2013 5th IEEE International Workshop on Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP)

Iterative root-MUSIC algorithm for DOA estimation

Mahdi Shaghaghi + and Sergiy A. Vorobyov +,∗

+ Department of Electrical and Computer Engineering, University of Alberta
Edmonton, AB T6G 2V4, Canada
∗ School of Electrical Engineering, Aalto University

Espoo, FI-00076 AALTO, Finland

Emails: [email protected] and [email protected]
(Invited Paper)

Abstract—This paper introduces a new high-resolution the signal and noise subspaces, we introduce a new iterative
subspace-based algorithm for direction-of-arrival estimation. The DOA estimation method. The algorithm in [9] is designed for
proposed method improves the quality of the estimation especially compressive measurements, whereas the proposed method here
in the case of small sample size by considering the structure of the is designed for signals sampled at the Nyquist rate. Further-
sample covariance matrix. The key idea is to identify undesirable more, the structural model of the sample covariance matrix is
terms in the sample covariance matrix which cause perturbations
considered here. It is observed that there are undesirable terms
in the estimation of the signal and noise subspaces. These terms
are then diminished in an iterative manner. The proposed method that cause perturbations on the estimations of the signal and
is studied by investigating the mean squared error, the detection noise subspaces. We deal with this problem by diminishing
probability, and the mean squared error in case of successful the undesirable terms in an iterative manner. The proposed
detection. It is shown that the new method outperforms the approach consists of a number of steps. The root-MUSIC
conventional methods. method is first used to estimate the DOAs. Next, the signal
and noise components are estimated and are used to reduce the
magnitude of the undesirable terms in the sample covariance
I. I NTRODUCTION
matrix. Then, updated estimations of the signal and noise
Direction-of-arrival (DOA) estimation has wide applica- vectors are obtained from the modified sample covariance
tions in radar and sonar systems, communications, astro- matrix. Given the updated estimations, the sample covariance
physics, seismology, etc. [1], [2]. Classical high-resolution matrix can be polished further. This process is pursued in an
methods for DOA estimation such as the multiple signal iterative manner.
classification (MUSIC) method [3], the root-MUSIC algorithm
It is worth mentioning that another iterative MUSIC
[4], and the estimation of signal parameters via rotational
method has been introduced in [10]. However, the approach
invariance techniques (ESPRIT) [5] are based on estimating
pursued here is different from [10] in that we collectively
the signal and noise subspaces from the sample covariance
consider the effect of all the sources to improve the quality of
matrix. The fidelity of the sample covariance matrix to the
the sample covariance matrix estimate over iterations, whereas
true data covariance matrix plays a critical role in a successful
the method in [10] is based on degenerating the observations
estimation. In real-life scenarios, only a limited number of
into a number of channels where each channel makes an
samples is available, and in the case that the number of samples
estimation for an individual source. The problem of the method
is comparable in magnitude to the observation dimension
in [10] is that if one of the sources is more powerful than the
(number of antenna array elements), the estimated subspaces
rest, it becomes dominant in all the channels and the algorithm
can largely deviate from the true subspaces. This problem
fails to detect the weaker sources. This issue does not raise in
has been recently addressed in [6]–[8] and it is dealt with
our method, as all the sources are taken into account together.
using random matrix theory. The introduced method is referred
to as G-MUSIC and it considers the asymptotic situation
when both the sample size and the number of array elements II. S YSTEM M ODEL AND BACKGROUND
tend to infinity at the same rate. It is then inferred that the
introduced method gives a more accurate description of the Consider L number of narrowband plane waves im-
situation where these two quantities are finite and comparable pinging on a uniform linear array (ULA) from directions
in magnitude [7]. θ1 , θ2 , · · · , θL . Without loss of generality, assume −π/2 ≤
θ1 ≤ θ2 ≤ · · · ≤ θL ≤ π/2. The array consists of M
In this paper, we consider a different approach to improve number of antennas separated from each other by a distance
the quality of the sample covariance matrix estimate in the of d ≤ λ/2, where λ is the wavelength of the plane wave
finite sample size region. Partially inspired by the work of [9], impinging on the array. Let the steering vector of the array
where the root-MUSIC method and the least squares technique a(θ) ∈ CM×1 be defined as
are used iteratively in order to improve the estimations of
 T
Supported in part by the Natural Science and Engineering Research Council
a(θ)  1, e−j2π(d/λ) sin(θ) , · · · , e−j2π(M−1)(d/λ) sin(θ)
(NSERC) of Canada. (1)

