This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.3004739, IEEE
Transactions on Signal Processing
1

Joint Transmit Beamforming for Multiuser MIMO

Communications and Radar
Xiang Liu, Tianyao Huang, Nir Shlezinger, Yimin Liu, Jie Zhou, and Yonina C. Eldar

Abstract—Future wireless communication systems are expected throughput of wireless communications operating in neigh-
to explore spectral bands typically used by radar systems, in or- bouring bands. To tackle this congestion, it has been recently
der to overcome spectrum congestion of traditional communica- proposed to allow wireless communications to share spectrum
tion bands. Since in many applications radar and communication
share the same platform, spectrum sharing can be facilitated by with radar systems, allowing both functionalities to simulta-
joint design as dual function radar-communications system. In neously operate over the same wide frequency bands [2]–[5].
this paper, we propose a joint transmit beamforming model for The common strategy to allow individual radar and commu-
a dual-function multiple-input-multiple-output (MIMO) radar nication systems to share spectrum with controllable mutual
and multiuser MIMO communication transmitter sharing the interference is to facilitate co-existence by some level of
spectrum and an antenna array. The proposed dual-function sys-
tem transmits the weighted sum of independent radar waveform cooperation [6]–[18]. These techniques include opportunistic
and communication symbols, forming multiple beams towards spectrum access [6], [7], transmit interference nulling [8],
the radar targets and the communication receivers, respectively. [9], adaptive receive interference cancellation [10]–[13] and
The design of the weighting coefficients is formulated as an optimization based beamforming design [14]–[18] to mitigate
optimization problem whose objective is the performance of the mutual interference. This approach typically requires the
the MIMO radar transmit beamforming, while guaranteeing
that the signal-to-interference-plus-noise ratio (SINR) at each individual radar and communication systems to either be
communication user is higher than a given threshold. Despite configured using some centralized entity, or alternatively, to
the non-convexity of the proposed optimization problem, it can exchange information, such as knowledge of the interference
be relaxed into a convex one, which can be solved in polynomial channel and radar waveform parameters, significantly increas-
time, and we prove that the relaxation is tight. Then, we propose ing the complexity of realizing such systems [19].
a reduced complexity design based on zero-forcing the inter-
user interference and radar interference. Unlike previous works, The difficulty associated with coordinating spectrum sharing
which focused on the transmission of communication symbols to radar and communication systems is notably reduced when
synthesize a radar transmit beam pattern, our method provides these functionalities operate on the same device. In fact, vari-
more degrees of freedom for MIMO radar and is thus able to ous emerging technologies, such as automotive vehicles [20],
obtain improved radar performance, as demonstrated in our sim- implement both radar sensing and data transmission from the
ulation study. Furthermore, the proposed dual-function scheme
approaches the radar performance of the radar-only scheme, same platform, as illustrated in Fig. 1. In such cases, spectrum
i.e., without spectrum sharing, under reasonable communication sharing can be realized by jointly designing a dual-function
quality constraints. radar-communications (DFRC) system [19]–[42]. One clear
Index Terms—Spectrum sharing, dual-function radar commu- advantage of DFRC methods over individual co-existing sys-
nication, MIMO radar, multiuser MIMO, transmit beamforming tems is that the functionalities share radio frequency (RF)
front-end and aperture, thus reducing the cost and weight
of hardware [43]. Moreover, radar and communication are
I. I NTRODUCTION naturally combined in a DFRC system, and no additional cost
is required for cooperation. Nonetheless, DFRC design has
The increasing demands on wireless communications net-
several associated challenges. From a hardware perspective,
works give rise to a growing need for spectrum sharing be-
the requirements of radar and communications may be quite
tween radar and communication systems. Nowadays, military
distinct in terms of, e.g., power amplifiers operation mode [19],
radars utilize numerous spectrum bands below 10 GHz, like
[24]. From the algorithmic side, properly combining radar and
S-band (2-4 GHz) and C-band (4-8 GHz), while spectrum
communications is a challenging task, and a broad range of
congestion is becoming a serious problem which limits the
strategies for doing so have been proposed in the literature,
This work was supported by the National Natural Science Foundation of see, e.g., the detailed survey in [20].
China under Grant 61801258 and 61571260, Futurewei Technologies, and the Early works on DFRC systems consider single-antenna
Air Force Office of Scientific Research under grant No. FA9550-18-1-0208.
Part of the work [1] were presented at 2019 IEEE International Conference devices. One way to implement such spectrum-sharing dual
on Signal, Information and Data Processing. X. Liu, T. Huang and Y. Liu function signaling is by utilizing orthogonal individual signals
are with the Department of Electronic Engineering, Tsinghua University, for radar and communications, as proposed in [21] which
Beijing, China (e-mail: [email protected], {huangtianyao;
yiminliu}@tsinghua.edu.cn). J. Zhou is with the Institute of Electronic Engi- studied time-division based DFRC systems. Alternatively, one
neering, China Academy of Engineering Physics, Mianyang, China (e-mail: can achieve both functions simultaneously by employing an
zhoujie [email protected]). N. Shlezinger and Y. C. Eldar are with the Faculty of appropriate integrated waveform, which can be utilized for
Mathematics and Computer Science, Weizmann Institute of Science, Rehovot,
Israel (e-mail: {nir.shlezinger; yonina}@weizmann.ac.il). both target detection and information transmission. For in-
stance, the probing capabilities of orthogonal frequency divi-

1053-587X (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Exeter. Downloaded on June 25,2020 at 20:01:43 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.3004739, IEEE
Transactions on Signal Processing
2

pulse, usually yielding low information rate of the same order

of radar pulse repetition frequency [19].
The second approach for implementing MIMO DFRC sys-
tems is based on transmit beamforming. Here, the spatial
degrees of freedom is exploited to synthesize multiple beams
towards several communication users and radar targets. As
opposed to information embedding strategies, transmit beam-
forming enables each function to use its individual waveform,
potentially supporting higher data rates and guaranteed radar
performance by utilizing conventional dedicated signals for
each functionality [19]. In this approach, the main design goal
is to properly beamform both the radar and communication
signals such that each can operate reliably.
In [40], [41], the array probing signal is designed to syn-
thesize radar and communication waveform towards different
directions. The method in [40], [41] considers waveform
synthesis at the main beam direction but cannot suppress the
Fig. 1. A dual function system in which communication and radar share the
transmit platform.
azimuth side-lobe of the radar transmit beam pattern. The
work [19] extended the method of [40], [41] to designing
sion multiplexing (OFDM) waveforms, which are widely used the array probing signal to match the radar transmit beam
for communication signaling, were studied in [22]–[26]. The pattern and minimize the interference power at multiple users.
combination of linear frequency modulation (LFM), which is a However, the methods in [19], [40], [41] only minimize
traditional radar waveform, with continuous phase modulation the interference power, and do not consider the signal-to-
(CPM) to realize a dual-function signal capable of conveying interference-plus-noise ratio (SINR) at each user, which is the
information was studied [27]–[30]. However, these schemes quantity dictating the communications rate. In [42], the authors
inherently result in performance loss for either the radar or studied transmit beamforming in DFRC systems with multiple
the communication [19]. For instance, LFM-CPM usually receivers, i.e., multiuser setup, in which the communication
exhibits higher side-lobes than standard LFM [27]. Moreover, waveform is utilized as a radar transmit waveform. In such a
a common problem emerging in these single-antenna schemes dual-function system, the available degrees of freedom (DoF)
is that radar systems with integrated waveforms usually form for the MIMO radar waveform, which affects the resulting
a single directional beam, which illuminates the radar target radar beam pattern, is equal to the number of communication
inside the beam. Therefore, single-antenna schemes are not users. Since the DoF of conventional MIMO radar, i.e., without
able to illuminate multiple targets and communicate with communication functionality, is at most the number of transmit
multiple users simultaneously as in Fig. 1. That leads to antennas, the resulting MIMO radar cannot utilize its full
notable degradation in signal-to-noise-ratio (SNR) when the DoF when the number of users is smaller than the number of
communication receivers are not physically located within the antennas, potentially leading to significant distortion of radar
radar main lobe. beam pattern. Furthermore, the resulting problem is a non-
Recognizing this limitation of single-antenna schemes, re- convex optimization problem, which is solved by sub-optimal
cent works on DFRC methods employ multiple-input-multiple- methods.
output (MIMO) systems, which provide higher degrees of In this paper, we design a transmit beamforming based
spatial freedom, and can simultaneously synthesize multiple MIMO DFRC system. As was done in [42], we design our
beams towards several communication users and radar targets. transmit beamforming to optimize both the radar transmit
These studies can be divided into two categories: information beam pattern and the SINR at the communication users. Unlike
embedding [31]–[37], [39] and transmit beamforming [19], [42], our proposed joint transmitter utilizes jointly precoded
[40]–[42]. In information embedding systems, radar is typi- individual communication and radar waveforms, allowing to
cally considered as the primary function, and the communi- extend the MIMO radar waveform DoF to its maximal value,
cation message is encoded into the MIMO radar waveform. i.e., the number of antennas. In fact, the previously proposed
For example, the works [34], [35] proposed to embed com- formulation of [42] can be regarded as a special case of the
munication bits by controlling the amplitude and phase of proposed one by nullifying the dedicated radar waveform.
radar spatial side-lobes. The works in [31]–[33] proposed to Furthermore, while we utilize individual signals for radar
convey the message in the form of index modulation, via the and communications, we exploit the fact that both signals
selection of active radar transmit antennas and the allocation of are transmitted from the same device, which also accom-
radar waveforms across active antennas. Embedding data bits modates the radar receiver. Consequently, the radar receiver
in the parameters of a radar waveform, e.g., phase, antenna has complete knowledge of the transmitted communication
index, and frequency, can yield communications in the form waveform, which is utilized for target detection in addition
of phase modulation [36], spatial modulation [37], and carrier to the dedicated radar signal. This approach contributes to the
frequency modulation [39]. However, such methods carry a power-efficiency of the DFRC system, further exploiting the
very limited number of communication symbols per radar inherent advantages of joint design over co-existing separate

1053-587X (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Exeter. Downloaded on June 25,2020 at 20:01:43 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.3004739, IEEE
Transactions on Signal Processing
3

radar and communication systems. Communication

Precoder
We formulate the design of the resulting precoding method Symbol
as a non-convex optimization problem, which we tackle using Shared
TX Array
two different methods: First, we show that it can be relaxed
into an equivalent semidefinite problem, where the latter can Radar waveform Precoder
be solved using conventional optimization tools, and prove that
the relaxation is tight. To circumvent the computational burden (a)
of recovering the optimal precoders from the relaxed formula- Communication
Precoder Shared TX Array
tion, we propose an additional design approach based on zero- Symbol
forcing the interference. Our numerical results demonstrate (b)
that, due to the increased DoF of the MIMO radar waveform,
our approach obtains improved radar transmit beam pattern Fig. 2. (a) The joint transmitter jointly precodes both communication
symbols and radar waveform; (b) The joint transmitter only utilizes precoded
compared to [42], under the same SINR constraints at the communication symbols.
communication users. Furthermore, we demonstrate that under
high SINR constraints our reduced complexity zero-forcing including 𝐾 parallel communication symbol streams to be
technique is capable of achieving comparable performance to communicated to 𝐾 users, respectively, while the 𝑀 ×𝐾 matrix
that of the optimal beamforming scheme, whose compuataion 𝑾𝑐 is the communication precoder. To achieve alias-free signal
is substantially more complex, indicating the potential of sampling and symbol transmission, the communication symbol
our approach in designing reliable DFRC beamforming at duration or the radar code duration, denoted by 𝑡 𝑠 , should
controllable complexity. satisfy 𝑡 𝑠 ≥ 1/(2𝐵) [44], where 𝐵 is the baseband bandwidth
The rest of this paper is organized as follows. Section II of the transmit platform. The maximal available radar delay
proposes the system model. Section III introduces the perfor- resolution is bounded by the symbol duration, and is given
mance metrics of MIMO radar and multiuser MIMO commu- by 1/(2𝐵). The maximal available symbol rate is 2𝐵. We
nication, respectively. Section IV establishes the optimization note that the scheme proposed in [42], which beamformed
model for joint beamforming, and proposes two methods for the communications symbols to be utilized for probing, can
designing precoders based on that formulation. Simulation be regarded as a special case of our system by letting the radar
results are presented in Section V. Finally, Section VI provides waveform be zero, namely transmitting only communication
concluding remarks. symbols, as depicted in Fig. 2(b).
Notations: In this paper, (·) 𝐻 , (·) 𝑐 and (·)𝑇 denote Hermi- Our goal is to design the matrices 𝑊𝑐 and 𝑊𝑟 in (1). Without
tian transpose, conjugate and transpose, respectively. Vectors loss of generality, we make the following assumptions: 1) Both
are denoted by bold lower class letters and matrices are radar and communication signals are zero-mean, temporally-
denoted by bold upper class letters. For a matrix 𝑨, the white and wide-sense stationary stochastic process; 2) The
(𝑖, 𝑗)-th elements of 𝑨 is denoted by [ 𝑨] 𝑖, 𝑗 , and [ 𝑨] 1: 𝑗 communication symbols are uncorrelated with radar wave-
denotes the sub-matrix containing the first 𝑗 columns of 𝑨. form, i.e.,
 
