Received January 21, 2019, accepted January 28, 2019, date of publication February 1, 2019, date of current version

February 20, 2019.

Digital Object Identifier 10.1109/ACCESS.2019.2896979

Direction-of-Arrival Estimation for Both

Uncorrelated and Coherent Signals
in Coprime Array
HAIYUN XU , DAMING WANG, BIN BA, WEIJIA CUI, AND YANKUI ZHANG
National Digital Switching System Engineering and Technological Research Center, Zhengzhou 450001, China
Corresponding author: Bin Ba ([email protected])
This work was supported in part by the National Natural Science Foundation of China, under Grant 61401513.

ABSTRACT Direction-of-arrival (DOA) estimation in coprime array involves the problem that there is the
coexistence of both uncorrelated signals and coherent signals in the multipath environment. An approach
to estimate DOA of uncorrelated and coherent signals separately, based on the coprime linear array and
coprime planar array, is proposed in this paper. The uncorrelated signals are estimated first, where the root
multiple signal classification (root-MUSIC) is applied in the coprime linear array and unitary estimating
signal parameters via rotational invariance techniques (Unitary-ESPRIT) is utilized in the coprime planar
array. Those subspace algorithms are of low complexity compared with the spectral peak search method.
We then eliminate the components of noises and uncorrelated signals and reconstruct a covariance matrix
of coherent signals. Finally, we use the root-MUSIC or Unitary-ESPRIT to resolve the one-dimensional or
two-dimensional DOAs of coherent signals, respectively. The simulation results demonstrate the computa-
tional efficiency and high accuracy of the proposed algorithm. The results also prove that this algorithm can
estimate the number of signals more than that of subarray sensors and separate two signals from the same
angle.

INDEX TERMS Direction-of-arrival, coprime linear array, coprime planar array, coherent signals.

I. INTRODUCTION highly correlated signals. In uniform non-sparse array, spatial

Direction-of-arrival (DOA) estimation for multiple uncor- smoothing divides an array into multiple overlapping subar-
related signals is a fundamental task in many applications rays, and combines all covariance matrices of subarrays to
such as radar [1], underwater acoustics [2], [3], indoor nav- recovery the rank of the signals subspace [12], [13]. Besides
igation, and wireless communication [4]. In most researches, spatial smoothing, in uniform non-sparse array, a processing
the uniform non-sparse arrays, such as uniform linear array that estimates the DOAs of uncorrelated signals and coherent
and uniform rectangular array (URAs) [5], with subspace signals, separately, has a favorable performance [14]–[17].
algorithms can obtain high-resolution DOA estimations. The The methods are all based on the processing that estimates
multiple signal classification (MUSIC) [6] is a spectral peak the uncorrelated signals, then eliminates the components of
search method. However, it owns large computational com- them, and resolves the coherent signals at last. Thus, they can
plexity. The root-MUSIC [7], estimating signal parameters estimate the number of signals more than that of sensors.
via rotational invariance techniques (ESPRIT) [8], propaga- Nowadays, the coprime arrays, a kind of sparse array,
tion method (PM) [9], and Unitary-ESPRIT [10], [11] aim including coprime linear arrays [18]–[20] and coprime planar
to reduce the complexity efficiently. The algorithms men- arrays [21], [22], which are respectively applied to estimate
tioned above are all based on uncorrelated signals. However, one-dimensional (1D) and two-dimensional (2D) DOAs,
there are usually the coherent signals or highly correlated have become a focus. Compared with a uniform non-sparse
signals in multipath environment, which results in the low array, a coprime array has a larger aperture with the same
rank of signals subspace. Thus, the subspace algorithms number of sensors so that it can acquire a higher accu-
become invalid to estimate DOAs for coherent signals and racy. When the signals are coherent, the coprime array can
also have effective methods. We consider that the coprime
array consists of two uniform arrays, then we respectively
The associate editor coordinating the review of this manuscript and
approving it for publication was Yue Zhang. use spatial smoothing in each subarray, and apply common

2169-3536 2019 IEEE. Translations and content mining are permitted for academic research only.
18590 Personal use is also permitted, but republication/redistribution requires IEEE permission. VOLUME 7, 2019
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
H. Xu et al.: DOA Estimation for Both Uncorrelated and Coherent Signals in Coprime Array

