] to globally synchronize the clocks in a sensor network,
and the time-slotted round-trip approach that was developed
in [20] and further improved or modified in [21] and [22] to
attain carrier synchronization for distributed beamforming in
multiuser wireless communication systems. These approaches
do not seem to be applicable to our problem of interest directly,
although a common feature shared by these approaches, that Fig. 2. Illustration of the round-trip phase synchronization method.
when in operation, a reference signal or message traverses in
loops through all the sensors, remains essential.
Likewise, the proposed round-trip phase synchronization 5) Upon receiving the signal transmitted from its upstream
scheme for coherent MIMO radar also employs an unmodulated neighbor, each sensor generates its local phase and fre-
beacon signal to travel through all these radar elements in a quency estimates from its observation, and calculates its
round-trip manner. Each node will record the phase information final phase based on its two-time phase estimates. In time
twice in order to compute the final phase. The equivalence of slot , relays the received signal to the next
the cumulative phase shift at each sensor, a quantity that is node in the cycle, i.e., .
directly related to the total propagation delay for the round-trip 6) Relay of the beacon signal ends after finishes its receipt
circuit, is the fundamental idea behind this proposed technique. of the signal from . generates a phase estimate from
This proposed technique, together with its analysis, employs its noisy observation, and then calculates its final phase.
pretty much the same assumptions as those in the master-slave
closed-loop method. For example, this scheme also assumes A. Algorithm Characterization and Timing Analysis
no a priori establishment of both time and frequency synchro-
In the following, we provide a detailed mathematical charac-
nization across all the radar elements, and assumes the channel
terization and timing analysis of this proposed synchronization
reciprocity. However, specific to this round-trip method, we
scheme. Again, we focus on the single-path and time-invariant
further assume that when one sensor transmits, the signal is only
channels for simplicity. We assume time slot starts at time
properly received by its intended receiver, i.e., its downstream
. Also, we assume all the time slots have the same duration
neighbor. This can be attained in practice, for example, by using
. should be long enough to ensure the proper reception and
directional transmissions or appropriate power control schemes
processing of the beacon signal. It, however, should also be rel-
at the transmitter (assuming each radar element loosely knows
atively short to reduce the synchronization overhead.
its and its immediate neighbors’ locations). In what follows, we
In time slot , node transmits a unit-amplitude sinu-
summarize this round-trip phase synchronization algorithm.
soidal beacon signal of duration to node . This beacon
1) Find a cycle which passes through all the radar ele-
signal is given by
ments. The radar elements are then labeled in order as
. Let be the initiating element.
(28)
2) In time slot , an unmodulated sinusoidal reference
signal is transmitted from to its downstream neighbor The signal received by can be expressed as
. This reference signal is further passed along the cycle
in a clockwise manner, as illustrated by Fig. 2. (29)
3) Upon receiving the reference signal, each radar element
generates its local phase and frequency estimates from where and . uses this
the noisy observation. In time slot , noisy observation to compute estimates of the received
which starts immediately after the conclusion of the re- frequency and phase. Such estimates are denoted as and
ceived signal, relays the received reference signal to the , respectively. It is worth noting that is the estimate of
next node in the cycle. will be the downstream neighbor the received phase at time when the signal is firstly
of . captured.
4) After the initiating node receives the signal from , Upon the end of receipt of the signal from node , i.e., at
it generates a phase estimate from its noisy observation. time , time slot begins immediately, and node starts
Time slot starts right after the conclusion of the re- to transmit. The signal that node transmits to its downstream
ceived signal, wherein transmits the original unmodu- neighbor is generated by using the phase and frequency es-
lated sinusoidal reference signal as in step 1) to sensor . timates and , and is given by
This signal is passed along the cycle in the counterclock-
wise direction, as shown in Fig. 2. (30)
5546 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 11, NOVEMBER 2011

We would like to point out that a periodic extension of the re- The signal received by can be expressed as
ceived signal can also be used as the transmit signal, as
was adopted in [21]. In particular, such a periodic extension can
be generated by using the frequency estimate and the phase
estimate to extrapolate the phase of at time . The where and . From
extrapolated phase is then given by . This extrap- , computes estimates of the received frequency and
olation approach, though, viable as it appears in practice, may phase, denoted as and , respectively. Then, can
result in degenerated accuracy in phase synchronization, as the compute its final phase through
error in estimating the carrier frequency will propagate and
(35)
accumulate over the course of extrapolation. For this reason, we
prefer to choose the transmit signal as per (30). In time slot , relays the beacon signal to .
Relay of the beacon signal is repeated for nodes The transmit signal can be written as
in time slot following the order of
. A general expression for the
signal transmitted from , is given by
The receive signal at is given by
(31)

where denotes the time when starts to relay the beacon


signal to , and is given by
where and
. This relay process continues along the cycle in the order of
(32)
until receives the signal
from . Thus, we can generalize the signal transmitted from
and are estimates of the received frequency and
, in time slot as
phase at node , respectively, upon receiving the signal trans-
mitted by sensor . Similarly, a general expression for the
received signal at , can be written as
where . and are
estimates of the frequency and phase, respectively, for the signal
received at and transmitted by . The signal received by
(33)
, can be written as
where follows from (32), and denotes the received
phase, and is given by

(34)
where we have
Additionally, the phase and frequency estimates obtained from (36)
(33) are denoted as and , respectively.
In time slot , transmits the following signal to Thus, can compute their final phase simply
through

(37)
The signal received at sensor is given by The whole signal relay procedure ends at time when fin-
ishes its reception of the beacon signal from . then gener-
ates a phase estimate from its received signal , and
calculates its final phase through

where and (38)