We let 𝑰 𝑛 and 0𝑚×𝑛 denote 𝑛-dimensional identity matrix and
E 𝒔[𝑛]𝒄 𝐻 [𝑛] = E (𝒔[𝑛]) E(𝒄 𝐻 [𝑛]) = 0 𝑀 ×𝐾 ; (2)
𝑚 × 𝑛 zero matrix, respectively. We use E(·) for the stochastic
expectation. For an integer 𝑛 > 0, the set consisting of all 𝑛- 3) Communication symbols intended to different users are
dimensional complex positive semidefinite matrices is denoted uncorrelated, namely,
by S𝑛+ .  
E 𝒄[𝑛]𝒄 𝐻 [𝑛] = 𝑰 𝐾 ; (3)
II. S IGNAL M ODEL 4) The individual radar waveform are generated by pseudo
Consider an antenna array shared by a colocated monostatic random coding [45]–[49], and thus are uncorrelated with each
MIMO radar system and a multiuser MIMO communication other, resulting in
transmitter as depicted in Fig. 1. In our work, both func-  
tionalities operate simultaneously by joint beamforming. The E 𝒔[𝑛] 𝒔 𝐻 [𝑛] = 𝑰 𝑀 . (4)
system diagram of our joint beamforming transmitter is shown Here, both signals are normalized to have unit power, and their
in Fig. 2(a), demonstrating that the transmited signal is a real power is encapsulated in their corresponding precoders 𝑾𝑟
weighted sum of communication symbols and radar waveform. and 𝑾𝑐 .
We consider an antenna array of 𝑀 elements, and let the In order to implement joint transmit beamforming, the
discrete-time transmit signal of this array at time index 𝑛 be precoders 𝑾𝑐 and 𝑾𝑟 are to be jointly designed in consid-
given by eration of the system performance. The performance metrics
𝒙 [𝑛] = 𝑾𝑟 𝒔[𝑛] + 𝑾𝑐 𝒄[𝑛], 𝑛 = 0, . . . , 𝑁 − 1. (1) of MIMO radar and multiuser MIMO communication are
detailed in Section III. In practice, the precoders should satisfy
Here, the 𝑀 × 1 vector 𝒔[𝑛] = [𝑠1 [𝑛], . . . , 𝑠 𝑀 (𝑛)] 𝑇 includes some constraints representing the transmit hardware. Here,
𝑀 individual radar waveforms, and the 𝑀 × 𝑀 matrix 𝑾𝑟 we require that the transmit waveform satisfies a per-antenna
is the beamforming matrix (or precoder) for radar waveform. power constraint, namely, that the transmit power of each
Similarly, 𝒄[𝑛] = [𝑐 1 [𝑛], . . . , 𝑐 𝐾 [𝑛]] 𝑇 is a 𝐾 × 1 vector antenna is identical. The per-antenna power constraint settles

1053-587X (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Exeter. Downloaded on June 25,2020 at 20:01:43 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.3004739, IEEE
Transactions on Signal Processing
4

with the common practice that radar waveforms should be of sight (LoS), the baseband signal at direction 𝜃 can be
transmitted with their maximal available power [50], and has expressed as
also been applied in multi-antenna communication systems 𝑦[𝑛; 𝜃] = 𝒂 𝐻 (𝜃)𝒙 [𝑛], (8)
[51]–[53]. We note that the per-antenna power constraints
where 𝒂(𝜃) is the array steering vector of direction 𝜃. When
can be extended to represent other power-related limitations,
the waveform is reflected from a point target located at angular
such as total power constraints, according to the hardware
direction 𝜃, the received signal can be written as
requirements.
To formulate the power constraint, define the covariance of 𝒓 [𝑛] = 𝛽𝒂 𝑐 (𝜃) 𝒂 𝐻 (𝜃)𝒙 [𝑛 − 𝑛 ′] + 𝒗 [𝑛], (9)
transmit waveform as
  where 𝛽 is the complex amplitude proportional to the radar-
𝑹 = E 𝒙 [𝑛]𝒙 𝐻 [𝑛] . (5) cross sections (RCS) of the target, 𝑛 ′ represents the discrete
time delay, and 𝒗 [𝑛] is additive zero-mean temporally-white
Substituting (1)-(4) into (5) yields the covariance 𝑹 as noise with covariance 𝑹 𝑣 .
Following the guidelines for MIMO radar probing signal
𝑹 = 𝑾𝑟 𝑾𝑟𝐻 + 𝑾𝑐 𝑾𝑐𝐻 . (6)
design stated in [50], the desired goals of MIMO radar transmit
The per-antenna power constraint implies that for each 𝑚 = beamforming include:
1, . . . , 𝑀 it holds that 1) Optimize the transmit power at given directions, or gener-
  ally match a desired beam pattern;
[𝑹] 𝑚,𝑚 = 𝑾𝑟 𝑾𝑟𝐻 + 𝑾𝑐 𝑾𝑐𝐻 𝑚,𝑚 = 𝑃𝑡 /𝑀, (7)
2) Decrease the cross correlation pattern among signals at sev-
where 𝑃𝑡 is the total transmit power. Under this constraint, eral given directions, which is essential for the performance
we discuss the radar and communication metrics for precoder of adaptive MIMO radar techniques.
design in the following section. Here, the transmit power (beam pattern) at angular direction
𝜃 is
 
III. P ERFORMANCE METRICS OF RADAR AND 𝑃(𝜃; 𝑹) = E |𝑦[𝑛; 𝜃]| 2
COMMUNICATION  
Based on the signal model of joint transmit beamforming, = E 𝒂 𝐻 (𝜃)𝒙 [𝑛]𝒙 𝐻 [𝑛] 𝒂(𝜃) = 𝒂 𝐻 (𝜃) 𝑹𝒂(𝜃), (10)
we aim to design the precoders in light of the following and the cross correlation pattern between direction 𝜃 1 and 𝜃 2
guidelines: For MIMO radar, the precoder is designed to syn- is defined as
thesize transmit beams towards radar targets of interests; For
multiuser MIMO communication, the precoder is designed to 𝑃𝑐 (𝜃 1 , 𝜃 2 ; 𝑹) = E (𝑦 ∗ [𝑛; 𝜃 1 ]𝑦[𝑛; 𝜃 2 ])
 
guarantee the receiving SINR at communication users. These = E 𝒂 𝐻 (𝜃 2 )𝒙 [𝑛]𝒙 𝐻 [𝑛] 𝒂(𝜃 1 ) = 𝒂 𝐻 (𝜃 2 ) 𝑹𝒂(𝜃 1 ). (11)
performance metrics of MIMO radar and multiuser MIMO
communication are properly formulated in Subsections III-A From (10) and (11), both the transmit beam pattern and cross
and III-B, respectively. correlation pattern are determined by the covariance 𝑹. Then,
properly beamforming of MIMO radar waveforms is achieved
by designing the covariance matrix 𝑹 [50], [54].
A. MIMO Radar Performance
To this aim, we use the loss function proposed in [50],
The main purpose of MIMO radar beamforming is to direct [54] to evaluate the radar performance, which is the weighted
the transmit beam towards several given directions, so that sum of two parts: beam pattern error and cross correlation.
one can obtain more information of the targets illuminated In particular, the first part is the mean square error (MSE)
by these beams. These directions are typically known to the between the obtained beam pattern and some desired beam
transmitter: When radar works in tracking mode, the beam di- pattern, given by
rection is inferred from the direction of the targets acquired at 𝐿
previous observations; When radar works in searching mode, 1Õ
𝐿 𝑟 ,1 ( 𝑹, 𝛼) = |𝛼𝑑 (𝜃 𝑙 ) − 𝑃(𝜃 𝑙 ; 𝑹)| 2 , (12)
the beam direction is given by the center of angular sector-of- 𝐿 𝑙=1
interest. Consequently, to formulate the performance metric
where 𝛼 is a scaling factor, 𝑑 (𝜃) is the given desired beam
associated with MIMO radar beamforming, we first express 𝐿
pattern, and {𝜃 𝑙 }𝑙=1 are sampled angle grids. The second part
the transmitted signal at each direction, and then develop a
is the mean-squared cross correlation pattern, expressed as
loss function evaluating the transmit beam pattern. Combining
𝑃−1 𝑃
the loss function and the per-antenna power constraint, we 2 Õ Õ 2
achieve an optimization problem which accounts for the radar 𝐿 𝑟 ,2 ( 𝑹) = 2
𝑃𝑐 (𝜃 𝑝 , 𝜃 𝑞 ; 𝑹) , (13)
𝑃 − 𝑃 𝑝=1 𝑞= 𝑝+1
performance.
In DFRC systems, the communication signals can also be where {𝜃 𝑝 } 𝑃𝑝=1 are the given directions of the targets. The
used for sensing, since the radar receiver has complete knowl- summation in (13) is normalized by 𝑃 22−𝑃 , as there exists 𝑃 2−𝑃
2

edge of the transmitted communication waveform. In this way,

pairs of distinct directions in the set {𝜃 𝑝 }. The loss function
the communication signal is not regarded as interference at
of radar is then
the radar receiver. Under the assumption that the transmit
waveform is narrow-band and the propagation path is line 𝐿 𝑟 ( 𝑹, 𝛼) = 𝐿 𝑟 ,1 ( 𝑹, 𝛼) + 𝑤 𝑐 𝐿 𝑟 ,2 ( 𝑹), (14)

1053-587X (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Exeter. Downloaded on June 25,2020 at 20:01:43 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.3004739, IEEE
Transactions on Signal Processing
5

where 𝑤 𝑐 is a weighting factor. As discussed in [50], [54], respectively. Since the users are generally not able to cooperate
the loss function 𝐿 𝑟 ( 𝑹, 𝛼) can be written as a positive- with each other, the off-diagonal elements of 𝑭𝑐 lead to inter-
semidefinite quadratic function of 𝑹 and 𝛼. user interference, which should be mitigated by precoding. At
Combining the loss functions in (12) and (13), the covari- the same time, since the users generally do not have any prior
ance of the transmitted signal in the absence of communication information on radar waveform, 𝑭𝑟 leads to interference from
constrains, i.e., in a radar-only setup, can be designed in radar. At the 𝑘-th user, the signal power is
light of the overall radar objective under per-antenna power  
2
constraints [50], i.e. E |[𝑭𝑐 ] 𝑘,𝑘 𝑐 𝑘 (𝑡)| 2 = [𝑭𝑐 ] 𝑘,𝑘 , (18)

min 𝐿 𝑟 ( 𝑹, 𝛼) (15a) the power of inter-user interference is

𝑹, 𝛼 Õ  Õ
2
subject to 𝑹 ∈ S𝑀+
, (15b) E |[𝑭𝑐 ] 𝑘,𝑖 𝑐 𝑖 (𝑡) 2 = [𝑭𝑐 ] 𝑘,𝑖 , (19)
𝑖≠𝑘 𝑖≠𝑘
[𝑹] 𝑚,𝑚 = 𝑃𝑡 /𝑀, 𝑚 = 1, . . . , 𝑀. (15c)
and the power of interference from radar is
We denote the optimal covariance of this problem by 𝑹0 .
Õ
𝑀  Õ𝑀
Generally, the performance requirement of multiuser MIMO E |[𝑭𝑟 ] 𝑘,𝑖 𝑠𝑖 (𝑡)| 2 =
2
[𝑭𝑟 ] 𝑘,𝑖 . (20)
communication, detailed in Subsection III-B, cannot be satis- 𝑖=1 𝑖=1
fied if the covariance of transmit waveform is 𝑹0 . In other
Therefore, the SINR at the 𝑘-th user is expressed as
words, there is an inherent radar performance loss due to
spectrum sharing with communications compared to the radar- |[𝑭𝑐 ] 𝑘,𝑘 | 2
only case. To address the communication performance of our 𝛾𝑘 = Í Í𝑀 . (21)
𝑖≠𝑘 |[𝑭𝑐 ] 𝑘,𝑖 | 2 + 𝑖=1 |[𝑭𝑟 ] 𝑘,𝑖 | 2 + 𝜎 2
DFRC system, we discuss the communication metric in the
next subsection. Two typical design criteria for multiuser beamforming are
[57], [58]:
• Throughput: maximizing the sum rate
B. Multiuser MIMO Communication Performance
𝐾
Õ
A common performance measure for multiuser broadcast 𝐶 (𝜸) = log2 (1 + 𝛾 𝑘 ) , (22)
communications is the SINR, which is directly related to 𝑘=1
the achievable rate under reduced complexity decoding [55,
• Fairness: maximizing the minimal SINR, referred to
Ch. 8]. Consequently, we design the precoders of MIMO
henceforth as the fairness SINR:
transmission to optimize the users’ SINR. To this aim, we
first present the communication signal model, and derive the 𝐹 (𝜸) = min{𝛾1 , . . . , 𝛾𝐾 }, (23)
expression of SINR with respect to the precoders (for radar
and communication signals) and the channel matrix, followed where 𝜸 = [𝛾1 , . . . , 𝛾𝐾 ]𝑇 .
In this work, we use fairness SINR
by a formulation of the combined objective which accounts 𝐹 (𝜸) as the performance metric for multiuser communication,
for communication performance. and require it to be higher than a given threshold Γ, guaran-
Consider a down-link multiuser MIMO transmission sce- teeing a minimal level of communication quality of service at
nario with 𝐾 < 𝑀 single antenna users observing the output each user, i.e.
of a frequency flat Gaussian noise channel. The channel output 𝛾 𝑘 ≥ Γ, 𝑘 = 1, . . . , 𝐾. (24)
at the 𝐾 users at time instance 𝑛, represented via the 𝐾 × 1 Moreover, the fairness beamforming is simpler in terms of
vector 𝒓 [𝑛], is given by computation complexity and can be solved in polynomial time,
𝒓 [𝑛] = 𝑯𝑾𝑐 𝒄[𝑛] + 𝑯𝑾𝑟 𝒔[𝑛] + 𝒗 [𝑛], (16) while the optimal throughput beamforming is NP hard [57].
We note that in the formulated joint beamforming problem in
where 𝑯 is the 𝐾 × 𝑀 narrow-band channel matrix and 𝒗 [𝑛] is Section IV, the fairness SINR requirement can be extended to
additive white Gaussian noise (AWGN) with covariance 𝜎 2 𝑰 𝐾 . having 𝐾 individual SINR constraints [59], namely,
In multiuser transmit beamforming, the precoder should
be designed to guarantee a certain level of SINR at the 𝛾 𝑘 ≥ Γ 𝑘 , 𝑘 = 1, . . . , 𝐾, (25)
users. Here, it is assumed that the transmit array knows the where Γ 𝑘 is the SINR threshold at the 𝑘-th user.
instantaneous downlink channel 𝑯. This knowledge can be Remark: The transmit power of radar signals is commonly
obtained for example, by exploiting wireless channel reci- much higher than that of a typical communication transmitter,
procity when operating in time-division duplex mode, i.e., the because the echoes reflected from the targets are attenuated
downlink channel is obtained via uplink channel estimation. with a two-way propagation loss. In the DRFC scheme, the
Alternatively, in frequency-division duplex mode, downlink high-power radar interference can induce a notable degradation
channel can be obtained via channel feedback from the users, in the ability of the communication receivers to recover the
see, e.g., [56]. Define the equivalent radar-to-user channel and transmitted symbols. Therefore, the SINR constraint (24) is
equivalent inter-user channel matrices as proposed to mitigate radar interference at users. When the
angular directions of communication users and targets are
𝑭𝑟 = 𝑯𝑾𝑟 , (17a)
sufficiently distinct, the radar waveform can form highly direc-
𝑭𝑐 = 𝑯𝑾𝑐 , (17b) tional beams towards the targets, while eliminating the radar

