886 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 13, NO.

5, SEPTEMBER 2019

An Overview of Enhanced Massive MIMO With

Array Signal Processing Techniques
Mingjin Wang , Student Member, IEEE, Feifei Gao , Senior Member, IEEE, Shi Jin , Member, IEEE,
and Hai Lin , Senior Member, IEEE

Abstract—In the past ten years, there have been tremendous re- I. INTRODUCTION
search progresses on massive MIMO systems, most of which stand
HE fifth generation (5G) wireless communications de-
from the communications viewpoint. A new trend to investigate
massive MIMO, especially for the sparse scenario like millimeter
wave (mmWave) transmission, is to re-build the transceiver design
T mands for a substantial increase in transmission throughput
and network coverage in order to support a broad range of
from array signal processing viewpoint that could deeply exploit emerging applications, such as smart phones, multimedia, social
the half-wavelength array and provide enhanced performances in networks, internet gaming, etc. One promising physical layer
many aspects. For example, the high dimensional channel could be technology of 5G is the massive multiple-input multiple-output
decomposed into small amount of physical parameters, e.g., angle of (MIMO) that scales up conventional MIMO by several orders
arrival (AoA), angle of departure (AoD), multi-path delay, Doppler of magnitude and advocates the use of a few hundred or thou-
shift, etc. As a consequence, transceiver techniques like synchro- sand antennas at the base station (BS) to greatly increase sys-
nization, channel estimation, beamforming, precoding, multi-user
access, etc., can be re-shaped with these physical parameters, as
tem capacity over available time-frequency resources [1]–[3].
opposed to those designed directly with channel state information It promises to reap all benefits of conventional MIMO, and
(CSI). Interestingly, parameters like AoA/AoD and multi-path de- hence significantly improves the spectrum and power efficien-
lay are frequency insensitive and thus can be used to guide the cies of wireless communications. In recent years, there has been
downlink transmission from uplink training even for FDD systems. tremendous theoretical research on massive MIMO systems,
Moreover, some phenomena of massive MIMO that were vaguely including capacity analysis [4], channel estimation [5], synchro-
revealed previously can be better explained now with array signal nization [6], beamforming technique [7], user scheduling [8],
processing, e.g., the beam squint effect. In all, the target of this spectral efficiency [9], etc.
paper is to present an overview of recent progress on merging array Practically, massive MIMO system chooses the antenna spac-
signal processing into massive MIMO communications as well as ing as half-wavelength to maintain an implementable array
its promising future directions.
aperture, which is quite different from the convention where
Index Terms—Massive MIMO, array signal processing, the antennas are often placed further away from each other to
mmWave, angle-delay-Doppler reciprocity, beam squint. achieve spatial diversity. With half-wavelength arrays, neverthe-
less, the array signal processing techniques for phased array in
Radar and Sonar applications [10]–[13] can possibly be applied
to enhance the quality of wireless data communications.
Array signal processing has long been used in military ap-
Manuscript received February 20, 2019; revised June 16, 2019; accepted July plications to extract the angles of Radar targets or to formulate
19, 2019. Date of publication August 12, 2019; date of current version September
20, 2019. This work was supported in part by the National Natural Science
narrow beams for jamming/anti-jamming [14]. The difference
Foundation of China under Grant 61831013 and Grant 61771274, in part by the from wireless communications is that the information contained
Beijing Municipal Natural Science Foundation under Grant 4182030 and Grant in signals is not cared but rather the AoA that represents the
L182042, in part by the R&D Project of the Science, Technology and Innovation target’s position. In fact, the famous term beamforming was
Commission of Shenzhen Municipality No. JCYJ20180306170617062, in part originated from array signal processing that means physically
by the National Science Foundation (NSFC) for Distinguished Young Scholars
of China under Grant 61625106, and in part by the National Natural Science formulating an electromagnetic beam towards the target. Later
Foundation of China under Grant 61531011. The guest editor coordinating on, the term beamforming was replanted in wireless communi-
the review of this article and approving it for publication was Prof. Zhi Tian. cations to represent the weights of multiple antennas that can
(Corresponding author: Feifei Gao.) maximize the signal-to-noise ratio (SNR) of a user while it does
M. Wang is with the Department of Automation, Tsinghua University, Beijing
100084, China (e-mail: [email protected]).
not necessarily formulate a physical beam over the space.
F. Gao is with the Department of Automation, Tsinghua University, Beijing The first trial to combine the area of array signal processing
100084, China, with the Institute for Artificial Intelligence, Tsinghua University, and wireless communications was probably the smart antennas
Beijing 100084, China, with the Beijing National Research Center for Informa- [15] that arose around the mid-to-late 1990’s when the Array-
tion Science and Technology (BNRist), Beijing 100084, China, and also with Comm deployed angle-based spatial division multiple access
the Key Laboratory of Digital TV System of Guangdong Province and Shenzhen
City, Research Institute of Tsinghua University in Shenzhen, Shenzhen 518057, (SDMA) base station (BS) in both Japan and Australia. Unfor-
China (e-mail: [email protected]). tunately, this first trial was not that successful due to two-fold
S. Jin is with the National Communications Research Laboratory, Southeast reasons: (i) Unlike in radar environment, microwave wireless
University, Nanjing 210096, China (e-mail: [email protected]). communications have too many multi-paths from reflection
H. Lin is with the Department of Electrical and Information Systems Graduate
School of Engineering, Osaka Prefecture University, Sakai 5998531, Japan
and scattering such that the AoAs of incoming paths are not
(e-mail: [email protected]). distinguishable; (ii) The number of antennas in conventional
Digital Object Identifier 10.1109/JSTSP.2019.2934931 MIMO is small, say 4 or 8, which cannot support accurate AoA
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
WANG et al.: OVERVIEW OF ENHANCED MASSIVE MIMO WITH ARRAY SIGNAL PROCESSING TECHNIQUES 887

estimation while would bring large side lobes when formulating II. CHANNEL CHARACTERISE OF LARGE ANTENNA ARRAY
the physical beams. Let us first briefly present how the wireless communica-
This time, owing to the large antenna size, massive MIMO
tions channels can be modeled from array signal processing
is well positioned for successfully deploying array signal pro- viewpoint. For ease of illustration, we consider the massive
cessing techniques since the spatial information of users can be MIMO systems with M  1 receive antennas at the BS in
precisely identified and very narrow beams can be formulated
the form of a uniform linear array (ULA), while the related
to multiplex different users. Meanwhile, the tendency to high discussion can be readily extended to planar array cases. The
frequency communications, e.g., millimeter wave (mmWave) carrier frequency is denoted by fc , and the carrier wavelength is
would face gigantic path loss such that the total number of
effective paths is small. Particularly, technical terms in array denoted by λc  c/fc , where c is the speed of light. The antenna
signal processing (e.g., angle of arrival/departure - AoA/AoD), spacing is denoted by d  λc /2. Meanwhile, we assume the
can be conceptualized in the terms of wireless communications, communications system has symbol period T and bandwidth W .
such as channel estimation, hybrid beamforming, interference
control, multiple access scheme, etc., which suggests an exciting A. Channel Modeling
research direction of leveraging advanced array signal process- Suppose a continuous-time passband signal R{s(t)ej2πfc t }
ing techniques to aid wireless communications. In fact, since is sent from far field user and arrives at BS via P multi-paths
massive MIMO is inherently built upon the multi-antenna array, (uplink). Each path can be characterized by four parameters:
it is expected that the mature array signal processing technolo-
the AoA θp ∈ [−π/2, π/2), the passband path gain β̃p ∈ R, the
gies could be exploited deeply into this architecture and provide
propagation delay to the first receive antenna τp , and the Doppler
more reliable designs as well as the enhanced performance
shift νp ∈ [−νmax /2, νmax /2), respectively. Here, τmax and
compared to those directly obtained from the communications
νmax are termed as the delay spread and (two-sided) Doppler
viewpoint.
spread of the channel, respectively. Moreover, the pth path ar-
In all, the purpose of this article is to provide an overview (m−1)d sin θp
of the state-of-the-art results that apply array signal processing rives at the mth receive antenna Δτm,p = c seconds
techniques for massive MIMO communication systems. We after it arrives at the first receive antenna, where c is the speed of
intend not to duplicate those transceiver techniques that have light. The received passband continuous-time signal at the mth
already been introduced in [16]–[19] from communications antenna can be described by [20], [21]1
viewpoint but rather focus on new designs from array signal 
P 
processing viewpoint. ỹm (t) = R β̃p s(t − (τp + Δτm,p ))
The paper is organized as follows. Section II presents the p=1
channel models of massive MIMO communication systems 
from array signal processing viewpoint where the channels ej2π(fc +νp )(t−(τp +Δτm,p )) . (1)
can be divided into spatial/frequency and narrowband/wideband
categories. A detailed overview of parameter estimation, both Removing the carrier frequency ej2πfc t , the baseband received
blind and training based, is provided in Section III, where it is signal is
pointed out that parameters like AoA/AoD, multi-path delay, and
Doppler shift are widely frequency insensitive and is helpful for 
P
 
downlink channel estimation even in FDD systems. Section IV ym (t) = βp e−j2π(fc +νp )Δτm,p ej2πνp t s t − τp − Δτm,p
p=1
discusses the synchronization issue in massive MIMO systems.

Section V describes angle-domain hybrid precoding and inter-
ference control architectures. We then introduce orthogonal time = hm (t, τ )s(t − τ )dτ, (2)
and frequency space (OTFS) modulation in massive MIMO
systems with angle-delay-Doppler 3D structured channels in where βp = β̃p e−j2π(fc +νp )τp is the equivalent baseband path
Section VI. With physical parameters, we present in Section VII gain, while the last equation represents the linear time varying
three new multiple access technologies and illustrate their rela- filter with the equivalent time varying channel impulse response
tionship with conventional time/frequency/spatial division mul- (CIR)
tiple access schemes. Section VIII investigates some artificial
intelligence (AI)-based signal processing techniques for massive 
P
hm (t, τ ) = βp e−j2π(fc +νp )Δτm,p
MIMO systems. Section IX concludes the paper and outline
p=1
future directions to further merge array signal processing with
massive MIMO communications. e j2πνp t
δ(τ − τp − Δτm,p ). (3)
Notations: Throughout this paper, vectors and matrices are
denoted by boldface lower-case and upper-case letters, respec- Taking the Fourier Transform of hm (t, τ ) over τ , we obtain
tively; the transpose, Hermitian, inverse, and pseudo-inverse of the frequency domain time varying channel response at the mth
the matrix A are denoted by AT , AH , A−1 and A† , respectively; antenna as
vec(A) represents column-major vectorization of the matrix A, 
P
i.e., the operation of stacking the columns of matrix A to form hm (t, f ) = βp e−j2π(fc +νp )Δτm,p ej2πνp t e−j2πf (τp +Δτm,p ) .
a vector; h0 , h1 , and h2 are the 0 -norm, 1 -norm and p=1
2 -norm of vector h, respectively; ∗ denotes the convolution (4)
operator; R represents the set of real numbers; C represents
the set of complex numbers; R(·) is the real part of a complex 1 When P goes to infinite, the sum function can be replaced by integration,
number, vector or matrix. i.e., the incident angles are in the continuous domain.
888 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 13, NO. 5, SEPTEMBER 2019

TABLE I
CLASSIFICATION OF WIRELESS CHANNELS ON THE BASIS OF
CHANNEL AND SIGNALING PARAMETERS

and the equivalent time domain MIMO channel vector is3


P
h(t, τ ) = βp a(θp , 0, νp )ej2πνp t δ(τ − τp ). (8)
Fig. 1. Illustration of beam squint effect in large-scale antenna array. p=1

It can be easily checked that h(t, f ) does not have beam squint
Stacking the channel from all M antennas, the frequency domain effect anymore.
time varying channel response vector of the whole antenna array
can be expressed as [20], [21]: C. Frequency Narrowband Modeling
When the bandwidth W is small or the multi-path delay is

P
small such that τp  T , for ∀p, p ∈ {1, . . . , P } (the multi-path
−j2πf τp j2πνp t
h(t, f ) = βp a(θp , f, νp )e e , (5) delay is negligible compared to symbol period), (5) reduces to
p=1

