Intelligent Reflecting Surface Enhanced Wireless Network Via Joint Active and Passive Beamforming
Intelligent Reflecting Surface Enhanced Wireless Network Via Joint Active and Passive Beamforming
Intelligent Reflecting Surface Enhanced Wireless Network Via Joint Active and Passive Beamforming
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2019.2936025, IEEE
Transactions on Wireless Communications
Abstract—Intelligent reflecting surface (IRS) is a revolutionary and complex signal processing are needed for achieving high-
and transformative technology for achieving spectrum and energy performance communication at mmWave frequencies, espe-
efficient wireless communication cost-effectively in the future. cially when massive MIMO is employed to exploit the small
Specifically, an IRS consists of a large number of low-cost passive
elements each being able to reflect the incident signal indepen- wavelengths. Moreover, adding an excessively large number
dently with an adjustable phase shift so as to collaboratively of active components such as small-cell BSs/relays/remote
achieve three-dimensional (3D) passive beamforming without the radio heads (RRHs) in wireless networks also causes a more
need of any transmit radio-frequency (RF) chains. In this paper, aggravated interference issue. As such, innovative research on
we study an IRS-aided single-cell wireless system where one IRS finding both spectrum and energy efficient techniques with
is deployed to assist in the communications between a multi-
antenna access point (AP) and multiple single-antenna users. low hardware cost is still imperative for realizing a sustainable
We formulate and solve new problems to minimize the total wireless network evolution with scalable cost in the future [4].
transmit power at the AP by jointly optimizing the transmit In this paper, intelligent reflecting surface (IRS) is proposed
beamforming by active antenna array at the AP and reflect as a promising new solution to achieve the above goal. Specif-
beamforming by passive phase shifters at the IRS, subject to ically, IRS is a planar array consisting of a large number of re-
users’ individual signal-to-interference-plus-noise ratio (SINR)
constraints. Moreover, we analyze the asymptotic performance configurable passive elements (e.g., low-cost printed dipoles),
of IRS’s passive beamforming with infinitely large number of where each of the elements is able to induce a certain phase
reflecting elements and compare it to that of the traditional active shift (controlled by an attached smart controller) independent-
beamforming/relaying. Simulation results demonstrate that an ly on the incident signal, thus collaboratively changing the
IRS-aided MIMO system can achieve the same rate performance reflected signal propagation [5]. Although passive reflecting
as a benchmark massive MIMO system without using IRS, but
with significantly reduced active antennas/RF chains. We also surfaces have found a variety of applications in radar systems,
draw useful insights into optimally deploying IRS in future remote sensing, and satellite/deep-space communications, they
wireless systems. were rarely used in mobile wireless communication. This is
Index Terms—Intelligent reflecting surface, joint active and mainly because traditional reflecting surfaces only have fixed
passive beamforming, phase shift optimization. phase shifters once fabricated, which are unable to cater to
the dynamic wireless channels arising from user mobility.
However, recent advances in RF micro electromechanical
I. I NTRODUCTION systems (MEMS) and metamaterial (e.g., metasurface) have
To achieve 1,000-fold network capacity increase and ubiq- made the reconfigurability of reflecting surfaces possible, even
uitous wireless connectivity for at least 100 billion devices by controlling the phase shifters in real time [6]. By smartly
in the forthcoming fifth-generation (5G) networks, a variety adjusting the phase shifts of all passive elements at the IRS,
of wireless technologies have been proposed and thoroughly the reflected signals can add coherently with the signals from
investigated in the last decade, including most prominently the other paths at the desired receiver to boost the received signal
ultra-dense network (UDN), massive multiple-input multiple- power or destructively at non-intended receivers to suppress
output (MIMO), and millimeter wave (mmWave) communi- interference as well as enhancing security/privacy [5].
cation [2]. However, the network energy consumption and It is worth noting that the proposed IRS differs significantly
hardware cost still remain critical issues in practical systems from other related existing technologies such as amplify-and-
[3]. For example, UDNs almost linearly scale up the circuit forward (AF) relay, backscatter communication, and active
and cooling energy consumption with the number of deployed intelligent surface based massive MIMO. First, compared to
base stations (BSs), while costly radio frequency (RF) chains the AF relay that assists in source-destination transmission
by amplifying and regenerating signals, IRS does not use a
Manuscript received October 12, 2018; revised February 23, 2019, May transmitter module but only reflects the received signals as a
15, 2019, and August 6, 2019; accepted August 12, 2019. Date of publication
xxx, 2019; date of current version xxx, 2019. This paper was presented in passive array, which thus incurs no transmit power consump-
part at the IEEE GLOBECOM 2018 [1]. The associate editor coordinating tion.1 Furthermore, active AF relay usually operates in half-
the review of this paper and approving it for publication was K. Kansanen. duplex (HD) mode and thus is less spectrally efficient than the
(Corresponding author: Qingqing Wu.)
The authors are with the Department of Electrical and Computer En- proposed IRS operating in full-duplex (FD) mode. Although
gineering, National University of Singapore, Singapore 117583 (e-mail:
[email protected]; [email protected]). 1 Although using devices like MEMs as mentioned previously to adjust the
Color versions of one or more of the figures in this paper are available phase shifts at the IRS requires some power consumption, it is practically
online at http://ieeexplore.ieee.org. negligible as compared to the much higher transmit power of active commu-
Digital Object Identifier xxx. nication devices.
1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2019.2936025, IEEE
Transactions on Wireless Communications
TABLE I
C OMPARISON OF IRS WITH OTHER RELATED TECHNOLOGIES .
