2D DOA Estimation for Uniform Rectangular Array

With One-bit Measurement

Yang Xiong Zeyang Li Fangqing Wen

School of Electronic and Information School of Electronic and Information School of Electronic and Information
Yangtze University Yangtze University Yangtze University
Jingzhou, China Jingzhou, China Jingzhou, China
201975086@[Link] 201700742@[Link] wfqltt@[Link]

Abstract—Direction-of-arrival (DOA) estimation is an which opens a new window for signal acquisition and
interesting research topic with various applications. Existing processing. Up to now, various CS-based sampling
algorithms provide superior estimation performance, at the cost frameworks have been developed to sample the analogy signal
of accurate quantified measurements. In this paper, we stress at sub-Nyquist rate, for example, random sampler, random
the problem of 2D DOA estimation for uniform rectangular demodulation, modulated wideband converter, time encoding
array using one-bit measurements. The relationship between the machine. A common characteristic that these architectures
covariance matrices of one-bit measurement and that of the share is that they need high-precision quantification. Unlike
accurately quantified measurement is analyzed in detail, from the above framework, the concept of one-bit CS has been
which we find the existing tensor algorithm can be directly
proposed [15]. One-bit CS refer to one-bit quantization, which
applied. As a result, a one-bit parallel factor analysis
(PARAFAC) estimator is proposed. Simulation results show the
means only one data bit need to be stored and processed, the
effectiveness of the proposed method. system complexity can be reduced accordingly. Due to this
reason, one-bit quantization is really suitable for massive
Keywords—One-bit quantification, array signal processing, multiple-input multiple-output (MIMO) systems [16]. As the
parallel factor analysis (PARAFAC), 2D DOA estimation. spectrum estimation with one-bit measurements is a sparse
inverse problem, signal processing problem with one-bit
I. INTRODUCTION measurements may suffer from lower computational
efficiency. More recently, a one-bit MUSIC framework was
Direction-of-arrival (DOA) estimation is one of the most
driven in [17]. It proven that the covariance matrix with one-
important branches of array signal processing [1]-[5]. It is one
bit measurements can be approximated by a scaled
of the key techniques in wireless communications, radars,
unquantized covariance matrix, thus the traditional subspace
sonars, et al. In the past decades, various super-resolution
algorithms can be directly applied. However, as it mentioned
estimators have been derived, for instance, multiple signal
in [17], one-bit quantization is lossy, so the performance of the
classification (MUSIC) [6], estimating signal parameters via
one-bit subspace algorithms will degrade.
rotational invariance technique (ESPRIT) [7], propagator
method [8], maximum likelihood (ML) [9], tensor-based Uniform rectangular array is a common array geometry.
approaches [10][11] and optimization-aware algorithms [12]. The DOA estimation problem in such scenario always involve
As is well known to us, MUSIC is computationally inefficient 2D spectrum estimation. In this paper, we stress the problem
as it requires extensive spectrum peak searching. Nor does ML of 2D DOA estimation for uniform rectangular array try with
estimator is efficient, since it needs exhaustive iteration. In one-bit measurements. To enhance the accuracy of parameter
comparison to MUSIC and ML, ESPRIT can obtain closed- estimation, the tensor based one-bit framework is proposed.
form solution, thus it is much more efficient than the latter. Firstly, the relationship between the covariance matrices of
However, in both MUSIC and ESPRIT, eigendecomposition one-bit measurement and that of the unquantized covariance
is unavoidable to obtain the signal subspace or the noise matrix is extended to tensor domain. We show that the array
subspace. Generally speaking, the complexity of one-bit data fulfills a parallel factor (PARAFAC) model.
eigendecomposition is on the third order of the matrix Therefore, a one-bit PARAFAC algorithm is proposed for 2D
dimension. In order to avoid this drawback, the propagator DOA estimation, in which closed-form and automatically
method is often recommend, which can obtain a similar paired parameters are achieved. Compared with the closed-
subspaces to the traditional eigendecomposition via least form one-bit ESPRIT algorithm, the proposed algorithm
squares (LS) method. Optimization-based algorithms can offers more accurate estimation performance. Numerical
obtain accurate spectrum estimation. Nevertheless, they are simulations are provided to verify the effectiveness of the
often too complex to practical engineering. Besides, tensor proposed framework.
approaches are often superior than the above matrix based
methods, as they have better de-noising performance than the Notation: Notation, capital letters, e.g., X , and lowercase
latter. Due to this reason, tensor estimator has drawn many letters, e.g., x in bold denote, respectively, matrices and
attentions for multiple parameter estimation. vectors. The superscript ()T , () H , () 1 and ()† represent
the operators of transpose, Hermitian transpose, inverse and
It should be noticed by the reader that the above mentioned pseudo-inverse, respectively. angle    returns the phase of a
algorithms are rely on accurate quantified measurements,
which means high-precision sampling is necessary. Although vector in radian. X  m, n  denote the  m, n  -th element of X
high-resolution quantization is prefect in terms of signal .  stands for the Kronecker product. The Khatri-Rao
recovery performance, it is likely to be impractical due to the product (column-wise Kronecker product) is denoted by  .
high hardware cost and system power consumption. Recently, The identity matrix is denoted by I .
compressed sensing (CS) theory has been proposed [13][14],
XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE
Authorized licensed use limited to: Southeast University. Downloaded on November 28,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.
 II. ONE-BIT QUANTIZATION AND SIGNAL MODEL  k2
where  k is defined as  k  . Because of all sensors are
As shown in Fig.1, we consider a uniform rectangular  n2
array with M  N -element sensors, which are omni- identical, a y , n  k ,k   a x , m  k ,k  =1 . Therefore, Eq.(3) can
directional with inter-element distance is d. Suppose that there
K
are K uncorrelated narrow-band sources appearing on the far- be simplified to  y2   Ry  mnmn   k2   n2 . Additional
field of the array. The angle pair is denoted by (  k ,k ), k 1

