Proceedings of the International MultiConference of Engineers and Computer Scientists 2008 Vol II

IMECS 2008, 19-21 March, 2008, Hong Kong

Direction of Arrival Estimation using a

Root-MUSIC Algorithm
H. K. Hwang, Zekeriya Aliyazicioglu, Marshall Grice, Anatoly Yakovlev

option. Certain systems such as a missile seeker or aircraft

Abstract—An array antenna system with innovative signal antenna have physical size limitations; therefore they have
processing can enhance the resolution of a signal direction of relatively wide mainlobe beamwidth. Consequently, the
arrival (DOA) estimation. Super resolution algorithms take resolution is quite poor. Also, if there are multiple signals
advantage of array antenna structures to better process the falling in the antenna mainlobe, it is difficult to distinguish
incoming signals. They also have the ability to identify multiple between them.
targets. This paper explores the eigen-analysis category of
Instead of using a single antenna, an array antenna system
super resolution algorithm. A class of Multiple Signal
Classification (MUSIC) algorithms known as a root-MUSIC with innovative signal processing can enhance the resolution
algorithm is presented in this paper. of signal DOA. An array sensor system has multiple sensors
distributed in space. This array configuration provides spatial
The root-MUSIC method is based on the eigenvectors of the
sensor array correlation matrix. It obtains the signal estimation samplings of the received waveform. A sensor array has
by examining the roots of the spectrum polynomial. The peaks better performance than the single sensor in signal reception
in the spectrum space correspond to the roots of the polynomial and parameter estimation. Its superior spatial resolution
lying close to the unit circle. provides a means to estimate the direction of arrival (DOA)
Statistical analysis of the performance of the of multiple signals. A sensor array also has applications in
processing algorithm and processing resource interference rejection [1], electronic steering [2], multi-beam
requirements are discussed in this paper. Extensive forming [3], etc. This technology is now widely used in
computer simulations are used to show the performance communications, radar, sonar, seismology, radio astronomy
of the algorithms. ,etc.
There are many different super resolution algorithms
Index Terms— Array antenna, Direction of arrival including spectral estimation, model based, and
estimation, Signal processing. eigen-analysis to name a few [4,5,6]. In this paper, we
concentrate the discussion on the application of estimating
the DOA of multiple signals. The focuses are on a class of
I. INTRODUCTION
Multiple Signal Classification (MUSIC) algorithms known
Accurate estimation of a signal direction of arrival as root-MUSIC and an extension of root-MUSIC. We present
(DOA) has received considerable attention in communication detailed MATLAB simulation results for each algorithm.
and radar systems of commercial and military applications.
Radar, sonar, and mobile communication are a few examples
of the many possible applications. For example, in defense
application, it is important to identify the direction of a II. ARRAY SENSOR SYSTEMS
possible threat. One example of commercial application is to
We use an array antenna with a 16 element uniform linear
identify the direction of a emergency cell phone call such that
array (ULA) in this paper. Fig. 1 shows the general
the rescue team can be dispatched to the proper location.
configuration for a ULA antenna having M elements
DOA estimation using a fixed antenna has many
arranged along a straight line with the distance between
limitations. Its resolution is limited by the antenna’s
sensor elements, be d = λ/2, where λ is the incoming signal
mainlobe beamwidth. Antenna mainlobe beamwidth is
wavelength. The angle of the incoming signal, θ, is measured
inversely proportional to its physical size. Improving the
relative to the antenna bore sight.
accuracy of angle measurement by increasing the physical
aperture of the receiving antenna is not always a practical

This work was supported in part by the Raytheon Bore sight

H. K. Hwang and Zekeriya Aliyazicioglu are with the Electrical and
Computer Engineering Department, California State Polytechnic University, Incoming
Pomona, CA 91768 USA, (e-mail: hkhwang@ csupomona.edu and Signal
θ
[email protected]).
Marshall Grice was a graduate student at California State Polytechnic d= λ/2
University, Pomona, CA. He is now with the Boeing, CA. (e-mail:
[email protected]) 1 2 3 4 M
Anatoly Yakovlev was an undergraduate student at California State Sensor Elements
Polytechnic University, Pomona, CA. He is now with Western Digital, San
Jose, CA 95138, USA (e-mail: [email protected]) Figure 1. ULA Antenna configuration