No. of trans-
Operating Hardware Energy xx con-
Technology Duplex mit RF chains Role
mechanism cost sumption
needed
Passive,
IRS Full duplex 0 Low Low Helper
reflect
Passive,
Backscatter Full duplex 0 Very low Very low Source
reflect
Active,
Half/full du-
MIMO relay receive andxxxx N High High Helper
plex
transmit
Active, Half/full du- Source/ Des-
Massive MIMO N Very high Very high
transmit/receive plex tination
1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2019.2936025, IEEE
Transactions on Wireless Communications
passive array with significantly reduced energy consumption suboptimal zero-forcing (ZF) based precoding at the AP to
and hardware cost. simplify the optimization of passive phase shifters, while in
Moreover, under the general multiuser setup, an IRS-aided this paper we optimize AP transmit precoding jointly with
system will benefit from two main aspects: the beamforming IRS’s phase shifts. As such, the algorithm proposed in [15]
of desired signal as in the single-user case as well as the is not applicable to solving the formulated problems in this
spatial interference suppression among the users. Specifically, paper.
a user near the IRS is expected to be able to tolerate more The rest of this paper is organized as follows. Section II
interference from the AP as compared to the user farther away introduces the system model and the problem formulation
from the IRS, because the phase shifts of the IRS can be for designing the IRS-aided wireless network. In Sections
tuned such that the interference reflected by the IRS can add III and IV, we propose efficient algorithms to solve the
destructively with that from the AP-user link at the near user formulated problems in the single-user and multiuser cases,
to suppress its overall received interference. This thus provides respectively. Section V presents numerical results to evaluate
more flexibility for designing the transmit beamforming at the performance of the proposed designs. Finally, we conclude
the AP for serving the other users outside the IRS’s covered the paper in Section VI.
region, so as to improve the SINR performance of all users Notations: Scalars are denoted by italic letters, vectors
in the system. Therefore, the transmit beamforming at the and matrices are denoted by bold-face lower-case and upper-
AP needs to be jointly designed with the phase shifts at case letters, respectively. Cx×y denotes the space of x × y
the IRS based on all the AP-IRS, IRS-users, and AP-users complex-valued matrices. For a complex-valued vector x, kxk
channels in order to fully reap the network beamforming denotes its Euclidean norm, arg(x) denotes a vector with each
gain. However, this design problem is difficult to be solved element being the phase of the corresponding element in x,
optimally in general, due to the non-convex SINR constraints and diag(x) denotes a diagonal matrix with each diagonal
as well as the signal unit-modulus constraints imposed by element being the corresponding element in x. The distribution
passive phase shifters. Although beamforming optimization of a circularly symmetric complex Gaussian (CSCG) random
under unit-modulus constraints has been studied in the re- vector with mean vector x and covariance matrix Σ is denoted
search on constant-envelope precoding [11], [12] as well as by CN (x, Σ); and ∼ stands for “distributed as”. For a
hybrid digital/analog processing [13], [14], such designs are square matrix S, tr(S) and S −1 denote its trace and inverse,
mainly restricted to either the transmitter or the receiver side, respectively, while S 0 means that S is positive semi-
which are not applicable to our considered joint active and definite. For any general matrix M , M H , rank(M ), and
passive beamforming optimization at both the AP and IRS. Mi,j denote its conjugate transpose, rank, and (i, j)th element,
To tackle this new problem, we first consider a single-user respectively. I and 0 denote an identity matrix and an all-
setup and apply the semidefinite relaxation (SDR) technique to zero matrix, respectively, with appropriate dimensions. E(·)
obtain a high-quality approximate solution as well as a lower denotes the statistical expectation. Re{·} denotes the real part
bound of the optimal value to evaluate the tightness of approx- of a complex number.
imate solutions. To reduce the computational complexity, we
further propose an efficient algorithm based on the alternating II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
optimization of the phase shifts and transmit beamforming
vector in an iterative manner, where their optimal solutions A. System Model
are derived in closed-form with the other being fixed. Then, As shown in Fig. 1, we consider the IRS-aided downlink
we extend our designs for the single-user case to the general communications in a single-cell network where an IRS is
multiuser setting, and propose two algorithms to obtain sub- deployed to assist in the communications from a multi-antenna
optimal solutions that also offer different tradeoffs between AP to K single-antenna users over a given frequency band.
performance and complexity. Numerical results demonstrate The set of the users is denoted by K. The number of transmit
that the required transmit power at the AP to meet users’ antennas at the AP and that of reflecting units at the IRS
SINR targets can be considerably reduced by deploying the are denoted by M and N , respectively. The IRS is equipped
IRS as compared to the conventional setup without using IRS with a controller that coordinates its switching between two
for both single-user and multiuser setups. In particular, for working modes, i.e., receiving mode for channel estimation
serving a single-user in the vicinity of the IRS, it is shown and reflecting mode for data transmission [8]. Due to the
that the AP’s transmit power decreases with the number of high path loss, it is assumed that the power of the signals
reflecting elements N at the IRS in the order of N 2 when N that are reflected by the IRS two or more times is negligible
is sufficiently large, which is consistent with the performance and thus ignored. To characterize the theoretical performance
scaling law derived analytically. Note that in [15], the authors gain brought by the IRS, we assume that the channel state
also considered the use of passive intelligent mirror (analogous information (CSI) of all channels involved is perfectly known
to IRS) to enhance the sum-rate in a multiuser system. This at the AP. In addition, the quasi-static flat-fading model is
paper differs from [15] in the following two main aspects. adopted for all channels. Since the IRS is a passive reflecting
First, to simplify the system model and algorithm design, device, we consider a time-division duplexing (TDD) protocol
[15] ignored the direct channels from the AP to users, while for uplink and downlink transmissions and assume channel
this paper considers the more general setting with the AP- reciprocity for the CSI acquisition in the downlink based on
user direct channels considered. Second, [15] adopted the the uplink training.
