756 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL.

70, 2022

Efficient Algorithms for Constant-Modulus Analog

Beamforming
Aakash Arora , Student Member, IEEE, Christos G. Tsinos , Senior Member, IEEE,
M. R. Bhavani Shankar , Senior Member, IEEE, Symeon Chatzinotas , Senior Member, IEEE,
and Björn Ottersten , Fellow, IEEE

Abstract—The use of a large-scale antenna array (LSAA) has I. INTRODUCTION

become an important characteristic of multi-antenna communica-
N MULTI-ANTENNA communication systems, an antenna
tion systems to achieve beamforming gains such as in designing
millimeter-wave (mmWave) systems to combat severe propagation
losses. In such applications, each antenna element has to be driven
I array is employed at the transmitter and/or at the receiver
to achieve beamforming, performance gains. For example, the
by a radio frequency (RF) chain for the implementation of fully- use of a large-scale antenna array (LSAA) has become a de-
digital beamformers, significantly increasing the hardware cost,
complexity, and power consumption. Therefore, constant-modulus cisive part of a millimeter wave (mmWave) system. This is
analog beamforming (CMAB) becomes a viable solution. In this pa- because the communication at mmWave frequencies suffers
per, we consider the scaled analog beamforming (SAB) or constant- from several propagation losses [2]–[4]. Therefore, to alleviate
modulus analog beamforming (CMAB) architecture and design these losses an LSAA is employed at the transmitter to achieve
the system parameters by solving two variants of beampattern beamforming gains. Some other applications of beamforming
matching problem. In the first case, both the magnitude and phase
of the beampattern are matched to the given desired beampattern design are in acoustic imaging for underwater exploration [5],
whereas in the second case, only the magnitude of the beampattern channel sounding in mmWave communications [6], etc. The
is matched. Both the beampattern matching problems are cast as a fully-digital implementation of a beamformer requires as many
variant of the constant-modulus least-squares (CLS) problem. We radio frequency (RF) chains as the number of antenna elements.
provide efficient algorithms based on the alternating majorization- Consequently, this places demands on hardware and increases
minimization (AMM) framework that combines the alternating
minimization and the MM frameworks and the conventional-cyclic implementation cost and power consumption. One possibility is
coordinate descent (C-CCD) algorithms to solve the problem in to employ analog phase-shifters but an LSAA requires as many
each case. We also propose algorithms based on a new modified- power amplifiers (PAs) as the number of antennas. This design
CCD (M-CCD) based approach. For all the developed algorithms further limits the applicability. To alleviate the requirement of
we prove convergence to a Karush-Kuhn-Tucker (KKT) point multiple PAs, an alternative is to use a variable gain amplifier
(or a stationary point). Numerical results demonstrate that the
proposed algorithms converge faster than the state-of-the-art so- (VGA) driving a phase-shifting network. Thus, the beamforming
lutions. Among all the algorithms, the M-CCD-based algorithms vector is constrained to have constant-modulus entries, where
have faster convergence when evaluated in terms of the number of the magnitude of each entry corresponds to the gain contribution
iterations and the AMM-based algorithms offer lower complexity. from the VGA. In this case, the transmitted signals present low
Index Terms—Analog beamforming, majorization- Peak-to-Average-Ratio (PAR) and they enable the use of power-
minimization, MM, alternating MM, AMM, cyclic coordinate efficient nonlinear amplifiers at the transmitter’s side. This
descent, CCD, large-scale antenna arrays, unit-modulus characteristic is highly desirable in LSAA-based systems [7],
constraints, nonconvex optimization, block cyclic coordinate [8] because they do not require highly linear PAs which are
descent, BCCD.
necessary for the fully-digital beamforming implementation [9],
[10]. The analog beamforming architectures have several other
advantages, for example, adjusting only the phases of the beam-
Manuscript received February 24, 2020; revised August 7, 2020, October forming vector constrains the power required to drive an antenna
12, 2020, March 20, 2021, and June 12, 2021; accepted June 16, 2021. Date
of publication July 8, 2021; date of current version February 10, 2022. The element to be a constant. These gains are more appealing when
associate editor coordinating the review of this manuscript and approving it the transmitter/receiver employs an LSAA.
for publication was Dr. Augusto Aubry. This work was supported in part by
the National Research Fund (FNR), Luxembourg under the AFR-PPP grant for
Ph.D. Project SPASAT (Ref.: 11607283), and in part by the CORE-PPP project A. Literature Review
PROSAT, ECLECTIC, CI-PHY and DISBuS. Part of this work was published
in IEEE Global Conference on Signal and Information Processing 2019 [1]. The constant-modulus analog beamforming (CMAB) prob-
(Corresponding author: Aakash Arora.) lem has been studied in the past [11]–[15], where the optimal
The authors are with the Interdisciplinary Centre for Security, Reliability
and Trust (SnT), University of Luxembourg, L-1855 Luxembourg, Luxembourg beamforming vector is designed by solving an optimization
(e-mail: [email protected]; [email protected]; [email protected]; problem subject to the unit-modulus constraints on the entries of
[email protected]; [email protected]). the beamforming vector. For example, an signal-to-interference-
This article has supplementary downloadable material available at https://doi.
org/10.1109/TSP.2021.3094653, provided by the authors. plus-noise ratio (SINR) maximization problem is considered
Digital Object Identifier 10.1109/TSP.2021.3094653 in [11]. However, due to the unit-modulus constraints, the
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
ARORA et al.: EFFICIENT ALGORITHMS FOR CONSTANT-MODULUS ANALOG BEAMFORMING 757

resulting optimization problem is nonconvex and in general In the context of radar sequence design, the majorization-
NP-hard [16]. Two algorithms based on the conjugate gradient minimization (MM) and the CD frameworks have been adopted
and Newton’s method to compute the beamforming weights in the literature to handle the unit-modulus or constant-modulus
are proposed in [11]. These methods are the special cases of constraints [19], [27], [30]. In many of these applications, the
Riemannian optimization on manifolds. A gradient search algo- problem is modeled as a convex/nonconvex quadratically con-
rithm is presented in [12], with the angle parameterization of strained quadratic program (QCQP). In [19], [30], MM algo-
the unit-modulus constraints to adaptively adjust the phases of rithms to solve the ISL minimization problem are proposed.
the entries of the beamforming vector. Receive beamforming In [27], a CD-based procedure is proposed to solve the problem.
is studied in [13] and the optimal beamforming weights are But the development of an MM and a CD-based algorithm
designed by minimizing the mean square error (MSE) between crucially depends on the specific structure of the problem. In
the array output and the desired signal. For null steering, a contrast to the existing problems solved using MM or CD-based
position-perturbation technique is presented in [14]. algorithms which optimize a unit-modulus vector, our problem
Recently, this problem has received attention from a beam- entails optimizing two coupled unit-modulus vector variables
pattern matching perspective, where the desired beampattern and a multiplicative scalar variable. Therefore, the existing al-
is matched using an analog phase shifting network [1], [15]. gorithms in the literature can not be directly applied to the beam-
This problem is cast as a unit-modulus least squares (ULS) pattern matching problem because of fundamental differences
problem. A closely related problem of unimodular radar se- in the problem formulation. Furthermore, beamforming design
quence (or code) design also arises in several active sensing with unit-modulus constraints is NP-hard [16] and hence, there
applications [17]–[27]. In these applications, different perfor- exists scope for efficient and scalable algorithms with better
mance metrics for example, in [22] the Peak Sidelobe Level performance compared to the existing ones.
(PSL), and Integrated Sidelobe Level (ISL) based designs are
proposed. Therein, the worst-case PSL/ISL is optimized under B. Contributions
the steering vector mismatches. In [23], the worst-case signal-
In this paper, we consider the scaled analog beamforming
to-interference-plus-noise ratio (SINR) is maximized over the
(SAB) architecture and design the beamforming vector by solv-
steering vector mismatches under the constant-modulus and
ing the beampattern matching problem subject to the constant-
similarity constraints for radar waveform synthesis. In [24], an
modulus constraints [15]. In SAB architecture, a common vari-
objective function consisting of a specific weighted beampattern
able gain amplifier (VGA) drives the phase-shifting network.
matching error and the space-frequency stopband energy is
Therefore, the magnitude of each entry of the beamforming
minimized subject to the constant-modulus constraints. In [25],
vector is a constant, representing the gain introduced by VGA.1
a quadratic function is maximized subject to the unit-modulus
We consider two variants of the beampattern matching problem,
and similarity constraints. In [26], a wideband multiple-input
which can be cast as the constant-modulus least squares (CLS)
multiple-output (MIMO) radar transmit beampattern design
problem. In the first problem, we consider both gain and phase
with spectral and constant-modulus constraints is considered and
of the beampattern as variables, whereas in the second problem
is solved through a sequence of constrained quadratic programs
we match only the magnitude of the beampattern [15]. Later, we
such that the constant-modulus constraint is achieved at the
show that the first problem formulation becomes a special case
convergence. A recent paper [27] considers the synthesis of
of the second one. This is affected by formulating the problem
constant-modulus waveforms by maximizing the SINR subject
with unit-modulus beamforming weight and beampattern phase
to multiple spectral compatibility constraints. To solve the prob-
vectors as well as a scalar corresponding to VGA gain as
lem, an iterative procedure based on the coordinate descent (CD)
variables.
framework is proposed.
For the considered formulations, we provide redefficient
In the literature, an approximate solution to the ULS problem
algorithms, and the convergence guarantees to an associated
is found by employing the semidefinite relaxation (SDR) tech-
stationary point. In this context, the contributions of the work
nique, which lifts the dimensionality of the search space from
include:
M to M 2 . As a result, a large number of design parameters r Algorithms: We propose efficient algorithms specializing
need to be optimized, resulting in increased memory usage
on the MM and the cyclic coordinate descent (CCD)
and storage requirements, which may be impractical. Another
optimization frameworks to the considered problem for-
disadvantage with SDR is that it empirically returns a rank-1
mulations. As mentioned earlier, the MM and the CCD
solution, but in cases when the solution is not rank-1, an ap-
(or in general CD) are optimization frameworks and the
propriate randomization technique should be employed [16],
algorithm development under each framework crucially
[28]; this further increases the overhead. To that end, the
depends on the structure of the specific problem. Herein, we
scalability of the SDR technique is a bottleneck in designing
propose efficient and scalable algorithms for LSAAs by ex-
beamformers for LSAA-based systems. Therefore, to keep the
ploiting the problem structure with theoretical convergence
computational complexity low, gradient-projection (GP) based
guarantees. There are no works on exploring different
algorithms are proposed in [15]. Another recent work proposed
optimization frameworks for the increasingly important
an alternating direction method of multipliers (ADMM) based
algorithm for beamforming in the context of wireless sensor
networks [29]. 1 All the elements of the beamforming vector have the same magnitude.
758 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 70, 2022

