1
A Convex Hull Based Approach for MIMO Radar
Waveform Design with Quantized Phases
Chenglin Ren, Zhaohui Ma, Fan Liu, Student Member, IEEE, Weichao Pi, and Jianming Zhou
Abstract—In this letter, we focus on designing constant-
modulus waveform with discrete phases for the multi-input
arXiv:1806.05621v1 [eess.SP] 14 Jun 2018
multi-output (MIMO) radar, where the signal-to-interference-
plus-noise ratio (SINR) is maximized in the presence of both the
signal-dependent clutter and the noise. Given the NP-hardness
of the formulated problem, we propose to relax the original
optimization as a sequence of continuous quadratic programming
(QP) subproblems by use of the convex hull of the discrete feasible
region, which yields approximated solutions with much lower
computational costs. Finally, we assess the effectiveness of the
proposed waveform design approach by numerical simulations.
Index Terms—waveform design, discrete phases, MIMO radar,
quadratic programming, convex hull
I. I NTRODUCTION
A CCORDING to the configuration of the antennas, the
MIMO radar is studied in two types: distributed MIMO
radar [1], [2] and colocated MIMO radar [3], [4]. Compared
with traditional phased array radar, MIMO radar offers extra
degrees of freedom by allowing individual waveforms to be
transmitted at each antenna, and therefore achieves enhanced
capabilities [5]–[7]. In addition, a well designed waveform
can significantly improve the SINR [8], [9] and probability of
detection [10].
The problem of waveform design for MIMO radar has
gained considerable attention in recent years [11]–[17]. In
practical scenarios, waveforms with constant modulus or low
peak-to-average-power ratio (PAPR) properties are needed to
avoid signal distortions, as the non-linear power amplifiers
are typically operated at the saturation region. Meanwhile, the
similarity constraint (SC) is also enforced, which employs a
reference waveform as the benchmark and allows the opti-
mized waveform to share some of the ambiguity properties of
the reference waveform. Relying on the semidefinite relaxation
(SDR) and rank-one decomposition techniques, the authors
of [14] introduce the sequential optimization procedures to
maximize the SINR accounting for the constant modulus
constraint (CMC) and the SC for the case of continuous
signal phase. In [16], a novel successive QCQP refinement
(SQR) algorithm is developed for the same optimization
problem in [14], which involves solving a sequence of con-
vex quadratically constrained quadratic programming (QCQP)
subproblems. Furthermore, by customizing the projection pro-
cedure for the convex QCQP subproblems, the Accelerated
Gradient Projection (AGP) is proposed in [17], which shares a
C. Ren, Z. Ma, F. Liu, W. Pi and J. Zhou are with the School of Information
and Electronics, Beijing Institute of Technology, Beijing, 100081, China
(e-mail: rencl3@bit.edu.cn; mzh 19890822@bit.edu.cn; liufan92@bit.edu.cn;
peao kelvin@163.com; zhoujm@bit.edu.cn)
comparable performance with the SQR method while notably
reducing the computational complexity. However, the works
aforementioned only address the continuous phase case. Given
the extensive use of digital phase shifters in many radar
systems, we consider in this letter the MIMO radar waveform
design for the case of quantized phases. Pioneered by [18], the
cases of both continuous and discrete phases are considered.
Based on the method of semidefinite program (SDP) relaxation
and randomization, the approximated solutions to the quadratic
optimization problem subject to the CMC and the SC are
presented. In [19], the optimization problem under the PAPR
constraint and energy constraint is resolved through the same
techniques.
In this letter, we optimize the SINR of the radar under both
CMC and SC, where the phases of the designed signals are
drawn from a discrete alphabet. By introducing the convex hull
of the finite feasible points, a novel continuous approximation
method (CAM) is proposed to relax the original problem
as a sequence of convex QP subproblems, which can be
solved via standard numerical tools. By doing so, the solution
of the original discrete problem can be then obtained by a
simple quantization procedure. Numerical results show that
the performance of the proposed approach can approximate
that of the continuous waveform design.
II. S YSTEM M ODEL
We consider a colocated narrow band MIMO radar system
equipped with NT transmit antennas and NR receive antennas,
where each antenna emits or receives N samples. The receive
waveform is given as [14]
M
X
r = α0 A(θ0 )s + αm A(θm )s + n (1)
m=1
where r = [rT1 , . . . , rTN ]T ∈ CNR N ×1 with rn ∈ CNR ×1 ,
n = 1, . . . , N being the n-th snapshot across the NR receive
antennas, s = [sT1 , . . . , sTN ]T ∈ CNT N ×1 stands for the
transmit waveform with sn ∈ CNT ×1 being the n-th snapshot
across the NT transmit antennas, n ∈ CNR N ×1 ∼ CN (0, σn2 I)
denotes circular white Gaussian noise, α0 and αm represent
the complex amplitudes of the target and the m-th interference
source, θ0 and θm represent the angles of the target and the
m-th interference source, respectively, and A(θ) stands for the
steering matrix of a Uniform Linear Array (ULA) with half-
wavelength separation between the antennas, which is given
as
A(θ) = IN ⊗ [ar (θ)at (θ)T ] (2)
 2