1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2019.2936025, IEEE
Transactions on Wireless Communications
The baseband equivalent channels from the AP to IRS, from to individual SINR constraints at all users. Accordingly, the
the IRS to user k, and from the AP to user k are denoted by problem is formulated as
G ∈ CN ×M , hH r,k ∈ C
1×N
, and hHd,k ∈ C
1×M
, respectively, K
with k = 1, · · · , K. It is worth noting that the reflected chan-
X
(P1) : min kwk k2 (3)
nel from the AP to each user via the IRS is usually referred W ,θ
k=1
to as a dyadic backscatter channel in RFID communications |(hH H 2
r,k ΘG + hd,k )wk |
[16], which behaves different from the AP-user direct channel. s.t. PK ≥ γk , ∀k, (4)
Specifically, each element of the IRS receives the superposed j6=k |(hH H 2 2
r,k ΘG + hd,k )wj | + σk
multi-path signals from the transmitter, and then scatters the 0 ≤ θn ≤ 2π, n = 1, · · · , N, (5)
combined signal with adjustable amplitude and/or phase as if
where γk > 0 is the minimum SINR requirement of user
from a single point source. Let θ = [θ1 , · · · , θN ] and define
k. Although the objective function of (P1) and constraints
a diagonal matrix Θ = diag(β1 ejθ1 , · · · , βN ejθN ) (with
in (5) are convex, it is challenging to solve (P1) due to the
j denoting the imaginary unit) as the reflection-coefficients
non-convex constraints in (4) where the transmit beamforming
matrix of the IRS, where θn ∈ [0, 2π) and βn ∈ [0, 1] denote
and phase shifts are coupled. In general, there is no standard
the phase shift2 and the amplitude reflection coefficient3 of
method for solving such non-convex optimization problems
the nth element of the IRS, respectively. The composite AP-
optimally. Nevertheless, in the next section, we apply the SDR
IRS-user channel is thus modeled as a concatenation of three
and alternating optimization techniques, respectively, to solve
components, namely, the AP-IRS link, IRS reflection with
(P1) approximately for the single-user case, which are then
phase shifts, and IRS-user link.
generalized to the multiuser case. Prior to solving problem
In this paper, we consider linear transmit precoding at
(P1), we present a sufficient condition for its feasibility as
the AP where each user is assigned with one dedicated
follows. Let H = [h1 , · · · , hK ] ∈ CM ×K where hH k =
beamforming vector. Hence, the complex baseband transmitted
PK hHr,k ΘG + h H
d,k , ∀k.
signal at the AP can be expressed as x = k=1 wk sk , where
Proposition 1: Problem (P1) is feasible for any finite user
sk denotes the transmitted data for user k and wk ∈ CM ×1
SINR targets γk ’s if rank(GH Hr + Hd ) = K.
is the corresponding beamforming vector. It is assumed that
Proof: If rank(GH Hr + Hd ) = K, the (right) pseudo
sk , k = 1, · · · , K, are independent random variables with
inverse of H H = HrH ΘG + HdH exists with Θ = I and the
zero mean and unit variance (normalized power). The signal
precoding matrix W at the AP can be set as
received at user k from both the AP-user and AP-IRS-user
1
channels is then expressed as W = H(H H H)−1 diag(γ1 σ12 , · · · , γk σk2 ) 2 . (6)
K
X It is easy to verify that the above solution allows all users to
yk = (hH H
r,k ΘG + hd,k ) wj sj + nk , k = 1, · · · , K, (1) achieve their corresponding γk ’s and thus (P1) is feasible.
j=1 Thanks to the additional AP-IRS-user link, the rank con-
dition in Proposition 1 is practically easier to be satisfied in
where nk ∼ CN (0, σk2 ) denotes the additive white Gaussian an IRS-aided system, as compared to that in the case without
noise (AWGN) at the user k’s receiver. Accordingly, the SINR the IRS, i.e., rank(Hd ) = K. For instance, if the AP-user
of user k is given by direct channels of two users lie in the same direction, then
|(hH H 2 rank(Hd ) = K does not hold. While the rank condition in an
r,k ΘG + hd,k )wk |
SINRk = PK , ∀k. (2) IRS-aided system may still hold since the combined AP-user
j6=k |(hH H 2 2
r,k ΘG + hd,k )wj | + σk channels (including both the AP-user direct and AP-IRS-user
reflected links) of these two users are unlikely to be aligned
too, due to the additional IRS reflected paths.
B. Problem Formulation
Let W = [w1 , · · · , wK ] ∈ CM ×K , Hr = III. S INGLE -U SER S YSTEM
[hr,1 , · · · , hr,K ] ∈ CN ×K , and Hd = [hd,1 , · · · , hd,K ] ∈ In this section, we consider the single-user setup, i.e.,
CM ×K . In this paper, we aim to minimize the total transmit K = 1, to draw important insights into the optimal joint
power at the AP by jointly optimizing the transmit beamform- beamforming design. In this case, no inter-user interference
ing at the AP and reflect beamforming at the IRS, subject is present, and thus (P1) is simplified to (by dropping the user
index)
2 To characterize the fundamental performance limits of IRS, we assume
that the phase shifts can be continuously varied in [0, 2π), while in practice (P2) : min kwk2 (7)
they are usually selected from a finite number of discrete values from 0 to w,θ
2π for the ease of circuit implementation. The design of IRS with discrete s.t. |(hH H 2
r ΘG + hd )w| ≥ γσ ,
2
(8)
phase shifts is addressed in our follow-up work [17].