beamforming problem [15]. In this context, the works

closer to the one pursued are [31]–[33] where the MM
and the CCD frameworks are utilized to solve different
optimization problems involving only one unit-modulus
constrained vector variables. Moreover, we show that the
optimization problem can be solved in closed-form in one
of the unit-modulus constrained vector variables as well
as the scaling variable. Therefore, we exploit the problem
structure and develop algorithmic solutions tailored to
the specific problems resulting in enhanced performance
gains in comparison to the existing works. Following we
summarize the algorithmic contributions:
r The MM-based algorithms utilize alternating minimiza- Fig. 1. CMAB or SAB architecture.
tion along with the MM framework. Thus, we call them
as alternating MM (AMM)-based algorithms. These, C. Organization of the Paper
algorithms differ from the standard MM formulation
The remainder of the paper is organized as follows. In Sec-
where the cost function is majorized for all variables,
tion II, the analog beamforming architecture is described and the
herein we only need majorization to handle the opti-
two problems are formulated. In Section III, we first describe
mization over the beamforming vector and the remain-
algorithmic frameworks followed by the proposed algorithmic
ing variables are optimized in closed-form. In partic-
solutions and convergence guarantees for the problem formula-
ular, this departs from the approach of [15] where the
tions presented in Section II. Simulation results are presented in
beampattern phase vector is optimized using a gradient-
Section IV and Section V concludes the work.
projection (GP) method and requires tuning an addi-
tional step-size parameter. This leads to effective and
efficient MM implementations. D. Notations Used
r Similarly, in the development of the CCD-based algo- The following notations are used throughout the paper. A
rithm, we do not solve all the associated single variable vector and a matrix are represented by a and A respectively.
sub-problems; only those associated with the beam- The i, j element of a matrix is denoted as A(i, j). The i-th
forming vector are solved componentwise, whereas the entry of a vector a is represented as a(i) or ai . The complex
other variables are updated block-wise as they admit exponential operation on each entry of a matrix is represented
a closed-form solution. This again leads to improved as e(jA) , the phase/argument of each element of a matrix is
CCD implementation than state-of-art. denoted as arg (A). The trace operator and the Frobenius norm
r Apart from the conventional-CCD (C-CCD) based ap- are represented as Tr(A) and AF ; a2 denotes the 2 norm
proach, we also propose modified-CCD (M-CCD) based of the vector. The real part of a scalar complex variable z
algorithms employing a new update rule and offering or a matrix varaible Z, is represented as Re(z). The symbol
faster convergence in comparison to the other algo- | · | denotes the modulus of a complex number. The Hermitian
rithms. operation, conjugate, and transpose of a matrix are denoted as
Convergence: We theoretically establish convergence guar- AH , A∗ , and AT respectively. The Schur-Hadamard product
antees for all the proposed algorithms. between two matrices is represented as A ◦ B; A  0 denotes
r Even though the constraints are nonconvex, we show a positive semi-definite (p.s.d.) matrix. The set of Hermitian
that the sequence of iterates generated by the AMM, positive semi-definite matrices is represented as Sn+ . A vector of
C-CCD, and M-CCD algorithms converge to a Karush- all ones and all zeros, each of size m are denoted as 1m×1 and
Kuhn-Tucker (KKT) point and is bounded. 0m×1 , respectively.
r We also prove that the solutions obtained by all the
algorithms satisfy the linear independence constraint II. ANALOG BEAMFORMING ARCHITECTURE AND PROBLEM
qualification (LICQ) (Proposition 3.1.1 in [34]) regu- FORMULATIONS
larity condition.
In this section, we first present the SAB architecture followed
Simulations: Numerical simulations under different beam-
by two beamforming design problem formulations.
pattern settings demonstrate the effectiveness of the pro-
posed algorithms. We analyze the evolution of beampat-
tern matching error with the number of iterations and A. SAB Architecture
study the scalability with the number of antennas. It is A transmitter equipped with an analog phase shifting network
observed that the proposed algorithms converge faster with driven by a common VGA serving multiple single-antenna users
a better beampattern matching accuracy in comparison is considered, as shown in Fig. 1. The use of a common VGA
with the state-of-the-art solutions existing in the literature. for all branches motivates the term SAB architecture. The gain
Moreover, increased performance gains are observed when of VGA is assumed to be an unknown and considered as a
LSAAs are employed. design variable. Therefore, the overall beamforming vector is
ARORA et al.: EFFICIENT ALGORITHMS FOR CONSTANT-MODULUS ANALOG BEAMFORMING 759

constrained to have constant-modulus entries, where the un- follow the same approach as well and formulate the following
known constant represents the gain introduced by the VGA. minimization problem,
Herein, we consider a uniform linear array (ULA) with M
P2 : minr∈R,w̃,u y ◦ u − Aw̃22
antenna elements and a spacing d = λ2 at the transmitter, where
λ represents the wavelength of operation. The array transmits subject to |w̃i | = r, ∀i ∈ [1, M ]
the same information to the users, therefore, this a broadcast
|uj | = 1, ∀j ∈ [1, N ].
beamforming scenario [35]. The array response of a ULA in a
direction θi is modeled as, In the subsequent section, we propose efficient algorithms to
 T solve problems P1 and P2 .
a(θi ) = 1 ejθi ej2θi . . . ej(M −1)θi . (1) Remark 1: It is important to note that problem P1 can be
viewed as a special case of problem P2 . Therefore, first, we
We consider an uniform discretization of the angular space propose efficient algorithms for solving problem P2 , and the
2(N −1)π T
into N points as θ = [0, 2π 4π
N , N ,..., N ] . Then, the ar- algorithms for solving problem P1 are derived afterward.
ray response from these directions can be written compactly
in matrix form as A(θ) = [a(θ1 ), a(θ2 ), . . . , a(θN )]H , where III. ALGORITHMIC SOLUTIONS FOR ANALOG BEAMFORMING
θi represents the i-th element of θ. The beampattern in the
direction θi takes the form y(θi ) = a(θi )H w̃, where w̃ is the Before proceeding towards the algorithm development, first,
beamforming vector to be designed. For notational convenience, we briefly discuss the algorithmic frameworks, MM, AMM, and
from now onwards we write y(θi ), a(θi ) and A(θ) as yi , ai and CCD. These preliminaries are added for the sake of improving
A, respectively. the readability.

B. Beampattern Matching A. Algorithmic Frameworks

The least-squares beampattern matching problem after con- 1) The MM Algorithm: The majorization-minimization
sidering the requirements in all {θi } is formulated as, method works on the principle of iteratively solving a sequence
of easier problems [36]–[39]. For example, let us consider the
P1 : minr∈R,w̃ y − Aw̃22 following minimization problem,
subject to |w̃i | = r, ∀i ∈ [1, M ], P3 : minx f (x)
where y is an N -dimensional vector denoting the desired re- subject to x ∈ G,
sponse along the directions represented by the elements of vector
where f : G → R is a continuous function, G is the constraint
θ and the variable r models the gain introduced by the VGA.
set and x is the unknown decision variable. In particular, an
The elements of the overall beamforming vector w̃ are con-
algorithm based on the MM framework starts with a feasible
strained to be constant modulus, thus representing phase-only
point x(0) ∈ G and iteratively solves the following problem,
beamforming.
x(k+1) ∈ arg min f˜(x; x(k) )
x
C. Beampattern Matching With Additional Degrees of
Freedom subject to x ∈ G, (3)
In the case of transmit beamforming, it may be required to where f˜(x; x(k) ) is the surrogate function majorizing the orig-
match only the magnitude of the beampattern because in general, inal objective function f (x) at x(k) and x(k) is the solution to
a receiver has to compensate for phase inconsistency [15]. In the above problem at k-th iteration. A valid surrogate function
this case, the system of equations for matching the beampattern for the minimization problem has the following properties,
magnitude can be expressed as,
f˜(x; x(k) ) ≥ f (x), ∀x ∈ G (4)
y = |Aw̃| , (2)
f˜(x(k) ; x(k) ) = f (x(k) ) (5)
where | · | denotes the entry-wise magnitude of the vector Aw̃.
Because of the non-differentiability of the modulus function, ∇f˜(x(k) ; x(k) ) = ∇f (x(k) ). (6)
the least-squares matching problem similar to problem P1 will Inequality (4) and equation (6) imply that the the surrogate
result in a non-differentiable objective function. To that end, we function is a tight upper bound of the original objective function.
can equivalently rewrite the system of equations in (2) as given Because of this property of the surrogate function, the objective
by, function value decreases with the number of iterations, ulti-
y ◦ u = Aw̃, mately converging to a stationary point of the original problem.
Equation (6), which is termed as the gradient consistency [40],
where u represents an additional vector with its i-th entry ui = ensures that the surrogate function and the objective function
H
ej arg(ai w̃) and the symbol ◦ denotes the element-wise product have the same gradients at x(k) .
between two vectors or matrices. Another way of modeling To formally prove the convergence of an MM algorithm,
this scenario is to consider |u| = 1N ×1 [15]. For simplicity, we first, we introduce the first-order optimality condition for the
760 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 70, 2022

