3232 IEEE ANTENNAS AND WIRELESS PROPAGATION LETTERS, VOL.

16, 2017

Phase-Only Pattern Synthesis for Linear

Antenna Arrays
Junli Liang, Xuhui Fan, Wen Fan, Deyun Zhou, and Jian Li, Fellow, IEEE

Abstract—This letter addresses the problem of phase-only an- is nonconvex because the feasible region for the weight vector
tenna beampattern synthesis. With the given magnitudes of the is an intersection of circles.
weight vector, we introduce a scaling factor to accurately represent To overcome these difficulties, this letter introduces scale
the shape constraints on both the mainlobe and sidelobe regions. variables to accurately represent the shape constraints on both
Moreover, we derive an iterative optimization method to solve the
mainlobe and sidelobe power levels, resulting in a coupled scale
resultant optimization problem efficiently. The performance of the
proposed method is demonstrated via numerical examples.
variable and phase vector variable optimization problem. More-
over, we represent the scale using the phase vector to reduce
Index Terms—Beampattern synthesis, phase-only, uncertain the coupled problem into an easily solved single phase vector
scale. optimization problem.
Throughout the letter, vectors and matrices are denoted by
I. INTRODUCTION boldface lowercase and uppercase letters, respectively. (·)∗ ,
HE task of phase-only antenna pattern synthesis is to opti- (·)T , and (·)H are the complex conjugate, transpose, and con-
T mize only the phases to match an expected radiation pattern
for an antenna array while fixing the excitation magnitudes. The
jugate transpose operators, respectively. 0m ×n represents the
m × n zero matrix. A square diagonal matrix with elements
corresponding topic has received significant attention in the liter- {a1 , . . . , an } is denoted by diag{[a1 , . . . , an ]}. {} and {},
ature [1]–[9] due to its wide applications in the millimeter-wave respectively, denote√ the real and imaginary parts of vectors or
communication, radar, sonar, etc. In [3] and [8], the semidefinite matrices. j = −1. Finally, | · | and ∠{·} are the magnitude
relaxation (SDR) [10] technique is applied to solve the relaxed and phase of a complex-valued scalar, respectively.
version of the phase-only beampattern synthesis problem by
dropping a rank-1 constraint. Since the number of optimization II. PROBLEM FORMULATION
variables is squared of the original problem, the SDR technique
Consider an array composed of M elements. Let a(θ) be the
is not suitable for large arrays. In [9], the nonconvex constant
steering vector for angle θ, and v = [v1 , v2 , . . . , vM ]T ∈ C M ×1
modulus constraint is relaxed to a convex constraint. However,
be the so-called weight vector. Then, the array beampattern
the relaxation cannot always ensure the constant modulus prop-
at θ is p(θ) = |vH a(θ)|2 . Let θn , n = 1, . . . , N1 , be the grid
erty for each weight element. Moreover, the power responses in
point of the angle region of interest for the mainlobe, and
the mainlobe region require careful selection [9]. Other beam-
L(θn ) ≥ 0 and U (θn ) > 0 stand for the desired lower and up-
pattern synthesis methods have been presented in [5]–[7].
per bounds of the corresponding radiated power. Additionally,
Unlike common beampattern synthesis problems of optimiz-
ing both magnitudes and phases [1], [6], [7], [11], [12], in {θn }N
n =N 1 +1 denote the N2 grid points in the sidelobe region
1 +N 2

