
- •Abstract
- •Keywords
- •Contents
- •Introduction
- •Background on Array Processing
- •2.1 INTRODUCTION
- •2.1.1 Propagation Delays in Uniform Linear Arrays
- •2.1.2 Narrowband Approximation
- •2.1.3 Matrix Equation for Array Data
- •2.1.4 Eigenstructure of the Spatial Covariance Matrix
- •2.2 ANTENNA BEAMFORMING BASICS
- •2.2.1 The Conventional Beamformer
- •2.2.2 The Minimum Variance Distortionless Response Beamformer
- •3.1 CLASSICAL METHODS FOR DIRECTION OF ARRIVAL ESTIMATION
- •3.1.1 Delay-and-Sum Method
- •3.1.2 Capon’s Minimum Variance Distortionless Response Method
- •3.2 SUBSPACE METHODS FOR DOA ESTIMATION
- •3.2.1 Multiple Signal Classification Algorithm
- •3.2.2 Orthogonal Vector Methods
- •3.2.3 The Root MUSIC Algorithm
- •3.2.4 The Minimum Norm Method
- •3.2.5 Estimation of Signal Parameters via Rotational Invariance Techniques
- •3.2.6 Linear Prediction
- •3.2.7 The Unitary ESPRIT for Linear Arrays
- •3.2.8 QR ESPRIT
- •3.2.9 Beamspace DOA Estimation
- •3.2.10 The DFT Beamspace ESPRIT
- •3.2.11 The Multiple Invariance ESPRIT
- •3.2.12 Unitary ESPRIT for Planar Arrays
- •3.2.13 Maximum Likelihood Methods
- •3.2.13.1 The Alternating Projection Algorithm for ML DOA Estimation
- •4.1 ADAPTIVE SIMULATION EXAMPLE
- •Appendix
- •Signal Generator
- •The MUSIC Algorithm
- •The ESPRIT Algorithm
- •MVDR Method and the Classical Beamformer
- •Code to Simulate the MUSIC, the ESPRIT, the MVDR, the Min-Norm, and the Classical DOA Algorithms
- •References
- •Additional References
- •List of Symbols
- •List of Acronyms
- •Author Biography

10 NARROWBAND DIRECTION OF ARRIVAL ESTIMATION FOR ANTENNA ARRAYS
of Rxx. Because stable SVD algorithms are available, several methods rely on decomposing the data matrix instead of diagonalizing the spatial covariance matrix to obtain the basis for the signal subspace. In practice, the N − r smallest eigenvalues will not be precisely σ 2; rather, they will all have small values compared to the signal eigenvalues. This is because the matrix Rxx is not known perfectly, but must be estimated from the data [12]. A common estimator for the spatial covariance matrix is the sample spatial covariance matrix, which is obtained by averaging rank-one data matrices of the form (xn xHn ), i.e.:
Rxx = |
1 |
KΣ−1xn xHn , |
(2.20) |
K |
|||
|
|
i = 0 |
|
where K is the total number of snapshots of data available from the sensors.
Although the discussion so far has focused on the uniform linear array, the principles of signal and noise subspaces also apply to other array geometries such as the uniform planar and the semispherical arrays.
2.2ANTENNA BEAMFORMING BASICS
In this section, a brief introduction to beamforming is provided. Classical direction of arrival estimation algorithms are based on beamforming concepts; therefore, this discussion is included to give the reader the necessary background to understand Sections 3.1.1 and 3.1.2.
If a weighted linear combination of the output of each element is taken, the array output can be computed by
y[n] = NΣ−1 wk xk [n] , |
(2.21) |
k = 0 |
|
where wk are the complex weights of the beamformer, which are shown in Figure 2.2. In vector notation, the output can be written as:
y[n] = wH xn, |
(2.22) |
where the N × 1 complex vector w contains the beamformer weights wk, k = 0, 1, …, N.
The array’s response to the incoming signals can be controlled by adjusting the elements of the weight vector wn. This process is referred to as spatial filtering or beamforming. Many methods exist to design or compute wn such that it produces a desired pattern. A beampattern is a plot of the gain of the beamformer in each possible direction.
Figure 2.3 shows a plot of the beampattern of a 10-element array where all the weights are equal to 1, for angles of arrival ranging from −90° to +90°.

Background on Array Processing 11
Figure 2.2: Narrowband beamforming system.
The magnitudes and weights of the beamformer determine the pattern and directivity of the antenna. Figure 2.2 and (2.10) can be used to write the output of the beamformer in the case that a single signal is present with angle of arrival θ.
|
1 N – 1 |
1 |
N – 1 |
|
1 N – 1 |
||
|
|
|
|
− j 2 nd sinθ |
|
|
− j 2 nd sinθ . (2.23) |
y[k] = N Σwn xn[k] = |
N Σwn s0[k]e |
= s0[k]N Σwn e |
|||||
|
|
n = 0 |
|
n = 0 |
|
|
n = 0 |
The signal s0[k] is scaled by the following function of θ :
W(θ0) = |
1 |
NΣ– 1 wne− jnω, and ω = 2 d sinθ, |
(2.24) |
N |
|||
|
|
n = 0 |
|
where W(θ) is known as the array factor or beampattern. notation as follows:
1 N −1
W(θ0) = N nΣ= 0 wne− jnω
The beampattern can be written in vector
= wH a(θ) , |
(2.25) |
where
w = |
1 |
* * |
* |
T , |
(2.26) |
N |
[w0 w1 |
· · · wN |
−1] |
|