peak search method to find the real DOAs [20]–[22]. Unfor-

tunately, one disadvantage of spatial smoothing is that it
reduces the maximum number of signals resolved. More-
over, common peak search also limits the number of signals
resolved less than that of subarray sensors. Hence, if we
combine those two methods to estimate coherent signals in
coprime array, we will not be able to estimate multiple sig-
nals with limited sensors. Two active estimation algorithms
FIGURE 1. Geometry of coprime linear array.
in coprime linear array of multiple-input multiple-output
(MIMO) system have been presented in [23] and [24]. But
the active detection has a higher cost than passive detection.
A passive estimation method in coprime linear array uses
fourth-order cumulants matrix (FCM), which can estimate
the number of signals more than that of sensors, but it needs
another supplementary sparse array [25].
In this paper, we introduce the methods from [15] and [17]
to coprime model to estimate both coherent and uncorrelated
signals, where we consider both coprime linear array and
coprime planar array. We present the algorithms steps and
prove that the methods in uniform non-sparse array can also
be applied to sparse array. Compared with spatial smoothing
in coprime array, we can estimate the number of signals big-
ger than that of subarray sensors. Compared with the existing
passive estimation algorithms, we need no supplementary
array in coprime linear array. Moreover, we have made some
FIGURE 2. Geometry of coprime planar array when M1 = 3, M2 = 4.
improvements to reduce the complexity. We use root-MUSIC
in coprime linear array. Partial spectral search (PSS) [21]
and 2D-ESPRIT [22] have been utilized in coprime planar
array, where the former is still based on spectral peak finding of coprime linear array is given by
and the latter needs to match parameters. Hence, we utilize h iT
ai (ϕ) = 1, e−jπMī sin ϕ , · · · , e−jπ(Mi −1)Mī sin ϕ , (1)
Unitary-ESPRIT in coprime planar array, which obtain the
lowest complexity than PSS and 2D-ESPRIT without param- where ϕ is the DOA of one signal. As for 2D coprime array,
eters matching. we combine two uniform rectangular arrays, where subarray
The paper is organized as follows. We first present two 1 has M1 × M1 sensors and subarray 2 has M2 × M2 sensors,
array models in Section 2. In Section 3, we explain our and the space distance between the two adjacent sensors is
approach to estimate DOAs, and show the steps of the pro- Mī λ/2. The steering vector of coprime planar array is denoted
posed algorithm. The computational complexity and the max- as
imum number of signals resolved analysis are presented in
Section 4. In Section 5, we show the results of simulations. bi (ϕ, θ) = ayi ⊗ axi , (2)
Finally, we summarize work in Section 6. Throughout this where
paper, IN represents the N dimensional unit array; (•)T , (•)∗ h iT
and (•)H respectively represent the transposition, conjuga- axi (ϕ, θ) = 1, e−jπMī sin θ cos ϕ , · · · , e−jπ(Mi −1)Mī sin θ cos ϕ ,
tion and conjugate transposition; diag [•] represents a vector
(3)
transforming to a diagonal matrix; ⊗ denotes the Kronecker h iT
−jπMī sin θ sin ϕ −jπ(Mi −1)Mī sin θ sin ϕ
product; d•e denotes the round up to the number. ayi (ϕ, θ) = 1, e ,··· ,e ,
(4)
II. SYSTEM MODEL and (ϕ, θ) are the azimuth angle and elevation angle of one
Considering the coprime array model, 1D and 2D array geom- signal, respectively.
etry are shown in Fig. 1 and 2, respectively. 1D coprime array Suppose that there are K narrowband far-field signals
is generally made up of two uniform linear arrays. Subarray impinging on the array. Assume P coherent signal groups,
1 has M1 sensors and subarray 2 has M2 sensors, where where the pth group has Lp signals. The coherent signal com-
M1 and M2 are the coprime integers (generally assuming ing from ϕp,` or (ϕp,` , θp,` ) is corresponding to the `th mul-
M1 < M2 ). We define the mth sensor location of the ith tipath propagation of Sp (t) with power σp2 for p = 1, · · · , P.
(i = 1, 2) subarray as (m − 1)Mī λ/2 (ī = 1, 2 and ī 6= i), The signals within each group are coherent to each other and
where λ denotes the wavelength. Hence, the steering vector uncorrelated to those in the different group. The total number

VOLUME 7, 2019 18591

 H. Xu et al.: DOA Estimation for Both Uncorrelated and Coherent Signals in Coprime Array

PP
of coherent signals is Kc = p=1 Lp . And the remaining However, to reduce the complexity burden, we apply root-
signals, Sk (t) coming from ϕk or (ϕk , θk ) with the power σk2 MUSIC [7] and Unitary-ESPRIT [10] to solve this problem.
for k = Kc + 1, · · · , K , are uncorrelated to each other. The In coprime linear array, through root-MUSIC, we can get
number of those signals is Ku = K − Kc . Thus, the received the estimation ϕ̂i,k (k = Kc + 1, · · · , K ). But because the
signals at the ith subarray can be represented as space between the adjacent sensors is more than half wave-
length, the ϕ̂i,k may be an ambiguous value. If we get the
Lp
P X
(1)
X one possible value of one signal, we can solute all possible
Xi (t) = ai ϕp,` βp,` Sp (t)


values, including the real value, by the relationship between
p=1 `=1
ambiguous values and real value. The relationship is given by
K
(1) Mī sin ϕi,real = 2m + Mī sin ϕi,ambiguous
X
+ ai (ϕk ) Sk (t) + ni (t)
k=Kc +1 (−Mī ≤ m ≤ Mī ). (12)
(1)
= Ai S (t) + ni (t) , (5) Then we define the two sets as
P Lp (1)
(2) 8i,k = {ϕ̂i,k,m |ϕ̂i,k,m = arcsin(sin ϕ̂i,k + 2m/Mī )}
XX
Xi (t) = bi ϕp,` , θp,` βp,` Sp (t)

p=1 `=1 (−Mī ≤ m ≤ Mī ), (13)

(1) (1) (1)
K
X (2)
8i = {8i,Kc +1 , · · · , 8i,K }. (14)
+ bi (ϕk , θk ) Sk (t) + ni (t)
k=Kc +1
Through finding the common values between the set 82,k
(2) and 81 , we can obtain the real DOA estimation ϕ̂kr of the
= Bi S (t) + ni (t) , (6) uncorrelated signal.
where the array manifold matrices are denoted as Similarly, as for coprime planar array, the possible estima-
tion (ϕ̂i,k , θ̂i,k ) can be resolved via Unitary-ESPRIT. Consider
Ai (ϕ) = [ai (ϕ1 ) · · · ai (ϕK )], (7) the relationship between ambiguous values and real value,
Bi (ϕ, θ) = [bi (ϕ1 , θ1 ) · · · bi (ϕK , θK )], (8) which is represented as (15), as shown at the top of the next
page. Thus, we can also have corresponding sets given by (16)
and βp,` is the complex fading coefficient of the `th coherent and (17), shown at the top of the next page.
signals in the pth group. The signal data vector Through finding the common values between the set 82,k
(2)
(2) (2) (2)
and 81 , and 22,k and 21 , we can calculate the real val-
T
S (t) = β1,1 S1 (t) , · · · , sK (t) ,

(9)
ues µ̂rk and ν̂kr . Furthermore, the DOA estimation (ϕ̂kr , θ̂kr )
where t = 1, · · · , J is sampling time and J is the number of (k = Kc + 1, · · · , K ) is denotes as
snapshots. And the noise vector is usually Gaussian random  q 
variables with zero means and variance σn2 . θ̂ r = arcsin µ̂r 2 + ν̂ r 2