ISBN: 978-988-17012-1-3 IMECS 2008

Proceedings of the International MultiConference of Engineers and Computer Scientists 2008 Vol II
IMECS 2008, 19-21 March, 2008, Hong Kong

For the a conventional antenna, the main lobe beam subspace and the remaining M-L eigenvectors will span the
width (MLBW) of an antenna manner is given by, in radians noise only subspace, QN. By eigen-analysis we can represent
the M-L smallest eigenvectors as
λ
MLBW = k (1)
Rq i = σ 2q i i = L + 1,..., M
D . (7)
where D is the diameter of the antenna array and k is a
proportionality constant, for most case k ≈ 1 [6]. Using (7) in (2), can be rewritten as

SDS H q i = 0 i = L + 1,..., M (8)

III. ROOT-MUSIC ALGORITHM
The root-MUSIC method relies on the following Since S is a full column rank matrix and D is diagonal, (8)
properties of the array correlation matrix: the space spanned becomes
by its eigenvectors may be partitioned into two orthogonal
subspaces, namely the signal plus noise subspace and the SHqi = 0 i = L + 1,..., M (9)
noise only subspace; the steering vectors corresponding to
the directional sources are orthogonal to the noise subspace or more explicitly
[7]. The MxM correlation matrix that contains L number of
incoming signals is formed by sk H q i = 0 i = L + 1,..., M
. (10)
R = SDS H + σ 2 I (2) k = 1,..., L

where σ 2 is the variance of the Gaussian white noise, D is Equation (10) proves the orthogonality between the signal
the signal power matrix and S is the signal direction matrix plus noise and the noise only subspaces. This is important
because it shows that the angle of the incoming signals can be
D = diag[ P1, P 2,...PL ] (3) found by searching for signal direction vectors that, when
projected onto the noise only subspace, give a zero result.
1 1 ... 1 Following this idea, if a polynomial, J(z), is constructed such
⎡ ⎤
⎢ e− j β (θ 1) − j β (θ 2 ) − j β (θ L ) ⎥ that
e ... e
S=⎢ ⎥ J ( z) = v HQNQN H v = 0 (11)
⎢ ... ... ... ... ⎥
⎢ − j ( M −1) β (θ 1) − j ( M −1) β (θ 2 ) − j ( M −1) β (θ L ) ⎥
⎣e e ... e ⎦ where the steering vector v is
(4)
T
and the phase delay between sensor elements is v = ⎡⎣1 z −1 z −2 " z −( M −1) ⎤⎦ (12)
(4) 2π d
β (θ i ) = sin(θ i ) (5) and
λ
2π d
j sin(θ )
Let λ1 ≥ λ2 ≥ ... ≥ λM be the eigenvalues of the z=e λ
. (13)
correlation matrix R, and υ1 ≥ υ2 ≥ ... ≥ υ L be the
eigenvalues for SDSH . Then from (2) Then the roots of J(z) contain the directional information
of the incoming signals. Ideally, the roots of J(z) would be on
the unit circle at locations determined by the directions of the
⎧υ + σ 2 i = 1, 2,...L incoming signals; however, due to the presence of noise the
λi = ⎨ i 2 (6)
⎩ σ i = L + 1,...M roots may not necessarily be on the unit circle. In this case,
the L closest roots to the unit circle are the roots that
correspond to the L incoming signals [9]. These selected
For high signal to noise ratios (SNR) υi  σ 2 . The
roots, by themselves, do not directly represent the incoming
eigenvalues can be used to determine the number of sources angle. For each root, the incoming angle is found by solving
that are detected by counting the number of comparatively (13).
large eigenvalues. Alternatively, Ref. [7] suggests a more
rigorous approach to determining the number of incoming ⎡ λ ⎤
sources that provides better detection performance when the θ k = arcsin ⎢ arg( zk ) ⎥
incoming SNR is not as high. For the purposes of this, the ⎣ 2π d ⎦ . (14)
incoming SNR is chosen to be sufficiently high as to not be in
a situation where the source number detection is ambiguous. Obviously, when the root-MUSIC algorithm is
Let q1, q2, … , qM be the eigenvectors associated with the implemented there is no prior knowledge of the incoming
signal directions or signal powers needed to construct the
decreasing ordered eigenvalues λ1 ≥ λ2 ≥ ... ≥ λM . From
correlation matrix using (2). Therefore the correlation matrix
(6) the first L eigenvectors will span the signal plus noise