. uses this noisy signal to generate an
estimate of the received phase, which is denoted as .
The counterclockwise cycle starts in time slot when B. Phase Synchronization Error Analysis
node concludes its receipt of the signal from and starts Now let us examine the final carrier phase at each radar ele-
to relay a fresh sinusoidal signal as in (28) back to . The ment as well as their accuracy. Estimates of the phase and fre-
transmit signal is explicitly written as quency at each sensor are inevitably subject to errors, and these
errors are closely related to the particular type of the estimators
employed. Here again, we adopt the ML estimators as described
YANG AND BLUM: PHASE SYNCHRONIZATION FOR COHERENT MIMO RADAR 5547

in Section III. Thus, the frequency and phase estimates, and where and are the cumulative phase errors induced
, can be calculated from (20) and (21), respectively. The es- by the inaccurate estimation of the carrier phase and frequency,
timation errors and , as defined, respectively, in (22) and respectively, and are both Gaussian distributed with zero mean.
(23), can be shown to be both zero-mean Gaussian distributed, Proof: See Appendix II.
with variance lower bounded by the CRLB bound given in (26). A close review of (42) reveals that the mean of the final phase
In accordance with the definition of in (22), we define , , is a function of the propagation delay between each
the notations pair of neighboring radar elements. It is thus essential to en-
sure that the channels remain invariant during the period when
phase synchronization is implemented. We would also like to
point out that, similar to the master-slave method, this round-trip
phase synchronization algorithm is also applicable to multipath
time-invariant channels. In that case, each propagation delay
in (42) needs to be properly adjusted so as to accommo-
Thus, in the clockwise cycle, the true frequency of the signal date the delay spread of the multipath channels.
propagating from to is written as When examining the detailed expressions for in
Appendix II, we find that , is a sum of
phase estimation error terms, but is equal to the sum of
phase estimation error terms divided by two. Suppose the
(39) phase estimator employed at each radar element is identical,
then the variance of is the same for ,
which is twice the variance of . Whereas, we find that the
Likewise, in the counterclockwise cycle, the true carrier fre- expression for for each radar element is more complicated
quency of the signal traversing from and received by than the corresponding . This is due to the fact that
is given by involves the propagation and accumulation of the related in-
dividual frequency estimation errors in the serial transmission
process. Finally, the variances of and can be easily
computed by using the covariance results, as given in (26), for
each constituent individual’s estimation error.
(40) Remark 2: Note that in order to implement this round-trip syn-
From (39) and(40), we can see that the frequency estimation chronization algorithm, cycles which can traverse all the radar
errors in upstream radar elements will propagate and build up sensors need to be identified first. This problem can be formu-
as the beacon signal proceeds to downstream radar elements. lated into a combinatorial optimization problem, which may be
Clearly, the sequential nature of the transmissions in this round- deemed as a special case of the minimum cost-to-time ratio cycle
trip algorithm accounts for the accumulated deviation of the true problem [29, Ch. 3, Sec. 13]. The minimum cost-to-time ratio
transmit frequency from the initial reference frequency . Sim- cycle problem, first formulated by Dantzig et al. in the 1960’s
ilar phenomenon also occurs to the carrier phase at each radar [30], has been considered to be a well-solved combinatorial
element. However, unlike the carrier frequency whose deviation problem, as over the years various polynomial-time algorithms
from the initial reference ( ) is only due to the cumulative fre- have been developed, for example, see [31] for a description and
quency estimation errors, the final phase at each radar element experimental analysis of some typical algorithms. Solutions to
is subject to the accumulation of both phase and frequency es- this combinatorial problem usually do not contain any repeated
timation errors. To this end, we summarize these results in the nodes. However, for our problem of interest, a particular sensor
following theorem. can appear more than once. In other words, loops within a cycle
Theorem 2: After running the round-trip phase synchroniza- are allowed for our application. This, however, will not affect the
tion algorithm, all the radar elements will reach a consensus of efficacy of this round-trip synchronization algorithm, as long as
the carrier phase (in the statistical mean sense). More specifi- the repeated node chooses the correct phase estimates to compute
cally, we have its final phase. For example, a repeated sensor in the cycle can
choose its first and last phase estimates to compute its final phase.
(41)

where is a constant and is given by V. BROADCAST CONSENSUS BASED PHASE SYNCHRONIZATION


In the last two sections, we have introduced and discussed the
master-slave closed-loop and the round-trip phase synchroniza-
(42)
tion algorithms. Note that both methods are not distributed in
nature. This can be easily verified for the master-slave approach,
With a little abuse of notation, let the resulting phase synchro- wherein a centralized entity is required. This also holds true for
nization error at each element be defined as . the round-trip method, since it requires all the radar elements
Then, can be expressed, in a general form, as to participate in the synchronization process at the same time
and in a sequential manner. For the same reason, the round-trip
method does not scale well either. Furthermore, the initiating
5548 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 11, NOVEMBER 2011