k k k
(18)
ϕ̂k = arctan ν̂k /µ̂k .
 r r r


III. DOA ESTIMATION FOR BOTH UNCORRELATED AND
COHERENT SIGNALS
From (5) and (6), we can obtain the received signal covariance B. DOA ESTIMATION FOR COHERENT SIGNALS
matrix of the ith (i = 1, 2) subarray as 1) ELIMINATE THE COMPONENTS OF NOISES AND
UNCORRELATED SIGNALS
(1) 1 (1) (1) H
i + σn IMi ,
2
RXi = X X = Ai RS AH (10) We first eliminate the components of the noises, following the
J i i steps as:
(2) 1 (2) (2) H
i + σn IMi2 ,
2
RXi = Xi Xi = Bi RS BH (11) (1) (1)
J Ri = R̂Xi − σ̂n2i IMi ≈ Ai RS AH
i , (19)
where RS = SSH /J .
If there exist the coherent signals, (2)
Ri =
(2)
R̂Xi − σ̂n2i IM 2 ≈ i ,
Bi RS BH (20)
i
the rank of RS is P + Ku , which is smaller than K . The con-
ventional spatial smoothing is used to increase rank of signal where σ̂n2i is the average value of Mi − Ku − P smallest
(1)
subspace to K . However, The algorithms in [15] and [17] eigenvalues of R̂Xi or Mi2 − Ku − P smallest eigenvalues of
divide covariance matrix into two parts, according to uncorre- (2)
R̂Xi . Moreover, the RS can be rewritten as
lated and coherent signals, and just needs to enhance the rank
RS1
 
of covariance matrix of coherent signals to Kc . In coprime
array, the processing of algorithms is changed and shown in ..
.
 
 
following parts.  
 RSKc ,

RS =  (21)
σ 2


A. DOA ESTIMATION FOR UNCORRELATED SIGNALS Kc +1
..
 
.
 
Generally, the MUSIC and 2D-MUSIC can be applied to  
find DOAs in linear array and rectangular array, respectively. σK
2

18592 VOLUME 7, 2019

H. Xu et al.: DOA Estimation for Both Uncorrelated and Coherent Signals in Coprime Array

(
Mī sin θi,real cos ϕi,real = 2mx + Mī sin θi,ambiguous cos ϕi,ambiguous (−Mī ≤ mx ≤ Mī )
(15)
Mī sin θi,real sin ϕi,real = 2my + Mī sin θi,ambiguous sin ϕi,ambiguous (−Mī ≤ my ≤ Mī )
(
(2)
8i,k = {µ̂i,k,mx |µ̂i,k,mx = 2mx /Mī + sin θ̂i,k cos ϕ̂i,k , −Mī ≤ mx ≤ Mī }
(2) (16)
2i,k = {ν̂i,k,my |ν̂i,k,my = 2my /Mī + sin θ̂i,k sin ϕ̂i,k , −Mī ≤ my ≤ Mī }
( (2) (2) (2)
8i = {8i,Kc +1 , · · · , 8i,K }
(2) (2) (2) (17)
2i = {2i,Kc +1 , · · · , 2i,K }

Lp
P X K
γpm ejπ(n−1)Mī sin ϕp,` + σk2 ejπ(n−m)Mī sin ϕk
(1)
X X
Ri (m, n) = σp2 βp,`
∗
(23)
p=1 `=1 k=Kc +1

(1) (1) (1)∗

ri (m, n) = Ri (m, n) − Ri (Mi − m + 1, Mi − n + 1) (25)

Lp
P X
(2) ∗ mx ,my jπMī sin θp,` ((nx −1) cos ϕp,` +(ny −1) sin ϕp,` )

X
Ri mx , my , nx , ny = σp2 βp,` ηp e
p=1 `=1
K
σk2 ejπMī sin θk ((nx −mx ) cos ϕk +(ny −my ) sin ϕk )
X
+ (28)
k=Kc +1
Lp
jπMī sin θp,`0 (1−mx ) cos ϕp,`0 +(1−my ) sin ϕp,`0
 
m ,m X
ηp x y = βp,l 0 e (29)
`0 =1
(2) (2) (2)∗
mx , my , nx , ny = Ri mx , my , nx , ny − Ri




ri Mi − mx + 1, Mi − my + 1, Mi − nx + 1, Mi − ny + 1 (30)
P X Lp
m ,m
σp2 ζp,lx y ejπMī sin θp,` ((nx −1) cos ϕp,` +(ny −1) sin ϕp,` )
(2)
X
mx , my , nx , ny =


ri (31)
p=1 `=1

p,`0 ( y)
m ,my XLp   jπM sin θ (1−m ) cos ϕ + 1−m sin ϕ 
p,`0 p,`0
ζp,`x = β ∗
β
p,` p,` 0 − β β ∗
p,` p,` e 0
ī x
(32)
`0 =1

(1) (1)∗
where Ri (m, n) and that of Ri (Mi − m + 1, Mi − n + 1) as
βp,1 S1
  K
σk2 e−jπ((Mi −n+1)−(Mi −m+1))Mī sin ϕk
X
1  . h
 .  β ∗ SH · · · β ∗ SH
i
RSp =  .  p,1 1 p,Lp p k=Kc +1
J  
K
βp,Lp Sp
σk2 ejπ(n-m)Mī sin ϕk ,
X
= (24)
βp,1 βp,1
∗ βp,1 βp,2
∗ βp,1 βp,L
∗ k=Kc +1
 
··· p
(1)
 βp,2 βp,1
 ∗ βp,2 βp,2
∗ ··· βp,2 βp,L
∗
p

 For m, n = 1, · · · , Mi , we then define ri (m, n) as (25),
(1)
 σp .
 2
= .. .. .. .. shown at the top of this page, and ri (m, n) can be rewritten


 . . . . 
 as
βp,Lp βp,1
∗ βp,Lp βp,2
∗ ··· βp,Lp βp,L
∗ P X Lp
p (1) m jπ(n−1)Mī sin ϕp,`
X
(22) ri (m, n) = σp2 ξp,` e , (26)
p=1 `=1
(1)
Thus, the element of Ri in the mth row and nth column is where
expressed as (23), shown at the top of this page, where γpm = Lp 
jπ(1−m)Mī sin ϕp,`0 jπ(1−m)Mī sin ϕp,`0
PLp X 
`0 =1 βp,` e ξp,` βp,` βp,`0 − βp,` βp,` .
. Because the covariance matrix m ∗ ∗
0 = 0 e
of uncorrelated signals is a Hermite matrix, we can get the `0 =1
relationship between the uncorrelated signals component of (27)