P

where a(θp , f, νp ) is the M × 1 wideband array steering vector: h(t, f ) = βp a(θp , f, νp )ej2πνp t , (9)
p=1
 f +νp d sin θp
a(θp , f, νp ) = 1, e−j2π(1+ fc ) λc , . . . , and the case is called frequency narrowband modeling. In-
terestingly, though the frequency dependent factor e−j2πf τp
f +νp (M −1)d sin θp T disappears (this is how the conventional frequency flat fading es-
e−j2π(1+ fc ) λc . (6) tablishes for small MIMO), the channel response vector h(t, f )
is still varying with frequency f due to the beam squint effect.
In orthogonal frequency division multiplexing (OFDM) sys- In this case, the inter symbol interference (ISI) in time domain
tem, the signal bandwidth is divided into N subcarriers. Ac- still exists and the OFDM modulation with certain amount of
cording to (6) the steering vectors at different subcarriers are cyclic prefix (CP) should, again, be applied [31].
affected by their frequencies, i.e., the steering vectors are dif-
ferent for different subcarriers even with the same AoAs. This D. Spatial and Frequency Narrowband Modeling
phenomenon is known as the beam squint effect [22]–[25], as If both τp W  1 and Δτm,p W  1 hold, for ∀p, p ∈ {1,
shown in Fig. 1.2 . . . , P }, ∀m, m ∈ {1, . . . , M }, then the channel response vec-
tor reduces to
B. Spatial Narrowband Modeling P
h(t) = βp a(φp , 0, νp )ej2πνp t , (10)
When the signal bandwidth W or the number of antennas M is p=1
small, such that Δτm,p  T , for ∀p, p ∈ {1, . . . , P }, ∀m, m ∈
{1, . . . , M } (the propagation delay across the array is negligible which yields the same frequency domain channels for all subcar-
compared to symbol period), the signals on all antennas can riers and truly presents to the flat fading environment. This case
be approximated by s(t − τp − Δτp,m ) ≈ s(t − τp ); namely, is then named as spatial- and frequency- narrowband modeling.
different receive antennas effectively observe the synchronized In comparison, the very general channel model (3) and (5) for
waveform. This case is called as narrowband modeling for array massive MIMO are named as spatial- and frequency-wideband
signal processing but should actually be more accurately named model or dual-wideband model [31], [32].
as spatial narrowband modeling in wireless communications.
In this case, one can stack the time domain baseband received E. Summary of Wideband-Narrowband Channel Modeling
signal from all M antennas and obtain the conventional massive According to the physical parameters, e.g., AoA, delay, and
MIMO model as [4]–[7], [26]–[30]: Doppler shifts, the channel of massive MIMO system can be
  classified into different categories, as shown in Tab. I:
y(t) = [y1 (t), y2 (t), . . . , yM (t)]T = h(t, τ ) ∗ s t − τp , (7) It needs to be mentioned that most massive MIMO (millime-
ter) literatures [26]–[30], [33]–[38], though assuming M (W ) to
be extremely large, only stick to the spatial narrowband model

2 The existing papers on beam squint effect have been hypothesized directly 3 Time domain MIMO channel vector would lose the steering vector type
from frequency domain of the broadband communications [19], [22]–[24], structure for the general modeling (2) due to the unsynchronized signal waveform
which did not reveal the fundamental reasons of this problem. at different antennas.
WANG et al.: OVERVIEW OF ENHANCED MASSIVE MIMO WITH ARRAY SIGNAL PROCESSING TECHNIQUES 889

(9) and (10), while ignoring the spatial wideband phenomenon Clearly, Rdl
h and Rh are not the same but are highly correlated.
ul
that actually corresponds to (8) and (6). This property is used to derive Rdl
h from Rh in an easier way
ul
Moreover, for extremely large MIMO system [39]–[41], the [52]. Hence, the conventional training spent on the downlink
wideband effect or beam squint must be considered, which was covariance estimation and feedback can be well saved.
unfortunately still ignored in their discussion.
G. Low-Rank CCM Property of Massive MIMO
F. Uplink Downlink Reciprocity for Both TDD and FDD
With array signal processing, one can well link the angle
It has been shown in [42]–[44] that the conductivity and domain property with the rank of the CCM. Considering a
relative permittivity of most materials remain unchanged if the finite scattering environment with P AoAs, θpul ∈ [θmin ul ul
, θmax ],
frequency of the electromagnetic wave does not vary much, p = 1, 2, . . . , P , it is proved in [53] that the rank of Rh satisfies
ul
say less than 1GHz. Hence, the physical AoA of the uplink
h)
rank(Rul
channel is roughly the same as the AoD of downlink channel in d
TDD/FDD systems, which is called the angle reciprocity [45]. ≤ | sin(θmax
ul
) − sin(θmin
ul
)| , as M → ∞,
M λc
In fact, angle reciprocity had already been observed in [46], [47] (13)
but was not tangibly utilized due to the low angle resolution of
the small MIMO. In turn, the path delay τp that is only related to which tells that the rank of Rul
h would be proportional to the
light speed and the path length is also reciprocal for uplink and angular spread (AS) | sin(θmax
ul
) − sin(θmin
ul
)| of all multi-paths
downlink channels in FDD systems. Meanwhile, it is obvious (similar property holds for downlink CCM). Obviously, Rul h
that the user’s velocity or the scatters’ velocity is independent would be of low rank when AS is small.5 The low-rank property
of the carrier frequency. of CCM has been leveraged by many existing works to han-
Hence, although there does not exist direct reciprocity be- dle the pilot contamination and to solve the downlink channel
tween uplink and downlink channels in FDD systems, the reci- estimation issue. The required narrow AS can be observed in
procity can still be built upon the physical parameters like angle, the following scenarios: (i) BS equipped with a large number
delay, and Doppler shift. Let us use the superscript (·)dl and (·)ul of antennas is always elevated at a very high altitude, such
to represent the downlink and uplink parameters, respectively. that there are few surrounding scatterers [54]–[56]; (ii) When
From previous discussion, there are massive MIMO system is employed at the mmWave band, the
νpdl νpul high path loss leads to limited number of the incoming paths
θpdl = θpul , τpdl = τpul , = . (11) [18], [33]–[35].
fcdl fcul
Equalities (11) are also named as the angle-delay-Doppler III. PARAMETERS ESTIMATION
reciprocity and is important for designing the massive MIMO
systems with advanced array signal processing techniques. For Channel estimation has long been deemed as a bottleneck for
example, to estimate downlink channel of FDD systems, one massive MIMO systems due to a large number of unknowns
only needs to estimate the channel gain βpdl if the uplink channel to be estimated, especially for the downlink case. For a rich
parameters are known [31], [45], [48], [49]. Hence, the typical scattering environment, the channel covariance is of full rank
overhead restriction of the downlink training in FDD systems and there is no way to save the channel estimation effort. In
can be removed. In the meantime, an over-the-air (OTA) test this case, the conventional least square (LS) or LMMSE based
was set up in [50] to assess the system performance of the channel estimation has to be applied for massive MIMO [6], [8],
downlink reconstruction that utilized the parameter reciprocity [26], [29].
in practical FDD scenarios. The OTA results shown that the Nevertheless, a major concern is that the massive MIMO
channel reconstructed by the parameter reciprocity is close to systems are normally launched in high attitude or in mmWave
that obtained from linear minimum mean square error (LMMSE) frequency band such that low-rank CCM or the channel spar-
estimator, and higher accuracy can be achieved by increasing the sity regularly holds [54]–[56]. With sparse assumption, many
number of antennas. channel estimation algorithms that aim to reduce the train-
In [51], Deepak et al. even demonstrated that in certain ideal ing overhead have been proposed from communication theory
scenario, it is possible to directly predict βpdl from the uplink βpul viewpoint, such as virtual MIMO channel representation (VCR)
without any additional effort, such that the downlink training model [59] and compressive sensing (CS) methods [60], [61]. A
can be completely removed. detailed overview of sparse channel estimation methods in mas-
Angle reciprocity could also facilitate the derivation of the sive MIMO can be found in [17], [62] and will not be discussed
downlink channel covariance matrix (CCM) from the uplink one here. The basic idea is to take discrete Fourier transform (DFT)
for FDD systems. In [52], the authors consider the time-invariant matrix as the dictionary and apply the conventional CS algorithm
spatial and frequency narrowband channel where the uplink and to recover the channel. However, such idea is equivalent to
downlink CCMs can be expressed as4 assuming that the P AoAs of the signal paths reside exactly
on fixed grids {0, 2π 2π 2π
M , 2 M , . . . , (M − 1) M }, which could only

P
Rul E{|βpul |2 }a(θp , fcul , 0)aH (θp , fcul , 0), provide an approximate channel model.6 This is known as grid
h =
p=1
5 Here, we assume: (i) The antenna spacings and the end-to-end distance do not

P
satisfy the Rayleigh distance criterion [57]; (ii) The mmWave massive MIMO
Rdl
h = E{|βpdl |2 }a(θp , fcdl , 0)aH (θp , fcdl , 0). (12) channel is not under the “best-case rain rate” [58]. Otherwise, mmWave massive
p=1 MIMO channels may also be of high rank.
6 If P is extremely large or the incident angles are in the continuous domain, it
4 When P goes to infinite, the sum function changes to integration which is necessary to approximate all incident angles with the equivalent finite discrete
corresponds to the continuous incident angles. incident angles.
890 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 13, NO. 5, SEPTEMBER 2019

mismatch [63], [64] and would cause power leakage problem. and temporal dimensionality of the multipath channels were
Consequently, the performance of the channel estimation with exploited [73]–[75]. Specifically, with the unitary ESPRIT, [75]
VCR would reach an error floor in high SNR regions. investigated joint/separated angle/delay estimation methodolo-
Alternatively, instead of estimating the channel h (commu- gies in 3D massive MIMO millimeter wave systems. In addition,
nication theory viewpoint), one can also choose to estimate [76] utilized the multi-dimensional unitary ESPRIT to extract
the parameters inside the channel, i.e., θp , τp , νp , βp , as has the azimuth and elevation parameters as well as the Doppler
been tried in conventional MIMO systems. However, since the frequencies from the noisy channel, where URA was adopted
number of antennas M in conventional MIMO is small and the at both the transmitter and receiver of the mmWave MIMO
number of paths P is large in microwave band, it is impossible systems.
to accurately extract those parameters. The blind algorithms highly rely on the statistical characteris-
Nevertheless, with massive antenna array and sparse assump- tics of the received signals and thus have high spectral efficiency
tion (especially the mmWave scenario), it is possible to apply [77]. However, these algorithms often require a large number
array signal processing techniques to extract precise AoA/AoD of received signals, which are restricted to slow time-varying
parameters from the channel. Moreover, It should be emphasized channels and also encounter high complexity [78].
that only half-wavelength arrays can estimate AoA without
ambiguity, while for those massive MIMO systems with larger B. Training Based Method
inter-antenna distances [1]–[3], [6], [8], [26], it is impossible to
apply AoA/AoD based signal processing. Fortunately, because Different from conventional array signal processing, wireless
of the size limitation, the majority of practical massive arrays are communications offers cooperate terminals and it is possible to
equipped with half-wavelength. Similarly, with broad communi- use training sequence for more efficient parameter estimation
cation bandwidth and the sparse assumption, it is also possible [79], [80], while leaving blind way as a supplement for param-
to apply array signal processing techniques to extract precise eters tracking.
delay and Doppler parameters [18], [33]–[36]. 1) CCM Based Method: Consider a finite scattering envi-
ronment for massive MIMO systems and assume that the AS of
each user is restricted within a narrow region. Based on (13), we
A. Blind Parameters Estimation define
The blind AoA estimation is a classical problem in the area rk  Rank(Rk )  M, (14)
of array signal processing, and there are many well-known
approaches based on the eigen-value decomposition (EVD) where Rk is the CCM of the kth user with size M × M . Then,
of the signal covariance matrix (not CCM). For example, the the channel of the kth user can be expanded by rk dominant
parametric algorithms MUSIC [11] and ESPRIT [12] have been eigenvectors that correspond to rk nonzero eigenvalues, which
already demonstrated their superior resolution compared to the would reduce the channel dimensions from M to rk . One could
non-parametric counterparts, say the DFT-based ones. mathematically demonstrate the low-rank property of CCM as
With massive MIMO configuration, [65] compared the Root- [81]
MUSIC with the ESPRIT algorithms at frequency 30 GHz with R k = U k Λk U H
k , (15)
a 1D ULA of 72 elements. It was shown that the root-MUSIC
algorithm outperformed the ESPRIT under good radio channel where Λk is the nonzero eigenvalue matrix of size rk × rk and
conditions. Meanwhile, Yong et al. [66] proposed a power U k is the subspace eigen-matrix of size M × rk .
profile based AoA estimation (PROBE) algorithm for 1D Lens- It has been shown that the CCMs of any two users with non-
embedded mmWave MIMO systems. In terms of computational overlapped AS are asymptotically orthogonal to each other [62],
complexity, the blind PROBE algorithm has an advantage over i.e.,
the MUSIC and ESPRIT algorithms. Due to the size restriction, UH
k U l → 0, for Ak ∩ Al = ∅, as M → ∞, (16)
the antenna arrays of massive MIMO systems were expected
to be implemented in more than one dimension [67]–[69]. where Ak and Al are the continuous AS intervals for the kth
The work [69] proposed an ESPRIT-based approach for 2D and the lth users, respectively. Hence, the pilot contamination
localization of multiple sources employing very large uniform could be removed for these users with non-overlapped AS even
rectangular arrays (URAs). To estimate nominal AoAs that were if they employ the same training sequence. Based on (16), [81]
coupled in the array steering vector, the array has to be divided proposed a downlink joint spatial division multiplexing (JSDM)
into at least three subarrays to decouple the 2D nominal AoAs. scheme, where a classical multiuser precoder was adopted to
However, such a 2D model is still used to estimate azimuth restrict each user’s beamforming vectors within the orthogonal
AoAs only, while is not adequate for 3D channel estimation complement of the channel subspaces of the others. Meanwhile,
or the so-called full-dimension MIMO (FD-MIMO) [70]. The [53] directly applied uplink channel training via the minimum
3D models take into account azimuth as well as the elevation mean square error (MMSE) estimator and proved that channels
of signal propagation. The works [71], [72] then applied the with non-overlapped AS can be estimated free of interference.
ESPRIT algorithm to estimate the 3D channel parameters, such By leveraging the low-rank property of CCMs and reducing
as the azimuth AoAs, the elevation AoAs, and the azimuth AoDs the effective dimensions of channels, downlink pilot contam-
in the mmWave massive MIMO systems. ination as well as downlink training and feedback overhead
On the other side, the delay factor can also be estimated via can be significantly reduced [62]. However, the acquisition of
array signal processing techniques, especially for frequency- CCM is a difficult task for multi-user massive MIMO systems
wideband systems. Compared with the methods only estimating because each user’s high-dimensional downlink CCM has to be
AoAs [65]–[69], [71], [72], joint AoA and delay estimation separately estimated and then fed back to BS. Furthermore, the
methods have shown significant superiority in terms of the accompanied computational complexity involved in the EVD of
accuracy of the reconstructed channel, since both the spatial high-dimensional CCMs for multiple users is hardly affordable.
WANG et al.: OVERVIEW OF ENHANCED MASSIVE MIMO WITH ARRAY SIGNAL PROCESSING TECHNIQUES 891