node (or the master node in the first method) in both approaches Note that the absence of a common time reference implies
may fail unexpectedly, making it difficult to complete the re- that none of these sensors know the true time, and as a result,
maining synchronization procedures. Thus, both methods are they do not know the “true” initial phase of their local oscillator.
not very robust against node failures. Thus, when a sensor receives a broadcast reference signal for the
As far as robustness and scalability goes, distributed algo- first time, we choose to update its phase using only the estimate
rithms have enormous advantages, and are thus more desirable of the received phase.
for phase synchronization in coherent MIMO radar systems.
The essence of these algorithms is that one or more sensors A. Algorithm Characterization
can communicate information with its immediate neighbors in a Let an undirected and connected graph consti-
round, and the computation is distributed over all the sensors in- tuted by the radar elements be given, where is the node set
volved. In particular, when information averaging is concerned, and is the edge set. The edge formation is related to sensor
these distributed algorithms are usually referred to as distributed communication connectivity. If two radar elements are within
consensus algorithms or gossip algorithms, e.g., [32]–[40]. Dis- the transmission range of each other, an edge is then said to be
tributed consensus algorithms have been applied to a variety of formed to connect these two nodes. Note that we do not allow
network-related scenarios. For example, they have been consid- self loops in the graph. The neighbors of are denoted by
ered in [39], [40] for clock or time synchronization in wireless . Let be the adjacency
sensor networks. matrix of , where for , if , and
Among the various distributed consensus algorithms, of par- otherwise. Further, we define a diagonal
ticular note is the broadcast-based consensus algorithm, which matrix with each diagonal entry given by .
has been studied in [36] and [37] without considering the sto- Thus, the graph Laplacian of is given by . We use
chastic disturbances, and lately in [38] by taking nonzero-mean to denote the column vector in with all entries equal
stochastic perturbations into account. In this section, we present to 0 except a 1 in position , and denote as the th largest
a broadcast consensus based phase synchronization algorithm eigenvalue of a symmetric matrix.
for coherent MIMO radar systems, which is a direct extension of As a priori time synchronization is not assumed, there is no
the work given in [38]. Similar to the aforedescribed two phase common time scale established among all these sensors. Each
synchronization approaches, this algorithm assumes no a priori radar element then operates according to its own local clock
time synchronization. For the ease of exposition, we assume the which ticks independently according to a Possion process with
carrier frequency synchronization is already established, while rate . For each element, the intertick times (the time between
noting that this assumption can be also relaxed. Thus, when a two consecutive clock ticks) are exponentially distributed with
sensor receives beacon signals from its neighbors, it only needs parameter , independent over time and across sensors. This is
to estimate the received phase. equivalent to a single clock which ticks according to a rate-
In this algorithm, each element broadcasts a reference signal Possion process and whose intertick times are i.i.d. exponen-
containing its local phase information to all its immediate neigh- tially distributed with parameter [41]. Clearly, the proba-
bors. Its neighbors then average their local phase with the re- bility that one particular sensor’s clock ticks at , given a clock
ceived phase. Thus, no global operations or information ex- tick does occur at , is equal to the probability that any other
changes are required in the process. As an inherent character- sensor’s clock ticks at , which is equal to in this case. On
istic, this algorithm exploits the broadcast nature of the wireless average, there are clock ticks per unit of absolute time. Since
communication environment, and obviates the need for sophis- phase values at these radar elements are updated in a discrete
ticated underlying media access control mechanisms. Specific manner only at each time instant , for convenience, we use
to this problem, our analysis will examine in detail how the the number of clock ticks (i.e., ) as a measure of time, instead
propagation delays among neighboring sensors and the phase of the absolute time (i.e., ). This amounts to discretizing the
errors induced by inaccurate local estimation affect the phase absolute time by denoting the interval as the th time
synchronization accuracy. To this end, we briefly outline this slot. However, we would like to note that in practice, to guar-
phase synchronization algorithm, whose mathematical charac- antee a high probability of which only one clock ticks during
terization and analysis will be furnished in the sequel. each time slot, caution should be executed when selecting and
1) Sensor elects to broadcast an unmodulated reference adjusting the unit of the absolute time.
signal. Let denote the phase
2) Each neighbor of , upon receiving the broadcast refer- values of these radar elements at the end of time slot . Sup-
ence signal, generates an estimate of the received phase. pose sensor ’s clock ticks at the beginning of the th
If it is the first time for to receive a broadcast refer- time slot. Sensor then broadcasts a reference signal to all its
ence signal, it updates its phase with a weighted value of neighboring sensors, which is similar to (14) but contains its
the phase estimate; otherwise, it updates its phase with up-to-date phase information. The phase value at each sensor is
a weighted average of its current phase and the phase updated according to the following equations:
estimate.
(43)
3) The phase of each of the remaining sensors, including ,
remains unchanged. (44)
4) This procedure continues, with each radar element having Here is usually referred to as the mixing parameter.
the same probability to broadcast in every round. denotes the phase estimate of the signal transmitted by
YANG AND BLUM: PHASE SYNCHRONIZATION FOR COHERENT MIMO RADAR 5549

sensor and received at sensor . From Section III, it follows The expectation of can be computed by
that

Thus, we can rewrite (43) into


Using the results stated in Lemma 1, it can be shown that the
sequence of will converge in expected value to the vector
of averages for every initial condition [33], [34] . In
(45)
other words, we have
Of note is that is zero-mean Gaussian distributed, and is
independent for different values of , , and .
Then using matrix-vector notation, we can reformulate (44)
and (45) all together more compactly into Additionally, this convergence is an almost sure convergence
[36], [37] .
(46) However, in the presence of the nonzero-mean perturbation
vector , will no longer converge to the vector
where is a random matrix, which, with prob- of averages . Fortunately, for our problem of interest, we
ability and assuming ’s clock ticks, takes the following are more concerned about whether or how well these different
value radar elements can synchronize to a same phase offset. This
same phase offset does not need to be , although it cannot
be either in this case because . Theoretically, for this
phase synchronization application, we have a lot of choices
when it comes to the specific value of this same phase offset
in (46) is the phase perturbation vector, whose to which each radar element needs to synchronize. Following
th element is given by the work in [35]–[38], here we adopt the average as the
common phase to be synchronized to. As a result, we define the
,
phase synchronization error as the deviation of the components
.
of from their average. This error, if denoted in component
It is easy to verify that , i.e., is a right eigenvector form, is
of corresponding to eigenvalue 1. Note that, however,
is not a left eigenvector of , i.e., .
We define the expectation of the random matrix as
. We also define a matrix . In the following Or the phase error vector is given by
lemma, we state some results about without proof, but note
that detailed information can be found in [38]. (48)
Lemma 1: The mean matrix is given by
It is easy to verify that the deviation vector evolves ac-
(47) cording to

which satisfies and . Further, for any (49)


, is a symmetric and nonnegative matrix, and satisfies
, where denotes the spectral radius of a In the following, we provide some synchronization result which
matrix [42]. is related to the asymptotic behavior of the expectation of .
Next, we define the expectation of the random perturbation Theorem 3: As this broadcast consensus based phase syn-
vector as . It is easy to verify that chronization algorithm iterates, the synchronization error at
is a column vector with its th element given by each element from their average , will converge in
expectation to a finite (steady-state) value, i.e.

(50)

Proof: See Appendix III.