12 NARROWBAND DIRECTION OF ARRIVAL ESTIMATION FOR ANTENNA ARRAYS
Figure 2.3: Polar plot of beampattern for a 10-element uniform linear array; D = λ /2.
is the vector containing the beamformer weights. The steering vector is defined as
┌ |
|
|
┐ |
|
a(θ) = │ |
1e− jω e− j 2ω . |
. . e− j(N− 1) |
T |
(2.27) |
ω│ . |
||||
└ |
|
|
┘ |
|
The beamformer output can also be written as a vector inner product as follows:
|
1 |
N −1 |
|
|
y[n] = |
Σ wkxk[n] = wH xn = s0[k]wH a(ș0) = s0[k]W(ș0). |
(2.28) |
||
N |
||||
|
|
k=0 |
|
The beamformer gain of the signal s0[k] is the beampattern evaluated at the angle of arrival corresponding to s0[k]. The beampattern gain can be evaluated as the vector inner product of the weight vector w and a(θ). If w and a(θ) are orthogonal, then wHa(θ) = 0 and hence the signal s0[k] is cancelled or nulled. Now, suppose two digitally modulated signals, s0(t) and s1(t), are present with angles of arrival θ0 and θ1, respectively. The beamformer output is given by
y[k] = s0[k]W(ș0) + s1[k]W(ș1) = s0[k]wHa(ș0) + s1[k]wHa(ș1). |
(2.29) |

Background on Array Processing 13
Figure 2.4: Demonstration of spatial aliasing with an eight-element uniform linear array with
D = 1.5λ.
If s0(t) is the signal of interest and s1(t) represents interference,w can be designed such that wHa(θ0) = 1 and wHa(θ1) = 0. The output is then
y[k] = s0[k]W(θ0) + s1[k]W(θ1) = s0[k] . |
(2.30) |
To avoid spatial aliasing in array processing, the element spacing must be at most λ /2 of the wavelength. If D > 0.5λ, aliasing occurs which manifests itself in terms of grating lobes. To illustrate spatial aliasing, we simulate a linear array with eight elements that are spaced 1.5λ apart. The beamformer weights are all equal to 1 and the beam pattern is shown in Figure 2.4. Because of aliasing, the array cannot distinguish between signals at 0°, −42°, and +42°.
2.2.1 The Conventional Beamformer
In the case of the conventional beamformer, all the weights are given a with a different phase, i.e.,
w = |
1 |
a(θ) = |
1 |
┌ |
e− jω e− j 2ω . . . e− j(N− 1)ω |
|
|
│1 |
|||
N |
|
||||
|
|
N └ |
|
magnitude of 1/N but each
┐T
│ . (2.31)
┘

14 NARROWBAND DIRECTION OF ARRIVAL ESTIMATION FOR ANTENNA ARRAYS
For a uniform linear array (Figure 1.3), the relationship between θ and ω is given in (2.24). Note that in the case where only one signal is present (r = 1) and neglecting noise, the beamformer output using the conventional beamformer is:
yk = wHxk = s0[k]wHa(θ) = s0[k] ( |
1 |
) a(θ)Ha(θ ) = s0[k]. |
(2.32) |
N |
In this case, the beamformer is said to be phase-aligned with the signal of interest and the signal of interest appears undistorted at the output. The signal-to-noise ratio (SNR) gain of the conventional beamformer can be computed by comparing the SNR at the output of a single element with the overall beamformer output. Assuming E[|vn[k]|2] = σ 2 and knowing |an(θ)|2 = 1, the SNR of the signal received by the kth element can be computed by first considering the signal model for element k.
|
xk[n] = ak(θ)s0 |
[k] + vk[n . |
|
|
(2.33) |
|||||
The signal power is given by |
|
|
|
|
|
|
|
|
|
|
s |
┌ |
0 |
|
2 |
┐ |
┌ |
|
2 |
┐ |
|
└| k |
| |
┘ |
└| 0 |
| |
┘ |
(2.34) |
||||
E |
= E│ a |
(θ)s |
[k] |
|
│= E│ s |
[k] |
|
│, |
and the noise power is |
|
|
┌ |
┐ |
|
En = E│|vn[k]|2 |
│= σ2. |
(2.35) |
└ |
┘ |
|
The SNR of the received signal by a single element is (Es /En)=(Es /σ 2).
Assuming E[vkvkH] = σ 2I and knowing a(θ)Ha(θ) = N, the SNR of the signal at the beam-
former output can be computed by first examining the beamformer output, i.e., |
|
||||||||||||||||
y = wH (a(θ)s |
[k] + v ) = s |
[k] + wH v . |
|
|
|
(2.36) |
|||||||||||
k |
|
|
|
|
0 |
|
|
|
k |
0 |
|
|
k |
|
|
|
|
The signal power is given by, |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
┌ |
|
|
┐ |
|
|
|
|
|
|
|
|
|
|
Es = E│|s0[k]|2 |
│ |
|
|
|
|
|
|
(2.37) |
||||||
|
|
|
|
|
|
|
└ |
|
|
┘ |
|
|
|
|
|
|
|
and the noise power is given by, |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
┌ |
|
2 |
┐ |
|
1 |
|
┌ |
|
|
|
|
┐ |
Nσ |
2 |
|
|
|
│ |
H |
│ |
│ |
|
|
│ |
|
H H |
|
|
. |
(2.38) |
|||||
│ |
|
|
|
|
|
|
|
|
|
||||||||
En = E w vk |
┘│ |
= |
N 2 |
E |
└ |
a(θ) |
vkv |
k |
a(θ)│= |
N 2 |
|
||||||
└│ |
|
│ |
|
|
|
|
|
┘ |
|
Hence, the SNR for the overall beamformer output is (Es /En)=(NEs /σ 2). The SNR has increased by a factor of N over the SNR at the output of a single element.