phase-only problems some degree of freedom of the weight and {U (θn )N N 1 +N 2

n =N 1 +1 } > 0 and {L(θn )n =N 1 +1 } = 0 as the de-
1 +N 2

vector is lost due to fixing the magnitudes. Especially, with sired upper and lower power bounds. Thus, the common beam-
improper magnitudes, it is difficult to ensure that the generated pattern synthesis problem without magnitude constraints can be
beampattern matches the desired beam shape, where there exists represented as
an uncertain scaling problem. The phase-only synthesis problem
Find v
Manuscript received August 13, 2017; revised October 21, 2017; accepted s.t. L(θn ) ≤ |vH a(θn )|2 ≤ U (θn ), n = 1, . . . , N (1)
November 6, 2017. Date of publication November 10, 2017; date of current
version December 11, 2017. This work was supported in part by the Natural
Science Foundation of China under Grant NSFC 61471295, in part by Central where N = N1 + N2 and both the magnitudes and phases of
University Funds under Grant G2016KY0308, Grant G2016KY0002, and Grant v are optimized to satisfy the desired power bounds in the
17GH030144, and in part by the Key Research Program of Frontier Sciences, mainlobe and sidelobe regions.
CAS, under Grant QYZDY-SSW-JSC035. (Corresponding author: Junli Liang.)
J. Liang, X. Fan, W. Fan, and D. Zhou are with the School of Electronics
When the magnitudes {Am }M m =1 of the weighting factors
and Information, Northwestern Polytechnical University, Xi’an 710065, China {vm }M
m =1 are given [3]–[5], [8], only the phases of v are op-
(e-mail: [email protected]; [email protected]; fanwen@ timized, and the corresponding beampattern synthesis problem
mail.nwpu.edu.cn; [email protected]). becomes
J. Li is with the Department of Electronic Engineering and Information Sci-
ence, University of Science and Technology of China, Hefei 230000, China, and
also with the Department of Electrical and Computer Engineering, University Find v
of Florida, Gainesville, FL 32611 USA (e-mail: [email protected]).
Color versions of one or more of the figures in this letter are available online s.t. L(θn ) ≤ |vH a(θn )|2 ≤ U (θn ), n = 1, . . . , N,
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/LAWP.2017.2771380 |vm | = Am , m = 1, . . . , M. (2)

1536-1225 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
LIANG et al.: PHASE-ONLY PATTERN SYNTHESIS FOR LINEAR ANTENNA ARRAYS 3233

However, when the magnitudes {Am }M m =1 are given, some

Motivated by the decomposability and superior convergence
freedom degree of the weight vector is lost, and we may properties of the alternating direction method of multiplier [16],
not be able to satisfy the bounded radiated power constraints we determine {ξ, φ, x, λ} via the following iteration steps.
|vH a(θn )|2 ∈ [L(θn ), U (θn )] for the radiated angles of inter- Step 1: Determine xn (t + 1) with given {ξ(t), φ(t), λ(t)}
est. In particular, this also holds for the constant modulus con- from
straints [9], [13]–[15], i.e., A1 = A2 = · · · = AM . To tackle the x(t + 1) = arg min L (ξ(t), φ(t), x, λ(t))
problem, we introduce a positive scaling factor β and formulate x
the optimizing problem as follows:
s.t. L(θn ) ≤ |xn | ≤ U (θn ),
2
n = 1, . . . , N. (8)
max β Overlooking the irrelevant terms to x and defining
β ,v

s.t. βL(θn ) ≤ |vH a(θn )|2 ≤ βU (θn ), n = 1, . . . , N, T λn (t)

x̃n (t) = ξ(t)e−j φ (t)
Aa(θn ) −
ρ
|vm | = Am , m = 1, . . . , M (3)
(8) can be simplified as N separable subproblems:
from which the obtained radiated powers |vH a(θn )|2 ∈
[βL(θn ), βU (θn )] can maintain the desired normalized beam min |xn − x̃n (t)|2
xn
shape well. Additionally, maximizing β is to avoid formulating
an ambiguity problem and also attain the largest radiated power. s.t. L(θn ) ≤ |xn |2 ≤ U (θn ) (9)
Let ξ = √1β > 0 and the solution to which is given by
v ⎧
⎪ j ∠{x̃ n (t)}
√ = ξAej φ (4) ⎨ U (θn )e , if |x̃n (t)|2 ≥ U (θn )
β xn (t + 1) = j ∠{x̃
L(θn )e n (t)}
, if |x̃n (t)|2 ≤ L(θn )
T ⎪
⎩
where A = diag{[A1 , . . . , AM ]} and φ = [φ1 , . . . , φM ] . x̃n (t), otherwise
Thus, the optimization problem in (3) is equivalent to (10)
min ξ for n = 1, . . . , N .
ξ ,φ
Step 2: Update {ξ(t + 1), φ(t + 1)} with given {x(t + 1),
T
s.t. L(θn ) ≤ |ξe−j φ Aa(θn )|2 ≤ U (θn ), n = 1, . . . , N. λ(t)} from
(5) {ξ(t + 1), φ(t + 1)} = arg min L (ξ, φ, x(t + 1), λ(t))
ξ ,φ
III. PROPOSED ALGORITHM N 
 T
To solve (5), we introduce auxiliary variables = arg min ξ + λrn (t){xn (t + 1) − ξe−j φ Aa(θn )}
ξ ,φ
T n =1
xn = ξe−j φ Aa(θn ) ρ T

