1548 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 62, NO.

6, MARCH 15, 2014

Real-Valued MUSIC for Efficient Direction

Estimation With Arbitrary Array Geometries
Feng-Gang Yan, Member, IEEE, Ming Jin, Shuai Liu, and Xiao-Lin Qiao

Abstract—Most of the existing methods for direction-of-arrival rotational invariance techniques (ESPRIT) [4], root-MUSIC
(DOA) estimation are based on numerical characteristics be- [5] and their derivations solve the problem of DOA estimation
hind the entire array output covariance matrix (AOCM). Since based on subspace decomposition which is usually accom-
the AOCM is generally a complex matrix, those approaches
plished by eigenvalue decomposition (EVD) or singular value
require tremendous complex computations accordingly. This
paper addresses the problem of DOA estimation with real-valued decomposition (SVD) on the AOCM. Since the inputs of sensor
computations by considering the real part of AOCM (R-AOCM) arrays are generally complex-valued signals, AOCM is a com-
and the imaginary part of AOCM (I-AOCM) separately. It is plex matrix accordingly. Therefore, those approaches require
shown that the null space of R-AOCM and that of I-AOCM are complex-valued operations, which may be computationally
the same subspace, which coincides with the intersection of the expensive for real-time applications.
original noise subspace and its conjugate subspace. Using such To reduce the complexity, real-valued (or unitary) estimators
a mathematical fact, a novel real-valued MUSIC (RV-MUSIC)
including U-MUSIC [6], U-ESPRIT [7], U-root-MUSIC [8],
estimator with a real-valued subspace decomposition on only
R-ACOM (or I-AOCM) instead of the entire ACOM is derived. unitary method of direction-of-arrival estimation (U-MODE)
Compared with most state-of-the-art unitary algorithms suitable [9] and unitary matrix pencil (U-MP) [10] usually exploit uni-
for only centro-symmetric arrays (CSAs), the proposed technique tary transformations [11] and forward/backword (FB) averaging
can be used with arbitrary array geometries. Unlike conventional [12] to transform AOCM to a real matrix, then estimate source
MUSIC with exhaustive spectral search, RV-MUSIC involves a DOAs with real-valued computations. Since one multiplication
limited search over only half of the total angular field-of-view with between two complex values generally requires four times that
a real-valued noise subspace, and hence reduces the complexity by
75%. Theoretical performance analysis on the mean square error
between two real ones, unitary methods can reduce about 75%
(MSE) and numerical simulations demonstrate that RV-MUSIC computational burdens as compared to their complex-valued
shows a very close accuracy to the standard MUSIC. versions. Another outstanding advantage of unitary algorithms
is that they also show improved accuracies as compared to com-
Index Terms—Arbitrary array geometries, direction-of-arrival
(DOA) estimation, real-valued MUSIC (RV-MUSIC), spectral plex-valued approaches. For example, it has been shown in [6]
search, subspace decomposition. that U-MUSIC has optimal Hermitian per-symmetric estimator
of AOCM in the sense of Euclidean distance, and hence out-
performs the standard MUSIC. In [8], both theoretical anal-
I. INTRODUCTION ysis and real-world experiments with sonar- and ultrasonic-data
have demonstrated that U-root-MUSIC shows better accuracies

I N array signal processing, numerical results behind the

entire array output covariance matrix (AOCM) are ex-
tensively derived by different algorithms to estimate the
than the conventional root-MUSIC as well. Despite their in-
creased estimation accuracies with reduced costs [13], almost all
of the state-of-the-art unitary methods are based on centro-sym-
direction-of-arrival (DOA) of multiple narrow-band sources. metrical arrays (CSAs), which severely limits their applications.
For example, beamfroming approaches such as the conven- For efficient DOA estimation with arbitrary array configura-
tional beamformer (CBF) [1] and Capon’s minimum variance tions, there have been several promising attempts to extend the
distortionless response (MVDR) beamfomer [2] use the ACOM concept of root-MUSIC to arbitrary array geometries. Tech-
and its inverse to find signal DOAs, respectively. On the other niques such as array interpolation (AI) [14] and beamspace
hand, subspace-based methods including multiple signal clas- transformation [15] have been developed to map the steering
sification (MUSIC) [3], estimation of signal parameters via vector of a nonuniform array (NUA) to that of a uniform
linear array (ULA) with Vandermonde structure, and then find
Manuscript received March 30, 2013; revised July 03, 2013 and December DOAs via polynomial rooting. Another two recently reported
15, 2013; accepted December 15, 2013. Date of publication January 09, 2014; approaches extending root-MUSIC to arbitrary arrays are the
date of current version February 26, 2014. The associate editor coordinating the
review of this manuscript and approving it for publication was Prof. Jean Pierre manifold separation technique (MST) [16] and Fourier-domain
Delmas. (FD) root-MUSIC [17]. For AI, it usually introduces mapping
F.-G. Yan is with the Department of Electronics and Information Engineering, errors which may cause increased bias [18] and excess vari-
Harbin Institute of Technology, Harbin 150001, China (e-mail: yfglion@gmail.
ance [19] while for MST and FD root-MUSIC, polynomial
com).
M. Jin, S. Liu, and X.-L. Qiao are with the Department of Electronics rooting with a sufficiently high order [17] is generally needed
and Information Engineering, Harbin Institute of Technology, Weihai to warrant that the truncation errors are small, which may cause
264209, China (e-mail: [email protected]; [email protected]; higher additional complexities than expected [20].
[email protected]).
Color versions of one or more of the figures in this paper are available online
We have shown in [21] that the exhaustive spectral search
at http://ieeexplore.ieee.org. involved in the conventional MUSIC can be compressed to a
Digital Object Identifier 10.1109/TSP.2014.2298384 limited angular sector instead of the total angular field-of-view