where IN is the N × N identity matrix, ⊗ denotes the

Kronecker product, at and ar stand for the transmit and the
receive steering vectors, respectively.
Aiming for maximizing the output SINR of the radar, we
jointly design the transmit waveform and the receive filter.
Without loss of generality, we assume that a linear Finite
Impulse Response (FIR) filter f ∈ CNR N ×1 is employed to
process the receive echo wave. The output of the filter r is
given as
M
X
r = f H r = α0 f H A(θ0 )s + αm f H A(θm )s + f H n (3)
m=1
H
where (·) denotes the Hermitian transpose. As a conse-
quence, the output SINR can be expressed as
 2
σ f H A(θ0 )s Fig. 1. The constellation points with Ω = 16 for k-th dimension.
SINR = (4)
f H T̂(s)f + f H f
2
where σ = E[|α0 | ]/σn2 with E[·] denoting the statistical Y(s) = AH (θ0 )[T̂(s) + I]−1 A(θ0 ). (10)
expectation, and According to [20], we can obtain a suboptimal SINR by
M
X assuming Y = Y(s) with a fixed s and optimizing s with the
T̂(s) = Im A(θm )ssH AH (θm ) (5) new Y iteratively, resulting in a sequence of subproblems as
m=1 follows
2
max sH Ys
where Im = E[|αm | ]/σn2 . s
p
s.t. |s(k)| = 1/ NT N (11)
III. P ROBLEM F ORMULATION arg s(k) ∈ [ωk , ωk + δ].
Based on the system model analyzed above, the waveform We consider the subproblem as (11) for each iteration in the
design with continuous phase is firstly considered, where the sequel. In addition, let q ∈ CNT N ×1 be the quantized transmit
CMC and the SC aforementioned have been imposed. Thus, waveform, and q(k) be the k-th entry of q. As a consequence,
the maximization of the SINR in (4) can be formulated as the maximization of the SINR in (11) for the continuous phase
 2 case can be quantized as
σ f H A(θ0 )s
max
f ,s f H T̂(s)f + f H f
p (6)
s.t. |s(k)| = 1/ NT N
ks − s0 k∞ ≤ ε
where s(k) stands for the k-th entry of s, k = 1, . . . , NT N ,
k·k∞ denotes the infinity norm, s0 represents the reference
waveform, and ε (0 ≤ ε ≤ 2) is a preset parameter to control
the similarity degree between s and s0 . In addition, by taking
into account the CMC, the SC can be rewritten as
arg s(k) ∈ [ωk , ωk + δ]
where ωk and δ are respectively given as
ωk = arg s0 (k) − arccos(1 − ε2 /2)
δ = 2 arccos(1 − ε2 /2)
where s0 (k) is the k-th entry of s0 . Noting that there is no
constraints on f as analyzed in [14], the maximization problem
of (6) can be equivalent to
max sH Y(s)s
s
p
s.t. |s(k)| = 1/ NT N
arg s(k) ∈ [ωk , ωk + δ]
where Y(s) is a positive-semidefinite matrix, which is given
as
max qH Yq
q
(12)
s.t. q∈Q
where Q is the discrete phase alphabet which is discussed
detailedly in the next section.
The optimization problem of (12) is non-convex and NP-
hard in general, whose optimal solution cannot be found in
polynomial time. To tackle this problem, we consider the
convex hull of the feasible points for each dimension, and
relax (12) to a continuous QP subproblem. In the next section,
(7) we introduce a novel algorithm—Continuous Approximation
Method (CAM) to approximate to the nearest feasible vector
qopt for the optimization problem of (12).
(8) IV. P ROPOSED A LGORITHM
To describe Q in (12), we suppose an extreme situation of
(9) with δ = 2π, i.e., the similarity parameter ε = 2, where
the SC vanishes and only the CMC is in effect. As shown
in Fig. 1, the feasible region of s(k) is a full circle with a
constant radius on the complex plane C. In the discrete phase
(9) case, we construct a constellation with Ω points on the circle,
i.e., q1 , . . . , qρ , . . . , qΩ , as an example with Ω = 16 shown in
Fig. 1, where the radian between any two adjacent points is
given as
2π
τ= . (13)
Ω
 3

