IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO.

2, FEBRUARY 2011 629

SPICE: A Sparse Covariance-Based Estimation

Method for Array Processing
Petre Stoica, Fellow, IEEE, Prabhu Babu, and Jian Li, Fellow, IEEE

Abstract—This paper presents a novel SParse Iterative Co- a small number of sources exists and therefore that only a few
variance-based Estimation approach, abbreviated as SPICE, to rows of the signal matrix
array processing. The proposed approach is obtained by the
minimization of a covariance matrix fitting criterion and is par-
ticularly useful in many-snapshot cases but can be used even in .. .. .. .. (2)
. . . .
single-snapshot situations. SPICE has several unique features not
shared by other sparse estimation methods: it has a simple and
sound statistical foundation, it takes account of the noise in the
data in a natural manner, it does not require the user to make are different from zero. The estimation problem is then to de-
any difficult selection of hyperparameters, and yet it has global cide from the data which rows of the above matrix are
convergence properties. nonzero. Indeed, once this is done, the solution to the location
Index Terms—Array processing, covariance fitting, direction-of- estimation problem, which is usually the main goal of array pro-
arrival (DOA) estimation, sparse parameter estimation. cessing, is immediate : if the row (let us say) of (2) is deemed
to be different from zero then we can infer that there is a corre-
sponding source at an estimated location equal to .
I. INTRODUCTION AND PRELIMINARIES The previous formulation of the location problem begs for the
use of basic ideas from the area of sparse parameter estimation,

C ONSIDER an array processing scenario in which the

main problem is to estimate the location parameters of a
number of narrowband sources that are present in the array’s
or rather simple extensions of those ideas to the present multi-
snapshot case. To describe these ideas briefly, let
viewing field. Let denote the set of possible locations, and (3)
let be a generic location parameter. Also, let denote
a grid that covers . We assume that the grid is fine enough ..
. (4)
such that the true location parameters of the existing sources
lie on (or, practically, close to) the grid. Under this reasonable
assumption we can use the following nonparametric model for (5)
the output of the array (see, e.g., [1]):
where the superscript denotes the conjugate transpose (we de-
note the matrix in (5) by to reserve the notation for an ex-
tended form of (5) that will be introduced in the next section).
(1)
A direct application of the -norm minimization principle (see,
e.g., [2]) to the present scenario described by (1)–(5) consists of
where is the total number of snapshots, is the number of estimating the matrix as the solution to the following con-
sensors in the array, is the th observed snapshot, strained minimization problem:
denotes the array transfer vector (aka manifold or
steering vector) corresponding to is the unknown (6)
signal impinging on the array from a possible source at , and
is a noise term.
where denotes the Euclidean norm for vectors and the Frobe-
A sparse (or semi-parametric) estimation method makes the
nius norm for matrices, and is a threshold that must be chosen
assumption, reminiscent of the parametric approach, that only
by the user. Note that the objective in (6) is equal to the -norm
of the vector , an observation that shows clearly that
Manuscript received June 11, 2010; revised August 09, 2010, October 22, this approach is a direct extension of the standard single-snap-
2010; accepted October 26, 2010. Date of publication November 01, 2010; date
of current version January 12, 2011. The associate editor coordinating the review shot approach of, e.g., [2].
of this manuscript and approving it for publication was Prof. Jean Pierre Delmas. A method for array processing based on (6) was pursued in
P. Stoica and P. Babu are with the Department of Information Technology, [3]. Note that (6) is easily recognized to be an SOCP (second
Uppsala University, Uppsala, SE 75105, Sweden (e-mail: [email protected]; prabhu.
[email protected]). order cone program), see, e.g., [4], which can be efficiently
J. Li is with the Department of Electrical and Computer Engineering, Univer- solved provided that , , and do not take on too large
sity of Florida, Gainesville, FL 32611 USA (e-mail: [email protected]). values. However, in the array processing application this is not
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. necessarily the case : indeed, while (tens to hun-
Digital Object Identifier 10.1109/TSP.2010.2090525 dreds) is reasonably small, and
1053-587X/$26.00 © 2010 IEEE
Authorized licensed use limited to: UNIVERSITY OF MANITOBA. Downloaded on August 04,2024 at 03:22:44 UTC from IEEE Xplore. Restrictions apply.
630 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 2, FEBRUARY 2011