1053-587X © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
YAN et al.: REAL-VALUED MUSIC 1549

by a compressed MUSIC (C-MUSIC) estimator which exploits II. PRELIMINARIES

the multiple orthogonality between the steering vector and the
intersection of a noise-like subspace cluster (NLSC), in which A. Signal Model and Subspace Decomposition
both tasks of subspace decomposition and spectral search are ac- A standard signal model for DOA estimation using a linear
complished by complex-valued computations. Thus, estimating array composed of omnidirectional sensors is given by [3]
source DOAs via real-valued computations with arbitrary array
geometries still remains unresolved.
In this paper, we present a novel real-valued MUSIC (RV-
MUSIC) estimator which finds signal DOAs by spectral search
over only half of the total angular field-of-view with a real- (3)
valued noise subspace, which can be used with no dependence where is the numbers of sources, is the number of snap-
on array structures. The developed method also shows increased
accuracies with reduced computational burdens as compared to shots. is the
C-MUSIC since the former can be taken as a real-valued ver- matrix of the signal direction vectors, is
sion of the latter. Another newly developed result in the present the angle set of unknown source DOAs, and
work shows that the real part of AOCM (R-AOCM) and the
imaginary part of AOCM (I-AOCM) can be used separately for
DOA estimates, whereas attentions used to be paid on the entire
AOCM by most existent methods.
Throughout the paper, matrices and vectors are denoted by (4)
upper- and lower-boldface letters, respectively. Complex- and
real-matrices are denoted by single-bar- and double-bar-upper is the steering vector, is center wavelength, is
boldface letters, respectively. In addition, the mathematical no- the coordinate of sensor , and ,
tations are denoted as follows and are the array output-, source waveform-, and
zero matrix (vector); additive Gaussian white noise (AGWN)-vectors, respectively.
The EVD of the entire AOCM
identity matrix;
th column of matrix ; (5)

th row of matrix ; can be expressed as

intersection; (6)
direct sum; where is source covariance matrix, and
conjugation; are the so-called signal- and noise-subspace matrices, respec-
tively. For practical situations, the theoretical AOCM in (5) is
transpose; unavailable, and it is usually estimated by
Hermitian transpose;
(7)
round down to integer;
Frobenius norm; Thus, the subspace decomposition is in fact given by
Kronecker delta;
(8)
inner product;
The subspace decomposition can also be obtained by per-
mathematical expectation; forming SVD on a direct-data matrix
real part of the embraced matrix; (9)
imaginary part of the embraced matrix;
as follows
rank of the embraced matrix.
(10)
For , and denote the null- and
column-spaces of , respectively, given by In a noise environment, the idea given by (9) is perturbed
by AGWN as , where
is the matrix of AGWN.
(1) Therefore, the SVD of is given by

(2) (11)

Noting that , we obtain from (8) and (11) that

where , are unknown coefficients. (12)

1550 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 62, NO. 6, MARCH 15, 2014

which implies that the subspace decomposition by the EVD of Remark 1: Assumption A1 is a standard restriction in the
is equivalent to that by the SVD of [21], [22]. literature on most DOA estimators including MUSIC [3]. For
non-coherent signals, this assumption can be easily satisfied.
B. Related Works Assumption A2 is commonly referred to as the so-called
Using some facts and rank- ambiguity restriction on array geometry (see
, the conventional MUSIC algorithm [3] suggests to [23], [24] and the references therein), which must be satisfied
estimate source DOAs by spectral search as follows in practical applications. It is worth noting that this assumption
is a sufficient condition for estimating signal DOAs without
(13) ambiguity since vectors are linearly
independent implied guarantees that , vectors
are linearly independent.
One of the most important advantages of the MUSIC algorithm
Although assumption A3 is too strict for conventional
is its easy implementation with arbitrary array configurations.
MUSIC, in which only is required, it is to be
However, since MUSIC involves a tremendous spectral search
shown shortly that this assumption allows a significant reduc-
step, it is computationally expensive for real-time applications.
tion on computational complexity as compared to MUSIC.
We have proposed another C-MUSIC estimator, which ex-
ploited NLSC (see [21] for detailed illustrations) to limit the
exhaustive spectral search of MUSIC to a small sector by B. Physical Analysis on AOCM, R-AOCM and I-AOCM

(14) For ideal and , we see clearly from (5) and (9) that

where is the intersection of NLSC, is the

compression time and is a small angular sector.
For MUSIC and C-MUSIC, both subspace decomposition
and spectral search require complex-valued computations. To
realize real-valued computations, the U-MUSIC [6] technique
transforms the standard MUSIC to a real-valued function as
(20)
(15)
Hence, R-AOCM and I-AOCM can be expressed as
where and are real vector and
matrix, respectively. Unfortunately, U-MUSIC can be used with
only CSAs, which severely limits its applications. (21)