dual-wideband systems, [31] utilized the 2D DFT-based algo-

rithm together with 2D rotation to estimate the AoAs and delays
of the multi-paths. With accurate AoA and delay estimation,
[31] further applied the angle-delay reciprocity to simplify the
downlink channel estimation in FDD systems.
3) Compressive Sensing Based Algorithms: CS provides a
sparse signal recovery method in the mmWave massive MIMO
systems and is a relatively new area of signal processing. See
[87]–[89] for the overviews of the fundamental theory in CS.
Based on spatial- and frequency-narrowband steering vector
a(θp , 0, 0), the channel vector in (10) can be represented by the
sparse format [90]:
h(t) = A(θ̄)x + n, (18)
where the M × L measurement matrix A(θ̄) is defined as
A(θ̄)  [a(θ̄1 , 0, 0), . . . , a(θ̄L , 0, 0)] and θ̄ is defined as θ̄ 
[θ̄1 , θ̄2 , . . . , θ̄L ], with θ̄i = −π + 2iπ
L , i ∈ {1, 2, . . . , L} divid-
ing the continuous angle space uniformly. Moreover, x ∈ C L×1
Fig. 2. Illustration of two power spectrums with M = 128 and P = 1. has the sparsity level of P , i.e., x has only P ≤ M non-zero
elements. Similar to beamspace method [59], [82], [83], the CS
2) DFT Based Algorithms: The spatial and frequency nar- model (18) assumes that the true channel parameters are exactly
rowband channel (10) naturally connects to Fourier transform aligned with the grids but with a finer grid interval when L ≥ M .
in that when M goes to infinity, the DFT of h(t) in (10) The AoA estimation problem is formulated as:
would only presents P non-zero peaks at the corresponding θ̂ = min x0 , s. t. h(t) − A(θ̄)x2 ≤ ξ, (19)
AoAs [45]. Hence the DFT-based representation of the massive x
MIMO channel with P non-zero supports are usually used as an where ξ is an error tolerance parameter that is related to the noise
approximation of h(t), i.e., the so called VCR [59]: statistics. Then, the CS-based sparse channel estimation boils
down to angle estimation and path gain estimation subproblems,
h(t) ≈ W H
M b, (17) which can be solved sequentially with reduced complexity [91].
Note that the on-grid CS model considers the θ̄ as the known
where W M ∈ C M ×M is the normalized DFT √ matrix whose parameter whose accuracy depends on L. It only provides the
2π
(m, n)th entry is [W M ]m,n = e−j M mn / M , 0 ≤ m, n ≤ true AoA estimation when L goes to infinity but the accom-
M − 1, and b ∈ C M ×1 is assumed to have P non-zero entries panied complexity is unaffordable. To deal with this problem,
with the position related to θp . The corresponding method is [92] treated θ̄ in (18) as the unknown continuous parameters
named as beamspace channel estimation [59], [82], [83]. that can take arbitrary values and covert the objective to not
However, DFT could only yield the AoA estimation with the only estimating the sparse signal, but also optimizing/refining
resolution of 1/M [84] and is a non-parametric AoA estimation the parameters θ̄. Then, the problem becomes a sparse signal
method. Consequently, due to the mismatch between DFT basis recovery problem with an off-grid parametric dictionary. i.e.,
and real θp , the power of the beamspace subchannels would
leak to their neighbors, and hence the true number of non-zero θ̂ off = min x0 , s. t. h(t) − A(θ)x2 ≤ ξ, (20)
x,θ
supports, i.e., the non-zero entries in b is much larger than P , as
shown in Fig. 2(b). The resultant phenomenon is also called grid where θ is the P × 1 unknown angle vector, and A(θ) is the
mismatch or power leakage effect [62], which leads to significant M × P off-grid parametric containing P columns of steering
channel estimation error or the increased training overhead in vectors.
order to estimate all the non-zero supports. For either the on-grid or the off-grid AoA estimation, many
The authors of [85] then proposed to add zeros at the end CS sparse signal recovery algorithms can be utilized and are
of h(t) before the DFT operation, such that the angle domain mainly divided into the following four categories:
is sampled in a much denser way. In this case, a better angle a) Convex relaxation algorithms: A famous convex relax-
estimation can be obtained compared to beamspace method. ation for the sparse recovery algorithm is the least absolute
On the other side, the authors [45] proposed to use a spatial shrinkage and selection operator (LASSO). Donoho [87] and
rotation to mitigate the power leakage effect by rotating the Candes [93] showed that one avenue for translating (19) into a
channel vector and concentrating most power in much fewer tractable problem is to replace  · 0 with its convex approxi-
grids. Compared to beamspace model, [45] and [85] targeted at mation  · 1 . The LASSO estimate of (18) is formulated as a
the real AoA estimation and was named as anglespace modeling Lagrangian relaxation of a quadratic program, which is given by
or the modified VCR. With accurate AoA estimation, [45] further [94]
applied the angle reciprocity to simplify the downlink channel
1
estimation for both TDD and FDD systems. The work [49] x̂ = λσ min x1 + h(t) − A(θ̄)x2 , (21)
then extended anglespace method to indoor mmWave massive x 2
communications with uniform rectangle array (URA). where σ is the standard derivation of the noise, and λ > 0 is
For frequency wideband systems, the author [86] utilized the regularization parameter. Sayeed et al. [95] then utilized the
the 2D DFT-based algorithms to capture the hidden sparsity LASSO-based algorithm for the sparse virtual channel estima-
of the channel and estimate the on-grid AoAs and delays. For tion, which exploited the inherent sparsity of mmWave channels
892 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 13, NO. 5, SEPTEMBER 2019

in the angle and delay domains. Eltayeb et al. [96] also developed d) Bayesian algorithms: The sparse Bayesian learning
a LASSO based array diagnosis techniques for the mmWave (SBL) principle could infer the sparse unknown signal from the
systems with large antenna arrays which jointly estimated the Bayesian viewpoint by considering the sparse priori [108]. A
AoAs/AoDs and the phase-shifts. typical SBL model is given by [108], [109]:
b) Greedy algorithms: Greedy iterative algorithms select
columns of A(θ̄) according to their inner product correlation p(h|x, ζ, A(θ̄)) = CN (A(θ̄)x, ζ −1 I M ), (24)
with the measurements h(t) in a greedy iterative manner [90].
where ζ  σ −2 denotes the noise precision. In general, ζ is
At each iteration, the sparse vector is updated from
assumed to follow Gamma prior distribution p(ζ) = Γ(v; χ, υ),
min h(t) − AΛ (θ̄)xΛ 2 , (22) where Γ(·) is the Gamma function, χ is the shape parameter,
xΛ and υ is the rate parameter.
Matteo et al. [110] first utilized the SBL method to estimate
where xΛ = AΛ (θ̄)† h(t), and Λ denotes the current esti- the AoA from the narrowband MIMO channels. Meanwhile,
mated support set. Thus, in each iteration, the residual h(t) − Zhang et al. [111] proposed a sparse Bayesian-based estimation
AΛ (θ̄)xΛ is orthogonal to the columns of A(θ̄) that are included algorithm for frequency wideband communications, which took
in the current estimated support [90]. Lau et al. [97] consid- advantage of the band occupation information by leveraging
ered the multi-user massive MIMO systems and deployed an the cluster property of the Dirichlet process. However, [110],
orthogonal matching pursuit (OMP)-based technique to reduce [111] are still on-grid sparse Bayesian algorithms. Das and
the training as well as the feedback overhead for AoA estimation. Terrence [112] then proposed an off-grid version of SBL based
However, the conditions on OMP estimation algorithms are relevance vector machine algorithm (SBLRVM) to estimate the
more restrictive than the restricted isometry condition (RIC) AoAs of multiple narrowband signals. Jian et al. [113] utilized
[98], which can easily lead to weak parameter estimation. Gui the off-grid SBL method to jointly obtain the super-resolution
et al. [99] then utilized a robust compressive sampling match- angle and delay for the accurate channel reconstruction in
ing pursuit (CoSaMP)-based algorithm to exploit the block- dual-wideband mmWave massive MIMO systems. The off-grid
structure sparsity in angular domain in order to further improve Bayesian algorithms have superior performance due to their use
channel estimation of massive MIMO systems. However, [97], of data-adaptive priors, which is suitable to extract the unknown
[99] still face the grid mismatch problem. Then, Wang et al. parameters i.e., AoAs and delays from the structured array.
[100] proposed a shift-invariant block-sparsity based channel
estimation algorithm that can jointly compute the off-grid an-
gles, the off-grid delays, and the complex gains of the wideband C. Prospects of Parameters Estimation
mmWave massive MIMO channels. The core idea of parametric massive MIMO channel modeling
c) Atomic norm algorithms: In order to circumvent grid is to estimate the physical channel parameters using various ef-
mismatch, [101] proposed an algorithm named atomic norm fective manners. However, it should be emphasized that although
denoising, expressed as the parametric channel estimation algorithms can greatly reduce
the computational complexity brought by the massive MIMO
1
x̂ = argmin A(θ̄)x − h(t)22 + γxA , (23) systems, they are very sensitive to the number of multi-path
x 2 and the accuracy of estimated parameters. There will be distinct
angle error when the estimated path number is wrong, especially
where γ presents the regularization parameter and  · A denotes in the discrete sparse scenes with finite incident paths. This
the atomic norm defined in [101]. Bhaskar et al. [101] proved is also a critical problem in traditional array signal processing
that (23) can be converted into a semi-definite programming that techniques. The commonly accepted algorithms for path number
can be computed by off-the-shelf solvers such as SeDuMi and estimation are minimum description length (MDL) and Akaike
SDPT3 in CVX toolbox. information criterion (AIC) [114], while MLD and AIC algo-
Atomic norm denoising can be applied to estimating the chan- rithms both have upper precision limit. At present, how to design
nel parameters with super-resolution [101] and was applied for a low-complexity as well as high-resolution algorithm is still a
super-resolution channel estimation in mmWave massive MIMO challenging work, and we hope researchers can conduct in-depth
systems by Wang et al. in [102]. To avoid the enlarged problem work in this area.
size due to the vectorization-based atomic norm minimization in
the 2D-angle scenarios, Tian et al. [103] proposed a decoupled
IV. SYNCHRONIZATION
atomic norm minimization technique and its theoretical results
were provided in [104]. Moreover, Tsai et al. [105] proposed an Apart from the channel estimation, another challenge of
atomic-norm-denoising-based method for mmWave FD-MIMO massive MIMO is the synchronization issue [115], [116].
channel estimation. Tirkkonen et al. [106] formulated mmWave Particularly for mmWave massive MIMO, the Doppler spread
MIMO channel estimation as atomic norm denoising problem is orders-of-magnitude larger than that of classical narrow-
and then designed the multi-user precoder. Besides, Chu et al. band wireless channels, which may deteriorate the system
[107] utilized the atomic norm minimization to manifest the performance [117].
channel sparsity in the continuous azimuth AoAs and AoDs, Consider wideband transmission employing OFDM modu-
which also provided a super-resolution channel estimators for lation. With perfect time and frequency synchronization in the
mmWave MIMO systems. space domain, the length of the cyclic prefix (CP) is usually
However, atomic norm denoising algorithm becomes slow set to be slightly larger than the delay span to mitigate time
when solving large scale problems. In order to provide a faster dispersion, while the length of the OFDM symbol is usually set
method for the semi-definite program, [101], [105] adopted to be inversely proportional to the Doppler spread to mitigate
the alternating direction method of multipliers (ADMM) to frequency dispersion [118], [119]. Embracing the beam squint
accelerate such process. effect, the authors of [31] showed that extra CP length is required
WANG et al.: OVERVIEW OF ENHANCED MASSIVE MIMO WITH ARRAY SIGNAL PROCESSING TECHNIQUES 893

to overcome the time delays across the large array aperture. As

a result, the overhead caused by CP will be much larger to deal
with the same delay spread for wideband massive MIMO system.
You et al. [117] proposed per-beam synchronization (PBS)
for mmWave massive MIMO-OFDM transmission, where both
delay and Doppler frequency spreads of the wideband MIMO
channels were reduced approximately by a factor of the number
of antennas. In the angle domain, Zhang et al. [120] designed a
frequency synchronization scheme for multi-user OFDM uplink
with a massive ULA at BS. By exploiting the angle information
of users, the carrier frequency offset (CFO) was estimated
for each user individually through a joint spatial-frequency
alignment procedure. Moreover, the authors in [121] and [122]
addressed the joint estimation issue of Doppler shifts and CFO
in high-mobility downlink massive MIMO systems, where a
high-resolution beamforming network was designed to separate
different Doppler and CFOs in the angle domain.

