Scribd
Upload
57 views

Presented By: Ercherio Marpaung (14S15056) Dewi Fortuna Simangunsong (14S15007)

This document presents research on generalized correlation subspaces (GCS) for direction-of-arrival estimation. It begins with an introduction to correlation subspaces and reviews prior work. It then defines the GCS, which depends on a density function rather than exact source locations. Closed-form expressions are derived for the GCS of 1D and multidimensional arrays. Numerical examples demonstrate GCS for nested arrays and prior source interval knowledge. DOA estimation using GCS involves 3 steps: 1) computing the GCS, 2) denoising the covariance matrix, and 3) estimating DOAs from the denoised matrix.

Uploaded by

Ercherio Marpaung
Copyright
© © All Rights Reserved
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
57 views

Presented By: Ercherio Marpaung (14S15056) Dewi Fortuna Simangunsong (14S15007)

This document presents research on generalized correlation subspaces (GCS) for direction-of-arrival estimation. It begins with an introduction to correlation subspaces and reviews prior work. It then defines the GCS, which depends on a density function rather than exact source locations. Closed-form expressions are derived for the GCS of 1D and multidimensional arrays. Numerical examples demonstrate GCS for nested arrays and prior source interval knowledge. DOA estimation using GCS involves 3 steps: 1) computing the GCS, 2) denoising the covariance matrix, and 3) estimating DOAs from the denoised matrix.

Uploaded by

Ercherio Marpaung
Copyright
© © All Rights Reserved
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
You are on page 1/ 30

Presented by: Ercherio Marpaung (14S15056)

Dewi Fortuna Simangunsong (14S15007)


• INTRODUCTION
• REVIEW OF CORRELATION SUBSPACE
• GENERALIZED CORRELATION SUBSPACE
• PROPERTISE OF GENERALIZED CORRELATION SUBSPACES
• The Constant Density Function
• The Constat Density Function with Known Shown Interval

• GENEALIZATION CORRELATION SUBSPACE IN MULTIPLE DIMENSION


• NUMERICAL EXAMPLE
• Generalized Correlation Subspaces
• DOA Estimation With Nested Arrays and Prior Source Intervals
• Computational Complexity

• CONCLUDING REMARKS
DIRECTION-OF-ARRIVAL (DOA) estimation has been a popular research field in array processing for
several
decades. This problem aims to estimate the source directions from sensor measurements. It arises in
many practical scenarios such as radio astronomy, radar, imaging, and communications. DOA
estimators such as MUSIC, ESPRIT, MODE , and SPICE, to name a few have been developed for this
purpose. These estimators are
functional for ULAs. It is known that, for an ULA with N physical sensors, these DOA estimators can
identify at most N − 1 uncorrelated sources
Another family of DOA estimators is applicable to sparse arrays, where the sensors are placed
nonuniformly. Some well known sparse arrays include minimum redundancy arrays,mnested arrays, and
coprime arrays. All these arrays have O(N2 ) consecutive elements in the difference coarray, with N being
the number of physical sensors. Due to this property, sparse arrays can identifyO(N2 ) uncorrelated sources
using N physical sensors and they lead to better spatial resolution than ULAs. DOA estimators for sparse
arrays consist of the augmented covariance matrix method, positive definite Toeplitz completion,spatial
smoothing MUSIC (SS MUSIC), dictionary methods and gridless approaches
The elegant concept of correlation subspaces was proposed by Rahmani and Atia, to improve DOA estimation
in a number of ways. For any given array geometry, the correlation subspace is uniquely determined, and
imposes some implicit constraints on the structure of the covariance, as we shall see. This subspace can be
utilized in denoising the sample covariance matrix. Then the source directions are estimated from the denoised
covariance matrix using off-the-shelf DOA estimators, such as the MUSIC algorithm. Note that the correlation
subspace depends on the array configurations and prior knowledge about the sources but is independent of the
choice of DOA estimators. Hence, a broad class of DOA estimators are applicable to the denoised covariance
matrix. However, the explicit expressions for the correlation subspace were not known, so its approximationwas
computed numerically. Furthermore, theway in which the correlation subspace is influenced by the array
configuration, and by partial knowledge about sources, was not explored.
The original work on correlation subspaces did not emphasize any knowledge about the sources
except that they are uncorrelated.
1. We show how to incorporate prior knowledge about source intervals with generalized
correlation subspaces. Furthermore, correlation subspaces for 1D and 2D arrays were analyzed
numerically in [while our work provides closed-form characterizations of generalized correlation
subspaces for 1D and multidimensional arrays.
2. Finally, in our work, covariance matrix denoising using the generalized correlation subspace is
implemented more efficiently than that in
Assume that D monochromatic sources impinge on an onedimensional sensor array

The sensor locations are ‘nd’, where n belongs to an integer set S ⊂Z, d = λ/2, and λ is the wavelength.
Let θi ∈[−π/2,π/2] be the DOA of the ith source
The normalized DOA of the i th source is defined as θi =(d/λ)sinθi ∈[−1/2,1/2].