III. THE PROPOSED ALGORITHM (22)

A. Basic Assumptions respectively. Noting that is composed of all the snapshots

The following basic assumptions on the data model (3) are of received datum, it must contain the entire information re-
considered to hold throughout the paper. garding source DOAs, which is separately stored in and
A1. Matrix is of full rank such that . From such a physical point of view, and
can be taken as two basic factors that are both necessary and
(16) hence, must be used simultaneously for DOA estimates. This
means that using only or may fail to find source
A2. The array steering vectors associated with any dif- DOAs since half of the original information is lost. However,
ferent angles , , i.e, this doesn’t mean that exploiting both and al-
, , are linearly independent. In other ways warrants success in DOA estimates.
words, the following equation Intuitively, due to the presences of both and
in (21) and (22), the information contained in either R-AOCM
(17) or I-AOCM is enough for finding source DOAs since both the
two basic factors are not lost in R-AOCM or I-AOCM. Hence,
it could be predicted that R-AOCM and I-AOCM can be used
holds if and only if separately, which hence gives a significant reduction on com-
plexity for computationally efficient DOA estimations.
(18)
Moreover, with the above intuitive explanations, differ-
ences between the accuracies of DOA estimators based on
where , are unknown scalars.
the entire AOCM and of those based on only R-AOCM or
A3. The number of signal DOAs is smaller than half that
only I-AOCM can be also predicted. Observing that R-AOCM
of sensors such that
contains the two second-order terms and
without the cross term while
(19)
on the contrary, I-AOCM contains only the cross term without
YAN et al.: REAL-VALUED MUSIC 1551

the two second-order terms, estimators using the entire AOCM which implies that is a symmetrical real-valued matrix,
may slightly outperform those using only R-AOCM or only whose EVD and SVD must require only real-valued computa-
I-AOCM. tions [6]. Therefore, we must have
In the sections to follow, we shall show in detail how to find
source DOAs by using R-AOCM and I-AOCM separately. (29)

C. Subspace Decompositions on R-AOCM and I-AOCM According to (24), it is clear that the dimension of must
be (see [21], where is also similar to
Let us consider the intersection of the original noise subspace the intersection of NLSC with two angular sectors). Hence, the
and the conjugate noise subspace , which is EVD (or SVD) of can be written as
given by
(30)
(23)
where subscripts and stand for the signal- and the
Since is a subset of , it also contains a part of noise-subspace, respectively, and and are two real
the vectors of . Therefore, we can use instead diagonal matrices composed of the significant- and the
of to estimate source DOAs. zero-eigenvalues of , respectively. Since
An important advantage of over is that the , it is clear that the column
former has a double orthogonality to at both the true space of must equal to the direct sum of the original signal-
DOAs and their mirror directions simultaneously, i.e., and the conjugate signal-subspace. In other words, we have

(24) (31)

On the other hand, the orthogonal basis of can be

This is because holds for arbitrary linear also computed by the subspace on equivalently. Noting
array geometries. Such a double orthogonality can help us limit that , we have
the exhaustive spectral search to only half of the total angular
field-of-view, and hence reduces a significant computational (32)
complexity as compared to the standard MUSIC.
Another outstanding advantage of over
which implies that is an antisymmetrical real-valued
is that the orthogonal basis of the former can be computed by
matrix. According to matrix theory [25], the SVD of
subspace decompositions on R-AOCM and I-AOCM with only
must require only real-valued computations (note that the EVD
real-valued computations while that of the latter resulted from
of may involve complex-valued computations [25]).
the EVD of the entire AOCM usually requires complex-valued
Therefore, we can write the SVD of as follows
computations. To see clearly about this, we give the following
theorem which reveals the relationship between , the
(33)
null space of R-AOCM and that of I-AOCM.
Theorem 1: Under assumptions A1 A3, we have where again, subscripts and denote the signal- and the noise-
subspace, respectively, and and are two real diagonal
(25) matrices composed of the significant- and the
zero-singular values of , respectively. Using the facts
where [21] and [25] gives

(26) (34)

Proof: See Appendix A. Hence, the columns of offer an orthogonal basis for the null
It follows directly from theorem 1 that , we space of . According to theorem 1, matrix can be
must have , which means that , further taken as an orthogonal basis of equivalently,
and we further have and we have

(27) (35)

Equation (27) can be identified as the characteristic one for the Thus, the SVD of can be rewritten as
real-valued matrix . Therefore, is an eigenvalue of
and is the eigenvector associated with . Since is (36)
an arbitrary vector of , matrix can be computed by
Thus, , and matrix can also be computed by the
the subspace decomposition on accordingly. Noting that
SVD of with only real-valued computations.
, we have
The above analysis reveals the relationship among subspace
decompositions on AOCM, conjugate AOCM, R-AOCM and
(28) I-AOCM, which is shown in Fig. 1 for clear illustrations.
1552 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 62, NO. 6, MARCH 15, 2014

By introducing two electronic angles

(39)

is transformed into as follows

Fig. 1. Relationship among subspace decompositions on AOCM, conjugate (40)

AOCM, R-AOCM and I-AOCM.

Consequently, we have , with which the

2-D steering vector also shows a double orthogonality to the
real-valued noise matrix . Thus, we can similarly exploit the
2-D-RV-MUSIC estimator. Noting that the definitions of and
in (39) can be changed with array geometries, RV-MUSIC
has no dependence on array configurations.
Detailed steps for implementing the proposed RV-MUSIC al-
gorithm are summarized in Algorithm 1.

