Received February 22, 2019, accepted February 28, 2019, date of publication March 6, 2019, date of current version

March 29, 2019.

Digital Object Identifier 10.1109/ACCESS.2019.2903166

Hybrid Precoding for Massive

mmWave MIMO Systems
XIANRU LIU 1 , XUEMING LI2 , SHU CAO3 , QINGYONG DENG 4,5 ,
RONG RAN 6 , (Member, IEEE), KIEN NGUYEN 7 , (Senior Member, IEEE),
AND PEI TINGRUI3
1 School of automation, Central South University, Changsha 410086, China
2 School of Information and Communication Engineering, Beijing University of Posts and Telecommunications, Beijing 100876, China
3 Key Laboratory of Hunan Province for Internet of Things and Information Security, Xiangtan University, Xiangtan 411105, China
4 School of Information and Communication Engineering, Beijing University of Posts and Telecommunications, Beijing 100876, China
5 Key Laboratory of Hunan Province for Internet of Things and Information Security, Xiangtan University, Xiangtan 411105, China
6 School of Electrical and Computer Engineering, Ajou University, Suwon 16499, South Korea
7 Graduate School of Engineering, Chiba University, Chiba 263-8522, Japan

Corresponding author: Qingyong Deng ([email protected])

This work was supported in part by the Open Project Program of the Key Laboratory of Universal Wireless Communications under Grant
2016-KFKT-2016104, in part by the Ministry of Education, Beijing University of Posts and Telecommunications, in part by the Natural
Science Foundation of China under Grant 61771414, Grant 61711540306, and Grant 61672447, and in part by the Natural Science
Foundation of Hunan Province of China under Grant 2017JJ2249 and Grant 2018JJ1025.

ABSTRACT Due to high costs and power consumptions, fully digital baseband precoding schemes are
usually prohibitive in millimeter-wave massive MIMO systems. Therefore, hybrid precoding strategies
become promising solutions. In this paper, we present a novel real-time yet high-performance precoding
strategy. Specifically, the eigenvectors corresponding to the larger eigenvalues of the right unitary matrix
after singular value decomposition on an array response matrix are used to abstract the angle information of
an analog precoding matrix. As the obtained eigenvectors correspond to the larger singular values, the major
phase information of channels is captured. In this way, the iterative search process for obtaining the analog
precoding vectors is avoided, and thus the hybrid precoding can be realized in parallel. To further improve
its spectral-efficiency, we enlarge the resultant vector set by involving more relevant vectors in terms of their
correlation values with the unconstrained optimal precoder, and a hybrid precoder is thus produced by using
the vector set. The simulation results show that our proposed scheme achieves near the same performance
as the orthogonal matching pursuit does, whereas it costs much fewer complexities than the OMP, and thus
can be realized in parallel.

INDEX TERMS Millimeter wave communication, MIMO, wireless communication, hybrid precoding.

I. INTRODUCTION To alleviate the shortage of frequency resource, millimeter

As the capacity demand on wireless communication is wave (mmWave) is deemed as a potential radio frequency
increasing rapidly, the communication systems operating in candidate. Due to its gigabit-per-second data rates [6]–[9],
microwave band are difficult to meet the quality-of-service mmWave will be deployed in the future fifth generation
(QoS) requirement. Hence, in the past several years, many (5G) wireless communication systems. Compared to the
advanced technologies such as relay cooperation, diversity currently used frequency bands, the main characteristic of
multiplexing and cognitive radio are utilized to enhance spec- mmWave communication is its tremendous increase in carrier
tral efficiency [1]–[5]. However, compared to the explosive frequency. It implies that mmWave signals experience an
increasing demand on system capacity, the obtained gain is orders-of-magnitude fading while propagating through the
too limited. Accordingly, exploring other radio frequency space. Therefore, it is difficult to ensure enough signal-to-
bands seems to be the essential solution to mitigate the noise ratio (SNR) for the conventional MIMO transceiver
crowded frequency bands. architecture. Fortunately, the substantially reduced wave-
length enables the facility to put large antenna arrays
The associate editor coordinating the review of this manuscript and in a much smaller space. Hence, mmWave systems can
approving it for publication was Wei Xu. integrate massive multiple-input multiple-output (MIMO)
2169-3536 2019 IEEE. Translations and content mining are permitted for academic research only.
VOLUME 7, 2019 Personal use is also permitted, but republication/redistribution requires IEEE permission. 33577
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 X. Liu et al.: Hybrid Precoding for Massive mmWave MIMO Systems