1053-587X (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Exeter. Downloaded on June 25,2020 at 20:01:43 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.3004739, IEEE
Transactions on Signal Processing
6

interference at the communication users via transmit beam- The selection of the threshold Γ affects the trade-off be-
forming methods with individual communication and radar tween the communication quality and radar performance.
waveform. Such beamforming capabilities may not be achiev- When Γ = 0, (26d) always holds, and the joint radar-
able when transmitting only the communication waveform as communication beamforming problem (26) reduces to the
in Fig. 2(b), motivating the usage of individually precoded radar-only optimization (15). When Γ > 0, compared with the
communication and radar waveforms following the model in radar-only transmit beamforming problem in (15), the precoder
Fig. 2(a). However, when the users and the radar targets 𝑾, which dictates the equivalent channels via (17), is restricted
are located in similar angular directions, canceling the radar by the SINR constraints in (26d). Therefore, there can be an
interference may result in failure to properly illuminate the inherent radar performance loss induced by the need to meet
radar targets. Nevertheless, the communication waveform is the communication performance guarantees, as compared to
completely known for the DFRC system and thus its reflected the radar-only case. If higher Γ is set, higher signal power
signal can also be utilized for target detection, indicating that and less interference is expected to be observed at the user
this challenge can be overcome by forming transmit beam with side, further restricting the precoding matrices. As a result, the
communication waveform to simultaneously cover the targets performance loss of MIMO radar becomes more significant if
and users for jointly radar sensing and data transmission. higher Γ is set.
This can be achieved by transmitting only the communication The optimization problem (26) is not convex because of the
waveform for both communicating and sensing, as in Fig. quadratic equality constraint in (26b) and is thus difficult to
2(b), which is a special case of the join beamforing model solve. Nonetheless, we show in Subsection IV-B that it can
in Fig. 2(a). Consequently, the joint beamforing model in be recast using semidefinite relaxation (SDR) such that the
Fig. 2(a) which is based on individually precoded radar and solution to the solvable relaxed problem is also the global
communication waveforms, is expected to be applicable to optimizer of the original non-convex (26), i.e., the relaxation
a broad range of scenarios, including those for which the is tight. To further reduce the computation complexity, we
model in Fig. 2(b) is suitable, as well as additional challenging propose a sub-optimal zero-forcing beamforming strategy in
setups, as numerically demonstrate in Section V. To achieve Subsection IV-C, which is shown to be able to approach the
this improved and robust dual-function capabilities, in the performance of the global solution to (26) in our numerical
following section we derive a joint beamforming scheme based study presented in Section V.
on the signal model of Fig. 2(b).
IV. J OINT T RANSMIT B EAMFORMING B. Joint Transmit Beamforming via SDR
With the proposed MIMO radar and communication perfor-
In this subsection, we tackle the non-convex problem (26)
mance metrics, we now turn to design a DFRC joint beam-
using an SDR strategy [60], [61]. To this aim, we first
forming scheme. We begin by formulating the joint transmit
explicitly write the relationship (26b) as a quadratic constraint
beamforming as an optimization problem with respect to the
with respect to each column of 𝑾. Let 𝒘 𝑖 denote the 𝑖-th
precoding matrices in Subsection IV-A. To solve this problem,
column of 𝑾, for 𝑖 = 1, . . . , 𝑀 + 𝐾. Then (26b) becomes
we propose a semidefinite relaxation (SDR) based optimiza-
tion scheme in IV-B, and a zero-forcing (ZF) methods which 𝑀
Õ +𝐾
cancels the inter-user interference and the radar interference 𝑹= 𝒘 𝑖 𝒘 𝑖𝐻 . (27)
in Subsections IV-C, respectively. 𝑖=1

A. Problem formulation Defining 𝑹𝑖 = 𝒘 𝑖 𝒘 𝑖𝐻 , we have

The goal of our joint DFRC beamforming is to optimize the 𝑀
Õ +𝐾
radar beam pattern under the transmit power and communica- 𝑹= 𝑹𝑖 , (28)
tion quality of service constraints. In particular, we minimize 𝑖=1
the loss function on radar beam pattern defined in (14), under
where we omit the rank-one constraints. The SINR constraints
the per-antenna power constraint (7) and the fairness SINR
in (26d) can be converted to linear constraints in the rank-
constraint (24) for each downlink user.
one matrices {𝑹𝑖 }. Letting 𝒉 𝐻 𝑘 denote the 𝑘-th row of 𝑯,
Let 𝑾 = [𝑾𝑐 , 𝑾𝑟 ] be the overall precoding matrix. The
𝑘 = 1, . . . , 𝐾, the entires of the equivalent channel matrices
precoding matrix can be obtained by solving the following
can be written as [𝑭𝑐 ] 𝑘,𝑖 = 𝒉 𝐻 𝐻
𝑘 𝒘 𝑖 , and [𝑭𝑟 ] 𝑘,𝑖 = 𝒉 𝑘 𝒘 𝑖+𝐾 .
optimization problem
Consequently, the SINR constriant becomes
min 𝐿 𝑟 ( 𝑹, 𝛼) (26a)
𝑾 ,𝛼
𝒉𝐻 𝐻
𝑘 𝒘 𝑘 𝒘 𝑘 𝒉𝑘
+
subject to 𝑹 = 𝑾𝑾 𝐻 ∈ S𝑀 , (26b) 𝛾𝑘 = Í 𝐻 𝐻 2
1≤𝑖 ≤𝑀 +𝐾 ,𝑖≠𝑘 𝒉 𝑘 𝒘 𝑖 𝒘 𝑖 𝒉 𝑘 + 𝜎
[𝑹] 𝑚,𝑚 = 𝑃𝑡 /𝑀, 𝑚 = 1, . . . , 𝑀, (26c)
𝒉𝐻𝑘 𝑹𝑘 𝒉𝑘
𝛾 𝑘 ≥ Γ, 𝑘 = 1, . . . , 𝐾, (26d) = Í 𝐻 2
1≤𝑖 ≤𝑀 +𝐾 ,𝑖≠𝑘 𝒉 𝑘 𝑹𝑖 𝒉 𝑘 + 𝜎
where (26a)-(26c) come from (15) addressing the radar perfor- 𝐻
𝒉𝑘 𝑹𝑘 𝒉𝑘
mance, and (26d) follows from considering the fairness SINR = ≥ Γ. (29)
requirement (24). 𝒉 𝑘 𝑹𝒉 𝑘 − 𝒉 𝐻
𝐻
𝑘 𝑹𝑘 𝒉𝑘 + 𝜎
2

1053-587X (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Exeter. Downloaded on June 25,2020 at 20:01:43 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.3004739, IEEE
Transactions on Signal Processing
7

We now cast (26) as an equivalent quadratic semidefinite be unique and convex optimization software may not give a
programming (QSDP) with rank-one constraints rank-one solution. Once the optimal solution 𝑹, ˆ 𝑹ˆ 1 , . . . , 𝑹ˆ 𝐾
are obtained, we use them to obtain the rank-one optimal
min 𝐿 𝑟 ( 𝑹, 𝛼) (30a)
𝑹, {𝑹𝑖 }, 𝛼 solution 𝑹˜ 1 , . . . , 𝑹˜ 𝐾 and the corresponding optimal precoder
𝑀
Õ +𝐾 𝒘˜ 1 , . . . , 𝒘˜ 𝐾 , as presented in Appendix A. First, we compute
+
subject to 𝑹 = 𝑹𝑖 ∈ S 𝑀 , (30b) 𝑹˜ 1 , . . . , 𝑹˜ 𝐾 and 𝑹,˜ 𝒘˜ 1 , . . . , 𝒘˜ 𝐾 via
𝑖=1

𝑹˜ = 𝑹, ˆ 𝒘˜ 𝑘 = 𝒉 𝐻 𝑹ˆ 𝑘 𝒉 𝑘 −1/2 𝑹ˆ 𝑘 𝒉 𝑘 , 𝑹˜ 𝑘 = 𝒘˜ 𝑘 𝒘˜ 𝐻 , (33)
[𝑹] 𝑚,𝑚 = 𝑃𝑡 /𝑀, 𝑚 = 1, . . . , 𝑀, (30c) 𝑘 𝑘
+
𝑹𝑖 ∈ S 𝑀 , rank( 𝑹𝑖 ) = 1, 𝑖 = 1, . . . , 𝐾 + 𝑀, for 𝑘 = 1, . . . , 𝐾. According to the proof of Theorem 1,
(30d) ˜ 𝑹˜ 1 , . . . , 𝑹˜ 𝐾 is optimal to (31) and hence is also optimal
𝑹,
 