Fig. 2. 2-D array on the X-Y plane, which is composed of omnidirec- Algorithm 1 The proposed RV-MUSIC algorithm
tional sensors with arbitrary array geometries.
Require: : snapshots of array output vector.
D. The Proposed Method 1: Initiation: , , .
With computed by the EVD of (or by the SVD of
2: for ; ; do
), the proposed RV-MUSIC estimator is given by
3: ;
4: end for
(37) 5: Compute ,
According to (24), the minima of over only (or );
half of total angular field-of-view are either the true DOAs or 6: for each (or ) do
their images. Because the steering vector is orthogonal to
the original noise subspace at only the true incident angles, 7: , ;
responding to the true DOAs are much larger than 8: end for
those associated with symmetrical mirror DOAs. Moreover, as
the number of the true DOAs, i.e., , is known in advance, we 9: for ; ; do
can use the standard MUSIC to select the true DOAs among 10: if then
the candidate angles by minimizing such that
estimation ambiguity is avoided. 11: ;
Although using the standard MUSIC to exclude the symmet- 12: else if then
rical mirror DOAs means that there is an additional EVD step
on involved in the proposed estimator. However, the com- 13: ;
plexity of this step is substantially lower than that of spectral 14: else
search [17], [21] since we only need to compute the product
for at most spectral points. 15: ;
Remark 2: The proposed RV-MUSIC algorithm can be di- 16: end if
rectly extended to estimate the two-dimensional (2-D) signal
17: end for
directions with ar-
bitrary plane array geometries. To demonstrate this clearly, we 18: return : a estimated angle set of the source DOAs.
take the plane array in Fig. 2 for example, with which
the 2-D steering vector is given by Comparisons of the primary real-valued computational com-
plexities of various algorithms are shown in Table I, where
stands for the total sample points of the standard MUSIC spa-
tial spectrum over . The common term
included in all the six algorithms gives the cost for computing
the subspace decomposition on a real-valued matrix of dimen-
(38) sions by using the fast subspace decomposition (FSD)
YAN et al.: REAL-VALUED MUSIC 1553

TABLE I Proof: See Appendix B.

COMPARISONS OF REAL-VALUED COMPUTATIONAL COMPLEXITY According to theorem 2 where is similar to , can be
also computed by the SVD of equivalently. Since
and , the SVD of can be written as

(44)

where and are composed

of the singular vectors of associated with the non-zero-
and zero-singular values respectively. With a similar
proof to that of theorem 2, it can be easily shown that .
Therefore, is given by the following theorem.
Theorem 3: Assume the elements of matrix are random
variables with zero means, then the perturbation of at high
SNR can be expressed by a linear function of matrix as fol-
lows

(45)

technique [26]. Another term involved in MVDR where

[2] gives the complexity of computing the inverse of [25].
(46)
Since the spectral search of the proposed RV-MUSIC is in-
volved over (or ) with , it Proof: See Appendix C.
only requires for computing its null Using the result of lemma 1 and that of theorem 3, a
spectrum. On the other hand, the standard MUSIC uses closed-form MSE expression for DOA estimation by the pro-
to compute its null spectrum over , it posed RV-MUSIC estimator is given by the following theorem.
costs for the spectral search step. Theorem 4: Assume the elements of matrix are random
Noting that [17], it is observed from Table I that variables with zero means and variance , then the MSE for the
about 75% computational complexities are reduced by the pro- estimation of incident angles by RV-MUSIC
posed estimator as compared to MUSIC. Hence, RV-MUSIC at high SNR is given by
shows a similar efficiency to the U-MUSIC [6] algorithm. It
is also seen from Table I that RV-MUSIC has much lower
complexities than C-MUSIC [21] with two angular sectors (47)
since complex-valued computations are required in C-MUSIC
for both subspace decomposition and spectral search. where
IV. PERFORMANCE ANALYSIS (48)
In this section, we use the theory of subspace perturbation and (49)
Taylor’s expansion series to derive a closed-from expression for (50)
the mean square error (MSE) of direction estimation by the pro-
posed RV-MUSIC estimator base on a high signal-to-noise ratio with denoting the first-derivative of
(SNR) assumption. Under such an assumption, the perturbation with respect to .
of is given by the following lemma [22]. Proof: See Appendix D.
Lemma 1: Assume the elements of matrix are random vari-
ables with zero means, then the perturbation of at high SNR V. SIMULATIONS
can be expressed by a linear function of matrix as follows
Computer simulations with 500 independent Monte Carlo
(41) trials are conducted to assess the performance of the proposed
estimator and to verify the derived MSE expression of DOA
As shown in Section III that can be computed by the EVD estimates by RV-MUSIC. For the performance comparison,
of (or by the SVD of ). However, it is difficult to the C-MUSIC algorithm with and angular
establish the perturbation of based on the EVD of (or sectors [21] and the unconditional Cramér-Rao Lower Bound
based on the SVD of ). To derive an expression for , (CRLB) given in [28] are also applied for references. For all
we give another theorem as follows. search-based algorithms, a coarse gird 1 was firstly used to get
Theorem 2: Using the original signal subspace matrix to candidate peaks, and a fine one 0.0053 was secondly applied
define a real matrix as follows around the candidate peaks for final DOA estimates.
First, we use a ULA of half-wavelength spaced
(42)
sensors to compare the root MSEs (RMSEs) of the proposed
Then, we have RV-MUSIC estimator with those of different algorithms, in-
cluding ESPRIT, conventional MUSIC, Capon’s MVDR, root-
(43) MUSIC, U-MUSIC [6], and C-MUSIC with sectors.
1554 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 62, NO. 6, MARCH 15, 2014

Fig. 3. RMSE for versus the SNR, where and Fig. 5. Experimental- and theoretical-RMSEs for versus the SNR,
sources at and are used on a ULA of sensors. where , sources at and are used on ULAs.

