Sparse Array Design for Dual-Function Radar-Communications System

Downloaded from: [Link] 2025-02-18 08:30 UTC

Citation for the original published paper (version of record):

Huang, H., Wu, L., Shankar, B. et al (2023). Sparse Array Design for Dual-Function
Radar-Communications System. IEEE Communications Letters, 27(5): 1412-1416.
[Link]

N.B. When citing this work, cite the original published paper.

© 2023 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained
for all other uses, in any current or future media, including reprinting/republishing this material for
advertising or promotional purposes, or reuse of any copyrighted component of this work in other
works.

(article starts on next page)

1412 IEEE COMMUNICATIONS LETTERS, VOL. 27, NO. 5, MAY 2023

Sparse Array Design for Dual-Function Radar-Communications System

Huiping Huang , Student Member, IEEE, Linlong Wu , Member, IEEE,
Bhavani Shankar , Senior Member, IEEE, and Abdelhak M. Zoubir , Fellow, IEEE

Abstract— The problem of sparse array design for dual- selection strategy by using a learning approach. This method
function radar-communications is investigated. Our goal is to requires a training process, which might be unavailable in
design a sparse array which can simultaneously shape desired some practical applications. The authors in [16] introduced a
beam responses and serve multiple downlink users with the
required signal-to-interference-plus-noise ratio levels. Besides,
realistic waveform constraint in the DFRC system, in order to
we also take into account the limitation of the radiated power improve the power efficiency. A surrogate subproblem instead
by each antenna. The problem is formulated as a quadratically of the original problem was solved, which might lead to a
constrained quadratic program with a joint-sparsity-promoting highly suboptimal solution. In [17] and [18], several types of
regularization, which is NP-hard. The resulting problem is solved DFRC systems were proposed which implement simultaneous
by the consensus alternating direction method of multipliers, beamformers associated with single and different sparse arrays
which enjoys parallel implementation. Numerical simulations
exhibit the effectiveness and superiority of the proposed method with shared aperture. However, these two works consider a
which leads to a more power-efficient solution. single-user case, and the corresponding methods cannot be
applicable to the case of multi-users. A weighted ℓ1,q -norm
Index Terms— Alternating direction method of multipliers,
dual-funtion radar-communications, sparse array design.
optimization based approach was developed in [19]. Note that
all the problems considered in [17], [18], and [19] are convex
I. I NTRODUCTION and were solved by the existing toolbox such as CVX.
Unlike in previous works, we propose a novel system model
D UAL-FUNCTION radar-communications (DFRC) sys-
tems have been recently widely investigated [1], [2],
[3], [4]. They find applications in a wide range of areas,
for DFRC systems, and formulate the corresponding sparse
array design problem as a quadratically constrained quadratic
including vehicular networks, indoor positioning and covert program (QCQP) regularized by a joint-sparsity-promoting
communications [5], [6]. To achieve accurate sensing and term. Besides the control on illumination beampattern and
high throughput, the DFRC systems probably require a large communication signal-to-interference-plus-noise ratio (SINR),
number of antennas [7], [8], [9], [10]. In practice, a base we also take into account the limitation of the radiated
station usually equips with less radio-frequency (RF) chains power by each antenna. We propose an algorithm based