transceiver elements to enhance signal gain and spectral connected structure for hybrid beamforming is proposed,
efficiency. Consequently, massive mmWave MIMO architec- where precoding matrix optimization problem is decomposed
ture has been adopted as a standard in 5G communication into multiple precoding sub-matrix optimization problems
systems [10]–[12], and becomes one of the research hotspots and near-optimal hybrid digital and analog precoders are
in the wireless communication domain. designed through factorizing the precoding sub-matrix for
For a massive mmWave MIMO system, to decrease the each sub-array. In [24], channel matrix is decomposed into an
complexity of the receiver and cancel the interference among angle domain basis matrix and a gain matrix, which perfectly
data streams, signals should be precoded before transmission. matches the structure of hybrid precoding. In [25], it is shown
When fully digital baseband precoding is used, it is not that the Gram matrix of frequency-selective channel can
difficult to precode multiple data streams to remove the inter- be decomposed into frequency-flat and frequency-selective
signal interference. However, an individual radio frequency components, which can be used for the analog and baseband
(RF) chain is required for each antenna, which is impractical precoders, respectively. All these research work are focused
when the number of antennas is large. On the other hand, on decomposing the channel matrix of the mmWave large-
when a pure analog precoding mode is taken, RF chain can scale MIMO systems through some mathematical methods.
be shared among different antennas, and the cost in terms As the constraints of analog matrix should be considered
of hardware and power consumption can be reduced greatly. during decomposition, the computational complexity is very
However, it is only applicable to the scenarios with single data high.
stream, so it is not beneficial to improve spectral efficiency. To decrease the complexity of hybrid precoder design,
Consequently, the methodology of combining the digital and some low-complexity hybrid precoding schemes have been
analog modes to form a hybrid precoding scheme is intro- proposed. A new view of multi-user hybrid massive mMIMO
duced into massive mmWave MIMO systems. It not only systems from array signal processing perspective is also pre-
determines precoders’ structure, but also affects system per- sented in [24], and a hybrid precoding and channel estima-
formances [13]. Hence, a hybrid precoder which can achieve tion strategies are proposed. In [26], an explicit analytical
good performance has attracted a lot of research attentions. sum-rate expression for generic channel covariance-based
In the past several years, much research are carried out beamforming scheme is derived, and a low-complexity
around massive MIMO. In [14], the minimal requested num- joint statistical beamforming and user scheduling algorithm
ber of RF chains and phase shifters in massive MIMO sys- is proposed. In [27], exploiting the spatial structure of
tem is analyzed. For distributed massive MIMO systems, mmWave channels, precoding/combining is formulated as a
the benefits of beamforming training scheme are analyzed sparse reconstruction problem, and a precoding algorithm is
in [15], and its spectral efficiency with linear beamforming developed to accurately approximate optimal unconstrained
under pilot contamination is presented in [16]. For mixed- precoder and combiner with low-cost RF hardware. Depend-
ADC/DAC multipair massive MIMO relaying systems, both ing on the known array geometry and a low training and
exact and approximate closed-form expressions of the achiev- feedback, a hybrid analog/digital precoding algorithm is pro-
able rate are derived and a power scaling law is presented posed in [28]. An angle-domain hybrid precoding is pro-
in [17]. For the similar system model, when two-way relaying posed in [29], and a novel channel tracking strategy is
with hardware impairments is considered, an approximation also presented. After the capacity optimization problem is
of the spectral efficiency with maximum ratio processing decomposed into a series of subproblems that are easier to
is derived in closed form in [18]. In [19], the design of be handled by considering each antenna array one by one,
directional hybrid precoding for multiuser mmwave com- a near-optimal iterative hybrid precoding scheme is proposed
munication systems with multiple eavesdroppers is investi- in [30] that is based on a more realistic subarray structure.
gated. To maximize the sum rate and the energy efficiency of An orthogonal matching and local search method is proposed
massive MIMO systems, a codebook based hybrid precoding to implement analog precoding in [31]. Based on the structure
scheme with low complexity is proposed in [20]. of mmWave channels, the design of a hybrid precoder is
In mmWave massive MIMO systems, the essence of hybrid formulated as a sparsity constrained least squares problem,
precoding is to use two constrained matrices to approach and a precoding algorithm is proposed to approximate the
the fully digital precoding matrix or inverted channel matrix. optimal unconstrained precoder in [32]. A hybrid precoding
Accordingly, the intuitive solution of hybrid precoding is processor based on parallel-index-selection and orthogonal
to decompose the fully digital precoding/channel matrix matching pursuit without matrix inversion is proposed in [33].
under given constraints. By directly decomposing the pre- For multiuser massive MIMO systems, a hybrid block diag-
designed unconstrained digital precoder/combiner, a hybrid onalization scheme is proposed to approach the capacity per-
RF and baseband precoding/combining scheme for multi- formance of the traditional baseband digital method in [34].
stream transmission is proposed in [21]. In [22], an iterative By exploring the idea of turbo equalizer together with the tabu
precoding algorithm is proposed for mmWave systems, where search algorithm, in [35], a novel hybrid precoding scheme is
all precoding operations are done in analog domain, and proposed to achieve near-optimal performance. In [36], after
mmWave system can adapt the rank of its transmission in the relevance of directional precoding structures is estab-
response to varying propagation conditions. In [23], a hybrid lished as a low-complexity and robust solution to meet the

33578 VOLUME 7, 2019

X. Liu et al.: Hybrid Precoding for Massive mmWave MIMO Systems

demands of data rate, a simple class of directional schedulers The rest of the paper is organized as follows. In Section II,
is proposed based on the channel directional structure, and after the system model of mmWave large-scale MIMO is
then the performance comparison between the single-user briefly introduced, the hybrid precoding problem is formu-
and multi-user scenarios is provided. lated. The proposed hybrid precoding scheme in parallel and
Though the performance of the aforementioned hybrid its improved version are presented specifically in Section III.
precoding schemes is very close to the optimal unconstrained We evaluate the performance of our solution schemes in
precoding, most of them require an iterative search. Hence, Section IV. Finally, Section V concludes the paper and
the real-time performance is not very good. In this paper, after presents our future research.
analyzing the relationship between antenna response matrix Notations: A is a matrix; a is a vector; a or A is a scalar;
and its right singular-value-decomposition (SVD) matrix, A(i) or A(i) is the ith column of A; (•)0 , (•)∗ , and (•)−1 denote
the most related vectors are selected out, which serves as the transpose, conjugate transpose, and inversion, respec-
analog precoder after each element is normalized. To further tively; E[•] denotes expectation operation.
improve system spectral efficiency, the selected vector set
is enlarged such that the most relevant base vector can be II. SYSTEM DESCRIPTION
involved in term of the correlation with the fully digital In this section, we first describe our considered system
precoding matrix. In this way, the iterative search manner model. Then, the characteristic of mmWave MIMO channel is
is not required any longer, and then the hybrid precoding analyzed.
scheme can achieve high spectral efficiency with a parallel
manner. The main contributions of this paper are summarized
as follows. A. SYSTEM MODEL
• As a baseline of hybrid precoder design, SVD is adopted We consider a single-user mmWave system in which one
to obtain the phase information of mmWave large- base station with NBS antennas transmits Ns data streams to a
scale MIMO channel. After SVD is carried out on receiver with NMS antennas simultaneously. To support the
the array response matrix, the right resultant matrix function of transferring multiple data streams in a parallel
consists of a group orthogonal basis, which can be mode, the transmitter and receiver are equipped with multiple
used to form array response vectors. Hence, the prob- radio frequency (RF) chains. The number of RF chains are
lem of designing hybrid precoding is converted into denoted as NRF and MRF , and they satisfy Ns ≤ NRF ≤ NBS
searching the most relevant basis vectors, which is and Ns ≤ MRF ≤ NMS constraints, respectively.
reflected on the value of eigenvalues, and can be realized In this paper, we focus on the precoder design at the
easily. transmitter due to the reciprocities between the transmitting
• Instead of taking advantage of the traditional optimiza- process and receiving process. The similar approach can
tion techniques used in hybrid precoding, such as OMP, be utilized in obtaining the combiner at the receiver, and
OMP+LS, the proposed hybrid precoding scheme is omitted for brevity. For the Ns transmitted data streams,
utilizes SVD on array response matrix, and the most before going through RF chains, they should be precoded
relevant antenna response vectors are selected as ana- in the baseband. Denote the baseband precoder, analog pre-
log precoding matrix directly after being normalized. coder and transmitted symbol vector as FBB , FRF , and s,
In this way, the iterative search process, which is a respectively, and their dimensions are NRF × Ns , NBS × NRF
major ingredient that leads high computational complex- and Ns × 1. Suppose the transmit power of s is normalized,
ity, is avoided, and a hybrid precoder can be obtained i.e., E[ss∗ ] = N1s INs . After precoding in the baseband domain,
directly. Consequently, hybrid precoding is promising to the resultant vector given as FBB s is the input of NRF RF
realize in parallel mode. chains for the up conversion. Next, an analog precoder FRF
• Unlike the schemes in [30], [31], and [27], where is applied for adjusting phase/angle to maximize the system
the analog precoding vectors are obtained one by one capacity or minimize the interference according to the system
through an iterative OMP, the proposed scheme can requirements. Therefore, the discrete-time transmitted sig-
select out all the analog precoding vectors simultane- nals is finally represented as x = FRF FBB s. As the precoder
ously. Consequently, the execution time should be short FRF achieved by an analog phase shifter only adjusts the
(i) (i)∗
so that the massive mmWave MIMO system becomes phases rather than the amplitudes of signals, (FRF FRF )l,l =
−1 (i)
more real-time. Nt should hold for any i, where FRF denotes the ith column
• The accuracy of hybrid precoding relative to the OMP vector of FRF and (.)l,l denotes the l th diagonal element of a
scheme can be controlled through setting the number of matrix. Meanwhile, the transmitted signals should guarantee
radio frequency chains. As we know, if all the right nor- the power constraint, which results in ||FRF FBB ||2F = Ns .
malized eigenvectors corresponding to nonzero eigen- For narrowband block-fading propagation scenarios,
values are selected out as analog precoding vectors, the received signals at the receiver can be shown as (1), where
all the antenna response vectors can be reconstructed. y is the NMS ×1 received signal vector, H is the channel matrix
Hence, the required number of radio frequency chains with NMS × NBS dimensions, ξ is the average received power,
can be determined according to the requested accuracy. and n is the additive white Gaussian noise (AWGN) vector

