World Academy of Science, Engineering and Technology

International Journal of Electronics and Communication Engineering

Vol:8, No:1, 2014

Performance Analysis of MUSIC, Root-MUSIC and

ESPRIT DOA Estimation Algorithm
N. P. Waweru, D. B. O. Konditi, P. K. Langat

Abstract—Direction of Arrival estimation refers to defining a ESPRIT and their variants. Eigenvalue decomposition based
mathematical function called a pseudospectrum that gives an algorithms involves manipulation of the signal autocorrelation
indication of the angle a signal is impinging on the antenna array. matrix to yield noise and signal subspaces from which angles
This estimation is an efficient method of improving the quality of
service in a communication system by focusing the reception and of arrival of impinging signals are extracted.
transmission only in the estimated direction thereby increasing This paper therefore, seeks to analyze the performance of
three DOA algorithms based on number of array elements and
Open Science Index, Electronics and Communication Engineering Vol:8, No:1, 2014 waset.org/Publication/9997885

fidelity with a provision to suppress interferers. This improvement

is largely dependent on the performance of the algorithm employed the Signal to Noise Ratio (SNR) to have an optimum choice
in the estimation. Many DOA algorithms exists amongst which are of algorithm and design for a given environment.
MUSIC, Root-MUSIC and ESPRIT. In this paper, performance of
these three algorithms is analyzed in terms of complexity, accuracy MUSIC, Root-MUSIC and ESPRIT algorithms were
as assessed and characterized by the CRLB and memory developed and simulated in MATLAB software for signals
requirements in various environments and array sizes. It is found operating at 2.4GHz for a ULA with inter-element spacing
that the three algorithms are high resolution and dependent on the of λ2 . In Section II of this paper, the signal model of an N
operating environment and the array size. element uniform linear array receiving M signals from
Keywords—Direction of Arrival, Autocorrelation matrix, directions θ1 , θ2 , . . . , θM is derived followed by the
Eigenvalue decomposition, MUSIC, ESPRIT, CRLB. Cramer-Rao Lower Bound (CRLB) for DOA estimation in
III. Sections IV-VI briefly describe the mathematical aspects
I. I NTRODUCTION behind the three algorithms. Simulation results are presented

W IRELESS communication is one of the fastest in Section VII followed by the discussion and conclusion in
growing fields in the engineering world. This has Sections VIII and IX respectively.
been necessitated by the advancement made in the research
and design of communication equipment [1], [2]. It started II. S IGNAL MODEL
with cellular communication, then the world-wide-web to the
extent where these services are accessible wherever
whenever. The Quality of Service (QoS) however,
deteriorates with distance between the transmitting node and
the receiver.
Estimating the direction of a transceiver is an effective
method of improving the QoS between a node and a
transceiver, by heightening and focusing the transmission
only to the direction of the receiver and vice versa for the
receiver [3]–[5]. This is achieved by the use antenna arrays
with some added capability to estimate the Direction of
Arrival (DOA) of all impinging signals.
The aforementioned improvement is to a large extent
dependent on the performance of the employed algorithm.
The performance of a DOA algorithm is in turn dependent
on the size of the array, number of impinging signals,
spacing between elements and the number of snapshots used
in the estimation process.
Many DOA techniques exist: quadratic type e.g CAPON
and those based on eigenvalue decomposition e.g. MUSIC, Fig. 1. Uniform linear array.
Waweru is with Electrical & Electronic Engineering Department of Dedan
Kimathi University of Technology, P.O Box 657-10100, Nyeri (corresponding
author, phone: +254724876363; e-mail: [email protected]). Consider a uniform linear array (ULA) having N identical
Konditi is with Electrical & Communication Engineering Department of elements, separated by a distance d as shown in Fig. 1. This
MultiMedia University College of Kenya, P.O Box 30305-00100, Nairobi. array is receiving a far field signal impinging the array at an
Langat is with Telecommunication & Information Engineering Department
of Jomo Kenyatta University of Agriculture and Technology, P.O Box 62000- angle θ to the array axis. Taking element 1 as the reference,
00200, Nairobi. the line path from the source to the ith element is shorter than