than antennas given the hardware cost consideration. For such on the consensus alternating direction method of multipliers
a configuration, it raises a question on how to adaptively (ADMM) to solve the resulting problem. Note that at each
switch the available RF chains to the corresponding subset of ADMM iteration, the primary variable has a closed-form
antennas [11], [12], [13], which can be interpreted as sparse solution, and the auxiliary variables can be solved efficiently
array design. in a parallel manner. Simulation results show its superior
In this work, we consider the sparse array design for performance compared to other examined methods.
the DFRC system. Similar problems have been studied It is worth mentioning that the sparse array beamforming
in [14], [15], [16], [17], [18], and [19]. The authors can be incorporated into the framework of hybrid beamform-
in [14] derived a Cramér-Rao bound for the cooperative ing, via decomposing the sparse array beamformer into the
radar-communications system, where they focused on target baseband beamformer with a particular selection preference
parameter estimation. The work [15] developed an antenna and the RF beamformer. Therefore, it can further improve the
power efficiency within the hybrid beamforming.
Manuscript received 14 February 2023; accepted 12 March 2023. Date The remainder of the letter is organized as follows. The
of publication 15 March 2023; date of current version 10 May 2023. The system model is established in Section II. The proposed
work of H. Huang was supported by the Graduate School CE within the
Centre for Computational Engineering at Technische Universität Darmstadt. algorithm is presented in Section III. Simulation results are
The work of Linlong Wu and Bhavani Shankar was supported in part by shown in Section IV, while Section V concludes the letter.
FNR CORE SPRINGER under Grant C18/IS/12734677 and in part by ERC Notations: Throughout this letter, bold-faced lower-case
AGNOSTIC under Grant EC/H2020/ERC2016ADG/742648. The associate
editor coordinating the review of this letter and approving it for publication (upper-case) letters denote vectors (matrices). Superscripts ·T ,
was T. Mao. (Corresponding author: Linlong Wu.) ·H , ·∗ , ·−1 denote transpose, Hermitian transpose, conjugate,
Huiping Huang was with the Department of Electrical Engineer- and inverse, respectively. E{·} denotes the expectation
ing and Information Technology, Technische Universität Darmstadt,
64283 Darmstadt, Germany. He is now with the Department of Electrical operator. |·| and ∠· are the modulus and phase, respectively,
Engineering, Chalmers University of Technology, 412 96 Gothenburg, Sweden both in an element-wise manner. C and R are the sets
(e-mail: huiping@[Link]). of complex and real numbers, respectively, and ȷ =
Linlong Wu and Bhavani Shankar are with the Interdisciplinary √
Centre for Security, Reliability and Trust (SNT), University of Luxem-
−1. ∥·∥0 , ∥·∥1 , and ∥·∥2 are ℓ0 -quasi-norm, ℓ1 -norm, and
bourg, 4365 Luxembourg City, Luxembourg (e-mail: [Link]@[Link]; ℓ2 -norm, respectively. ⊗ denotes the Kronecker product. ⊘ is
[Link]@[Link]). the element-wise division operator. IN and OM ×N are the
Abdelhak M. Zoubir is with the Department of Electrical Engineering and N × N identity matrix and M × N zero matrix, respectively.
Information Technology, Technische Universität Darmstadt, 64283 Darmstadt,
Germany (e-mail: zoubir@[Link]). 1 and 0 are the all-ones and all-zeros vector of appropriate
Digital Object Identifier 10.1109/LCOMM.2023.3257739 size, respectively.
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see [Link]
HUANG et al.: SPARSE ARRAY DESIGN FOR DUAL-FUNCTION RADAR-COMMUNICATIONS SYSTEM 1413