978-1-4673-3146-3/13/$31.00 ©2013IEEE 53
 2013 5th IEEE International Workshop on Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP)

where (·)T stands for the transposition operator. At time instant Comparing (7) with (4), it can be observed that the expansion
k ∈ N, the received vector x(k) ∈ CM×1 is given by  consists of four terms while the model for R comprises
of R
two summands. The first two terms of R  given by (7) can

L
x(k) = a(θi )si (k) + n(k) (2) be considered as estimations for the two summands of R,
i=1 which represent the signal and noise components, respectively.
The last two terms of R  in (7) are undesirable by-products
where si (k) ∈ C is the amplitude of the i-th wave (source)
which can be viewed as estimations for the correlation between
and n(k) ∈ CM×1 is the noise vector at time k. By ar-
the signal and noise vectors. In the system model under
ranging the amplitudes of the sources in the vector s(k) =
study, we consider the noise vectors to be zero-mean and
[s1 (k), s2 (k), · · · , sL (k)]T ∈ CL×1 and forming the Van- also independent of the signal vectors. Therefore, the signal
dermonde matrix A = [a(θ1 ), a(θ2 ), · · · , a(θL )] ∈ CM×L , and noise components are uncorrelated to each other. As a
the model (2) can be rewritten in matrix-vector form as result, for large enough number of samples N , the last two
x(k) = As(k) + n(k). (3) terms in (7) tend to zero. However, it is favorable in practical
applications to estimate the DOAs based on a finite number
We assume that the sources are uncorrelated. Furthermore, of samples, so that real-time decisions can be made as fast
we consider the noise vector n(k) to be independent from as possible. In this case, the last two terms in (7) may have
the sources and other noise vectors and to have the circularly significant values, which causes the estimations of the signal
symmetric complex normal distribution NC (0, σn2 I M ) where and noise subspaces to deviate from the true signal and noise
I M is the identity matrix of size M . subspaces.
The root-MUSIC algorithm estimates the DOAs based on
One possible approach to remedy the effect of the unde-
the structure of the covariance matrix of the received signal.
sirable components is to consider the eigenvalue perturbation
Considering the system model (3), the data covariance matrix
caused by these terms. The incorporation of this knowledge
R ∈ CM×M is given by
into the estimation method may result in better estimations of
R  E{x(k)x(k)H } = ASAH + σn2 I M (4) the signal and noise subspaces. In this paper, we tackle the
H problem by reducing the effect of the undesirable terms in an
where S = E{s(k)s(k)H } ∈ CL×L , and (·) and E{·} stand iterative manner.
for the Hermitian transposition and the expectation operators,
respectively. The steps of the proposed method are listed in Table
I. The algorithm starts by computing the sample covariance
Let N number of snapshots (samples) be available. The
conventional method for estimating the covariance matrix from matrix R (0) where the superscript (·)(0) refers to the estimation
the samples x(k) (1 ≤ k ≤ N ) is given by made at the initialization step. The “←” notation means that
the value obtained from the right-hand side is assigned to