International Scholarly and Scientific Research & Innovation 8(1) 2014 209 ISNI:0000000091950263
 World Academy of Science, Engineering and Technology
International Journal of Electronics and Communication Engineering
Vol:8, No:1, 2014

that to the 1st element. If the received signal at sensor 1 is two: array steering vector and the noise subspace, therefore is
x1 (t) = s(t), it is delayed at sensor i by a null for a particular Angle of Arrival (AOA).
(i − 1)d sin θ For a uniform linear array with N elements and M signals
τi = (1) s1 (t),s2 (t),. . . ,sM (t) arriving from directions θ1 ,θ2 ,. . . θM ,
c
and in the presence of noise n(t), the received signal x(t) is
Then the received signal at sensor i is given by

x(t) = As(t) + n(t) (7)

xi (t) = e−jωτi x1 (t) = e−jkd(i−1)sinθ s(t)
(2)
= ej(i−1)ψ s(t) Defining an N×N autocorrelation matrix of the received signal
Rxx as
Putting received signals from all N elements together.
⎡ ⎤ ⎡ ⎤ Rxx = E{x(t)xH (t)} = ARss AH + σ02 I (8)
x1 (t) 1
⎢ x2 (t) ⎥ ⎢ e−jψ ⎥
⎢ ⎥ ⎢ ⎥ where:
⎢ .. ⎥ ⎢ .. ⎥
x(t) = ⎢
⎢ . ⎥=⎢
⎥ ⎢ . ⎥ s(t) = a(θ)s(t)
⎥ (3) Rss = E{s(t)sH (t)} = diag{σ12 , ....., σM
2
} (9)
⎢ .. ⎥ ⎢ .. ⎥
Open Science Index, Electronics and Communication Engineering Vol:8, No:1, 2014 waset.org/Publication/9997885

⎣ . ⎦ ⎣ . ⎦
−j(N −1)ψ Rxx has N eigenvalues [λ1 , λ2 , . . . , λN ] and N associated
xN (t) e
eigenvectors making a subspace E=[e1 , e2 , . . . , eN ]. Sorting
where a(θ)is called the steering vector. the N eigenvalues from the smallest to the largest, the subspace
If there are M signal sources, received by the array arriving E can be decomposed into two subspaces:
at angles θ1 , θ2 , . . . , θM , we get a signal model
E = [e1 , ..., eM , eM +1 , ..., eN ] (10)

 

x(t) = As(t) (4) EN ES

where the vector A is the array steering vector having the = [E N E S ] (11)
following vandermode structure.
⎡ ⎤ E N is the N ×(N − M ) noise subspace composed of the
1 1 ... 1 eigenvectors associated with the noise, whereas E S is the
⎢ e−j2ψ1 e−j2ψ2 ... e−j2ψM ⎥ N × M signal subspace composed of the eigenvectors
⎢ ⎥
A=⎢ .. .. . . ⎥ (5) associated with the arriving signal.
⎣ . . . . .
. ⎦
−j(N −1)ψ1 −j(N −1)ψ2 −j(N −1)ψM Due to the orthogonality of the noise subspace and the
e e ... e
array steering vector at the angles of arrival θ1 ,θ2 ,. . . θM , the
III. C RAMER -R AO L OWER B OUND matrix product aH (θ)EN EN H
a(θ) is zero for this angles. The
reciprocal of this matrix product creates sharp peaks at the
Cramer-Rao Lower Bound provides an algorithm angle of arrival. Thus the MUSIC pseudospectrum is given
independent means of assessing and comparing the accuracy as
and performance of a DOA algorithm. The CRLB on the 1
variance of direction estimation errors provides a useful P (θ) = H (12)
| a (θ)EN EN H a(θ) |
characterization of the achievable accuracy of the DOA
system. This is achieved by comparing the Mean Square V. ROOT-MUSIC
Error(MSE) with the CRLB [6], [7]. This is a variant of MUSIC algorithm that employs more
The CRLB theorem states that for a length N vector of information than MUSIC [10]. Unlike MUSIC which involves
received signal x dependent on a set of parameters P, and plotting the pseudospectrum against the angles and searching
corrupted by additive noise, the variance of an unbiased for the peaks, ROOT-MUSIC involves finding the roots of a
estimate of the pth estimate is greater than the cramer-rao polynomial.
lower bound.
Starting with the pseudospectrum of MUSIC algorithm
The CRLB of a DOA estimation problem is given by
6 1
var(θ) ≥ CRLB = (6) P (θ) = (13)
snr[N (N 2 − 1)(kd)2 sin2 θ] | aH (θ)EN EN
H a(θ) |