frequency-flat quasi-static channel from the transmitter to the

m-th user. The CSI, {hm }m , is assumed to be perfectly
available at the transmitter. The maximum power radiated by
each antenna n is given by Pn . Therefore, we have [7]
M
X
H
wm En wm ≤ Pn , ∀n = 1, 2, . . . , N, (6)
m=1
where En is the N ×N all-zero matrix except the n-th diagonal
entry being 1. The received SINR of user m is defined below,
Fig. 1. Illustration of a DFRC system. and a certain minimum SINR needs to be guaranteed, i.e.,
2
hH
m wm
SINRm ≜ P 2 ≥ γm , ∀m = 1, . . . , M,
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION j̸=m |hH 2
m wj | + σm
(7)
We consider a DFRC system, as indicated in Fig. 1. The P PM
transmitter, consisting of N antennas, emits M signals, each where j̸=m is short notation for j=1,j̸=m .
to a single user. The weight vector wm ∈ CN is designed to Now we consider the scenario where only K ≤ N
transfer the data symbol, sm (t), to user m, ∀m = 1, 2, . . . , M . RF chains are available, and thus only K antennas can be
All the M transmitted signals are also received by a target. simultaneously utilized in the DFRC system. Therefore, wm ,
The total transmitted signal can then be formulated as ∀m, are sparse, and moreover, they share the same sparsity
M
X pattern. To fulfill this requirement, we need:
x(t) = wm sm (t), (1) ∥w∥
e 0 ≤ K, (8)
m=1
T
where t is the time index. We assume that the data symbols, where we ≜ [w e1 , w
e2 , . . . , w
eN ] with its component defined as
sm (t) ∀m = 1, 2, . . . , M , are mutually uncorrelated, and each wen ≜ ∥[w1 (n), w2 (n), . . . , wM (n)]∥2 ∀n = 1, 2, . . . , N , and
sm (t) is zero-mean, spatially white with unit variance, i.e., wm (n) being the n-th entry of vector wm , ∀m = 1, 2, . . . , M .
The quality of service (QoS) problem aims at minimizing
E{|sm (t)|2 } = 1, ∀m = 1, 2, . . . , M, (2a) the total transmit power (TxPower), with several constraints
E{|si (t)s∗j (t)|} = 0, ∀i, j = 1, 2, . . . , M, and i ̸= j. (2b) described above. To this end, the QoS problem is given as
For the radar system, we wish to focus the transmit energy M
X
on a certain angular region, say B, in order to have a higher min TxPower ≜ ∥wm ∥22 (9a)
{wm }
probability of detecting the target(s) lying therein. To this end, m=1

we expect that the array response within B is not less than a s.t. 4), (6), (7), and (8). (9b)
preset threshold, while the response outside B (denoted as B̄) By replacing the non-convex ℓ0 -quasi-norm in (8) with the ℓ1 -
is not higher than a small threshold. Mathematically, norm, and incorporating it into Problem (9) as a regularization
E xH (t)a(θi )aH (θi )x(t) ≥ ϵp , ∀θi ∈ B, term, we have

(3a)
M
E x (t)a(θj )aH (θj )x(t) ≤ ϵs , ∀θj ∈ B̄,
 H
(3b) X
min ∥wm ∥22 + η∥w∥e 1 s.t. (4), (6), and (7), (10)
where a(θ) denotes the steering vector of θ, ϵp and ϵs are two {wm }
m=1
preset thresholds corresponding to the mainlobe and sidelobe, where η is a parameter controlling the sparsity of the solution.
respectively. Substituting (1) and (2) into (3), we have
M
X
H
III. P ROPOSED M ETHOD
wm a(θi )aH (θi )wm ≥ ϵp , ∀θi ∈ B, (4a)
m=1
The difficulties of solving Problem (10) are twofold. First,
M the discorporated sparsity has a joint structure among all
X
H
wm a(θj )aH (θj )wm ≤ ϵs , ∀θj ∈ B̄. (4b) wm [20], which is different from the classical sparsity
m=1 and consequently, the existing solutions cannot be directly
For the downlink communication systems, the received data extended. Second, the constraints (4a) and (7) are nonconvex.
by the m-th user is given as In what follows, we first deal with the joint-sparsity
M
structure of the solution by reformulating the problem, and
ym (t) = hH wm sm (t) +
X
hH then, we develop an algorithm based on the consensus ADMM.
m wj sj (t) + nm (t),
| m {z } | {z }
j=1,j̸=m
communication signal noise
| {z }
interference signals
A. Problem Reformulation
(5) First of all, we define a long column vector w ∈ CM N
as w ≜ [w1T , w2T , . . . , wM
T T
] and define M matrices Φm ∈
where nm (t) is the additive white Gaussian noise with mean R N ×M N
as Φm ≜ [ON ×(m−1)N , IN , ON ×(M −m)N ]. Then,
2
zero and known variance σm , ∀m = 1, 2, . . . , M . Besides, Problem (10) can be reformulated as
hm ∈ CN denotes the channel state information (CSI) vector,
which models the propagation loss and phase shift of the min ∥w∥22 + η∥w∥2,1 (11a)
w
1414 IEEE COMMUNICATIONS LETTERS, VOL. 27, NO. 5, MAY 2023