V. BEAMFORMING AND PRECODING

The precoding techniques for massive MIMO systems with
both sub-6 GHz and mmWave frequency band have been well
studied in [17], [19], [123]–[125]. In this section, we will not go
through those works that have already been described from com-
munications viewpoints but rather to introduce those designed
from the array signal processing viewpoint. As mentioned in
Section II, the array signal processing based design is mainly Fig. 3. Illustration of two different beamforming methods.
applicable for large arrays and sparse environments. In fact many
beamforming designs for mmWave originate from array signal
then the objective function could be minimized. If MRF < P ,
processing techniques.
then additional effort should be applied to minimize the ob-
jective function, say the beamforming design [129]. Matching
A. Hybrid Precoding the columns in F RF with the true steering vector a(θpdl , 0, 0)
In conventional MIMO systems, each antenna is driven by is named as anglespace beamforming, and the corresponding
its own dedicated radio frequency (RF) chain and thus the full beams are shown in Fig. 3(a), [85], [130], [131]. Alternatively,
digital beamformer (DB) can be applied [126]. However, DB is if F RF only supports M fixed directions, then the best beam-
not suitable for practical massive MIMO systems since the huge forming vectors should be selected from a predetermined code-
number of RF chains brings high hardware cost, high system book, i.e., beamspace beamforming, [82], [83], [132], [133].
complexity, and large power consumption. The hybrid analog Obviously, the beamspace precoding can be achieved with low
and digital beamforming architecture is then proposed, where complexity but can only be viewed as “angle on the grid” ap-
M antennas share the same MRF  M RF chains [127], [128]. proach and will suffer from power leakage [62], [85], as shown in
The analog beamforming matrix is usually realized by a low- Fig. 3(b). Once the analog beamforming matrix F RF is obtained,
cost phase shifter or switching network that imposes a constant the digital precoder f BB can be derived following minimum
modulus constraint on each element. mean square error (MMSE) criterion [85], [134], [135] or the
For conciseness, we illustrate the beamforming design with a zero-forcing (ZF) criterion [131].
single user only, as did in [19]. Under spatial and frequency It has been shown in [19] that the hybrid analog and digital
narrowband environment, the downlink transmission can be beamforming design for multiuser massive MIMO system can
expressed as reduce the number of RF chains and achieve near-optimal per-
formance as that of fully-digital beamforming structure when
ỹ = hH F RF f BB s + n, (25) carefully designed. Moreover, there are many other solutions
that design the hybrid structure with AoAs/AoDs. Details can
where h = P p=1 βp a(θp , 0, 0) is the downlink channel, F RF
dl dl
be found in [19] and will not be re-stated here.
is the M × MRF analog precoding matrix with constant modulus
entries, while f BB is the MRF × 1 digital precoding vector. The
objective is to design the hybrid analog and digital beamforming B. Interference Control
that can maximize the overall system spectral efficiency, i.e., From communications viewpoint, the user interference is
2 caused by the non-zero inner product between the channel vector
min |h − F RF f BB |
F RF ,f BB and the beamforming vector, while the cancellation of the inter-
1 ference to a specific user is achieved by adjusting the weight on
s.t. |[F RF ]i | = √ , i = 1, 2, . . . , M. (26) different antennas. From array signal processing viewpoint, such
M interference is due to the side lobes of the beamforming vector
The optimization (26) can be well addressed from array signal and leaks towards the undesired user, while the interference
processing viewpoint in that when columns of F RF contain cancellation is equivalent to formulating a physical null over
the steering vector a(θpdl , 0, 0) and elements of f BB match βp , the beam pattern towards the AoAs of the undesired users. To be
894 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 13, NO. 5, SEPTEMBER 2019

space (OTFS) considers all of these three parameters in massive

MIMO systems. Leveraging the basis expansion model (BEM)
for the channel [140], OTFS first transforms the time-varying
multipath channel into a two-dimensional time-independent
channel in the delay-Doppler domain [141], [142]. The OTFS
rearranges the data sequence of length Ne No into a 2D data
block with the size of Ne × No , where Ne and No are the
number of units along the delay dimension and Doppler di-
mension, respectively. This block data is transmitted with total
duration Ne T seconds and total bandwidth No f Hz, e.g,
a sampling of the time and frequency axes at intervals T
and f , respectively. OTFS is different from previous work
in that it multiplexes data in the delay-Doppler domain, and
each transmitted symbol experiences a near-constant channel
gain even in cases with high Doppler fading [143], massive
Fig. 4. A design example, maximize the magnitude response in the main lobe, MIMO [144], or at high frequencies such as millimeter waves
or equivalently r, while suppressing the ripple into a given level ε. [145]. The relatively constant channel gain over all symbol
transmissions obtained in OTFS massive MIMO system greatly
specific, the lobe corresponding to the direction of the maximum reduces the overhead and complexity associated with physical
gain is referred to as the main lobe while lobes with much smaller layer adaptation, where signal processing can play a role. There
gain are the side lobes. Usually, there exists a large ripple in are many other discussions including signal detection [146],
main/side lobes of hybrid beamforming structure, which would modulator design [147], interference cancellation [148], and
cause interferences between different users. Suppressing the rip- performance analysis [149].
ple is beneficial to achieve a uniform beamforming performance, In this part, we focus on the angle-delay-Doppler 3D struc-
which can guarantee user fairness to some extent. tured sparsity of the OTFS channel in massive MIMO sys-
Ideally, one would expect the ripple of each beam to be as tems. The downlink channel H of the OTFS massive MIMO
small as possible while the main lobe power as large as possible. is a 3D tensor of dimension Ne × No × M , whose (ne , no , m)
However, these two goals generally conflict with each other th entry (ne ∈ {1, 2, . . . , Ne }, no ∈ {1, 2, . . . , No }, and m ∈
unless the number of antennas M goes to infinity. To reach a {1, 2, . . . , M }) is given by [150]
reasonable solution, a small ripple in both the main lobe and 
P
the side lobes are accepted but the ripple must be limited to a Hne ,no ,m = βp Ψ(ne Ts − τp )ΥNo
small positive real number ε, as shown in Fig. 4. The design p=1
of the training beamforming vector can be obtained from the
following optimization problem [136]–[138]: (no − νp No T − No /2 − 1)
max q
rs,b × ΥM (m − M (1 + sin θp )/2 − 1), (28)
f qs,b ,rs,b
q

where Ts is the system sampling interval, Ψ(τ ) is the

s. t. |a(θp , 0, 0)f qs,b − rs,b
q q
| ≤ ε, (θp ∈ Is,b ) band-limited pulse shaping filter (e.g., the Hanning window),
j2π m−1
ΥM (x)  M m=1 e
M , and n , n , m corresponding to the
e o
|a(θp , 0, 0)f qs,b | ≤ ε, (θp ∈ Is,b
q
), (27) delay, Doppler, and angle index, respectively. The function
q ΥM (x) has the property |ΥM (x)| ≈ 0 when |x|  1. Then
where rs,b is the main lobe power, f qs,b is the training beam-
q Hne ,no ,m has the dominant elements only if ne = τp /Ts ,
forming vector, and Is,b is the beam interval. no = νp No T + No /2 + 1, and m = M (1 + sin θp )/2 + 1,
Alkhateeb et al. [132] have realized the importance of con- as shown in the Fig. 5 [150]. Since the number of signifi-
trolling the ripple. However, the designed interference control cant propagation paths is typically limited, the 3D time-variant
method from LS principle in [132] cannot well suppress the channel is sparse along the delay dimension. Also, the Doppler
ripple and thus cause a large performance loss on non-quantized frequency of a path is usually much smaller than the system
points in the beamspace domain. Raghavan et al. [139] proposed bandwidth, the 3D channel is block-sparse along the Doppler
a maximin optimization to achieve a well-shaped training beam, dimension, Besides, with the assumption of narrow AS, the 3D
which essentially looked for a flat array gain response in the channel is burst-sparse along the angle dimension [151]. The
coverage area of the designed beam and thus can realize a small lengths of non-zero bursts can be regarded as constant, but the
ripple in the main lobe as well as side lobes. To reduce the user start position of each nonzero burst is unknown, as shown in
interference and increase the beamforming performance, Zhang Fig. 5. This 3D sparse property indicates that the high dimension
et al. [138] proposed a training codebook design by shaping channel can be efficiently estimated via limited channel param-
the geometric pattern of each beam and formulating the beam eters with low overhead. However, the existing OTFS works
design as an optimization problem with both the ripple and the estimated the channel parameters via on-grid methods where
transition band being taken into consideration. the power leakage effect still exists. Hence, future researches are
still demanded for developing off-grid parametric algorithms.
VI. ORTHOGONAL TIME AND FREQUENCY SPACE
OF MASSIVE MIMO VII. MULTIPLE ACCESS
In addition to exploiting angle, delay and Doppler separ- The key advantage of massive MIMO is to tremendously
atively, a new modulation named orthogonal time frequency enhance the system capacity by simultaneously serving multiple
WANG et al.: OVERVIEW OF ENHANCED MASSIVE MIMO WITH ARRAY SIGNAL PROCESSING TECHNIQUES 895

Fig. 7. Illustration of the ADMA to eliminate the blockage problem in 3D full

Fig. 5. Angle-delay-Doppler 3D channel, which is burst-sparse along the space.
angle dimension, sparse along the delay dimension, and block-sparse along the
Doppler dimension [150].
Nevertheless, having orthogonal CSI from two users does
not necessarily mean that the angular spreads of the users are
non-overlapping, especially for conventional small MIMO. Ac-
tually, the conventional spatial domain multiple access (SDMA)
is universal for any number of antennas while ADMA can
only be applied when AoAs of users are distinguishable, say in
massive MIMO and sparse scenario. Nevertheless, the advantage
of ADMA is drawn from the frequency insensitivity of the
angular parameters, i.e., even if the frequency changes (uplink
and downlink in FDD systems), the multi-users access designed
from ADMA is still applicable. Moreover, the ADMA would
link the users’ transmission with their individual locations such
that many geometric theories could be utilized to design the
multi-user scheduling, even for multiple BS stations or the cell-
free case [48]. An example that adopts two BS and geometric
Fig. 6. ADMA for users with non-overlapping angular spreads. theory to eliminate the blockage problem is shown in Fig. 7 [48].

users in the same frequency and time resource. With the knowl- B. Delay Division Multiple Access (DDMA)
edge of CSI, the conventional 1G to 4G mobile communication Similarly, the frequency domain channel and the delay domain
networks generally assign orthogonal time, frequency, or spatial parameters formulate the second Fourier transform pair. For
resource for different users. Nevertheless, these multiple access the conventional narrowband case, the bandwidth W is too
schemes can be re-visited under the massive MIMO configura- small to discriminate multi-path delays, such that one can only
tion with the aid of array signal processing. rely on the frequency domain orthogonality to design multiple
access schemes. However, under the broadband transmission
(mmWave) and sparse scenario, the delay τp can be well ex-
A. Angle Division Multiple Access (ADMA) tracted from array signal processing techniques. It can be easily
From (10), it can be observed that the spatial domain channel shown that users with non-overlapping delays under broadband
and angle domain parameters formulate a Fourier transform pair, configuration will exhibit orthogonal frequency domain chan-
especially when M goes to infinity. For conventional MIMO nels, which can be used for distinguishing users. The corre-
case, M is too small to discriminate AoAs, such that one can only sponding multiple access scheme is named as delay division
rely on the spatial channel h to design multiple access schemes. multiple access (DDMA). Interestingly, delay parameters can
However, with massive array and sparse scenario, as explained be artificially controlled as compared to AoA parameters. For
previously, the AoAs can be well extracted and could be used example, even if two users have overlapping delay parameters,
for discriminating users too. It has been implicitly shown in one user can purposely postpone its transmission such that the
[53], [81] that the orthogonal eigen-space based multiple access two users would have non-overlapping delays, as shown in
scheme is equivalent to using non-overlapping angular spread Fig. 8. However, the orthogonal frequency domain channels
to distinguish different users. In [45], [85], non-overlapping from two users do not necessarily mean that the delays of the
angular spread were explicitly exploited for multiple users users are non-overlapping, especially for narrowband systems.
access and formulated the so-called angle division multiple Actually, the advantage of the conventional frequency domain
access (ADMA), as shown in Fig. 6. It was actually proved in multiple access (FDMA) is its universal applicability for any
[45], [85] that users with non-overlapping angular spread would bandwidth while DDMA can only be applied when the delays of
result in orthogonal spatial channels under massive MIMO different users are distinguishable. Nevertheless, the advantage
configurations. of DDMA is also drawn from the frequency insensitivity of the
896 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 13, NO. 5, SEPTEMBER 2019

Fig. 8. DDMA for users with non-overlapping delays.