VOLUME 7, 2019 33579

 X. Liu et al.: Hybrid Precoding for Massive mmWave MIMO Systems

following i.i.d distribution CN(0, σn2 ). given as (5), where W and H are the numbers of antenna
elements, k = 2π λ (λ is the corresponding wavelength) and
y = ξ HFRF FBB s + n.
p
(1)
d is the inter-element spacing between two neighbor anten-
Without loss of generality, we assume that the channel nas, N = WH is the size of the antenna array. For simplicity,
matrix H is known at both the transmitter and receiver.1 we use At and Ar to represent the matrices consisting of
At the receiver(s), the received signals are processed in the all array response vectors at (•) and ar (•) respectively. These
similar way to recover the original transmitted data streams. matrix representations will be used in the sequel.
Accordingly, the finally processed signals can be shown 1
as (2), where WBB and WRF denote the baseband and analog aUPA (φ, θ) = √ [1, · · · , ekd(msin(φ)sin(θ )+ncos(θ )) ,
N
combiners, respectively.
· · · , ekd((W −1)sin(φ)sin(θ )+(H −1)cos(θ )) ]T . (5)
ŷ = ξ W∗BB W∗RF HFRF FBB s + W∗BB WRF ∗ n.
p
(2)
III. HYBRID PRECODING STRATEGY
Similar to the precoder, the analog combiner WRF at
In this section, after analyzing the relation between the
the receiver only provides the phase adjustments. So the
channel matrix and antenna array response vectors, the
elements of WRF should have the same amplitudes, i.e.
(i) (i)∗ problem of hybrid precoding is formulated. Then, a novel
(WRF WRF )l,l = Nr−1 . Given the system model (2), its
low-complexity hybrid precoding strategy and its improved
spectral efficiency is described as (3) [27], where Rn =
version are proposed.
σn2 W∗BB W∗RF WRF WBB is the noise covariance matrix after
the received signals are combined at the receiver. Therefore,
A. PROBLEM FORMULATION
for mmWave massive MIMO systems, the task of hybrid
precoding is to maximize R through designing FBB , FRF , To maximize R in (3), FBB , FRF , WBB and WRF should be
WBB and WRF jointly. optimal. Nevertheless, due to the non-convex constraints of
FRF and WRF , it is extremely difficult to tackle the optimiza-
ξ −1 ∗ tion problem maximizing R via FRF , WBB and WRF . Con-
R = log2 (|INs + R WBB W∗RF HFRF FBB
Ns n sequently, the problem of designing analog/digital precoders
× F∗BB F∗RF H∗ WRF WBB |). (3) and combiners is usually decoupled into two independent
sub-problems. One is designing precoders for the transmitter,
B. CLUSTER-BASED CHANNEL MODEL another is designing combiners for the receiver, and both of
Due to the characteristics of severe path-loss and high level them follow the similar approach to obtain solutions. At the
of correlation among antennas, the traditional channel model transmitter, for data streams s, the achievable mutual infor-
is not applicable for mmWave large-scale MIMO transceiver mation of the mmWave massive MIMO system can be shown
array. Consequently, a narrow-band clustered channel rep- as (6).
resentation based on the extended Saleh-Valenzuela model ξ
is usually adopted [38]–[40], where the channel matrix H I(FRF , FBB ) = log2 (|I + HFRF FBB F∗BB F∗RF H∗ |). (6)
NS σn2
is assumed to be a sum of the contributions of Ncl dif-
ferent clusters, each of which has Nray propagation paths. To maximize I(FRF , FBB ) in (6), the classical methods usu-
Accordingly, the channel model of mmWave massive MIMO ally consist of two stages. First, considering the entirety
systems can be abstracted as (4), where γ is a normalization of FRF FBB and omitting all hardware limitations, the mutual
factor, αil is the complex gain of the l th propagation path in the information can be maximized by Fopt = V(:, 1 : Ns ),
ith cluster. Functions 3r (•) and 3t (•) represent the gains of where the SVD of H is U6V∗ . Fopt is the fully optimal
transmit and receive antenna, respectively, that correspond to baseband precoding matrix, which consists of the Ns col-
different departure and arrival angles. ar (•) and at (•) denote umn vectors in V that correspond to the highest singular
the normalized respective receive and transmit array response values in 6. Second, considering the hardware constraints,
vectors when azimuth and elevation angles vary, and they are FRF and FBB are designed using some optimization strategies
independent [27]. Moreover, φilr (θilr ) and φilt (θilt ) present the to approach Fopt . The commonly used method of designing
arrival and departure azimuth (elevation) angles, respectively, FRF is to select the most relevant array response vectors. As a
of the l th ray in the ith cluster. result, we formulate the optimization problem as [27],
X
H=γ αi,l 3r (φilr , θilt )3t (φilr , θilt )ar (φilr , θilt )at (φilr , θilt )∗ . arg min {||Fopt − At F̂BB ||F }
At ,F̂BB
i,l
(4) subject to ||diag(F̂BB F̂∗BB )||0 = NRF