IV. MUSIC defining C = EN EN H

, the denominator of equation 12 above
can be rewritten as
Multiple Signal Classification (MUSIC) is a popular high
1
resolution algorithm based on eigenstructure technique. The P (θ) = H (14)
main idea behind this DOA algorithm is that of performing | a (θ)Ca(θ) |
eigenvalue decomposition on the correlation matrix [8], [9], the mth element am (θ) of the array steering vector is
separating it into two subspaces: signal subspace and the noise defined as
subspace. Since the signal subspace is spanned by the array
steering vector of the received signals, this makes the steering
vector orthogonal to the noise subspace. The product of the am (θ) = e−jkdm sin θ , m = 0, 1, . . . , N − 1 (15)

International Scholarly and Scientific Research & Innovation 8(1) 2014 210 ISNI:0000000091950263
 World Academy of Science, Engineering and Technology
International Journal of Electronics and Communication Engineering
Vol:8, No:1, 2014

The denominator, thus can be rewritten as


N −1 N
 −1
aH (θ)Ca(θ) = e−jkdm sin θ Cmn ejkdn sin θ
m=0 n=0
−1
(16)

N
jkd sin θ
= C e
=−N +1

where C is the sum of the elements along the th diagonal

of C.
Letting z = e−jkd sin θ , equation 16 above simplifies to


N −1
D(z) = C z  (17)
=−N +1

The roots of D(z) that lie closest to the unit circle correspond
Open Science Index, Electronics and Communication Engineering Vol:8, No:1, 2014 waset.org/Publication/9997885

to the poles of the MUSIC pseudospectrum. These 2(N − 1)

roots can be written as
Fig. 2. ULA decomposition in ESPRIT algorithm.
zi =| zi | ejarg(zi ) , i = 1, 2, . . . , 2(N − 1) (18)
Choosing those roots inside the unit circle whose magnitude
| zi | 1, and comparing ejarg(zi ) to e−jkd sin θ gives From equation 20, correlation matrices R11 and R22 of the
signals in the two subarrays can be estimated as

argzi R11 = E{x1 (t)xH
θi = −sin−1 (19) 1 (t)}
kd (23)
R22 = E{x2 (t)xH
2 (t)}
VI. ESPRIT
Eigen-decomposing R11 and R22 result in two signal
Estimation of Signal Parameters via Rotational Invariance
subspace E1 and E2 respectively. Defining a 2M × 2M
Technique (ESPRIT) algorithm involves decomposing an N
matrix C from the two subspaces such that
element array into two identical subarrays each with S element ⎡ H ⎤
[8], [11]. The objective of ESPRIT algorithm is to estimate the E1  
angle of arrival by determining the rotation operator Φ. C=⎣ ⎦ E1 E2 = EC ΛEC H
(24)
The separation distance between the two subarrays is Δ E2H
(measured in wavelengths). Fig. 2 show a ten element linear
array and possible subarray configurations for S=9, 7 and 5 EC is a 2M × 2M matrix obtained by eigenvalue
elements respectively (Δ=1, 3 and 5). The 1st element in the decomposition of C such that λ1 ≥ λ2 ≥ . . . ≥ λ2M and
subarray is the first element in the first subarray whereas the Λ = diag{λ1 λ2 . . . λ2M }.
(Δ + 1)th element of the original sensor is the first element Partitioning EC into four M × M submatrices such that
in the second subarray.  
E11 E12
Letting EC = (25)
E21 E22
N-Number of elements in the original array
S-Number of elements in each subarray The rotation operator is estimated as
M-Number of signals hitting the subarray.
−1
Also letting x1 (t) and x2 (t) be the received signal in the two Φ = −E12 E22 (26)
subarrays, corrupted by additive white gaussian noise n1 (t)
and n2 (t) respectively. From M eigenvalues of Φ, angles of arrival can be estimated
as 
−1 arg(λi )
x1 (t) = As(t) + n1 (t) θi = sin (27)
(20) kΔ
x2 (t) = AΦs(t) + n2 (t)
VII. SIMULATION RESULTS
where x1 (t), x2 (t), n1 (t) and n2 (t) are M × 1 matrices. A
is the S × M steering matrix and the variable Φ is M × M Simulations were done in MATLAB software for angles
diagonal matrix called the rotation operator. of arrival θ1 = −64◦ , θ2 = 0◦ , θ1 = 23◦ and θ1 = 58◦
respectively. The array size was held to 8 elements as the
Φ = diag{ejψ1 , ejψ2 , . . . , ejψM } (21)
values of SNR were varied from 0-100dB in steps of 20dB.
where This was repeated holding SNR to 50dB and varying the array
ψi = −2kΔ sin θi ; 1≤i≤M (22) size from 5-100 elements.