Fig. 9. Massive MIMO with PDMA.

delay parameters, which says that even if the frequency changes

(uplink and downlink in FDD systems), the multi-users access
designed from DDMA is still applicable.

C. Doppler Division Multiple Access (DoDMA)

Lastly, it can be observed that the time domain channel and
the Doppler domain parameters formulate the third Fourier
transform pair. Hence, we could theoretically predict that it is
possible to extract the Doppler parameters of different users and
then discriminate users with non-overlapping Doppler param-
eters. For example, for wideband communications in a sparse
environment, the symbol duration T is very small such that a
large amount of symbols (corresponding to large array for AoAs
or large bandwidth for delays) can be observed in a short time, Fig. 10. PDMA and traditional multiple access technologies to construct
during which period the Doppler parameters remain the same. X-DMA.
In this case, the Doppler parameters νp can be well extracted
and used for discriminating users too. The resultant scheme is different situations, one can flexibly choose various combi-
name as Doppler division multiple access (DoDMA) or velocity nations of the above X-DMA to maximize the utilization of
division multiple access (VDMA). It can be easily showed the time/frequency/spatial resource. For example, with mas-
that users with non-overlapping Doppler shifts will result into sive MIMO and narrowband transmission where AoAs can be
orthogonal time domain channel. However, orthogonal time discriminated but the delays cannot, one can choose ADMA
domain channels from two users do not necessarily mean that combined with FDMA for multi-user access. Moreover, for
the Doppler shifts of the users are non-overlapping. Actually, multi-user OTFS massive MIMO systems, if angle/delay/
the advantage of the conventional time domain multiple access Doppler are utilized simultaneously for multiple access, then
(TDMA) is universal for any Doppler/velocity while DoDMA it belongs to a type of PDMA.
can only be applied when Doppler shifts of users are distin- Nevertheless, it must be pointed out that introducing ADMA,
guishable. Nevertheless, the advantage of DoDMA is also drawn DDMA, and DoDMA with the aid of array signal processing
from the frequency insensitivity (with a known scaling factor) do not mean we can achieve three completely new degree of
of the Doppler parameters, which says that even if the frequency freedoms. For example, SDMA and ADMA are inherently from
changes (uplink and downlink in FDD systems), the multi-users the same Fourier pair and hence it is not possible to apply SDMA
access designed can be adapted to the new frequency band easily. and ADMA simultaneously.
However, DoDMA may be difficult to implement in practice
because once users are moving, their Doppler parameters would
be varying and may cause a frequent update of the multiple VIII. ARTIFICIAL INTELLIGENCE METHODS FOR
access strategy. MASSIVE MIMO
In recent years, artificial intelligence (AI), e.g., machine
D. Path Division Multiple Access (PDMA) learning (ML) and deep learning (DL), have achieved great
The previously introduced ADMA, DDMA, and DoDMA success in the fields of computer vision, natural language pro-
are all based on the physical parameters of the multi-paths, cessing, speech recognition, etc. [153]. Researchers in the field
which together can be named as path division multiple access of wireless communications are eager to apply AI technologies
(PDMA).7 A diagram to represent different multiple access in all levels of the wireless systems that could create intelligent
schemes as well as their relationships is shown in Fig. 9. communications to meet various demands for 5G beyond [154].
The logical relationship between the aforementioned six Inspired by the recent advances in ML/DL paradigm, ML/DL-
multiple access techniques is shown in Fig. 10. According to based wireless communication techniques have aroused consid-
erable interest among the academic and industrial communities
[153]–[155]. In this section, we investigate ML/DL-based signal
7 The term PDMA has been used in lens array [152] but in fact only refers to processing techniques in massive MIMO systems, including
angle domain. channel estimation, beamforming, etc.
WANG et al.: OVERVIEW OF ENHANCED MASSIVE MIMO WITH ARRAY SIGNAL PROCESSING TECHNIQUES 897

A. Learning Based Channel Estimation

In [156], the authors exploited a learned denoising-based
approximate message passing (LDAMP) network to solve the
channel estimation problem, which was the first attempt to use
DL technology for beamspace channel estimation. The LDAMP
network regards the channel matrix as a 2D natural image and in-
corporate the denoising convolutional neural network (DnCNN)
into the iterative signal recovery algorithm for channel estima-
tion. With large number of channel matrices being training data,
the LDAMP can be applied to a variety of selection networks.
In order to estimate AoA with high resolution, Huang et al.
[157] utilized a deep neural network (DNN) that integrated the Fig. 11. The structure of array signal processing techniques for massive MIMO
sparsity channels.
massive MIMO into DL. The DNN was employed to conduct
off-line learning and online learning procedures, which learned
the statistics of the wireless channel and the spatial structures
in the angle domain. In addition, an online DNN based channel parameters of channels that characterize the channels responses,
estimation algorithm for doubly selective fading channels was i.e., angle, multi-path delay, and Doppler can be well estimated
proposed in [158]. With properly selected inputs, the DNN can with large array and broad bandwidth. Moreover, we especially
not only exploit the features of channel variation from previous point out three new multiple access schemes as well as their
channel estimations but also extract additional features from relationships to the conventional TDMA FDMA, and SDMA.
pilots and received signals [158]. The whole principle of applying array signal processing for
massive MIMO system is summarized in Fig. 11, and the key
B. Learning Based Beamforming advantages of applying array signal processing are summarized
here:
To solve the analog beam selection problem for hybrid beam- r The number of the unknown parameters can be reduced
forming in mmWave massive MIMO systems, Long et al. [160] such that the training overhead can be greatly shortened,
proposed a data-driven ML solution by resorting to support and the estimation accuracy is higher than the conventional
vector machine (SVM). In order to reduce the complexity of approaches that simply assume channel sparsity;
classification, the AoAs and AoDs of multi-paths were used r Angle, delay, and Doppler parameters are widely frequency
as entries of feature vectors, and a large number of samples insensitive and the so designed transceiver techniques like
from the mmWave channels were collected as the training data. channel estimation, synchronization, precoding, and user
Then, Antón-Haro and Mestre [161] exploited AoA information scheduling, etc., can be well adapted from one frequency
and learning approaches to perform beam selection in hybrid band to another.
mmWave communication systems. To compare with different Besides, with array signal processing, one can deeply under-
degrees of complexity/sophistication, [161] adopted three lean- stand the RF communications rationale and explain the effect
ing methods including two ML approaches, SVM, k-nearest for spatial- and frequency-wideband effect that was previously
neighbors, and one DL approach, the multi-layer perceptron. ignored from communications viewpoint. It is then not difficult
The beam selection task was finally changed into a multi-class to realize that most existing algorithms were designed only with
classification problem and was solved via those three schemes. frequency- and spatial- narrowband channel models. Extensions
Meanwhile, Klautau et al. [162] investigated the performance of to wideband channels in many cases are important and necessary,
several ML schemes (SVM, Adaboost, Decision trees, Random especially for mmWave, Terahertz massive MIMO systems and
forests) but also DNN and reinforcement learning for beam se- extremely large MIMO systems.
lection techniques on vehicle-to-infrastructure using mmWave. Moreover, to fully utilize the array signal processing tech-
niques in wireless communications, the problems faced by
C. Prospects of Learning Based Methods the traditional array signal processing should also be studied.
For example, the parameter-based channel models discussed
From the brief introduction, it can be seen that the trial of in previous sections are sensitive to array response errors and
learning networks utilized in parametric massive MIMO sys- how to calibrate large scale array in FDD systems is still an
tems can achieve appealing performance in channel estimation, open problem [163]. Further works are needed to investigate the
beamforming, etc. Many other extensions are also attracting, areas of array signal processing like target tracking, array posi-
like CSI feedback/reconstruction [159], signal detection, chan- tioning, array shape design, array calibration, etc., that can all
nel coding/decoding, end-to-end communications systems (see be re-shaped into wireless communications. Though we cannot
[153]–[155] for more suggested future works). Based on the spend more space to illustrate all these works, we hope this short
previous sections, we firmly believe that these AI methods could overview paper can give a clear picture of why and how the array
be applied from array signal processing viewpoints and would signal processing can be merged into wireless communications.
achieve all the advantages discussed preciously. We would also like to encourage researchers from both areas to
devote more efforts into this promising direction.
IX. CONCLUSION AND FUTURE WORKS
The purpose of this paper is to provide an overview on array REFERENCES
signal processing techniques that have been successfully applied [1] T. L. Marzetta, “Noncooperative cellular wireless with unlimited numbers
in the context of massive MIMO systems. These methods rely of base station antenna,” IEEE Trans. Wireless Commun., vol. 9, no. 11,
on sparse and low-rank signal models, for which the physical pp. 3950–3600, Nov. 2010.
898 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 13, NO. 5, SEPTEMBER 2019