k  1,2,, K , where  k and k represent the elevation angle assumption is that the powers of the signals are the same, i.e.,
and azimuth angle of the k-th source, respectively. Besides,  k = , then we have
we assume that source signals and noise are uncorrelated, and K

 a  ,    a  ,   a  ,    a  , 

*

both of them are modeled as independent, zero-mean, circular, y ,n1 k k x ,m1 k k y ,n 2 k k x ,m 2 k k

y  k 1
complex Gaussian random processes. Then the received signal m1, n 1 ym 2, n 2
K   1
of the (m,n)-th sensor can be expressed as (5)
z which implies  y  1 . More importantly, it can be find
m1, n1 ym 2 ,n 2

that the magnitudes with respect to the real part and

imaginary part of the correlation coefficient decreases as 
decreases.
 Next, we focus on Rx . It has been pointed out that
0 y xm , n  t  is zero-mean, unit variance [17], i.e., E  xm , n  t   0 ,
 
 2 x =1 , the ( m1n1 , m2 n2 )-th entity of Rx can be expressed as
m ,n

=  Rx m n , m n . It is obvious that if m1n1  m2 n2 ,

M
x m1, n1 xm 2 ,n 2 1 1 2 2
subarray1

x subarray2 subarray N x m1,n1 xm 2 , n 2

  Ry  m n , m n  1 . According to the inverse sine law
1 1 2 2

Fig.1 Illustration of 2D DOA estimation for uniform

x 

2 arcsin e  ym1,n1 ym 2 ,n 2  , we have x =
rectangular array m1,n 1 xm 2 , n 2
 m 1, n1 xm 2 , n 2

ym , n  t     a y , n 
K


, k   a x ,m  k , k  sk  t   nm , n  t 
2
 
arcsin real  y y  
 j arcsin imag  y m1,n 1 m 2 ,n 2
   m1,n 1 ym 2 , n 2
  . Hence
k 1
k
(1) the relationship between Rx and Ry is
 m  1, 2,  , M ; n  1, 2,  , N 
2 1 
a x ,m  k , k   e
j 2  m 1 d sin  k cos k / 
, ay, n  k ,k   e j 2  n 1 d sin  sin k /  Rx  arcsin e  R y  (6)

k
a 
donate the spatial response of the k-the signal with respective  
to x-axis and y-axis, respectively.  is the wavelength. Now we can get the relation that Ry  a sin e  Rx  , where
2  
nm , n  t  is the associate noise. The one-bit quantized
sin e z sin  real z  j sin  imag z , a is an unknown scaler.
measurement is defined as
Furthermore, it is proven that in [17] in SNR is sufficiently
xm , n  t  
1
2