Note that when uniform planar array structure is consid- ||At F̂BB ||2F = Ns . (7)
ered, both ar (•) and at (•) can be expressed as the format where At is the antenna response matrix with the dimension
1 For practical mmWave massive MIMO systems, the channel state infor- of NBS × Ncl Nray (Ncl and Nray are the numbers of scattering
mation (CSI) can be obtained through channel estimation technology at cluster and propagation paths per cluster, respectively, in clus-
receiver, and feedback to transmitter. tered channel model, which will be described briefly in the

33580 VOLUME 7, 2019

X. Liu et al.: Hybrid Precoding for Massive mmWave MIMO Systems

next subsection), and || • ||0 denotes pseudo norm returning and an approximation analog precoder can be constructed.
the number of non-zero elements. Implicitly, by searching the In this way, hybrid precoding can be realized without using
most suitable antenna response vectors in At , the constraints any iterative search process, and computational complexity
of the analog precoder can be neglected. can be reduced greatly.
Once the analog precoder FRF FBB has been designed, For antenna response matrix A0t , when it is decomposed
the combiner WBB WRF at the receiver can be obtained via the using SVD, it can be shown as A0t = U6V∗ , where
similar approach. Let us assume Wopt = WBB WRF and relax the dimensions of A0t , U, 6 and V are Ncl Nray × NBS ,
the hardware limitations, the optimal Wopt can be achieved Ncl Nray ×Ncl Nray , Ncl Nray ×NBS and NBS ×NBS , respectively.
by minimizing the mean squared error (MMSE) between To see more clear relationship between A0t and V, let us
the transmitted and processed received signals [37]. Then, assume β = U6. Then, let β i , Ai and Vi denote the
the problem of designing hybrid combiners can be formulated ith row/row/column of β, A0t and V, respectively, for the sake
as (8), where y is the received signal vector, and RF is a of convenience. Consequently, the following relation can be
constant-gain phase-only entries with NMS × MRF size. obtained.
arg min {||E[yy∗ ]1/2 Wopt − WRF WBB ||F } Ai = [β i V1∗ β i V∗2 , · · · , β i V∗NBS ] = β i [V∗1 V∗2 , · · · , V∗NBS ].
WRF ,WBB
subject to WRF ∈ RF . (8) (10)

Note that the process to solve (8) and (6) at the transmitter It implies that the orthogonal basis V∗i ∈ ∀i can be used to
and receiver are similar, and hence we only present the strat- express Ai . For (10), as β i can be removed into the part of
egy to design hybrid precoder in this paper, and the method digital precoder, the remaining task of the hybrid precoding
to optimize the hybrid combiners is omitted for the sake of is to select out the most relevant orthogonal vectors from V.
brevity. Therefore, searching the most relevant vectors in At succes-
Note that the column vectors of the fully digital precod- sively is equivalent to obtaining the most suitable orthogonal
ing matrix Fopt are orthogonal [27]. However, they are not basis which can be realized in the parallel model.
necessarily complete orthogonal bases for the channel’s row For V∗i , though its norm is one, the norms of its element is
space since some eigenvectors corresponding to non-zero not necessary one. Hence, its elements should be normalized.
eigenvalues are not selected out due to the constraint in the let V∗i = [V1i , V2i , · · · , VNBS i ]. Then, it can be rewritten as
number column vectors (only NS column vectors are taken
[V1i , V2i , · · · , VNBS i ]
out). On the other hand, from (4), it can be seen that the
V1i V2i VNBS i
linear combination of all array response vectors at (φilt , θilt ) =[ , ,··· , ]
forms the channel matrix H. Hence, it can be concluded that |V1i | |V2i | |VNBS i |
|V1i | ×0 × · · · ,
 
the linear combination of at (φilt , θilt ) can be used to present ×0
Fopt accurately. Therefore, we can argue that hybrid precoder  0 ×|V2i | × · · · , ×0 
× · · · × · · · × · · · , × · · · , (11)

design is to find two smaller matrices FRF and FBB such that
Fopt = FRF FBB . Moreover, FBB should be unit-amplitude 0 ×0 × · · · , ×|VNBS i |
which implies that the column vectors of FBB is able to V1i VN
, V2i , · · · , |VNBS i | ] can be realized with the
i
present Fopt . Design (7) can be equally converted into the in which [ |V 1i | |V2i | BS
following problem analog phase shifter, and
|V1i | 0 · · · ,
opt opt
 
(FRF , FBB ) = arg min||Fopt − FRF FBB ||F , 0
(i)  0 |V2i | · · · , 0 
subject to FRF ∈ {at (φilt , θilt ), ∀i, l},  ··· ··· ··· , ··· 
 
||FRF FBB ||2F = Ns . (9) 0 0 · · · , |VNBS i |

B. SOLUTION STRATEGY can be removed from the part of baseband precoder. From the
To solve the optimization problem in (9), its essence is to above analysis, if all the vectors corresponding to non-zero
select out the NRF most relevant antenna response vectors eigenvalues are taken out for the construction of analog phase
from {at (φilt , θilt ), ∀i, l}, which is denoted as At for nota- shifter, the performance of the hybrid precoding in this way is
tional simplicity. In the existing hybrid precoding schemes, equal to that of the OMP scheme. Obviously, as the number of
optimal matching strategies are commonly used, such RF chains is limited,2 the construction of At is probably not
as [27] and [31]. Though the performance of these hybrid perfect. Consequently, its performance is degraded compared
precoding schemes is very close to that of the fully digital to the optimal iterative search scheme.
baseband precoding scheme, its computational complexity Based on the above analysis, the proposed hybrid precod-
is exponentially increasing with the number of transceiver ing scheme can be summarized as Algorithm 13 .
antennas, which leads to huge latency. In fact, for the antenna 2 It can be seen that the number of RF chains is equal to the number of the
response matrix At , if an orthogonal basis can be used to selected vectors from At .
present it, the corresponding phase information is obtained, 3 norm(•) function means normalizing each elements of the matrix.