3 In practice, each element of the IRS is usually designed to maximize the 0 ≤ θn ≤ 2π, n = 1, · · · , N. (9)
signal reflection. Thus, we set βn = 1, ∀n, in the sequel of this paper for
simplicity. Note that this scenario is different from the traditional backscatter Although much simplified, problem (P2) is still a non-convex
communication where the RFID tags usually need to harvest a certain amount optimization problem since the left-hand-side (LHS) of (8)
of energy from the incident signals for powering their circuit operation and
thus a much smaller amplitude reflection coefficient than unity is resulted in is not jointly concave with respect to w and θ. In the next
general. two subsections, we solve (P2) by applying the SDR and
1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2019.2936025, IEEE
Transactions on Wireless Communications
alternating optimization techniques, respectively, which will be solvers such as CVX [20]. Generally, the relaxed problem (19)
extended to the general multiuser system in the next section. may not lead to a rank-one solution, i.e., rank(V ) 6= 1, which
implies that the optimal objective value of problem (19) only
A. SDR serves an upper bound of problem (17). Thus, additional steps
are needed to construct a rank-one solution from the obtained
We first apply SDR to solve problem (P2), which also helps
higher-rank solution to problem (19), while the details can be
obtain a lower bound of the optimal value of (P2) for evalu-
found in [1] and thus are omitted here. It has been shown that
ating the performance gaps from other suboptimal solutions.
such an SDR approach followed by a sufficiently large number
For any given phase shift θ, it is known that the maximum-
of randomizations guarantees at least a π4 -approximation of the
ratio transmission (MRT) is the optimal transmit beamforming
√ (hH ΘG+hH )H optimal objective value of problem (19) [19].
solution to problem (P2) [18], i.e., w∗ = P khrH ΘG+hdH k ,
r d
where P denotes the transmit power of the AP. Substituting
w∗ to problem (P2) yields the following problem B. Alternating Optimization
To achieve lower complexity than the SDR-based solution p-
min P (10) resented in the preceding subsection, we propose an alternative
P,θ
1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2019.2936025, IEEE
Transactions on Wireless Communications
as the effective channel perceived by the nth reflecting element The received power is then given by
at the IRS. Therefore, (28) suggests that the nth phase shift N
should be tuned such that the phase of the signal that passes X 2
Pu = P |hr,n ||gn | . (30)
through the AP-IRS and IRS-user links is aligned with that of n=1
the signal over the AP-user direct link to achieve coherent
signal combining at the user. Furthermore, it is interesting Since |hr,n | and |gn | are statistically independent
√ and fol-
to note that the obtained phase θn∗ is independent of the low
√ Rayleigh distribution with mean values π%h /2 and
amplitude of hn,r . As a result, the optimal transmit power π%g /2, respectively, we have E(|hr,n ||gn |) = π%h %g /4.
PN
γσ 2 By using the fact that n=1 |hr,n ||gn |/N → π%h %g /4 as
is given by P ∗ = k(hH ΘG+h H )w̄k2 from (P2). Next, we
r d N → ∞, it follows that
optimize the transmit beamforming direction for given θ in
(28). As in Section III-A, the combined AP-user channel is P π 2 %2h %2g
∗ Pu → N 2 . (31)
given by hH H
r ΘG + hd and hence MRT is optimal, i.e., w̄ = 16
H H H
(hr ΘG+hd )
khH H . The above alternating optimization approach This thus completes the proof.
r ΘG+hd k
is practically appealing since both the transmit beamforming The power scaling law with the optimal IRS phase design
and phase shifts are obtained in closed-form expressions, in Proposition 2 is highly promising since it implies that
without invoking the SDP solver. Its convergence is guaranteed by using a large number of reflecting units at the IRS, we
by the following two facts. First, for each subproblem, the can scale down the transmit power of the AP by a factor
optimal solution is obtained which ensures that the objective of 1/N 2 without compromising the user received SNR. The
value of (P2) is non-increasing over iterations. Second, the fundamental reason behind such a “squared gain” is that the
optimal value of (P2) is bounded from below due to the IRS not only achieves the transmit beamforming gain of order
SNR constraint. Thus, the proposed algorithm is guaranteed N in the IRS-user link as in the conventional massive MIMO
to converge. [10], but also captures an inherent aperture gain of order N
by collecting more signal power in the AP-IRS link, which,
C. Power Scaling Law with Infinitely Large Surface however, cannot be achieved by scaling up the number of
transmit antennas in massive MIMO due to the fixed total
Next, we characterize the scaling law of the average re- transmit power. Moreover, for the two benchmark cases with
ceived power at the user with respect to the number of reflect- unit and random phase shifts at the IRS, a received power
ing elements, N , in an IRS-aided system with N → ∞. For gain of order N is also achieved. This shows the practical
simplicity, we assume M = 1 with G ≡ g to obtain essential usefulness of the IRS, even without requiring any channel
insight. By ignoring the AP-user direct channel, the user’s knowledge for optimally setting the phase shifts. Note that the
received power is given by Pu = P |hH |2 = P |hH 2
r Θg| . received noise power in the IRS-aided system remains constant
We consider three different phase shift solutions, i.e., 1) unit as N increases and thus the corresponding user receive SNR
phase shift where Θ = I; 2) random phase shift where θn ’s also has the same squared gain as the received signal power
in Θ are uniformly and randomly distributed in [0, 2π); and with increasing N .
3) optimal phase shift which is obtained by the above two Next, we show the performance scaling law of an FD AF
proposed algorithms (both optimal for M = 1). relay aided system under the same setup as the above IRS-
Proposition 2: Assume hH r ∼ CN (0, %2h I) and g ∼ aided system. The relay is equipped with N transmit and N
2
CN (0, %g I). As N → ∞, it holds that receive antennas and the direct channel from the AP to the
user can be similarly ignored when N is asymptotically large.
2 2
N P %h %g , for Θ = I or random Θ, We assume that the relay adopts linear receive and transmit
Pu → 2 2 2 (29) beamforming vectors, denoted by xH r and xt , respectively.
N 2 P π %h %g , for optimal Θ. In addition, perfect self-interference cancellation (SIC) is
16
assumed at the relay so that the obtained performance serves
Proof: The three cases are discussed as follows: as an upper bound for the practical case with imperfect SIC.