s.t. wH Aθi w ≥ ϵp , ∀θi ∈ B, (11b) In what follows, we show how to solve w and vl from (15a)
H
w A θ j w ≤ ϵs , ∀θj ∈ B̄, (11c) and (15b), respectively. We start by solving w from (15a). It is
H straightforward to see that, by calculating the derivative of the
w Bn w ≤ Pn , ∀n = 1, 2, . . . , N, (11d) objective function of (15a) with respect to (w.r.t.) w and setting
wH Cm w it to 0, we obtain the solution to (15a) as
≥ γm , ∀m = 1, 2, . . . , M, (11e)
wH Cm̄ w + σm
2 L
ρ X
where we define Aθi ≜
PM H H w
b = (vl + ul ). (16)
m=1 ΦP m a(θi )a (θi )Φm , 2 + ρL
PM M l=1
Aθj ≜ m=1 Φm a(θj )a (θj )Φm , Bn ≜ m=1 ΦH
H H
m En Φm ,
H H
P H H
On the other hand, to solve vl from (15b), we firstly consider
Cm ≜ Φm hm hm Φm , Cm̄ ≜ j̸=m Φj hm hm Φj , and the unconstrained minimization problem:
PM 2
we have utilized m=1 ∥w ∥
m 2 = ∥w∥22 , wm = Φm w, η ρ
PN q PM min f (vl ) ≜ ∥vl ∥2,1 + ∥vl − w + ul ∥22 . (17)
and ∥w∥2,1 ≜ n=1 2
m=1 |wm (n)| . Further, since the
vl L 2
H 2
denominators, w Cm̄ w + σm , are always positive, (11e) can The derivative of f (vl ) w.r.t. vl is calculated as
hη i
be written as ∇vl f (vl ) = (IM ⊗ G) + ρIM N vl − ρ(w − ul ),
wH Dm w ≥ γm σm
2
, ∀m = 1, 2, . . . , M, (12) L
where G ∈ RN ×N is a diagonal matrix with diagonal being
where Dm ≜ Cm − γm Cm̄ , ∀m = 1, 2, . . . , M . We observe v
u M
that (11b), (11c), (11d), and (12) take the form: wH Fw ≤ f . uX
Therefore, Problem (11) can be reformulated as 1⊘ t  |v (1 + (m − 1)N )|2 ,
l
m=1
min ∥w∥22 + η∥w∥2,1 (13a) T
w
v
u M
s.t. wH Fl w ≤ fl , ∀l = 1, 2, . . . , L, (13b)
uX
...,t |vl (N + (m − 1)N )|2  .
where the subscript ·l is used to indicate the l-th constraint m=1
in Problem (11), and L is the total number of constraints. By setting the derivative to 0, we obtain
It is worth noting that Fl corresponding to −Aθi is hη i
negative semidefinite, and Fl corresponding to −Dm could be (IM ⊗ G) + ρIM N vl = ρ(w − ul ). (18)
indefinite. Therefore, in general, Problem (13) is non-convex η L 
and thus NP-hard [21]. To solve this problem, we propose an
Since L (IM ⊗ G) + ρIM N is real-valued, ∠vl =
∠(w − ul ). Thus, we only need to calculate the modulus of
algorithm based on the consensus ADMM, which is capable
vl , using
of handling all the constraints in parallel. hη i
(IM ⊗ G) + ρIM N |vl | = ρ|w − ul |. (19)
L
By exploring the structure of IM ⊗ G, we observe that there
B. Proposed Algorithm are N blocks each containing M equal entries. Extracting the
rows of the equal entries yields
Firstly, we formulate Problem (13) by introducing L 
η