(depending on the desired resolution and the dimension of This assumption on the noise term in (1) is reasonable in most
(1D, 2D, 3-D etc.)) can be rather large. For such large dimen- applications. Let us also assume that the signals and the
sional problems the currently available SOCP software (see, noise are uncorrelated with each other for any and that
e.g., [5]) is too slow to use. In an attempt to overcome this com-
putational problem, [3] suggested a way to reduce the number (9)
of columns of and in (6) to manageable values by means
of a singular value decomposition operation. Then the data snapshots are uncorrelated
The approaches that rely on (6), such as the one in [3], suffer with one another and have the following covariance matrix:
from a number of problems. First, the motivation of (6) is
more clearly established in the noise-free case than in the more
practical noisy data case. Second, and likely related to the first
problem, there exist no clear-cut guidelines for the selection of
in (6) (see, e.g., [6] for a critical discussion on this aspect). .. .. .. ..
. . . .
Finally, solving (6) as an SOCP may be too time consuming
for some applications in which , and especially take
on large values, and the available techniques for reducing the
dimensions in (6) (such as that suggested in [3]) require the .. .. .. .. .. ..
choice of further hyperparameters, besides , and even so they . . . . . . ..
.
are only approximate.
In this paper we present a new method of sparse parameter
.. .. .. .. .. .. ..
estimation in models of the type of (1). This method, which is . . . . . . .
called SPICE (SParse Iterative Covariance-based Estimation),
is obtained using a novel covariance-based fitting approach that (10)
was recently introduced in [7] where the focus was on time-
series data (the single-snapshot case) as opposed to the array where
data (the multi-snapshot scenario) considered here. SPICE has a
number of useful features that are hardly shared by other sparse
estimation methods. In particular, SPICE does not suffer from
(11)
the problems described in the previous paragraph. Indeed, as
will be shown in the next section, SPICE has a sound (covari-
ance-based) statistical motivation which makes it possible to use .. .. .. .. .. ..
the method in noisy data scenarios without the need for choosing . . . . . .
any hyperparmeters. Additionally, the SPICE algorithm has a
simple form, and yet it enjoys global convergence properties.
We will show that the covariance fitting problem that SPICE .. .. .. .. .. .. ..
. . . . . . .
solves can be reformulated as an SOCP of the form of (6) but
with , with and replaced by other matrices with
only columns (typically ), and with row-augmented
matrices and that, unlike and in (6), take account of .. .. .. .. .. ..
the noise term in the data (1). This SOCP formulation of the . . . . . .
SPICE estimation criterion shows that the -norm minimiza- (12)
tion problem in (6) can be given a simple statistical motivation
based on covariance fitting, provided that the matrices in (6) are .. .. .. .. .. .. ..
. . . . . . .
suitably defined, see the rest of this paper for details.

II. SPICE ESTIMATION CRITERION The assumption in (9) that the source signals are spatially un-
correlated, which led to the above expression for the covariance
Let us assume that matrix , does not always hold: in some applications the signals
can be correlated or even coherent. However the SPICE method
proposed in this paper is robust to this assumption—we refer to
.. .. .. .. (7) [7] (see also [1]) for a theoretical explanation of this robustness
. . . . property, which will be illustrated numerically in Section V.
We will consider the following covariance fitting criterion for
the purpose of parameter estimation (see, e.g., [8] and the refer-
where stands for the expectation operator, and ences therein):
if
(8) (13)
elsewhere.
Authorized licensed use limited to: UNIVERSITY OF MANITOBA. Downloaded on August 04,2024 at 03:22:44 UTC from IEEE Xplore. Restrictions apply.
STOICA et al.: SPICE: A SPARSE COVARIANCE-BASED ESTIMATION METHOD FOR ARRAY PROCESSING 631