International Scholarly and Scientific Research & Innovation 8(1) 2014 211 ISNI:0000000091950263
 World Academy of Science, Engineering and Technology
International Journal of Electronics and Communication Engineering
Vol:8, No:1, 2014
Open Science Index, Electronics and Communication Engineering Vol:8, No:1, 2014 waset.org/Publication/9997885

(a) MUSIC for varying SNR (b) MUSIC for varying snapshots

(c) MUSIC for varying array size (d) MUSIC for a 5 element array

(e) MUSIC for a 200 element array (f) MUSIC for two angles closely spaced
Fig. 3. Performance of MUSIC algorithm

VIII. D ISCUSSION of this algorithms in an environment with varying SNR. For

Fig. 3 shows the performance of MUSIC algorithm in low values of SNR, 0dB, the spikes depicting the arrival of a
various operating conditions. Fig. 3a shows the performance

International Scholarly and Scientific Research & Innovation 8(1) 2014 212 ISNI:0000000091950263
 World Academy of Science, Engineering and Technology
International Journal of Electronics and Communication Engineering
Vol:8, No:1, 2014
Open Science Index, Electronics and Communication Engineering Vol:8, No:1, 2014 waset.org/Publication/9997885

(a) Root-MUSIC for varying SNR (b) Root-MUSIC for varying array size

(c) Root-MUSIC for two closely spaced angles-30dB (d) Root-MUSIC for two closely spaced angles-1000B

(e) Root-MUSIC at 23◦ for varying array size (f) Root-MUSIC at 23◦ for varying SNR
Fig. 4. Performance of Root-MUSIC

signal from certain direction are small and the response is arrival. As the values of SNR increase, however, the
almost flat. It is thus difficult to exactly extract the angles of resolution of the algorithm is observed to improve

International Scholarly and Scientific Research & Innovation 8(1) 2014 213 ISNI:0000000091950263
 World Academy of Science, Engineering and Technology
International Journal of Electronics and Communication Engineering
Vol:8, No:1, 2014

considerably and the spikes become more definite. This is between the eigenvalues associated with the signal and those
attributed to the fact that for low SNR the difference associated with the noise become smaller and the peaks
Open Science Index, Electronics and Communication Engineering Vol:8, No:1, 2014 waset.org/Publication/9997885

(a) MSE of MUSIC for varying SNR (b) MSE of MUSIC for varying array size

(c) MSE of Root-MUSIC for varying SNR (d) MSE of Root-MUSIC for varying array size

(e) MSE of ESPRIT for varying SNR (f) MSE of ESPRIT for varying array size
Fig. 5. MSE in various environments

International Scholarly and Scientific Research & Innovation 8(1) 2014 214 ISNI:0000000091950263
 World Academy of Science, Engineering and Technology
International Journal of Electronics and Communication Engineering
Vol:8, No:1, 2014