B. Phase Synchronization Error Analysis
Equation (50) indicates that is a function of the
In the absence of the phase perturbation vector , this Laplacian matrix of the network, the mixing parameter ,
phase synchronization algorithm is governed by and the mean of the phase perturbation vector . Since is
connected, cannot be zero. Thus, Theorem 3 implies that this
algorithm, in general, cannot achieve average consensus, i.e.,
5550 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 11, NOVEMBER 2011

we have . In other words, generally speaking, this and is lower bounded by


phase synchronization algorithm will lead to
(54)
(51) where in (53), is a parameter defined as

This is in stark contrast to the previous two phase synchroniza-


tion algorithms which can synchronize the phase of each ele-
ment to the same value in expectation, as evidenced by (27) and satisfies ; is given by the equation at the bottom
and(41) in Theorem 1 and Theorem 2, respectively. of the page, and .
However, it should be noted that when , which is Due to space limitations, a proof of Theorem 4 is not pro-
equivalent to vided here, but we direct readers to [38] for more information.
Of note is that the upper bound given by(53) is closely related to
(52) , which is the second smallest eigenvalue of and is
also referred to as the algebraic connectivity of [43]. For the
special case of , the lower bound given by (54) becomes
we have , thus the average consensus is achieved, zero, and the upper bound(53) reduces to
and all the sensors will eventually synchronize in expectation to
the same phase value. The condition stipulated in (52) indicates
that the network should be well structured, such that the sum
of delay (or distance) from one sensor to all its neighbors is the Remark 3: We would like to point out that the bounds given
same as others. Those types of networks are referred to as equal in Theorem 4 are very informative for practical implementa-
delay networks in [15] or time-delay balanced networks in [40]. tion, as they reveal the relative range over which the phase at
Examples of this type of networks include, but are not limited each sensor deviates from their average. More specifically, the
to, ring networks with equal distance and hypercube networks lower bound indicates that the broadcast consensus based phase
with equal distance degree [40]. synchronization algorithm, in general, has a nonreducible error
To further examine the performance of this phase synchro- floor. The upper bound, on the other hand, tells the value of
nization algorithm, we consider the total mean-square phase the maximum possible phase synchronization error. Therefore,
synchronization error, or the mean-square phase error for short. for a given set of parameters, one can easily obtain an estimate
Let us define , and denote the of the resulting phase synchronization error through Theorem
mean-square phase error as . Then, we have 4. If that error is tolerable, one can certainly choose to employ
this phase synchronization algorithm; otherwise, one can seek
other alternatives. Of further note is that since this algorithm
is iterative in nature, the repeated information exchange among
radar elements can result in considerable communication over-
head which may merit special attention when implementing this
algorithm in practice.

F. DISCUSSION
This indicates that the mean-square deviation is pro-
portional to the average pairwise expected deviation among the A. Algorithm Comparison
node states, and can thus be interpreted as a measure of how far
Among these three approaches, the master-slave closed-loop
the components of are from consensus. In the following
method is the simplest one. It, however, can lead to the best
theorem, we provide asymptotic upper and lower bounds on
phase synchronization accuracy. This is because there are only
.
two (the least number) nodes which participate in each round of
Theorem 4: The limit of the mean-square phase error
the phase synchronization process, i.e., the master node and one
is upper bounded by slave node. Thus, the propagating accumulation of each indi-
vidual’s phase and frequency estimation errors is very limited,
and is less of a concern. Moreover, in this method, the phase
offset induced by the propagation channel is explicitly com-
(53)
pensated via the closed-loop operation, which further improves
YANG AND BLUM: PHASE SYNCHRONIZATION FOR COHERENT MIMO RADAR 5551

the phase synchronization accuracy. Therefore, this approach the channel propagation delay, which may result from, for ex-
is particularly useful when a very precise phase synchroniza- ample, the ambiguity in the location of certain radar elements or
tion is required. It is generally applicable to situations where the mobile scatters and other propagation anomalies, can lead
a relatively powerful centralized entity exists or can be identi- to intolerable phase variations across some or all the radar el-
fied. It can also be repeatedly used in systems having tree-like ements. This is particularly true when high frequency carriers
topologies or hierarchical structures. To complete the whole are chosen as the reference signals. One way to alleviate this
phase synchronization process, a total of transmissions unfavorable effect is to use relatively low carrier frequency. If
are needed, including the broadcast transmission made by the the carrier frequency is low enough, the location uncertainty of
master node. the MIMO radar elements will not play an important role. An-
Unlike the master-slave closed-loop method, both the round- other way is to recalibrate these sensors periodically to track the
trip and the broadcast consensus based algorithms do not re- channel changes. This may induce some extra overhead, but the
quire a powerful master entity. However, the round-trip scheme accompanying benefits are in general well worth it. Besides, the
does need a global initiator to start the phase synchronization variations of the propagation channels among all these MIMO
process. In theory, the global initiator can be any radar sensor radar elements, though dependent on the specific operating en-
in the system. This round-trip method is able to synchronize all vironment, are usually very slow; it takes a relatively long time
the radar sensors to the same mean carrier phase, regardless of to reach the point where the physical changes in the medium are
the topology of these sensors as long as a cycle traversing all so pronounced that the resulting phase mismatch seriously dete-
these sensors can be identified. But since sensors in the system riorates the performance of the coherent processing, especially
participate in the phase synchronization process all together in a for applications such as target localization. Therefore, very fre-
cooperative fashion, this round-trip method may suffer from the quent phase resynchronization may not be needed.
propagating accumulation of each individual’s phase and fre- Another practical issue that merits attention is the phase dy-
quency estimation errors, as was illustrated in Section IV. As namics or phase instability [45] in the coherent MIMO radar
a result, this method may only be suitable to coherent MIMO system. In the previous sections, for simplicity we ignore the
radar systems with a limited number of sensors. In other words, phase shift due to the oscillator phase noise at each sensor by as-
it does not scale very well. However, of particular note is, that if suming such a shift is very small and negligible. However, note
the clustering infrastructure is introduced as in [44], this round- that the oscillator phase noise is always present, and is an ir-
trip method, when coupled with the master-slave closed-loop reducible error even in the absence of estimator errors. Without
approach, is also applicable to coherent MIMO radar systems loss of generality, let us take the reference signal defined by (14)
with a relatively large number of radar sensors. This round-trip as one example, where is a fixed phase reference. Consid-
method needs a total of transmissions to complete the phase ering a practical oscillator and assuming negligible amplitude
synchronization process. noise, the reference signal in (14) is more generally expressed
The broadcast consensus based method, as we mentioned ear- as
lier, is distributed and localized in nature. It needs neither a cen-
tralized entity nor a particular global initiator. It is scalable and
applicable to coherent MIMO radar systems composed of a large
number of densely scattered sensors. It is also robust to sensor where is a function of time. Oscillator phase noise can
failures. But unfortunately, the functionality of this approach is cause to drift from the desired values in a random manner,
highly dependent on, and varies from topology to topology of and can degrade the accuracy of phase synchronization. A quan-
the radar sensors. For arbitrary topologies, this approach may titative understanding of the impact of phase noise in oscilla-
be only useful when a coarse phase synchronization is required. tors requires a rigorous characterization of the phase noise’s be-
This is because Theorem 3 indicates that this algorithm, in gen- havior. In recent years, extensive research has been done in this
eral, cannot synchronize all the sensors to the same mean carrier regard; well-known references include, among others, [25] and
phase. However, for well structured sensor placements, such as [26]. In particular, [26] presents a theory for phase noise that is
those which satisfy (52), this method is able to synchronize all valid for any oscillator, regardless of the operating mechanisms.
the sensors in expectation to (or very close to) the same phase According to this theory, for a single noise source, can be
value, and thus may be used for the purpose of fine phase syn- characterized by [26]
chronization. Since this algorithm runs in an iterative manner,
it generally requires more transmissions than the other two ap- (55)
proaches to complete the whole synchronization process. Thus,
it has to be implemented at a relaxed time constraint as opposed where is an external source, noise, or otherwise; is a
to a stringent one. periodic function dependent on the oscillator’s topology, which
models the way a particular input source affects the oscillator
B. Some Practical Issues phase. Note that (55) is derived using the nonlinear perturbation
As can be seen, in each of the proposed phase synchroniza- technique. When is a stationary white or modulated-white
tion algorithms, the propagation delay for the channel between noise source, such as shot and thermal noise, it can be shown
any pair of MIMO radar elements that are involved in the ref- that becomes, asymptotically with , a Gaussian random
erence signal exchange, can be critical in determining the re- variable with a constant mean (i.e., the intended phase) and a
sulting phase synchronization accuracy. A small uncertainty in variance that is linearly increasing with time, i.e., ,
5552 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 11, NOVEMBER 2011