auxiliary variables {vl ∈ CM N }L l=1 , and settle the original
+ ρ |vl(n) | = ρ|cl(n) |, (20)
L∥vl(n) ∥2
variable w and the auxiliary variables {vl }l in a separable
fashion, as ∀n = 1, 2, . . . , N , where vl(n) ≜ [vl (n), vl (n + N ),
L . . . , vl (n + (M − 1)N )]T ∈ CM , and cl(n) ∈ CM contains
ηX
min ∥w∥22 + ∥vl ∥2,1 (14a) the corresponding entries of (w − ul ). From (20), we have
w,{vl } L i=1
ρL∥vl(n) ∥2
s.t. vlH Fl vl ≤ fl , vl = w, ∀l = 1, 2, . . . , L. (14b) |vl(n) | = |cl(n) |. (21)
η + ρL∥vl(n) ∥2
Next, we form the scaled-form augmented Lagrangian Performing the element-wise square operation, we have
function related
PL to the aboveP problem, as: L(w, {vl }, {ul }) = ρ2 L2 ∥vl(n) ∥22
L
∥w∥22 + Lη l=1 ∥vl ∥2,1 + ρ2 l=1 ∥vl − w + ul ∥22 − ∥ul ∥22 , |vl(n) |2 = 2 |cl(n) |2 .


η + ρ2 L2 ∥vl(n) ∥22 + 2ηρL∥vl(n) ∥2
where ρ > 0 stands for the augmented Lagrangian parameter,
and ul is the scaled dual variable corresponding to the equality (22)
constraint vl = w in Problem (14). Hence, we further have
Finally, the consensus ADMM updating equations can be
written down as ∥vl(n) ∥22 = 1T |vl(n) |2 (23a)
L ρ2 L2 ∥vl(n) ∥22
2 ρX = 2 1T |cl(n) |2 .
w ← arg min ∥w∥2 + ∥vl − w + ul ∥22 , (15a) η + ρ2 L2 ∥vl(n) ∥22 + 2ηρL∥vl(n) ∥2
w 2
l=1
( η ρ (23b)
arg min ∥vl ∥2,1 + ∥vl − w + ul ∥22
vl ← vl L 2 (15b) The above equation leads to
s.t. vlH Fl vl ≤ fl , ρ2 L2 ∥vl(n) ∥22 + 2ηρL∥vl(n) ∥2 + η 2 − ρ2 L2 1T |cl(n) |2 = 0,
ul ← ul + vl − w. (15c) (24)
HUANG et al.: SPARSE ARRAY DESIGN FOR DUAL-FUNCTION RADAR-COMMUNICATIONS SYSTEM 1415

the left-hand side of which is a simple quadratic function w.r.t. Algorithm 1 Consensus ADMM for Solving Problem (13)
∥vl(n) ∥2 , and its unique1 root is given as Input: η, ρ, kmax , Fl ∈ CM N ×M N , fl , ∀l = 1, 2, . . . , L
q Output: w b ∈ CM N
ρL 1T |cl(n) |2 − η bl(0) ← vl(init) , u
Initialize: v b (0)
l ← u(init)
l ,k←0
∥vl(n) ∥2 = . (25) 1: while k < kmax do
ρL XL  
Then, by substituting (25) into (21), we obtain 2: b (k+1) ← 2+ρL
w ρ
bl(k) + u
v b (k)
l
l=1
q
ρL 1T |cl(n) |2 − η 3: for each l = 1, 2, . . . , L do
|vl(n) | = q |cl(n) |. (26) 4: cl ← w b (k)
b (k+1) − u l
(k+1)
ρL 1T |cl(n) |2 5: ∠v̄l ← ∠cl
6: for each n = 1, 2,√ . . . , N do

In Parallel
In (26), we define |vl(n) | = 0 if |cl(n) | = 0. The solution for (k+1) ρL 1T |cl(n) |2 −η
7: |v̄l(n) | ← √ T |cl(n) |
(18), referred to as v̄l , is finally obtained by combining vl(n) ρL 2
1 |cl(n) |