minimization of a continuously differentiable function from of the cyclic coordinate descent (CCD) algorithm. For more
Proposition 3 in [30] as, information, one may refer to [34] and references therein.
Proposition III.1 (Proposition 3 in [30]): Consider f : Rn → 4) Modified BCCD Algorithm: Consider problem P4 and
R be a continuously differentiable function and if x(∞) is a local select m out of n blocks of the decision variable x. Let U =
minimum of f over a subset G of Rn , then {1, 2, . . . , n} and U1 = {n1 , n2 , . . . , nm } denote the index set
 T   for all the blocks of decision variable x and the index set of
∇f x(∞) y ≥ 0, ∀y ∈ TG x(∞) , (7) the selected blocks, respectively, where each ni ∈ [1, n] and
ni = nj , ∀i = j. Since, the constraints are separable in each xi
where TG (x(∞) ) denotes the tangent cone of G at x(∞) . we propose to consider the following nested algorithm.
A vector x satisfying the optimality condition (7) is referred (0)
1) Initialize all xi , ∀i ∈ [1, n] to a feasible point from their
to as a stationary point. For more insights on the MM framework respective constraint sets and set k = 0.
and its convergence properties, one may refer to [30], [36]–[39] 2) For all i ∈ U \ U1
and references therein. (k)
a) Solve the subproblem for xi given the remaining
2) Alternating Majorization-Minimization (AMM) Algo-
updated blocks of x.
rithm: Let us consider the following minimization problem,
(1) Solve the subproblems for xj , ∀j ∈ U1 given the
P4 : minx f (x1 , x2 , . . . , xn ) remaining updated blocks of x.
b) Increment i to i + 1 and go to step 2a.
subject to xi ∈ Gi , ∀i = 1, 2, . . . , n
3) If the convergence condition is met, then stop. Otherwise,
where f : G → R is the objective function, G = G1 × G2 × increment k to k + 1 and go to step 2.
· · · × Gn is the constraint set and x = (x1 , x2 , . . . , xn ) is the In this algorithmic setup, the variables xi , ∀i ∈ U1 are up-
decision variable partitioned into n-blocks with each block dated more frequently than the remaining block variables. As
xi having dimensions ni × 1. At the k-th iteration, following will be seen later, this modification achieves a lower objective
sub-problem is solved, per-iteration value than the C-CCD-based approach. For the
  specific case of beampattern matching problems P1 and P2 ,
(k+1) (k+1) (k+1) (k)
xi ∈ arg min gi xi ; x1 , . . . , xi−1 , xi , . . . , x(k)
n we will establish convergence guarantees to a stationary point.
xi
We refer to the algorithms developed based on this approach as
subject to xi ∈ Gi , (8) M-CCD-based algorithms.
for all i = 1, 2, . . . , n and the blocks are updated in a cyclic order A maximum block improvement (MBI) technique can also be
(j) adopted to solve problems P1 and P2 [41]–[43]. But selecting the
and xi denotes the update available for block xi at j-th itera-
maximum improvement block requires the computation of the
tion. Similar to the conventional MM framework, here function
(k+1) (k+1) (k) (k) objective function with respect to each block variable at every
gi (xi ; x1 , . . . , xi−1 , xi , . . . , xn ) is a tight majorizer
iteration of an MBI selection rule-based algorithm [41], [42].
of the original objective function in block variable xi and
This significantly increases the computational complexity of the
satisfies the properties from (4)-(6). The AMM algorithm can
algorithm especially for the large dimensional block variables
also be interpreted as a block-successive upper minimization
and when a large number of blocks are to be optimized as in the
(BSUM) [40] / block MM algorithm [38]. It is important to
case of beamforming design with LSAAs.
note that in the AMM algorithm, nonconvex constraints can be
accommodated. Moreover, only those objective functions are
majorized for which the subproblems are not easy to minimize. B. Algorithms for Problem P2
Otherwise, the original objective function is minimized with It can be seen that problem P2 is nonconvex because of the
respect to each block. Therefore, the convergence proof in [40] multiplicative variables u and w, and the constant-modulus
is adapted accordingly. A sequential optimization algorithm with constraints. Even if, the variables r and u are known the problem
a maximum block improvement (MBI) technique [41], [42] is is shown to be NP-hard [16]. We reconsider problem P2 and
also proposed in [43] for dealing with resource allocation in write it in a more amenable form as given by,
wireless networks and radar systems.
3) Block Cyclic Coordinate Descent (BCCD) Algorithm: In P6 : mins∈C,w,u f (s, u, w) = y ◦ u − sAw22
the BCCD approach, problem P4 is solved block-wise. Similar
subject to w ∈ AM , u ∈ AN ,
to the AMM framework, here, the problem is solved for each
block while keeping the remaining blocks fixed. The most where A = {x ∈ C| |x|2 = 1} and AM denotes the M -ary
common form of the BCCD algorithm is given by, Cartesian product, the constant modulus constraints on vector
 
(k+1) (k+1) (k+1) (k) w̃ in P2 are equivalently replaced by an unconstrained complex
P5 : x i ∈ arg min f x1 , x2 , xi , xi+1 , . . . , x(k)
n
xi ∈Gi factor s multiplying w and variable u is a vector having the unit-
modulus entries. The beamforming vector is now constrained to
subject to xi ∈ Gi ,
have the unit-modulus entries and the magnitude of variable s
for all i = 1, 2, . . . , n. Therefore, each minimization step con- represents the gain introduced by the VGA.
siders the previously computed minimizers. When block size 1) AMM-Based Algorithm: Problem P6 is still nonconvex
reduces to one, the resulting algorithm is known by the name and cannot be solved jointly in all the variables. But, in present
ARORA et al.: EFFICIENT ALGORITHMS FOR CONSTANT-MODULUS ANALOG BEAMFORMING 761

form the objective function of problem P6 is partially convex, the following Lemma from [30], [36], to construct a majorizing
meaning, given one variable it is convex in the other. Thus, function for the objective function of problem (11).
we exploit the partial convexity of the objective function and Lemma III.2 (Lemma 1 in [30]): The quadratic function of the
propose an alternating minimization scheme. form wH Sw, with S being a Hermitian matrix is majorized by
At (k + 1)-th iteration, problem P6 is solved with-respect-to wH Tw + 2Re(wH (S − T)wk ) + wkH (T − S)wk at the point
variable s by assuming the solutions w(k) and u(k) for variables wk , where T is a Hermitian matrix such that T  S.
w and u, respectively. Then, the variable u is updated given Lemma III.2 can be easily proven using second order Taylor
s(k+1) and w(k) . Finally, the problem is solved for variable expansion and subsequently replacing the Hessian matrix S by
w using the solution, s(k+1) and u(k+1) for variables s and another Hermitian matrix T such that T  S. For a general twice
u, respectively. This procedure results in three subproblems as differentiable function, Lemma III.2 is also known by the name
given below: of quadratic upper bound principle as mentioned in equation
(4.6) under Section 4.6 in Chapter 4 of [37].
s(k+1) = arg min f (s, u(k) , w(k) ) (9)
s∈C For obtaining the solution for variable w, once again we
expand the objective function of problem (11) as given by,
u(k+1) = arg min f (s(k+1) , u, w(k) ) (10)
u∈AN
f (s(k+1) , u(k+1) , w) = (u(k+1) )H YH Yu(k+1)
w (k+1)
= arg min f (s (k+1)
,u (k+1)
, w). (11)  
w∈AM − 2Re s(k+1) (u(k+1) )H YH Aw
Working on the aforementioned idea, the sub-problem with
respect to variable s is convex and admits the following closed- + |s(k+1) |2 Aw22 . (15)
form solution, The first term in (15) is independent of w and the third term
(w(k) )H AH Yu(k) is convex in w. Therefore, using Lemma III.2 we majorize the
sk+1 =   , (12) third term and obtain a tight upper bound, as given by,
Aw(k) 2
2
wH P̃w ≤ λmax (P̃)wH w
where Y = diag(y) is a diagonal matrix. Now, considering  
variable s to be given as s(k+1) , we need to solve problem (10). + 2Re wH (P̃ − λmax (P̃)I)w(k)
But this minimization problem is not straightforward to solve
because of the unit-modulus constraints on vector u. To get an + (w(k) )H (λmax (P̃)I − P̃)w(k) , (16)
update for u, we assume s and w to be given and expand the
objective function of problem (10), where P̃ = |s(k+1) |2 AH A, I is a M × M identity matrix and
  we have chosen matrix T = λmax (P̃)I = |s(k+1) |2 λmax (P)I,
uH YH Yu − 2Re s(k+1) uH YH Aw(k) where P = AH A. The function λmax (P) represents the
 2 maximum eigenvalue of matrix P. Therefore, the function,
  f (s(k+1) , u(k+1) , w) is majorized as,
+ |s(k+1) |2 Aw(k)  . (13)
2
The first and the last terms in the objective function above are f (s(k+1) , u(k+1) , w)
independent of u, therefore, after ignoring the constant terms ≤ f˜(w; s(k+1) , u(k+1) , w(k) ) (17)
we arrive at the following formulation,  
  = yH y − 2Re s(k+1) uH YH Aw(k)
P7 : minu −Re s(k+1) uH YH Aw(k)
+ λmax (P̃)wH w
subject to u ∈ AN .  
It is seen that problem P7 admits the following closed-form + 2Re wH (P̃ − λmax (P̃)I)w(k)
solution,
+ (w(k) )H (λmax (P̃)I − P̃)w(k) . (18)
(k+1) j arg (s(k+1) Y H Aw(k) )
u =e . (14)
The first and fifth terms on the right-hand side of the inequal-
Now, assuming the solutions s(k+1) and u(k+1) for the variables ity (17) are independent of w, and the third term is also a
s and u, respectively, we need to solve problem (11). But this constant, w22 = M , due to the unit-modulus property of the
minimization problem is shown to be NP-hard [16]. Therefore, beamforming vector. After ignoring the constant terms on the
we propose to adopt the AMM framework and construct a right-hand-side (RHS) of (17), the majorized problem to solve
majorizing function for the objective function of problem (11). problem (11) for variable w is formulated as,
More precisely speaking, we propose an AMM algorithm that  
is similar to the block MM framework, except for one change. P8 : minw −Re wH b(k+1)
That is, sub-problems that are difficult to directly minimize are
subject to w ∈ AM ,
approximated using a majorizing function, and for the remaining
variables solution is directly computed in closed-form. We refer where b(k+1) = (s(k+1) )∗ Ỹu(k+1) − |s(k+1) |2 Qw(k) , Ỹ =
to this as an AMM algorithm. To this end, we propose to use AH Y, Q = (P − λmax (P)I) and w(k) is the solution available
762 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 70, 2022