Fig. 3. Two example topologies for the simulation study. (a) A grid topology of 36 nodes. (b) A ring topology of 25 nodes.

where is a scalar constant [26]. Suppose the period of is variables7, and studied the impact of the phase synchronization
, can be expressed as [26] errors on the target localization accuracy in coherent MIMO
radar, both analytically and numerically. [5] presents some
interesting and elegant results, which dovetail our work on
phase synchronization very well (or the other way around).
We thus do not repeat or duplicate these results in this paper
Clearly, the parameter is also dependent on the specific os-
to illustrate how the phase synchronization errors affect the
cillator circuit. The methods for computing and its values for
performance of coherent MIMO radar. Interested readers are
some practical electronic oscillators can be found in [26]. Once
referred to [4]–[7] for further information. Also, as the first two
the parameter is identified, by incorporating at each
synchronization approaches can be implemented and verified
radar element into the analysis of the final phase synchroniza-
with relative ease, for brevity, in this section we only focus
tion error, one should be able to easily assess the phase noise’s
on the broadcast consensus based algorithm and study its
impact on the accuracy of each proposed phase synchronization
convergence behaviors.
algorithm. To counter the effect of phase noise in oscillators, pe-
Since the broadcast consensus based algorithm is applicable
riodic phase synchronization remains indispensable. How often
to coherent MIMO radar systems composed of a large number
the phase synchronization procedure needs to be repeated, how-
of densely scattered sensors, as we discussed in Section VI, we
ever, is dependent on, among other factors, the aggregate impact
consider here two example topologies8. The first one, as shown
of the phase noise across all the local oscillators on the phase
in Fig. 3(a), is a grid topology of 36 sensors, and the distance be-
synchronization accuracy, as well as the phase accuracy or sta-
tween any two neighboring nodes along the same horizontal or
bility requirements of the coherent MIMO radar system and the
vertical line is . The second one, as illustrated by Fig. 3(b),
specific application. It is indicated in [12, p. 261] that the phase
contains 25 sensors which are equally distributed on a circle
accuracy or stability requirements for coherent processing by a
with a radius of . We set the reliable transmit-receive range
monostatic or bistatic radar can range from less than one de-
gree to many tens of degrees of RF phase over a coherent pro- to . Thus, the grid topology in Fig. 3(a) does not satisfy (52),
cessing interval, depending on the type and duration of coherent while the ring topology in Fig. 3(b) does. In our simulations,
processing. We conjecture that similar phase accuracy or sta- the frequency of the reference signal is set to
bility requirements may hold as well for coherent processing by rad/second. We assume the SNR value of the received reference
a MIMO radar system. signal at each sensor and during each time slot is the same, and
is equal to . The sample number is set
to 10. Thus, through (26), the CRLB bound of the phase esti-
VII. SIMULATION RESULTS
mate can be computed, which equals 0.0172. We set the mixing
In this section, we provide some simulation results to parameter to 0.5. For each example topology, we randomly
demonstrate the efficacy of the phase synchronization algo- generate a set of “true” initial phase values at these radar sen-
rithms. While it is desirable to study the relation between the sors, which are over the radian range . We conduct 2000
performance of the phase synchronization algorithms and the independent trials under the same set of initial values, and then
performance of coherent MIMO radar for applications such as 7This seems to be a valid assumption according to our analysis of the phase
target localization or detection, this line of research has been synchronization errors.
already conducted in [4]–[7]. For example, in [5], the authors 8Note these two example topologies are considered here for the purpose of
modeled the phase synchronization errors as zero-mean random illustration only; they might not necessarily be applicable in practice.
YANG AND BLUM: PHASE SYNCHRONIZATION FOR COHERENT MIMO RADAR 5553

Fig. 4. Simulation results for the grid topology in Fig. 3(a). (a) Ensemble average of versus for , , and . (b) Ensemble average of versus .