In Parallel
(k+1)
(which can be calculated as |vl(n) |eȷ∠vl(n) ). Then, the solution 8:
(k+1) (k+1)
v̄l(n) ← |v̄l(n) |e l(n)
ȷ∠v̄

to (15b), denoted by v bl , is found via the following theorem. 9: end for

(k+1) (k+1)
Theorem: If ρ satisfies ρ2 ≫ Lη , then v bl can be solved via: 10: Construct v̄l using v̄l(n) , ∀n
(k+1)H (k+1)
bl ← arg min ∥vl − v̄l ∥22
v s.t. vlH Fl vl ≤ fl . (27) 11: if v̄l Fl v̄l( ≤ fl then v bl(k+1) ← v̄l(k+1)
(k+1) 2
vl arg min ∥vl − v̄l ∥2
12: else v bl(k+1) ← vl
Proof: See Appendix A. □ H
s.t. vl Fl vl ≤ fl
Remark 1: Note that η is related to the sparsity of the 13: end if
(k+1) (k) (k+1)
solution of Problem (10), and L is the total number of 14: u
bl ←ubl + v bl −w b (k+1)
constraints in Problem (13). Both of them are known for a 15: end for
specific problem. Hence, it is easy to choose a ρ such that 16: k ←k+1
ρ η 17: end while
2 ≫ L. 18: b ←w
w b (k)
Remark 2: If v̄l satisfies the constraint of (27), i.e.,
v̄lH Fl v̄l ≤ fl , it is easy to have v
bl = v̄l . Otherwise, notice
that Problem (27) is a QCQP with one constraint, which can
be solved optimally despite that Fl may be indefinite [21].
So far, we have presented how to solve w and vl from
(15a) and (15b), respectively. Note that {vl }L L
l=1 and {ul }l=1
can be calculated in parallel. The complete consensus ADMM
for solving Problem (13) is summarized in Algorithm 1,
in which kmax is used to terminate the iteration, and the
superscript ·(k) denotes the corresponding variable at the k-th
iteration.
Fig. 2. Beampattern comparison.

IV. S IMULATION R ESULTS tuning parameter is2 η = 0.1, and the augmented Lagrangian
In this section, we evaluate the performance of the proposed parameter is ρ = 50 (ρ/2 ≫ η/L is satisfied). The value of
algorithm compared with the feasible point pursuit successive vl(init) is given by any feasible point, while u(init)
l = 0 and
convex approximation (FPP-SCA) method [22]. FPP-SCA was kmax = 100. The beampatterns are drawn in Fig. 2, which
proposed for general QCQP and it is adapted to our problem. indicates that both the proposed and FPP-SCA methods have
Note that unlike our solution, no closed-form solution of FPP- beamlobes in the radar mainlobe B and the users directions.
SCA is given. Two metrics are adopted: the TxPower defined In addition, the proposed method has much higher response
in (9a) and the radar-side mainlobe-to-sidelobe response ratio within B than the FPP-SCA method.
2
m=1 |wm a(θi )|
P PM H
θi ∈B Next, we examine the TxPower and MSRR versus the
(MSRR) defined as MSRR = P PM
|w a(θ )|2
H .
θj ∈B̄ m=1 m j number of selected sensors, i.e., K. We also consider a strategy
We first consider a transmit system with a uniform linear of randomly selecting K sensors. The parameters are the
array of N = 10 antennas and K = 8 or 10 RF chains. There same as those in the last example. 500 Monte-Carlo trials are
are M = 2 users located at −45◦ and 45◦ . The mainlobe performed. The results are plotted in Fig. 3. It is seen that
and sidelobe are B = [−5◦ , 5◦ ] and B̄ = [−90◦ , −60◦ ] ∪ our proposed method has the lowest transmit power and the
[−30◦ , −20◦ ] ∪ [20◦ , 30◦ ] ∪ [60◦ , 90◦ ]. The thresholds for the highest MSRR, among all tested methods.
mainlobe and sidelobe response are ϵp = 10 and ϵs = 0.5. The Finally, we test the TxPower and MSRR versus the number
maximum power radiated by each antenna is Pn = 40 dBm, of users, i.e., M . The number of selected antennas is fixed as
2
∀n, the same as [7]. The noise variance is set to σm = 1, ∀m. K = 8, and the other parameters are unchanged as in the last
The threshold for the received SINR is γm = 10 dB, ∀m.
The number of constraints in Problem (13) is L = 38, the 2 We do not study the relationship between η and the sparsity of the solution,
because of the space limitation of the letter. Instead, when we obtain w, b
1 Note that the quadratic function in (24) has two roots, one positive and one we choose K antennas corresponding to the largest (in an ℓ2,1 -norm sense)
negative. In our case, the negative one is omitted, since its root ∥vl(n) ∥2 ≥ 0. K components. We have good results when the tuning parameter is η = 0.1.
1416 IEEE COMMUNICATIONS LETTERS, VOL. 27, NO. 5, MAY 2023