[2] E. G. Larsson, O. Edfors, F. Tufvesson, and T. Marzetta, “Massive MIMO [28] R. R. Muller, L. Cottatellucci, and M. Vehkapera, “Blind pilot decontam-
for next generation wireless systems,” IEEE Commun. Mag., vol. 52, ination,” IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 773–786,
no. 2, pp. 186–195, Feb. 2014. Oct. 2014.
[3] F. Rusek et al., “Scaling up MIMO: Opportunities and challenges with [29] J. C. Guey and L. D. Larsson, “Modeling and evaluation of MIMO
very large arrays,” IEEE Signal Process. Mag., vol. 30, no. 1, pp. 40–60, systems exploiting channel reciprocity in TDD mode,” in Proc. IEEE
Jan. 2013. 60th Veh. Technol. Conf., 2004, pp. 4265–4269.
[4] A. Yang, Y. Jing, C. Xing, Z. Fei, and J. Kuang, “Performance analysis [30] J. Choi, D. J. Love, and P. Bidigare, “Downlink training techniques for
and location optimization for massive MIMO systems with circularly FDD massive MIMO systems: Open-loop and closed-loop training with
distributed antennas,” IEEE Trans. Wireless Commun., vol. 14, no. 10, memory,” IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 802–814,
pp. 5659–5671, Oct. 2015. Oct. 2014.
[5] B. Hassibi and B. M. Hochwald, “How much training is needed in [31] B. Wang, F. Gao, J. Shi, H. Lin, and G. Y. Li, “Spatial- and frequency-
multiple-antenna wireless links?,” IEEE Trans. Inf. Theory, vol. 49, no. 4, wideband effects in millimeter-wave massive MIMO systems,” IEEE
pp. 951–963, Apr. 2013. Trans. Signal Process., vol. 64, no. 6, pp. 1461–1476, May 2018.
[6] H. V. Cheng and E. G. Larsson, “Some fundamental limits on frequency [32] B. Wang et al., “Spatial-wideband effect in massive MIMO with ap-
synchronization in massive MIMO,” in Proc. Asilomar Signals, Syst. plication in mmWave systems,” IEEE Commun. Mag., vol. 56, no. 12,
Comput., Nov. 2013, pp. 1213–1217. pp. 134–141, Dec. 2018.
[7] A. F. Molisch et al., “Hybrid beamforming for massive MIMO: A survey,” [33] M. Xiao et al., “Millimeter wave communications for future mobile
IEEE Commun. Mag., vol. 55, no. 9, pp. 134–141, Sep. 2017. networks,” IEEE J. Sel. Areas Commun., vol. 35, no. 9, pp. 1909–1935,
[8] E. Björnson, E. G. Larsson, and M.Debbah, “Massive MIMO for maximal Sep. 2017.
spectral efficiency: How many users and pilots should be allocated?,” [34] T. S. Rappaport, S. Sun, R. Mayzus, and H. Zhao, “Millimeter wave
IEEE Trans. Wireless Commun., vol. 15, no. 4, pp. 1293–1308, Feb. 2016. mobile communications for 5G cellular: It will work!” IEEE Access,
[9] N. Shlezinger and Y. C. Eldar, “On the spectral efficiency of noncoop- vol. 1, no. 1, pp. 335–349, May 2013.
erative uplink massive MIMO system,” IEEE Trans. Wireless Commun., [35] M. K. Samimi and T. S. Rappaport, “Ultra-wideband statistical channel
vol. 67, no. 3, pp. 1956–1971, Mar. 2019. model for non line of sight millimeter-wave urban channels,” in Proc.
[10] J. Li, P. Stoica, Z. Wang, “On robust Capon beamforming and diagonal IEEE Global Commun. Conf., Dec. 2014, pp. 3483–3489.
loading,” IEEE Trans. Signal Process., vol. 51, no. 7, pp. 1702–1715, [36] T. S. Rappaport, G. R. MacCartney, M. K. Samimi, and S. Sun, “Wide-
Jul. 2003. band millimeter-wave propagation measurements and channel models for
[11] R. O. Schmidt, “A signal subspace approach to multiple emitter location future wireless communication system design,” IEEE Trans. Commun.,
and spectral estimation,” Ph.D. Dissertation, Dept. Elect. Eng., Stanford vol. 63, no. 9, pp. 3029–3056, Sep. 2015.
Univ., Stanford, CA, USA, May 1981. [37] A. Younis, N. Abuzgaia, R. Mesleh, and H. Haas, “Quadrature spatial
[12] R. Roy and T. Kailath, “ESPRIT-estimation of signal parameters via modulation for 5G outdoor millimeter-wave communications: Capacity
rotational invariance techniques,” IEEE Trans. Acoust., Speech, Signal analysis,” IEEE Trans. Wireless Commun., vol. 16, no. 5, pp. 2882–2890,
Process., vol. 37, no. 7, pp. 984–995, Jul. 1989. May 2017.
[13] C. Sturm and W. Wiesbeck, “Waveform design and signal processing [38] A. L. Swindlehurst, E. Ayanoglu, P. Heydari, and F. Capolino,
aspects for fusion of wireless communications and radar sensing,” Proc. “Millimeter–wave massive MIMO: The next wireless revolution?,” IEEE
IEEE, vol. 99, no. 7, pp. 1236–1259, Jul. 2011. Commun. Mag., vol. 52, no. 9, pp. 56–62, Sep. 2014.
[14] R. A. Monzingo and T. W. Miller, Introduction to Adaptive Arrays. New [39] S. Payami and F. Tufvesson, “Channel measurements and analysis for
York, NY, USA: Wiley, 1980. very large array systems at 2.6 GHz,” in Proc. IEEE 6th Eur. Conf.
[15] R. H. Roy III and B. Ottersten, “Spatial division multiple access wireless Antennas Propagat., Mar. 2012, pp. 433–437.
communication systems,” U.S. Patent 5, 515,378, May 7, 1996. [40] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and spectral effi-
[16] J. G. Andrews et al., “What will 5G be?,” IEEE J. Sel. Areas Commun., ciency of very large multiuser MIMO systems,” IEEE Trans. Commun.,
vol. 32, no. 6, pp. 1065–1082, Jun. 2014. vol. 61, no. 4, pp. 1436–1449, Apr. 2013.
[17] L. Lu, G. Y. Li, A. L. Swindlehurst, A. Ashikhmin, and R. Zhang, “An [41] A. Amiri, M. Angjelichinoski, E. de Carvalho, and R. W. Heath, “Ex-
overview of massive MIMO: Benefits and challenges,” IEEE J. Sel. Topics tremely large aperture massive MIMO: Low complexity receiver archi-
Signal Process., vol. 8 no. 5, pp. 742–758, Oct. 2014. tectures,” in Proc. IEEE Global Commun. Conf. Workshops, Dec. 2018,
[18] Y. Niu, Y. Li, D. Jin, L. Su, and A. V. Vasilakos, “A survey of millimeter pp. 1–6.
wave communications (mmWave) for 5G: opportunities and challenges,” [42] International Telecommunications Union Recommendation, “Propaga-
Wireless Netw., vol. 21, no. 8, pp. 2657–2676, Nov. 2015. tion data and prediction methods for the planning of indoor radiocom-
[19] R. W. Heath, N. Gonzalez-Prelcic, S. Rangan, W. Roh, and A. M. Sayeed, munication systems and radio local area networks in the frequency range
“An overview of signal processing techniques for millimeter wave MIMO 900 MHz to 100 GHz,” ITU-R, P.1238–7, 2012.
systems,” IEEE J. Sel. Topics Signal Process., vol. 10, no. 3, pp. 436–453, [43] Y. D. J. Bultitude and T. Rautiainen, “IST-4-027756 WINNER II D1.
Apr. 2016. 1.2 V1. 2 WINNER II Channel Models,” NOKIA, Tech. Rep., Sep. 30,
[20] R. B. Ertel, P. Cardieri, K. W. Sowerby, T. S. Rappaport, and J. H. Reed, 2007.
“Overview of spatial channel models for antenna array communication [44] METIS, “Mobile and wireless communications enablers for the twenty-
systems,” IEEE Pers. Commun., vol. 5, no. 1, pp. 10–22, Feb. 1998. twenty information society,” D1.4 METIS Channel Models, 2015. [On-
[21] E. S. Gopi, “Mathematical model of the time-varying wireless channel,” line]. Available: https://www.metis2020.com/documents/deliverables/
in Digital Signal Processing for Wireless Communication Using Matlab. [45] H. Xie, F. Gao, S. Zhang, and S. Jin, “A unified transmission strategy
Cham, Switzerland: Springer, 2016, pp. 1–50. for TDD/FDD massive MIMO systems with spatial basis expansion
[22] T. S. Chu and H. Hashemi, “True-time-delay-based multi-beam arrays,” model,” IEEE Trans. Veh. Technol., vol. 66, no. 4, pp. 3170–3184,
IEEE Trans. Microw. Theory Techn., vol. 61, no. 8, pp. 3072–3082, Apr. 2016.
Aug. 2013. [46] A. Kuchar, M. Tangemann, and E. Bonek, “A real-time DOA-based smart
[23] J. H. Brady and A. M. Sayeed, “Wideband communication with high- antenna processor,” IEEE Trans. Veh. Technol., vol. 51, no. 6, pp. 1279–
dimensional arrays: New results and transceiver architectures,” in Proc. 1293, Nov. 2002.
IEEE Int. Conf. Commun., May 2015, pp. 1042–1047. [47] K. Hug, K. Kalliola, and J. Laurila, “Spatial reciprocity of uplink and
[24] X. Gao, L. Dai, S. Zhou, A. M. Sayeed, and L. Hanzo, “Beamspace chan- downlink radio channels in FDD systems,” in Proc. COST, May 2002,
nel estimation for wideband millimeter-wave MIMO with lens antenna pp. 66–77.
array,” in Proc. IEEE Int. Conf. Commun., May 2018, pp. 1–6. [48] H. Xie, B. Wang, F. Gao, and S. Jin, “A full-space spectrum-sharing
[25] B. Wang, F. Gao, S. Jin, H. Lin, and G. Y. Li, “Spatial-wideband effect strategy for massive MIMO cognitive radio systems,” IEEE J. Sel. Areas
in massive MIMO systems,” in Proc. 23rd Asia-Pacific Conf. Commun., Commun., vol. 34, no. 10, pp. 2537–2549, Oct. 2016.
Dec. 2017, pp. 1–6. [49] D. Fan, F. Gao, G. Wang, Z. Zhong, and A. Nallanathan,” Angle do-
[26] A. Pitarokoilis, S. K. Mohammed, and E. G. Larsson, “On the optimality main signal processing-aided channel estimation for indoor 60-GHz
of single-carrier transmission in large-scale antenna systems,” IEEE TDD/FDD massive MIMO systems,” IEEE J. Sel. Areas Commun.,
Wireless Commun. Lett., vol. 1, no. 4, pp. 276–279, Aug. 2012. vol. 35, no. 9, pp. 1948–1961, Sep. 2017.
[27] Z. Chen and C. Yang, “Pilot decontamination in wideband massive [50] Y. Han, T. H. Hsu, C. K. Wen, K. K. Wong, and S. Jin, “Efficient downlink
MIMO systems by exploiting channel sparsity,” IEEE Trans. Wireless channel reconstruction for FDD multi-antenna systems,” IEEE Trans.
Commun., vol. 15, no. 7, pp. 5087–5100, Jul. 2016. Wireless Commun., vol. 18, no. 6, pp. 3161–3176, Apr. 2019.
WANG et al.: OVERVIEW OF ENHANCED MASSIVE MIMO WITH ARRAY SIGNAL PROCESSING TECHNIQUES 899

[51] D. Vasisht, S. Kumar, H. Rahul, and D. Katabi, “Eliminating channel [75] R. Shafin, L. Liu, Y. Li, A. Wang, and J. Zhang, “Angle and delay
feedback in next-generation cellular networks,” in Proc. Assoc. Comput. estimation for 3-D massive MIMO/FD-MIMO systems based on para-
Machinery Special Interest Group. Data Commun. Conf., Aug. 2016, metric channel modeling,” IEEE Trans. Wireless Commun., vol. 16, no. 8,
pp. 398–411. pp. 5370–5383, Aug. 2017.
[52] H. Xie, F. Gao, S. Jin, J. Fang, and Y. C. Liang, “Channel estimation for [76] R. O. Adeogun, P. D. Teal, and P. A. Dmochowski, “Channel prediction
TDD/FDD massive MIMO systems with channel covariance computing,” for millimeter wave MIMO systems in 3D propagation environments,”
IEEE Trans. Wireless Commun., vol. 17, no. 6, pp. 4206–4218, Jun. 2018. in Proc. 24th Int. Conf. Telecommun., May 2017, pp. 1–5.
[53] H. Yin, D. Gesbert, M. Filippou, and Y. Liu, “A coordinated approach [77] W. Zhang, F. Gao, and Q. Yin, “Blind channel estimation for MIMO-
to channel estimation in large-scale multiple-antenna systems,” IEEE J. OFDM systems with low order signal constellation,” IEEE Commun.
Sel. Areas Commun., vol. 31, no. 2, pp. 264–273, Feb. 2013. Lett., vol. 19, no. 3, pp. 499–502, Jan. 2015.
[54] R. C. de Lamare, “Massive MIMO systems: Signal processing challenges [78] I. Barhumi, G. Leus, and M. Moonen, “Optimal training design for
and future trends,” URSI Radio Sci. Bull. vol. 2013, no. 347, pp. 8–20, MIMO OFDM systems in mobile wireless channels,” IEEE Trans. Signal
Dec. 2013. Process., vol. 51, no. 6, pp. 1615–1624, Jun. 2003.
[55] C. Sun, X. Gao, S. Jin, M. Matthaiou, Z. Ding, and C. Xiao, “Beam divi- [79] M. Biguesh and A. B. Gershman, “Training-based MIMO channel es-
sion multiple access transmission for massive MIMO communications,” timation: A study of estimator tradeoffs and optimal training signals,”
IEEE Trans. Commun., vol. 63, no. 6, pp. 2170–2184, Jun. 2015. IEEE Trans Signal Process., vol. 54, no. 3, pp. 884–893, Mar. 2006.
[56] Z. Zhang, X. Wang, K. Long, A. V. Vasilakos, and L. Hanzo, “Large- [80] L. Tong, B. M. Sadler, and M. Dong, “Pilot-assisted wireless transmis-
scale MIMO based wireless backhaul in 5G network,” IEEE Wireless sions,” IEEE Signal Process. Mag., vol. 21, no. 6, pp. 12–25, Nov. 2004.
Commun., vol. 22, no. 5, pp. 58–66, Oct. 2015. [81] A. Adhikary, J. Nam, J.-Y. Ahn, and G. Caire, “Joint spatial division
[57] D. Gesbert, H. Bolcskei, D. A. Gore, and A. J. Paulraj, “Outdoor MIMO and multiplexing-the large-scale array regime,” IEEE Trans. Inf. Theory,
wireless channels: Models and performance predication,” IEEE Trans. vol. 59, no. 10, pp. 6441–6463, Oct. 2013.
Commun., vol. 50, no. 12, pp. 1926–1934, Dec. 2002. [82] J. Brady, N. Behdad, and A. M. Sayeed, “Beamspace MIMO for
[58] Y. P. Zhang, P. Wang, and A. Goldsmith, “Rainfall effect on the per- millimeter-wave communications: System architecture, modeling, anal-
formance of millimeter-wave MIMO systems,” IEEE Trans. Wireless ysis, and measurements,” IEEE Trans. Antennas Propagat., vol. 61, no. 7,
Commun., vol. 14, no. 9, pp. 4857–4866, Sep. 2015. pp. 3814–3827, Jul. 2013.
[59] A. M. Sayeed, “Deconstructing multiantenna fading channels,” IEEE [83] A. Sayeed and J. Brady, “Beamspace MIMO for high-dimensional mul-
Trans. Signal Process., vol. 50, no. 10, pp. 2563–2579, Oct. 2002. tiuser communication at millimeter-wave frequencies,” in Proc. IEEE
[60] W. U. Bajwa, J. Haupt, A. M. Sayeed, and R. Nowak, “Compressed chan- Global Commun. Conf., Dec. 2013, pp. 3679–3684.
nel sensing: A new approach to estimating sparse multipath channels,” [84] H. L. V. Trees, Optimum Array Processing: Part IV of Detection, Esti-
Proc. IEEE, vol. 98 no. 6, pp. 1058–1076, Jun. 2010. mation and Modulation Theory. Hoboken, NJ, USA: Wiley, 2002.
[61] C. Steffens, Y. Yang and M. Pesavento, “Multidimensional sparse re- [85] H. Lin, F. Gao, S. Jin, and G. Y. Li, “A new view of multi-user hybrid
covery for MIMO channel parameter estimation,” in Proc. Eur. Signal massive MIMO: Non-orthogonal angle division multiple access,” IEEE
Process. Conf., Budapest, Hungary, 2016, pp. 66–70. J. Sel. Areas Commun., vol. 35, no. 10, pp. 2268–2280, Jul. 2017.
[62] H. Xie, F. Gao, and S. Jin, “An overview of low-rank channel estimation [86] Z. Chen and C. Yang, “Pilot decontamination in wideband massive
for massive MIMO systems,” IEEE Access, vol. 4, pp. 7313–7321, MIMO systems by exploiting channel sparsity,” IEEE Trans. Wireless
Nov. 2016. Commun., vol. 15, no. 7, pp. 5087–5100, Jul. 2016.
[63] Y. Chi, L. L. Scharf, A. Pezeshki, and A. R. Calderbank, “Sensitivity to [87] D. L. Donoho, “Compressed sensing,” IEEE Trans. Inf. Theory, vol. 52,
basis mismatch in compressed sensing,” IEEE Trans. Signal Process., no. 4, pp. 1289–1306, Apr. 2006.
vol. 59, no. 5, pp. 2182–2195, May 2011. [88] R. G. Baraniuk, “Compressive sensing,” IEEE Signal Process. Mag.,
[64] P. Stoica and P. Babu, “Sparse estimation of spectral lines: Grid selection vol. 24, no. 4, pp. 118–121, Jul. 2007.
problems and their solutions,” IEEE Trans. Signal Process., vol. 60, no. 2, [89] A. M. Bruckstein, D. L. Donoho, and M. Elad, “From sparse solutions of
pp. 962–967, Feb. 2012. systems of equations to sparse modeling of signals and images,” SIAM
[65] A. Patwari and G. R. Reddy, “1D direction of arrival estimation using Rev., vol. 51, no. 1, pp. 34–81, Feb. 2009.
root-MUSIC and ESPRIT for dense uniform linear arrays,” in Proc. IEEE [90] Y. C. Eldar, “Sampling Theory: Beyond Bandlimited Systems,” Cam-
Recent Trends Electron., Inf. Commun. Technol., May 2017, pp. 667–672. bridge, UK: Cambridge University Press, Apr. 2015.
[66] G. Y. Suk, Y. G. Lim, H. B. Yilmaz, J. N. Shim, D. K. Kim, and C. [91] Y. Wang, Z. Tian, S. Feng, and P. Zhang, “A fast channel estimation
B. Chae, “Low complexity DoA estimation in millimeter wave MIMO approach for millimeter-wave massive MIMO systems,” in Proc. IEEE
with RF lens,” in Proc. IEEE Wireless Commun. Netw. Conf., Apr. 2018, Global Conf. Signal Inf. Process., Dec. 2016, pp. 1413–1417.
pp. 1–6. [92] J. Fang, F. Wang, Y. Shen, H. Li, and R. S. Blum, “Super-resolution
[67] B. Ottersten, P. Stoica, and R. Roy, “Covariance matching estimation compressed sensing for line spectral estimation: An iterative reweighted
techniques for array signal processing applications,” Digit. Signal Pro- approach,” IEEE Trans. Signal Process., vol. 64, no. 18, pp. 4649–4662,
cess., vol. 8, no. 3, pp. 185–210, Jul. 1998. Sep. 2016.
[68] B. L. Ng et al., “Fulfilling the promise of massive MIMO with 2D [93] E. Candes, J. Romberg, and T. Tao, “Robust uncertainty principles. Exact
active antenna array,” in Proc. IEEE Global Commun. Conf. Workshops, signal reconstruction from highly incomplete frequency information,”
Dec. 2012, pp. 691–696. IEEE Trans. Inf. Theory, vol. 52, no. 2, pp. 489–509, Feb. 2006.
[69] A. Hu, T. Lv, H. Gao, Z. Zhang, and S. Yang, “An ESPRIT-based [94] E. J. Candes and Y. Plan, “Near-ideal model selection by 1 minimiza-
approach for 2-D localization of incoherently distributed sources in tion,” Ann. Statist., vol. 37, pp. 2145–2177, 2009.
massive MIMO systems,” IEEE J. Sel. Topics Signal Process., vol. 8, [95] P. Schniter and A. Sayeed, “Channel estimation and precoder design
no, 5. pp. 996–1011, Oct. 2014. for millimeter-wave communications: The sparse way,” in Proc. IEEE
[70] Y. H. Nam et al., “Full-dimension MIMO (FD-MIMO) for next genera- Asilomar Conf. Signals, Syst. Comput., Nov. 2014, pp. 273–277.
tion cellular technology,” IEEE Commun. Mag., vol. 51, no. 6, pp. 172– [96] M. E. Eltayeb, T. Y. Al-Naffouri, and R. W. Heath, “Compressive sensing
179, Jun. 2013. for millimeter wave antenna array diagnosis,” IEEE Trans. Commun.,
[71] L. Liu, Y. Li, and J. Zhang, “DoA estimation and achievable rate analysis vol. 66, no. 6, pp. 2708–2721, Jun. 2018.
for 3D millimeter wave massive MIMO systems,” in Proc. IEEE Signal [97] X. Rao and V. K. N. Lau, “Distributed compressive CSIT estimation and
Process. Advances Wireless Commun., Jun. 2014, pp. 6–10. feed-back for FDD multi-user massive MIMO systems,” IEEE Trans.
[72] R. Shafin, L. Liu, J. Zhang, and Y. C. Wu, “DoA estimation and capacity Signal Process., vol. 62, no. 12, pp. 3261–3271, Jun. 2014.
analysis for 3-D millimeter wave massive-MIMO/FD-MIMO OFDM [98] E. J. Candès, “The restricted isometry property and its implications for
systems,” IEEE Trans. Wireless Commun., vol. 15, no. 10, pp. 6963–6978, compressed sensing,” Compte Rendus Math., vol. 346, no. 9–10, pp. 589–
Jul. 2016. 592, May 2008.
[73] M. Wax and A. Leshem, “Joint estimation of time delays and directions [99] Q. Guo, G. Gui, and F. Li, “Block-partition sparse channel estimation
of arrival of multiple reflections of a known signal,” IEEE Trans. Signal for spatially correlated massive MIMO systems,” in Proc. 8th Int. Conf.
Process., vol. 45, no. 10, pp. 2477–2484, Oct. 1997. Wireless Commun. Signal Process., Oct. 2016, pp. 1–4.
[74] S. Geng, J. Kivinen, X. Zhao, and P. Vainikainen, “Millimeter-wave [100] M. Wang, F. Gao, M. F. Flanagan, N. Shlezinger, and Y. C. Eldar, “A
propagation channel characterization for short-range wireless commu- block sparsity based estimator for mmWave massive MIMO channels
nications,” IEEE Trans Veh. Technol., vol. 58, no. 1, pp. 3–13, Jan. 2009. with beam squint,” Preprint, 2019, arXiv: 1904.12272.
900 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 13, NO. 5, SEPTEMBER 2019