VOLUME 7, 2019 33581

 X. Liu et al.: Hybrid Precoding for Massive mmWave MIMO Systems

Algorithm 1 Low-Complexity Hybrid Precoding Algorithm

1: Obtain Fopt through SVD on H.
2: A0t = U6V∗
3: FRF = [V1 , V2 , · · · , VNRF ]∗ .
4: FRF = norm(FRF )
5: FBB = (F∗RF FRF )−1 F∗RF Fopt
6: return FRF , FBB

Algorithm 2 The Improved Low-Complexity Hybrid Precod-

ing Process
1: Require : Fopt .
2: A0t = U6V∗ .
3: F̃RF = [V1 , V2 , · · · , Vn ]∗ .
4: 9 = F̃0RF Fopt .
5:  = diag(99 ∗ ). FIGURE 1. The spectral efficiency of four precoding schemes with the
6: [Y , I ] = sort(,0 descend 0 ). varying SNR, where Ns = 1 and Ns = 3 are set under the conditions
NBS = 64, NMS = 16 and NRF = MRF = 4, respectively.
7: F̂RF = F̃RF (:, I (1 : NRF )).
8: FRF = norm(FRF ˆ ).
9: FBB = (F∗RF FRF )−1 F∗RF Fopt .
In the simulation, the cluster-based channel model is consid-
10: return FRF , FBB .
ered, in which Ncl = 8 and there are Nray = 10 rays per
cluster. Both the arrival and departure azimuth and elevation
angles of each cluster follows the Laplacian distribution. Sim-
For mmWave massive MIMO transceiver, as NRF << NBS ilar to [27] and [41], the spread angle in azimuth and elevation
and NRF << Ncl Nray , it is incomplete that the NRF most angles are also set as 20◦ and 60◦ , respectively. Moreover, it is
relevant orthogonal basis vectors are used to approximate also assumed that all clusters have the equal transmit power.
the antenna response matrix At . Furthermore, in the OMP In practical mmWave massive MIMO systems, the base sta-
hybrid precoding schemes, just the NRF relevant antenna tions are usually more capable to configure larger massive
response vectors in At are selected out as the hybrid precoder. transceiver antenna array compared to the mobile terminals.
Consequently, there is a gap between the performance of Hence, in our simulation, we let the number of antennas at
Algorithm 1 and that of the fully digital baseband precoding the base station be larger than that of the mobile user, and
scheme. As a compromise, an enlarged candidate orthogonal the scenarios with NBS = 64, NMS = 16 and NBS = 256,
basis set is considered next, which means more orthogonal NMS = 64 are considered.
basis vectors are selected according to the correlation with the Fig. 1 and Fig. 2 show the spectral efficiency with the
fully digital baseband precoder Fopt . This idea is detailedly varying SNR under the condition that NBS = 64, NMS = 16
demonstrated in Algorithm 2, where n is the parameter used for NRF = MRF = 4 and NRF = MRF = 8, respectively. For
to set the scale of the candidate orthogonal basis vectors and the purpose of comparison, the simulation results of the fully
NRF ≤ n ≤ NBS . digital precoding and OMP schemes are presented as well.
In order to identify the impact of the number of data streams
C. REAL-TIME ANALYSIS on the system spectral efficiency, we also consider Ns = 1
To see the computational complexity of the proposed hybrid and 3 cases.
precoding scheme, we compare it with the classical hybrid From Fig. 1 and Fig. 2, we can observe that the spectral
precoding schemes in [27] and [31]. From Algorithm 1 and efficiency achieved by Algorithm 1 and Algorithm 2 approach
Algorithm 2, it can be seen that most of the computing to that of the OMP scheme with respect to the number of
operations are concentrated on the SVD, sorting and matrix RF chains. For example, in Fig. 2, with NRF = MRF = 8,
inversion (but just be required once). The traditional hybrid the spectral performance of Algorithm 2 is almost as same
precoding schemes, such as OMP and OMP+LS, usually as that of the OMP and Fully Digital Precoding schemes.
consist of an iterative search process, in which sorting, matrix That is because, the more RF chains are available, the more
inversion and normalization operations are repeated for each basis vectors can be used to generate the more accurate
iteration, and are time-consuming. Consequently, the pro- representative matrix of the antenna response matrix At .
posed hybrid precoding is more real-time than the traditional It is not surprise that Algorithm 2 always outperforms
schemes. Algorithm 1. But we notice that their performance gap
decreases as the number of RF chain increases. The essen-
IV. PERFORMANCE EVALUATION tial of our proposed schemes is how to select the relevant
In this section, numerical simulations are presented to show basis vectors to construct a new matrix which can accurately
the performance of our proposed hybrid precoding schemes. represent the antenna response matrix At . In Algorithm 1,

33582 VOLUME 7, 2019

X. Liu et al.: Hybrid Precoding for Massive mmWave MIMO Systems

FIGURE 2. The spectral efficiency of four precoding schemes with the

FIGURE 4. The spectral efficiency of four precoding schemes with the
varying SNR, where Ns = 1 and Ns = 3 are set under the conditions
varying SNR, where Ns = 1 and Ns = 3 are set under the conditions
NBS = 64, NMS = 16 and NRF = MRF = 8, respectively.
NBS = 256, NMS = 64 and NRF = MRF = 8, respectively.

NRF = 4 and NRF = 8, respectively. In Fig. 3, we can observe

that the performance gap between the proposed Algorithm 2
and the OMP gets bigger compared to Fig. 1 due to the
increased number of antennas. It is because that when the size
of the antenna array increases, the number of transmission
paths increase. Consequently, the components that construct
At becomes more complex, and it is more difficult to con-
struct At by using the orthogonal vectors. However, when
we increase the number of RF chains, as illustrated in Fig. 4,
Algorithm 2 approaches the OMP for both Ns = 1 and 3 cases,
while costs much less complexities.
In Fig. 5, we consider three cases that Ns = 1, 2 and 3
under the conditions that NBS = 64, NMS = 16 and
NRF = MRF = 8. From simulation results, we can observe
FIGURE 3. The spectral efficiency of four precoding schemes with the
varying SNR, where Ns = 1 and Ns = 3 are set under the conditions that the spectral efficiencies of the proposed two algorithms
NBS = 256, NMS = 64 and NRF = MRF = 4, respectively. increase in term of the number of data streams, which fol-
lows the general behaviors of MIMO systems. In Fig. 6,
we preserve NRF = MRF = 8 and only increase the
the number of basis vectors of At , which is the same as that of numbers of transmit and receive antennas up to 256 and 64,
RF chains, is used to produce an analogy precoder. Since the respectively. The spectral efficiency of the proposed
channel matrix H consists of vectors of At , the optimal anal- Algorithm 2 increases as the number of data streams
ogy precoder should be heavily related to At . But the number increases. But the proposed Algorithm 1 with Ns = 3
of RF chain limits the number of basis vector candidates does not achieve the best performance. That is because the
which we can select to construct the representative matrix, precoder via Algorithm 1 with the degree of freedom of
there is not guarantee that the selected vectors can contribute NRF = MRF = 8 is not robust enough to deal with additional
enough to produce a good precoder. However, in Algorithm 2, interference caused by the increased data streams.
a larger number of eigenvectors are allowed to be correlated Therefore, in order to learn the impact of the number of
with the optimal precoder, so more vector candidates can be data streams on the spectral efficiency thoroughly, we illus-
used to determine the most relevant basis vectors. Therefore, trate the spectral efficiencies with respect to the number of
a better precoder is possible to be produced compared to data streams when SNR is given as 0dB in Fig. 7. From simu-
Algorithm 1, which is confirmed by simulation results. Fur- lation results, we can observe that the fully digital precoding
thermore, when the number of RF chains increases, the prob- scheme and OMP follow the general behaviors of MIMO sys-
ability of selecting the most relevant basis vectors becomes tems that the spectral efficiency near linearly increases as the
larger, and hence the spectral efficiency gap is reduced. number of data streams at the middle or high SNR regimes.
To show the impact of antenna array size on the spectral Our proposed Algorithm 2 behaves the same as the OMP
efficiency, the simulation results for the cases NBS = 256 when Ns is less than 6, but starts to degrade when Ns is over 6,
and NMS = 64 are presented in Fig. 3 and Fig. 4 for and the proposed Algorithm 1 degrades from Ns = 2. It is not