In this case, the user receive SNR can be expressed as
• When Θ = I, we have hH = hH r g. By invoking
the Lindeberg-Lévy central limit theorem [21], we have P Pr kxH 2 H
r gk khr xt k
2
kxt k2 kxr k2
hH 2 2
r g ∼ CN (0, N %h %g ) as N → ∞. As the equivalent γF D = Pr σr2 kxr k2 khH 2 2 2 H 2 , (32)
r xt k +P σ kxt k kxr gk
H
channel h is a random variable, the average user re- kxt k2 kxr k2 + σr2 σ 2
ceived power is given by Pu = P E(|hH |2 ) → N P %2h %2g .
where Pr and σr2 denote the transmit power and the noise
For the case of random phase shifts with θn ∈ [0, 2π),
power at the relay, respectively. It is not difficult to show that
we have hH = hH r ḡ where ḡ = Θg. As Θ is a unitary hr
the optimal solution maximizing γF D satisfies x∗t = kh rk
and
matrix, it follows that ḡ has the same distribution as g, ∗ g ∗ ∗
xr = kgk . Substituting xt and xr into (32), we have
i.e., ḡ ∼ CN (0, %2g I). Thus we attain the same result as
Θ = I. P Pr kgk2 khr k2
• For optimal Θ where the solution is given by (28), we γF D = . (33)
PN Pr σr2 khr k2 + P σ 2 kgk2 + σr2 σ 2
have |hH | = |hH r Θg| = n=1 |h r,n ||gn |, where hr,n
and gn are the n-th elements in hH r and g, respectively.
Then, we have the following proposition.
1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2019.2936025, IEEE
Transactions on Wireless Communications
1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2019.2936025, IEEE
Transactions on Wireless Communications
solution of (P4), due to the common phase shifting matrix (V ) Algorithm 1 Alternating optimization algorithm.
for all users. 1: Initialize the phase shifts θ = θ 1 and set the iteration
In the proposed alternating optimization algorithm, we solve number r = 1.
problem (P1) by solving problems (P3) and (P4) alternately 2: repeat
in an iterative manner, where the solution obtained in each 3: Solve problem (P3) for given θ r , and denote the optimal
iteration is used as the initial point of the next iteration. The solution as W r .
details of the proposed algorithm are summarized in Algorithm 4: Solve problem (P4) or (P4’) for given W r , and denote
1. In particular, the algorithm starts with solving problem (P3) the solution after performing Gaussian randomization
for given θ instead of solving (P4) for given W . This is as θ r+1 .
deliberately designed since (P3) is always feasible for any 5: Update r = r + 1.
arbitrary θ, provided that rank(GH ΘHr + Hd ) = K, while 6: until The fractional decrease of the objective value is
this may not be true for (P4) with arbitrary W . On the below a threshold > 0 or problem (P4)/(P4’) becomes
other hand, as solving (P4) only attains a feasible solution, infeasible.
it remains unknown whether the objective value of (P3) will
monotonically decrease or not over iterations in Algorithm 1.
Intuitively, if the feasible solution obtained by solving (P4) where the slack variable αk can be interpreted as the “SINR
achieves a strictly larger user SINR than the corresponding residual” of user k in phase shift optimization. Note that (P4)
SINR target γk for user k, then the transmit power of user and (P4’) have the same set of feasible V , while (P4’) is more
k and hence the total transmit power in problem (P3) can be efficient than (P4) in terms of the converged solution, as will
properly reduced without violating all the SINR constraints. be verified in Section V-B by simulation.
More rigorously, the convergence of Algorithm 1 is ensured
by the following proposition. B. Two-Stage Algorithm
Proposition 4: The objective value of (P3) is non-
increasing over the iterations by applying Algorithm 1. Inspired by the combined channel gain maximization prob-
Proof: Denote the objective value of (P3) based on a lem (13) in the single-user case, we next propose a two-stage
feasible solution (θ, W ) as f (θ, W ). As shown in step 4 of algorithm with lower complexity compared to the alternating
Algorithm 1, if there exists a feasible solution to problem (P4), optimization algorithm by decoupling the joint beamforming
i.e., (θ r+1 , W r ) exists, it is also feasible to problem (P3). As design problem (P1) into two beamforming subproblems, for
such, (θ r , W r ) and (θ r+1 , W r+1 ) in step 3 are the feasible optimizing the phase shifts and transmit beamforming, respec-
solutions to (P3) in the rth and (r+1)th iterations, respectively. tively. Specifically, the phase shifts at the IRS are optimized
(a) (b) in the first stage by solving a weighted effective channel gain
It then follows that f (θ r+1 , W r+1 ) ≥ f (θ r+1 , W r ) = maximization problem. This aims to align with the phases of
f (θ r , W r ), where (a) holds since for given θ r+1 in step 3 different user channels so as to maximize the beamforming
of Algorithm 1, W r+1 is the optimal solution to problem gain of the IRS, especially for the users near to the IRS. In
(P3); and (b) holds because the objective function of (P3) is the second stage, we solve problem (P3) to obtain the optimal
regardless of θ and only depends on W . MMSE-based transmit beamforming with given phase shifts
To achieve better converged solution, we further transform θ.