at k-th iteration. It can be shown that problem P8 admits the where wi

(k+1)
denotes the update of wi at k + 1-th iteration,
following closed-form solution, A−i is the matrix formed after removing the i-th column from
w = ej arg (b
(k+1)
). (19) the matrix A, ãi is the i-th column of matrix A, w−i is the vector
(k)
formed by removing the i-th element from vector w and w−i
The complete algorithm summarizing the steps is presented in denotes the update available for vector w−i after k-th iteration. It
Algorithm 1. is important to note that the minimization problem (25) is solved
for each component wi .
Algorithm 1: MM Based CMAB Design with Additional As already shown, for the given values of variables u and w,
Degrees of Freedom. the minimization problem (20) with respect to variable s admits
Input: The matrix A, Y, w(0) ∈ A and u(0) ∈ A the closed-form solution as given in (12) with appropriate change
Output: w, s of iteration indices. For variable u, we do not have to update its
Set k = 0, P = AH A, Y  = AH Y, β = λmax (AH A) component variables sequentially, as problem (21) with respect
1: repeat index over k = 0 : N  − 1 to the block variable u admits the closed-form solution as given
(k) H  (k)
2: sk+1 = (w Aw ) Yu
2 ; by (14). Now, we consider the minimization problem (25) with
2
(k+1) H (k)
respect to the each component wi of the vector w, as given by,
3: u(k+1) = ej arg (s Y w )
;  2
(k+1) (k+1) ∗
) Ỹu(k+1) − |s(k+1) |2 Qw(k)  (k) 
4: b = (s P9 : minwi Yu(k+1) − s(k+1) hi − s(k+1) wi ãi 
(k+1) 2
5: w(k+1) = ej arg (b )
;
6: until convergence subject to wi ∈ A.
(k)  (k+1)  (k)
where hi = ( j<i wj ãj + j>i wj ãj ). Now, upon
2) CCD-Based Algorithm: Here, we present CCD-based al- expanding the objective function of problem P9 and ignoring
gorithms. We provide two versions of the CCD-type algorithms. the constant terms, we obtain,
The first algorithm follows the C-CCD-based approach, that is  
all the decision variables are concatenated into one vector as P10 : minwi −Re wi∗ (s(k+1) )∗ ãH i (Yu (k+1)
− s (k+1) (k)
h )
[s, uT , wT ]T and (M + N + 1) scalar subproblems are solved
subject to wi ∈ A.
at every iteration or lesser subproblems depending upon the
block size chosen. As already mentioned in Section III-A3, in It is evident that problem P10 admits the following closed-form
the second approach, we update variables s and u after updating solution,
each component, wi of the beamforming vector w. This modifi-  
(k)
j arg (s(k+1) )∗ ãH
i (Yu
(k+1)
−s(k+1) hi
cation results in faster convergence, for more information on the wi = e , (26)
advantages of M-CCD the reader is referred to the simulation
results. Later in this section, we also prove the convergence for all i in [1, M ].
guarantees to a stationary point for both approaches. Steps for M-CCD: For the M-CCD approach, we assume that
Steps for C-CCD: First, we write the update steps associated the w(k) and u(k) are the solutions available for variables w
with the C-CCD-based algorithm. and u, respectively at the k-th iteration. With the slight abuse
(k)
of notation, we use ui to denote the i-th inner update of u
s(k+1) = arg min f (s, u(k) , w(k) ) (20) at the k-th outer iteration of the algorithm. Then, the following
s∈C
sub-problems need to be solved to solve at every iteration,
u(k+1) = arg min f (s(k+1) , u, w(k) ) (21)
u∈AM (k+1) (k) (k) (k)
s1 = arg min f (s, u(k) , w1 , w2 , . . . , wM ) (27)
(k+1) (k) (k) s∈C
w1 = arg min f (s(k+1) , u(k+1) , w1 , w2 , . . . , wM )
w1 ∈A (k+1) (k) (k) (k)
(22) u1 = arg min f (s, u, w1 , w2 , . . . , wM ) (28)
u∈AN
(k+1) (k+1)
w2 = arg min f (s(k+1) , u(k+1) , w1 , w2 , w1
(k+1)
= arg min f (s1
(k+1) (k)
, u(k+1) , w1 , w2 , . . . , wM )
(k)
w2 ∈A w1 ∈A
(k) (k) (29)
w3 , . . . , w M ) (23)
(k+1) (k+1) (k+1) (k) (k)
.. s2 = arg min f (s, u1 , w1 , w2 , . . . , wM ) (30)
. s∈C
(k+1) (k+1) (k+1) (k+1) (k)
(k+1) (k+1) u2 = arg min f (s2 , u, w1 , w2 , w3 ,
wi = arg min f (s(k+1) , u(k+1) , w1 ,..., u∈AN
wi ∈A
(k+1) (k) (k) (k)
wi−1 , wi , wi+1 , . . . , wM ), (24) . . . , wM ) (31)
(k+1) (k+1) (k+1) (k+1)
.. w2 = arg min f (s2 , u2 , w1 , w2 ,
. w2 ∈A

(k+1) (k+1) (k) (k)

wM = arg min f (s(k+1) , u(k+1) , w1 ,..., w3 , . . . , w M ) (32)
wM ∈A
(k+1) ..
wM −1 , wM ), (25) .
ARORA et al.: EFFICIENT ALGORITHMS FOR CONSTANT-MODULUS ANALOG BEAMFORMING 763

(k+1)
sM
(k+1)
= arg min f (s, uM −1 , w1
(k+1)
, w2
(k+1)
, The algorithms proposed for solving problem P2 can be tailored
s∈C for P11 as well. Therefore, for brevity, we do not provide the
(k+1) (k) detailed derivation steps for obtaining the algorithms.
. . . , wM −1 , wM ) (33)
1) AMM-Based Algorithm: In this case, the solution to prob-
(k+1) (k+1) (k+1) lems (9) and (11) become
uM = arg min f (sM , u, w1 ,
u∈AN
(w(k) )H AH y
(k+1) (k)
. . . , wM −1 , wM ) (34) s(k+1) =   , (36)
Aw(k) 2
2
(k+1) (k+1) (k+1) (k+1) (k+1)
wM = arg min f (sM , uM , w1 , w2 , and
wM ∈A

(k+1) (k+1)
. . . , wM −1 , wM ). (35) w = ej arg (c ), (37)
where c(k+1) = (s(k+1) )∗ AH y − |s(k+1) |2 Qw(k) , respec-
After the completion of (k + 1)-th iteration, we denote
(k+1) (k+1) (k+1) (k+1) tively. The overall algorithm summarizing the steps is presented
s(k+1) = sM , w(k+1) = [w1 , w2 , . . . , w M ]T in Algorithm 3.
(k+1) (k+1)
and u = uM . As can be seen from the above updates
the variables s and u are updated after obtaining the solution Algorithm 3: MM Based CMAB Design.
of each wi . Each sub-problem in the aforementioned steps
Input: The matrix A, y and w0 ∈ A
is solved using the solutions obtained from C-CCD based
Output: w, s
approach.
SET k = 0, P = AH A, ỹ = AH y, β = λmax (AH A),
Depending upon where we update variables s and u, we obtain
Q = P − βI
two different algorithms as summarized in Algorithm 2. To
1: repeat index over k = 0 : N  − 1
distinguish between the two algorithms we denote conventional (w (k) H
) H

and modified CCD-based algorithms as Algorithm 2 (C-CCD) 2: s(k+1) = Aw(k) 2 y ;

A
2
and Algorithm 2 (M-CCD), respectively. As pointed out earlier 3: c(k+1) = (s(k+1) )∗ ỹ − |s(k+1) |2 Qw(k) ;
(k+1)
in Section III-A4, here we also expect that the M-CCD algorithm 4: Compute w(k+1) = ej arg (c )
;
to have faster convergence than the other algorithms. This will 5: until convergence
be seen shortly in Section IV from simulation results.
The convergence of Algorithm 3 is proven as a special case
Algorithm 2: Conventional/Modified-Cyclic Coordinate of the convergence of Algorithm 1, which will be presented in
Descent (C/M-CCD)-Based CMAB Design with Additional Section III-B1. The reader is referred to the proof of Theorem
Degrees of Freedom. III.3 in Appendix A.
Input: The matrix A, y and w0 ∈ A 2) CCD-Based Algorithm: Here, we present two CCD based
Output: w, s algorithms as a special case of the ones in Section III-B2.
Set k = 0, u = 1 When |u| = 1N ×1 , the solution to problem (20) is given by (36)
1: repeat index over k = 0 : N  − 1 with appropriate change of iteration indices and the solution to
2: Update s using (12); Conventional version problem (25) is given by,
3: Update u using (14); Conventional version wi = ej arg ((s
(k+1) ∗ H
) ãi (y−s(k+1) d(k) ))
, ∀i ∈ [1, M ], (38)
4: for i = 1 to M do
(k)  (k+1)  (k)
5: Update s using (12); Modified version where di = ( j<i wj ãj + j>i wj ãj ). The steps of
6: Update u using (14); Modified version the algorithms are presented in Algorithm 4. Likewise, in the
7: Update wi using (26); case of problem P2 , here we also expect the M-CCD-based
8: end for algorithm to perform better than the C-CCD-based algorithm.
9: until convergence The convergence of Algorithm 4 for both modified as well
as conventional approaches are proven as a special case of the
convergence of Algorithm 2, to be presented in Section III-B2.

D. Convergence Analysis
C. Algorithms for Problem P1
In this section, we present convergence guarantees of the
As mentioned earlier problem P1 can be considered as a spe-
proposed algorithms based on the AMM, C-CCD and M-CCD
cial case of P2 , that is when ui = 1, ∀i ∈ [1, N ] and eliminating
frameworks.
its update from the algorithms. Similar to problem P2 , problem
1) Convergence Analysis of AMM-Based Algorithms: First,
P1 can be reformulated as,
we prove the convergence guarantees to a stationary point for
Algorithm 1 in Theorem III.3. Because the convergence of
P11 : mins∈C,w g(s, w) = y − sA(θ)w22
Algorithm 3 follows the same steps except for one change, that
subject to w ∈ AM . is by substituting u = 1N ×1 and ignoring its update in the proof
764 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 70, 2022