Suppose that there exists a point v̆l ∈ {vl : vlH Fl vl ≤ fl },

such that f (v̆l ) < f (evl ). Thus, by using (29), we obtain that
ρ 2 ρ 2
2 ∥v̆ l − v̄ l ∥2 < 2 ∥e
v l − v̄l ∥2 , which contradicts (28). This
implies that f (e vl ) ≤ f (vl ) holds for all feasible vl , that is,
v
el is the solution to Problem (15b). This completes the proof.

R EFERENCES
[1] K. V. Mishra et al., “Toward millimeter-wave joint radar communica-
tions: A signal processing perspective,” IEEE Signal Process. Mag.,
vol. 36, no. 5, pp. 100–114, Sep. 2019.
Fig. 3. TxPower (left) and MSRR (right) versus K. [2] F. Liu et al., “Joint radar and communication design: Applications, state-
of-the-art, and the road ahead,” IEEE Trans. Commun., vol. 68, no. 6,
pp. 3834–3862, Jun. 2020.
[3] S. Shi et al., “Distributed 5G NR-based integrated sensing and
communication systems: Frame structure and performance analysis,” in
Proc. 30th Eur. Signal Process. Conf. (EUSIPCO), Belgrade, Serbia,
Aug. 2022, pp. 1062–1066.
[4] Z. Cheng et al., “Hybrid beamforming in mmWave dual-function radar-
communication systems: Models, technologies, and challenges,” 2022,
arXiv:2209.04656.
[5] C. Yang and H.-R. Shao, “WiFi-based indoor positioning,” IEEE
Commun. Mag., vol. 53, no. 3, pp. 150–157, Mar. 2015.
[6] H. Wymeersch et al., “5G mmWave positioning for vehicular networks,”
Fig. 4. TxPower (left) and MSRR (right) versus M . IEEE Wireless Commun., vol. 24, no. 6, pp. 80–86, Dec. 2017.
[7] E. Chen and M. Tao, “ADMM-based fast algorithm for multi-group
multicast beamforming in large-scale wireless systems,” IEEE Trans.
example. The results are depicted in Fig. 4, which again show Commun., vol. 65, no. 6, pp. 2685–2698, Jun. 2017.
that our algorithm leads to a more power-efficient solution. [8] A. M. Elbir et al., “A family of deep learning architectures for
channel estimation and hybrid beamforming in multi-carrier mm-wave
massive MIMO,” IEEE Trans. Cognit. Commun. Netw., vol. 8, no. 2,
V. C ONCLUSION pp. 642–656, Jun. 2022.
[9] E. Raei et al., “MIMO radar transmit beampattern shaping for spectrally
We studied the problem of sparse array design for dense environments,” IEEE Trans. Aerosp. Electron. Syst., early access,
dual-function radar-communications. Our design aimed at Aug. 2022, doi: 10.1109/TAES.2022.3196333.
maintaining good control in both mainlobe and sidelobe, and [10] A. Kassaw et al., “Fractional programming for energy efficient power
control in uplink massive MIMO systems,” in Proc. 27th Eur. Signal
also to keep the signal-to-interference-plus-noise ratio for the Process. Conf. (EUSIPCO), A Coruña, Spain, Sep. 2019, pp. 1–5.
users larger than a certain level. In addition, we considered the [11] S. A. Hamza and M. G. Amin, “Sparse array beamforming design for
limitation of the radiated power by each antenna. The problem wideband signal models,” IEEE Trans. Aerosp. Electron. Syst., vol. 57,
no. 2, pp. 1211–1226, Apr. 2021.
was formulated as a quadratically constrained quadratic [12] T. Wei et al., “Sparse array beampattern synthesis via majorization-based