problem (P4) into an optimization problem with an explicit Let v = [ejθ1 , · · · , ejθN ]H ∈ CN ×1 and Φk =
objective to obtain a generally more efficient phase shift diag(hH N ×M
r,k )G ∈ C , ∀k. The weighted sum of the com-
solution to reduce the transmit power. The rationale is that bined channel gain of all users is expressed as
for the transmit beamforming optimization problem, i.e., (P3),
K K
all the SINR constraints are active at the optimal solution. As X X
tk khH H 2
r,k ΘG + hd,k k = tk kv H diag(hH H 2
r,k )G + hd,k k
such, optimizing the phase shift to enforce the user achievable
k=1 k=1
SINR to be larger than the SINR target in (P4) directly K
leads to the transmit power reduction in (P3) (e.g., by simply X
tk kv H Φk + hH 2
= d,k k , (52)
scaling down the power of transmit beamforming). To this end, k=1
problem (P4) is transformed into the following problem
K
where we set the weights to be tk = γk1σ2 , k = 1, ..., K,
k
motivated by constraint (11). Based on (52), the phase shifts
X
(P4’) : max αk (48)
V ,{αk }
k=1
can be obtained by solving the following problem
K
X K
s.t. tr(Rk,k V ) + |bk,k |2 ≥ γk
X
tr(Rk,j V ) (P5) : max tk kv H Φk + hH
d,k k
2
(53)
v
j6=k k=1
XK s.t. |vn | = 1, n = 1, · · · , N. (54)
+ γk ( |bk,j |2 + σk2 ) + αk , ∀k, (49)
j6=k
Note that for K = 1, (P5) is equivalent to problem (13) for
the single-user case in Section III-A. However, in the multiuser
Vn,n = 1, n = 1, · · · , N + 1, (50)
case, due to the same set of phase shifts applied for all users
V 0, αk ≥ 0, ∀k, (51) with different channels, the combined channel power gains
1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2019.2936025, IEEE
Transactions on Wireless Communications
here for brevity. With the phase shifts obtained from (P5),
Fig. 2. Simulation setup of the single-user case (top view).
the MMSE-based transmit bemaforming is then obtained by
solving (P3). Compared to the alternating optimization based
algorithm proposed in Section IV-A, the two-stage algorithm 20
V. S IMULATION R ESULTS 12
1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2019.2936025, IEEE
Transactions on Wireless Communications
10
user link is dominant in the latter case. Moreover, it can be Lower bound
5
observed that the AP-IRS MRT scheme behaves oppositely SDR
4.5 Alternating optimization
as the user moves away from the AP toward IRS. Finally,
AP−user MRT
Fig. 3 also shows that if the transmit beamforming is not 4
AP−IRS MRT
designed properly, the performance achieved by using the IRS 3.5
Random phase shift
Without IRS
may be even worse than that of the case without the IRS, e.g., 3
with the AP-IRS MRT scheme for d ≤ 35 m. This further
2.5
demonstrates that the proposed joint beamforming designs can
dynamically adjust the AP’s beamforming to strike an optimal 2
1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2019.2936025, IEEE
Transactions on Wireless Communications
11
9
U8
1 bps/Hz increase by doubling N U5 U3
8
AP
U6
7
IRS
Achievable rate (bps/Hz)
d0
6 U4 d2
d1
5 2 bps/Hz increase by doubling N U7 U1 U2
3 IRS
FD AF relay Fig. 6. Simulation setup of the multiuser case (top view).
2 HD AF relay
1
Under the above setup, we plot the achievable rate in
0
0 200 400 600 800 1000 1200 1400 1600
bits/second/Hertz (bps/Hz) versus N in Fig. 5, where the HD
Number of elements/antennas at the IRS/AF relay, N AF relay is also considered as a benchmark. It is observed
that when N is small, the IRS-aided system is able to achieve
Fig. 5. Comparison between the IRS and the FD/HD AF relay. the same rate as the FD/HD AF relay-aided system by using
more reflecting elements. However, since the IRS’s elements
are passive, no transmit RF chains are needed for them
fading). For example, for the same user SNR, a transmit power and thus the cost is much lower as compared with that of
of 2 dBm is required at the AP when N = 30 while this active antennas for the AF relay requiring transmit RF chains.
value is reduced to −4 dBm when N = 60, which suggests Furthermore, one can observe that by doubling N from 400
an around 6 dB gain by doubling the number of reflecting to 800, the achievable rate of using IRS increases about 2
elements. In contrast, the transmit power required by using bps/Hz whereas that of using the FD AF relay only increases
the random phase shift decreases with increasing N in a much about 1 bps/Hz. This is due to their different SNR gains (N 2
slower rate, because without reflect beamforming the average versus N ) with increasing N as revealed in Propositions 2
signal power of the reflected signal is comparable to that of the and 3. As a result, it is expected that the IRS-aided system
signal from the AP-user direct link in this case. Finally, it is will eventually outperform the FD/HD AF relay-aided system
observed that the above gains diminish as the user moves away when N is sufficiently large, as shown in Fig. 5.
from the IRS. For example, for the case of d = 15 m shown in
Fig. 4 (c) where the AP-user direct link signal is much stronger
than that of the IRS-user link, the required transmit power is B. Multiuser System
insensitive to the number of reflecting elements. For the case Next, we consider a multiuser system with eight users,
of d = 41 m shown in Fig. 4 (b) when the user is neither close denoted by Uk , k = 1, · · · , 8, and their locations are shown
to the AP nor close to the IRS, it is observed that the transmit in Fig. 6. Specifically, Uk ’s, k = 2, 4, 6, 8, lie evenly on a half
power gain of the proposed schemes is generally lower than circle centered at the reference antenna of the IRS with radius
N 2 . This is because in this case the signal power received d2 = 3 m, which are usually considered as “cell-edge” users,
at the IRS is compromised as the AP transmit beamforming as compared to Uk ’s, k = 1, 3, 5, 7, which lie evenly on a
is steered to strike a balance between the AP-IRS link and circle centered at the reference antenna of the AP with radius
the AP-user direct link. In practice, the number of reflecting d1 = 20 m. Since the IRS can be practically deployed in LoS
elements can be properly selected depending on the IRS’s with the AP and “cell-edge” users, we set αAI = αIu = 2.8,
location as well as the target user SNR/AP coverage range. βAI = βIu = 3 dB, respectively. We compare our proposed
3) Comparison with AF Relay: Next, we compare the two algorithms (named as Alternating optimization w/ IRS and
achievable rates of the IRS versus the FD AF relay based Two-stage algorithm w/ IRS, respectively) in Section IV with
on the results derived in Section III. We consider the setup in the two conventional designs in the case without the IRS, i.e.,
Fig. 2 with d0 = d = 100 m, dv = 1 m, σr2 = −80 dBm, MMSE and zero-forcing (ZF) based beamforming [23], [27].