N
  1
R x(k)x(k)H (5)
the left-hand side. Next, the initial values for the DOAs are
N estimated using the root-MUSIC algorithm based on R  (0) .
k=1
Then, the iterative process begins. First, the Vandermonde
where R  ∈ CM×M is the sample data covariance matrix. Next, matrix is formed using the previous estimations of the DOAs.
the eigenvalues of R are computed and sorted in nondecreasing The superscript (·)() refers to the estimation made at the -
order. The eigenvectors corresponding to the first (M − L) th iteration. Next, we estimate the amplitudes of the sources
eigenvalues are then arranged as the columns of a matrix such that the squared norm of the difference between the
denoted by G ∈ CM×(M−L) . The range space of G  represents observation and the estimation is minimized. This problem can
the estimation of the noise subspace. Recalling (1) and defining be formulated as
z  ej2π(d/λ) sin(θ) , the steering vector can be rewritten as ()
 T 
ŝ(k)() = arg min x(k) − A s22 . (8)
s
a(z) = 1, z −1 , · · · , z −(M−1) . (6)
The minimization of (8) can be done using the least squares
For the root-MUSIC method, the roots of the equation (LS) technique and the corresponding solution is
G
aT (z −1 )G  H a(z) = 0 are used to obtain the estimates of  −1

H
H
the DOAs denoted by θ̂1 , θ̂2 , · · · , θ̂L [4]. ()
ŝ(k) = 
A
()

A
()
A ()
x(k). (9)
III. P ROPOSED M ETHOD
The noise component is then estimated as the difference
Let us start by expanding (5) using (3) as follows
between the estimated signal and the observation made by the

N array.
= 1
R (As(k) + n(k)) (As(k) + n(k))
H
N After estimating the signal and noise vectors, we estimate
k=1
1   the undesirable last two terms in (7) and deduct them from
1 
N N
=A s(k)s(k)H AH + n(k)n(k)H + the initial sample covariance matrix R  (0) . Note that, in the
N N algorithm, the estimations of the undesirable terms are scaled
k=1 k=1
1 
N  1 
N  down by a scaling factor γ which is considered to be a real
A s(k)n(k)H + n(k)s(k)H AH .(7) number between zero and one. This scaling factor is introduced
N N due to the fact that the estimations of the undesirable terms are
k=1 k=1

54
 2013 5th IEEE International Workshop on Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP)

TABLE I. I TERATIVE ROOT-MUSIC A LGORITHM The interelement spacing is set to d = λ/2 and the num-
ber of snapshots is N = 10. Each source vector s(k) is
Inputs: considered to be uncorrelated with other source vectors and
M, d, λ, N, L, γ, and to have the circularly symmetric complex normal distribution
received vectors x(1), x(2), · · · , x(N ) NC (0, σs2 I L ). The
 signal-to-noise
 ratio (SNR) is defined as
Outputs: SNR  10 log10 σs2 /σn2 .
Estimations θ̂1 , θ̂2 , · · · , θ̂L The performance of the proposed algorithm is compared
initialization: with that of the G-MUSIC and root-MUSIC methods consid-
R (0) ← 1 N x(k)x(k)H ering the mean squared error (MSE), the detection probability,
 N k=1


(0) and the conditional mean squared error (CMSE). The detection
(0) (0) (0) 
θ̂1 , θ̂2 , · · · , θ̂L ← root-MUSIC R , L probability is defined as the probability of successful detection
←1 which is in turn defined as the event that both DOA’s are
repeat estimated within one degree of their corresponding true values,
    i.e., the difference between the true value of each DOA and its
 () ← a θ̂(−1) , a θ̂(−1) , · · · , a θ̂(−1)
A 1 2 L estimated value is less than 1×(π/180). The conditional mean
for k = 1, 2, · · · , N squared error (CMSE) is then defined as the expected value of

H −1
H
() ()
theestimation error conditioned
 on successful
 detection, i.e.,
()
ŝ(k) ← 
A 
A A () x(k) L () 2 
E i=1 θi − θ̂i 2  successful detection . The reason for
() defining the CMSE in the above manner is to further inves-
n̂(k)() ← x(k) − A  ŝ(k)()
tigate the accuracy of the algorithms after making successful
end  
  detection. We estimate the MSE, the detection probability, and
 () N
()
U ←A 1
k=1 ŝ(k)
()
n̂(k)() H the CMSE using the Monte Carlo method with 105 number of
N
 H  trials. The iterative root-MUSIC algorithm is conducted for 3
() (0)
 