where denotes the positive definite square-root of Appendix A, and is therefore convex (see, e.g., [9]). Solving (18)
as an SDP, however, is not recommended due to the fact that the
(14) available SDP solvers are too computationally intensive for the
values of , , and encountered in many array processing
and the inverses of and are assumed to exist. Note that applications. Consequently we adopt a different line of attack
exists under weak conditions: indeed, while the “true” values of that consists of replacing the problem of minimizing (18) by a
many in may be zero, one typically has related problem, as described next.
which renders nonsingular. Regarding , the in- It follows from (17) that a consistent (in ) estimate of the
verse of this matrix exists with probability one as long as right-hand side of this equation is given by . Therefore, we
. However, for the sample covariance matrix is sin- can think of reformulating the problem of minimizing as the
gular and therefore (13) cannot be used. In the latter case, one following constrained minimization:
can estimate the parameters by minimizing the following
criterion, instead of (13): (20)
(15)
where
The above criterion is the one used in [7] where the focus was
on the time series case with . In the present array pro-
cessing scenario, however, we prefer to consider (13) because (21)
usually the condition is satisfied and because (13) has
a statistically stronger motivation than (15): indeed, under cer- Interestingly, the problems (18) and (20) are not only asymptoti-
tain conditions specified in [8], it can be shown that the param- cally equivalent (as increases, and under the condition that
eter estimates minimizing (13) are statistically asymptotically in (10) can represent the true covariance matrix) but they are ex-
(in ) efficient, whereas the estimates obtained from (15) are actly equivalent (in general) in the sense that their solutions are
suboptimal. Nevertheless, we should note that what we do in scaled versions of each other (note that a scaling of has no
the following for (13) applies with minor modifications to (15) effect on source location estimation). This equivalence property
as well, see Remark 1 below—this observation can be useful in is proved in Appendix B where we also show that the problem
those array processing applications in which and there- obtained from (18) by constraining the first term to (a con-
fore in which (15) should be used in lieu of (13). straint suggested, once again, by asymptotic considerations), in-
A simple calculation shows that stead of constraining the second term as in (20), is equivalent to
(18) and (20) as well.
The problem (20) is also an SDP, and therefore convex.
Furthermore, note that the linear constraint in (20) is of the
(weighted) -norm type, and thus it can be expected to be
sparsity inducing for the solution to (20). Apparently, the crite-
(16)
rion in (20) was never considered in the previous literature on
where sparse parameter estimation. On the other hand, interestingly
enough, this type of criterion occurs frequently in the seemingly
unrelated literature on optimal experiment design (see, e.g.,
(17)
[10] and the references of that paper). The latter literature (in
particular [10]) has served as a source of inspiration for [7] and,
It follows from (16) and (17) that the minimization of is equiv- to some extent, for this paper as well.
alent to the minimization of the function
III. SPICE UPDATING FORMULAS
(18) Let and consider the following problem:

Remark 1: A similar calculation shows that in the case of (22)

(15) the function to be minimized with respect to the unknown
powers is The solution to (22) is given (for fixed ) by

(23)
(19) and the corresponding minimum value of the function in (22) is

Owing to the analogy between (18) and (19), it should come as

no surprise that what we do in the following for (18) can also be the original objective in (20) (24)
done for (19) if necessary (e.g., for ).
The problem of minimizing in (18) with respect to To prove the above assertion, observe that (23) follows if we can
can be easily shown to be an SDP (semi-definite program), see show that (hereafter the notation , with and being
Authorized licensed use limited to: UNIVERSITY OF MANITOBA. Downloaded on August 04,2024 at 03:22:44 UTC from IEEE Xplore. Restrictions apply.
632 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 2, FEBRUARY 2011

Hermitian matrices of appropriate dimensions, means that the From this, it follows that the minimization of the objective in
difference matrix is positive semi-definite): (22) with respect to (s.t. and )
gives (for fixed ):

(25) (31)

By standard properties of partitioned matrices (see, e.g., [11,

Lemma A.3]) and the fact that , (25) holds if and only if and the corresponding minimum value of the objective is
the following partitioned matrix is positive semi-definite:
(32)

Equation (31) above provides the solution to the second step

of the cyclic algorithm, whereas the solution to the first step is
given by (23). Combining (23) and (31) leads to the updating
(26) formulas of the SPICE algorithm in which only the powers
(that are the quantities of main interest) appear explicitly:
The central matrix in (26) can be rewritten as

(33)
(27)

and thus it is obviously positive semi-definite (because it has (34)

the form , with ). Therefore, (25) (and
hence (23)) is proved and then (24) follows via substitution.
To summarize, we have proved above (see (24)) that the mini- where the index denotes the iteration number, and is the
mization of the objective in (22) with respect to , for any fixed matrix made from . The algorithm can be initialized
, leads to the original function of in (20). Then it fol- with the power estimates obtained by means of the periodogram
lows that the minimization of (22) with respect to and method (see, e.g., [1]):
yields the same as the minimization of (20). The usefulness
of this observation lies in the fact that the minimization of the
(35)
augmented function in (22) can be conveniently done by means
of a cyclic algorithm that minimizes (22) with respect to , for
fixed , then minimizes (22) with respect to , for given , and Remark 2: The SPICE algorithm for the alternative covari-
so forth until convergence. The solution to the first step of this ance fitting criterion in (15) (or (19)) can be readily derived by
algorithm was already derived, see (23). The solution needed in paralleling the above calculations. The result is an updating for-
the second step can also be obtained in closed form. To show mula similar to (33) above with the only difference that
how this can be done, let in (33) and (34) should be replaced by and in (21) by
.
.. (28)
. B. The Case of Identical
In some applications, it is known that the noise components in
and observe that the different elements of the array output vector have the same
variance:

(29) (36)

Using this information is important since, based on it, we can

reduce the number of powers that need to be estimated. To de-
A. The Case of Different
rive the necessary modification of SPICE, that takes (36) into
By the Cauchy–Schwarz inequality account, first observe that the minimization of (22) with respect
to is not affected by (36). However the minimization with re-
spect to and , for fixed , is slightly different. Under
the above constraint on , (29) becomes

(30) (37)

Authorized licensed use limited to: UNIVERSITY OF MANITOBA. Downloaded on August 04,2024 at 03:22:44 UTC from IEEE Xplore. Restrictions apply.
STOICA et al.: SPICE: A SPARSE COVARIANCE-BASED ESTIMATION METHOD FOR ARRAY PROCESSING 633

This function is to be minimized with respect to and Because SPICE monotonically decreases the objective
, subject to function (due to its cyclic operation) and as the minimization
problem it solves is convex, we can expect that the algorithm
(38) has global convergence properties. That this is indeed the case
follows from the general analysis in [10] where it was proved
where that, under weak conditions (essentially requiring that
and that the matrix stays positive definite as the iteration
proceeds), the limit points of the SPICE iterative process are
(39)
global solutions to (20).

By the same argument as above, see (30)–(32), the solution to IV. SOCP FORMULATION OF SPICE
this optimization problem is
SPICE takes account of the noise without introducing any hy-
(40) perparameters to be selected by the user. Indeed, SPICE is a
fully data adaptive approach that estimates both and
from the available data snapshots. As already mentioned, to our
knowledge, the SPICE criterion was never used in the previous
literature on sparse parameter estimation. It is therefore some-
(41) what unexpected that SPICE can be related to (an extended ver-
sion of) the classical -norm minimization approach in (6), as
where we explain in the rest of this section.
As indicated in Section II the SPICE estimation problem,
(42) (20), can be cast and solved as an SDP. However it can also
be reformulated as an SOCP which is a special case of SDP. It
is this reformulation of (20) as an SOCP that reveals the connec-
and the corresponding minimum function (for fixed ) is tion between SPICE and a certain extended -norm minimiza-
tion problem. To reformulate SPICE as an SOCP we proceed
in the following way. Instead of minimizing (22) cyclically (or
(43) alternately) with respect to and , we minimize this function
with respect to for arbitrary . The result of this minimiza-
tion operation is already available from the calculations in the
Inserting the expression (23) for in (40) and (41), we obtain previous section where it was shown that the minimizing powers
the following modified SPICE algorithm for the case of (36): are given by (31) and the corresponding minimum function by
(32). It remains to minimize (32) with respect to (under the
(44) constraint in (22)):

(48)
(45)

Once is obtained from (48), can be calculated using (31).

The above problem, which is an SOCP (see Appendix A),
is similar to the -norm minimization problem (6). However,
(46) there are also significant differences between these two prob-
lems, as well as between the ways in which they have been ob-
Initial estimates of the powers, for , can still be tained, as follows.
obtained with the periodogram (see (35)), and can be ini- i) The -norm in (48) has a weighted form, with the
tialized for instance as the average of the smallest values of weights determined by the data. It is well known that
each multiplied by . To motivate this choice for is an estimate of the inverse power
note that, at least for sufficiently large, we have corresponding to the th point of the location grid (see,
with the equality being likely to e.g., [1]). Therefore, the smaller the said power the larger
hold for the smallest values of , e.g., for the smallest such the and consequently more weight is put on mini-
values. This observation implies that (below are the mizing the associated in (48), which intuitively is the
smallest values of the set and the corresponding way it should be. Furthermore, the -norm objective in
manifold vectors) (48) follows naturally from the statistically sound covari-
ance fitting criterion in (13), in contrast with the similar
objective in (6) whose motivation is essentially heuristical.
(47)
ii) The matrices and in (48) comprise extra rows that
account for the noise in the data, unlike the similar quan-
can be expected to be a reasonable estimate of . tities and in (6).
Authorized licensed use limited to: UNIVERSITY OF MANITOBA. Downloaded on August 04,2024 at 03:22:44 UTC from IEEE Xplore. Restrictions apply.
634 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 2, FEBRUARY 2011