sign   ym, n  t     jsign   ym , n  t     (2) low, then
2 Ry
 2
Rx 
 1   I (7)
where if   ym , n  t    0 , sign   ym, n  t    returns 1, otherwise   a
Although a is unknown, Rx and R y share the same
it returns -1.   ym , n  t   and   ym , n  t   represent,
eigenvectors. As a result, the traditional subspace-based
respectively, the real part and imaginary part of the complex
algorithms, such as MUSIC and ESPRIT, can be directly
value. Now we focus on the relationship between the
applied to Rx for 2D DOA estimation. Now we consider L
covariance matrix Ry of ym, n  t  and the covariance matrix
one-bit snapshots are available. Moreover, the one-bit
Rx of the one-bit measurement xm , n  t  . The ( p, p )-th ( quantified data is stacked into a matrix as
p  1, 2, , mn ) diagonal element of Ry is X  O   Ay  Ax  S T  E  (8)
K
 y2   R y  mnmn   a y , n  k ,k   a x , m  k ,k   k2   n2 Ax = a x 1 ,1  , a x  2 ,2  ,, a x  K ,K     M  K
2
(3) where ,
k 1

where  k2 and  n2 account for the power of the k-th signal and Ay = a y 1 ,1  , a y  2 ,2  ,, a y  K ,K    N K
are the
noise, respectively. If m1n1  m2 n2 , the correlation coefficient response matrices, with the steering vectors are
of ( m1n1 , m2 n2 )-th element of R y can be expressed as follows a x  k ,k  = a x ,1  k ,k  , a x ,2  k ,k  ,, a x ,M  k , k   and
K * a y  k , k  = a y,1  k , k  , ay,2  k ,k  ,, a y,N  k ,k   , respectively.
 a yn1 k , k   a xm  k , k    a yn  k , k   a xm  k , k    k
 2 
S   L K is the unquantized signal measurement matrix, E is
1 2
y m1, n 1 ym 2, n 2
 k 1
K K

 a  , fk  k  1  a  , f 
2 2
m k m k k k  1 the array noise sample matrix. According to the result of (7),
k 1 k 1
(8) can be rewritten as
(4)

Authorized licensed use limited to: Southeast University. Downloaded on November 28,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.
 X    Ay  Ax  S T   N (9) Then the estimation Ax , Ay and S are unique up to
permutation and scaling of columns. The permutation and
where  and  are scalers, S  U T S , N = U T E , U   L L
scaling effect in this paper can be formulated as
is a unitary matrix.
 Aˆ x  Ax ΠΔ1 +E1

 Sˆ  SΠΔ2 +E2 (18)
III. THE PROPOSED FRAMEWORK ˆ
A
 y  Ay ΠΔ3 + E 3
To further explore the tensor nature of the array
measurement, the noiseless array signal X can be expressed where Π is a permutation matrix. E1 , E 2 and E 3 are the
as a trilinear model (or a third-order PARAFAC model) fitting errors, and Δ1 , Δ2 and Δ3 are diagonal scaling
K
xm , n ,l   Ax  m, k  Ay  n, k  S  l , k  matrices with Δ1 Δ2 Δ3  I K .
k 1 (10) After the Ax and Ay are estimated by trilinear
 m  1,, M ; n  1,, N ; l  1,, L  decomposition, 2D DOA can be estimated by the
where Ax  m, k  , Ay  n, k  and S  l , k  are the  m, k  -element Vandermonde feature of the response matrices. For example,
of the x-axis direction matrix Ax , the  n, k  -element of the y- we can get the phase of the a x  k ,k  as h  angle  a x  k , k  
axis direction matrix Ay and the  l , k  - elements of S ,  0, 2 d sin  k cos k /  ,, 2 d  M  1 sin  k cos k /   . It is
T

respectively. Actually, matrix X can be regarded as the easy to find

result of slicing the tensor x along the source signal Pc x  h (19)
direction. According to the symmetry of trilinear model, the with
slice format of 3D array x along other two dimensions can
1 0 
be obtained. In detail, these slice matrices can be written as 1 
2 d /   c  c x0 
 Ax D1  Ay   P (20)
     x  vk 
 Ax D2  Ay   T  
X   Ay  Ax  S  N x = 
T
 S + Nx (11) 1 2  M  1  d /  
   Obviously, c x is a scaler that we do not care about,
  0

 Ax DN  Ay   vk = sin  k cos k . The LS solution of c x can be calculated via

 S T D1  Ax   cˆ x0 
   P hˆ (21)
†
 T 
S D2  Ax   T  vˆk 
Y   Ax  S  ATy  N y =  A  Ny (12)
   y
 T  where ĥ accounts for the estimation of h . It is seen that vˆk is
 S DM  Ax  the estimation of sin  k cos k . Similarly, we can get the
 Ay D1  S T   estimation of sin  k sin k as uˆ k . Then 2D DOA estimation can
 