R ←R −γ U + U () () iterations, and the results are compared with the G-MUSIC

and root-MUSIC methods.
  ()
() () ()
θ̂1 , θ̂2 , · · · , θ̂L ← root-MUSIC R , L  We first investigate the performance of the proposed algo-
rithm by considering the MSE versus SNR as shown in Fig. 1.
←+1 It can be seen that the proposed method offers an SNR gain
until stopping criterion is satisfied of about 1 dB over the root-MUSIC method after the first
Assign the last estimated DOAs to θ̂1 , θ̂2 , · · · , θ̂L . iteration. Further improvement of the MSE after the second
and the third iterations is negligible. The G-MUSIC method
has a poor performance in this scenario.
not perfect, and therefore, these estimations are scaled down The detection probability is investigated next. Fig. 2 depicts
before being subtracted from the sample covariance matrix. the detection probability versus SNR. It can be seen that the
Next, given the updated sample covariance matrix R  () , the iterative root-MUSIC algorithm outperforms the root-MUSIC
DOAs are estimated using the root-MUSIC method and are method over a wide range of SNR values by about 1 to 2 dB
fed back to the next iteration. The value of γ is chosen after the first iteration. The detection probability after the
such that the spectral norm of the undesirable terms in the second and the third iterations remains almost the same as
H 
 (2) (2)  the result of the first iteration.
second iteration, i.e., U + U  is minimized. This
2 Finally Fig. 3 illustrates the performance of the algorithms
minimization is performed by considering different values for
in terms of the CMSE. It can be seen that the iterative root-
γ at the first iteration (e.g. γ = 0.1, 0.2, · · · , 1) and making
MUSIC algorithm outperforms the root-MUSIC method with
DOA estimations for each value of γ. Then, each set of the
an improvement of more than 2 dB at the low-SNR values
DOA estimations corresponding to each value of γ is tested
region. Unlike the previous two experiments where there has
in the second iteration, and the value of γ which results in the
been unnoticeable improvement after the first iteration, the
minimum norm of the undesirable terms is selected. The same
CMSE improves at the second iteration compared to the first
γ is used for the rest of the iterations.
iteration. This indicates that more accurate estimations can be
The algorithm continues until some stopping criterion is made at the second iteration. The result after the third iteration
satisfied. For example, the iterative process can be stopped is almost the same as the output of the second iteration.
when a predetermined fixed number of iterations is performed From the above simulations, it can be observed that the
or
Lthe distance between two consecutive estimations, i.e., iterative root-MUSIC algorithm introduces less estimation er-
(−1) () 2
i=1  θ̂ i − θ̂ i 2 is less than a given threshold value. ror compared to the G-MUSIC and root-MUSIC methods.
Moreover, the proposed algorithm offers a better detection
IV. S IMULATION R ESULTS probability, and when the detection is successful, the iterative
root-MUSIC algorithm is more precise than the G-MUSIC and
In this section, the performance of the proposed iterative
root-MUSIC methods.
root-MUSIC algorithm is investigated. We consider L = 2
sources impinging on an array of M = 10 antenna elements In the above experiments, the value of γ is selected once,
from directions θ1 = 35 × (π/180) and θ2 = 37 × (π/180). and is kept unchanged after the first iteration. We performed

55
 2013 5th IEEE International Workshop on Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP)

0 −34

−36
−10
−38

−20 −40

CMSE
−42
MSE

−30
−44

−40 −46
G−MUSIC G−MUSIC
root−MUSIC −48 root−MUSIC
−50 I−root−MUSIC, iter = 1 I−root−MUSIC, iter = 1
I−root−MUSIC, iter = 2 −50 I−root−MUSIC, iter = 2
I−root−MUSIC, iter = 3 I−root−MUSIC, iter = 3
−60 −52
6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34
SNR (dB) SNR (dB)
Fig. 1. MSE versus SNR after 1, 2, and 3 iterations for the iterative root- Fig. 3. CMSE versus SNR after 1, 2, and 3 iterations for the iterative root-
MUSIC (I-root-MUSIC) algorithm compared with the G-MUSIC and root- MUSIC (I-root-MUSIC) algorithm compared with the G-MUSIC and root-
MUSIC methods. MUSIC methods.