[101] B. N. Bhaskar, G. Tang, and B. Recht, “Atomic norm denoising with [124] W. Xu, J. Liu, S. Jin, and X. Dong, “Spectral and energy efficiency of
applications to line spectral estimation,” IEEE Trans. Signal Process., multi-pair massive MIMO relay network with hybrid processing,” IEEE
vol. 61, no. 23, pp. 5987–5999, Dec. 2013. Trans. Commun., vol. 65, no. 9, pp. 3794–3809, Sep. 2017.
[102] Y. Wang, P. Xu, and Z. Tian, “Efficient channel estimation for massive [125] X. Gao, L. Dai, S. Han, I. Chih-Lin, and R. W. Heath, “Energy-efficient
MIMO systems via truncated two-dimensional atomic norm minimiza- hybrid analog and digital precoding for mmWave MIMO systems with
tion,” in Proc. IEEE Int. Conf. Commun., May 2017, pp. 1–6. large antenna arrays,” IEEE J. Sel. Areas Commun., vol. 34, no. 4, pp. 998–
[103] Z. Tian, Z. Zhang, and Y. Wang, “Low-complexity optimization for two- 1009, Apr. 2016.
dimensional direction-of-arrival estimation via decoupled atomic norm [126] S. Kutty and D. Sen, “Beamforming for millimeter wave communica-
minimization,” in Proc IEEE Int. Conf. Acoust., Speech Signal Process., tions: An inclusive survey,” IEEE Commun. Surveys Tuts., vol. 18, no. 2,
Mar. 2017, pp. 3071–3075. pp. 949–973, Jan. 2016.
[104] Z. Zhang, Y. Wang, and Z. Tian, “Efficient two-dimensional line spec- [127] Y. Cui, X. Fang, and L. Yan, “Hybrid spatial modulation beamforming for
trum estimation based on decoupled atomic norm minimization,” Signal mmWave railway communication systems,” IEEE Trans. Veh. Technol.,
Process., vol. 163, pp. 95–106, 2019. vol. 65, no. 12, pp. 9597–9606, Dec. 2016.
[105] Y. Tsai, L. Zheng, and X. Wang, “Millimeter-wave beamformed full- [128] X. Chen, Z. Zhang, L. Wu, J. Dang, P. Lu, and C. Sun, “Robust beam
dimensional MIMO channel estimation based on atomic norm minimiza- management scheme based on simple 2-D DOA estimation,” in Proc. 9th
tion,” IEEE Trans. Commun., vol. 66, no. 12, pp. 6150–6163, Dec. 2018. Int. Conf. Wireless Commun. Signal Process., Oct. 2017, pp. 1–6.
[106] J. Deng, O. Tirkkonen, and C. Studer, “Mmwave channel estimation via [129] A. B. Gershman, N. D. Sidiropoulos, S. Shahbazpanahi, M. Bengtsson,
atomic norm minimization for multi-user hybrid precoding,” in Proc. and B Ottersten, “Convex optimization-based beamforming,” IEEE Sig-
IEEE Wireless Commun. Netw. Conf.), Apr. 2018, pp. 1–6. nal Process. Mag., vol. 27, no. 3, pp. 62–75, May 2010.
[107] H. Chu, L. Zheng, and X. Wang, “Super-resolution mmWave chan- [130] J. Eisenbeis, M. Krause, T. Mahler, S. Scherr, and T. Zwick, “Path based
nel estimation using atomic norm minimization,” Preprint, 2018, MIMO channel model for hybrid beamforming architecture analysis,” in
arXiv:1801.07400. Proc. 11th German Microw. Conf., Freiburg, Germany, 2018, pp. 311–
[108] D. P. Wipf and B. D. Rao, “An empirical Bayesian strategy for solving 314.
the simultaneous sparse approximation problem,” IEEE Trans. Signal [131] J. Zhao, F. Gao, W. Jia, S. Zhang, S. Jin and H. Lin, “Angle domain hybrid
Process., vol. 55, no. 7, pp. 3704–3716, Jul. 2007. precoding and channel tracking for millimeter wave massive MIMO Sys-
[109] T. L. Hansen, M. A. Badiu, B. H. Fleury, and B. D. Rao, “A sparse tems,” IEEE Trans. Wireless Commun., vol. 16, no. 10, pp. 6868–6880,
Bayesian learning algorithm with dictionary parameter estimation,” in Oct. 2017.
Proc. IEEE 8th Sensor Array Multichannel Signal Process. Workshops, [132] A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, “Channel
Jun. 2014, pp. 385–388. estimation and hybrid precoding for millimeter wave cellular sys-
[110] M. Carlin, P. Rocca, G. Oliveri, F. Viani, and A. Massa, “Directions- tems,” IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 831–846,
of-arrival estimation through Bayesian compressive sensing strategies,” Oct. 2014.
IEEE Trans. Antennas Propagat., vol. 61, no. 7, pp. 3828–3838, [133] Z. Xiao, P. Xia, and X. G. Xia, “Channel estimation and hybrid precoding
Jul. 2013. for millimeter-wave MIMO systems: A low-complexity overall solution,”
[111] L. Wang, L. Zhao, G. Bi, C. Wan, L. Zhang, and H. Zhang, “Novel wide- IEEE Access, vol. 5, pp. 16100–16110, Jan. 2017.
band DOA estimation based on sparse Bayesian learning with Dirichlet [134] C. B. Peel, B. M. Hochwald, and A. L. Swindlehurst, “A
process priors,” IEEE Trans. Signal Process., vol. 64, no 2, pp. 275–289, vector-perturbation technique for near-capacity multiantenna multiuser
Jun. 2016. communication–Part I: Channel inversion and regularization,” IEEE
[112] A. Das and T. J. Sejnowski, “Narrowband and wideband off-grid Trans. Commun., vol. 53, no. 1, pp. 195–202, Jan. 2005.
direction-of-arrival estimation via sparse Bayesian learning,” IEEE J. [135] J. Du, W. Xu, C. Zhao, and L. Vandendorpe, “Hybrid beamforming design
Oceanic Eng., vol. 43, no. 1, pp. 108–118, Jan. 2018. for multiuser massive MIMO-OFDM systems,” in Proc. 15th Int. Symp.
[113] M. Jian, F. Gao, Z. Tian, S. Jin, and S. Ma, “Angle-domain aided UL/DL Wireless Commun. Syst., Lisbon, Portugal, 2018, pp. 1–6.
channel estimation for wideband mmWave massive MIMO systems with [136] S. Hur, T. Kim, D. J. Love, J. V. Krogmeier, T. A. Thomas, and A. Ghosh,
beam squint,” IEEE Trans. Wireless Commun., vol. 18, no. 7, pp. 3515– “Millimeter wave beamforming for wireless backhaul and access in small
3527, May 2019. cell networks,” IEEE Trans. Commun., vol. 61, no. 10, pp. 4391–4403,
[114] M. Wax and T. Kailath, “Detection of signals by information theoretic Oct. 2013.
criteria,” IEEE Trans. Acoust., Speech, Signal Process., vol. ASSP-33, [137] C. Liu, M. Li, S. V. Hanly, I. B. Collings, and P. Whiting, “Mil-
no. 2, pp. 387–392, Apr. 1985. limeter wave beam alignment: Large deviations analysis and design
[115] W. Roh et al., “Millimeter-wave beamforming as an enabling technology insights,” IEEE J. Sec. Areas commun., vol. 35, no. 7, pp. 1619–1631,
for 5G cellular communications: Theoretical feasibility and prototype Jul. 2017.
results,” IEEE Commun. Mag., vol. 52, no. 2, pp. 106–113, Feb. 2014. [138] J. Zhang, Y. Huang, Q. Shi, J. Wang, and L. Yang, “Codebook design
[116] J. G. Andrews, T. Bai, M. N. Kulkarni, A. Alkhateeb, A. K. Gupta, and R. for beam alignment in millimeter wave communication systems,” IEEE
W. Heath, “Modeling and analyzing millimeter wave cellular systems,” Trans. Commun., vol. 65, no. 11, pp. 4980–4995, Nov. 2017.
IEEE Trans. Commun., vol. 65, no. 1, pp. 403–430, Jan. 2017. [139] V. Raghavan, J. Cezanne, S. Subramanian, A. Sampath, and O. Koymen,
[117] L. You, X. Gao, G. Y. Li, X. G. Xia, and N. Ma, “BDMA for millimeter- “Beamforming tradeoffs for initial UE discovery in millimeter-wave
wave/terahertz massive MIMO transmission with per-beam synchroniza- MIMO systems,” IEEE J. Sel. Topics Signal Process., vol. 10, no. 3,
tion,” IEEE J. Sel. Areas Commun., vol. 35, no. 7, pp. 1550–1563, pp. 543–559, Apr. 2016.
Jul. 2017. [140] G. B. Giannakis and C. Tepedelenlioglu, “Basis expansion models and
[118] T. Hwang, C. Yang, G. Wu, S. Li, and G. Y. Li, “OFDM and its diversity techniques for blind identification and equalization of timevary-
wireless applications: A survey,” IEEE Trans. Veh. Technol., vol. 58, ing channels,” Proc. IEEE, vol. 86, no. 10, pp. 1969–1986, Oct. 1998.
no. 4, pp. 1673–1694, May 2009. [141] R. Hadani et al., “Orthogonal time frequency space modulation,” in Proc.
[119] E. Dahlman, S. Parkvall, and J. Skold, “4G: LTE/LTE-advanced for IEEE Wireless Commun. Netw. Conf., Mar. 2017, pp. 1–6.
mobile broadband,” 2nd ed. Waltham, MA, USA: Academic, 2014. [142] R. Hadani and A. Monk, “OTFS: A new generation of modulation
[120] W. Zhang, F. Gao, S. Jin, and H. Lin, “Frequency synchronization for addressing the challenges of 5G,” Preprint, 2018, arXiv:1802.02623.
uplink massive MIMO systems,” IEEE Trans. Wireless Commun., vol. 17, [143] V. Khammammetti and S. K. Mohammed, “OTFS-based multiple-access
no. 1, pp. 235–249, Jan. 2018. in high Doppler and delay spread wireless channels,” IEEE Wireless
[121] Y. Ge, W. Zhang, F. Gao, and H. Min, “Angle-domain approach for Commun. Lett., vol. 8, no. 2, pp. 528–531, Apr. 2019.
parameter estimation in high-mobility OFDM with fully/partly calibrated [144] S. Rakib and R. Hadani, “Methods of operating and implementing
massive ULA,” IEEE Trans. Wireless Commun., vol. 18, no. 1, pp. 591– wireless OTFS communciations systems,” U.S. Patent 9, 634, 719,
607, Jan. 2019. Apr. 25, 2017.
[122] W. Guo, W. Zhang, P. Mu, F. Gao, and H. Lin, “High-mobility wideband [145] R. Hadani et al., “Orthogonal time frequency space (OTFS) modula-
massive MIMO communications: Doppler compensation, analysis and tion for millimeter-wave communications systems,” in Proc. IEEE Int.
scaling law,” 2018, arXiv:1807.07438. Microw. Symp., Jun. 2017, pp. 681–683.
[123] W. Tan, M. Matthaiou, S. Jin, and X. Li, “Spectral efficiency of DFT- [146] M. K. Ramachandran and A. Chockalingam, “MIMO-OTFS in high-
based processing hybrid architectures in massive MIMO,” IEEE Wireless Doppler fading channels: signal detection and channel estimation,” in
Commun. Lett., vol. 6, no 5, pp. 586–589, Oct. 2017. Proc. IEEE Global Commun. Conf., Dec. 2018, pp. 206–212.
WANG et al.: OVERVIEW OF ENHANCED MASSIVE MIMO WITH ARRAY SIGNAL PROCESSING TECHNIQUES 901

