IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO.

6, JUNE 2023 7627

Joint Beamforming and Channel Reconfiguration for

RIS-Assisted Millimeter Wave Massive
MIMO-OFDM Systems
Hanyu Wang , Student Member, IEEE, Jun Fang , Senior Member, IEEE, and Hongbin Li , Fellow, IEEE

Abstract—In this paper, we consider the problem of joint ac- transmission of multiple data streams. To overcome the above
tive and passive beamforming for reconfigurable intelligent sur- challenges, RIS has been recently introduced to improve the
face (RIS)-assisted millimeter-wave (mmWave) multiple-input- spectral efficiency and coverage of mmWave systems [7], [8],
multiple-output (MIMO) orthogonal frequency-division multiplex-
ing (OFDM) systems. The objective is to maximize the spectral [9], [10]. RIS is a planar meta-surface equipped with a large num-
efficiency by jointly optimizing the RIS’s reflection coefficients ber of programmable passive reflecting elements. With the aid
and the hybrid precoder/combiner at the base station (BS)/mobile of a smart controller, each reflecting element can independently
station (MS). By resorting to classical hybrid precoder/combiner reflect the incident signal with a reconfigurable amplitude and
optimization techniques, this joint beamforming problem can be
phase shift, thus helping realize a favorable propagation channel
converted into an optimization concerning only the design of the
reflection coefficients. Nevertheless, the objective function of the for high-capacity and robust mmWave links [11], [12], [13], [14],
resulting passive beamforming problem does not have an explicit [15], [16].
expression of the reflection coefficients. To address this difficulty, A key problem of interest in RIS-assisted mmWave commu-
the inherent structure of the effective channel is exploited, based nications is to jointly devise the reflection coefficients and the
on which an approximation of the truncated singular value decom-
transmit precoder to optimize the system performance. Such a
position (SVD) of the effective channel is obtained, and a favorable
propagation environment can be realized by directly manipulating problem is also referred to as joint active and passive beamform-
the singular values of the effective channel. Simulation results show ing, and has been extensively studied over the past few years [8],
that the proposed method can configure a favorable channel with a [9], [17], [18], [19], [20], [21], [22], [23], [24], [25], [26], [27],
small condition number. Moreover, the proposed method presents [28], [29], [30], [31]. Most studies on joint active and passive
a clear performance advantage over state-of-the-art algorithms,
beamforming, however, focused on single-input-single-output
while with a low computational complexity.
(SISO) or multiple-input-single-output (MISO) systems, e.g.
Index Terms—Joint active and passive beamforming, mmWave [8], [17], [18], [19], [20], [21], [22], [23], which cannot be
communications, reconfigurable intelligent surface (RIS). straightforwardly extended to MIMO systems. In addition, con-
sidering the rich spectrum resources available at the mmWave
I. INTRODUCTION band, mmWave systems are very likely to operate on wideband
channels with frequency selectivity [32]. Therefore it is of
ILLIMETER wave (mmWave) communication is re-
M garded as a promising technology for future cellular
networks [1], [2], [3]. Nevertheless, due to reduced diffraction
practical significance to study joint beamforming for massive
MIMO OFDM systems. Specifically, in [29], [30], an alternating
optimization-based method was proposed to address the joint
and high penetration loss, mmWave communication has limited beamforming problem for massive MIMO-OFDM systems, in
coverage and is vulnerable to blockage by obstacles [4], [5], [6]. which the transmit precoder and the RIS’s reflection coefficients
Also, the sparse scattering characteristics of mmWave channels are optimized in an alternating manner. Nevertheless, the opti-
make it difficult to exploit the spatial diversity for simultaneous mization of the reflection coefficients is still very involved. To
address this difficulty, [29] proposed to alternatively optimize
Manuscript received 6 August 2022; revised 29 December 2022 and 4 Febru- each reflection coefficient, with other reflection coefficients
ary 2023; accepted 6 February 2023. Date of publication 9 February 2023;
date of current version 20 June 2023. The work of Hongbin Li was supported fixed. Such a sequential optimization approach, however, incurs
in part by the National Science Foundation under Grants ECCS-1923739 and a prohibitively high computational complexity. Another ap-
ECCS-2212940. This work was supported in part by the National Key R\&D proach [30] addresses this difficulty by finding a lower bound of
Program of China under Grant 2018YFB1801500, and in part by the National
Science Foundation of China under Grant 61829103. The review of this article the objective function and employs a majorization-minimization
was coordinated by Prof. Jia-Chin Lin. (Corresponding author: Jun Fang.) technique to maximize the lower bound. Optimizing the lower
Hanyu Wang and Jun Fang are with the National Key Laboratory bound, however, may lead to a solution that is far from the
of Science and Technology on Communications, University of Elec-
tronic Science and Technology of China, Chengdu 611731, China (e-mail: optimum. In addition to the AO-based method, a low-complexity
[email protected]; [email protected]). joint design method was proposed in [31] for joint beamforming
Hongbin Li is with the Department of Electrical and Computer Engineering, for RIS-assisted massive MIMO-OFDM systems. The basic idea
Stevens Institute of Technology, Hoboken, NJ 07030 USA (e-mail: hong-
[email protected]). is to align the passive beamforming vector with the dominant
Digital Object Identifier 10.1109/TVT.2023.3243389 path of the reflected channel. Nevertheless, as only the dominant
0018-9545 © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: University of Leeds. Downloaded on December 04,2024 at 21:29:32 UTC from IEEE Xplore. Restrictions apply.
7628 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 6, JUNE 2023