VOLUME 7, 2019 33583

 X. Liu et al.: Hybrid Precoding for Massive mmWave MIMO Systems

FIGURE 5. The spectral efficiency of the proposed hybrid precoding FIGURE 7. The spectral efficiency of four precoding schemes with the
schemes with the varying SNR, where Ns = 1, Ns = 2 and Ns = 3 are set increasing Ns , where SNR = 0dB is set under the conditions NBS = 256,
under the conditions NBS = 64, NMS = 16 and NRF = MRF = 8, NMS = 64 and NRF = MRF = 8, respectively.
respectively.

FIGURE 8. The spectral efficiency with the increasing number of selected

array response vectors to correlate with the fully digital precoding matrix,
FIGURE 6. The spectral efficiency of the proposed hybrid precoding where SNR = −10dB and SNR = 0dB are set under the conditions
schemes with the varying SNR, where Ns = 1, Ns = 2 and Ns = 3 are set NBS = 256, NMS = 64 and NRF = MRF = 8, respectively.
under the conditions NBS = 256, NMS = 64 and NRF = MRF = 8,
respectively.

consider NBS = 256, NMS = 64, NRF = MRF = 8 and

difficult to understand that given At , the selected orthogonal Ns = 3. From simulation results, we can see that when the
vectors which are relevant to Fopt , become the loose represen- number of candidate vectors is greater than 12, the spectral
tatives of the vectors of Fopt when the number of data stream efficiencies achieved by Algorithm 2 approach to a constant.
increases. Therefore, the interference cancellation capability For NRF = MRF = 8, only four additional vector candi-
of the precoder is degraded, which results in the perfor- dates are added to the selection set and obtain near-optimal
mance loss. The issue can be addressed by increasing the performance, which causes negligible complexity increments
number of candidate vectors used to generate the precoder compared to Algorithm 1.
(i.e., Algorithm 2), but the complexity is increased as well.
There arises a tradeoff between the interference cancellation V. CONCLUSION
capability of the precoder via Algorithm 2 and the number In this paper, we propose a novel hybrid precoding scheme
of basis vectors used to generate the representative matrix. for a mmWave massive MIMO system with the cluster-
Namely, a tradeoff between performance and complexity. based channel model. Unlike the existing precoding schemes,
Fortunately, since NRF ≥ 2 Ns is required to approach which generally require an iterative and successive search
the optimal according to the results given in [14], our pro- processes to achieve the analog phase shifters, our proposed
posed Algorithm 2 is implementable for the most reasonable scheme exploits the correlation values between the eigen-
scenarios. vectors of the array response matrix and the digital optimal
Finally, we illustrate the impact of the number of candidate precoder in selecting the most relevant basis vectors which
vectors used in producing precoders in Fig. 8, where we still are used to reconstruct the representative matrix of the array