VOLUME 7, 2019 18593

 H. Xu et al.: DOA Estimation for Both Uncorrelated and Coherent Signals in Coprime Array

Hence, we remove the components of uncorrelated signals C. ALGORITHM STEPS CONCLUSION

via (25). The main steps for the proposed algorithms in coprime lin-
This processing can also be applied to the 2D symmet- ear array and coprime planar array can be summarized as
(2)
ric rectangular array. The element of Ri in the (Mi (mx − follows:
1) + my )th row and (Mi (nx − 1) + ny )th column is denoted
as (28), as shown at the top of the previous page, where Algorithm 1 DOA Estimation Algorithm in Coprime Linear
m ,m
ηp x y is given by (29), shown at the top of the previous page. Array
For mx , my , nx , ny = 1, · · · , Mi , we define the new value 1) Calculate the covariance matrix via (5) and estimate
(2)
ri mx , my , nx , ny as (30), shown at the top of the previous the DOAs of uncorrelated signals using root-MUSIC

page, which can be rewritten as (31), as shown at the top of in each subarray.
m ,m
the previous page, where ζp,`x y is given by (32), shown at the 2) Find the common values of two subarrays and obtain
top of the previous page. Thus, we eliminate the component the real DOAs as ϕ̂kr (k = Kc + 1, · · · , K ).
of uncorrelated signals in coprime planar array via (30). 3) Eliminate the component of noises and uncorrelated
signals via (19) and (25).
(1)
2) COHERENT SIGNALS IN COPRIME LINEAR ARRAY 4) Define a new matrix ri and construct the new covari-
In order to construct covariance matrix of coherent signals, ance matrix of coherent signals via (33).
whose rank is Kc , we utilize the feature that the covariance 5) Utilize root-MUSIC to resolve the DOAs of coherent
matrix of linear array is a Toeplitz matrix, and reconstruct a signals in each subarray and find the common values
(1)
Toeplitz matrix based on ri (m, n). Define the new covari- of two subarrays as ϕ̂kr (k = 1, · · · , Kc ).
ance matrix as
 (1) (1)
· · · ri (m, Mi )


ri m, M̄i Algorithm 2 DOA Estimation Algorithm in Coprime Planar
(1)
RY (m) =  .. .. ..
. . .
 
i
 Array
(1) (1) 1) Calculate the covariance matrix via (6) and use
ri (m, 1) ··· ri (m, M̄i )
Unitary-ESPRIT to solve the possible values of each
= Āi 6 (1) (m) ĀH
i , (33)
subarray.
where M̄i = (M 2) Find the common values of two subarrays and obtain
(1)
i + 1)/2. If Mi is an even number, we just
the real DOAs as (ϕ̂kr , θ̂kr ) (k = Kc + 1, · · · , K ).
set ri m, M̄i = 0. The new array manifold Āii =
h 3) Eliminate the component of noises and uncorrelated
J ai ϕ1,1 , · · · , ai ϕP,LP , J = OM̄i ×(M̄i −1) , IM̄i
 ∗
∗


and
h i signals via (20) and (30).
(2)
6 (1) (m) = diag σ12 ξ1,1
m , · · · , σ 2ξ m
P P,LP . When Kc < M̄i , 4) Define a new matrix ri and Fi,ms , construct Yi , and
(1) calculate the new covariance matrix of coherent signals
the rank of RYiis generally Kc . (2)
(1) as RYi .
As a result, we can apply root-MUSIC to RYi to solve 5) Utilize Unitary-ESPRIT to resolve the DOAs of coher-
the closed-form solutions of DOAs. Through (13) and (14), ent signals in each subarray and find the common val-
we can calculate all possible DOAs and find the common ues of two subarrays as (ϕ̂kr , θ̂kr ) (k = 1, · · · , Kc ).
values ϕ̂kr (k = 1, · · · , Kc ).

3) COHERENT SIGNALS IN COPRIME PLANAR ARRAY

In order to construct the covariance matrix of coherent signals IV. ANALYSIS OF MAXIMUM NUMBER OF SIGNALS
based on a rectangular array, we first construct a new matrix RESOLVED AND COMPUTATIONAL COMPLEXITY
Fi,ms of size Ms12 × M as (34), shown at the bottom of the
s2
A. ANALYSIS OF MAXIMUM NUMBER OF SIGNALS
next page, where ms = 1, · · · , Ms2 with Ms1 +Ms2 −1 = Mi , RESOLVED
and 1 ≤ mx , my ≤ Mi . Arrange all Fi,ms to a new matrix Fi Here we analyze the maximum number of signals that can be
as Fi = [Fi,1 , · · · , Fi,Ms2 ]. Hence, construct the covariance resolved by the proposed algorithm. In coprime linear array,
matrix of coherent signals as using the proposed algorithm to estimate the uncorrelated
signals, we should meet the requirement that Ku +P ≤ M1 −1.
(2)
i .
RYi = Fi FH (35) To estimate the coherent signals, the requirement changes
to Kc ≤ d(M1 − 1)/2e. If P = 1, the Kc + Ku can reach
(2)
2 ≥ K and M 2 > K , the rank R is K , which sat-
When Ms2 c s1 c Yi c the maximum value M1 − 2 + d(M1 − 1)/2e. In conclusion,
isfies the requirement of using Unitary-ESPRIT. We provide the proposed algorithm can estimate more number of signals
Theorem 1, the proof of which is postponed into Appendix A. than the number of subarray sensors when there are coherent
And we can calculate (ϕ̂kr , θ̂kr ) (k = 1, · · · , Kc ) via (16),(17) signals.
and (18). Furthermore, in coprime planar array, when Ku + P ≤
2 ≥ K and M 2 > K , the rank of R(2)
Theorem 1: If Ms2 c s1 c Yi M12 − 1, we can first estimate the DOAs of uncorrelated
is Kc . signals. After constructing an Ms1 × Ms1 covariance matrix,

18594 VOLUME 7, 2019

H. Xu et al.: DOA Estimation for Both Uncorrelated and Coherent Signals in Coprime Array

we can estimate maximum Kc = Ms1 2 − 1 coherent signals,

2 , where
while the total coherent signals are no more than Ms2
Ms1 + Ms2 − 1 = M1 . The maximum K requires that
Ms1 and Ms2 are big enough, and P is small enough. Thus,
when Ms1 = d(M1 + 1)/2e, Ms2 = M1 + 1 − Ms1 , and
P = 1, the number of resolved signals is maximum as K =
min{Ms22 , M 2 − 1} + M 2 − 2, which can also be larger than
s1 1
the number of subarray sensors.