path is utilized, this method does not fully exploit the spatial di-
versity of the reflected channel, which prevents it from achieving
a higher spectral efficiency.
In this paper, we study the problem of joint active and pas-
sive beamforming for RIS-assisted mmWave MIMO-OFDM
systems, where an RIS is deployed to assist the downlink
transmission from the BS to the MS. Specifically, the objec-
tive is to maximize the spectral efficiency by jointly optimiz-
ing the RIS’s reflection coefficients and the hybrid precoder
(combiner) at the BS (MS). By resorting to classical hybrid
precoder/combiner optimization techniques, we show that this
complicated problem can be converted into an optimization
concerning only the design of the reflection coefficients. Nev-
ertheless, the objective function of this passive beamforming Fig. 1. RIS-assisted MIMO-OFDM downlink communication system..
problem is a function of the singular values of the effective
(or equivalent) channel, which does not have an explicit ex-
pression of the reflection coefficients. To address this difficulty, The rest of the paper is organized as follows. In Section II,
we start by analyzing the geometric model of the effective the system model and the formulation of the active/passive
channel and show that it can serve as a good approximation beamforming problem are discussed. The passive beamforming
of the truncated SVD of the effective channel. Thus the singular problem is studied in Section III, where a SVD-approximation-
values of the effective channel can be approximated by the path based method is developed. In Section IV, based on the op-
gains associated with the reflected paths which have an explicit timized reflection coefficients, the design of active hybrid pre-
expression of the reflection coefficients. A manifold-based opti- coder/combiner is studied. The computational complexity of the
mization algorithm is then developed to optimize the reflection proposed method is analyzed in Section V. Simulation results
coefficients. are presented in Section VI, followed by concluding remarks in
A major contribution of this work is that we established an Section VII.
explicit mathematical relationship between the singular values Notations: Vectors and matrices are denoted by lowercase
of the effective channel and the reflection coefficients. This and uppercase letters, respectively. R and C respectively denote
enables us to customize a favorable propagation channel with the real space and complex space. The transpose, conjugate,
a small condition number by directly manipulating its singular conjugate transpose, Moore-Penrose pseudo inverse, determi-
values. Note that the idea of the current work is partially inspired nant, rank and trace of a matrix X are represented by X T ,
by our previous work [25]. Nevertheless, there are several im- X ∗ , X H , X † , det(X), rank(X) and tr(X), respectively. The
portant distinctions between this work and [25]. Firstly, different element in the ith row and jth column of X is denoted by
from [25] which is concerned with single-carrier narrowband X[i, j], while X[{im }, {jn }] denotes the submatrix whose
channels, this paper considers the joint beamforming problem (m, n)th element is equal to (im , jn )th element of X. diag(X)
with wideband frequency-selective mmWave channels. Sec- denotes the vector that contains the diagonal elements of X,
ondly, in the current work, we consider a more general scenario while diag(x) represents the diagonal matrix that contains the
where a direct link between the BS and the MS is available. elements of vector x on its main diagonal. | · | and  · F indicate
In this case, the problem is more complicated as the effective the absolute value of a scalar and the Frobenius norm of a
channel is a sum of the direct channel and the reflected channel, matrix, respectively. The real part of a complex number are
and we will need to determine how many and which reflected denoted by R{·}. The identity matrix is denoted by I. The real
paths should be chosen for data transmission before optimizing and complex Gaussian distributions with mean μ and variance
the passive beamforming vector. σ 2 is denoted by N (μ, σ 2 ) and CN (μ, σ 2 ), respectively. The
The work [26] also customizes a favorable propagation chan- Kronecker product and Hadamard product are denoted by ⊗
nel via exploiting the inherent structure of the effective chan- and ◦, respectively.
nel. It considers a scenario consisting of multiple spatially
distributed RISs. For each RIS, only the dominant path of the II. SYSTEM MODEL AND PROBLEM FORMULATION
corresponding BS-RIS-MS channel is selected for data trans-
mission. Therefore the reflection coefficients of each RIS can be A. System Model
easily determined by simply aligning the passive beamforming Consider a mmWave MIMO-OFDM system, where an RIS
vector with the dominant path. Different from [26], our work is employed to assist data transmission from the BS to the MS
considers a single RIS scenario. The objective is to carefully as illustrated in Fig. 1. We assume a hybrid analog and digital
devise the reflection coefficients of the RIS such that it can fully beamforming structure at both the BS and the MS, in which the
exploit the spatial diversity of the cascade channel to improve BS is equipped with Nt antennas and Rt (Rt < Nt ) RF chains
the spectral efficiency of the system. This leads to a passive and the MS is equipped with Nr antennas and Rr (Rr < Nr )
beamforming design problem that is more complicated than that RF chains. The total number of OFDM tones (subcarriers) is
in [26]. P . For each subcarrier, the BS sends Ns data steams to the MS