[147] A. Farhang, A. RezazadehReyhani, L. E. Doyle, and B. Farhang- Borou- Feifei Gao (M’09–SM’14) received the B.Eng. de-
jeny, “Low complexity modem structure for OFDM-based orthogonal gree from Xi’an Jiaotong University, Xi’an, China,
time frequency space modulation,” IEEE Wireless Commun. Lett., vol. 7, in 2002; the M.Sc. degree from McMaster Univer-
no. 3, pp. 344–347, Jun. 2018. sity, Hamilton, ON, Canada, in 2004; and the Ph.D.
[148] P. Raviteja, K. T. Phan, Y. Hong, and E. Viterbo, “Interference cancella- degree from the National University of Singapore,
tion and iterative detection for orthogonal time frequency space modula- Singapore, in 2007. He was a Research Fellow with
tion,” IEEE Trans. Wireless Commun., vol. 17, no. 10, pp. 65601–6515, the Institute for Infocomm Research (I2R), A*STAR,
Oct. 2018. Singapore in 2008 and an Assistant Professor with the
[149] A. RezazadehReyhani, A. Farhang, M. Ji, R. R. Chen, and B. Farhang- School of Engineering and Science, Jacobs Univer-
Boroujeny, “Analysis of discrete-time MIMO OFDM-based orthogonal sity, Bremen, Germany from 2009 to 2010. In 2011,
time frequency space modulation,” in Proc. IEEE Int. Conf. Commun., he joined the Department of Automation, Tsinghua
May 2018, pp. 1–6. University, Beijing, China, where he is currently an Associate Professor.
[150] W. Shen, L. Dai, J. An, P. Fan, and R. W. Heath, “Channel estimation He has authored or coauthored over 120 refereed IEEE journal papers and over
for orthogonal time frequency space (OTFS) massive MIMO,” Preprint, 120 IEEE conference proceeding papers. His research areas include communica-
2019, arXiv:1903.09441. tion theory, signal processing for communications, array signal processing, and
[151] A. Liu, V. K. N. Lau, and W. Dai, “Exploiting burst-sparsity in massive convex optimizations, with particular interests in MIMO techniques, multicarrier
MIMO with partial channel support information,” IEEE Trans. Wireless communications, cooperative communication, and cognitive radio networks.
Commun., vol. 15, no. 11, pp. 7820–7830, Nov. 2016. Prof. Gao has served as an Editor for IEEE TRANSACTIONS ON
[152] Y. Zeng, L. Yang, and R. Zhang, “Multi-user millimeter wave MIMO with WIRELESS COMMUNICATIONS, IEEE SIGNAL PROCESSING LETTERS, IEEE
full-dimensional lens antenna array,” IEEE Trans. Wireless Commun., COMMUNICATIONS LETTERS, IEEE WIRELESS COMMUNICATIONS LETTERS, In-
vol. 17, no. 4, pp. 2800–2814, Apr. 2018. ternational Journal on Antennas and Propagations, and China Communications.
[153] T. Wang, C. K. Wen, H. Wang, F. Gao, T. Jiang, and S. Jin, “Deep He has also served as the Symposium Cochair for the 2019 IEEE Conference
learning for wireless physical layer: Opportunities and challenges,” China on Communications (ICC), the 2018 IEEE Vehicular Technology Conference
Commun., vol. 14, no. 11, pp. 92–111, Nov. 2017. Spring (VTC), the 2015 ICC, the 2014 IEEE Global Communications Confer-
[154] Z. Qin, H. Ye, G. Y. Li, and B. H. F. Juang, “Deep learning in physical layer ence, and the 2014 IEEE VTC Fall,
communications,” IEEE Wireless Commun., vol. 26, no. 2, pp. 93–99,
Apr. 2019.
[155] Q. Mao, F. Hu, and Q. Hao, “Deep learning for intelligent wireless net-
works: A comprehensive survey,” IEEE Commun. Surveys Tuts., vol. 20,
Shi Jin (S’06–M’07) received the B.S. degree in com-
no. 4, pp. 2595–2621, Jun. 2018. munications engineering from the Guilin University
[156] H. He, C. K. Wen, S. Jin, and G. Y. Li, “Deep learning-based channel
of Electronic Technology, Guilin, China, in 1996, the
estimation for beamspace mmWave massive MIMO systems,” IEEE
M.S. degree from the Nanjing University of Posts and
Wireless Commun. Lett., vol. 7, no. 5, pp. 852–855, May 2018.
Telecommunications, Nanjing, China, in 2003, and
[157] H. Huang, J. Yang, H. Huang, Y. Song, and G. Gui, “Deep learning for the Ph.D. degree in communications and information
super-resolution channel estimation and DOA estimation based massive
systems from Southeast University, Nanjing, in 2007.
MIMO system,” IEEE Trans Veh. Technol., vol. 67, no. 9, pp. 8549–8560.
From June 2007 to October 2009, he was a Research
Jan. 2018.
Fellow with the Adastral Park Research Campus,
[158] Y. Yang, F. Gao, X. Ma, and S. Zhang, “Deep learning-based channel
University College London, London, U.K. He is cur-
estimation for doubly selective fading channels,” IEEE Access, vol. 7,
rently with the Faculty of the National Mobile Com-
pp. 36579–36589, Mar. 2019.
munications Research Laboratory, Southeast University, Nanjing, China. His
[159] T. Wang, C. K. Wen, S. Jin, and G. Y. Li, “Deep learning-based CSI research interests include space time wireless communications, random matrix
feedback approach for time-varying massive MIMO channels,” IEEE
theory, and information theory. He serves as an Associate Editor for the IEEE
Wireless Commun. Lett. vol. 8, no. 2, pp. 416–419, Apr. 2019.
TRANSACTIONS ON WIRELESS COMMUNICATIONS, and IEEE COMMUNICATIONS
[160] Y. Long, Z. Chen, J. Fang, and C. Tellambura, “Data-driven-based analog
LETTERS, and IET Communications. Dr. Jin and his coauthors have been awarded
beam selection for hybrid beamforming under mm-Wave channels,” IEEE the 2011 IEEE Communications Society Stephen O. Rice Prize Paper Award in
J. Sel. Topics Signal Process., vol. 12, no. 2, pp. 340–352, May 2018.
the field of communication theory and a 2010 Young Author Best Paper Award
[161] C. Antón-Haro and X. Mestre, “Learning and data-driven beam selection
by the IEEE Signal Processing Society.
for mmWave communications: An angle of arrival-based approach,”
IEEE Access, vol. 7, pp. 20404–20415, Jan. 2019.
[162] A. Klautau, P. Batista, N. González-Prelcic, Y. Wang, and R. W. Heath,
“5G MIMO data for machine learning: Application to beam selection us-
ing deep learning,” in Proc. IEEE Inf. Theory Appl. Workshops, Feb. 2018, Hai Lin (M’05–SM’12) received the B.E. degree
pp. 1–9. from Shanghai JiaoTong University, Shanghai, China,
[163] D. Fan, F. Gao, B. Ai, G. Wang, Z. Zhong, and A. Nallanathan, “Self- in 1993, the M.E. degree from the University of
positioning for UAV swarm via RARE direction-of-arrival estimator,” in the Ryukyus, Nishihara, Japan, in 2000, and the
Proc. IEEE Global Commun. Conf., Dec. 2018, pp, 1–6. Dr. Eng. degree from Osaka Prefecture University,
Sakai, Japan, in 2005. Since 2000, he was a Research
Associate with the Graduate School of Engineering,
Osaka Prefecture University, where he is a Professor
now. His research interests are in signal processing
for communications, wireless communications, and
Mingjin Wang (S’17) received the B.Eng. degree statistical signal processing.
from Xi’an Jiaotong University, Xi’an, China, in Dr. Lin has acted many times as Technical Program Committee Member for
2017. He is currently working toward the Ph.D. IEEE International Conference on Communications (ICC), Global Communica-
degree at the Tsinghua National Laboratory for In- tions Conference (GLOBECOM), Wireless Communications and Networking
formation Science and Technology (TNList), Depart- Conference (WCNC), Vehicular Technology Conference (VTC), etc. He served
ment of Automation, Tsinghua University, Beijing, as a Technical Program Cochair for the Wireless Communications Symposium
China. His current research interests include signal of IEEE ICC 2011, the Signal Processing for Communications Symposium
processing for communications and broadband wire- of IEEE ICC 2013, and the Wireless Communications Symposium of IEEE
less communications, with a focus on Massive MIMO GLOBECOM 2013. Dr. Lin was an Editor for the IEEE TRANSACTIONS ON
systems, array signal processing, and compressed WIRELESS COMMUNICATIONS, and now is an Editor for the IEEE TRANSACTIONS
sensing in communications. ON VEHICULAR TECHNOLOGY. He is a member of the IEICE.