The constellation points are distributed over the whole contin-

uous feasible region for each dimension. Thus, each sample
of the continuous transmit waveform can be quantized to the
finite feasible points which can be given as
p
qρ = exp(jρτ )/ NT N (14)
where ρ = 1, . . . , Ω. For notational convenience, let
n p  o

ΓΩ = qρ = exp(jρτ )/ NT N  ρ = 1, . . . , Ω (15)
represent the discrete phase alphabet for each dimension. As
shown in Fig. 2 (a), by connecting all the feasible points,
a regular polygon is formulated in the tint area, which is
the convex hull of ΓΩ . Further, let ΛΩ be the discrete phase
alphabet for all dimensions, which is given as (a)

Λ Ω = ΓΩ ⊙ · · · ⊙ ΓΩ ⊙ · · · ⊙ ΓΩ (16) 90 0.2
| {z } 120 60
the number of ΓΩ is NT N 0.15

where ⊙ denotes the Cartesian product. Obviously, q ∈ ΛΩ , 150 0.1 30

q(k) ∈ ΓΩ . 0.05

However, for the general situation of (9) with δ less than

180 0
2π, the CMC and the SC are both involved. We introduce
a positive even integer parameter η to indicate the similarity
between q and q0 , where q0 represents the quantized reference 210 330