Fig. 5. Simulation results for the ring topology in Fig. 3(b). (a) Ensemble average of versus for . (b) Ensemble average of versus .

compute the ensemble average of some parameters of interest acceptable under the application’s performance requirements,
as well as their corresponding theoretical results. another phase synchronization algorithm which leads to a
In Fig. 4(a), we plot the ensemble average of the phase better accuracy may be used instead. In Fig. 4(b), we plot the
synchronization error for example sensors, , , ensemble average of the squared phase error . We also plot
and . These sensors have 3, 5, and 8 immediate neighbors, the theoretical value of the upper and lower bounds on the limit
respectively, and they actually represent all the three types of of the mean-square phase error , which are computed
sensors in this grid topology. For each example sensor, we also through (53) and (54), respectively. It is observed that after the
compute the corresponding theoretical limit value of the mean algorithm converges, the mean-square phase error does fall in
phase synchronization error through (50). Their theoretical between the upper and lower bounds. This corroborates our
values are different from each other under this simulation theoretical analysis.
setup, and are plotted in Fig. 4(a) for comparison purpose. We Similarly, for the example ring topology illustrated in
can see that as the algorithm proceeds, the simulated value of Fig. 3(b), we plot both simulation and theoretical results for
the mean phase synchronization error at each sensor closely and in Fig. 5(a) and (b), respectively. In
approaches their respective theoretical limit value. But note Fig. 5(a), we pick up as the example sensors.
that since the mean phase synchronization error varies from Since this topology satisfies (52), we have . This means
sensor to sensor and not all of them approach 0, the applica- that the theoretical limit value of for each sensor in
bility of this algorithm in practice will be highly dependent on this topology is the same, and is equal to 0. Also, the lower
whether those nonzero mean phase synchronization errors are bound on the limit of reduces to 0. Again, as shown
acceptable or not. If those phase synchronization errors are not in both Figs. 5(a) and 5(b), the theoretical values exhibit a good
5554 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 11, NOVEMBER 2011

match with the experiment results. A comparison of the results (61)


plotted in Figs. 4 and 5 clearly indicates that the functionality
of this broadcast consensus based algorithm is highly related to where in (57), we apply (17); in (58), we use (16). Since ,
the topology of the radar sensors. , , and are all zero-mean Gaussian distributed,
is also Gaussian distributed and satisfies . Thus,
VIII. CONCLUSION we have . This then completes the proof.
Practical realization of the coherent processing in widely
separated MIMO radar systems requires the development of APPENDIX II
implementable techniques to ensure a common notion of phase PROOF OF THEOREM 2
among all the distributed radar elements. In this work, we pro- Proof: By plugging9 (22) and (23) into (34) and recursively
posed and studied three effective approaches to achieve phase applying (34), the carrier phase of the signal received by and
synchronization in coherent MIMO radar systems. The first sent by in the clockwise cycle can be expressed as
one is the master-slave closed-loop approach, which employs
a master-slave architecture. This method is very simple, and
is suitable for fine phase synchronization. But it may be easily
plagued by the master sensor failure. The second one is the
round-trip algorithm, which employs an unmodulated beacon
signal to travel through all the radar elements in a round-trip (62)
manner. This method is applicable to arbitrary topologies
of the radar sensors as long as a cycle traversing all these In a similar manner, we can also obtain an expression for the
sensors can be identified, but its performance is susceptible to carrier phase at in the counterclockwise cycle, which
the propagating accumulation of each individual’s phase and is given by
frequency estimation errors. The third one is the broadcast
consensus based algorithm, which is scalable, and is distributed
and localized in nature. But this algorithm may take a number
of transmissions to reach the convergence, for it runs in an
iterative fashion. And its performance is highly dependent on
the underlying topology. Distinct as they are, these algorithms
are all time asynchronous, and in particular, the master-slave
closed-loop algorithm and the round-trip algorithm also require
no a priori frequency synchronization. These algorithms should
cater to different needs arising as a result of different number
of MIMO radar elements, different sensor placements, as well (63)
as different operational preferences. Furthermore, using the
same analytical framework, we mathematically characterized Thus, the final phase at , can be expressed
each phase synchronization method, derived and analyzed as
some relevant figures of merit such as the aggregate phase
(64)
synchronization error. We discussed some issues that may
arise in practice when implementing these algorithms, and also where is a constant phase as given in (42). denotes
provided simulation results that are focused on the broadcast the phase error resulting from the inaccurate estimation of the
consensus based algorithm. These simulation results indicate a carrier phase, and is written by
good match with our theoretical analysis.

APPENDIX I
PROOF OF THEOREM 1
Proof: Given the definitions of and all other parame- denotes the phase error due to the inaccurate estimation of
ters, we have that for the carrier frequency, and is written as

(56)
(57)

(58)

(59)

(60) 9Appropriate subscript or parameter initialization is assumed.


YANG AND BLUM: PHASE SYNCHRONIZATION FOR COHERENT MIMO RADAR 5555

Now let us consider two special cases of an , as de-


noted by (38) and(35), respectively. Since
and , we can expand in
(38) into

Since the estimation errors and , as defined, respec-


tively, in (22) and(23), are both zero-mean Gaussian distributed,
it is easy to see that and are both Gaussian distributed
Plugging (39) with , (62) with , and (63) with with zero mean, and thus (41) is obtained. We then complete this
into the above equation, can be further expanded into proof.

APPENDIX III
where we have PROOF OF THEOREM 3
Proof: Repeatedly applying (49), we obtain

Since are i.i.d drawn with equal probability,


we have

Similarly, we can also expressly compute in (35). So the


final phase at can be expressed as

where the last equality follows from the property in Lemma 1


that is a left eigenvector of corresponding to eigenvalue
where we have 1, i.e., or . Since and ,
are statistically independent, we obtain

Therefore, the expectation of the deviation vector can


be computed through
Finally, for those two cases of and that are
not included in (64), it is easy to show that the final phase at
and can be expressed in three parts. For completeness, we
give the final expressions for and directly as follows:
where we define .
Let the eigendecomposition of be given by
, where is a matrix whose columns are the
eigenvectors of , and is a diagonal matrix whose diag-
onal elements are the corresponding eigenvalues. It then follows
that and
5556 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 11, NOVEMBER 2011