iii) The inequality constraint in (6) is replaced in (48) by a hy-

perparameter-free equality constraint where is used
in lieu of in (6). Note that the matrix in (48) has
columns instead of columns as in (6), which evi-
dently makes (48) more convenient computationally (typ-
ically ). This dimensional reduction also follows
naturally from the employed covariance fitting criterion,
whereas the idea suggested in [3] for reducing the number
of columns of and in (6) was motivated on more or
less heuristical grounds and its use depends on the selec-
tion of an additional hyperparameter, besides .
The reader might wonder what is the relationship between
in (48) and in (6). To clarify this aspect, observe that we can
write the data (1) in the following matrix form:
Fig. 1. Inter-element spacings (in units of =2) for the NULA with N = 100.
(49)

where According to our albeit limited experience, the SPICE mul-

tiplicative algorithm in (33) and (34) (or (44)–(46)) is usually
(50) visibly faster than the SOCP-based algorithm that was outlined
in this section. However, the SOCP formulation of the SPICE es-
It follows from (49) that we can assume, as in (6), that timation problem remains important for understanding the con-
provided that is “sufficiently large” ; furthermore, a nection between the proposed approach and the standard one
proper selection of would require information about the noise based on -norm minimization.
variance(s) and the noise distribution, which is usually unavail-
able, and even then the inequality constraint in (6) could not in V. NUMERICAL ILLUSTRATIONS AND CONCLUDING REMARKS
general be made to hold surely for any finite practically conve- In this section, we illustrate numerically the performance of
nient value of . On the other hand, it also follows from (49) the proposed methods and compare it with the performance of
that some standard methods from the literature. We will consider
both a uniform linear array (ULA) and a nonuniform linear
(51)
array (NULA). In the ULA case, the sensors are uniformly
placed with a spacing of , where denotes the wavelength
and therefore that of the sources. The inter-element spacings in the case of NULA
are as shown in Fig. 1. The number of sensors in the array is
(52) for ULA and for NULA. In the ULA case the
number of snapshots is whereas it is in the
NULA case. The steering vector for the ULA, corresponding
The above equation clarifies the relationship between and . to a direction of arrival (DOA) equal to , is given by
Unlike (6), the equality constraint (52) holds deterministically
and it does not involve any hyperparameter. Note also that if ..
. (54)
is row-sparse then so is the matrix in (52). Re-
garding the other matrix appearing in (52), viz. ,
(observe that is a constant). The steering vector for
it tends to zero as the signal-to-noise ratio (SNR) increases ;
the NULA can be similarly defined. The interval for the DOA is
however, for small or medium SNR values this matrix is not
. We use a uniform grid to cover ,
row-sparse, which may be perceived as a downside of (52), al-
with a step of 0.1 , which means that .
though a relatively minor one as the matrix can still be con-
The source signals , see (1), have constant modulus,
sidered to be row-sparse because .
which is usually the situation in communications applications.
Whenever the constraint is enforced, the
We will consider cases with both uncorrelated and coherent
above reformulation of the SPICE estimation problem should
be changed as follows. First, is obtained as the solution of the sources. The noise term in (1) is chosen to be white, both tem-
following SOCP (see (43)): porally and spatially, and Gaussian distributed with zero mean
and variance .

A. Fixed Sources : DOA Estimation

In this subsection we consider DOA estimation of fixed
(53) sources using the ULA. The data samples were simulated using
(1) with three sources at 10 , 40 , 55 and
(see Appendix A for the formulation of the above problem as an the following signals ,
SOCP). Then and are determined using (40)–(42). and , where the phases were
Authorized licensed use limited to: UNIVERSITY OF MANITOBA. Downloaded on August 04,2024 at 03:22:44 UTC from IEEE Xplore. Restrictions apply.
STOICA et al.: SPICE: A SPARSE COVARIANCE-BASED ESTIMATION METHOD FOR ARRAY PROCESSING 635

independently and uniformly distributed in . In the

simulation with coherent sources, the sources at and were
the coherent ones (i.e., they had identical phases), whereas the
source at was uncorrelated to them. The noise variance was
varied to control the signal-to-noise ratio defined as

SNR dB

The following methods were used for DOA estimation :

: the periodogram (PER), see (35);