(32). To show that 𝑹, ˜ 𝑹˜ 1 , . . . , 𝑹˜ 𝐾 is also optimal to (30),
1 + Γ−1 𝒉 𝐻 𝐻 2
𝑘 𝑹 𝑘 𝒉 𝑘 ≥ 𝒉 𝑘 𝑹𝒉 𝑘 + 𝜎 , 𝑘 = 1, . . . , 𝐾, (30e)
we construct rank-one matrices { 𝑹˜ 𝑖 }𝑖 ≥𝐾 +1 as 𝑹˜ 𝑖 = 𝒘˜ 𝑖 𝒘˜ 𝑖𝐻 ,
where (30e) is derived from (29). We observe that in problem where the vectors 𝒘˜ 𝑖 for 𝑖 > 𝐾 are calculated by the Cholesky
(30), the individual matrices {𝑹𝑖 }𝑖 ≥𝐾 +1 have no effect on the decomposition [66]
SINR constraints and are only encapsulated in the overall
𝐾
Õ
covariance matrix 𝑹. Therefore, we can remove the variables
{𝑹𝑖 }𝑖 ≥𝐾 +1 , and (30) is relaxed to 𝑾𝑟 𝑾𝑟𝐻 = 𝑹˜ − 𝒘˜ 𝑘 𝒘˜ 𝐻
𝑘 , (34)
𝑘=1
min 𝐿 𝑟 ( 𝑹, 𝛼) (31a)
𝑹,𝑹1 ,...,𝑹 𝐾 , 𝛼 where 𝑾𝑟 = [ 𝒘˜ 𝐾 +1 , . . . , 𝒘˜ 𝐾 +𝑀 ] is a lower triangular matrix.
𝐾
Õ From (34), it can be verified that constraint (30b) holds for
+ + ˜ 𝑹˜ 1 , . . . , 𝑹˜ 𝐾 +𝑀 . Therefore, 𝑹,
˜ 𝑹˜ 1 , . . . , 𝑹˜ 𝐾 +𝑀 is a feasible
subject to 𝑹 ∈ S𝑀 , 𝑹− 𝑹 𝑘 ∈ S𝑀 , (31b) 𝑹,
𝑘=1 solution to (30) and hence is also an optimal solution to (30).
[𝑹] 𝑚,𝑚 = 𝑃𝑡 /𝑀, 𝑚 = 1, . . . , 𝑀, (31c) Furthermore, the precoding matrix 𝑾 ˜ = [ 𝒘˜ 1 , . . . , 𝒘˜ 𝐾 +𝑀 ] is a
+ solution to (26).
𝑹𝑘 ∈ S𝑀 , rank( 𝑹 𝑘 ) = 1, 𝑘 = 1, . . . , 𝐾, (31d)
  We summarize the procedure to compute the precoding
1 + Γ−1 𝒉 𝐻 𝐻 2
𝑘 𝑹 𝑘 𝒉 𝑘 ≥ 𝒉 𝑘 𝑹𝒉 𝑘 + 𝜎 , 𝑘 = 1, . . . , 𝐾. (31e) matrix 𝑾 in Algorithm 1. The main computational burden in
The optimization problem (31) is still non-convex because Algorithm 1 stems from solving the QSDP (32). Specifically,
of the rank-one constraints. Omitting these constraints leads given a solution accuracy 𝜖, the worst case complexity to solve
to the following relaxation: the QSDP (32) with the primal-dual interior-point algorithm
in [67], [68] is O (𝐾 6.5 𝑀 6.5 log(1/𝜖)).
min 𝐿 𝑟 ( 𝑹, 𝛼) (32a)
𝑹,𝑹1 ,...,𝑹 𝐾 , 𝛼
𝐾
Õ Algorithm 1 Joint transmit beamforming via SDR
+ + Input:
subject to 𝑹 ∈ S𝑀 , 𝑹− 𝑹 𝑘 ∈ S𝑀 , (32b)
𝑘=1 Total transmit power 𝑃𝑡 ;
[𝑹] 𝑚,𝑚 = 𝑃𝑡 /𝑀, 𝑚 = 1, . . . , 𝑀, (32c) Power of AWGN at users 𝜎 2 ;
𝑹𝑘 ∈ +
S𝑀 , 𝑘 = 1, . . . , 𝐾, (32d) Expression of the MIMO radar loss function 𝐿 𝑟 ( 𝑹, 𝜶);
  Instantaneous downlink channel 𝑯;
1 + Γ−1 𝒉 𝐻 𝐻 2
𝑘 𝑹 𝑘 𝒉 𝑘 ≥ 𝒉 𝑘 𝑹𝒉 𝑘 + 𝜎 , 𝑘 = 1, . . . , 𝐾. (32e) SINR threshold Γ.
Output:
This relaxed optimization model (32) is a convex QSQP,
The overall precoding matrix 𝑾.
because the target function is a positive-semidefinite quadratic
Steps:
form and all the constraints are either linear or semidefinite. ˆ 𝑹ˆ 1 , . . . , 𝑹ˆ 𝐾 by solving
1: Compute the optimal value of 𝑹,
The global optimum of (32) can be obtained in polynomial
(32) with convex optimization solvers.
time with convex optimization toolboxes [62]–[65].
2: Compute 𝒘 ˜ 1 , . . . , 𝒘˜ 𝐾 via (33).
The relaxation used in SDR is tight if the optimal
3: Compute 𝒘 ˜ 𝐾 +1 , . . . , 𝒘˜ 𝐾 +𝑀 via (34).
𝑹1 , . . . , 𝑹𝐾 for (32) are exactly rank-one, i.e., the solution ˜ = [ 𝒘˜ 1 , . . . , 𝒘˜ 𝐾 +𝑀 ].
4: Set the overall precoding matrix 𝑾
to the relaxed problem is also a solution to the original non-
convex problem. While such relaxations are not necessarily
tight, the SDR used in obtaining (32) from (31) is tight, as
stated in the following theorem:
C. Joint Transmit Beamforming via ZF
Theorem 1. There exists a global optimum for (32), denoted
˜ 𝑹˜ 1 , . . . , 𝑹˜ 𝐾 , satisfying The computational burden associated with obtaining the
by 𝑹,
precoder via Algorithm 1 motivates seeking a reduced com-
rank( 𝑹˜ 𝑘 ) = 1, 𝑘 = 1, . . . , 𝐾. plexity sub-optimal beamforming strategy. In this subsection,
we focus on ZF beamforming. ZF methods facilitate obtaining
Proof. See Appendix A. 
closed-form, tractable, and interpretable precoders [58], [69].
We note that Theorem 1 only states that the rank-one global In addition to its relative simplicity, from a communications
optimum exists. Generally, the global optimum to (32) may not perspective, ZF beamforming is known to asymptotically

1053-587X (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Exeter. Downloaded on June 25,2020 at 20:01:43 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.3004739, IEEE
Transactions on Signal Processing
8

approach the sum-capacity in broadcast channels [70], indi- detained in the proof of Theorem 2. Here, we briefly give
cating its potential to approach optimal performance in setups the final expressions. First, we recover an 𝑀 × 𝑀 matrix
involving multi-user communications. 𝑳 𝑟 which satisfies 𝑹˜ = 𝑳 𝑟 𝑳 𝑟𝐻 . This can be obtained using,
We design the precoders to eliminate the inter-user inter- e.g., Cholesky decomposition, though 𝑳 𝑟 does not have to be
ference and radar interference, obtained by restricting 𝑭𝑐 to a triangular and any matrix satisfying 𝑹˜ = 𝑳 𝑟 𝑳 𝑟𝐻 may be used
diagonal matrix and 𝑭𝑟 to a zero matrix, i.e. ˜ Then, the resulting precoder 𝑹˜ is
to calculate 𝑾.
√ √
h i𝑇
𝑭𝑐 = diag 𝑝 1 , . . . , 𝑝 𝐾 , 𝑭𝑟 = 0𝐾 ×𝑀 . (35) ˜ = 𝑳 𝑟 𝑸 𝐻 𝑸𝑇
𝑾 ℎ 𝑓 , (41)
1:𝑀
Here, 𝑝 𝑘 is the signal power at the 𝑘-th user, for 1 ≤ 𝑘 ≤ 𝐾.
Enforcing the interference to be canceled facilitates achieving where 𝑸 ℎ and 𝑸 𝑓 are obtained by applying row QR
high SINR values at the users. In our numerical study in decomposition
 √ to 𝑯𝑳 𝑟 and 𝑭, respectively. Since 𝑭 =
Section V we demonstrate that the achievable performance diag( 𝒑),
˜ 0𝐾 ×𝑀 is diagonal, it holds that 𝑸 𝑓 = 𝑰 𝑀 +𝐾 , and
under the additional ZF constraint approaches that of the thus (41) is simplified to
global solution to (26), obtained with increased computational  
𝑾˜ = 𝑳 𝑟 𝑸 𝐻 , 0 𝑀 ×𝐾 . (42)
burden via Algorithm 1, when the SINR threshold is high. ℎ
In ZF beamforming, the SINR constraint (26d) is refor- According to the proof of Theorem 1, 𝑾, ˜ 𝑹˜ is a feasible
mulated as Γ1 𝑝 𝑘 ≥ 𝜎 2 , and the corresponding optimization solution to (36). Since 𝑹˜ is the global optimum to (36), 𝑾˜ is
problem (26) becomes also globally optimal to (36).
min 𝐿 𝑟 ( 𝑹, 𝛼) (36a) The resulting ZF beamforming method is summarized below
𝑾 ,𝛼
as Algorithm 2. The main computational burden in Algo-
+
subject to 𝑹 = 𝑾𝑾 𝐻 ∈ S𝑀 , (36b) rithm 2 stems from solving the QSDP problem (40), as is
[𝑹] 𝑚,𝑚 = 𝑃𝑡 /𝑀, 𝑚 = 1, . . . , 𝑀, (36c) also the case in Algorithm 1. Given a solution accuracy
 √  𝜖, the worst case complexity to solve the QSDP problem
𝑯𝑾 = diag( 𝒑), 0𝐾 ×𝑀 , (36d)
(40) with the primal-dual interior-point algorithm in [67],
1
𝑝 𝑘 ≥ 𝜎 2 , 𝑘 = 1, . . . , 𝐾. (36e) [68] is O (𝑀 6.5 log(1/𝜖)). Compared to the recovering the
Γ global solution via the SDR-based Algorithm 1, the worst-
The ZF beamforming optimization (36) is still non-convex. case computation complexity for ZF beamforming is lower
The following theorem shows that it can be converted to by a factor of 𝐾 6.5 . This computational complexity reduction
convex problem. stems from the fact that the optimization problem (40) involves
Theorem 2. Given a covariance matrix 𝑹 ∈ S𝑛+ and a full only one semidefinite constraint, while the problem (32), from
rank 𝐾 × (𝐾 + 𝑀) matrix 𝑭, there exists a matrix 𝑾 satisfying which Algorithm 1 originates, involves 𝐾 + 2 = O (𝐾) such
(36b) and constraints.
𝑯𝑾 = 𝑭 (37)
Algorithm 2 Joint transmit beamforming via ZF
if and only if Input:
𝑯𝑹𝑯 𝐻 = 𝑭𝑭 𝐻 . (38) Total transmit power 𝑃𝑡 ;
Power of AWGN at users 𝜎 2 ;
Proof. See Appendix B. 
Expression of the MIMO radar loss function 𝐿 𝑟 ( 𝑹, 𝜶);
Theorem 2 indicates that constraints (36b) and (36d) are Instantaneous downlink channel 𝑯;
equivalent to SINR threshold Γ.
𝑯𝑹𝑯 𝐻 = diag ( 𝒑) , (39) Output:
 √  The overall precoding matrix 𝑾. ˜
by letting 𝑭 = diag( 𝒑), 0𝐾 ×𝑀 . Using (39), the globally Steps:
optimal 𝑹 to (36) is found by 1: Compute the optimal 𝑹, ˜ and 𝒑˜ by solving optimization
min 𝐿 𝑟 ( 𝑹, 𝛼) (40a) problem (40) with convex optimization solvers.
𝑹, 𝛼 2: Compute the Cholesky decomposition of 𝑹˜ as 𝑹˜ = 𝑳 𝑟 𝑳 𝑟𝐻 .
subject to 𝑹 ∈ S+𝑀 , 𝑯𝑹𝑯 𝐻 = diag ( 𝒑) , (40b) 3: Given 𝑯𝑳 𝑟 , calculate 𝑸 ℎ with the row QR decomposition
[𝑹] 𝑚,𝑚 = 𝑃𝑡 /𝑀, 𝑚 = 1, . . . , 𝑀, (40c) (59) shown later in Appendix B.
4: Compute the overall precoding matrix 𝑾 ˜ using (42).
1
𝑝 𝑘 ≥ 𝜎 2 , 𝑘 = 1, . . . , 𝐾. (40d)
Γ
Similar to (32), the optimization (40) is a convex QSDP, and We next discuss how the selection of the SINR threshold
the global optimum of (40) can be obtained in polynomial Γ affects trade-off between communications and radar when
time. As we show in the sequel, the overall complexity of ZF using ZF beamforming. As noted in the discussion following
beamforming is substantially lower than that of recovering the the original optimization problem (26), the radar loss function
global optimum via Algorithm 1. here decreases as Γ increases, i.e., the less restrictive the com-
The solution of (40), i.e., the matrix 𝑹˜ and the vector 𝒑,
˜ munication constraints are, the better the radar functionality
are used to construct the optimal precoding matrix 𝑾, ˜ As can perform. However, there are two phenomenons which are

1053-587X (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Exeter. Downloaded on June 25,2020 at 20:01:43 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.3004739, IEEE
Transactions on Signal Processing
9

explained in the sequel, that are different under ZF beam- V. N UMERICAL RESULTS
forming compared to the original optimization problem (26): In this section we numerically evaluate the proposed joint
1) When Γ approaches zero, the radar performance achieved is beamforming methods, i.e. SDR beamforming (Algorithm 1)
generally different from the radar-only optimal performance; and ZF beamforming (Algorithm 2), in a simulation study. We
2) The radar loss function and the obtained fairness SINR begin by analyzing the achievable radar beampattern of the
remain constant if Γ is lower than some positive value. proposed schemes, compared to the DFRC method of [42] in
To understand phenomenon 1), we specialize the ZF op- Subsection V-A. Then, we compare ZF and SDR beamforming
timization problem (40) for the case of Γ = 0, resulting in in Subsection V-B, while the comparison of SDR beamforming
and the DFRC method of [42] in terms of their inherent radar-
communication tradeoffs is presented in Subsection V-C.
min 𝐿 𝑟 ( 𝑹, 𝛼) (43a)
𝑹,𝒑, 𝛼 In the experiments reported in this section, we use the
+ following settings: The transmit array is a uniform linear array
subject to 𝑹 ∈ S𝑀 , (43b)
[𝑹] 𝑚,𝑚 = 𝑃𝑡 /𝑀, 𝑚 = 1, . . . , 𝑀, (43c) with half wavelength spaced elements. The number of transmit
𝐻
antennas is 𝑀 = 10 and the total transmit power 𝑃𝑡 = 1.
𝑯𝑹𝑯 = diag ( 𝒑) . (43d) For MIMO radar transmit beamforming, the ideal beam pat-
Here, we note that this formulation is distinct from the tern consists of three main beams, whose the directions are
radar-only optimization problem (15), since, even when the 𝜃 1 = −40◦ , 𝜃 2 = 0◦ and 𝜃 3 = 40◦ . The width of each ideal
SINR can take any value, we still force the interference to beam is Δ = 10◦ , and thus the desired beam pattern is