1
and noise subspaces. An iterative approach has been introduced
0.9 to remedy this problem, and simulation results have shown the
effectiveness of the proposed method. It has been shown that
0.8 the introduced method has a better performance in terms of the
mean squared error, it has a higher detection probability, and
Detection Probability

0.7 it offers more accurate estimations compared to the G-MUSIC

and root-MUSIC methods.
0.6

0.5
R EFERENCES
[1] H. L. Van Trees, Optimum Array Processing: Detection, Estimation, and
0.4 Modulation Theory. Part IV, New York: Wiley, 2002.
[2] D. G. Manolakis, V. K. Ingle, and S. M. Kogon, Statistical and Adap-
0.3 tive Signal Processing: Spectral Estimation, Signal Modeling, Adaptive
G−MUSIC Filtering and Array Processing. Boston, MA: McGraw-Hill, 2000.
root−MUSIC
0.2 [3] R. O. Schmidt, “Multiple emitter location and signal parameter estima-
I−root−MUSIC, iter = 1 tion,” IEEE Trans. Antennas Propagat., vol. AP-34, no. 3, pp. 276–280,
0.1 I−root−MUSIC, iter = 2 Mar. 1986.
I−root−MUSIC, iter = 3 [4] A. J. Barabell, “Improving the resolution performance of eigenstructure-
0 based direction-finding algorithms,” in Proc. ICASSP, Boston, MA, Apr.
6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 1983, pp. 336–339.
SNR (dB) [5] R. Roy and T. Kailath, “ESPRIT–Estimation of signal parameters via
Fig. 2. Detection probability versus SNR after 1, 2, and 3 iterations for the rotational invariance techniques,” IEEE Trans. Acoust., Speech, Signal
iterative root-MUSIC (I-root-MUSIC) algorithm compared with the G-MUSIC Processing, vol. 37, pp. 984–995, Jul. 1989.
and root-MUSIC methods. [6] X. Mestre, “Improved estimation of eigenvalues and eigenvectors of
covariance matrices using their sample estimates,” IEEE Trans. Inf.
Theory, vol. 54, no. 11, pp. 5113–5129, Nov. 2008.
further simulations with the value of γ being updated at [7] X. Mestre and M. A. Lagunas, “Modified subspace algorithms for DOA
each iteration. However, it did not offer further improvement estimation with large arrays,” IEEE Trans. Signal Process., vol. 56, no. 2,
compared to the case that the same γ is used at all of the pp. 598–614, Feb. 2008.
iterations. [8] P. Vallet, P. Loubaton, and X. Mestre, “Improved subspace estimation
for multivariate observations of high dimension: the deterministic signal
case,” IEEE Trans. Inf. Theory, vol. 58, no. 2, pp. 1043–1068, Feb. 2012.
V. C ONCLUSION [9] M. Shaghaghi and S. A. Vorobyov, “Improved model-based spectral
compressive sensing via nested least squares,” in Proc. ICASSP, Prague,
We have introduced a new high-resolution subspace-based Czech Republic, May 2011, pp. 3904–3907.
method for DOA estimation. The proposed method is based [10] Z. Tian, “Iterative music: coherent signal estimation, performance
on the fact that in the case of finite number of samples, the analysis,” in Proc. ICASSP, Orlando, FL, USA, May 2002, pp. III-3041–
undesirable terms in the sample covariance matrix can have III-3044.
significant values and cause deviations in the estimated signal

56