33584 VOLUME 7, 2019

X. Liu et al.: Hybrid Precoding for Massive mmWave MIMO Systems

response matrix. Therefore, the iterative search process is no [19] Y. Huang, J. Zhang, and M. Xiao, ‘‘Constant envelope hybrid precoding for
longer required and the proposed scheme can significantly directional millimeter-wave communications,’’ IEEE J. Sel. Areas Com-
mun., vol. 36, no. 4, pp. 845–859, Apr. 2018.
reduce the complexity and be realized in parallel. Meanwhile, [20] S. He, J. Wang, Y. Huang, B. Ottersten, and W. Hong, ‘‘Codebook-based
the proposed scheme achieves a comparable performance hybrid precoding for millimeter wave multiuser systems,’’ IEEE Trans.
as OMP. Signal Process., vol. 65, no. 20, pp. 5289–5304, Oct. 2017.
[21] W. Ni, X. Dong, and W. S. Lu, ‘‘Near-optimal hybrid processing for
massive MIMO systems via matrix decomposition,’’ IEEE Trans. Signal
REFERENCES Process., vol. 65, no. 15, pp. 3922–3933, Aug. 2017.
[1] L. Mohjazi, M. Dianati, G. K. Karagiannidis, S. Muhaidat, and [22] O. E. Ayach, R. W. Heath, Jr., S. Rajagopal, and Z. Pi, ‘‘Multimode pre-
M. Al-Qutayri, ‘‘RF-powered cognitive radio networks: Technical chal- coding in millimeter wave MIMO transmitters with multiple antenna sub-
lenges and limitations,’’ IEEE Commun. Mag., vol. 53, no. 4, pp. 94–100, arrays,’’ in Proc. IEEE Global Commun. Conf. (GLOBECOM), Atlanta,
Apr. 2015. GA, USA, Dec. 2013, pp. 3476–3480.
[2] S. Wang, R. Ruby, V. C. M. Leung, and Z. Yao, ‘‘A low-complexity power [23] D. Zhang, Y. Wang, X. Li, and W. Xiang, ‘‘Hybridly connected structure for
allocation strategy to minimize sum-source-power for multi-user single- hybrid beamforming in mmWave massive MIMO systems,’’ IEEE Trans.
AF-relay networks,’’ IEEE Trans. Commun., vol. 64, no. 8, pp. 3275–3283, Commun., vol. 66, no. 2, pp. 662–674, Feb. 2018.
Aug. 2016. [24] H. Lin, F. Gao, S. Jin, and G. Y. Li, ‘‘A new view of multi-user
[3] J. Mietzner, R. Schober, L. Lampe, W. H. Gerstacker, and P. A. Hoeher, hybrid massive MIMO: Non-orthogonal angle division multiple access,’’
‘‘Multiple-antenna techniques for wireless communications—A compre- IEEE J. Sel. Areas Commun., vol. 35, no. 10, pp. 2268–2280, Oct. 2017.
hensive literature survey,’’ IEEE Commun. Surveys Tuts., vol. 11, no. 2, [25] W. M. Chan, T. Kim, H. Ghauch, and M. Bengtsson, ‘‘Subspace estimation
pp. 87–105, 2nd Quart., 2009. and hybrid precoding for wideband millimeter-wave MIMO systems,’’ in
Proc. 50th Asilomar Conf. Signals, Syst. Comput., Pacific Grove, CA,
[4] Q. Deng et al., ‘‘Dynamic spectrum sharing for hybrid access in
USA, Nov. 2016, pp. 286–290.
OFDMA-based cognitive femtocell networks,’’ IEEE Trans. Veh. Tech.,
[26] C. Zhang, Y. Huang, Y. Jing, S. Jin, and L. Yang, ‘‘Sum-rate anal-
vol. 67, no. 11, pp. 10830–10840, Nov. 2018.
ysis for massive MIMO downlink with joint statistical beamforming
[5] S. Wang, R. Ruby, V. C. M. Leung, Z. Yao, X. Liu, and Z. Li, ‘‘Sum- and user scheduling,’’ IEEE Trans. Wireless Commun., vol. 16, no. 4,
power minimization problem in multisource single-AF-relay networks: pp. 2181–2194, Apr. 2017.
A new revisit to study the optimality,’’ IEEE Trans. Veh. Technol., vol. 66, [27] O. El Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath, Jr.,
no. 11, pp. 9958–9971, Nov. 2017. ‘‘Spatially sparse precoding in millimeter wave MIMO systems,’’ IEEE
[6] P. B. Papazian, G. A. Hufford, R. J. Achatz, and R. Hoffman, ‘‘Study Trans. Wireless Commun., vol. 13, no. 3, pp. 1499–1513, Mar. 2014.
of the local multipoint distribution service radio channel,’’ IEEE Trans. [28] A. Alkhateeb, R. W. Heath, Jr., and G. Leus, ‘‘Achievable rates of multi-
Broadcast., vol. 43, no. 2, pp. 175–184, Jun. 1997. user millimeter wave systems with hybrid precoding,’’ in Proc. IEEE Int.
[7] T. S. Rappaport, G. R. MacCartney, M. K. Samimi, and S. Sun, ‘‘Wideband Conf. Commun., London, U.K., Jun. 2015, pp. 1232–1237.
millimeter-wave propagation measurements and channel models for future [29] J. Zhao, F. Gao, W. Jia, S. Zhang, S. Jin, and H. Lin, ‘‘Angle domain
wireless communication system design,’’ IEEE Trans. Commun., vol. 63, hybrid precoding and channel tracking for millimeter wave massive MIMO
no. 9, pp. 3029–3056, Sep. 2015. systems,’’ IEEE Trans. Wireless Commun., vol. 16, no. 10, pp. 6868–6880,
[8] R. C. Daniels and R. W. Heath, Jr., ‘‘60 GHz wireless communica- Oct. 2017.
tions: Emerging requirements and design recommendations,’’ IEEE Veh. [30] L. Dai, X. Gao, J. Quan, S. Han, and C.-L. I, ‘‘Near-optimal hybrid
Technol. Mag., vol. 2, no. 3, pp. 41–50, Sep. 2007. analog and digital precoding for downlink mmWave massive MIMO sys-
[9] J. G. Andrews, T. Bai, M. N. Kulkarni, A. Alkhateeb, A. K. Gupta, and tems,’’ in Proc. IEEE Int. Conf. Commun., London, U.K., Jun. 2015,
R. W. Heath, Jr., ‘‘Modeling and analyzing millimeter wave cellular sys- pp. 1334–1339.
tems,’’ IEEE Trans. Commun., vol. 65, no. 1, pp. 403–430, Jan. 2017. [31] C. Rusu, R. Méndez-Rial, N. González-Prelcic, and R. W. Heath, Jr., ‘‘Low
[10] A. I. Sulyman, A. T. Nassar, M. K. Samimi, G. R. MacCartney, Jr., complexity hybrid sparse precoding and combining in millimeter wave
T. S. Rappaport, and A. Alsanie, ‘‘Radio propagation path loss models for MIMO systems,’’ in Proc. IEEE Int. Conf. Commun. (ICC), London, U.K.,
5G cellular networks in the 28 GHz and 38 GHz millimeter-wave bands,’’ Jun. 2015, pp. 1340–1345.
IEEE Commun. Mag., vol. 52, no. 9, pp. 78–86, Sep. 2014. [32] O. E. Ayach, R. W. Heath, Jr., S. Abu-Surra, S. Rajagopal, and Z. Pi, ‘‘Low
[11] M. Elkashlan, T. Q. Duong, and H.-H. Chen, ‘‘Millimeter-wave commu- complexity precoding for large millimeter wave MIMO systems,’’ in Proc.
nications for 5G: Fundamentals: Part I [guest editorial],’’ IEEE Commun. IEEE Int. Conf. Commun. (ICC), Canada, U.K., Jun. 2012, pp. 3724–3729.
Mag., vol. 52, no. 9, pp. 52–54, Sep. 2014. [33] Y.-Y. Lee, C.-H. Wang, and Y.-H. Huang, ‘‘A hybrid RF/baseband precod-
ing processor based on parallel-index-selection matrix-inversion-bypass
[12] J. G. Andrews et al., ‘‘What will 5G be?’’ IEEE J. Sel. Areas Commun.,
simultaneous orthogonal matching pursuit for millimeter wave MIMO sys-
vol. 32, no. 6, pp. 1065–1082, Jun. 2014.
tems,’’ IEEE Trans. Signal Process., vol. 63, no. 2, pp. 305–317, Jan. 2015.
[13] J. Zhang, L. Dai, X. Li, Y. Liu, and L. Hanzo, ‘‘On low-resolution ADCs
[34] W. Ni and X. Dong, ‘‘Hybrid block diagonalization for massive multiuser
in practical 5G millimeter-wave massive MIMO systems,’’ IEEE Commun.
MIMO systems,’’ IEEE Trans. Commun., vol. 64, no. 1, pp. 201–211,
Mag., vol. 56, no. 7, pp. 205–211, Jul. 2018.
Jan. 2016.
[14] T. E. Bogale, L. B. Le, A. Haghighat, and L. Vandendorpe, ‘‘On the number [35] X. Gao, L. Dai, C. Yuen, and Z. Wang, ‘‘Turbo-like beamforming based
of RF chains and phase shifters, and scheduling design with hybrid analog- on tabu search algorithm for millimeter-wave massive MIMO systems,’’
digital beamforming,’’ IEEE Trans. Wireless Commun., vol. 15, no. 5, IEEE Trans. Veh. Technol., vol. 65, no. 7, pp. 5731–5737, Jul. 2016.
pp. 3311–3326, May 2016. [36] V. Raghavan, S. Subramanian, J. Cezanne, A. Sampath,
[15] J. Li, D. Wang, P. Zhu, and X. You, ‘‘Benefits of beamforming training O. H. Koymen, and J. Li, ‘‘Single-user versus multi-user precoding for mil-
scheme in distributed large-scale MIMO systems,’’ IEEE Access, vol. 6, limeter wave MIMO systems,’’ IEEE J. Sel. Areas Commun., vol. 35, no. 6,
pp. 7432–7444, 2018. pp. 1387–1401, Jun. 2017.
[16] J. Li, D. Wang, P. Zhu, J. Wang, and X. You, ‘‘Downlink spectral efficiency [37] T. Kailath, A. H. Sayed, and B. Hassibi, Linear Estimation.
of distributed massive MIMO systems with linear beamforming under pilot Upper Saddle River, NJ, USA: Prentice-Hall, 2000.
contamination,’’ IEEE Trans. Veh. Technol., vol. 67, no. 2, pp. 1130–1145, [38] H. Xu, V. Kukshya, and T. S. Rappaport, ‘‘Spatial and temporal character-
Feb. 2018. istics of 60-GHz indoor channels,’’ IEEE J. Sel. Areas Commun., vol. 20,
[17] J. Zhang, L. Dai, Z. He, B. Ai, and O. A. Dobre, ‘‘Mixed-ADC/DAC no. 3, pp. 620–630, Apr. 2002.
multipair massive MIMO relaying systems: Performance analysis and [39] A. M. Sayeed, ‘‘Deconstructing multiantenna fading channels,’’ IEEE
power optimization,’’ IEEE Trans. Commun., vol. 67, no. 1, pp. 140–153, Trans. Signal Process., vol. 50, no. 10, pp. 2563–2579, Oct. 2002.
Jan. 2019. [40] V. Raghavan and A. M. Sayeed, ‘‘Sublinear capacity scaling laws for sparse
[18] J. Zhang, X. Xue, E. Björnson, B. Ai, and S. Jin, ‘‘Spectral effi- MIMO channels,’’ IEEE Trans. Inf. Theory, vol. 57, no. 1, pp. 345–364,
ciency of multipair massive MIMO two-way relaying with hardware Jan. 2011.
impairments,’’ IEEE Wireless Commun. Lett., vol. 7, no. 1, pp. 14–17, [41] Z. Pi and F. Khan, ‘‘An introduction to millimeter-wave mobile broadband
Feb. 2018. systems,’’ IEEE Commun. Mag., vol. 49, no. 6, pp. 101–107, Jun. 2011.