: the iterative adaptive approach (IAA), see [12];
: the multiple signal classification method (MUSIC), see
e.g., [1];
: SPICE, see (15), (19) as well as Remark 2 and [7];
: the enhanced SPICE or, for short, SPICE-plus
(SPICE+), see (44)–(46).
Note that IAA was shown in [12] to outperform the method of
[3] and its variations in the literature, which is why we selected
IAA for the present comparison. As a cross-check and for com-
parison’s sake we have also computed the SPICE+ estimates by
means of an SDP solver (see [5] for the latter). The so-obtained
estimates were identical (within numerical accuracy) to those
provided by the SPICE+ algorithm; however, as expected, the
SDP solver was either much slower or, worse, it could not be
executed for instance due to memory problems. Fig. 2. Estimation performance of PER, IAA, MUSIC, SPICE and SPICE+
In Fig. 2, we show the average root mean-square errors (the ULA case): RMSE versus SNR for the DOA estimates obtained in 1000
Monte Carlo runs. (a) Uncorrelated sources. (b) Coherent sources.
(RMSEs) of the DOA estimates obtained with in
1000 Monte Carlo runs, for several SNR values:
suming two mobile sources, one moving linearly from 30 to
60 and the other from 60 to 30 , in steps of 0.03 , over a
course of 1000 data snapshots. The uncorrelated signals of these
two sources were given by and
where denotes the estimate of in the th Monte Carlo , where the phases were independently and
run. In this figure, the top plot shows the RMSEs for uncor- uniformly distributed in .
related sources and the bottom plot shows the RMSEs for co- In this example, only the methods , , and are
herent sources. As can be seen from this figure, the PER esti- considered. At any given time instant , the sample covariance
mates of the DOAs have poor accuracy in all cases under study, matrix is formed from the most recent 100 data snapshots
due to a significant bias (which does not decrease as the SNR and the DOA estimates are recomputed by each of the three
increases). Regarding IAA, this method provides competitive methods. As before, the SPICE methods, and , are
estimates only for SNR 0 dB. The parametric method of initialized with the PER estimate. Alternatively, one may
MUSIC, which requires knowledge on the number of sources think of initializing the SPICE methods with their respective
in the data, yields reasonably accurate DOA estimates in the estimates obtained from the previous data window; however,
uncorrelated source case (at least for SNR 0 dB) but com- as discussed in Section III-B, for global convergence SPICE
pletely fails in the coherent source case (as expected). The pro- should be initialized with a dense estimate rather than a sparse
posed methods of SPICE and SPICE+ give the best performance one. The SPICE methods were iterated only five times except
in the cases considered: in particular, their threshold SNR ap- at where they were applied for the first time and were
pears to be lower by some 10 dB than the threshold SNR of IAA iterated till convergence. We have also tried a larger number of
and of MUSIC. For uncorrelated sources, SPICE+ is more ac- iterations, such as 20, but noticed no significant improvement
curate than SPICE, whereas SPICE outperforms SPICE+ in the in the DOA estimates [compare Fig. 3(c) and (d)]. The SNR in
coherent source case. From a computational standpoint SPICE+ this simulation was 20 dB (however, very similar results were
converged faster than SPICE presumably due to the data-depen- obtained for SNR 10 dB and 0 dB).
dent weights used by SPICE+. Fig. 3 shows the plots of DOA estimates versus for the three
considered methods. In each plot, the estimates shown at any
B. Mobile Sources : DOA Tracking time were obtained as the locations of the two largest peaks
In this subsection, we consider DOA estimation of mobile in the spatial spectrum provided by the method in question.
sources using the NULA. The data samples were generated as- As can be seen from the figure, PER performs poorly as it
Authorized licensed use limited to: UNIVERSITY OF MANITOBA. Downloaded on August 04,2024 at 03:22:44 UTC from IEEE Xplore. Restrictions apply.
636 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 2, FEBRUARY 2011

Fig. 3. Tracking performance of PER, SPICE and SPICE+ (the NULA case). The solid lines denote the trajectories of the true DOAs: (a) PER; (b) SPICE with
five iterations; (c) SPICE+ with five iterations; and (d) SPICE+ with 20 iterations.