According to Lemma 1, we have . Thus, it is easy [18] R. Mudumbai, G. Barriac, and U. Madhow, “On the feasibility of
to see that will converge to 0, as . For the distributed beamforming in wireless networks,” IEEE Trans. Wireless
Commun., vol. 6, no. 4, pp. 1–10, Apr. 2007.
same reason, we have [19] Q. Li and D. Rus, “Global clock synchronization in sensor networks,”
in Proc. IEEE Conf. Comput. Commun., Mar. 2004, pp. 564–574.
[20] D. R. Brown, III, G. Prince, and J. A. McNeill, “A method for carrier
frequency and phase synchronization of two autonomous cooperative
transmitters,” in Proc. IEEE Workshop on Signal Process. Adv. Wire-
less Commun., New York, Jun. 2005, pp. 260–264.
[21] D. R. Brown, III and H. V. Poor, “Time-slotted round-trip carrier
Thus, the steady-state value of can be written as synchronization for distributed beamforming,” IEEE Trans. Signal
Process., vol. 56, no. 11, pp. 5630–5643, Nov. 2008.
[22] Q. Wang and K. Ren, “Time-slotted round-trip carrier synchronization
in large-scale wireless networks,” in Proc. IEEE Int. Conf. Commun.,
May 2008, pp. 5087–5091.
[23] Q. He, R. S. Blum, and A. M. Haimovich, “Non-coherent MIMO radar
for location and velocity estimation: More antennas means better per-
formance,” IEEE Trans. Signal Process., vol. 58, no. 7, pp. 3661–3680,
Jul. 2010.
where the second equality follows from (47) in Lemma 1. [24] M. Skolnik, Introduction to Radar Systems, 3rd ed. New York: Mc-
Graw-Hill, 2001.
[25] A. Hajimiri and T. H. Lee, “A general theory of phase noise in electrical
REFERENCES oscillators,” IEEE J. Solid-State Circuits, vol. 33, no. 2, pp. 179–194,
Feb. 1998.
[1] A. M. Haimovich, R. S. Blum, and L. J. Cimini, Jr, “MIMO radar with [26] A. Demir, A. Mehrotra, and J. Roychowdhury, “Phase noise in oscilla-
widely separated antennas,” IEEE Signal Process. Mag., vol. 25, pp. tors: A unifying theory and numerical methods,” IEEE Trans. Circuits
116–129, Jan. 2008. Syst.: Fund. Theory Appl., vol. 47, no. 5, pp. 655–674, May 2000.
[2] N. H. Lehmann, A. M. Haimovich, R. S. Blum, and L. Cimini, “High [27] H. Fu and P. Y. Kam, “MAP/ML estimation of the frequency and phase
resolution capabilities of MIMO radar,” in Proc. 40th Asilomar Conf. of a single sinusoid in noise,” IEEE Trans. Signal Process., vol. 55, no.
Signals, Syst., Comput., Oct. 2006, pp. 25–30. 3, pp. 834–845, Mar. 2007.
[3] H. Godrich, A. M. Haimovich, and R. S. Blum, “Target localization ac- [28] J. Tribolet, “A new phase unwrapping algorithm,” IEEE Trans. Acoust.,
curacy gain in MIMO radar-based systems,” IEEE Trans. Inf. Theory, Speech, Signal Process., vol. 25, no. 2, pp. 170–177, Apr. 1977.
vol. 56, no. 6, pp. 2783–2803, Jun. 2010. [29] E. L. Lawler, Combinatorial Optimization: Networks and Matroids.
[4] Q. He and R. S. Blum, “Cramer-Rao bound for MIMO radar target New York: Holt, Rinehart and Winston, 1976.
localization with phase errors,” IEEE Signal Process. Lett., vol. 17, [30] G. B. Dantzig, W. O. Blattner, and M. R. Rao, “Finding a cycle in
no. 1, pp. 83–86, Jan. 2010. a graph with minimum cost to time ratio with application to a ship
[5] H. Godrich, A. M. Haimovich, and H. V. Poor, “An analysis of phase routing problem,” in Theory of Graphs, P. Rosenstiehl, Ed. New
synchronization mismatch sensitivity for coherent MIMO radar sys- York: Dunod, Paris, Gordon and Breach, 1967, pp. 77–84.
tems,” in Proc. IEEE Int. Workshop on Computat. Adv. Multi-Sens. [31] A. Dasdan, “Experimental analysis of the fastest optimum cycle ratio
Adaptive Process., Aruba, Dutch Antilles, Dec. 2009, pp. 153–156. and mean algorithms,” ACM Trans. Design Autom. Electron. Syst., vol.
[6] H. Godrich and A. M. Haimovich, “Localization performance of co- 9, no. 4, Oct. 2004.
herent MIMO radar systems subject to phase synchronization errors,” [32] R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and coop-
in Proc. 4th Int. Symp. Commun., Contr. Signal Process., Limassol, eration in networked multi-agent systems,” Proc. IEEE, vol. 95, no. 1,
Cyprus, Mar. 3–5, 2010. pp. 215–233, Jan. 2007.
[7] M. Akçakaya and A. Nehorai, “MIMO radar detection and adaptive [33] L. Xiao and S. Boyd, “Fast linear iterations for distributed averaging,”
design under a phase synchronization mismatch,” IEEE Trans. Signal in Syst. Contr. Lett.. New York: Elsevier, 2004, vol. 53, pp. 65–78.
Process., vol. 58, no. 10, pp. 4994–5005, Oct. 2010. [34] S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah, “Randomized gossip al-
[8] V. V. Shakhguildyan and S. S. Sviridenko, “Phase synchronization gorithms,” IEEE Trans. Inf. Theory, vol. 52, pp. 2508–2530, Jun. 2006.
systems studies,” IEEE Trans. Commun., vol. COM-30, no. 10, pp. [35] L. Xiao, S. Boyd, and S.-J. Kim, “Distributed average consensus with
2260–2263, Oct. 1982. least-mean-square deviation,” J. Parallel Distrib. Comput., vol. 67, no.
[9] B. D. Steinberg, “Phase synchronizing a nonrigid, distributed, transmit- 1, pp. 33–46, 2007.
receive radar antenna array,” IEEE Trans. Aerosp. Electron. Syst., vol. [36] F. Fagnani and S. Zampieri, “Randomized consensus algorithms over
AES-18, no. 5, pp. 609–620, Sep. 1982. large scale networks,” IEEE J. Sel. Areas Commun., vol. 26, no. 4, pp.
[10] E. H. Attia, “Phase synchronizing large antenna arrays using the spatial 634–649, May 2008.
correlation properties of radar clutter,” Ph.D. degree, Univ. Penn., , [37] T. C. Aysal, M. E. Yildiz, A. D. Sarwate, and A. Scaglione, “Broadcast
1985. gossip algorithms for consensus,” IEEE Trans. Signal Process., vol.
[11] E.-A. Lee and C. N. Dorny, “A broadcast reference technique for 57, pp. 2748–2761, Jul. 2009.
self-calibrating of large antenna phased arrays,” IEEE Trans. Antennas [38] Y. Yang and R. S. Blum, “Broadcast-based consensus with nonzero-
Propag., vol. 37, no. 8, pp. 1003–1010, Aug. 1989. mean stochastic perturbations,” IEEE Trans. Inf. Theory, unpublished.
[12] N. J. Willis, Bistatic Radar, 2nd ed. New York: SciTech , 2005. [39] L. Schenato and G. Gamba, “A distributed consensus protocol for clock
[13] P. Lopez-Dekker, J. J. Mallorqui, P. Serra-Morales, and J. synchronization in wireless sensor network,” in Proc. 46th IEEE Conf.
Sanz-Marcos, “Phase synchronization and Doppler centroid esti- Decision Contr., New Orleans, LA, Dec. 2007, pp. 2289–2294.
mation in fixed receiver bistatic SAR systems,” IEEE Trans. Geosci. [40] G. Xiong and S. Kishore, “Analysis of distributed consensus time
Remote Sens., vol. 46, no. 11, pp. 3459–3471, Nov. 2008. synchronization with Gaussian delay over wireless sensor networks,”
[14] B. P. Ng, J. P. Lie, M. H. Er, and A. Feng, “A practical simple geometry EURASIP J. Wireless Commun. Netw., vol. 2009, p. 9, 2009.
and gain/phase calibration technique for antenna array processing,” [41] D. Bertsekas and R. Gallager, Data Networks, 2nd ed. Englewood
IEEE Trans. Antennas Propag., vol. 57, no. 7, pp. 1963–1972, Jul. Cliffs, NJ: Prentice-Hall, 1992.
2009. [42] R. A. Horn and C. R. Johnson, Matrix Analysis. Cambridge, U.K.:
[15] W. C. Lindsey, F. Ghazvinian, W. C. Hagmann, and K. Dessouky, Cambridge Univ. Press, 1985.
“Network synchronization,” Proc. IEEE, vol. 73, no. 10, pp. [43] M. Fiedler, “Algebraic connectivity of graphs,” Czechoslovak Math. J.,
1445–1467, Oct. 1985. vol. 23, no. 2, pp. 298–305, 1973.
[16] B. Sundararaman, U. Buy, and A. D. Kshemkalyani, “Clock synchro- [44] Y. Yang, R. S. Blum, and B. M. Sadler, “A distributed and energy-ef-
nization for wireless sensor networks: A survey,” Ad Hoc Netw., vol. ficient framework for Neyman-Pearson detection of fluctuating signals
3, no. 3, pp. 281–323, 2005. in large-scale sensor networks,” IEEE J. Sel. Areas Commun., vol. 28,
[17] Y. Tu and G. J. Pottie, “Coherent cooperative transmission from mul- no. 7, pp. 1149–1158, Sep. 2010.
tiple adjacent antennas to a distant stationary antenna through AWGN [45] J. Rutman, “Characterization of phase and frequency instabilities in
channels,” in Proc. IEEE Veh. Technol. Conf., AL, 2002, vol. 1, pp. precision frequency sources: fifteen years of progress,” Proc. IEEE,
130–134. vol. 66, no. 9, pp. 1048–1075, Sep. 1978.
YANG AND BLUM: PHASE SYNCHRONIZATION FOR COHERENT MIMO RADAR 5557