+ ({xn (t + 1) − ξe−j φ Aa(θn )})2
for n = 1, . . . , N , and rewrite (5) in the following equivalent 2
form: N 
 T

min ξ + λin (t){xn (t + 1) − ξe−j φ Aa(θn )}

ξ ,φ n =1
ρ 
s.t. L(θn ) ≤ |xn |2 ≤ U (θn ), n = 1, . . . , N, +
T
({xn (t + 1) − ξe−j φ Aa(θn )})2 (11)
T
2
xn = ξe−j φ Aa(θn ), n = 1, . . . , N. (6) where the variables ξ and φ are coupled.
Based on (6), we construct the augmented Lagrangian as To decouple them, we rewrite (11) as
follows: min E(φ)ξ 2 + F (φ)ξ + G (12)
ξ ,φ
N 
 T
L (ξ, φ, x, λ) = ξ + λrn {xn − ξe−j φ Aa(θn )} where
n =1 N
 ρ  −j φT
ρ T E(φ) = e Aa(θn )aH (θn )Aej φ (13)
+ ({xn − ξe−j φ Aa(θn )})2 2 n =1
2
N  N
ρ
 ∗
T λn (t) T
+ λin {xn − ξe−j φ Aa(θn )} F (φ) = 1 − xn (t + 1) + e−j φ Aa(θn )
2 n =1 ρ
n =1
ρ  N
ρ
T
+ ({xn − ξe−j φ Aa(θn )})2 (7) λn (t)
2 − xn (t + 1) + aH (θn )Aej φ
2 n =1 ρ
where ρ > 0 is a user-defined step size, and λrn and λin are the La- (14)
grange multipliers corresponding to the real and imaginary parts
T
of the constraint xn = ξe−j φ Aa(θn ) for n = 1, . . . , N . Ad- ρ
N
λn (t) 2
ditionally, x = [x1 , . . . , xN ] , λ = [λ1 , . . . , λN ]T , and λn =
T G= |xn (t + 1) + | . (15)
2 n =1 ρ
λrn + jλin for n = 1, . . . , N .
3234 IEEE ANTENNAS AND WIRELESS PROPAGATION LETTERS, VOL. 16, 2017