ISBN: 978-988-17012-1-3 IMECS 2008

Proceedings of the International MultiConference of Engineers and Computer Scientists 2008 Vol II
IMECS 2008, 19-21 March, 2008, Hong Kong

must be estimated using only the information available from 2π ( d + ei )

the sensor array. There are several methods commonly used β (θ i ) = sin(θ i )
λ (17)
to perform this estimation such as temporal averaging, spatial
smoothing or, a hybrid combination of both temporal
When these data are applied to the root-MUSIC algorithm,
averaging and spatial smoothing [8]. In this paper, we use
(14) no longer has the correct value for d and obviously will
only the temporal averaging method.
return a result that has some increased error.
The estimated correlation matrix using the temporal
averaging method with k snapshots is given as B. Root-Music Simulation Results
Statistical results, Fig.4 , show that two incoming signals
Φ = E[ A H A ] (15) are clearly identified even as the separation between the two
signals is well below the conventional main lobe beam width.
where the incoming data matrix A is It can be seen that as the spacing between signals decreases
the variance of the estimates increases. The average
estimation error is in all cases nearly zero.
⎡ u1 (1) u1 (2) ... u1 ( k ) ⎤
⎢ u (1) u (2) ...
⎥
u2 ( k )
A=⎢ 2 2 ⎥
⎢ ... ... ... ... ⎥ 1 Beam Width Separation:
Average Estimated Angle =
⎢ ⎥ 39.9972 46.0041
u
⎣ M (1) u (2) ... uM ( k ) ⎦
M
. (16) Average Estimation Error=
0.0028 -0.0041
with ui(k) being the ith sensor output at time k. Mean Square Error =
0.0009 0.0012
The estimated correlation matrix, Φ , asymptotically
approaches the correlation matrix, R as the number of
snapshots increases. Therefore in order to have an accurate
estimation of the correlation matrix the observation time must
(a)
be sufficiently long. The long observation times are not ideal
for radar signal processing applications; however there are
many applications where this does not pose a problem. 1/2 Beam Width Separation:
Average Estimated Angle =
Correlation matrix estimation techniques like the spatial 39.9988 42.9995
smoothing method are better suited for use in time sensitive Average Estimation Error=
systems. 0.0012 0.0005
Mean Square Error =
0.0023 0.0025
A. Sensor Spacing and Phase Sensitivity
The root-MUSIC algorithm assumes that each sensor is
perfectly spaced relative to the other sensors in the array. (b)
While this holds true for the theoretical case, perfect sensor
spacing is difficult to achieve when the algorithm is actually 1/4 Beam Width Separation:
implemented even with modern construction techniques. Average Estimated Angle =
39.9999 41.4987
The method by which the algorithm is modified to model Average Estimation Error=
these variations is rather straight forward. The sensor 0.0001 0.0013
spacing problem is characterized by Fig.3. Mean Square Error =
0.0047 0.0042

Incoming
Signal (c)
Figure 4. Statistical comparison for 2 signals at various separation angles
θ
d+e3 d+e4
d+e1 d+e2 d+eM
d= λ/2 Fig.5 shows the effect of increasing the number of snap
shots used for the temporal averaging correlation matrix
1 2 3 4 M
Variation in Distance Ideal Sensor estimation. The estimation variance decreases with increased
between Sensor Elements Elements observation times. The average estimation error does not
Figure 3. Non-ideal ULA sensor spacing seem to be very sensitive to the observation time
The spacing error of each sensor, ei, is a Gaussian
random variable added to the ideal spacing. Taking this error
into account, (5) is used to create the phase shift between
sensor elements for the incoming signals becomes

ISBN: 978-988-17012-1-3 IMECS 2008

Proceedings of the International MultiConference of Engineers and Computer Scientists 2008 Vol II
IMECS 2008, 19-21 March, 2008, Hong Kong