Fig. 4. RMSE for versus the numbers of snapshots, where the

and sources at and are used on Fig. 6. Experimental- and theoretical-RMSEs for versus the num-
a ULA of sensors. bers of snapshots, where , sources at and
are used on ULAs.

Fig. 3 demonstrates the RMSEs against the SNR, where the number of snapshots, where the number of snapshots varies over
SNR varies over a wide range from 20 dB to 30 dB. It can a wide range from to . It can be seen again from
be concluded from the figure that U-MUSIC, despite its lim- Fig. 4 that the proposed technique performs similarly to the stan-
ited applications for only CSAs, is the most accurate one among dard MUSIC and to C-MUSIC with angular sectors. It
the presented seven algorithms, which shows a RMSE closest can be also seen that RV-MUSIC shows much better perfor-
to the CRLB. It can be also seen from the figure that the stan- mances than ESPRIT and Capon’s MVDR beamformer, espe-
dard MUSIC, root-MUSIC, C-MUSIC with angular sec- cially in scenarios with small numbers of snapshots.
tors as well as the proposed method show similar performances Fig. 5 and Fig. 6 compare RMSEs of DOA estimation by
to each other, which is much better than ESPRIT and Capon’s the proposed RV-MUSIC with different numbers of antennas,
method. Noting that the accuracy of C-MUSIC decreases as in terms of both based on the experimental- and the theoretical-
increases and C-MUSIC with angular sectors reduces results given by (47). We can observe clearly from the figurers
the complexity by only 50% [21] while RV-MUSIC saves that that the simulated results and the analytic expectations agree
by about 75%, the proposed method hence shows improved ac- with each other closely when , which verifies the
curacy with reduced complexity as compared to the C-MUSIC theoretical analysis in Section IV.
technique. Fig. 7 and Fig. 8 plot the RMSEs as functions of the SNR
To see more clearly the performance of the new approach, and as those of the number of snapshots, respectively, where
Fig. 4 plots RMSEs of different algorithms as functions of the the number of sources varies from to the its maximum
YAN et al.: REAL-VALUED MUSIC 1555

Fig. 7. RMSE for versus the SNR, where and a ULA of Fig. 9. RMSE for versus the SNR, where and
sensors is used. sources at and are used on a MRLA of sensors.

Fig. 8. RMSE for versus the numbers of snapshots, where the

Fig. 10. RMSE for versus the numbers of snapshots, where the
and a ULA of sensors is used.
and sources at and are used on a
MRLA of sensors.

. As seen clearly
from the figures that the differences between the RMSEs of as functions of the numbers of snapshot, where the amounts of
RV-MUSIC and those of MUSIC increase as increases. This snapshot varies from to .
is mainly caused by the fact that the dimensions of are We see from Fig. 9 that the five techniques show very
whiles those of are , close accuracies to each other. More Specifically, the standard
thus the difference between the former and the latter increases MUSIC and the proposed method slightly outperform MDVR
as increases, and the relationship between RV-MUSIC and and C-MUSIC with sectors, especially for small numbers
MUSIC is similar to that between C-MUSIC and MUSIC [21]. of snapshots. On the other hand, C-MUSIC with angular
Second, we examine the performance of the proposed sectors shows close performances to the standard MUSIC as
approach with NUAs, where ESPRIT, U-MUSIC and the con- well as to Capon’s MVDR and the proposed method. From
ventional root-MUSIC cannot be exploited for DOA estimates Fig. 10, it is seen clearly that in scenarios of small amounts
any more. We use the minimum-redundancy linear arrays of snapshots, our method has much better accuracies than the
(MRLAs) [29] to compare Capon’s MDVR beamformer, the MVDR beamformer and C-MUSIC with angular sectors
standard MUSIC and C-MUSIC with as well as while for large numbers of snapshots, the proposed method
angular sectors. In the simulation, the unconditional CRLB is shows a similar performance to C-MUSIC with sectors.
also applied for comparison reference. Noting that C-MUSIC with or sector involves a
In Fig. 9, we fix the number of snapshots as and higher complexity than RV-MUSIC [21], the proposed method
display RMSEs of different algorithms as functions of the SNR provides an improved performance-to-complexity tradeoff as
while in Fig. 10, we fix the , and plot RMSEs compared to the C-MUSIC technique.
1556 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 62, NO. 6, MARCH 15, 2014

Fig. 11. Experimental- and theoretical-RMSEs for versus the SNR, Fig. 13. Resolution probability against the SNR, where and
where , sources at and are used on closely-spaced sources at and are used on a ULA with
MRLAs. sensors.