N M ). Whilst the complexity of M-CCD based Algorithm 2

Algorithm 4: Conventional/Modified-Cyclic Coordinate
increases to O(N M 2 + N M ).
Descent (C/M-CCD)-Based CMAB Design.
3) Algorithm 3 (MM): The complexity of computing each
Input: The matrix A, y, w0 ∈ A and s0 ∈ C
step of Algorithm 3 is similar to that of Algorithm 1, except for
Output: w, s
the removal of update of variable u while updating wi ’s. Thus,
Set k = 0,
there will be a reduction by a factor O(N M ).
1: repeat index over k = 0 : N  − 1
4) Algorithm 4 (C-CCD and M-CCD): Similar to Algorithm
2: for i = 1 to M do
2, the computational complexity for the C-CCD based Algorithm
3: Update s using (36); Modified version
2 is modified due to the removal of the update of variable u while
4: Update wi using (38);
updating wi . In comparison to 2 (C-CCD), the complexity of
5: end for
M-CCD based Algorithm 3 increases to O(N M 2 + N M ).
6: Update s using (36); Conventional version
7: until convergence
IV. SIMULATION RESULTS
In this section, we provide numerical simulations to evaluate
of Theorem III.3. Therefore, the convergence of Algorithm 3 the performance of the proposed algorithms and show their
can be easily proven as a special case of Algorithm 3. potential in different scenarios. Specifically, we consider two
Theorem III.3: Let {s(k) , w(k) , u(k) } be a sequence gener- beamforming scenarios, namely pencil and sector beamforming.
ated by Algorithm 1. Then, every limit point of the sequence is For each scenario, we compare the proposed algorithms to
a KKT-point of problem P6 . the ones in [15] and the semidefinite relaxation (SDR) based
Proof: See Appendix A.  algorithm. The performance is evaluated based on the evolution
2) Convergence Analysis of CCD-Based Algorithms: We es- of the beampattern matching error with the number of iterations
tablish the convergence guarantees to a stationary for Algorithm and the scalability with the number of antennas. All the results
2 for both the conventional as well as modified CCD based are averaged over 10 000 Monte-Carlo runs. The advantages of
approaches in Theorem III.4. Similar, to Section III-D1 the having additional degrees of freedom as shown in problem P2
convergence for Algorithm 4 for both conventional as well as in comparison to problem P1 are already mentioned in [15].
modified approaches follow the same steps with u = 1N ×1 and Therefore, we consider problem P1 for the pencil beamforming
removing its update in the proof of Theorem III.4. Therefore, and problem P2 for the sector beamforming scenario.
we consider the general case of Algorithm 4.
Theorem III.4: Let {s(k) , w(k) , u(k) } be the sequence gener- A. Pencil Beamforming
ated by Algorithm 2 using either the conventional or the modified
CCD approaches. Then, every limit point of {s(k) , w(k) , u(k) } Herein, we consider the scenario with M = 100, antenna
is a KKT-point of problem P6 . elements, the angle space is uniformly discretized into N = 36
Proof: See Appendices B and C for C-CCD and M-CCD based points in radians. The beampattern vector y is generated accord-
algorithms, respectively, in the supplemental information.  ing to the following equation,

1 if i ∈ I,
E. Complexity Analysis yi = (39)
0 otherwise,
1) Algorithm 1 (MM): From Algorithm 1, it is seen that
the computational complexity is mainly affected by the matrix- where I ∈ [1, N ] denotes the index set for non-zero entries of
matrix product and the eigenvalue computation. The maximum vector y with cardinality of card(I) = K. For fairness, all the
eigenvalue of Hermitian positive semidefinite matrix P is easy to algorithms are initialized to the same feasible starting point. We
compute using Krylov–Schur Algorithm [44]. Since the matrix choose K = 4 entries to construct the index set I from [1, N ],
P and the maximum eigenvalue of P is computed outside the and select the corresponding K angles from θ. The index set
loop, we analyze the worst-case per iteration complexity by chosen in this case is, I = {5, 10, 15, 20}. We then obtain y
considering the naive implementation of matrix-vector products. according to (39). Fig. 2 shows the objective function variation of
The quantity Aw22 can be computed in O(N 2 ) operations. problem P11 with the number of iterations. It is observed that the
The product Pw is computed in O(M 2 ) operations, whereas the proposed algorithms converge faster with a lower beampattern
phase computation can be done in O(M ) operations. In addition matching error in comparison to the state-of-the-art algorithms
to this, variable u can be computed in O(N M ) operations. in the literature [15] and as well as the SDR-based algorithm.
2) Algorithm 2 (C-CCD and M-CCD): The computational It is important to highlight the fact that the M-CCD-based
complexity for the C-CCD approach in Algorithm 2 is dominated algorithm outperforms the other algorithms in the comparison.
by the computation of elements wi ’s of the beamforming vector The modified update for the variable s reduces significantly the
w. Each wi is computed in O(N M ) operations, thus, to compute objective function in comparison to the C-CCD-based algorithm.
all the entries of w, O(N M 2 ) operations are required. For the M- One of the potential reasons is being the fact that, for each vari-
CCD-based approach, the computation of each wi is followed by able of problem P6 given the rest of the variables, the problem
an update of s and u, both of them can be computed in O(N M ) admits a closed-form solution. Therefore, updating s inside the
operations. The per-iteration complexity becomes O(N M 2 + inner loop of Algorithm 4 (M-CCD) significantly decreases the
ARORA et al.: EFFICIENT ALGORITHMS FOR CONSTANT-MODULUS ANALOG BEAMFORMING 765

Fig. 2. Objective function evolution with the number of iterations in pencil

beamfomring scenario, N = 36 and M = 100. Fig. 4. Variation of the beampattern matching error with the number of antenna
elements in pencil beamforming scenario, N = 36.

Fig. 3. Sampled beampatterns obtained from different algorithms (refer to Fig. 5. Variation of the average per-iteration runtime with the number of
(39)) for pencil beamforming, N = 36 and M = 100. antenna elements in pencil beamforming scenario, N = 36.

objective function in comparison to Algorithm 4 (C-CCD). The with respect to the number of antenna elements as observed
sampled beampatterns obtained are shown in Fig. 3. Kindly note in [15]. Once, again the M-CCD approach is superior to the
that the algorithms for pattern matching consider only the sample other algorithms. The average per-iteration runtime required by
points as presented in (39). Under this formulation, it is seen that the algorithms with increasing the number of antennas is shown
the proposed algorithms match the original beampattern better in Fig. 5. It is seen that Algorithm 3 uses a similar amount of av-
than the existing algorithms. erage per-iteration central processing unit (CPU) time to that of
Next, we study the scalability of the proposed algorithms with Algorithm 2 in [15] albeit having superior beampattern matching
increasing the number antennas. The stopping criteria chosen for accuracy as observed from Fig. 4. It is important to highlight that
the algorithms is either wkw
−wk−1 2
k 2
≤ 10−6 or a maximum num- the average per-iteration CPU time required by all the proposed
ber iterations are reached. The average beampattern matching algorithms is significantly lesser than the SDR-based algorithm.
error variation with M is shown in Fig. 4, where the number of The CCD-based approaches presented in Algorithm 4 require
antenna elements are varied from 2 to 100. The index set chosen more time to converge, the reason behind this is the increased
in this case is, I = {5, 10, 15, 20}. It is observed that increasing per-iteration complexity. The modified approach results in far
the number of antenna elements results in a better beampattern better beampattern approximation as shown in Fig. 4 than all the
approximation. Moreover, the proposed algorithms achieve sev- algorithms but it has higher per-iteration complexity. Therefore,
eral orders of lower beampattern matching error in comparison to the AMM-based method provides a good trade-off between
Algorithm 2 from [15] and as well as the SDR-based algorithm. the performance and the computational complexity. On the one
It is also seen that the SDR-based algorithm does not scale well hand, it has better beampattern matching accuracy in comparison
766 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 70, 2022

Fig. 6. Objective function evolution with the number of iterations in sector Fig. 7. Sampled beampatterns obtained from different algorithms (refer to
beamforming scenario, N = 144 and M = 250. (39)) for sector beamforming scenario, N = 144 and M = 250.

to the Algorithm 2 in [15] but lesser than the CCD based

algorithms, whereas, on the other hand, it takes similar CPU
time to converge to that of Algorithm 2 from [15].

B. Sector Beamforming
In the sectored beamforming scenario, we quantize the
angle space with N = 144 points in radians and consider
M = 250. Similar, to the previous case the beampattern vec-
tor y is generated by (39), with the following index set,
I = {1, 2, . . . , 18, 55, 56, . . . , 90127128, . . . , 144}. The objec-
tive function variation of problem P6 with the number of itera-
tions is shown in Fig. 6. It is seen that the proposed algorithms
have superior performance to the one in [15] and Algorithm 2
(M-CCD) achieves lesser beampattern matching error within a
small number of iterations in comparison to the other algorithms. Fig. 8. Variation of the beampattern matching error with the number of antenna
Once again, it is also seen that the proposed M-CCD-based elements in sector beamforming, N = 144.
algorithm outperforms the other algorithms in comparison. This
performance gain comes from the modified update of variables
s and u as both the subproblems admit closed-form solutions. better beampattern approximation. The performance is further
Thus, updating s and u while updating wi in Algorithm 2 reinforced if the number of antenna elements is significantly
(M-CCD) significantly decreases the objective function in com- larger than the cardinality of the angular discretization. Once
parison to Algorithm 2 (C-CCD). This results in much better again, the proposed algorithms outperform the one in [15] as
beampattern matching accuracy. well as the SDR-based algorithm. The SDR-based algorithm
The sampled beampatterns obtained are shown in Fig. 7. does not scale well with the number of antenna elements as
Please note that the algorithms for pattern matching consider observed in [15].
only the sample points as presented in (39). Under this formu- The average per-iteration runtime required by the algorithms
lation, it is seen that the proposed algorithms match the original is shown in Fig. 9. It is observed that the Algorithm 1 based
beampattern better than the existing algorithms. on the AMM approach uses a lesser average per-iteration CPU
The scalability of the algorithms with the number of antenna time than Algorithm 3 from [15] albeit with better beampattern
elements is shown in Fig. 8, the number of antenna elements approximation accuracy. All the proposed algorithms require
is varied from 2 to 350. It is observed that the beampattern less per-iteration CPU time than the SDR-based algorithm.
matching error drops by several orders of magnitude when As expected, Algorithm 2 (modified as well as conventional
an LSAA is employed. This results in a better beampattern CCD) uses more CPU time per-iteration because of the elements
matching accuracy. All the algorithms perform similar until of the analog beamforming vector w that are updated sequen-
M = 144, which is the length of the beampattern vector y. tially followed by the computation of variables s and u. More-
After this point, the proposed algorithms result in significantly over, the M-CCD-based Algorithm 2 is slightly computationally
ARORA et al.: EFFICIENT ALGORITHMS FOR CONSTANT-MODULUS ANALOG BEAMFORMING 767

r The M-CCD-based algorithms converged with the least

beampattern matching error.
r The AMM-based Algorithm 3 performed even better in
terms of average CPU time to converge, thereby, providing
a good trade-off between performance and complexity.
Our results demonstrated the aforementioned gains of the
proposed algorithms by utilizing a new approach based on the
MM, alternating minimization, and CCD frameworks.