B. ANALYSIS OF COMPUTATIONAL COMPLEXITY

We analyze the computational complexity of proposed algo-
rithm, which is divided into two parts: coprime linear array
and coprime planar array.
We present an algorithm using root-MUSIC to reduce
the complexity in coprime linear array. Compared with
the algorithm using MUSIC, the difference of complexity
between them is the complexity of subspace algorithms.
The root-MUSIC
costs O M13 + M23 + (d(M1 + 1) /2e)3 +
(d(M2 + 1) /2e)3 while the
MUSIC

1 + 1) /
costs O (((d(M
2e)2 + (d(M2 + 1) /2e)2 Gϕ +O M12 + M12 , where Gϕ

denotes the number of spectral points.

Similarly, the difference of complexity between the
algorithm using Unitary-ESPRIT and algorithm using 2D-
MUSIC is the complexity of subspace algorithms. That of for-

mer is O M16 + M26 + (d(M1 + 1) /2e)6 + (d(M2 + 1) 6

/2e)



is O (dM1 /2e) + (dM2 /2e) Gϕ Gθ
4 4
and that of latter
+O M14 + M24 , where Gϕ , Gθ denote the number of spec-
tral points.
For the sake of clarity, we compare the complexities of
methods in Fig. 3, where we set M2 = M1 +1, 1θ = 1ϕ, and
Gθ = 90◦ /1θ + 1, Gϕ = 360◦ /1ϕ + 1. The figure proves
that root-MUSIC and Unitary-ESPRIT can truly reduce the FIGURE 3. Complexity comparison versus the number of subarray
complexity compared with spectral peak search methods. sensors. (a) Coprime linear array. (b) Coprime planar array.
Moreover, the spectral peak search method can obtain the
higher accuracy with a smaller searching step, which means
costing higher complexity. signals coexist. Consider that a coprime planar array is
also a planar array, thus, the unknown parameter vec-
tor in [26] has been changed to ψ = ϕ T , θ T , β Tr , β Ti ,

V. SIMULATION RESULTS T
where ϕ = ϕ1,1 , · · · , ϕP,LP , ϕKc +1 , · · · ϕK , θ =

A. CRAMER-RAO BOUND OF COPRIME ARRAY T
θ1,1 , · · · , θP,LP , θKc +1 , · · · θK , and β r , β i are the real

In this section, the Cramer-Rao bound (CRB) of coprime
array is presented, which is plotted as a benchmark. The part and the imaginary part  of the vector β =
T
β1,2 , · · · , β1,L1 , β2,2 , · · · , βP,LP . We denote the mth ele-

CRB for two-dimensional DOA estimation in planar array
has been derived in [26], when uncorrelated and coherent ments in ψ as ψm , thus the general expression of the (m, n)th

(2) (2)
mx , my , 1, ms mx , my , Ms2 , ms



ri ··· ri
 .. .. 

 . ··· . 

(2) (2)
ri mx , my , Ms1 , ms mx , my , Ms1 + Ms2 − 1, ms ,



··· ri
 
 
Fi,ms

= .. .. 
(34)
. ··· .

 
 (2) (2)
 ri mx , my , 1, Ms1 + ms − 1, mx , my , Ms2 , Ms1 + ms − 1



··· ri 
.. ..
 
. .
 
 ··· 
(2) (2)
mx , my , Ms1 , Ms1 + ms − 1, · · · ri mx , my , Ms1 + Ms2 − 1, Ms1 + ms − 1



ri

VOLUME 7, 2019 18595

 H. Xu et al.: DOA Estimation for Both Uncorrelated and Coherent Signals in Coprime Array

element in Fisher information matrix (FIM) can be expressed

as
∂ 2L
 
FIM ψm ψn = −E . (36)
∂ψm ∂ψn
Moreover, the FIMs corresponding to each vector included
in ψ are denoted as FIMϕϕ , FIMθ θ , FIMβr βr , and FIMβi βi ,
respectively, and to the cross terms between each vector are
FIMϕθ , FIMϕβr , FIMϕβi , FIMθ ϕ , FIMθβr , FIMθβi , FIMβr ϕ ,
FIMβr θ , FIMβr βi , FIMβi ϕ , FIMβi θ and FIMβi βr . As a result,
the whole FIM is given as
 
FIMϕϕ FIMϕθ FIMϕβr FIMϕβi
 FIMθϕ FIMθ θ FIMθβr FIMθβi 
FIMψψ =  FIMβr ϕ FIMβr θ FIMβr βr FIMβr βi .


FIMβi ϕ FIMβi θ FIMβi βr FIMβi βi

(37)
Define H = FIM−1ψψ , and we can obtain the CRBs of azimuth
angle and coherent angle, respectively, as
v
u
u1 X K
CRBϕ = t Hkk , (38)
K
k=1
v
u
u1 X 2K
CRBθ = t Hkk , (39)
K
k=K +1

where Hk,k denotes the (k, k)th element of H.