Fig. 12. Experimental- and theoretical-RMSEs for versus the num-

bers of snapshots, where , sources at and Fig. 14. Resolution probability against the number of snapshots, where the
are used on MRLAs. and closely-spaced sources at and
are used on a ULA with sensors.

To verify the theoretical MSE expression of DOA estimates

by RV-MUSIC with MRLAs, Fig. 11 and Fig. 12 compare the It is observed that U-MUSIC, as addressed in [6], has the highest
simulated RMSEs with those computed by (47) with different resolution among the four estimators. On the other hand, for low
numbers of antennas. It is seen clearly from the two figures that SNRs and small numbers of snapshots, the proposed method
there is a close match between the simulated results and their shows an improved resolution for two closely-spaced sources
theoretical expectations, especially for , which as compared to the standard MUSIC. It can be also concluded
verifies again the theoretical analysis in Section IV. from the two figures that C-MUSIC with sectors shows
Next, we use a ULA composed of half-wavelength a little better resolution than RV-MUSIC.
spaced sensors to compare resolution probabilities of different Finally, we use a ULA to plot the simulation times of DOA
algorithms in Fig. 13 and Fig. 14, including the conventional estimates by different algorithms as functions of the number of
MUSIC, U-MUSIC, C-MUSIC with sector and the pro- sensors in Fig. 15. The simulated results are given by a PC with
posed method. In the simulation, we use closely-spaced Intel(R) Core(TM) Duo T5870 2.0 GHz CPU and 1GB RAM
sources at and , which are said to be success- by running the Matlab codes in the same environment. It can be
fully resolved if and only if [30] seen from Fig. 15 that RV-MUSIC is the most efficient method
among the five algorithms with a simulation time being about 4
(51)
times lower than that of MUSIC.
YAN et al.: REAL-VALUED MUSIC 1557

which is equivalent to
(A.5)
According to (1) and (A.5), the columns of must belong to
. Thus, it follows directly from (2) that
(A.6)

On the other hand, assume that , then we

have . Using (5), can be rewritten as

(A.7)

Fig. 15. Simulation time against the number of sensors, where , the Therefore, we have
, sources at and are used on a ULA,
and are sampled by points with a grid 0.1 . (A.8)

where and . Expanding and

VI. CONCLUSIONS as weighted sums of the columns of and , respec-
We have proposed a novel efficient RV-MUSIC algorithm tively, (A.8) can be rewritten as
for DOA estimation with arbitrary array geometries in this
paper, which exploits the subspace decomposition on only (A.9)
R-AOCM (or I-AOCM) instead of the entire AOCM. This is
based on a newly developed result showing that the null space
of R-AOCM and that of I-AOCM in fact equal to the intersec- where the fact is used, and and are the -th
tion of the original noise subspace and its conjugate subspace. element of and , respectively. According to assumption A2,
Theoretical analysis and simulation results demonstrate that (A.9) holds if and only if , (noting
RV-MUSIC shows a similar performance to the standard that ), which means that
MUSIC while its complexity is about four times lower than
that of MUSIC. Compared with C-MUSIC which can be taken (A.10)
as a complex-valued version of RV-MUSIC, the real-valued
Because is invertible, it follows from (A.10) that
computations in RV-MUSIC also lead to increased accuracy
, and hence , . This
with reduced computations in some scenarios. Future research
implies that , and we have
includes to develop unitary estimators with arbitrary arrays as
well as with fewer sensors. (A.11)

APPENDIX A Combining (A.6) and (A.11), we finally have

PROOF OF THEOREM 1
Since , we must (A.12)
have and .
These facts together with and By subtracting (A.3) from (A.2), it can be similarly proved
lead to that , which completes the proof.

(A.1) APPENDIX B
Now, postmultiplying the left- and the right-side of (5) by as PROOF OF THEOREM 2
well as using (A.1), which gives It is beneficial to rewrite as follows
(A.2) (B.1)
Similarly, we also obtain the following equation by postmulti-
Now, assume , then we have and
plying the both sides of the conjugate version of (5) by
. Thus, we further have and
(A.3) , which leads to .
This implies that , and we finally have
Adding (A.2) and (A.3), we have
(A.4) (B.2)
1558 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 62, NO. 6, MARCH 15, 2014

On the other hand, suppose , then we have APPENDIX D

, which is equivalent to PROOF OF THEOREM 4
. Taking norm for the both sides of For the sake of notational simplicity, let us define
the above equation gives
. Because and are
the projections of onto and , respec- (D.1)
tively, we have and . (D.2)
Therefore, ,
which gives . Thus, (D.3)
, .
Hence, , and we finally have Using (45) and the zero means of additive noise, it can be easily
(B.3) proven with a high SNR assumption that

Combining (B.2) and (B.3) gives , which (D.4)

completes the proof.