αAI = 3.2, αIu = 2, βAI = 0, βIu = ∞, and M = 1. To Specifically, the transmit power of the MMSE-based scheme
focus on the comparison with large N , the direct link from without the IRS is obtained by solving (P3) with Θ = 0,
the AP to the user is ignored, and perfect SIC is assumed for while that of the ZF-based scheme without the IRS is given by
the FD AF relay. As such, the SNRs and the corresponding tr(P (HdH Hd )−1 ) where P = diag(σ12 γ1 , · · · , σK 2
γK ). The
achievable rates for the IRS-aided and the FD AF relay-aided transmit power required by using the random phase shift at
systems can be obtained based on (30) and (32), respectively. the IRS and MMSE beamforming at the AP is also plotted as
For a fair comparison, we assume that both systems have the a benchmark. Before comparing their performances, we first
same total transmit power budget P = 5 mW (for single link show the convergence behaviour of the proposed Algorithm
only). Since the IRS is passive, all the transmit power is used 1 in Fig. 7 by setting M = 4 and considering that only Uk ,
at the AP, whereas since the AF relay is active like the AP, k = 1, 2, 3, 4, are active (need to be served) with γ = 20 dB.
an optimal power allocation between them is required which The phase shifts are initialized using the two-stage algorithm.
can be obtained by exhaustive search. It is observed that the transmit power required by the proposed
1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2019.2936025, IEEE
Transactions on Wireless Communications
12
35
user direct channel (i.e., hH
d,k ) for user k, which is defined as
Solving P4’ [18], [27]
Solving P4
34.5
!
|hH d,k wk |
ρk , E , ∀k. (57)
Transmit power at the AP (dBm)
34 khHd,k kkwk k
1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2019.2936025, IEEE
Transactions on Wireless Communications
13
10
0
5
0 −10
−5
−20
−10
−15 −30
−4 0 4 8 12 16 20 24 −4 0 4 8 12 16 20 24
User SINR target, γ (dB) User SINR target, γ (dB)
(a) Signal and interference powers at user 1 (far from the IRS). (b) Signal and interference powers at user 2 (near the IRS).
1 1
Alternating optimization w/ IRS Alternating optimization w/ IRS
MMSE beamforming w/o IRS MMSE beamforming w/o IRS
0.9 0.9
ρ2
ρ1
0.85 0.85
ZF beamforming ZF beamforming
0.8 0.8
−4 0 4 8 12 16 20 24 −4 0 4 8 12 16 20 24
User SINR target, γ (dB) User SINR target, γ (dB)
(c) ρ1 (d) ρ2
Fig. 9. Illustration for the collaborative working mechanism of the AP and IRS.
is concluded that the transmit beamforming directions for 32 Two−stage algorithm w/ IRS
MMSE beamforming w/o IRS
the users with the IRS are drastically different from those 30 ZF beamforming w/o IRS
in the case without the IRS, depending on their different
28
distances with the IRS. The above results further demonstrate
the necessity of jointly optimizing the transmit beamforming 26
22
2) AP Transmit Power versus Rician Factor of G: In
Fig. 10, we plot the AP’s transmit power required by the 20
1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2019.2936025, IEEE
Transactions on Wireless Communications
14
1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2019.2936025, IEEE
Transactions on Wireless Communications
15
system with M = 40. This validates the practical throughput [3] S. Zhang, Q. Wu, S. Xu, and G. Y. Li, “Fundamental green tradeoffs:
gain of IRS even by taking into account a moderate delay for Progresses, challenges, and impacts on 5G networks,” IEEE Commun.
Surveys Tuts., vol. 19, no. 1, pp. 33–56, First Quarter 2017.
its coordination with the AP. [4] Q. Wu, G. Y. Li, W. Chen, D. W. K. Ng, and R. Schober, “An overview
of sustainable green 5G networks,” IEEE Wireless Commun., vol. 24,
VI. C ONCLUSIONS no. 4, pp. 72–80, Aug. 2017.
[5] Q. Wu and R. Zhang, “Towards smart and reconfigurable environment:
In this paper, we proposed a novel approach to enhance Intelligent reflecting surface aided wireless networks,” IEEE Commun.
the spectrum and energy efficiency as well as reducing the Mag., 2019. [Online] Available: https://arxiv.org/abs/1905.00152.
[6] T. J. Cui, M. Q. Qi, X. Wan, J. Zhao, and Q. Cheng, “Coding
implementation cost of future wireless communication systems metamaterials, digital metamaterials and programmable metamaterials,”
by leveraging the passive IRS via smartly adjusting its signal Light: Science & Applications, vol. 3, no. 10, e218, Oct. 2014.
reflection. Specifically, given the user SINR constraints, the [7] S. Hu, F. Rusek, and O. Edfors, “Beyond massive MIMO: The potential
of data transmission with large intelligent surfaces,” IEEE Trans. Signal
active transmit beamforming at the AP and passive reflect Process., vol. 66, no. 10, pp. 2746–2758, May 2017.
beamforming at the IRS were jointly optimized to minimize [8] L. Subrt and P. Pechac, “Intelligent walls as autonomous parts of smart
the transmit power in an IRS-aided multiuser system. By indoor environments,” IET Communications, vol. 6, no. 8, pp. 1004–
1010, May 2012.
applying the SDR and alternating optimization techniques, [9] X. Tan, Z. Sun, J. M. Jornet, and D. Pados, “Increasing indoor spectrum
efficient algorithms were proposed to trade off between the sharing capacity using smart reflect-array,” in Proc. IEEE ICC, May
system performance and computational complexity. It was 2016.
[10] H. Q. Ngo, E. Larsson, and T. Marzetta, “Energy and spectral efficiency
shown for the single-user system that the receive SNR increas- of very large multiuser MIMO systems,” IEEE Trans. Commun., vol. 61,
es quadratically with the number of reflecting elements of the no. 4, pp. 1436–1449, Apr. 2013.