VOLUME 7, 2019 33585

 X. Liu et al.: Hybrid Precoding for Massive mmWave MIMO Systems

XIANRU LIU received the master’s and Ph.D. RONG RAN (M’05) received the Ph.D. degree
degrees in control science and engineering from from the Department of Electrical and Electronic
Central South University, Changsha, China, Engineering, Yonsei University, in 2009. In 2009,
in 2006 and 2011, respectively, where she is she joined ETRI, Daejeong, and has worked on
currently an Instructor. Her research interests the IEEE 802. 16m standardization. She was a
include computer vision and image understanding, Research Associate with The Hong Kong Uni-
machine learning, wireless cooperative communi- versity of Science and Technologies, in 2010.
cation, and cognitive networks. Since 2014, she has been a Faculty of the Electri-
cal and Computer Engineering Department, Ajou
University. Her current research interests include
5G wireless communications, machine/deep learning, and sparse signal pro-
XUEMING LI received the B.E. degree in elec-
cessing. She is a member of the KICS.
tronics engineering from the University of Sci-
ence and Technology of China, in 1992, and the
Ph.D. degree in electronics engineering from the
Beijing University of Posts and Telecommunica-
tions (BUPT), in 1997. From 1997 to 1999, he
was a Postdoctoral Researcher with the Institute of
Information Science, Beijing Jiaotong University.
He has been with BUPT, since 1999. In 2002, he
was a Guest Lecturer with Karlsruhe University,
Germany. His current research interests include digital image processing, KIEN NGUYEN (S’08–M’12–SM’16) received
video coding, and multimedia telecommunication. He has undertaken many the B.E. degree in electronics and telecommu-
state and enterprise Research and Development projects. He has authored nication from the Hanoi University of Science
three books and over 50 papers in the field of multimedia information and Technology, Vietnam, and the Ph.D. degree
processing and transmission. He is currently a Senior Member of the Chinese in informatics from the Graduate University for
Institute of Electrics, and a Senior Member of the China Society of Image Advanced Studies, Hayama, Japan, in 2004 and
and Graphics. 2012, respectively. He joined the National Insti-
tute of Information Communications Technology,
SHU CAO received the bachelor’s degree in Yokosuka, Japan, in 2014, as a Researcher. His
electronic science and technology from the research interests include novel, software-based,
Hunan Institute of Engineering, China, in 2016. evolvable networking technologies for the next generation of mobile net-
Her research interests include millimeter-wave works and the Internet of Things. He is a member of the IEICE.
communication, massive MIMO, and machine
learning.

QINGYONG DENG received the B.Eng. degree in

electrical information engineering, and the M.Eng. PEI TINGRUI has majored in signal and informa-
degree in signal and information processing from tion processing and received the Ph.D. degree from
Xiangtan University, in 2004 and 2009, respec- the Beijing University of Posts and Telecommuni-
tively. He is currently pursuing the Ph.D. degree cations, in 2004. He visited Japan as a Researcher,
with the Beijing University of Posts and Telecom- from 2006 to 2007. He is currently a Professor and
munications. Since 2004, he has been a Lec- a Doctoral Supervisor. He focuses on the research
turer with the College of Information Engineering, of wireless sensor networks, ad hoc, mobile com-
Xiangtan University. His current researches focus munication networks, and social computing.
on cognitive networks, machine-to-machine com-
munication, smart grid, and wireless communication.

33586 VOLUME 7, 2019