APPENDIX C Hence, is a consistent estimate for at high

PROOF OF THEOREM 3 SNR, and we can obtain the second-order approximation of the
derivative of about the true value as follows
It follows from (B.1) that the perturbation of is given by (see [21], [27], and references therein)
(C.1)
(D.5)
where the second-order terms and are
neglected. Since is obtained by the SVD of , it can where higher-order terms are neglected, and the first- and
be similarly concluded from Lemma 1 that at high SNR can second-order derivatives of with respect to
be expressed by a linear function of as follows are denoted by
and , respectively. By
(C.2)
performing a forward derivation, it can be shown that
Inserting (C.1) into (C.2) as well as using (41), we have
(D.6)

(C.3) Inserting (C.8) into (D.6) and using the notations in (D.3) gives

As is the intersection of and , we

must have and .
Thus, the first term of (C.3) can be simplified as

(C.4)

Similarly, the third term of (C.3) is given by

(C.5) (D.7)

Therefore, (C.3) can be simplified as

Noting that and , it fol-
lows from (D.7) that
(C.6)

Noting that and are the orthogonal projections

onto and , respectively, we must have

(C.7)

Inserting (C.7) into (C.6), is finally given by (D.8)

(C.8)
Now, consider the first term of (D.8). Expanding
which completes the proof. and as weighted sums of the columns and the rows
YAN et al.: REAL-VALUED MUSIC 1559

of , respectively, and using the zero means of AGWN Substituting (D.9)–(D.13) into (D.8) as well as using the fact
noise , we finally obtain

(D.14)

which completes the proof.

ACKNOWLEDGMENT
The authors would like to thank the anonymous reviewers for
their many insightful comments and suggestions, which helped
(D.9) improve the quality and readability of this paper.
where and are the th and th element of and ,
REFERENCES
respectively; and and are the th row and the th column
[1] J. Krim and M. Viberg, “Two decades of array signal processing re-
of , respectively. In a similar way, we can prove that search: The parametric approach,” IEEE Signal Process. Mag., vol. 13,
no. 3, pp. 67–94, Jul. 1996.
(D.10) [2] J. Capon, “High-resolution frequency-wavenumber spectrum anal-
ysis,” Proc. IEEE, vol. 57, pp. 1408–1418, Aug. 1987.
According to (D.9) and (D.10), it is clear that the first term of [3] R. O. Schmidt, “Multiple emitter location and signal parameter estima-
tion,” IEEE Trans. Antennas Propag., vol. AP-34, no. 3, pp. 276–280,
(D.8) equals to zero. Mar. 1986.
Now, let us consider the second term of (D.8). Since the vari- [4] R. Roy and T. Kailath, “ESPRIT-estimation of signal parameters via
ance of noise is , we can similarly write that rotational invariance techniques,” IEEE Trans. Signal Process., vol. 37,
no. 7, pp. 984–995, Jul. 1989.
[5] B. D. Rao and K. V. S. Hari, “Performance analysis of root-MUSIC,”
IEEE Trans. Acoust., Speech, Signal Process., vol. 37, pp. 1939–1949,
Dec. 1989.
[6] K. C. Huarng and C. C. Yeh, “A unitary transformation method for
angle-of-arrival estiamtion,” IEEE Trans. Signal Process., vol. 39, pp.
975–977, April 1991.
[7] Z. Guimei, C. Baixiao, and Y. Minglei, “Unitary ESPRIT algorithm
for bistatic MIMO radar,” Electron. Lett., vol. 48, no. 3, pp. 179–181,
Feb. 2012.
[8] M. Pesavento, A. B. Gershman, and M. Haardt, “Unitary root-MUSIC
with a real-valued eigendecomposition: A theoretical and experimental
performance study,” IEEE Trans. Signal Process., vol. 48, no. 5, pp.
1306–1314, May 2000.
(D.11) [9] A. B. Gershman and P. Stoica, “On unitary and forward-backward
MODE,” Digit. Signal Process., vol. 9, no. 2, pp. 67–75, Feb. 1999.
[10] N. Yilmazer, J. Koh, and T. K. Sarkar, “Utilization of a unitary
transform for efficient computation in the matrix pencil method to
find the direction of arrival,” IEEE Trans. Signal Process., vol. 54, pp.
175–181, Jan. 2006.
[11] M. Haardt and F. Romer, “Enhancements of unitary ESPRIT for non-
circular sources,” in Proc. ICASSP, Feb. 2004, pp. 101–104.
[12] D. A. Linebarger, R. D. DeGroat, and E. M. Dowling, “Efficient direc-
tion-finding methods employing forward-backward averaging,” IEEE
Trans. Signal Process., vol. 42, no. 8, pp. 2136–2145, Aug. 1994.
[13] M. Haardt and J. A. Nossek, “Unitary ESPRIT: How to obtain in-
creased estimation accuracy with a reduced computational burden,”
IEEE Trans. Signal Process., vol. 43, no. 5, pp. 1232–1242, May 1995.
(D.12) [14] B. Friedlander, “The root-MUSIC algorithm for direction finding with
interpolated arrays,” Signal Process., vol. 30, pp. 15–29, 1993.
[15] C. P. Mathews and M. D. Zoltowski, “Eigenstructure techniques for
2-D angle estimation with uniform circular arrays,” IEEE Trans. Signal
Process., vol. 42, no. 9, pp. 2395–2407, Sep. 1994.
[16] F. Belloni, A. Richter, and V. Koivunen, “DoA estimation via manifold
separation for arbitrary array structures,” IEEE Trans. Signal Process.,
vol. 55, no. 10, pp. 4800–4810, Oct. 2007.
[17] M. Rbsamen and A. B. Gershman, “Direction-of-arrival estimation for
nonuniform sensor arrays: From manifold separation to Fourier domain
MUSIC Methods,” IEEE Trans. Signal Process., vol. 57, pp. 588–599,
Feb. 2009.
[18] P. Hyberg, M. Jansson, and B. Ottersten, “Array interpolation and
bias reduction,” IEEE Trans. Signal Process., vol. 52, no. 10, pp.
2711–2720, Oct. 2004.
[19] P. Hyberg, M. Jansson, and B. Ottersten, “Array interpolation andDOA
MSE reduction,” IEEE Trans. Signal Process., vol. 53, no. 12, pp.
4464–4471, Dec. 2005.
(D.13) [20] J. Zhuang, W. Li, and A. Manikas, “Fast root-MUSIC for arbitrary
arrays,” Electron. Lett., vol. 46, no. 2, Feb. 2010.
[21] F. G. Yan, M. Jin, and X. L. Qiao, “Low-complexity DOA estima-
where and (because tion based on compressed MUSIC and its performance analysis,” IEEE
) are used in the last step of (D.13). Trans. Signal Process., vol. 61, no. 8, pp. 1915–1930, Apr. 2013.
1560 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 62, NO. 6, MARCH 15, 2014