half-wavelength element spacing. In the first two experiments,

Algorithm 1: Phase-only array pattern synthesis.
we focus on the phase-only nulling and wide mainlobe beam-
Initialization: λ(0), ξ(0), φ(0), and T ; pattern design problems, whereas we consider a more difficult
for t = 0, . . . , T problem of constant modulus beampattern synthesis in the third
Obtain x(t + 1) using the steps described in (8)–(10); experiment.
Determine ξ(t + 1) and φ(t + 1) using (11)–(17); In actual applications, the calibrated amplitude settings are
Update {λrn (t + 1), λin (t + 1)} using (18) and (19); employed as the weight magnitudes. Similar to [8] and [9],
end for t = T . in the simulations we generate sinusoid-type magnitudes and
Output vopt = Aej φ . unimodular magnitudes for the first two experiments and last
experiment, respectively. In all experiments, we set λ(0) = 0,
When φ is given, the optimal ξ is given by ρ = 1, and T = 300. Additionally, a random amplitude in the
interval [0, 1] and 32 random phases in the interval [0, 2π] are
F (φ) employed for initializations of ξ(0) and φ(0), respectively.
ξ(φ) = − . (16)
2E(φ)
Inserting (16) into (12) and overlooking the constant term, A. Exp. 1: Phase-Only Nulling
we have In the first experiment, we explore the ability of the proposed
F 2 (φ) method to synthesize a notching beampattern for suppressing
min − . (17) interferences, where the −50 dB notch lies between −40◦ and
φ 4E(φ)
−35◦ , and the 0 dB mainlobe points at 0◦ . In addition, the re-
Here, we consider applying the Broyden–Fletcher–Goldfarb– gions [−90◦ , −5◦ ] and [5◦ , 90◦ ] are set as the sidelobe regions
Shanno (BFGS) algorithm [17], which can be realized via the with a maximal of −19 dB power constraint. The given magni-
MATLAB code “fminunc,” to solve the nonlinear optimization tudes are displayed with green on the top of Fig. 1(a). We also
problem in (17). implement the methods in [3] and [8] for comparison purposes,
Inserting the obtained φ(t + 1) from BFGS into (16), we and show the errors between the desired and actually computed
F (φ(t+1))
update ξ as ξ(t + 1) = − 2E (φ(t+1)) .
magnitudes at the bottom of Fig. 1(a). Additionally, the nor-
Step 3: Update Lagrange multipliers {λrn , λin } as malized beampatterns synthesized by these methods are plotted
in Fig. 2(a). From these figures, it can be seen that both the
λrn (t + 1) = λrn (t) proposed method and that in [8] perform well, whereas [3] has
T poor transition regions and cannot attain the given magnitudes.
+ ρ{(xn (t + 1) − ξ(t + 1)e−j φ (t+1)
Aa(θn )} (18)
λin (t + 1) = λin (t) B. Exp. 2: Wide Mainlobe With Upper and Lower Bounds
T
+ ρ{xn (t + 1) − ξ(t + 1)e−j φ (t+1)
Aa(θn )} (19) In the second experiment, we investigate the capacity of the
proposed method for generating a wide mainlobe for detecting
for n = 1, . . . , N . targets with inaccurate directions. In this experiment, the main-
Steps 1–3 are repeated until a predefined maximum iteration lobe region is set as [−10◦ , 10◦ ], and the upper and lower con-
number T is reached. The proposed method is summarized in straints, i.e., 0.35 and −0.35 dB, are set for the power constraints
Algorithm 1. in the mainlobe region, increasing the nonconvexity of the op-
timization problem. Both [−90◦ , −14◦ ] and [14◦ , 90◦ ] are the
A. On Computational Complexity sidelobe regions with a maximal of −14 dB power constraint.
The given magnitudes and synthesized normalized beampatterns
Regarding the computational complexity, we consider the are plotted in Figs. 1(b) and 2(b), respectively, which show that
major part, namely, multiplications involved in one iteration. the methods in [3] and [8] cannot provide a satisfactory main-
Both steps 1 and 3 require O{M N }, whereas the BFGS method lobe region and cannot attain the given magnitudes, whereas the
in step 2 requires O{M 2 }. Therefore, the proposed method proposed method can work well.
requires O{2M N + M 2 } in one iteration. According to the
CVX toolbox [18], the CVX computation in [3], [8], and [9]
requires O{max{N1 + N2 , M }4 M 0.5 log( 1 )}, where  is the C. Exp. 3: Constant Modulus Pattern Synthesis
given default value in [18], i.e., 2.220446049250313−8 . In the third experiment, we consider a more challenging prob-
lem, i.e., to synthesize a wide-mainlobe beampattern but with
B. On Convergence constant modulus constraints. In this experiment, the mainlobe
The proposed method is based on the alternating direction region is set as [−15◦ , 15◦ ] with the upper and lower bound
method of multipliers and BFGS method [16], [17], the former constraints as 0.3 and −0.3 dB, respectively, whereas both
of which has superior convergence property (see [16] for de- [−90◦ , −20◦ ] and [20◦ , 90◦ ] are the sidelobe regions with a max-
tails), whereas the latter has the Q-linear convergence property imal of −14 dB power constraint. The method in [9] is also im-
[17]. Therefore, the proposed method converges. plemented for comparison purposes. The attained magnitudes
and the normalized beampatterns are plotted in Figs. 1(c) and
2(c), respectively. The proposed method can work well for the
IV. NUMERICAL EXAMPLES
constant modulus constrained beampattern synthesis problem.
In this section, numerical examples are presented to evaluate However, as discussed in [9], the relaxation cannot always en-
the performance of the proposed method. In all of the three ex- sure the satisfaction of the constant modulus constraint for each
periments, we consider a 32-element uniform linear array with weighting factor.
LIANG et al.: PHASE-ONLY PATTERN SYNTHESIS FOR LINEAR ANTENNA ARRAYS 3235

Fig. 1. (a) Magnitudes and error in Exp. 1; (b) magnitudes in Exp. 2; and (c) magnitudes in Exp. 3.

Fig. 2. (a) Synthesized patterns in Exp. 1; (b) synthesized patterns in Exp. 2; and (c) synthesized patterns in Exp. 3.