cannot track either the source moving from 30 to 45 or the

one moving from 45 to 30 . Note that at each time instant,
the data window (made of the most recent 100 data snapshots)
comprises two sets of DOAs each with a width of 3 . For
instance, at , the data window will contain signals
with DOAs lying in the interval 39 42 and, respectively,
in 48 51 . PER often picks wrongly the two peaks only
from the set with larger DOAs. On the other hand, the SPICE
methods correctly pick the two peaks from both sets which
leads to the band-like appearance of the corresponding plots in
Fig. 3(b)–(d); observe that, as expected, the width of the bands
in these figures (measured along the DOA axis) is equal to 3 .
Note also that the true DOA trajectories are well approximated
by the upper edge of the bands in Fig. 3(b)–(d) corresponding
to the increasing DOA and by the lower edge of the bands
associated with the decreasing DOA—this behavior, which was
expected in view of the above discussion, is in fact quite stable
Fig. 4. Tracking performance of SPICE+ for a sinusoidal trajectory (the NULA
with respect to the SNR [in simulations not shown here, we case). The solid line denotes the true DOA trajectory.
have observed that decreasing the SNR from 20 to 0 dB caused
only 1 DOA estimate, out of 2000, to lie outside bands similar
to those in Fig. 3(b)–(d)]. To provide further insight into this APPENDIX A
type of behavior, a single source moving along a sinusoidal SOME SDP AND SOCP FORMULATIONS
trajectory was considered. Fig. 4 shows the DOA estimates
for this source obtained with SPICE+. It is clear from this First, we show that the problem of minimizing (18), s.t.
plot that the width of the band decreases till , where , can be formulated as an SDP. The proof
it is nearly zero, and then starts increasing again; hence, as that the same is true for (20) is similar and therefore its details
expected, the width of the band is proportional to the slope are omitted. Let
of the DOA variation. Furthermore in this case, too, the true
DOA trajectory is well approximated by the upper edge of the ..
band (for increasing DOA) and the lower edge (for decreasing . (55)
DOA).
Authorized licensed use limited to: UNIVERSITY OF MANITOBA. Downloaded on August 04,2024 at 03:22:44 UTC from IEEE Xplore. Restrictions apply.
STOICA et al.: SPICE: A SPARSE COVARIANCE-BASED ESTIMATION METHOD FOR ARRAY PROCESSING 637

and let APPENDIX B

(56) EQUIVALENCE OF (18) AND (20)
Let and denote the minimization problems corre-
Using this notation we can rewrite the function in (18) as sponding to (18) and (20). Therefore

(57) (64)

and
Let be auxiliary variables satisfying ,
or equivalently (65)

(58) Note that we have replaced by 1 in the right-hand side of the

constraint in (65) for notational convenience (this replacement
Then the minimization problem under discussion can be stated has only a scaling effect on the solution of ). For complete-
as ness, we will also consider the following problem obtained by
constraining the first term in (64) to one:

(66)

First, we prove the equivalence of and . By making use of

(59) an auxiliary variable we can rewrite (65) and (66) as follows:

which is an SDP (as is a linear function of ) [4], [9]. and (67)

Next, we show that (48) can be cast as an SOCP. Let
denote auxiliary variables that satisfy and (68)

(60) Let (note that ) and reformulate as

Then simply observe that (48) can be rewritten as and (69)

where is made from . It follows from (67) and (69) that

if is the solution to then is the solution to ,
and thus the proof of the equivalence between and is
concluded.
Next, we consider and . By writing down the
(61)
Karush–Kuhn–Tucker conditions for it can be shown
which is an SOCP [4]. that the two terms of the objective in (64) must be equal to one
Finally, we show how to reformulate (53) as an SOCP. Let another at the minimizing . A more direct proof of this
still denote some auxiliary variables (but now only fact runs as follows. Assume that is the optimal solution
of them) and constrain them to satisfy of and that at we have

(70)

for some ; note that the scaling factor in (70) must nec-
(62) essarily be positive, which is why we wrote it as ; below we
will let denote the square root of . We will show in
Using we can restate (53) in the following form: the following that the previous assumption leads to a contradic-
tion : cannot be the solution to if . To do so let
and observe that, at , the two terms in the ob-
jective of are identical:

(71)

Making use of (70) and (71) along with the assumption that
.. is the solution to , we obtain the following inequality:
.
(72)

(63) or, equivalently,

which, once again, is an SOCP. (73)

Authorized licensed use limited to: UNIVERSITY OF MANITOBA. Downloaded on August 04,2024 at 03:22:44 UTC from IEEE Xplore. Restrictions apply.
638 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 2, FEBRUARY 2011

which cannot hold, and hence the solution to must satisfy [12] T. Yardibi, J. Li, P. Stoica, M. Xue, and A. B. Baggeroer, “Source
(70) with . Using this fact we can reformulate as localization and sensing: A nonparametric iterative adaptive approach
based on weighted least squares,” IEEE Trans. Aerosp. Electron. Syst.,
follows: vol. 46, no. , pp. 425–443, Jan. 2010.

