wavelength of the carrier and d is the distance between two
sensors, s
k
(t)isthekth signal received at the reference
sensor, and n
m
(t) is the underlying noise at sensor x
m
(t).
The vector form of (9) can be expressed as
xðtÞ¼ΑðθÞsðtÞþnðtÞð10Þ
where A(θ) is the M P matrix of the array steering
vectors A(θ)¼[a(θ
1
),a(θ
2
),…,a(θ
p
)], in which aðθ
k
Þ¼
½1; e
jð2π=λÞd sin ðθ
k
Þ
; …; e
jð2π=λÞdðM 1Þ sin ðθ
k
Þ
T
, x(t) is the M 1
vector of signals received by the array sensors x(t)¼ [x
1
(t),
x
2
(t),…,x
M
(t)]
T
, s(t) is the P 1 vector of the signals s(t)¼
[s
1
(t),s
2
(t),…,s
p
(t)]
T
, and n(t) is the M 1 vector of the
noise n (t)¼[n
1
(t),n
2
(t),…,n
M
(t)]
T
.
3.2. Noncircular signals
Circularity can be derived from the geometrical inter-
pretation of signals [17]. A complex random variable is said
to be circular if its statistical properties are rotational
invariant for an arbitrary phase ψ. In this paper, we only
consider the first and the second order statistical proper-
ties of the signals. A complex random signal z is said to be
circular at the order 2 if both of the mean E{z} and the
elliptic covariance E{z
2
} equal zero. Since the signals are
often assumed to be zero-mean, the noncircularity of z can
be equated to Efz
2
ga0. Circularity is a common hypoth-
esis for the narrowband signals in the classic direction
finding algorithms, but we still can easily find numerous
noncircular signals like AM (amplitude modulated) or
BPSK signals.
For an arbitrary signal z, we acquire
Efz
2
g¼ρe
jϕ
Efzz
n
g¼ρe
jϕ
s
2
z
ð11Þ
where ϕ is the noncircularity phase and ρ is the noncircu-
larity rate which satisfies 0r ρ r 1. Obviously, circular
signal is the special case with ρ¼ 0. In this paper, we
consider only the noncircular signals with ρ¼ 1, e.g., BPSK
and AM signals. In this case, the noncircular signal z can be
expressed as [6]
z ¼ e
ðjϕ=2Þ
z
0
ð12Þ
where z
0
is the zero phase version of z.
Therefore, for the noncircular signal z, we can incorpo-
rate the unconjugated spatial covariance function E{z
2
}
with the conventional covariance function E{zz
*
} to gain
more abundant second order statistical information.
4. Proposed solution in the SαS framework
4.1. The extended covariation based NC-MUSIC algorithm
In this section, we will derive the extended covariation
based matrix of the sensor outputs and show it can be
applied with subspace techniques to gain the bearing
information for noncircular sources under SαS noise
environments.
Considering the noncircular signals with the noncircu-
larity rate ρ¼1, from Eq. (12), we can describe the signal
s(t) in Eq. (10) as
sðtÞ¼Φ
1=2
s
0
ðtÞð13Þ
where s
0
(t)¼[s
01
(t),s
02
(t),…,s
0P
(t)]
T
, in which the fs
0i
ðtÞg
P
i ¼ 1
is the corresponding zero phase version of the signal
fs
i
ðtÞg
P
i ¼ 1
, and Φ
1=2
¼ diagfe
jϕ
1
=2
; e
jϕ
2
=2
; …; e
jϕ
P
=2
g, in which
fϕ
i
g
P
i ¼ 1
is representing the noncircularity phase of the
signal fs
i
ðtÞg
P
i ¼ 1
. Suppose only noncircular signals are
involved in s(t), we can form the extended array output
x
nc
(t) as the combination of x(t) and its conjugate version
x
*
(t)
x
nc
ðtÞ¼
xðtÞ
x
n
ðtÞ
"#
ð14Þ
Definition 1. The 2M 2M extended covariation matrix
C
nc
of the extended output vector x
nc
(t) can be defined as
C
nc
¼
C
1
C
2
C
3
C
4
"#
ð15Þ
where C
1
, C
2
, C
3
, and C
4
are the four M M sub-matrices of
C
nc
, and the corresponding (i,j)th entries of these sub-
matrices are taking the values [x
i
(t),x
j
(t)]
α
, ½x
i
ðtÞ; x
n
j
ðtÞ
α
,
½x
n
i
ðtÞ; x
j
ðtÞ
α
and ½x
n
i
ðtÞ; x
n
j
ðtÞ
α
, respectively.
Theorem 1. The matrix form C
nc
can be expressed as
C
nc
¼
A0
0A
n
Γ
0s
0
0 Γ
0s
"#
I
P
Φ
Φ
n
I
P
"#
A0
0A
n
H
þκ
n
I
2M
ð16Þ
where Γ
0s
¼diag{γ
01
,γ
02
,…,γ
0P
} is the diagonal covariation
matrix for the signal s(t), in which γ
0k
¼[s
0k
(t),s
0k
(t)]
α
k¼1,2,…,P. κ
n
¼[n
i
(t),n
i
(t)]
α
i¼1,2,…,M is the covariation
of the noise. See Appendix A for the proof of Theorem 1.
We can further simplify the expression (16) as
C
nc
¼
A
A
n
Φ
n
Γ
0s
A
A
n
Φ
n
H
þκ
n
I
2M
ð17Þ
Clearly, the expression for the extended covariation
matrix C
nc
is identical to the well-known expression R
nc
which stands for the extended covariance matrix in
NC-MUSIC [4]
R
nc
¼
A
A
n
Φ
n
R
0s
A
A
n
Φ
n
H
þs
2
n
I
2M
ð18Þ
where R
0s
is the diagonal signal covariance matrix and
s
2
n
is the variance of the underlying noise.
Accordingly, similar to NC-MUSIC, we can formulate the
extended steering vector
B ¼
A
A
n
Φ
n
¼½bðθ
1
; ϕ
1
Þ; bðθ
2
; ϕ
2
Þ; …; bðθ
P
; ϕ
P
Þ ð19Þ
where bðθ
i
; ϕ
i
Þ¼½
aðθ
i
Þ a
n
ðθ
i
Þe
jϕ
i
T
i ¼ 1; 2; ⋯; P. The
DOAs can be obtained by minimizing the following spatial
spectrum g(θ,ϕ)overθ and ϕ
gðθ; ϕÞ¼b
H
ðθ; ϕÞV
n
V
H
n
bðθ; ϕÞ; ð20Þ
in which V
n
are the corresponding left singular vectors
associated with the noise subspace of C
nc
. To alleviate the
computational complexity, like NC-MUSIC, the 2-D search
of g(θ,ϕ)overθ and ϕ can be substituted with the 1-D
search only over θ for the local minima of the following
Z. Jinfeng, Q. Tianshuang / Signal Processing 98 (2014) 252–262254