waveform. By using η, we rewrite (8) as follows

γk = arg q0 (k) − ητ /2
(17)
ϕ = ητ
where q0 (k) is the k-th entry of q0 . For each dimension, η
stands for the number of feasible points around q0 (k), which
is obviously less than Ω. According to the second equation in
(8), the actual similarity tolerance ε between q and q0 can be
expressed as p
ε = 2[1 − cos(ϕ/2)]. (18)
For instance, Fig. 2 (a) shows the feasible points similar to
q0 (k) in the case of η = 6, which can be enumerated as a
finite set
Γη (k) = {p1 (k), . . . pη/2+1 (k), . . . , pη+1 (k)} (19)
where the angle of p1 (k) is γk , pη/2+1 (k) is q0 (k), and
pη+1 (k) is the mirror constellation point of p1 (k) with respect
to q0 (k). The convex hull of Γη (k) is shown as a darker
polygon marked by Θ(k). Noting that Γη (k) ⊆ ΓΩ , any given
pµ (k), µ = 1, . . . , η + 1, can be expressed as
p circle is sopt (k), which obviously locates in the convex hull.
pµ (k) = exp{j[γk + (µ − 1)τ ]}/ NT N . (20)
Furthermore, let
Λη = Γη (1) ⊙ · · · ⊙ Γη (k) ⊙ · · · ⊙ Γη (NT N ) (21)
represent the feasible vector of q for all dimensions. It is
obvious that Λη ⊆ ΛΩ , and especially, when η = Ω,
Λη = ΛΩ .
We focus on the feasible points of (12) for each dimension,
where we relax the finite feasible points set Γη (k) to the
convex hull itself, resulting in the following QP problem
max sH Ys
s (22)
s.t. s∈∆
240 300
270
(b)
Fig. 2. (a) The convex hull of ΓΩ with Ω = 16 and the convex hull of
Γη (k) with η = 6 for k-th dimension; (b) A simulation example of the k-th
entry of the optimal solution sopt for the CAM approach.
where ∆ is the Cartesian product of the convex hull for all
dimensions, which can be expressed as
∆ = Θ(1) ⊙ · · · ⊙ Θ(k) ⊙ · · · ⊙ Θ(NT N ). (23)
Obviously, the optimization problem of (22) is convex, which
can be solved via numerical solvers, e.g., the CVX toolbox.
As shown in Fig. 2 (a), sopt (k) is the k-th entry of the optimal
solution sopt of (22). In Fig. 2 (b), we take a simulation
example of sopt (k) for the CAM approach with Ω = 16 and
η = 6. The red stars indicate the feasible points. The blue
We provide the analytical formulas of the convex hull Θ(k)
for each dimension in the sequel. For notational convenience,
we omit the dimension number k. As shown in Fig. 2 (a), the
convex hull has η + 1 edges which are highlighted in red, i.e.,
L1 : p1 p2 , . . . , Lµ : pµ pµ+1 , . . . , Lη : pη pη+1 , and Lη+1 :
pη+1 p1 . Let us define the middle point of pµ pµ+1 as
pµ + pµ+1
mµ = , µ = 1, . . . , η. (24)
2
According to the basic plane analytic geometry, given any s ∈
C, the line Lµ can be expressed as
Line Lµ : fµ (s) = Re(m∗µ (s − mµ )) = 0. (25)
 4

20 0 20

19 -10 19

18 18
-20
17 17
-30
16 16
SINR [dB]

SINR [dB]

SINR [dB]
-40
15 15
-50
14 14
-60
13 13
CAM, =16, =6
-70 AGP, =16, =6
12 12
CAM, =32, =12
-80 AGP, =32, =12
11 11
CAM CAM CAM, =48, =18
AGP AGP AGP, =48, =18
10 -90 10
2 4 6 8 10 12 14 16 18 20 -80 -60 -40 -20 0 20 40 60 80 4 6 8 10 12 14 16 18 20 22 24
Iteration index Angle [Degrees] Number of samples

(a) (b) (c)

Fig. 3. Performance comparison between the CAM in discrete phase case and the AGP in continuous phase case: (a) SINR in each iteration; (b) Beampattern;
(c) SINR with increasing number of samples.

In addition, for the (η + 1)-th edge, the middle point mη+1 is where k = 1, ..., NT , n = 1, ..., N . For the discrete phase case,
pη+1 + p1 the reference waveform q0 is obtained by quantizing s0 . The
, mη+1 = (26) numbers of the transmit and the receive antennas are NT = 4,
2
NR = 8, respectively. In addition, we consider a scenario
and the corresponding line Lη+1 is given as with three fixed signal-dependent clutters and additive white
Gaussian disturbance with variance σn = 0dB. The target is
Line Lη+1 : fη+1 (s) = Re(m∗η+1 (s − mη+1 )) = 0. (27) 2
located at an angle θ0 = 15◦ with a reflecting power of |α0 | =
As a consequence, the k-th convex hull for each dimension 10dB and three fixed interference sources located at θ1 =
can be separately formulated in two cases. For the first one, −50◦ , θ2 = −10◦ and θ3 = 40◦ reflecting a power of |α1 |2 =
2 2
when 0 ≤ ϕ < π, namely 0 ≤ η < Ω/2 , Θ(k) is given by |α2 | = |α3 | = 30dB.
 As shown in Fig. 3 (a), we firstly compare the SINR in
fµ (s) ≤ 0, µ = 1, . . . , η each iteration between the CAM and the AGP with N = 8,
(28)
fη+1 (s) ≥ 0. Ω = 16 and η = 6. According to (13), (17) and (18), we have
For the case of π ≤ ϕ ≤ 2π, namely Ω/2 ≤ η ≤ Ω, Θ(k) is ϕ = 3π/4 and ε ≈ 1.1. The SINR resulting from the CAM
given by is slightly inferior to the AGP for the continuous phase case.