The algorithm does exhibit a rather strong sensitivity to the

1/4 Beam Width Separation positional accuracy of the sensor placement; however with
32 Snap Shots: proper array calibration these effects could be minimized.
Average Estimated Angle =
39.9999 41.4987
Average Estimation Error=
0.0001 0.0013 IV. CONCLUSION
Mean Square Error =
0.0047 0.0042 We have presented the root-MUSIC method based on the
eigenvector of the sensor array correlation matrix to estimate
angle of incoming signals. We give extensive computer
(a)
simulation results to demonstrate the performance of the
algorithms, which enhance the DOA estimation.
1/4 Beam Width Separation
The simulation results of the root-MUSIC algorithm show
96 Snap Shots: the following results
Average Estimated Angle = 1. The capability to resolve multiple targets with
39.9992 41.5008
Average Estimation Error=
separation angles smaller the main lobe beam width
0.0008 -0.0008 of the array antenna.
Mean Square Error = 2. The estimation variance can be reduced by increasing
0.0015 0.0016 the number of snapshots in correlation matrix
estimation
(b) 3. The estimation variance increases as the angle
separation between signals becomes smaller
Figure 5. Variation of snap shot comparison 4. The estimation variance depends on the direction of
In Fig.6, We compare the simulation results of ideally the signal. A signal coming from the bore sight has
spaced sensor elements against the results of a simulation minimum estimation variance.
where the sensor elements have a 1% random variance in
their spacing; in other words σ = λ / 200 .
ACKNOWLEDGMENT
The authors would like to thank the Raytheon Corporation
for its support of this investigation.

REFERENCES
[1]. Anne Lee, Lijia Chen, H. K. Hwang and et al. “Simulation Study of
Wideband Interference Rejection using Adaptive Array Antenna”.
IEEE Aerospace Conference, March 5-12, 2005
[2]. M. G. M. Hussain, Performance Analysis and Advancement of
Figure 6. Comparison of Non-Ideal Sensor Spacing Self-Steering Arrays for Nonsinusoidal Waves, IEEE Trans. on
Electromagnetic Compatibility, May 1988
Figure 6. Comparison of Non-Ideal Sensor Spacing [3]. X. Zhang and D. Su, Digital Processing System for Digital Beam
Forming Antenna, IEEE International Symposium on Microwave,
Antenna Propagation and EMC Technologies, 2005
Increasing the amount of spacing variance from 1% to 5% [4]. Jianmin Zhu, Megan Chan and H. K. Hwang, “Simulation Study on
shows an increased error variance in Fig.7. It is worth noting Adaptive Antenna Array” IEEE International Signal Processing
that while the performance of the algorithm decreases with Conference, Dallas, 2003
[5]. Marshall Grice, Jeff Rodenkirch, Anatoly Yakovlev, H. K. Hwang, Z.
increased sensor spacing error the algorithm is still able to Aliyazicioglu, Anne Lee, “Direction of Arrival Estimation using
successfully distinguish both incoming signal directions. Advanced Signal Processing”, RAST Conference , Istanbul-Turkey,
2007.
[6]. Skolnik, Merrill. Introduction to RADAR Systems. 3rd ed. New York:
Mc Graw Hill, 2001.
[7]. Godara, Lal Chand, Smart Antennas. Boca Raton: CRC Press, 2004
[8]. Kuo, Chen Yu. Wideband Signal Processing for Super Resolution DOA
Estimation. Pomona: California State Polytechnic University Pomona,
2006.
[9]. Forsythe, Keith. “Utilizing Waveform Features for Adaptive
Beamforming and Direction Finding with Narrowband Signals.”
Lincoln Laboratory Journal 10.2 (1997): 99-126.

Figure 7. MUSIC Spectrum with Two Signals from (65°, 15°) and (65°, 25
Figure 7. MUSIC Spectrum with Two Signals from (65°, 15°)
and (65°, 25°)

The simulation results of the root-MUSIC algorithm

clearly demonstrate the ability to resolve multiple targets
with separation angles smaller then the main lobe beam width
of the array thus proving its super-resolution capabilities.

ISBN: 978-988-17012-1-3 IMECS 2008