APPENDIX A
PROOF OF THEOREM III.3
Before presenting the convergence proof, we first note that
the optimization problem w.r.t u is not observable for the zero
entries of y. Hence, the focus is on the case of non-zero entries,
yi . Now, we prove the following Lemma VI.1 for the uniqueness
of the minimizer of the optimization sub-problems with respect
Fig. 9. Variation of the average per-iteration runtime with the number of
antenna elements in sector beamforming scenario, N = 144. to variables s and u which is used in the proof later in this section.
Lemma VI.1 (Uniqueness of Minimizers): The objective
f (s, w, u) has an unique minimizer for
r A. s given w, u, and
r B. u given s, w.
expensive in comparison with the conventional one because
of the sequential updates of variables s and u in conjunction Proof: We show the proof into two parts as follows:
with the components wi ’s. Therefore, AMM-based Algorithm Optimization of s given u and w: The considered objective
2 again provides a trade-off between the performance and the function, y ◦ u − sAw22 is strictly convex in s since the
computational complexity. second-order derivative satisfies, 2Aw22 > 0; this is because
A is full rank (Vandermonde) and w = 0. As a result, it has the
following unique minimizer,
V. CONCLUSION wH AH Yu
s∗ = . (40)
In this paper, we studied the problem of designing constant- Aw22
modulus analog beamforming (CMAB) systems equipped with
Optimization of u given s and w: The objective function in
large-scale antenna arrays (LSAAs). Two beampattern matching
variable u can be written as,
problems were considered to design the parameters of an analog

beamforming system based on the scaled analog beamforming f (s, u, w) = y22 + sAw22 − 2Re uH (sYH Aw) . (41)
(SAB) or CMAB architecture. In the beampattern matching
Noting that all the entries of vector u are unit-modulus, the
problem, the least-squares error between the desired and the de-
minimum of (41) occurs when u is aligned with sYH Aw
signed beampatterns was minimized subject to the unit-modulus
leading to,
constraints. In the first case, both the magnitude and phase of
H
the beampattern were matched to the given desired beampattern u∗ = ej arg(sY Aw)
. (42)
whereas in the second case, the magnitude of the beampattern H

is matched, resulting in coupled unit-modulus vectors. We show that u∗ = ej arg(sY Aw) is the unique minimizer, by
H

For each problem, we proposed efficient alternating contradiction. Let u+ = ej(arg(sY Aw)+θ) be another mini-
majorization-minimization (AMM) and conventional-cyclic co- mizer, where at least one entry of θ is non-zero and is not an
ordinate descent (C-CCD) based algorithms. In addition to the integer multiple of 2π (otherwise, u∗ = u+ ). Then, from (42),
conventional-CCD (C-CCD) based approach, we also proposed we have,

modified-CCD (M-CCD) algorithms. The convergence of the f (s, u∗ , w) = y22 + sAw22 − 2 |sYH Aw|i (43)
proposed algorithms to stationary points of the optimization i
problems was theoretically established. In numerical simula- 
tions, different beamforming scenarios were considered for the f (s, u+ , w) = y22 + sAw22 −2 |sYH Aw|i cos(θ i ).
i
comparison of the algorithms. We list the following observations
(44)
exhibiting the performance gains of the proposed algorithms:
r The proposed algorithms outperformed the state-of-the- Since both u∗ and u+ minimize f (s, u, w), it follows from
art solutions existing in the literature, resulting in better (43), (44) that, f (s, u∗ , w) = f (s, u+ , w). This is possible only
beampattern matching accuracy. if cos(θi ) = 1, ∀i ∈ [N ], which implies that all the entries of
r The M-CCD-based algorithms converged faster in com- θ are zero or some integer multiples of 2π going against the
parison to the other algorithms when visualized with the assumption. This leads to a contradiction and hence proving the
number of iterations. uniqueness. 
768 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 70, 2022

Now we are ready to prove Theorem III.3. and f˜(w; s(k) , u(k) , w(k) ), we get,
Proof: First, we recall from (9)-(11) the following update
order of the iterates from the algorithm at (k + 1)-th iteration, f (s, u(∞) , w(∞) ) ≥ f (s(∞) , u(∞) , w(∞) ). (56)

s(k+1) = arg min f (s, w(k) , u(k) ) (45) The inequality (56) implies that, s(∞) is a block-wise minimizer
s
of the function f (·) and therefore, satisfies the partial KKT
u(k+1) = arg min f (s(k+1) , w(k) , u) (46) conditions with respect to the variable s, given by,
u
∇s f (s(∞) , u(∞) , w(∞) ) = 0. (57)
w(k+1) = arg min f˜(w; s(k+1) , w(k) , u(k+1) ) (47)
w∈A
2) Convergence for u: We now focus on u, we begin with
where f˜(w; s(k+1) , w(k) , u(k+1) ) denotes the majorizing func- the following claim.
tion of the original objective function f (s, w, u) after the up- Claim 1 : Both s(kj ) and s(kj +1) converge to s(∞) .
dates of the variables s and u. The majorizing function also First, we show that s(kj +1) converges. This follows directly
depends upon the solution w(k) obtained at the k-th iteration. [w(kj ) ]H AH Yu(kj ) )
from the equation s(kj +1) = and t(∞) =
Aw(kj ) 22
(∞) H H (∞)
[w ] A Yu
A. Descent of the Objective function limj→∞ s(kj +1) → Aw(∞) 22
. We show that t(∞) =
s(∞) . Consider the following,
The following can be easily shown,
f (s, u(kj ) , w(kj ) ) ≥ f (s(kj +1) , u(kj ) , w(kj ) ). (58)
f (s(k) , u(k) , w(k) ) ≥ f (s(k+1) , u(k) , w(k) ) (48)
Now, letting j → ∞ and using the continuity of the functions
≥ f (s(k+1) , u(k+1) , w(k) ) (49) f (s, w, u) and u(kj ) → u(∞) , w(kj ) → w(∞) , and s(kj +1) →
= f˜(w(k) ; s(k+1) , u(k+1) , w(k) ) (50) t(∞) , it follows that,

≥ f˜(w(k+1) ; s(k+1) , u(k+1) , w(k) ) (51) f (s, u(∞) , w(∞) ) ≥ f (t(∞) , u(∞) , w(∞) ). (59)

≥ f (s(k+1) , u(k+1) , w(k+1) ). (52) This implies that t(∞) , in addition to s(∞) , is also the mini-
mizer of f (s, u(∞) , w(∞) ). Since f (s, u(∞) , w(∞) ) has a unique
Inequality (48) and (49) above follows from the updates of s and minimizer from Lemma VI.1, it follows that t(∞) = s(∞) . This
u, respectively. Equation (50) holds because the approximated implies limj→∞ s(kj +1) = s(∞) . We use this in the next step.
function is a valid majorizer in the block w, inequality (51) Block Minimization: To prove that u(∞) is a block-wise
follows from (47), inequality (52) follows from the descent minimizer of f (s, u, w) given s and w, we first write the
property of the MM framework. Therefore, the function se- following set of inequalities,
quence {f (s(k) , u(k) , w(k) )} decreases monotonically and thus,
f (s(kj +1) , w(kj ) , u) (60)
converges.
≥ f (s(kj +1) , u(kj +1) , w(kj ) ) (61)
B. Convergence to a Stationary Point
= f˜(w(kj ) ; s(kj +1) , u(kj +1) , w(kj ) ) (62)
Now, we show that the sequence generated by the algorithm
converges to a KKT-point of the problem. Assume a convergent ≥ f˜(w(kj +1) ; s(kj +1) , u(kj +1) , w(kj ) ) (63)
subsequence {s(kj ) , u(kj ) , w(kj ) } → {s(∞) , u(∞) , w(∞) }. It is
important to note that s(∞) ∈ C, u(∞) ∈ AN and w(∞) ∈ AM ≥ f (s(kj +1) , u(kj +1) , w(kj +1) ) (64)
because the respective constraint sets are closed. We now show ≥ f (s(kj+1 ) , u(kj+1 ) , w(kj+1 ) ). (65)
that s(∞) , u(∞) and w(∞) are the block-wise minimizers of the
function f (s, u, w) with respect to s, u and w, respectively. These inequalities are obtained using standard manipulations. It
1) Convergence for s: To demonstrate this, we present the should be noted that (60) differs from (53) in the fact that the
following set of inequalities, LHS of inequality uses a different subsequence of s. Using the
fact s(kj +1) converges to s(∞) as j → ∞ from Claim 1, and
f (s, u(kj ) , w(kj ) ) letting j → ∞, and given that s(kj+1 ) , s(kj +1) both converge to
≥ f (s(kj +1) , u(kj ) , w(kj ) ) (53) s(∞) , it can be shown that,

.. f (s(∞) , u, w(∞) ) ≥ f (s(∞) , u(∞) , w(∞) ), (66)

.
The inequality (66) implies that, u(∞) is a block-wise minimizer
≥ f (s(kj +1) , u(kj +1) , w(kj +1) ) (54) of f (·) and, therefore, satisfies the partial KKT conditions with
respect to the variable u, given by,
≥ f (s(kj+1 ) , u(kj+1 ) , w(kj+1 ) ). (55)
∇u f (s(∞) , u(∞) , w(∞) ) + 2ν ◦ u(∞) = 0, (67)
These inequalities are based on the descent of the objective
function from (48)-(52). Now, letting j → ∞, using the con- where ν ∈ RN ×1 is the dual-variable vector associated with the
vergence of the subsequences and the continuity of f (s, u, w) unit-modulus constraints on vector u.
ARORA et al.: EFFICIENT ALGORITHMS FOR CONSTANT-MODULUS ANALOG BEAMFORMING 769