fµ (s) ≤ 0, µ = 1, . . . , η Fig. 3 (b) shows the beampattern of the optimal waveform
(29) for both methods, using the same parameters as in Fig. 3 (a).
fη+1 (s) ≤ 0.
The simulation reveals that the CAM achieves a comparable
Finally, by comparing the Euclidean distance between beampattern performance with that of the continuous phase
sopt (k) and the feasible points in Γη (k), the feasible point with case.
the minimal distance is checked for each dimension. Thus, we In Fig. 3 (c), we compare the SINR for both methods with
obtain the optimal feasible vector qopt . increasing number of samples N . In order to normalizing the
Remark: The complexity of the CAM mainly comes from actual similarity tolerance ε, we change Ω and η with a fixed
computation of solving the QP problem. By using the interior- ratio 8/3, i.e., Ω = 16 and η = 6, Ω = 32 and η = 12,
point
√ algorithm, the total number of iterations needed is Ω = 48 and η = 18, leading to the same ε as Fig. 3 (a).
O( NT N log(1/ε)), and each iteration can be executed in The AGP results in a consistent SINR with different Ω and
O(NT3 N 3 ) arithmetic operations [21]. Meanwhile, considering η. Meanwhile, the SINR of the CAM exhibits an acceptable
the finite discrete alphabet, the optimization problem of (12) fluctuation, which is only 1dB lower than its AGP counterpart
can be also solved by the exhaustive search method, while the substantially.
worst case complexity of such method is exponential.

VI. C ONCLUSION
V. N UMERICAL R ESULTS
In this letter, we propose a novel approach for MIMO
In this section, we provide numerical results to evaluate
radar waveform design with constant modulus and discrete
the performance in terms of the SINR and the beampattern
phases. By constructing the polygon convex hull of the feasible
between the CAM in discrete phase case and the AGP [17]
constellation points, we develop a CAM algorithm to relax
in continuous phase case, and use chirp waveform as our
the discrete optimization problem as sequential convex QP
benchmark. For the continuous phase case, the reference
subproblems. The optimal solution is then obtained by quan-
waveform s0 ∈ CNT N ×1 can be obtained by stacking the
tizing the solution of QP to its nearest neighbor. Compared
columns of S0 , which we choose as the following orthogonal
with the exhaustive search, the computational complexity has
chirp waveform matrix without loss of generality
been significantly reduced. Numerical results reveal that the
2 proposed approach only yields slight performance-loss with
exp[j2πk(n − 1)/N ] exp[jπ(n − 1) /N ]
S0 (k, n) = √ (30) respect to that of the continuous phase case.
NT N