IRS, which is more cost efficient than the conventional massive [11] S. K. Mohammed and E. G. Larsson, “Single-user beamforming in large-
scale MISO systems with per-antenna constant-envelope constraints: The
MIMO or multi-antenna AF relay. While for the multiuser sys- doughnut channel,” IEEE Trans. Wireless Commun., vol. 11, no. 11, pp.
tem, it was shown that IRS-enabled interference suppression 3992–4005, Nov. 2012.
can be jointly designed with the AP transmit beamforming to [12] S. Zhang, R. Zhang, and T. J. Lim, “Constant envelope precoding for
MIMO systems,” IEEE Trans. Commun., vol. 66, no. 1, pp. 149–162,
improve the performance of all users in the system, even for Jan. 2018.
those that are far away from the IRS. Extensive simulation [13] O. El Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath,
results under various practical setups demonstrated that by “Spatially sparse precoding in millimeter wave MIMO systems,” IEEE
Trans. Wireless Commun., vol. 13, no. 3, pp. 1499–1513, Mar. 2014.
deploying the IRS and jointly optimizing its reflection with [14] F. Sohrabi and W. Yu, “Hybrid digital and analog beamforming design
the AP transmission, the wireless network performance can for large-scale antenna arrays,” IEEE J. Sel. Topics Signal Process.,
be significantly improved in terms of energy consumption, vol. 10, no. 3, pp. 501–513, Apr. 2016.
[15] C. Huang, A. Zappone, M. Debbah, and C. Yuen, “Achievable rate
coverage as well as achievable rate, as compared to the maximization by passive intelligent mirrors,” in Proc. IEEE ICASSP,
conventional systems without using the IRS. Useful insights on Apr. 2018.
optimally deploying the IRS and its delay-performance trade- [16] C. Boyer and S. Roy, “Backscatter communication and RFID: Coding,
energy, and MIMO analysis,” IEEE Trans. Commun., vol. 62, no. 3, pp.
off were also drawn to provide useful guidance for practical 770–785, Mar. 2014.
design and implementation. [17] Q. Wu and R. Zhang, “Beamforming optimization for wireless network
In practice, if IRS is equipped with receive RF chains, the aided by intelligent reflecting surface with discrete phase shifts,” [On-
line] Available: https://arxiv.org/abs/1906.03165.
commonly used pilot-assisted channel estimation methods can [18] D. Tse and P. Viswanath, Fundamentals of wireless communication.
be similarly applied to the IRS as shown in Section V-B; Cambridge university press, 2005.
otherwise, it is infeasible for the IRS to directly estimate the [19] A. M.-C. So, J. Zhang, and Y. Ye, “On approximating complex quadratic
optimization problems via semidefinite programming relaxations,” Math-
channels with its associated AP/users. For the latter (more ematical Programming, vol. 110, no. 1, pp. 93–110, Jun. 2007.
challenging) case, a viable approach may be to design the [20] M. Grant and S. Boyd, “CVX: MATLAB software for disciplined convex
IRS’s passive beamforming based on the feedback from the programming,” 2016. [Online] Available: http://cvxr.com/cvx.
[21] H. Cramér, Random variables and probability distributions. Cambridge
AP/users that receive the signals reflected by the IRS, which University Press, 2004, vol. 36.
is worth investigating in the future work. In addition, after this [22] A. Wiesel, Y. C. Eldar, and S. Shamai, “Linear precoding via conic
paper was submitted, we became aware of another parallel optimization for fixed MIMO receivers,” IEEE Trans. Signal Process.,
vol. 54, no. 1, pp. 161–176, Jan 2006.
work [28], which shows that the IRS-aided wireless system [23] M. Bengtsson and B. Ottersten, “Optimal and suboptimal transmit
is more energy-efficient than the conventional multi-antenna beamforming,” Handbook of antennas in wireless communications, Boco
AF relay system with HD operation. Although the spectrum Raton, FL: CRC, 2001.
[24] M. Schubert and H. Boche, “Solution of the multiuser downlink beam-
efficiency can be further improved by using the FD AF forming problem with individual SINR constraints,” IEEE Trans. Veh.
relay, effective SIC is required, which incurs additional energy Technol., vol. 53, no. 1, pp. 18–28, Jan. 2004.
consumption. As such, it is worthy of further comparing the [25] Z.-Q. Luo and W. Yu, “An introduction to convex optimization for
communications and signal processing,” IEEE J. Sel. Areas Commun.,
energy efficiency of IRS with the FD AF relaying in future vol. 24, no. 8, pp. 1426–1438, Aug. 2006.
work. [26] R. Zhang, C. C. Chai, and Y.-C. Liang, “Joint beamforming and power
control for multiantenna relay broadcast channel with QoS constraints,”
R EFERENCES IEEE Trans. Signal Process., vol. 57, no. 2, pp. 726–737, Feb. 2009.
[27] T. Yoo and A. Goldsmith, “On the optimality of multiantenna broad-
[1] Q. Wu and R. Zhang, “Intelligent reflecting surface enhanced wireless cast scheduling using zero-forcing beamforming,” IEEE J. Sel. Areas
network: Joint active and passive beamforming design,” in Proc. IEEE Commun., vol. 24, no. 3, pp. 528–541, Mar. 2006.
GLOBECOM, Dec. 2018. [28] C. Huang, A. Zappone, G. C. Alexandropoulos, M. Debbah, and
[2] F. Boccardi, R. W. Heath, A. Lozano, T. L. Marzetta, and P. Popovski, C. Yuen, “Large intelligent surfaces for energy efficiency in wireless
“Five disruptive technology directions for 5G,” IEEE Commun. Mag., communication,” [Online] Available: https://arxiv.org/abs/1810.06934.
vol. 52, no. 2, pp. 74–80, Feb. 2014.
1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2019.2936025, IEEE
Transactions on Wireless Communications
16
1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.