 Ay D2  S T   T be accomplished via
Z   S  Ay  Ax  N z =   Ax  N z (13)
 
T
ˆ  arcsin uˆ 2  vˆ 2
    k k k
   (22)
 Ay DL  S   ˆk  arctan  uˆk / vˆk 
T

where Dp  A  returns a diagonal matrix with the diagonal Noticed that Aˆ x and Aˆ y share the same permutation, thus the
elements are the p-th row of A . estimated 2D DOAs are automatically paired.
As a popular technique to decomposition a trilinear
model, the trilinear alternating least squares (TALS)
algorithm has been widely utilized. Usually, TALS updates IV. SIMULATION RESULTS
one factor matrix via least squares (LS) technique, and treat To verify the effectiveness of the proposed algorithm,
other factor matrices as known parameters. The iteration will numerical simulations are carried out. In the simulations, we
repeat until the algorithm convergence. According to (11)- assume that there exist K  3 uncorrelated sources with 2D
(13), the LS updates of S , Ay and Ax are given by DOAs are 15 ,10  ,  25 , 20  and  35 ,30  , and L snapshots
†
Sˆ T   Ay  Ax  X (14) are collected. We consider the uniform rectangular array with
an M  N omni-directional sensors, and d   / 2 . The signal-
Aˆ T   A  S  Y
†
(15)
y x
to-noise ratio (SNR) in the simulation is defined as the ratio of
†
Aˆ   S  Ay  Z
T
x (16) signal power to noise power before one-bit quantization.
Herein, the root mean square error (RMSE) is utilized for
After TALS, the LS estimations of S , Ay and Ax can be performance assessment, which is defined as
achieved. Let the k-rank of Ax , Ay and S are k A , k A and 1 K 1 Q  ˆ
   ˆ 
  k  , where ˆk , n ,
2 2

k S , then if [19]
x y
RMSE    k ,n  k
K k 1 Q n 1 
k ,n

k Ax +k Ay +k S  2 K  2 (17) ˆk , n represent n-th Monte Carlo trial of  k and  k estimates,

Authorized licensed use limited to: Southeast University. Downloaded on November 28,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.
respectively, Q is the total number of Monte Carlo trials. In In the third example, we test the RMSE performance of
our simulation, Q is set to 500. the proposed framework versus Snapshot number L, where
In the first example, we plot the scatter results of the M=12, N=8 and SNR=0. Comparison result is illustrated in
proposed one-bit PARAFAC framework in Fig.2 and Fig.3, Fig.5. Clearly, RMSE on 2D DOA would improve with L
where M=12, N=8 and L=1000 are considered, and SNRs are increasing. An interesting observation is that O-PARAFAC
set to 5dB and 10dB, respectively. From which we observed algorithm provides more accurate 2D DOA estimation than
2D DOAs can be estimated and correctly paired from one-bit O-ESPRIT, as the tensor structure has been utilized in the
measurements. proposed O-PARAFAC framework. Also, both one-bit
algorithms offer higher RMSE than the U-ESPRIT and U-
PARAFAC, since the one-bit measurement is lossy.
elevation/deg

RMSE(deg)
Fig 2. Scatter results with SNR= 5dB

Fig 5. RMSE vs. L for DOA estimation

Fig. 6 illustrates the RMSE comparison of versus M,
elevation/deg

where N=8, L=1000 and SNR=0. It is shown that estimation

performance would improve with the increasing M. Also, we
can see the performance corresponding to O-PARAFAC is
between that of the O-ESPRIT and U-ESPRIT.

Fig 3. Scatter results with SNR= 10dB

RMSE(deg)

In the second example, we depict the RMSE

performance of the proposed one-bit PARAFAC (marked
with O-PARAFAC) framework versus SNR, where M=12,
N=8 and L=1000. For comparison purpose, the traditional
ESPRIT with one-bit measurement (marked with O-ESPRIT)
, the traditional ESPRIT with un-quantified measurement
(marked with U-ESPRIT) and the traditional PARAFAC with
un-quantified measurement (marked with U-PARAFAC) are
Fig 6. RMSE vs. M for DOA estimation
added. It is seen that at low SNR regions (SNR<0), there is
no visible performance difference between O-PARAFAC V. CONCLUSIONS
and U-ESPRIT. But when SNR increase, both the one-bit
algorithms provide worse performance then the un-quantified In this paper, we have stressed the problem of 2D DOA
algorithms, and situation become worse when SNR is larger estimation for uniform rectangular array try with one-bit
than 0dB. Besides, the proposed O-PARAFAC algorithm measurements. The relationship between the covariance
achieves better RMSE than O-ESPRIT, since the tensor matrices of one-bit measurement and that of the accurately
structure has been exploited. quantified measurement is extended to tensor domain, and a
one-bit PARAFAC framework is proposed. Since one-bit
quantization system is more flexible than the high-precision
quantization, the proposed framework will have a bright
prospect in further applications, especially for massive
RMSE(deg)