It is assumed that the sources and noise are zero-mean and uncorrelated.
Namely,lets =[A1,...,AD,𝑛𝑆𝑇 ]T THEN E[s]= 0 E[ss] = diag(p1,...,pD,pn,...,pn),
&
Rearranging thematrix
The covariance elements leads
of xS to expressed as
can be

where the correlation vectors c(θi) are defined as


where the correlation
The correlation subspace
subspace matrix S computed
was originally is defined by
as the following definition
To begin with, consider finite snapshot sensor
measurements
x(k) for k = 1, . . . , K.
The algorithm first denoises the sample covariance matrix
R using the following convex program (P1):
The algorithm first denoises the sample covariance matrix R using the following convex program
(P1):

∗
The final stage is to perform DOA estimation on 𝑅𝑝1 It was shown in [M. Rahmani and G. K.
Atia, “A subspace method for array covariance matrix estimation,” in Proc. IEEE Sensor
Array Multichannel Signal Process. Workshop, Jul. 2016, pp. 1–5.] that, the MUSIC algorithm
∗
on 𝑅𝑝1 can exhibit better estimation performance than that on Rs. It should be noted that the
DOA estimators are not restricted to the MUSIC algorithm. Other estimators, such as
ESPRIT, MODE, and SPICE, can also be exploited. This result shows that the structure of
the covariance matrix, as specified by the correlation subspace, helps to improve the
estimation performance.
Summarizing, the DOA estimator associated with the correlation subspace is
composed of the following three steps

Step 1: Correlation subspace

Step 2: Denoising.

Step 3: DOA estimation


In order to derive closed-form expressions for CS, we will first propose the generalized correlation
subspace GCS(ρ), which is parameterized by a nonnegative density function ρ(θ). Then it will be
shown that GCS(ρ) depends only on the support of ρ(θ), and is invariant to its exact shape. Using
this property, we can derive simple and closed forms for CS by selecting some density functions
ρ(θ) such that 1) CS = GCS(ρ) and 2) it is straight forward to derive closed-form expressions for
GCS(ρ).

the generalized correlation subspace is invariant to the exact shape of ρ(θ)


under that
support, as indicated by the following lemma:
The density functions ρ1 (θ) and ρ2 (θ) are illustrated in Fig. 1. It can be
observed that these density functions share the same support [−1/2, 1/2].
Furthermore,Corollary 1 also enables us to analyze the correlation
subspace readily through the generalized correlation subspace GCS(ρ2 ).
Example 1: Assume the sensor locations are characterized by an integer set S = {0, 2}.
According the difference coarray becomes D = {−2, 0, 2}. Next we will evaluate the matrix J,.
First we consider the matrices I(m) for m ∈ D as follows
Finally, we will verify in this example. Starting the correlation vector is given
by

VS
In this case, the entry of SD(1[− 1/2,1/2]) associated withcoarray locations m1,m2 ∈ D becomes

Proof: Itwas proved in [44] that the columns of J are


orthogonal.
It suffices to consider the norms of the individual columns
of J. For the coarray locationm ∈ D, the norm of the
associated which proves this lemma
column is
Example 2: Assume the sensor locations are given by S = {0, 2}, as in Example 1.

Then the weight functions w(m) are given by

the generalized correlation subspace matrix S(1[−1/2,1/2]) can be expressed as

indicates that the matrix 𝐉𝐖 −1/2 Type equation here. corresponds to the orthonormal
eigenvectors while the diagonal matrix W is associated with the eigenvalues. In particular,
the positive eigenvalues and the associated eigenvectors of S(1[−1/2,1/2]) are given by
Properties of S(1[−1/2,1/2]): The following list some properties regarding the eigen-structure of S(1[−1/2,1/2]).
Some items are direct consequences of the properties of the weight function w(m), as in [46,
Lemma 1]
the generalized correlation subspace GCS(1[−1/2,1/2]) becomes

Theorem 1: Let the matrix J be defined as in Definition 5.


Then the correlation subspace satisfies

The dimension of the correlation subspace is


the size of the difference coarray, i.e., dim(CS) = |D|.
Example 3: Consider the super nested array with N1 = N2 = 5[21]. The sensor locations are depicted in
Fig. 2(d). The
difference coarray becomesD = {−29, . . . , 29} (hole-free) and |D| = 59.We also choose the parameter α
= 0.1. The eigenvalues λk and the eigenvectors are illustrated in Fig. 3 . It was shown in [37] that the
eigenvalues λk are
distinct, and the first α|D| eigenvalues are close to, but less than one, where · is the floor function. As the
index k exceeds α|D|, the magnitude of the eigenvalues decays exponentially. This property indicates
that the matrix SD(1[−α/2,α/2]) can be well-approximated by a matrix of rank L. Namely,
Theorem 2: Let the density function be 1[−α/2,α/2] for 0 < α < 1. Assume the difference coarray D is
hole-free. Then we have

Hole-Free Difference Coarrays and I = [θmin, θmax]: Let us consider another scenario where
the known interval is notccentered around the origin, such as [θmin, θmax] with −0.5 <θmin < θmax
< 0.5. In this case, the density function can be
the generalized correlation subspace
NUMERICAL EXAMPLE
 Generalized Correlation Subspaces

 DOA Estimation With Nested Array and Prior


Source Intervals
The estimation performance is measured in terms of
root mean-squared errors (RSME), defined as
 Computational Complexity
CONCLUDING REMARKS

 Generalized correlation subspaces with applications


to DOA estimation for uncorrelated sources and
known source intervals.
 Generalized correlation subspaces have closed-form
characterizations in terms of array profiles such as
sensor locations and difference coarrays
Chun-Lin Liu and P. P. Vaidyanathan, Fellow “Correlation Subspaces:
Generalizations and Connection to Difference Coarrays” Digit. Signal
Process., July. 2017.

You might also like