Yang Yang received the B.Eng. degree in informa- Rick S. Blum (S’83–M’84–SM’94–F’05) received
tion engineering from Xi’an Jiaotong University, the B.S. degree in electrical engineering from the
Xi’an, China, in 2001, the M.Eng. degree in elec- Pennsylvania State University, University Park, in
trical engineering from the National University of 1984 and the M.S. and Ph.D. degrees in electrical
Singapore in 2004, and the Ph.D. degree in electrical engineering from the University of Pennsylvania in
engineering from Lehigh University, Bethlehem, 1987 and 1991, respectively.
PA, in 2009. He was a member of Technical Staff at General
From 2002 to 2004, he was with the Institute for In- Electric Aerospace, Valley Forge, PA (1984–91) and
focomm Research, Singapore, as a Research Scholar. he graduated from GE’s Advanced Course in Engi-
During summer 2007, he was an intern with Bell Lab- neering. Since 1991, he has been with the Electrical
oratories, Alcatel-Lucent, Crawford Hill, Holmdel, and Computer Engineering Department, Lehigh Uni-
NJ. He is currently a Postdoctoral Research Associate with Lehigh University. versity, Bethlehem, PA, where he is currently a Professor and holds the Robert
His research interests include communications, signal processing, and their ap- W. Wieseman Chaired Research Professorship in Electrical Engineering. He
plications to distributed sensor processing, wireless sensor networks, and en- holds several patents. His research interests include signal processing for com-
ergy/power systems. munications, sensor networking, radar, and sensor processing.
Dr. Blum is on the editorial board for the Journal of Advances in Information
Fusion of the International Society of Information Fusion. He was an Asso-
ciate Editor for the IEEE TRANSACTIONS ON SIGNAL PROCESSING (2000–2002)
and for the IEEE COMMUNICATIONS LETTERS. He has edited special issues for
IEEE TRANSACTIONS ON SIGNAL PROCESSING, IEEE JOURNAL OF SELECTED
TOPICS IN SIGNAL PROCESSING, and the IEEE JOURNAL ON SELECTED AREAS
IN COMMUNICATIONS. He is a Member of the Sensor Array and Multichannel
(SAM) Technical Committee (2009-present); the Signal Processing for Commu-
nications (SPC) Technical Committee (1999–2001); and the Communications
Theory Technical Committee of the IEEE Communications Society. He was on
the IEEE Communications Society’s Awards Committee. He is an IEEE Third
Millennium Medal winner, an IEEE Signal Processing Society Distinguished
Lecturer (2011–2012), and a member of Eta Kappa Nu and Sigma Xi. He was
awarded an ONR Young Investigator Award (1997) and an NSF Research Ini-
tiation Award (1992). His IEEE Fellow Citation “for scientific contributions to
detection, data fusion and signal processing with multiple sensors” acknowl-
edges some early contributions to the field of sensor networking.

View publication stats

You might also like