We can consider the linear array as a special planar array,
where only the
 azimuth can  be solved. Hence, we just need
change ψ = ϕ T , β Tr , β Ti , and the FIM and CRB of azimuth
angle are respectively given by
  FIGURE 4. Distribution of estimated values. (a) Coprime linear array.
FIMϕϕ FIMϕβr FIMϕβi (b) Coprime planar array.
FIMψψ = FIMβr ϕ FIMβr βr FIMβr βi . (40)
FIMβi ϕ FIMβi βr FIMβi βi same azimuth. The other is that the two signals, from the same
azimuth, are both uncorrelated signals or belong to the groups
v
u
u1 X K
CRBϕ = t Hkk . (41) of coherent signal. With the proposed algorithm, we can
K separate two signals in different steps in the former situation.
k=1
But we cannot obtain the results in the latter situation because
B. SIMULATION EXPERIMENTS their peaks are totally overlapped in spatial spectrum in the
To measure the accuracy of the algorithms, we define the root same estimating step.
mean square error (RMSE) as Hence, we assume that there are two coherent signals at
v ϕ1,1 = 0.7◦ and ϕ1,2 = 25.8◦ , and one signal uncorrelated
to the other two signals at ϕ3 = 0.7◦ with a signal-to-
u
Q
u 1 X 2
u
RMSE = t ξ − ξ̂ q . (42) noise ratio (SNR) of 15dB and J = 500. We set M1 = 7
QK
q=1 and M2 = 9. The results of proposed algorithm are listed
in Table 1. Through finding the nearest values between sets
where Q, ξ and ξ̂ q are the number of Monte Carlo simulations, of two subarrays, we can ensure that ϕ̂1,1 = 0.6979◦ , ϕ̂1,2 =
the real values and the qth estimated values, respectively. 25.82◦ and ϕ̂3 = 0.6993◦ .
Both conclusions of two situations can be applied to
1) DOA ESTIMATION OF TWO DISTINCT SIGNALS FROM THE coprime planar array, where the two signals come from
SAME ANGLE the same azimuth and elevation angle. The signals are
There are two possible situations when two distinct sources from (ϕ1,1 , θ1,1 ) = (45◦ , 45◦ ), (ϕ1,2 , θ1,2 ) = (64◦ , 54◦ ),
from the same azimuth (for simplicity we only consider and (ϕ3 , θ3 ) = (45◦ , 45◦ ), and M1 = 4, M2 = 5.
coprime linear array here). One situation is that one of the The estimated values are shown in Table 2. Obviously,
coherent signals and one uncorrelated signal come from the (µ̂r1 , ν̂1r ) = (0.5002, 0.5001), (µ̂r2 , ν̂2r ) = (0.3545, 0.7271),

18596 VOLUME 7, 2019

H. Xu et al.: DOA Estimation for Both Uncorrelated and Coherent Signals in Coprime Array

TABLE 1. The detailed values of results in coprime linear array.

TABLE 2. The detailed values of results in coprime planar array.

and (µ̂r3 , ν̂3r ) = (0.5002, 0.4998), and then (θ̂1,1 , ϕ̂1,1 ) =

(44.967◦ , 45.007◦ ), (θ̂1,2 , ϕ̂1,2 ) = (64.016◦ , 54.007◦ ), and
(θ̂3 , ϕ̂3 ) = (44.997◦ , 45.009◦ ).
In conclusion, the proposed algorithm is valid to estimate
the DOAs, even if there are two distinct signals from the
same angle, where one is an uncorrelated signal and the other
belongs to a group of coherent signals.

2) FEASIBILITY DEMONSTRATION WHEN THE NUMBER OF

SIGNALS IS BIGGER THAN THAT OF SUBARRAY SENSORS
In the second simulation, we consider the case of Ku = 5,
P = 1 and Kc = 3 with a SNR of 15dB in coprime linear
array, where M1 = 7 and M2 = 9. Next, we conduct
the simulation in coprime planar array where M1 = 4 and
M2 = 5. We set Ku = 14, P = 1, and Kc = 4 with a FIGURE 5. RMSE versus SNR in coprime linear array.

SNR of 15dB. The estimated values of 20 times Monte Carlo

simulations are shown in Fig. 4 (a) and (b). The figure shows
that the proposed algorithms can realize the estimation of the the range from 0.5◦ , 0.1◦ , and 0.01◦ .The results are shown
true DOAs, when there is the coexistence of both uncorrelated in Fig. 5 and Fig. 6, respectively.
and coherent signals, under the condition that the number of The Fig. 5 demonstrates that FCM is not sensitive to SNR.
signals is bigger than that of subarray sensors. Because there exists the fence effect in spatial spectrum and
the real angles are not located at the searching grid, the accu-
racy of MUSIC with 1ϕ = 0.5◦ do not improve when SNR is
3) RMSE COMPARISON UNDER DIFFERENT SNRs high. Moreover, the smaller searching step can acquire higher
In the third simulation, we compare the RMSE of proposed accuracy. Hence, the RMSE of MUSIC with 1ϕ = 0.01◦ is
algorithm using root-MUSIC with that of algorithm using smaller than that of MUSIC with 1ϕ = 0.1◦ , but the gap
MUSIC and FCM under SNRs from −5dB to 15dB at 5dB between them is not big. In low SNR, the RMSE of FCM is
intervals and J = 500, where M1 = 7 and M2 = 9. lowest, while that of root-MUSIC is highest. However, when
There are two coherent signals at ϕ1,1 = −40.6◦ and SNR is high, root-MUSIC has close accuracy as MUSIC
ϕ1,2 = 25.8◦ , and one signal uncorrelated to the other with small searching steps, and higher accuracy than FCM.
two at ϕ3 = 0.2◦ . Furthermore, we compare the proposed Although the RMSE of MUSIC with 1ϕ = 0.01◦ is closest
algorithm using Unitary-ESPRIT with the algorithm using to the CRB, considering the complexity, root-MUSIC is more
2D-MUSIC, where M1 = 4 and M2 = 5. Assume that computationally efficient than MUSIC.
two coherent signals come from (ϕ1,1 , θ1,1 ) = (25.3◦ , 28.1◦ ) As for coprime planar array, the RMSE of 2D-MUSIC with
and (ϕ1,2 , θ1,2 ) = (64.4◦ , 54.2◦ ), and one uncorrelated 1ϕ = 0.5◦ is highest due to fence effect. The proposed
signal comes from (ϕ3 , θ3 ) = (45.2◦ , 45.7◦ ). We conduct algorithm using Unitary-ESPRIT obtains higher RMSE than
100 simulations for each SNR and vary searching step in 2D-MUSIC with small searching steps in low SNR, but the

VOLUME 7, 2019 18597

 H. Xu et al.: DOA Estimation for Both Uncorrelated and Coherent Signals in Coprime Array

FIGURE 6. RMSE versus SNR in coprime planar array. (a) Elevation.

(b) Azimuth. FIGURE 8. RMSE versus J in coprime planar array. (a) Azimuth.
(b) Elevation.

4) RMSE COMPARISON UNDER THE DIFFERENT NUMBER

OF SNAPSHOTS
Consider situations used in simulation 3 again. The number
of snapshots is varied in the range from J = [20, 50,
100, 200, 500, 1000, 2000, 5000], when SNR = 15dB, and
the results are shown in Fig. 7 and Fig. 8, respectively. With
the increase of number of snapshots, the accuracy of the
proposed algorithms becomes higher. Moreover, the decline
gradually reaches a plateau. The figures also prove the
conclusions in simulation 3.