3) Convergence for w: For the minimization with respect to majorizing function in (17) should have the same gradient at
w, we start with the following claim. (s(∞) , w(∞) , u(∞) ).
Claim 2 : Both u(kj ) and u(kj +1) converge to u(∞) . Regularity Condition: The regularity condition, linear in-
To prove the claim, first, we show that u(kj +1) converges to dependence constraint qualification (LICQ) implies that the
(∞)
u as j → ∞ as it depends upon s(kj +1) , w(kj ) and it has gradients of the active inequality constraints and the gradients of
been shown that the objective function is uniquely minimized. equality constraints are linearly independent at a feasible point
Thus, we can write, (Proposition 3.1.1 in [34]). This condition is automatically sat-
(kj +1)
Y H Aw(kj ) )
isfied at the solution as there are no constraints on variable s and
z(∞) = lim u(kj +1) = lim ej arg(s (68) the constraint sets involving w, u satisfy the LICQ condition.
j→∞ j→∞
(∞)
For w(∞) and u, the constraints |wi |2 = 1, for all i ∈ [M ] and
Y H Aw(∞) )
= ej arg(s . (69) |uj |2 = 1, for all j ∈ [M ] are decoupled among the entries wi ’s
and uj ’s, respectively. All the constrained optimization variables
We now show that z(∞) = u(∞) . For this, we present the fol-
can be combined in to a single vector as, x := [wT , uT ]T .
lowing inequality that follows from the minimization w.r.t u,
Now the gradient of constraint |xi |2 − 1 = 0 with respect to
f (s(kj +1) , u, w(kj ) ) ≥ f (s(kj +1) , u(kj +1) , w(kj ) ). (70) x is computed as, gi (x) = 2ei , for all i ∈ [M + N ], where
ei ∈ R(M +N )×1 is a vector with i-th entry being 1 and rest of
Now, letting j → ∞ and using the continuity of the func-
the entries are zeros. Clearly all the gradient vectors are linearly
tions f (s, w, u) and s(kj +1) → s(∞) (Claim 1), w(kj ) → w(∞)
independent. This implies that each point of the constraint set
(choice of subsequence), and u(kj +1) → z(∞) (as shown in
of problem P6 is regular.
(69)), it follows that,
Thus, combining the partial KKT-conditions along each block
f (s(∞) , u, w(∞) ) ≥ f (s(∞) , z(∞) , w(∞) ). (71) (57), (67) and (79), we get,
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
This implies that z(∞) , in addition to u(∞) , is also the min- ∇s f (s(∞) , u(∞) , w(∞) ) 0 0
imizer of f (s(∞) , u, w(∞) ). However, since f (s(∞) , u, w(∞) ) ⎢ (∞) (∞) (∞) ⎥ + 2 ⎢ ⎥ ◦ ⎢ ⎥ = 0
⎣ ∇u f (s , u , w ) ⎦ ⎣ν ⎦ ⎣ u ⎦ M +N +1 .
has a unique minimizer for u given w(∞) , s(∞) , it follows that (∞) (∞)
∇w f (s , u , w ) (∞)
λ w
z(∞) = u(∞) . This implies limj→∞ u(kj +1) = u(∞) . We use
(80)
this in the next step.
Block Minimization: To prove block minimizer, the proof Thus, we conclude that every limit point of the sequence gener-
follows the earlier presented ones, except for the fact that the ated by the algorithm is a KKT-point of the problem. 
inequality is for the majorizer,
ACKNOWLEDGMENT
f˜(w; s(kj +1) , u(kj +1) , w(kj ) )
The authors would like to thank the associate editor and the
≥ f˜(w(kj +1) ; s(kj +1) , u(kj +1) , w(kj ) ) (72) reviewers for providing the useful comments.
≥ f (s(kj +1) , u(kj +1) , w(kj +1) ) (73)
REFERENCES
≥ f (s(kj+1 ) , u(kj+1 ) , w(kj+1 ) ) (74) [1] A. Arora, C. G. Tsinos, B. Shankar Mysore R, S. Chatzinotas, and B. Ot-
tersten, “Majorization-minimization algorithms for analog beamforming
= f˜(w(kj+1 ) ; s(kj+1 ) , u(kj+1 ) , w(kj+1 ) ), (75) with large-scale antenna arrays,” in Proc. IEEE Global Conf. Signal Inf.
Process., Nov. 2019, pp. 1–5.
where (73) follows from the upper bound property (4) of the [2] A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, “Channel estimation
majorizing function function. Now, taking the limit j → ∞, and and hybrid precoding for millimeter wave cellular systems,” IEEE J. Sel.
by the convergence of s(kj +1) , u(kj +1) we obtain, Topics Signal Process., vol. 8, no. 5, pp. 831–846, Oct. 2014.
[3] O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath, “Spa-
f˜(w; s(∞) , u(∞) , w(∞) ) (76) tially sparse precoding in millimeter wave MIMO systems,” IEEE Trans.
Wireless Commun., vol. 13, no. 3, pp. 1499–1513, Mar. 2014.
[4] A. Arora, C. G. Tsinos, B. S. M. R. Rao, S. Chatzinotas, and B. Ot-
≥ f˜(w(∞) ; s(∞) , u(∞) , w(∞) ). (77) tersten, “Hybrid transceivers design for large-scale antenna arrays us-
ing majorization-minimization algorithms,” IEEE Trans. Signal Process.,
The inequality (72) implies that, w(∞) is a minimizer with vol. 68, pp. 701–714, 2020.
respect to the variable w for function f˜, as given by, [5] V. Murino and A. Trucco, “Three-dimensional image generation and
processing in underwater acoustic vision,” Proc. IEEE, vol. 88, no. 12,
∇w f˜(w(∞) ; s(∞) , u(∞) , w(∞) ) + 2λ ◦ w(∞) = 0 (78) pp. 1903–1948, Dec. 2000.
[6] E. Ben-Dor, T. S. Rappaport, Y. Qiao, and S. J. Lauffenburger, “Millimeter-
∇w f (s(∞) , u(∞) , w(∞) ) + 2λ ◦ w(∞) = 0, (79) wave 60 GHz outdoor and vehicle AOA propagation measurements using
a broadband channel sounder,” in Proc. IEEE Global Telecommun. Conf.,
Dec. 2011, pp. 1–6.
where λ ∈ RM ×1 is the dual-variable vector associated with [7] C. Mollén, E. G. Larsson, and T. Eriksson, “Waveforms for the massive
the unit-modulus constraints. In (78) we have used the gradient MIMO downlink: Amplifier efficiency, distortion, and performance,” IEEE
consistency condition between the objective function and its Trans. Commun., vol. 64, no. 12, pp. 5050–5063, Dec. 2016.
[8] S. K. Mohammed and E. G. Larsson, “Constant-envelope multi-user pre-
majorizer, given in (6), to obtain (79). This arises from the coding for frequency-selective massive MIMO systems,” IEEE Wireless
requirement that the original objective function in (15) and the Commun. Lett., vol. 2, no. 5, pp. 547–550, Oct. 2013.
770 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 70, 2022