TABLE I
ACCURACY OF ESPRIT FOR VARYING SNR 1◦ and θ2 =0◦ in different environments. From this, it is clear
SNR -64◦ 0◦ 23◦ 58◦ that Root-MUSIC is a high resolution DOA algorithm which
0dB -62.8826 1.3855 35.1460 57.0785 can estimate two closely spaced angles of arrival as two angles
10dB -65.4913 -0.0669 28.2430 57.5258 and not one.
20dB -64.3173 0.1507 24.2929 58.9783
30dB -64.0421 0.0655 23.2738 58.3887 Fig. 4e and 4f depicts the accuracy of the Root-MUSIC for
40dB -64.0120 -0.0120 22.8796 58.0220 a signal impinging on the array specifically at 23◦ for varying
50dB -64.0092 0.0001 23.0749 57.9863 values of array size and SNR. It can be observed that the
60dB -63.9981 0.0026 23.0057 57.9986
70dB -64.0011 0.0004 22.9964 58.0001 performance of this algorithm is less dependent on the array
80dB -63.9997 0.0004 22.9997 58.0010 size but largely on the operating environment but it anyhow
90dB -63.9999 0 22.9994 58.0000 improves with the improvement of either.
100dB -64.0000 0 22.9997 57.9999
Tables I and II shows the response of the ESPRIT algorithm
to varying values of SNR and array size. As in the other two
TABLE II algorithms, the resolution is very poor for low values of SNR
ACCURACY OF ESPRIT FOR VARYING ARRAY SIZE
and improves as the SNR values increase. Minimal variation in
N -64◦ 0◦ 23◦ 58◦
6 -64.0064 0.0007 22.6700 57.8722
the accuracy of the algorithm with the array size is observed.
Open Science Index, Electronics and Communication Engineering Vol:8, No:1, 2014 waset.org/Publication/9997885

10 -64.0052 -0.0068 23.0135 57.9998 Fig. 5 summarizes the performance of the three algorithms
20 -63.9983 0.0025 22.9432 58.0063 in terms of the Mean Square Error (MSE) and in comparison
30 -63.9910 0.0006 22.9462 58.0328
40 -64.6071 0.0069 22.6469 56.5691
with CRLB as the values of the SNR and the array sizes are
50 -63.9274 -0.0014 23.0007 57.9558 varied. Four angles of arrival θ1 = −36◦ , θ2 = −3◦ , θ1 =
60 -64.0218 0.0118 23.4792 58.0016 0◦ and θ1 = 87◦ (two closely spaced and one near endfire).
70 -63.8838 0.0117 22.7634 57.3365
80 -62.9261 0.0001 22.0148 53.5281
From Fig. 5a-5e, it can be observed that the MSE diminishes
90 -64.0007 0.0021 23.0129 57.9459 with increase in both SNR and array size and the performance
100 -63.9965 -0.0003 23.0709 58.0131 approaches the CRLB. However, variation of MSE with the
array size is flat and approaching the CRLB behaviour for
ESPRIT algorithm as captured in Fig. 5f.
therefore become smaller with respect to the noise levels.
With increase in SNR, the difference between the two sets of IX. CONCLUSION
eigenvalues is substantial and the peaks are bigger with From the discussion above, where three DOA algorithms
respect to the noise levels. are developed in MATLAB and their response to various
Fig. 3b depicts the response of the MUSIC algorithm to parameters presented, it can be concluded that the three
varying number of snapshots. It can be seen from the figure algorithms are high resolution algorithms. This is derived
that for 20 snapshots the response has less pronounced spikes. from the precision with which the angles of arrival are
The resolution is seen to improve with increase in the number estimated. The three algorithm are highly sensitive to the
of snapshots from 20 to 200,000. The number of snapshots signal to noise ratio SNR where the resolution of the
affect the correlation between the received signals. For less algorithms is found to improve with the increase in the SNR.
snapshots, the received signals seem more correlated making Resolution is also found to considerably increase with the
it difficult to distinguish between them. number of elements forming the antenna array.
The response of MUSIC algorithm to the array size is as MUSIC algorithms is suitable for an array with few array
shown in Fig. 3c. From the figure, it can be seen that for 5 elements in an environment with high SNR with average
elements, the spikes are very definite and exactly correspond snapshots whereas root-MUSIC requires relatively more
the the angles of arrival. As the array size increase, the elements than MUSIC in a high SNR environment again
resolution improves and the extraction of the angle of arrival with average snapshots. ESPRIT algorithm on the other hand
becomes easier. An extract of the pseudospectrum for N=5 is less dependent on the array size and can perform relatively
and N=200 are shown in Fig. 3d and 3e respectively. well than the other two in an environment having low SNR.
MUSIC pseudospectrum for two closely spaced signals
separated by 1◦ i.e. θ1 =-1◦ and θ2 =0◦ is captured in Fig. 3f R EFERENCES
for N=8 and SNR=100dB. From the figure, the two angles of [1] S. Kamboj and R. Dahiya, “Adaptive antenna array for satellite
arrival can be exactly extracted making MUSIC algorithm a communication systems,” Proceedings of the International
high resolution DOA algorithm. Multiconference of Engineers and Computer Scientists, vol. 2,
March 2008,Hong Kong.
Fig. 4a shows the behaviour of root-MUSIC algorithm for [2] A. Iozsa and A. Vesa, “Direction of arrival estimation for
varying values of SNR. From the figure, it can be seen that uniform sensor arrays,” International Symposium on Electronics and
the accuracy is poor for 0dB SNR and improves as the SNR Telecommunication(ISETC), no. 978-1-4244-8460-7/10, 2010.
[3] A. Iozsa and A. Vesa, “The esprit algorithm. variants and
increases from 0dB to 100dB. The response of Root-MUSIC precision,” International Symposium on Electronics and
to variation in the array size is shown in Fig. 4b. It can be Telecommunication(ISETC), no. 978-1-4244-8460-7/10, 2010.
observed from the figure that the variation of the estimated [4] A. A. Noori, S K. Gharghan, and A. A. Wahab, “Study of signal
estimation parameters via rotational invariance technique by using ants
angle of arrival is minimal with increase in the array size. colony optimization algorithm,” Eng. & Tech. Journal, vol. 29, no. 4,
Fig. 4c and 4d presents two signals separated by 1◦ , i.e θ1 =- pp. 736–749, 2011.