VI. CONCLUSIONS
FIGURE 7. RMSE versus J in coprime linear array. The paper has presented the DOA estimation algorithm in
coprime array when there is the coexistence of both coherent
gap between them becomes smaller with SNR increasing. and uncorrelated signals. The paper has described coprime
They have the close RMSE in high SNR. Also given the linear array model for 1D DOA estimation and coprime pla-
complexity, even though the 2D-MUSIC with 1ϕ = 0.01◦ nar array model for 2D DOA estimation. We then present the
is closest to the CRB, the Unitary-ESPRIT is more favorable associated algorithms, analyze the computational complexity
because it costs much lower complexity to acquire the highly and the maximum number of signals resolved, and conduct
precise estimated values. the simulation experiments, respectively. We prove that both

18598 VOLUME 7, 2019

H. Xu et al.: DOA Estimation for Both Uncorrelated and Coherent Signals in Coprime Array

2 ≥ K . Thus, the rank of R (2)

proposed algorithms in two models can estimate signals of is of full of rank when Ms2 c Yi is Kc ,
which the number is bigger than that of subarray sensors, which satisfies the requirement of using subspace algorithms.
when there exist the coherent signals. And compared with This completes the proof of Theorem 1.
spectral peak search method, the root-MUSIC in coprime lin-
ear array and Unitary-ESPRIT in coprime planar array have REFERENCES
greatly reduced the complexity. They obtain the higher RMSE [1] J. E. Ball and N. H. Younan, ‘‘Radar and radio signal processing,’’ Elec-
in low SNR than spectral peak search methods, while they tronics, vol. 6, no. 3, p. 64, 2017.
[2] Y. Wang, S. Ma, G. L. Liang, and Z. Fan, ‘‘Chirp spread spectrum of
can obtain the close accuracy when SNR is high. Moreover, orthogonal frequency division multiplexing underwater acoustic commu-
the proposed method can separate two signals coming from nication system based on multi-path diversity receive,’’ Acta Phys. Sinca,
the same angle, when one of them is a uncorrelated signal and vol. 63, no. 4, p. 044302, 2014.
[3] Y. Zhou, Y. Wu, D. Chen, and F. Tong, ‘‘Compressed sensing estimation
the other belongs to a group of coherent signals. of underwater acoustic MIMO channels based on temporal joint sparse
recovery,’’ J. Electron. Inf. Technol., vol. 38, no. 3, pp. 1920–1927, 2016.
APPENDIX [4] A. Gaber and A. Omar, ‘‘Sub-nanosecond accuracy of TDOA estimation
using matrix pencil algorithms and IEEE 802.11,’’ in Proc. IEEE. Int.
PROOF OF THEOREM 1 Symp. Wireless Commun. Syst., Aug. 2012, pp. 646–650.
The Fi,ms in (34) can be reformulated as [5] P. Heidenreich, A. M. Zoubir, and M. Rubsamen, ‘‘Joint 2-D DOA esti-
h i mation and phase calibration for uniform rectangular arrays,’’ IEEE Trans.
ζ Dxi ζ · · · Dxi s2 ζ ,
s −1 M −1
Fi,ms = Bci Dmyi (43) Signal Process., vol. 60, no. 9, pp. 4683–4693, Sep. 2012.
[6] R. O. Schmidt, ‘‘Multiple emitter location and signal parameter esti-
m ,my m ,m mation,’’ IEEE Trans. Antennas Propag., vol. 34, no. 3, pp. 276–280,
where ζ = [σ12 ζ1,1x , · · · , σP2 ζP,Lx P y ], and Mar. 1986.
  [7] D. Zhang, Y. S. Zhang, G. M. Zheng, C. Q. Feng, and J. Tang, ‘‘Improved
dx1,1 DOA estimation algorithm for co-prime linear arrays using root-MUSIC
.. algorithm,’’ Electron. Lett., vol. 53, no. 18, pp. 1277–1279, Sep. 2017.
Dxi =  .
 
[8] C. Zhou and J. Zhou, ‘‘Direction-of-arrival estimation with coarray

dxP,LP ESPRIT for coprime array,’’ Sensors, vol. 17, no. 8, p. 1779, 2017.
 −jπM sin θ cos [9] B. Byun and D. Yoo, ‘‘Improved direction of arrival estimation based on
e ī 1,1 ϕ1,1 
coprime array and propagator method by noise power spectral density
.. ,
estimation,’’ J. Adv. Navigat. Tech., vol. 20, pp. 367–373, 2016.
= .
 
[10] M. D. Zoltowski, M. Haardt, and C. P. Mathews, ‘‘Closed-form 2-D
e−jπMī sin θP,LP cos ϕP,LP angle estimation with rectangular arrays in element space or beamspace
via unitary ESPRIT,’’ IEEE Trans. Signal Process., vol. 44, no. 2,
(44) pp. 316–328, Feb. 1996.
  [11] M. Haardt and J. A. Nossek, ‘‘Unitary ESPRIT: How to obtain increased
dy1,1
estimation accuracy with a reduced computational burden,’’ IEEE Trans.
..
Dyi =  .
 
 Signal Process., vol. 43, no. 5, pp. 1232–1242, May 1995.
[12] T.-J. Shan, M. Wax, and T. Kailath, ‘‘On spatial smoothing for direction-
dyP,L of-arrival estimation of coherent signals,’’ IEEE Trans. Acoust., Speech,
 −jπM sin θ sin ϕ P  Signal Process., vol. ASSP-33, no. 4, pp. 806–811, Apr. 1985.
e ī 1,1 1,1
[13] C.-C. Yeh, J.-J. Lee, and Y.-Y. Chen, ‘‘Estimating two-dimensional angles
.. .
= . of arrival in coherent source environment,’’ IEEE Trans. Acoust., Speech
 