For our PC (Intel i7-6700 CPU, 64 bit, RAM 16 GB), the [5] R. Shore, “Nulling a symmetric pattern location with phase-only weight
run times of the first experiment by [3], [8], and the proposed control,” IEEE Trans. Antennas Propag., vol. AP-32, no. 5, pp. 530–533,
May 1984.
method are 0.54, 72.22, and 34.32 s, respectively. For the second [6] O. M. Bucci, G. Mazzarella, and G. Panariello, “Reconfigurable arrays
experiment, the run times of these methods are 0.75, 75.45, and by phase-only control,” IEEE Trans. Antennas Propag., vol. 39, no. 7,
35.79 s, respectively; whereas for the third experiment, the run pp. 919–925, Jul. 1991.
times of [9] and the proposed method are 36.51 and 35.62 s, [7] R. Vescovo, “Reconfigurability and beam scanning with phase-only con-
respectively. Obviously, [3] is the most efficient one, but at the trol for antenna arrays,” IEEE Trans. Antennas Propag., vol. 56, no. 6,
pp. 1555–1565, Jun. 2008.
cost of the unsatisfactory pattern shapes. The proposed method [8] B. Fuchs, “Application of convex relaxation to array synthesis problems,”
is implemented more efficiently than both [8] and [9] and has IEEE Trans. Antennas Propag., vol. 62, no. 2, pp. 634–640, Feb. 2014.
the most satisfactory pattern shapes. [9] P. Cao, J. S. Thompson, and H. Haas, “Constant modulus shaped beam
synthesis via convex relaxation,” IEEE Antennas Wireless Propag. Lett.,
vol. 16, pp. 617–620, 2017.
V. CONCLUSION [10] Z.-Q Luo, W.-K. Ma, A. M.-C. So, Y. Ye, and S. Zhang, “Semidefinite re-
laxation of quadratic optimization problems,” IEEE Signal Process. Mag.,
This letter introduced a new algorithm to solve phase-only an- vol. 27, no. 3, pp. 20–34, May 2010.
tenna beampattern synthesis problems. To improve the method [11] W.-Q. Wang, “Range-angle dependent transmit beampattern synthesis for
linear frequency diverse arrays,” IEEE Trans. Antennas Propag., vol. 61,
efficiently, we have decoupled effectively the scaling variable no. 8, pp. 4073–4081, Aug. 2013.
and the phase vector in the iterative optimization approach. In [12] J. Xiong, W.-Q. Wang, H. Z. Shao, and H. Chen, “Frequency diversity
terms of the pattern shape, the proposed approach can outper- array transmit beampattern optimization with genetic algorithm,” IEEE
form existing methods, as illustrated via numerical examples. Antennas Wireless Propag. Lett., vol. 16, pp. 469–472, 2017.
[13] J. Liang, H. C. So, J. Li, and A. Farina, “Unimodular sequence design
based on alternating direction method of multipliers,” IEEE Trans. Signal
REFERENCES Process., vol. 64, no. 20, pp. 5367–5381, Oct. 2016.
[14] P. Stoica, J. Li, and Y. Xie, “On probing signal design for MIMO radar,”
[1] H. Lebret and S. Boyd, “Antenna pattern synthesis via convex optimiza- IEEE Trans. Signal Process., vol. 55, no. 8, pp. 4151–4161, Aug. 2007.
tion,” IEEE Trans. Signal Process., vol. 45, no. 3, pp. 526–531, Mar. [15] H. He, J. Li, and P. Stoica, Waveform Design for Active Sensing Systems:
1997. A Computational Approach. Cambridge, U.K.: Cambridge Univ. Press,
[2] O. M Bucci, G. D. Elia, and G. Mazzarella, and G. Panariello, “Antenna 2012.
pattern synthesis: A new general approach,” Proc. IEEE, vol. 82, no. 3, [16] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed
pp. 358–371, Mar. 1994. optimization and statistical learning via the alternating direction method
[3] P. J. Kajenski, “Phase only antenna pattern notching via a semidefinite of multipliers,” Found. Trends Mach. Learn., vol. 3, no. 1, pp. 1–122,
programming relaxation,” IEEE Trans. Antennas Propag., vol. 60, no. 5, 2010.
pp. 2562–2565, May 2012. [17] R. Fletcher, “A new approach to variable metric algorithms,” Comput. J.,
[4] P. J. Kajenski, “Phase-only monopulse pattern notching via semidefinite vol. 13, no. 3, pp. 317–322, 1972.
programming,” in Proc. 2012 IEEE Int. Symp. Antennas Propag., 2012, [18] CVX Toolbox: Version 2.1, Mar. 2017. [Online]. Available:
pp. 1–2. http://cvxr.com/cvx/