International Scholarly and Scientific Research & Innovation 8(1) 2014 215 ISNI:0000000091950263
 World Academy of Science, Engineering and Technology
International Journal of Electronics and Communication Engineering
Vol:8, No:1, 2014

[5] M. Barkar, R. M. Vani, and P. V. Hunagund, “Eigen structures based

direction of arrival estimation algorithms for smart antennas systems,”
in IJCSNS International Journal of Computer Science and Network
Security, vol. 9, pp. 96–100, Nov 2009.
[6] Engin Tuncer and Benjamin Friedlander, Classical and Modern
Direction-of-Arrival Estimation, ch. 1, pp. 10–13. Elsevier Inc., 2009.
[7] Y. Wu, H. Liu, and H. C. So, “Fast and accurate direction-of-arrival
estimation for a single source,” Progress In Electromagnetics Research
C, vol. 6, pp. 13–20, 2009.
[8] A. Vesa, “Direction of arrival estimation using music and root-music
algorithm,” in 18th Telecommunication forum TELFOR, November
2010.
[9] X. Gu and Y. H. Zhang, “Resolution threshold analysis of music
algorithm in radar range imaging,” Progress In Electromagnetics
Research B, vol. 31, pp. 297–321, 2011.
[10] Z. Aliyazicioglu, H. K. Hwang, M. Grice, and A.Yakovlev, “Sensitivity
analysis for direction of arrival estimation using a root-music algorithm,”
Engineering Letters, vol. 13, pp. 353–360, 2008.
[11] C. R. Dongarsane, A. N. Jadhav, and S. M. Hirikude, “Performance
analysis of esprit algorithm for smart antenna system,” International
Open Science Index, Electronics and Communication Engineering Vol:8, No:1, 2014 waset.org/Publication/9997885

Journal of Communication Networks and Security, vol. 1, pp. 34–37,

2012.

International Scholarly and Scientific Research & Innovation 8(1) 2014 216 ISNI:0000000091950263