[22] F. Li, H. Liu, and R. J. Vaccaro, “Performance analysis for DOA Ming Jin received the B.E., M.S., and Ph.D. degrees
estimation algorithms: Unification, simplification, and observations,” in information and communication engineering from
IEEE Trans. Aerosp. Electron. Syst., vol. 29, no. 4, pp. 1170–1184, Harbin Institute of Technology (HIT), Harbin, China,
Oct. 1993. in 1990, 1998, and 2004, respectively.
[23] C. Proukakis and A. Manikas, “Study of ambiguities of linear arrays,” From 1998 to 2004, he was with the Department
in Proc. ICASSP, Apr. 1994, vol. 4, pp. 549–552. of Electronics Information Engineering, HIT. Since
[24] K. C. Tan and Z. Goh, “A detailed derivation of arrays free of higher 2006, he has been a Professor of The School of In-
rank ambiguities,” IEEE Trans. Signal Process., vol. 44, no. 2, pp. formation and Electricity Engineering, HIT, Weihai,
351–359, Feb. 1996. China. His current interests are in the areas of array
[25] G. H. Golub and C. H. Van Loan, Matirx Computations. Baltimore, signal processing, parallel signal processing, and
MD, USA: Johns Hopkins Univ. Press, 1996. radar polarimetry.
[26] Xu and Kailath, “Fast subspace decomposition,” IEEE Trans. Signal
Process., vol. 42, no. 3, pp. 539–551, Mar. 1994.
[27] P. Stoica and T. Soderstrom, “Statistical analysis of a subspace method
for bearing estimation without eigendecomposition,” Proc. Inst. Electr.
Eng., vol. 139, no. 4, pt. F, pp. 301–305, 1992. Shuai Liu was born in Heilongjiang Province,
[28] P. Stoica and A. Nehorai, “Performance study of conditional and China, in 1980. He received the B.E. and M.S.
unconditional direction-of-arrival estimation,” IEEE Trans. Acoust., degrees from Northwestern Polytechnical University
Speech, Signal Process., vol. 38, pp. 1783–1795, Oct. 1990. China, in 2002 and 2005, respectively, and received
[29] C. Chambers et al., “Temporal and spatial sampling influence on the the Ph.D degree in information and communication
estimates of superimposed narrowband signals: When less can mean engineering from Harbin Institute of Technology
more,” IEEE Trans. Signal Process., vol. 44, no. 12, pp. 3085–3098, (HIT), China, in 2013.
2004. Since 2013, he has been an Associate Professor
[30] Q. T. Zhang, “Probability of resolution of the MUSIC algorithm,” IEEE of The School of Information and Electricity Engi-
Trans. Signal Process., vol. 43, no. 4, pp. 978–987, Apr. 1994. neering, HIT, Weihai, China. His current interests are
in the area of conformal array and polarization sensi-
tive array signal processing.

Feng-Gang Yan (S’11–M’14) received the B.E., Xiao-Lin Qiao was born in the Inner Mongolia Au-
M.S., and Ph.D. degrees in information and commu- tonomous Region, China, in June 1948. He received
nication engineering from Xi’an Jiaotong University, the B.E., M.S., and Ph.D. degrees in information and
Xi’an, the Graduate School of Chinese Science of communication engineering from Harbin Institute of
Academic, Beijing, and Harbin Institute of Tech- Technology (HIT), Harbin, China, in 1976, 1983, and
nology (HIT), Harbin, in 2005, 2008, and 2013, 1991, respectively.
respectively. Dr. Qiao was with the Department of Electronics
From July 2008 to March 2011, he was a Research Information Engineering of HIT from 1983 to 1993.
Associate of the Fifth Research Institute of China Since 1994, he has been a Professor of The School
Aerospace Science and Technology Corporation of Information and Electricity Engineering, HIT.
(CASC), where his research was mainly focused During 1994–2011, he was the President of HIT,
on the processing of remote sensing images. Since October 2013, he is a Weihai, China. In the past 15 years, he has authored and co-authored nearly
Teacher with the Department of Electronics Information Engineering, HIT, 70 publications. His research interests are in the areas of signal processing,
Weihai, China. His current research interests include array signal processing wireless communication, special radar, parallel signal processing, and radar
and statistical performance analysis. polarimetry.