program, and solved by the consensus alternating direction ADMM,” in Proc. IEEE 94th Veh. Technol. Conf. (VTC-Fall), Norman,
method of multipliers. The proposed algorithm was able to CA, USA, Sep. 2021, pp. 1–5.
be implemented in parallel. Simulation results demonstrated [13] H. Huang et al., “Sparse array beamformer design via ADMM,” 2022,
arXiv:2208.12313.
better performance of the proposed algorithm than the other [14] L. Wang et al., “Transmitter selection and receiver placement for target
tested methods. parameter estimation in cooperative radar-communications system,” IET
Signal Process., vol. 16, no. 7, pp. 776–787, Sep. 2022.
[15] Z. Xu et al., “Cramér-rao bound and antenna selection optimization
A PPENDIX A for dual radar-communication design,” in Proc. IEEE Int. Conf.
P ROOF OF T HEOREM Acoust., Speech Signal Process. (ICASSP), Singapore, May 2022,
pp. 5168–5172.
Solving (27) is equal to finding a point in {vl : [16] X. Zhang et al., “Joint antenna selection and waveform design for
vlH Fl vl ≤ fl }, such that it is closest (in an ℓ2 -norm sense) coexistence of MIMO radar and communications,” EURASIP J. Adv.
to v̄l . Hence, the theorem equivalently states that the solution Signal Process., vol. 2022, no. 1, pp. 1–27, Dec. 2022.
[17] X. Wang et al., “Sparse transmit array design for dual-function radar
to Problem (15b) is the point closest (in an ℓ2 -norm sense) to communications by antenna selection,” Digit. Signal Process., vol. 83,
v̄l , provided that ρ/2 ≫ η/L. To show this, we denote v el as pp. 223–234, Dec. 2018.
the point in {vl : vlH Fl vl ≤ fl }, such that [18] X. Wang et al., “Dual-function MIMO radar communications system
design via sparse array optimization,” IEEE Trans. Aerosp. Electron.
∥e
vl − v̄l ∥2 ≤ ∥vl − v̄l ∥2 (28) Syst., vol. 55, no. 3, pp. 1213–1226, Jun. 2019.
[19] A. Ahmed et al., “Antenna selection strategy for transmit beamforming-
holds for any vl ∈ {vl : vlH Fl vl ≤ fl }. Our goal is to show based joint radar-communication system,” Digit. Signal Process.,
vl ) ≤ f (vl ), for any vl ∈ {vl : vlH Fl vl ≤ fl }.
f (e vol. 105, Oct. 2020, Art. no. 102768.
The Lagrangian parameter ρ is chosen as ρ = Cη/L, where [20] H. Huang et al., “Off-grid direction-of-arrival estimation using second-
order Taylor approximation,” Signal Process., vol. 196, Jul. 2022,
C is a constant. Then, |g(vl ) − f (vl )| → 0 as C → ∞ (i.e., Art. no. 108513.
ρ/2 ≫ η/L), where g(vl ) ≜ ρ2 ∥vl − w + ul ∥22 . Moreover, [21] K. Huang and N. D. Sidiropoulos, “Consensus-ADMM for general
ρ quadratically constrained quadratic programming,” IEEE Trans. Signal
f (vl ) = g(vl ) = ∥vl − v̄l ∥22 , (29) Process., vol. 64, no. 20, pp. 5297–5310, Oct. 2016.
2 [22] O. Mehanna et al., “Feasible point pursuit and successive approximation
as long as C → ∞. Note that, in the second equality above, of non-convex QCQPs,” IEEE Signal Process. Lett., vol. 22, no. 7,
we used the fact that v̄l = w − ul as C → ∞. pp. 804–808, Jul. 2015.