be cancelled. This restriction is reflected in the additional 1, 𝜃 𝑝 − Δ2 ≤ 𝜃 ≤ 𝜃 𝑝 + Δ2 , 𝑝 = 1, 2, 3,
constraint (43d) imposed on 𝑹, namely that 𝑯𝑹𝑯 𝐻 should 𝑑 (𝜃) = (45)
0, otherwise.
be a diagonal matrix. The optimal radar-only covariance 𝑹0 ,
𝐿
which is not forced to satisfy this interference cancelling In (12), the direction grids {𝜃 𝑙 }𝑙=1 are obtained by uniformly
constraint, generally does not satisfy it, i.e. 𝑯𝑹0 𝑯 𝐻 is not sampling the range of −90 to 90◦ with resolution of 0.1◦ .
◦

a diagonal matrix. If 𝑹0 is not a feasible solution of problem The radar loss in (14) accounts for both objectives equally,
(43), the radar-only optimal performance cannot be achieved. namely, the weighting factor is set to 𝑤 𝑐 = 1. The multi-user
In order to explain phenomenon 2), we again focus on the communications channel obeys a Rayleigh fading model, i.e.,
ZF optimization specialized to the case of no SINR constraints the entries of 𝑯 are i.i.d. standard complex normal random
in (43), and denote its solution by {𝑹II , 𝛼II , 𝒑 II }. Given 𝒑 II , variables, and the channel output at each user is corrupted
the resulting fairness SINR is given by with an additive white Gaussian noise of variance 𝜎 2 = 0.01.
In our simulations we use SINR threshold values Γ vary-
ΓII = min{ 𝒑 II }/𝜎 2 . (44) ing from 4dB to 24dB, and number of users simulated is
𝐾 = 2, 4, 6. We simulate different Γ and 𝐾 to test the
In problem (40), if the given Γ is not greater than ΓII , i.e. 0 ≤
impact of these parameters on the performance of the proposed
Γ ≤ ΓII , the constraint (40d) always holds and can be regarded
joint beamforming methods. For each value of Γ and 𝐾,
as being invariant to the actual solution {𝑹II , 𝛼II , 𝒑 II }. In this
the performance is averaged over 1000 Monte Carlo tests.
case, {𝑹II , 𝛼II , 𝒑 II } is still a feasible solution for (40). Thus,
The individual radar waveform and communication symbols
the minimized radar loss function is equal to 𝐿 𝑟 ( 𝑹II , 𝛼II ), and
comprising the transmitted signal 𝒙 [𝑛] in (1) are generated as
the obtained fairness SINR is equal to ΓII , for ZF beamforming
random quadrature-phase-shift-keying modulated sequences,
derived under every SINR constraint satisfying Γ ≤ ΓII .
and the transmit signal block size set to is 𝑁 = 1024.
Here we compare the two proposed beamforming methods. The MATLAB CVX toolbox [71], [72] is used to solve
The key difference between them is whether to completely the QSDP problems (32) and (40). We compare our joint
eliminate the interference. As a globally optimal method, the beamforming schemes with the DFRC beamforming method
radar performance of SDR beamforming should be better proposed in [42], in which only communication symbols are
than that of the sub-optimal ZF beamforming under the same precoded. Specially, we use gradient projection method to
communication requirement. However, the performance gap solve the sum-square penalty (SSP) problem which only pre-
may become small with reasonably large Γ, because the codes communication symbols under per-antenna constraint in
interference is expected to be eliminated under strict constraint [42]. In the sum-square penalty problem in [42], the weighing
on SINR. In this case, ZF beamforming is preferable since its factors are 𝜌1 = 1, 𝜌2 = 2 and the given SINR at each user is
corresponding QSDP problem has a much simpler form. To equal to the SINR threshold Γ in (26) .
explain the performance gap when the given Γ is low, we note
that it is unnecessary to completely eliminate the interference,
which restricts the precoder in a null space. Thanks to the A. MIMO Radar Transmit Beam Pattern
more degrees of freedom for designing 𝑾, SDR beamforming First, we numerically evaluated the MIMO radar transmit
enjoys better radar performance. In addition, as Γ goes to beam patterns 𝑃(𝜃; 𝑹) defined in (10) for SDR beamforming,
zero, the radar performance of SDR beamforming goes to the ZF beamforming, and the SSP approach [42]. The transmit
optimal radar-only performance, while the radar performance beam patterns for Γ = 12 dB are depicted in Fig. 3 for 𝐾 = 2
of ZF beamforming cannot and stays constant when Γ is lower and in Fig. 4 for 𝐾 = 4. The optimal radar-only beam pattern,
than some value. obtained from (15), are also evaluated for comparison.

1053-587X (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Exeter. Downloaded on June 25,2020 at 20:01:43 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.3004739, IEEE
Transactions on Signal Processing
10

Observing Fig. 3, we note that when 𝐾 = 2, the average

10
beam pattern for SDR beamforming and ZF beamforming
approaches that of the optimal radar-only beamforming, while
the SSP beamformer of [42] only synthesizes two main beams 5

towards 0◦ and 40◦ . The fact that the SSP beamformer is

unable to steer three main beams for 𝐾 = 2 stems from 0

its decreased MIMO radar DoF. In particular, as noted in

(10), the MIMO transmit beam pattern is determined by the -5
covariance of transmit waveform, and thus the DoF for MIMO
radar transmit beamforming is given by the rank of covariance
-10
matrix. In the SSP approach, only communication symbols are
precoded and thus the DoF cannot be larger than the number of
users 𝐾, namely, the rank of covariance matrix cannot exceed -15
-90 -60 -30 0 30 60 90
𝐾. In our scheme however, both communication symbols and
radar waveform are precoded, and thus the DoF can be as high
as its maximal value 𝑀, i.e., the covariance can have full rank. Fig. 3. Transmit beam pattern of MIMO radar, for Γ = 12 dB and 𝐾 = 2.
Numerically solving (15) using the CVX toolbox reveals that
rank of the optimal radar-only covariance 𝑹0 is 4. In other
6
words, the required DoF to achieve the optimal performance
of radar is 4. As a result, if 𝐾 = 2 < 4, the SSP approach 4

does not have enough DoF to form three main beams as in 2

the optimal radar beam pattern, explaining the degraded beam 0

pattern observed in Fig. 3. Our scheme are capable of forming
-2
beam patterns which are close to the optimal radar beam
pattern, since the available DoF in our schemes is 𝑀 = 10 > 4. -4

When 𝐾 = 4, the SSP approach has enough DoF and is thus -6

capable of forming a beam pattern comparable to the optimal -8

radar beam pattern, as shown in Fig. 4.
-10
We also observe in Figs. 3-4 that the main-lobe power of
the ZF beamforming is lower than that of the SDR beam- -12
-90 -60 -30 0 30 60 90
forming, implying an expected radar performance loss for ZF
beamforming compared to SDR beamforming. This follows
since when the SINR threshold Γ is not very high, one can Fig. 4. Transmit beam pattern of MIMO radar, for Γ = 12 dB and 𝐾 = 4.
achieve the desired SINR level without canceling the inter-
Radar performance is evaluated using the beam pattern MSE,
ference, allowing to further optimize the radar beam pattern
defined as the MSE between the obtained MIMO radar trans-
by proper optimization. In ZF beamforming, the interference
mit beam pattern and the optimal radar-only beam pattern, and
is completely eliminated
 regardless of the SINR threshold,
√ is written as
namely 𝑯𝑾 = diag( 𝒑), 0𝐾 ×𝑀 even if Γ is low. This
𝐿
additional constraint limits the DoF of 𝑾 and introduces the 1Õ
radar performance loss compared to the SDR beamforming. MSE = |𝑃(𝜃 𝑙 ; 𝑹0 ) − 𝑃(𝜃 𝑙 ; 𝑹)| 2 , (46)
𝐿 𝑙=1
Nevertheless, in order to fully compare ZF beamforming to
SDR beamforming, one must also account for the communica- where 𝑃(𝜃 𝑙 ; 𝑹0 ) is the optimal radar-only beam pattern with
tion performance. In particular, ZF beamforming can provide 𝑹0 obtained from (15). Low beam pattern MSE indicates
improved communication rates compared SDR beamforming improved MIMO radar transmit beamforming. The numeri-
due to the fact that it completely eliminates the interference cally compared beam pattern MSE values versus the SINR
regardless of the specified SINR threshold. Our numerical threshold Γ are depicted in Fig. 5. As expected, the beam
results detailed in the sequel show that the obtained SINR pattern MSE increases with the increment of Γ, implying that
of ZF beamforming may be much higher than Γ, while the the more restrictive SINR demands naturally come at the cost
obtained SINR by the SDR beamforming is generally quite of radar performance. The results in Fig. 5 validate three
close to Γ. To understand the inherent tradeoffs of the proposed characters of the two proposed joint beamforming schemes:
schemes, in the following subsection we compare SDR and ZF 1) SDR beamforming achieves improved radar performance
beamforing in terms of both their radar and communication compared the the sub-optimal ZF strategy; 2) The performance
performance measures. gap between the two methods notably narrows at high SINR
constraints, i.e., as Γ increases; 3) When Γ is lower than
some value, radar performance of the ZF beamforming stays
B. Comparison Between ZF and SDR Beamforming constant, as discussed in Subsection IV-C. It is also observed
In this subsection, we compare the radar performance and Fig. 5 that the more communication receivers the DFRC
communication performance of SDR and ZF beamforming. system has to communicate with reliably, i.e., as 𝐾 increases,

1053-587X (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Exeter. Downloaded on June 25,2020 at 20:01:43 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.3004739, IEEE
Transactions on Signal Processing
11

100 40

35
10-1

30
10-2

25

10-3

20

-4
10
15

10-5
10 11 12 13 14 15 16 17 18 10

5
Fig. 5. Beam pattern MSE versus SINR threshold Γ. 10 11 12 13 14 15 16 17 18

the higher the beam pattern MSE is, again indicating the
inherent tradeoff between radar and communications in DFRC Fig. 6. Achievable sum rate versus SINR threshold Γ.
systems. In particular, it is observed that the impact of 𝐾 on
the beam pattern MSE is more significant than the impact of Γ,
20
namely, the demand to support an increased number of users
is more restrictive in terms of radar performance compared to 15
the requirement to provide improved SINR at each user.
The communication performance is evaluated in terms of 10

the achievable sum rate defined in (22). The resulting values

5
are depicted in Fig. 6, where we observe that ZF beamforming
achieves higher communciation rate compared to SDR beam- 0

forming, despite its performance loss for radar. This follows

-5
since, as discussed in the previous subsection, ZF beamform-
ing typically yields SINR values higher than the imposed -10
threshold Γ, as it nullifies the interference regardless of the
value of Γ. Conversely, SDR beamforming, which aims at -15
2 4 6 8 10 12 14 16 18
improving radar performance without imposing any structure
on the resulting interference, does so by tunning its SINR to
be close to the threshold Γ, allowing to further improve radar Fig. 7. Interference-to-noise ratio versus SINR threshold Γ, SDR beamform-
performance without violating the SINR constraint. Hence, ing.
the achievable sum rate of SDR beamforming demonstrates (32) and (36) may become infeasible. To calculate the feasible
an approximate linear increase with the SINR constraint Γ in probability under a given Γ and 𝐾, we ran multiple Monte
Fig. 6. Carlo tests, randomizing a new channel realization in each test.
From Figs. 5 and 6, it is observed that performance of The feasible probability is calculated by dividing the number
the two methods coincides as Γ increases. For large values of feasible tests by the total number of tests.
of Γ, the interference tends to be naturally eliminated by the The relationship between the feasible probability and Γ is
SDR beamforming in order to meet the SINR constraints. To demonstrated in Fig. 8, for 𝐾 = 2, 4, 6. It is observed in
demonstrate this property, we depict in Fig. 7 the interference- Fig. 8 that the feasible probability is roughly the same for
to-noise ratio at the first user versus SINR threshold for ZF and SDR beamforming, and that both curves decrease
SDR beamforming. Observing Fig. 7, we note that when as the number of users and SINR threshold increases. This
Γ is high enough, the interference power becomes much implies that our optimization approach may fail with very high
dominant than the noise power, and thus the interference can SINR restrictions, and thus for practical applications, the SINR
be effectively ignored. Therefore, under high SINR conditions, threshold should be carefully set. If the given threshold is too
it is reasonable to completely eliminate the interference, and high, the two problems may be infeasible and our method will
ZF beamforming is asymptotically optimal. fail to return any meaningful solution. Nevertheless, this result
We note that optimization problems tackled by SDR beam- shows that the feasibility can almost always be ensured if Γ is
forming (32) and ZF beamforming (36) are not always feasi- lower than some value under Rayleigh channel. We also note
ble. As the total transmit power is fixed to be 𝑃𝑡 , the signal that this infeasible situation can be avoided if one changes the
power at users should have an upper bound. Correspondingly, SINR constraints into a part of penalty functions, i.e. (32a) or
the achievable feasible SINR should have an upper bound. (40a), as done in the scheme in [42]. We leave the analysis of
If the given Γ is too high, the joint beamforming problem this modification for future investigation.