and
(74) Petre Stoica (SM’91–F’94) is a Professor of systems modeling in the Depart-
or, equivalently (as ), ment of Information Technology, Uppsala University, Uppsala, Sweden.
More details about him can be found at http://www.it.uu.se/katalog/ps.
and (75)

where , as before, and is a new auxiliary Prabhu Babu received the M.Sc. degree in electrical engineering from the In-
dian Institute of Technology, Delhi, India, in 2007.
variable. Comparing (75) and (67) concludes the proof of the He is currently working towards the Ph.D. degree in electrical engineering
equivalence of and . with applications in signal processing at the Department of Information Tech-
nology, Uppsala University, Uppsala, Sweden.
REFERENCES
[1] P. Stoica and R. Moses, Spectral Analysis of Signals. Upper Saddle
River, NJ: Prentice-Hall, 2005. Jian Li (S’88–M’90–SM’97–F’05) received the M.Sc. and Ph.D. degrees in
[2] R. Tibshirani, “Regression shrinkage and selection via the lasso,” J. R. electrical engineering from The Ohio State University (OSU), Columbus, in
Stat. Soc. .B, Methodol., vol. 58, no. 1, pp. 267–288, 1996. 1987 and 1991, respectively.
[3] D. Malioutov, M. Cetin, and A. Willsky, “A sparse signal reconstruc- From April to June 1991, she was an Adjunct Assistant Professor with the
tion perspective for source localization with sensor arrays,” IEEE Department of Electrical Engineering, OSU. From July 1991 to June 1993,
Trans. Signal Process., vol. 53, no. 8, pp. 3010–3022, Aug. 2005. she was an Assistant Professor with the Department of Electrical Engineering,
[4] M. Lobo, L. Vandenberghe, S. Boyd, and H. Lebret, “Applications of University of Kentucky, Lexington. Since August 1993, she has been with the
second-order cone programming,” Linear Algebra Appl., vol. 284, no. Department of Electrical and Computer Engineering, University of Florida,
1-3, pp. 193–228, 1998. Gainesville, where she is currently a Professor. In fall 2007, she was on
[5] J. Sturm, “Using SeDuMi 1.02, a MATLAB toolbox for optimization sabbatical leave at the Massachusetts Institute of Technology, Cambridge.
over symmetric cones,” Optim. Methods Softw., vol. 11, no. 1, pp. Her current research interests include spectral estimation, statistical and array
625–653, 1999. signal processing, and their applications.
[6] A. Maleki and D. Donoho, “Optimally tuned iterative reconstruction al- Dr. Li was a member of the Editorial Board of Signal Processing from
gorithms for compressed sensing,” IEEE J. Sel. Topics Signal Process., 2005 to 2007. She has been a member of the Editorial Board of Digital Signal
vol. 4, no. 2, pp. 330–341, 2010. Processing—A Review Journal since 2006. She was an Executive Committee
Member of the 2002 International Conference on Acoustics, Speech, and
[7] P. Stoica, P. Babu, and J. Li, “New method of sparse parameter estima-
Signal Processing, Orlando, FL, in May 2002. She was an Associate Editor of
tion in separable models and its use for spectral analysis of irregularly
the IEEE TRANSACTIONS ON SIGNAL PROCESSING from 1999 to 2005 and an
sampled data,” IEEE Trans. Signal Process., vol. 59, no. 1, pp. 35–47, Associate Editor of the IEEE Signal Processing Magazine from 2003 to 2005.
Jan. 2011. She is presently a member of the Sensor Array and Multichannel Technical
[8] B. Ottersten, P. Stoica, and R. Roy, “Covariance matching estima- Committee, IEEE Signal Processing Society. She received the 1994 National
tion techniques for array signal processing applications,” Digital Signal Science Foundation Young Investigator Award and the 1996 Office of Naval
Process., vol. 8, no. 3, pp. 185–210, 1998. Research Young Investigator Award. She is a coauthor of papers that received
[9] Y. Nesterov and A. Nemirovskii, Interior-Point Polynomial Algorithms the First and Second Place Best Student Paper Awards, respectively, at the 2005
in Convex Programming. Philadelphia, PA: SIAM, 1994. and 2007 Annual Asilomar Conferences on Signals, Systems, and Computers,
[10] Y. Yu, “Monotonic convergence of a general algorithm for computing Pacific Grove, CA. She is also a coauthor of the paper that received the M. Barry
optimal designs,” Ann. Stat., vol. 38, no. 3, pp. 1593–1606, 2010. Carlton Award for the best paper published in the IEEE TRANSACTIONS ON
[11] T. Söderström and P. Stoica, System Identification. Englewood AEROSPACE AND ELECTRONIC SYSTEMS in 2005. She is a Fellow of IET and a
Cliffs, NJ: Prentice-Hall, 1989. member of Sigma Xi and Phi Kappa Phi.

Authorized licensed use limited to: UNIVERSITY OF MANITOBA. Downloaded on August 04,2024 at 03:22:44 UTC from IEEE Xplore. Restrictions apply.