R EFERENCES
[1] E. Fishler, A. Haimovich, R. S. Blum, L. J. Cimini, D. Chizhik, and
R. A. Valenzuela, “Spatial diversity in radars-models and detection
performance,” IEEE Trans. Signal Process., vol. 54, no. 3, pp. 823–838,
March 2006.
[2] A. M. Haimovich, R. S. Blum, and L. J. Cimini, “MIMO radar with
widely separated antennas,” IEEE Signal Process. Mag., vol. 25, no. 1,
pp. 116–129, 2008.
[3] K. W. Forsythe, D. W. Bliss, and G. S. Fawcett, “Multiple-input
multiple-output (MIMO) radar: performance issues,” in Proc. 38th
Asilomar Conf. Signals, Syst. Comput., Nov 2004, pp. 310–315 Vol.1.
[4] J. Li and P. Stoica, “MIMO radar with colocated antennas,” IEEE Signal
Process. Mag., vol. 24, no. 5, pp. 106–114, Sept 2007.
[5] N. H. Lehmann, E. Fishler, A. M. Haimovich, R. S. Blum, D. Chizhik,
L. J. Cimini, and R. A. Valenzuela, “Evaluation of transmit diversity in
MIMO-radar direction finding,” IEEE Trans. Signal Process., vol. 55,
no. 5, pp. 2215–2225, May 2007.
[6] C. Y. Chen and P. P. Vaidyanathan, “MIMO radar space-time adaptive
processing using prolate spheroidal wave functions,” IEEE Trans. Signal
Process., vol. 56, no. 2, pp. 623–635, Feb 2008.
[7] L. Xu, J. Li, and P. Stoica, “Target detection and parameter estimation
for MIMO radar systems,” IEEE Trans. Aerosp. Electron. Syst., vol. 44,
no. 3, pp. 927–939, July 2008.
[8] L. K. Patton, S. W. Frost, and B. D. Rigling, “Efficient design of radar
waveforms for optimised detection in coloured noise,” IET Radar, Sonar,
Navig., vol. 6, no. 1, pp. 21–29, January 2012.
[9] P. Setlur and M. Rangaswamy, “Waveform design for radar STAP in
signal dependent interference,” IEEE Trans. Signal Process., vol. 64,
no. 1, pp. 19–34, Jan 2016.
[10] M. R. Bell, “Information theory and radar waveform design,” IEEE
Trans. Inf. Theory, vol. 39, no. 5, pp. 1578–1597, Sep 1993.
[11] A. D. Maio, S. D. Nicola, Y. Huang, D. P. Palomar, S. Zhang, and
A. Farina, “Code design for radar STAP via optimization theory,” IEEE
Trans. Signal Process., vol. 58, no. 2, pp. 679–694, Feb 2010.
5
[12] Y. C. Wang, X. Wang, H. Liu, and Z. Q. Luo, “On the design of constant
modulus probing signals for MIMO radar,” IEEE Trans. Signal Process.,
vol. 60, no. 8, pp. 4432–4438, Aug 2012.
[13] A. Aubry, A. D. Maio, B. Jiang, and S. Zhang, “Ambiguity function
shaping for cognitive radar via complex quartic optimization,” IEEE
Trans. Signal Process., vol. 61, no. 22, pp. 5603–5619, Nov 2013.
[14] G. Cui, H. Li, and M. Rangaswamy, “MIMO radar waveform design
with constant modulus and similarity constraints,” IEEE Trans. Signal
Process., vol. 62, no. 2, pp. 343–353, Jan 2014.
[15] A. Aubry, A. D. Maio, Y. Huang, M. Piezzo, and A. Farina, “A new
radar waveform design algorithm with improved feasibility for spectral
coexistence,” IEEE Trans. Aerosp. Electron. Syst., vol. 51, no. 2, pp.
1029–1038, April 2015.
[16] O. Aldayel, V. Monga, and M. Rangaswamy, “Successive QCQP re-
finement for MIMO radar waveform design under practical constraints,”
IEEE Trans. Signal Process., vol. 64, no. 14, pp. 3760–3774, July 2016.
[17] C. Ren, F. Liu, L. Zhou, J. Zhou, W. Luo, and S. Yang, “MIMO
radar waveform design with practical constraints: A low-complexity
approach,” arXiv:1805.03787, May 2018.
[18] A. De Maio, S. D. Nicola, Y. Huang, Z. Q. Luo, and S. Zhang, “Design
of phase codes for radar performance optimization with a similarity
constraint,” IEEE Trans. Signal Process., vol. 57, no. 2, pp. 610–621,
Feb 2009.
[19] A. De Maio, Y. Huang, M. Piezzo, S. Zhang, and A. Farina, “Design of
optimized radar codes with a peak to average power ratio constraint,”
IEEE Trans. Signal Process., vol. 59, no. 6, pp. 2683–2697, June 2011.
[20] B. Friedlander, “Waveform design for MIMO radars,” IEEE Trans.
Aerosp. Electron. Syst., vol. 43, no. 3, pp. 1227–1238, July 2007.
[21] R. D. C. Monteiro and I. Adler, “Interior path following primal-dual
algorithms. part ii: Convex quadratic programming,” Mathematical
Programming, vol. 44, no. 1, pp. 43–66, May 1989. [Online].
Available: https://doi.org/10.1007/BF01587076