antenna systems.
REFERENCES
[1] Shi Z., Zhou C., Gu Y., et al. Source estimation using
coprime array: a sparse reconstruction perspective, IEEE
Sensors Jo., 2017, 17(3): 755-765.
[2] Cao M., Vorobyov S. A., Hassanien A., Transmit array
interpolation for DOA estimation via tensor
decomposition in 2-D MIMO radar, IEEE Trans. Signal
Fig 4. RMSE vs. SNR for DOA estimationa Process., 2017, 65(19): 5225–5239.

Authorized licensed use limited to: Southeast University. Downloaded on November 28,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.
[3] Zhou C., Gu Y., Fan X., et al. Direction-of-arrival
estimation for coprime array via virtual array
interpolation, IEEE Trans. Signal Process., 2018, 66(22):
5956-5971.
[4] Zhou C., Gu Y., Shi Z. et al. Off-grid direction-of-arrival
estimation using coprime array interpolation, IEEE
Signal Process. Lett., 2018, 25(11): 1710-1714.
[5] Liu T., Wen F., Zhang L., et al. Off-grid DOA estimation
for colocated MIMO radar via reduced-complexity sparse
Bayesian learning, IEEE Access, 2019, 7: 99907 – 99916.
[6] Liao B. Fast angle estimation for MIMO radar with
nonorthogonal waveforms, IEEE Trans. Aero. Elec. Sys.,
2018, 54(4): 2091-2096.
[7] Wen F., Shi J., Zhang Z. Direction finding for bistatic
MIMO radar with unknown spatially colored noise,
Circuit. Syst. Signal Process., 2020, 39(5):2412-2424.
[8] Dong Y., Dong C., Xu J., Zhao G., Computationally
efficient 2-D DOA estimation for L-shaped array with
automatic pairing, IEEE Antenna Wirel. Propaga. Lett.,
2016, 15: 1669-1672.
[9] Meng X., Jiang Z. A Generalized sparse Bayesian
learning algorithm for one-bit DOA estimation. IEEE
Commun. Lett., 2018, 22(7):1414-1417.
[10] Wang X., Wang W., Li X., et al. A tensor-based subspace
approach for bistatic MIMO radar in spatial colored
noise. Sensors, 2014, 14(3): 3897-3907.
[11] Guo Y., Wang X., Wang W., et al., Tensor-based angle
estimation approach for strictly noncircular sources with
unknown mutual coupling in bistatic MIMO radar.
Sensors, 2018, 18(9), 2788.
[12] Wang X., Huang M., Wu X., et al. Direction of arrival
estimation for MIMO radar via unitary nuclear norm
minimization. Sensors, 2017, 17(4): 939.
[13] Candès E. J., Wakin M. B. An introduction to
compressive sampling. IEEE Signal Process. Mag., 2008,
25(2): 21-30.
[14] Duarte M. F., Davenport M. A., Takhar D., et al. Single-
pixel imaging via compressive sampling. IEEE Signal
Process. Mag., 2008, 25(2): 83-91.
[15] Yan M, Yang Y, Osher S. Robust 1-bit compressive
sensing using adaptive outlier pursuit. IEEE Trans. Signal
Process., 2012, 60(7): 3868-3875.
[16] Liu C. L., Vaidyanathan P. P. One-bit sparse array DOA
estimation, IEEE International Conference on Acoustics,
Speech and Signal Processing (ICASSP), 2017: 3126-
3130.
[17] Huang X., Liao B. One-bit MUSIC. IEEE Signal Process.
Lett., 2019, 26(7): 961 - 965.
[18] Jacovitti G., Neri A. Estimation of the autocorrelation
function of complex Gaussian stationary processes by
amplitude clipped signals. IEEE Trans. Inform. Theory,
1994, 40(1): 239-245.
[19] Zhang X., Xu Z., Xu L., et al. Trilinear decomposition-
based transmit angle and receive angle estimation for
multiple-input multipleoutput radar, IET Radar Sonar
Navigat., 2011, 5(6): 626-631.

Authorized licensed use limited to: Southeast University. Downloaded on November 28,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.