1053-587X (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Exeter. Downloaded on June 25,2020 at 20:01:43 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.3004739, IEEE
Transactions on Signal Processing
12

1 100

0.9
10-2
0.8

0.7
10-4
0.6

0.5 10-6

0.4
10-8
0.3

0.2
10-10
0.1

0 10-12
0 5 10 15 20 25 0 2 4 6 8 10 12 14 16 18 20

Fig. 8. Feasible probability the SDR beamforming (32) and ZF beamforming Fig. 9. Beam pattern MSE versus SINR threshold Γ.
(36) versus SINR threshold Γ.
radar receive signal is corrupted with Gaussian noise with
C. Comparing SDR beamforming with SSP DFRC Method covariance 𝑹 𝑣 = 𝜎𝑟2 𝑰, where 𝜎𝑟2 = 1. The Capon spatial
Finally, we compare our proposed SDR beamforming spectrum at the 20-th range resolution bin and the range profile
method to the SSP DFRC scheme previously proposed in [42], at direction 0◦ in one test are demonstrated in Fig. 10, for
which utilizes a single precoded communication waveform for 𝐾 = 2 and Γ = 12dB. In 10, the range profile and Capon spatial
both communicating and sensing. To that aim, we evaluate spectrum for the radar-only case, the SSP approach and SDR
their tradeoff between the communication performance, en- beamforming are compared. From Fig. 10, it is observed that
capsulated in the achieved fairness SINR defined in (23) and the range and angular resolution for SDR beamforming is close
the radar beam pattern MSE defined in (46). The numerically to that for the radar-only case. The performance degradation
evaluated tradeoffs for number of users 𝐾 = 2, 4, 6 are of the SSP approach resulting from the lack of radar DoF is
depcited in Fig. 9. As discussed in Subsection V-A, our scheme significant, since the SSP approach cannot form a notable peak
notably outperform the SSP approach for 𝐾 = 2, as clearly around the coordinate (20, 0◦ ), see Figs. 10(c) and 10(d), and
demonstrated in Fig. 9. Since the SSP approach only precodes the amplitude estimation error at the 20-th range resolution
communication symbols, it cannot provide enough degrees bin is very large, see Fig. 10(d). When 𝐾 = 2, the reflected
of freedom for MIMO radar transmit beamforming, and the signal from the three targets at the 20-th range resolution
obtained beam pattern MSE is significant. When 𝐾 = 4, 6, bin in the SSP approach are linearly dependent, and thus
our SDR beamforming technique still outperforms the SSP the cross correlation defined in (11) cannot be suppressed
approach, although the gain is less notable compared to 𝐾 = 2. effectively. Therefore, the performance of adaptive MIMO
The fact that SDR beamforming outperforms the SSP method radar processing technique for the SSP approach decreases
of [42] even when the latter is capable of exploiting the full significantly.
MIMO radar DoF stems from the following reasons: 1) The The second simulation is conducted to evaluate the spatial
SSP problem is non-convex and the obtained solution may processing performance of MIMO radar, including the angle
be a local optimum; 2) In the SSP problem, the radar lost estimation accuracy and target detection performance. We
function, defined as k 𝑹 − 𝑹0 k𝐹2 , does not directly reflect the simulate three radar targets located at directions 𝜃 1 = −40◦ ,
performance of radiation beam pattern. 𝜃 2 = 0◦ , and 𝜃 3 = 40◦ , respectively. These targets are in the
Since the beam pattern MSE is not the only performance same range resolution bin and the complex amplitude of the
measure for radar, we also analyze the sensing capabilities at targets are all 1. The targets’ reflected signal is corrupted with
radar receiver. To extract the range and angular profile of radar additive noise whose covariance is 𝑹 𝑣 = 𝜎𝑟2 𝑰. The angle of the
targets from the received radar signal, we first perform range targets is estimated by finding the peaks of the Capon spatial
compression [73] to obtain the range profile, and then use the spectrum. The angle estimation performance is evaluated by
least square (LS) Capon method [74], [75] to calculate the the root-mean-square-error (RMSE), defined as
spatial spectrum in each range resolution bin. v
u
t  Õ
1 3
 


The first simulation is conducted to examine the range RMSE = E (𝜃 𝑝 − 𝜃ˆ 𝑝 ) 2 , (47)
resolution and angular resolution of the MIMO radar. In the 3 
 𝑝=1 
simulation, there are five targets in the field of view of radar.
The coordinate of targets in radar polar coordinate system is where 𝜃 𝑝 is the real angle and 𝜃ˆ 𝑝 is the estimated angle for
defined by the discrete time delay 𝑛 ′ (or the range resolution the 𝑝-th target, for 𝑝 = 1, . . . , 3. The generalized likelihood
bin index) and the angular direction 𝜃 as defined in (9). In our ratio test proposed in [76] is applied to detect the target. To
parameter setting, the coordinate of these targets are (10, 0◦ ), demonstrate the target detection performance, we study the
(20, −40◦ ), (20, 0◦ ), (20, 40◦ ) and (30, 0◦ ), respectively, and relationship between the detection probability and the transmit
the complex amplitude 𝛽 in (9) for each target is 1. The SNR given by 𝑃𝑡 𝑁/𝜎𝑟2 , under a fixed false alarm probability

1053-587X (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Exeter. Downloaded on June 25,2020 at 20:01:43 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.3004739, IEEE
Transactions on Signal Processing
13

1.2 1.2

0.11
1 1

Capon spectrum
Range profile

0.8 0.8 0.1

0.6 0.6
0.09
0.4 0.4

0.2 0.2 0.08

0 0
0 5 10 15 20 25 30 35 40 -90 -60 -30 0 30 60 90 0.07
Range resolution bin index Angle (degree)

(a) (b) 0.06

1.2 1.2
0.05
1 1
Capon spectrum
0.04
Range profile

0.8 0.8

0.6 0.6 0.03

0 2 4 6 8 10 12 14 16 18 20
0.4 0.4

0.2 0.2

0 0 Fig. 11. RMSE for angle estimation with LS-Capon method versus SINR
0 5 10 15 20 25 30 35 40 -90 -60 -30 0 30 60 90

Range resolution bin index Angle (degree) threshold Γ.

(c) (d)
1.2 1.2 1

1 1 0.9
Capon spectrum
Range profile

0.8 0.8
0.8
0.6 0.6
0.7
0.4 0.4

0.6
0.2 0.2

0.5
0 0
0 5 10 15 20 25 30 35 40 -90 -60 -30 0 30 60 90

Range resolution bin index Angle (degree) 0.4

(e) (f) 0.3

0.2
Fig. 10. Capon spatial spectrum at the 20-th range resolution bin and the
range profile at direction 0◦ , for 𝐾 = 2 and Γ = 12 dB. (a) Range profile for 0.1
the radar-only case. (b) Capon spatial spectrum for the radar-only case. (c)
Range profile for the SSP method. (d) Capon spatial spectrum for the SSP 0
10 15 20 25
method. (e) Range profile for the SDR method. (f) Capon spatial spectrum
for the SDR method.

𝑃 𝑓 𝑎 . To calculate the detection probability, we ran 1000 Monte Fig. 12. Detection probability versus transmit SNR under false alarm
probability 𝑃 𝑓 𝑎 = 10−4 , for Γ = 12 dB.
Carlo tests to produce randomized Gaussian noise for each
channel realization, and thus the total number of tests is 106 . Γ = 12 dB and 𝑃 𝑓 𝑎 = 10−4 . From [42], it is noted that there
The numerically evaluated tradeoff between angle estima- exists detection performance loss for simultaneous multiuser
tion RMSE and achieved fairness SINR for SDR beamforming information transmission compared to the radar-only case.
and the SSP DFRC system of [42] is depicted in Fig. 11 for If 𝐾 = 2, the detection performance of SDR beamforming
𝐾 = 2, 4, 6. Here the RMSE of the SSP approach for 𝐾 = 2 is notably outperforms that of the the SSP approach, because
not evaluated since it frequently fails to detect the targets near the SSP approach usually cannot provide enough DoF to form
the true angle direction of the targets. The angle estimation three beams to cover the three target. Hence, reflected signal
RMSE in radar-only case is also displayed for comparison. from one of the targets may experience notable SNR loss,
Observing Fig. 11, we note that the angle estimation RMSE significantly reducing the detection probability. If 𝐾 = 4, 6,
tends to increase with the fairness SINR, again indicating that although the detection performance of SDR beamforming and
the improved communication performance induces some loss the SSP approach is close, we note that the SDR beamforming
on the radar performance. If 𝐾 = 2, the angle estimation can achieve better communication quality. In particular, the
performance of SDR beamforming is almost identical to the SDR beamforming guarantees the achieved fairness SINR
performance in radar-only case, indicating that the proposed is higher than Γ, while in the SSP approach the SINR is
DFRC system achieves angle estimation performance close to considered as a penalty term in the penalty function and
that of the radar-only scheme. The RMSE of angle estimation the obtained SINR at users is generally less than Γ in our
slightly increases if more communication users are under ser- simulation.
vice. It is also noted that under most considered fairness SINR
values, our proposed SDR beamforming achieves improved VI. C ONCLUSION
angle estimation RMSE compared to the SSP method. In this paper, we proposed two joint beamforming ap-
The numerically evaluated detection probability versus proaches for MIMO radar and multiuser MIMO communi-
transmit SNR for SDR beamforming, SSP DFRC system cation sharing spectrum and transmit array. The precoders of
of [42] and the radar-only case is depicted in Fig. 12, for the individual radar waveform and communication symbols

1053-587X (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Exeter. Downloaded on June 25,2020 at 20:01:43 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.3004739, IEEE
Transactions on Signal Processing
14

are designed to optimize the performance of MIMO radar ˜ 𝑹˜ 1 , . . . , 𝑹˜ 𝐾

With the derivation above, it is verified that 𝑹,
transmit beamforming while meeting SINR constraints at com- is a feasible solution, and furthermore, it is also a global
munication users. To solve the proposed optimization problem optimum to (32), completing the proof. 
efficiently, we proposed an algorithm based on SDR, which
allows to accurately recover the optimal beamforming scheme. B. Proof of Theorem 2
To reduce the computational burden of SDR beamforming,
we proposed a reduced complexity sub-optimal strategy based When the conditions (36b) and (37) hold, it follows that
on joint beamforming with zero-forced inter-user and radar 𝑯𝑹𝑯 𝐻 = 𝑯𝑾𝑾 𝐻 𝑯 𝐻 = 𝑭𝑭 𝐻 , (53)
interference. Simulation results showed that the radar beam
patterns and angle estimation performance obtained by the i.e. (38) holds, proving the necessity.
proposed dual-function system is comparable to those of the Next, we prove that condition (38) is also sufficient. Assume
optimal radar-only scheme. We also observed the advantage of that condition (38) holds. We will then construct a 𝑾 that
our method over the previous dual-function work that precodes satisfies (36b) and (37). To this aim, we recall that the QR
only multiuser communication symbols from simulations in decomposition [66] of a 𝑛 × 𝑚 matrix 𝑩 with 𝑛 ≥ 𝑚 is defined
terms of their inherent radar-communication tradeoffs. These as  
performance gains are most notable when the number of ′ 𝑼𝑎
𝑩 = 𝑷 𝑎 𝑼𝑎 = 𝑷 𝑎 , (54)
communication users is relatively small, and particularly less 0 (𝑛−𝑚)×𝑚
than the MIMO radar DoF required to meet the desired where 𝑼𝑎 is a 𝑚 × 𝑚 upper triangular matrix, 𝑷 𝑎′ is a 𝑛 × 𝑚
transmit beam pattern. matrix with orthogonal unit columns, and 𝑷 𝑎 is a 𝑛 × 𝑛 unitary
matrix. Then, define the row QR decomposition of a 𝑚 × 𝑛
A PPENDIX matrix 𝑨 = 𝑩𝑇 with 𝑚 ≤ 𝑛 as
A. Proof of Theorem 1  
𝑨 = 𝑳 𝑎 𝑸 𝑎′ = 𝑳 𝑎 , 0𝑚×(𝑛−𝑚) 𝑸 𝑎 , (55)
Let 𝑹,ˆ 𝑹ˆ 1 , . . . , 𝑹ˆ 𝐾 be an arbitrary global optimum to (32).
˜ 𝑹˜ 1 , . . . , 𝑹˜ 𝐾 from where 𝑳 𝑎 = 𝑼𝑇𝑎 is a 𝑚 × 𝑚 lower triangular matrix, 𝑸 𝑎′ =
We prove the theorem by constructing 𝑹,
ˆ ˆ ˆ (𝑷 𝑎′ )𝑇 is a 𝑚 × 𝑛 matrix with orthogonal unit rows, and 𝑸 𝑎 =
𝑹, 𝑹1 , . . . , 𝑹𝐾 with

𝑷𝑇𝑎 is a 𝑛 × 𝑛 unitary matrix, i.e.
𝑹˜ = 𝑹, ˆ 𝒘˜ 𝑘 = 𝒉 𝐻 𝑹ˆ 𝑘 𝒉 𝑘 −1/2 𝑹ˆ 𝑘 𝒉 𝑘 , 𝑹˜ 𝑘 = 𝒘˜ 𝑘 𝒘˜ 𝐻 , (48)
𝑘 𝑘
𝑸 𝑎 𝑸 𝑎𝐻 = 𝑸 𝑎𝐻 𝑸 𝑎 = 𝑰 𝑛 . (56)
for 𝑘 = 1, . . . , 𝐾. It is clear that 𝑹˜ 1 , . . . , 𝑹˜ 𝐾 are positive
semidefinite and are rank-one. We note that
 