Signal Process., vol. 37, no. 1, pp. 153–155, Jan. 1989.
e−jπMī sin θP,LP sin ϕP,LP [14] F.-M. Han and X.-D. Zhang, ‘‘An ESPRIT-like algorithm for coherent
(45) DOA estimation,’’ IEEE Antennas Wireless Propag. Lett., vol. 4, no. 12,
pp. 443–446, Dec. 2005.
Bci is denoted as Bci = [ã∗yi (ϕ1,1 , θ1,1 ) ⊗ ã∗xi (ϕ1,1 , θ1,1 ), · · · , [15] X. Xu, Z. Ye, Y. Zhang, and C. Chang, ‘‘A deflation approach to direction
of arrival estimation for symmetric uniform linear array,’’ IEEE Antennas
ã∗yi (ϕP,LP , θP,LP ) ⊗ ã∗xi (ϕP,LP , θP,LP )], where ãxi = J̃axi , ãyi = Wireless Propag. Lett., vol. 5, no. 1, pp. 486–489, 2006.
J̃ayi , and J̃ = [IMs1 , 0Ms1 ,M −Ms1 ]. Thus, [16] Z. Ye and X. Xu, ‘‘DOA estimation by exploiting the symmetric config-
uration of uniform linear array,’’ IEEE Trans. Antennas Propag., vol. 55,
no. 12, pp. 3716–3720, Dec. 2007.
 M −1

1 dx1,1 · · · dx1,1s2 [17] X. Xu and Z. Ye, ‘‘Two-dimensional direction of arrival estimation by
. .. .. .. 
Fi,ms = Bci Dyi 6 (2)  . . ,
exploiting the symmetric configuration of uniform rectangular array,’’ IET
. . . (46)

Radar, Sonar Navigat., vol. 6, no. 5, pp. 307–313, Jun. 2012.
Ms2 −1 [18] P. Pal and P. P. Vaidyanathan, ‘‘Coprime sampling and the music algo-
1 dxP,Lp · · · dxP,LP
rithm,’’ in Proc. IEEE Digit. Signal Process. Workshop, IEEE Signal
Process. Educ. Workshop, Sedona, AZ, USA, Jan. 2011, pp. 289–294.
where 6 (2) = diag(ζ ). Next, Fi can be rewritten as [19] P. P. Vaidyanathan and P. Pal, ‘‘Sparse sensing with co-prime samplers
and arrays,’’ IEEE Trans. Signal Process., vol. 59, no. 2, pp. 573–586,
Fi = Bci 6 (2) Di , (47) Feb. 2011.
[20] C. Zhou, Z. Shi, Y. Gu, and X. Shen, ‘‘DECOM: DOA estimation with
where Di = [dy1,1 ⊗ dx1,1 , · · · , dyP,LP ⊗ dxP,LP ], dxp,` = combined MUSIC for coprime array,’’ in Proc. IEEE Int. Conf. Wireless
M −1 M −1
[1, · · · , dxp,`s2 ]T , and dyp,` = [1, · · · , dyp,`s2 ]T . At last, Commun. Signal Process. (WCSP), Oct. 2013, pp. 1–5.
[21] Q. Wu, F. Sun, P. Lan, G. Ding, and X. Zhang, ‘‘Two-dimensional
the covariance matrix of coherent signals is given by direction-of-arrival estimation for co-prime planar arrays: A partial spec-
(2) tral search approach,’’ IEEE Sensors J., vol. 16, no. 14, pp. 5660–5670,
i = Bci 6 Di Di 6 Bci .
(2) H (2)H H
RYi = Fi FH (48) Jul. 2016.
[22] W. Zheng, X. Zhang, and H. Zhai, ‘‘Generalized coprime planar array
Given the conclusion in [12], we can find that Bci is of full geometry for 2-D DOA estimation,’’ IEEE Commun. Lett., vol. 21, no. 5,
column rank when Ms12 > K , 6 (2) is of full rank, and D DH pp. 1075–1078, May 2017.
c i i

VOLUME 7, 2019 18599

 H. Xu et al.: DOA Estimation for Both Uncorrelated and Coherent Signals in Coprime Array

[23] S. Qin, Y. D. Zhang, and M. G. Amin, ‘‘DOA estimation of mixed coherent BIN BA received the M.S. and Ph.D. degrees
and uncorrelated targets exploiting coprime MIMO radar,’’ Digit. Signal from the National Digital Switching System
Process., vol. 61, pp. 26–34, Feb. 2017. Engineering and Technological Research Center,
[24] E. BouDaher, F. Ahmad, and M. G. Amin, ‘‘Sparsity-based direction find- Zhengzhou, China, in 2012 and 2015, respectively,
ing of coherent and uncorrelated targets using active nonuniform arrays,’’ where he is currently working in communications
IEEE Signal Process. Lett., vol. 22, no. 10, pp. 1628–1632, Oct. 2015. and information system. His main research inter-
[25] Y. Hu, Y. Liu, and X. Wang, ‘‘DOA estimation of coherent signals on ests include wireless communication theory, signal
coprime arrays exploiting fourth-order cumulants,’’ Sensors, vol. 17, no. 4,
processing, and parameter estimation.
p. 682, 2017.
[26] Y. Zhang et al., ‘‘Estimation of two-dimensional direction-of-arrival for
uncorrelated and coherent signals with low complexity,’’ IET Radar Sonar
Navigat., vol. 4, no. 4, pp. 507–519, 2010.

WEIJIA CUI received the M.S. and Ph.D. degrees

from the National Digital Switching System
HAIYUN XU received the B.S. degree from Engineering and Technological Research Center,
the National Digital Switching System Engineer- Zhengzhou, China, in 2001 and 2007, respectively,
ing and Technological Research Center (NDSC), where he is currently working in communica-
Zhengzhou, China, in 2016, where he is cur- tions and information system. His main research
rently pursuing the M.S. degree in communica- interests include wireless communication theory,
tions and information system. His main research satellite and mobile communication, and signal
interests include wireless communication theory, processing.
signal processing, and parameter estimation.

DAMING WANG was born in 1971. He is YANKUI ZHANG received the M.S. degree
currently a Professor with the National Digital from the National Digital Switching System
Switching System Engineering and Technologi- Engineering and Technological Research Center,
cal Research Center, Zhengzhou, China. His main Zhengzhou, China, in 2016, where he is currently
research interests include wireless communication pursuing the Ph.D. degree in communications and
theory, satellite and mobile communication, and information system. His main research interests
signal processing. include array signal processing and parameter
estimation.

18600 VOLUME 7, 2019