[9] D. Spano, S. Chatzinotas, S. Andrenacci, J. Krause, and B. Ottersten, “Per- [33] M. A. Kerahroodi, A. Aubry, A. De Maio, M. M. Naghsh, and M.
antenna power minimization in symbol-level precoding for the multibeam Modarres-Hashemi, “A coordinate-descent framework to design low
satellite downlink,” Int. J. Satell. Commun. Netw., vol. 37, no. 1, pp. 15–30, PSL/ISL sequences,” IEEE Trans. Signal Process., vol. 65, no. 22,
2019. pp. 5942–5956, Nov. 2017.
[10] D. Spano, M. Alodeh, S. Chatzinotas, and B. Ottersten, “Symbol-level [34] D. P. Bertsekas, Nonlinear Programming, 2nd ed. Belmont, MA, USA:
precoding for the nonlinear multiuser MISO downlink channel,” IEEE Athena Scientific, 1999.
Trans. Signal Process., vol. 66, no. 5, pp. 1331–1345, Mar. 2018. [35] M. Alodeh et al., “Symbol-level and multicast precoding for multiuser
[11] S. T. Smith, “Optimum phase-only adaptive nulling,” IEEE Trans. Signal multiantenna downlink: A state-of-the-art, classification, and challenges,”
Process., vol. 47, no. 7, pp. 1835–1843, Jul. 1999. IEEE Commun. Surveys Tuts., vol. 20, no. 3, pp. 1733–1757, Jul.-
[12] P. Thompson, “Adaptation by direct phase-shift adjustment in narrow-band Sep. 2018.
adaptive antenna systems,” IEEE Trans. Antennas Propag., vol. AP-24, [36] D. R. Hunter and K. Lange, “A tutorial on MM algorithms,” Amer. Statist.,
no. 5, pp. 756–760, Sep. 1976. vol. 58, no. 1, pp. 30–37, 2004.
[13] C. Baird and G. Rassweiler, “Adaptive sidelobe nulling using digitally [37] K. Lange, MM Optimization Algorithms. Philadelphia, PA, USA: Society
controlled phase-shifters,” IEEE Trans. Antennas Propag., vol. AP-24, for Industrial and Applied Mathematics, 2016.
no. 5, pp. 638–649, Sep. 1976. [38] Y. Sun, P. Babu, and D. P. Palomar, “Majorization-minimization algorithms
[14] T. H. Ismail and M. M. Dawoud, “Null steering in phased arrays by in signal processing, communications, and machine learning,” IEEE Trans.
controlling the element positions,” IEEE Trans. Antennas Propag., vol. 39, Signal Process., vol. 65, no. 3, pp. 794–816, Feb. 2017.
no. 11, pp. 1561–1566, Nov. 1991. [39] K. Lange, D. R. Hunter, and I. Yang, “Optimization transfer using surrogate
[15] J. Tranter, N. D. Sidiropoulos, X. Fu, and A. Swami, “Fast unit-modulus objective functions,” J. Comput. Graph. Stat., vol. 9, no. 1, pp. 1–20, 2000.
least squares with applications in beamforming,” IEEE Trans. Signal [40] M. Razaviyayn, M. Hong, and Z.-Q. Luo, “A unified convergence analysis
Process., vol. 65, no. 11, pp. 2875–2887, Jun. 2017. of block successive minimization methods for nonsmooth optimization,”
[16] S. Zhang and Y. Huang, “Complex quadratic optimization and semidefinite SIAM J. Optim., vol. 23, no. 2, pp. 1126–1153, 2013.
programming,” SIAM J. Optim., vol. 16, no. 3, pp. 871–890, 2006. [41] B. Chen, S. He, Z. Li, and S. Zhang, “Maximum block improvement and
[17] P. Stoica, H. He, and J. Li, “New algorithms for designing unimodular polynomial optimization,” SIAM J. Optim., vol. 22, no. 1, pp. 87–107,
sequences with good correlation properties,” IEEE Trans. Signal Process., 2012.
vol. 57, no. 4, pp. 1415–1425, Apr. 2009. [42] Z. Li, A. Uschmajew, and S. Zhang, “On convergence of the maximum
[18] M. Soltanalian and P. Stoica, “Designing unimodular codes via quadratic block improvement method,” SIAM J. Optim., vol. 25, no. 1, pp. 210–233,
optimization,” IEEE Trans. Signal Process., vol. 62, no. 5, pp. 1221–1234, 2015.
Mar. 2014. [43] A. Aubry, A. De Maio, A. Zappone, M. Razaviyayn, and Z. Luo, “A
[19] Y. Li and S. A. Vorobyov, “Fast algorithms for designing unimodular wave- new sequential optimization procedure and its applications to resource
form(s) with good correlation properties,” IEEE Trans. Signal Process., allocation for wireless systems,” IEEE Trans. Signal Process., vol. 66,
vol. 66, no. 5, pp. 1197–1212, Mar. 2018. no. 24, pp. 6518–6533, Dec. 2018.
[20] Y. Li and S. A. Vorobyov, “Efficient single/multiple unimodular waveform [44] G. Stewart, “A Krylov-Schur algorithm for large eigenproblems,” SIAM
design with low weighted correlations,” in Proc. IEEE Int. Conf. Acoust., J. Matrix Anal. Appl., vol. 23, no. 3, pp. 601–614, 2002.
Speech, Signal Process., 2017, pp. 3226–3230.
[21] Y. Li and S. A. Vorobyov, “Joint space-(slow) time transmission with
unimodular waveforms and receive adaptive filter design for radar,” in
Proc. IEEE Int. Conf. Acoust., Speech, Signal Process., 2018, pp. 3276–
3280.
[22] A. Aubry, A. De Maio, and Y. Huang, “MIMO radar beampattern design
via PSL/ISL optimization,” IEEE Trans. Signal Process., vol. 64, no. 15,
Aakash Arora (Student Member, IEEE) received
pp. 3955–3967, Aug. 2016.
the M.S. (Research) degree in electrical engineering
[23] X. Yu, G. Cui, L. Kong, J. Li, and G. Gui, “Constrained waveform design
from the Indian Institute of Technology Delhi, India,
for colocated MIMO radar with uncertain steering matrices,” IEEE Trans.
in 2017. He is currently working toward the Ph.D.
Aerosp. Electron. Syst., vol. 55, no. 1, pp. 356–370, Feb. 2019.
degree in computer science from the Interdisciplinary
[24] X. Yu, G. Cui, J. Yang, and L. Kong, “Wideband MIMO radar beam-
Centre for Security, Reliability and Trust, University
pattern shaping with space-frequency nulling,” Signal Process., vol. 160,
of Luxembourg, Luxembourg. His research inter-
pp. 80–87, 2019.
ests include optimization algorithms, statistical signal
[25] G. Cui, X. Yu, G. Foglia, Y. Huang, and J. Li, “Quadratic optimization
processing, wireless communications, and machine
with similarity constraint for unimodular sequence synthesis,” IEEE Trans.
learning.
Signal Process., vol. 65, no. 18, pp. 4756–4769, Sep. 2017.
[26] B. Kang, O. Aldayel, V. Monga, and M. Rangaswamy, “Spatio-spectral
radar beampattern design for coexistence with wireless communication
systems,” IEEE Trans. Aerosp. Electron. Syst., vol. 55, no. 2, pp. 644–657,
Apr. 2019.
[27] A. Aubry, A. De Maio, M. A. Govoni, and L. Martino, “On the design of
multi-spectrally constrained constant modulus radar signals,” IEEE Trans.
Signal Process., vol. 68, pp. 2231–2243, 2020. Christos G. Tsinos (Senior Member, IEEE) received
[28] Z. Luo, W. Ma, A. M. So, Y. Ye, and S. Zhang, “Semidefinite relaxation the Diploma degree in computer engineering and in-
of quadratic optimization problems,” IEEE Signal Process. Mag., vol. 27, formatics, the M.Sc. and Ph.D. degrees in signal pro-
no. 3, pp. 20–34, May 2010. cessing and communication systems, and the M.Sc.
[29] C. G. Tsinos and B. Ottersten, “An efficient algorithm for unit-modulus degree in applied mathematics from the University
quadratic programs with application in beamforming for wireless sen- of Patras, Greece, in 2006, 2008, 2013, and 2014,
sor networks,” IEEE Signal Process. Lett., vol. 25, no. 2, pp. 169–173, respectively. From 2014 to 2020, he held a Postdoc-
Feb. 2018. toral Researcher, Research Associate and Research
[30] J. Song, P. Babu, and D. P. Palomar, “Optimization methods for design- Scientist positions with the University of Patras and
ing sequences with low autocorrelation sidelobes,” IEEE Trans. Signal University of Luxembourg, Luxembourg. He has re-
Process., vol. 63, no. 15, pp. 3998–4009, Aug. 2015. cently been elected as an Assistant Professor with
[31] L. Wu, P. Babu, and D. P. Palomar, “Transmit waveform/receive filter the University of Athens, Greece. He was involved in a number of different
design for MIMO radar with multiple waveform constraints,” IEEE Trans. Research and Development projects funded by national and/or EU funds. His
Signal Process., vol. 66, no. 6, pp. 1526–1540, Mar. 2018. current research interests include optimization and machine learning for signal
[32] L. Wu, P. Babu, and D. P. Palomar, “Cognitive radar-based sequence design processing and communications. He is a Member of the Technical Chamber of
via SINR maximization,” IEEE Trans. Signal Process., vol. 65, no. 3, Greece.
pp. 779–793, Feb. 2017.
ARORA et al.: EFFICIENT ALGORITHMS FOR CONSTANT-MODULUS ANALOG BEAMFORMING 771

M. R. Bhavani Shankar (Senior Member, IEEE) Björn Ottersten (Fellow, IEEE) received the M.S.
received the master’s and Ph.D. degrees in electrical degree in electrical engineering and applied physics
communication engineering from the Indian Institute from Linköping University, Linköping, Sweden, in
of Science, Bangalore, India, in 2000 and 2007, re- 1986 and the Ph.D. degree in electrical engineering
spectively. From 2007 to September 2009, he was from Stanford University, Stanford, CA, USA, in
a Postdoc with the ACCESS Linnaeus Centre, Sig- 1990. He has held Research positions with the De-
nal Processing Lab, Royal Institute of Technology partment of Electrical Engineering, Linköping Uni-
(KTH), Sweden. In October 2009, he joined SnT, versity, the Information Systems Laboratory, Stan-
as a Research Associate and is currently a Research ford University, the Katholieke Universiteit Leuven,
Scientist with SnT. From 2006 and 2007, he was Leuven, Belgium, and the University of Luxembourg,
with Beceem Communications, Bangalore, India, as Luxembourg. From 1996 to 1997, he was the Director
a Staff Design Engineer, working on Physical Layer algorithms for WiMAX of Research with ArrayComm, Inc., a start-up in San Jose, CA, USA, based on
compliant chipsets. In 2004, he was a Visiting Student with the Communication his patented technology. In 1991, he was appointed as a Professor of signal
Theory Group, ETH Zurich, headed by Prof. Helmut Bölcskei. Prior to joining processing with the Royal Institute of Technology (KTH), Stockholm, Sweden.
Ph.D., he worked on Audio Coding algorithms in Sasken Communications, He is the Head of the Department for Signals, Sensors, and Systems, KTH,
Bangalore, India, as a Design Engineer from 2000 to 2001. His research and the Dean of the School of Electrical Engineering, KTH. He is currently
interests include design and optimization of MIMO communication systems, the Director for the Interdisciplinary Centre for Security, Reliability and Trust,
automotive radar and array processing, polynomial signal processing, satellite University of Luxembourg. He was the recipient of the IEEE Signal Process-
communication systems, resource allocation, game theory, and fast algorithms ing Society Technical Achievement Award, the EURASIP Group Technical
for structured matrices. Achievement Award, and the European Research Council advanced research
He is currently on the Executive Committee of the IEEE Benelux joint grant twice. He has co-authored journal papers that was the recipient of the
chapter on communications and vehicular technology and serves as handling IEEE Signal Processing Society Best Paper Award in 1993, 2001, 2006, 2013,
Editor of the Elsevier Signal Processing and a Member of the EURASIP and 2019, and 8 IEEE conference papers best paper awards. He is a Board
Technical Area Committee on Theoretical and Methodological Trends in Signal Member of the IEEE Signal Processing Society, the Swedish Research Council
Processing. He was the co-recipient of the 2014 Distinguished Contributions to and currently serves on the boards of EURASIP and the Swedish Foundation
Satellite Communications Award, from the Satellite and Space Communications for Strategic Research. He was the Editor-in-Chief of the EURASIP Signal
Technical Committee of the IEEE Communications Society. Processing, and served on the Editorial Boards of the IEEE TRANSACTIONS
ON SIGNAL PROCESSING, IEEE SIGNAL PROCESSING MAGAZINE, IEEE OPEN
JOURNAL FOR SIGNAL PROCESSING, EURASIP Journal of Advances in Signal
Processing, and Foundations and Trends in Signal Processing. He is a Fellow
Symeon Chatzinotas (Senior Member, IEEE) re- of EURASIP.
ceived the M.Eng. in telecommunications from Aris-
totle University of Thessaloniki, Greece and the
M.Sc. and Ph.D. in electronic engineering from Uni-
versity of Surrey, UK, in 2003, 2006 and 2009, re-
spectively.
He is currently Full Professor/Chief Scientist I
and Head of the SIGCOM Research Group at SnT,
University of Luxembourg. He is coordinating the re-
search activities on communications and networking,
acting as a PI for more than 20 projects and main
representative for 3GPP, ETSI, DVB. In the past, he has been a Visiting Professor
at the University of Parma, Italy, lecturing on “5G Wireless Networks”. He was
involved in numerous R&D projects for NCSR Demokritos, CERTH Hellas and
CCSR, University of Surrey. He is currently in the editorial board of the IEEE
TRANSACTIONS ON COMMUNICATIONS, IEEE OPEN JOURNAL OF VEHICULAR
TECHNOLOGY and the International Journal of Satellite Communications and
Networking. He was the co-recipient of the 2014 IEEE Distinguished Contribu-
tions to Satellite Communications Award and Best Paper Awards at EURASIP
JWCN, CROWNCOM, ICSSC. He has coauthored more than 450 technical
papers in refereed international journals, conferences and scientific books.