We now show that 𝑹, ˜ 𝑹˜ 1 , . . . , 𝑹˜ 𝐾 is also a global optimum   𝑳 𝑎𝐻
𝑨 𝑨 𝐻 = 𝑳 𝑎 , 0𝑚×(𝑛−𝑚) 𝑸 𝑎 𝑸 𝑎𝐻
to (32). Since the target function 𝐿( 𝑹, 𝛼) is determined by 𝑹 0 (𝑛−𝑚)×𝑚 (57)
and 𝑹ˆ = 𝑹,˜ we only need to validate that 𝑹, ˜ 𝑹˜ 1 , . . . , 𝑹˜ 𝐾 is a = 𝑳 𝑎 𝑳 𝑎𝐻 .
feasible solution to (32).
First, one can derive that Observing the left hand side of (38), we proceed by writing
the Cholesky decomposition of 𝑹 as
𝒉𝐻 ˜ 𝐻
˜ 𝑘 𝒘˜ 𝐻
𝑘 𝑹𝑘 𝒉𝑘 = 𝒉𝑘 𝒘
𝐻 ˆ
𝑘 𝒉𝑘 = 𝒉𝑘 𝑹𝑘 𝒉𝑘 (49)
𝑹 = 𝑳 𝑟 𝑳 𝑟𝐻 , (58)
by substituting (48). Thus
    and writing the row QR decomposition to 𝑯𝑳 𝑟 as
1 + Γ−1 𝒉 𝐻 ˜
𝑘 𝑹𝑘 𝒉𝑘 = 1 + Γ
−1
𝒉𝐻 ˆ
𝑘 𝑹𝑘 𝒉𝑘  
(50) 𝑯𝑳 𝑟 = 𝑳 ℎ , 0𝐾 ×( 𝑀 −𝐾 ) 𝑸 ℎ , (59)
2 2
≥ 𝒉𝐻 ˆ 𝐻 ˜
𝑘 𝑹𝒉 𝑘 + 𝜎 = 𝒉 𝑘 𝑹𝒉 𝑘 + 𝜎 ,
where 𝑸 ℎ is a 𝑀 × 𝑀 unitary matrix and 𝑳 ℎ is a 𝐾 × 𝐾 lower
namely constraint (32e) holds for 𝑹, ˜ 𝑹˜ 1 , . . . , 𝑹˜ 𝐾 . triangular matrix. Applying (57), we rewrite the left hand side
ˆ
Next, we show that 𝑹 𝑘 − 𝑹 𝑘 ∈ S𝑀˜ + . For any 𝒗 ∈ C 𝑀 , it
of (38) as
holds that


−1 𝐻 2
𝑯𝑹𝑯 𝐻 = 𝑯𝑳 𝑟 𝑳 𝑟𝐻 𝑯 𝐻 = 𝑳 ℎ 𝑳 ℎ𝐻 . (60)
𝒗 𝐻 𝑹ˆ 𝑘 − 𝑹˜ 𝑘 𝒗 = 𝒗 𝐻 𝑹ˆ 𝑘 𝒗 − 𝒉 𝐻 ˆ
𝑘 𝑹𝑘 𝒉𝑘 𝒗 𝑹ˆ 𝑘 𝒉 𝑘 . (51)
Similarly, applying row QR decomposition to 𝑭 yields
According to the Cauchy-Schwarz inequality, one has  

𝐻
𝑭 = 𝑳 𝑓 , 0𝐾 ×𝑀 𝑸 𝑓 , (61)
2
𝒉𝐻 ˆ ˆ 𝐻 ˆ
𝑘 𝑹𝑘 𝒉𝑘 𝒗 𝑹𝑘 𝒗 ≥ 𝒗 𝑹𝑘 𝒉𝑘 , (52)

and then
so 𝒗 𝐻 𝑹ˆ 𝑘 − 𝑹˜ 𝑘 𝒗 ≥ 0 holds for any 𝒗 ∈ C 𝑀 , i.e. 𝑹ˆ 𝑘 − 𝑹˜ 𝑘 ∈ 𝑭𝑭 𝐻 = 𝑳 𝑓 𝑳 𝐻𝑓 , (62)
S𝑀+ . It therefore follows that
according to (57). In (61), 𝑸 𝑓 is a (𝑀 + 𝐾)-dimension unitary
𝐾
Õ 𝐾
Õ 𝐾
Õ
+
matrix and 𝑳 𝑓 is a 𝐾 × 𝐾 lower triangular matrix.
𝑹˜ − 𝑹˜ 𝑘 = 𝑹ˆ − 𝑹ˆ 𝑘 + 𝑹ˆ 𝑘 − 𝑹˜ 𝑘 ∈ S𝑀 , Here, we note that both 𝑯𝑹𝑯 𝐻 and 𝑭𝑭 𝐻 are positive
𝑘=1 𝑘=1 𝑘=1
definite given that 𝑭 is a full rank 𝐾 × (𝐾 + 𝑀) matrix ,
˜ 𝑹˜ 1 , . . . , 𝑹˜ 𝐾 . Finally,
namely, the constraint (32b) holds for 𝑹, indicating that the diagonal elements of 𝑳 ℎ and 𝑳 𝑓 are all
ˆ ˜
since 𝑹 = 𝑹, (32c) also holds for 𝑹.˜ non-zero real numbers.

1053-587X (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Exeter. Downloaded on June 25,2020 at 20:01:43 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.3004739, IEEE
Transactions on Signal Processing
15

Since 𝑳 ℎ and 𝑳 𝑓 are lower triangular matrices, we find that [7] ——, “Opportunistic sharing between rotating radar and cellular,” IEEE
(60) is the Cholesky decomposition of 𝑯𝑹𝑯 𝐻 , and (62) is the Journal on Selected Areas in Communications, vol. 30, no. 10, pp. 1900–
1910, November 2012.
Cholesky decomposition of 𝑭𝑭 𝐻 . Since 𝑯𝑹𝑯 𝐻 = 𝑭𝑭 𝐻 and [8] S. Sodagari, A. Khawar, T. C. Clancy, and R. McGwier, “A projection
the Cholesky decomposition of a positive definite matrix is based approach for radar and telecommunication systems coexistence,”
unique [66], we have that in 2012 IEEE Global Communications Conference (GLOBECOM), Dec.
2012, pp. 5010–5014.
𝑳ℎ = 𝑳 𝑓 , (63) [9] J. A. Mahal, A. Khawar, A. Abdelhadi, and T. C. Clancy, “Spectral coex-
istence of MIMO radar and MIMO cellular system,” IEEE Transactions
if we require that the diagonal elements of 𝑳 ℎ and 𝑳 𝑓 are on Aerospace and Electronic Systems, vol. 53, no. 2, pp. 655–668, Apr.
2017.
positive real numbers. [10] H. Deng and B. Himed, “Interference mitigation processing for
We can now construct the matrix 𝑾 as 𝑾 = 𝑳 𝑟 𝑸 𝑤 to spectrum-sharing between radar and wireless communications systems,”
satisfy (36b), where 𝑸 𝑤 is a 𝑀 × (𝑀 + 𝐾) matrix obeying IEEE Transactions on Aerospace and Electronic Systems, vol. 49, no. 3,
pp. 1911–1919, Jul. 2013.
𝑸𝑤 𝑸𝐻 𝑤 = 𝑰 𝑀 . Since we also require 𝑾 to meet (37), the [11] G. Meager, R. A. Romero, and Z. Staples, “Estimation and cancellation
matrix 𝑸 𝑤 should satisfy that of high powered radar interference for communication signal collection,”
in 2016 IEEE Radar Conference (RadarConf), May 2016, pp. 1–4.
𝑯𝑳 𝑟 𝑸 𝑤 = 𝑭. (64) [12] N. Nartasilpa, S. Shahi, A. Salim, D. Tuninetti, N. Devroye, D. Erri-
colo, D. P. Zilz, and M. R. Bell, “Let’s share CommRad: Co-existing
To this aim, the matrix 𝑸 𝑤 is constructed as 𝑸 𝑤 = 𝑸 ℎ𝐻 𝑸ˆ 𝑓 , communications and radar systems,” in 2018 IEEE Radar Conference
h i𝑇 (RadarConf18), Apr. 2018, pp. 1278–1283.
where the 𝑀 × (𝑀 + 𝐾) matrix 𝑸ˆ 𝑓 = 𝑸𝑇𝑓 denotes the [13] N. Nartasilpa, A. Salim, D. Tuninetti, and N. Devroye, “Communications
1:𝑀 system performance and design in the presence of radar interference,”
first 𝑀 rows of 𝑸 𝑓 , and satisfies IEEE Transactions on Communications, vol. 66, no. 9, pp. 4170–4185,
Sep. 2018.
𝑸ˆ 𝑓 𝑸ˆ 𝐻𝑓 = 𝑰 𝑀 (65) [14] M. Rihan and L. Huang, “Optimum co-design of spectrum sharing be-
tween MIMO radar and MIMO communication systems: An interference
according to (56). Thus, the matrix 𝑾 is computed as alignment approach,” IEEE Transactions on Vehicular Technology, pp.
1–1, 2018.
𝑾 = 𝑳 𝑟 𝑸 ℎ𝐻 𝑸ˆ 𝑓 . (66) [15] J. Qian, M. Lops, L. Zheng, X. Wang, and Z. He, “Joint system design
for coexistence of MIMO radar and MIMO communication,” IEEE
Using (66), we can calculate 𝑾 from 𝑹 and 𝑭 with 𝑳 𝑟 , 𝑸 ℎ Transactions on Signal Processing, vol. 66, no. 13, pp. 3504–3519, Jul.
and 𝑸 𝑓 obtained by applying matrix decomposition, i.e. (58), 2018, wOS:000435193800008.
(59) and (61), respectively. [16] B. Li, A. P. Petropulu, and W. Trappe, “Optimum co-design for spectrum
sharing between matrix completion based MIMO radars and a MIMO
To prove (36b) and (36d), we substitute (66) into these two communication system,” IEEE Transactions on Signal Processing,
equations, yielding vol. 64, no. 17, pp. 4562–4575, Sep. 2016.
[17] B. Li and A. P. Petropulu, “Joint transmit designs for coexistence of
(𝑎) (𝑏)
𝑾𝑾 𝐻 = 𝑳 𝑟 𝑸 ℎ𝐻 𝑸ˆ 𝑓 𝑸ˆ 𝐻𝑓 𝑸 ℎ 𝑳 𝑟𝐻 = 𝑳 𝑟 𝑳 𝑟𝐻 = 𝑹, (67) MIMO wireless communications and sparse sensing radars in clutter,”
IEEE Transactions on Aerospace and Electronic Systems, vol. 53, no. 6,
and pp. 2846–2864, Dec. 2017.
[18] F. Liu, C. Masouros, A. Li, T. Ratnarajah, and J. Zhou, “MIMO radar and
(𝑐)  
𝑯𝑾 = 𝑯𝑳 𝑟 𝑸 ℎ𝐻 𝑸ˆ 𝑓 = 𝑳 ℎ , 0𝐾 ×( 𝑀 −𝐾 ) 𝑸 ℎ 𝑸 ℎ𝐻 𝑸ˆ 𝑓 cellular coexistence: A power-efficient approach enabled by interference
(68) exploitation,” IEEE Transactions on Signal Processing, vol. 66, no. 14,
(𝑑)   (𝑒)   (𝑓 ) pp. 3681–3695, Jul. 2018.
= 𝑳 ℎ , 0𝐾 ×( 𝑀 −𝐾 ) 𝑸ˆ 𝑓 = 𝑳 𝑓 , 0𝐾 ×𝑀 𝑸 𝑓 = 𝑭,
[19] F. Liu, L. Zhou, C. Masouros, A. Li, W. Luo, and A. Petropulu, “To-
ward dual-functional radar-communication systems: Optimal waveform
respectively, where (𝑎) follows from (65) and 𝑸 ℎ𝐻 𝑸 ℎ = 𝑰 𝑀 design,” IEEE Transactions on Signal Processing, vol. 66, no. 16, pp.
(56); (𝑏) stems from (58); (𝑐) is due to (59); (𝑑) applies 4264–4279, Aug. 2018.
again 𝑸 ℎ𝐻 𝑸 ℎ = 𝑰 𝑀 ; (𝑒) uses (63); and ( 𝑓 ) follows from (61). [20] D. Ma, N. Shlezinger, T. Huang, Y. Liu, and Y. C. Eldar, “Joint
radar-communications strategies for autonomous vehicles,” IEEE Signal
Therefore, the condition (38) is also sufficient, completing the Processing Magazine, early access, 2020.
proof.  [21] J. Moghaddasi and K. Wu, “Multifunctional transceiver for future radar
sensing and radio communicating data-fusion platform,” IEEE Access,
R EFERENCES vol. 4, pp. 818–838, 2016.
[22] C. Sturm, E. Pancera, T. Zwick, and W. Wiesbeck, “A novel approach
[1] X. Liu, T. Huang, Y. Liu, and J. Zhou, “Joint transmit beamforming to OFDM radar processing,” in 2009 IEEE Radar Conference, 2009.
for multiuser MIMO communication and MIMO radar,” in 2019 IEEE [23] C. Sturm, T. Zwick, and W. Wiesbeck, “An OFDM system concept for
International Conference on Signal, Information and Data Processing joint radar and communications operations,” in VTC Spring 2009 - IEEE
(ICSIDP), December 2019. 69th Vehicular Technology Conference, Apr. 2009, pp. 1–5.
[2] B. Paul, A. R. Chiriyath, and D. W. Bliss, “Survey of RF communi- [24] C. Sturm and W. Wiesbeck, “Waveform design and signal processing
cations and sensing convergence research,” IEEE Access, vol. 5, pp. aspects for fusion of wireless communications and radar sensing,”
252–270, 2017. Proceedings of the IEEE, vol. 99, no. 7, pp. 1236–1259, Jul. 2011.
[3] DARPA, “Shared spectrum access for radar and communications [25] T. Zhang and X. Xia, “OFDM synthetic aperture radar imaging with
(SSPARC),” [Online], 2016, available: https://www.darpa.mil/program/ sufficient cyclic prefix,” IEEE Transactions on Geoscience and Remote
shared-spectrum-access-for-radar-and-communications. Sensing, vol. 53, no. 1, pp. 394–404, Jan. 2015.
[4] Federal Communications Commission, “FCC proposes innovative small [26] X. Liu, Y. Liu, X. Wang, and J. Zhou, “Application of communication
cell use in 3.5 GHz band,” [Online], 2012, available: https://www.fcc. OFDM waveform to SAR imaging,” in 2017 IEEE Radar Conference
gov/document/fcc-proposes-innovative-small-cell-use-35-ghz-band. (RadarConf), May 2017, pp. 1757–1760.
[5] L. Zheng, M. Lops, Y. C. Eldar, and X. Wang, “Radar and communi- [27] C. Sahin, J. Jakabosky, P. M. McCormick, J. G. Metcalf, and S. D. Blunt,
cation co-existence: an overview,” IEEE Signal Processing Magazine, “A novel approach for embedding communication symbols into physical
vol. 36, no. 5, pp. 85–89, 2019. radar waveforms,” in 2017 IEEE Radar Conference (RadarConf), May
[6] R. Saruthirathanaworakun, J. M. Peha, and L. M. Correia, “Opportunistic 2017, pp. 1498–1503.
primary-secondary spectrum sharing with a rotating radar,” in 2012 In- [28] Q. Li, K. Dai, Y. Zhang, and H. Zhang, “Integrated waveform for a
ternational Conference on Computing, Networking and Communications joint radar-communication system with high-speed transmission,” IEEE
(ICNC), Jan 2012, pp. 1025–1030. Wireless Communications Letters, pp. 1–1, 2019.

1053-587X (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Exeter. Downloaded on June 25,2020 at 20:01:43 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.3004739, IEEE
Transactions on Signal Processing
16

[29] C. Sahin, J. G. Metcalf, and S. D. Blunt, “Filter design to address range [52] M. Vu, “MISO capacity with per-antenna power constraint,” IEEE
sidelobe modulation in transmit-encoded radar-embedded communica- Transactions on Communications, vol. 59, no. 5, pp. 1268–1274, May
tions,” in 2017 IEEE Radar Conference (RadarConf), May 2017, pp. 2011.
1509–1514. [53] S. Loyka, “The capacity of Gaussian MIMO channels under total and
[30] Y. Zhang, Q. Li, L. Huang, C. Pan, and J. Song, “A modified waveform per-antenna power constraints,” IEEE Transactions on Communications,
design for radar-communication integration based on LFM-CPM,” in vol. 65, no. 3, pp. 1035–1043, 2017.
2017 IEEE 85th Vehicular Technology Conference (VTC Spring), Jun. [54] D. R. Fuhrmann and G. S. Antonio, “Transmit beamforming for MIMO
2017, pp. 1–5. radar systems using signal cross-correlation,” IEEE Transactions on
[31] X. Wang, A. Hassanien, and M. G. Amin, “Sparse transmit array design Aerospace and Electronic Systems, vol. 44, no. 1, pp. 171–186, Jan.
for dual-function radar communications by antenna selection,” Digital 2008.
Signal Processing, vol. 83, pp. 223–234, DEC 2018. [55] A. El Gamal and Y.-H. Kim, Network information theory. Cambridge
[32] A. Hassanien, E. Aboutanios, M. G. Amin, and G. A. Fabrizio, “A university press, 2011.
dual-function MIMO radar-communication system via waveform per- [56] T. Yoo, N. Jindal, and A. Goldsmith, “Multi-antenna downlink channels
mutation,” Digital Signal Processing, vol. 83, pp. 118–128, DEC 2018. with limited feedback and user selection,” IEEE Journal on Selected
[33] X. Wang, A. Hassanien, and M. G. Amin, “Dual-function MIMO radar Areas in Communications, vol. 25, no. 7, pp. 1478–1491, Sep. 2007.
communications system design via sparse array optimization,” IEEE [57] Y. Liu, Y. Dai, and Z. Luo, “Coordinated beamforming for MISO
Transactions on Aerospace and Electronic Systems, pp. 1–1, 2018. interference channel: Complexity analysis and efficient algorithms,”
IEEE Transactions on Signal Processing, vol. 59, no. 3, pp. 1142–1157,
[34] A. Hassanien, M. G. Amin, Y. D. Zhang, and F. Ahmad, “Dual-function
Mar. 2011.
radar-communications: Information embedding using sidelobe control
[58] A. Wiesel, Y. C. Eldar, and S. Shamai, “Zero-forcing precoding and
and waveform diversity,” IEEE Transactions on Signal Processing,
generalized inverses,” IEEE Transactions on Signal Processing, vol. 56,
vol. 64, no. 8, pp. 2168–2181, Apr. 2016.
no. 9, pp. 4409–4418, Sep. 2008.
[35] A. Hassanien, M. G. Amin, Y. D. Zhang, and F. Ahmad, “Phase-
[59] M. Schubert and H. Boche, “Solution of the multiuser downlink beam-
modulation based dual-function radar-communications,” IET Radar,
forming problem with individual SINR constraints,” IEEE Transactions
Sonar Navigation, vol. 10, no. 8, pp. 1411–1421, 2016.
on Vehicular Technology, vol. 53, no. 1, pp. 18–28, Jan. 2004.
[36] A. Hassanien, M. G. Amin, Y. D. Zhang, and B. Himed, “A dual-function [60] Z.-q. Luo, W.-k. Ma, A. M.-c. So, Y. Ye, and S. Zhang, “Semidefinite
MIMO radar-communications system using PSK modulation,” in Proc. relaxation of quadratic optimization problems,” IEEE Signal Processing
EUSIPCO, Aug. 2016. Magazine, vol. 27, no. 3, pp. 20–34, May 2010.
[37] D. Ma, T. Huang, Y. Liu, and X. Wang, “A novel joint radar and [61] D. P. Palomar and Y. C. Eldar, Convex optimization in signal processing
communication system based on randomized partition of antenna array,” and communications. Cambridge university press, 2010.
in 2018 IEEE International Conference on Acoustics, Speech and Signal [62] L. Vandenberghe and S. Boyd, “Semidefinite programming,” SIAM
Processing (ICASSP), April 2018, pp. 3335–3339. Review archive, vol. 38, no. 1, p. 49, 1996.
[38] T. Huang, N. Shlezinger, X. Xu, D. Ma, Y. Liu, and Y. C. Eldar, “Multi- [63] K. C. Toh, M. J. Todd, and R. H. Tütüncü, “SDPT3 — a Matlab soft-
carrier agile phased array radar,” arXiv preprint arXiv:1906.06289, 2019. ware package for semidefinite programming, version 1.3,” Optimization
[39] T. Huang, N. Shlezinger, X. Xu, Y. Liu, and Y. C. Eldar, “MAJoRCom: Methods and Software, vol. 11, pp. 545–581, 1999.
A dual-function radar communication system using index modulation,” [64] R. H. Tütüncü, K.-C. Toh, and M. J. Todd, “Solving semidefinite-
IEEE Transactions on Signal Processing, early access, 2020. quadratic-linear programs using SDPT3,” Mathematical Programming,
[40] P. M. McCormick, S. D. Blunt, and J. G. Metcalf, “Simultaneous radar vol. 95, no. 2, pp. 189–217, 2003.
and communications emissions from a common aperture, part I: Theory,” [65] K.-C. Toh, M. J. Todd, and R. H. Tütüncü, On the Implementation
in 2017 IEEE Radar Conference (RadarConf), May 2017, pp. 1685– and Usage of SDPT3 – A Matlab Software Package for Semidefinite-
1690. Quadratic-Linear Programming, Version 4.0. Boston, MA: Springer
[41] P. M. McCormick, B. Ravenscroft, S. D. Blunt, A. J. Duly, and US, 2012, pp. 715–754.
J. G. Metcalf, “Simultaneous radar and communication emissions from [66] X. Zhang, Matrix Analysis and Applications. Cambridge, United
a common aperture, part II: Experimentation,” in 2017 IEEE Radar Kingdom: Cambridge University Press, 2017.
Conference (RadarConf), May 2017, pp. 1697–1702. [67] J.-W. Nie and Y.-X. Yuan, “A predictor–corrector algorithm for QSDP
[42] F. Liu, C. Masouros, A. Li, H. Sun, and L. Hanzo, “MU-MIMO commu- combining dikin-type and newton centering steps,” Annals of Operations
nications with MIMO radar: From co-existence to joint transmission,” Research, vol. 103, no. 1, pp. 115–133, Mar. 2001.
IEEE Transactions on Wireless Communications, vol. 17, no. 4, pp. [68] K.-C. Toh, “An inexact primal–dual path following algorithm for convex
2755–2770, Apr. 2018. quadratic SDP,” Mathematical Programming, vol. 112, no. 1, pp. 221–
[43] P. Kumari, J. Choi, N. González-Prelcic, and R. W. Heath, “IEEE 254, Mar. 2008.
802.11ad-based radar: An approach to joint vehicular communication- [69] G. Dimic and N. D. Sidiropoulos, “On downlink beamforming with
radar system,” IEEE Transactions on Vehicular Technology, vol. 67, greedy user selection: Performance analysis and a simple new algo-
no. 4, pp. 3012–3027, Apr. 2018. rithm,” IEEE Transactions on Signal Processing, vol. 53, no. 10, pp.
[44] H. Nyquist, “Certain topics in telegraph transmission theory,” Proceed- 3857–3868, Oct. 2005.
ings of the IEEE, vol. 90, no. 2, pp. 280–305, Feb 2002. [70] T. Yoo and A. Goldsmith, “On the optimality of multi-antenna broadcast
scheduling using zero-forcing beamforming,” IEEE Journal on Selected
[45] S. Guangmin, L. Guosui, and G. Hong, “Signal analysis and process-
Areas in Communications, vol. 24, no. 3, pp. 528–541, Mar. 2006.
ing for random binary phase coded pulse radar,” Journal of Systems
[71] M. Grant and S. Boyd, “CVX: Matlab software for disciplined convex
Engineering and Electronics, vol. 15, no. 4, pp. 520–524, Dec 2004.
programming, version 2.1,” http://cvxr.com/cvx, Mar. 2014.
[46] E. H. Kim and K. H. Kim, “Random phase code for automotive MIMO
[72] ——, “Graph implementations for nonsmooth convex programs,” in
radars using combined frequency shift keying-linear FMCW waveform,”
Recent Advances in Learning and Control, ser. Lecture Notes in Control
IET Radar, Sonar Navigation, vol. 12, no. 10, pp. 1090–1095, 2018.
and Information Sciences, V. Blondel, S. Boyd, and H. Kimura, Eds.
[47] G. V. K. Sharma and K. R. Rajeswari, “Four-phase orthogonal code Springer-Verlag Limited, 2008, pp. 95–110.
design for MIMO radar systems,” in 2012 National Conference on [73] J. Li, L. Xu, P. Stoica, K. W. Forsythe, and D. W. Bliss, “Range com-
Communications (NCC), Feb 2012, pp. 1–4. pression and waveform optimization for MIMO radar: A Cramer–Rao
[48] D. V. Sarwate and M. B. Pursley, “Crosscorrelation properties of bound based study,” IEEE Transactions on Signal Processing, vol. 56,
pseudorandom and related sequences,” Proceedings of the IEEE, vol. 68, no. 1, pp. 218–232, Jan. 2008.
no. 5, pp. 593–619, May 1980. [74] L. Xu, J. Li, and P. Stoica, “Radar imaging via adaptive MIMO
[49] S. W. Golomb and G. Gong, Signal Design for Good Correlation: For techniques,” in Proc. EUSIPCO, 2006.
Wireless Communication, Cryptography, and Radar. Cambridge, United [75] ——, “Target detection and parameter estimation for MIMO radar
Kingdom: Cambridge University Press, 2005. systems,” IEEE Transactions on Aerospace and Electronic Systems,
[50] P. Stoica, J. Li, and Y. Xie, “On probing signal design for MIMO radar,” vol. 44, no. 3, pp. 927–939, Jul. 2008.
IEEE Transactions on Signal Processing, vol. 55, no. 8, pp. 4151–4161, [76] I. Bekkerman and J. Tabrikian, “Target detection and localization using
Aug. 2007. MIMO radars and sonars,” IEEE Transactions on Signal Processing,
[51] W. Yu and T. Lan, “Transmitter optimization for the multi-antenna vol. 54, no. 10, pp. 3873–3883, Oct. 2006.
downlink with per-antenna power constraints,” IEEE Transactions on
Signal Processing, vol. 55, no. 6, pp. 2646–2660, June 2007.

1053-587X (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Exeter. Downloaded on June 25,2020 at 20:01:43 UTC from IEEE Xplore. Restrictions apply.