Authorized licensed use limited to: University of Leeds. Downloaded on December 04,2024 at 21:29:32 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: JOINT BEAMFORMING AND CHANNEL RECONFIGURATION 7629

simultaneously. Specifically, the transmitted signal associated B. Channel Model

with the pth subcarrier can be expressed as
We adopt the following geometric wideband mmWave chan-
z p = F RF F BB,p sp , (1) nel model to characterize the direct-link channel and the re-
flection channel. Specifically, the BS-MS channel in the delay
where sp ∈ CNs denotes the pth subcarrier’s symbol vector domain can be expressed as [32], [34]
Rt ×Ns
that satisfies E{sp sHp } = I, F BB,p ∈ C is the baseband

precoding matrix associated with pth subcarrier, and F RF ∈ Nr Nt 
Ld

CNt ×Rt is the RF precoding matrix that is common to all G(τ ) = ˜l aMS (θld )aH
BS (γl )δ(τ − τl ),
d d
(6)
Ld
subcarriers. Also, we impose a transmit power constraint for l=1

each subcarrier, i.e. F RF F BB,p 2F ≤ ρ.

The procedure to generate the transmitted signal (1) is elabo- where ˜l is the complex path gain, γld , θld and τld represent the
rated as follows. Firstly, the symbol vector sp at each subcarrier angle of departure (AoD), the angle of arrival (AoA) and the
is precoded using a baseband precoding matrix F BB,p . The time delay associated with lth path respectively, aBS (γld ) and
symbol blocks are transformed to the time-domain using P -point aMS (θld ) respectively denote the BS’s and MS’s array response
inverse discrete Fourier transform (IDFT). A cycle prefix is then vectors, and δ(·) represents the Dirac-delta function.
added to the symbol block, and finally a common RF precoder Similarly, the BS-RIS channel T (τ ) and the RIS-MS channel
F RF is applied to all subcarriers. After the signal received at R(τ ) in the delay domain are modeled as
the MS, a common RF combiner W RF ∈ CNr ×Rr is applied to 
Nt M 
Lt
all subcarriers. Then the cyclic prefix is removed and symbols
T (τ ) = α̃l aRIS (ϑa,l , ϑe,l )aH
BS (γl )δ(τ − τl ),
t
are converted back to the frequency domain by performing a Lt
l=1
discrete Fourier transform (DFT). Finally, the received signal (7)
associated with pth subcarrier is combined using a baseband

combining matrix W BB,p ∈ CRr ×Ns . Therefore, the received Nr M 
Lr

signal associated with the pth subcarrier can be expressed as R(τ ) = β̃l aMS (θl )aH
RIS (χa,l , χe,l )δ(τ − τl ),
r
Lr
l=1
yp = W H
BB,p W RF H p F RF F BB,p sp + W BB,p W RF np ,
H H H
(2) (8)

where np ∈ CNr denotes the additive Gaussian noise with where α̃l (β̃l ) and τlt (τlr ) respectively represent the complex
zero mean and variance σ 2 , and H p ∈ CNr ×Nt denotes the path gain and the time delay, γl (θl ) denotes the AoD (AoA)
frequency-domain channel matrix associated with the pth sub- associated with the BS(MS), ϑa,l (χa,l ) and ϑe,l (χe,l ) denote
carrier. the azimuth and elevation AoA (AoD) associated with the RIS,
The transmitted signal arrives at the receiver via propagating aRIS (ϑa,l , ϑe,l ) and aRIS (χa,l , χe,l ) denote the RIS’s array re-
through the direct BS-MS channel and the BS-RIS-MS reflection sponse vectors.
channel. The RIS is a uniform planar array (UPA) consisting Accordingly, the frequency-domain channels associated with
of M reflecting elements, where each of them behaves like a the pth subcarrier can be respectively written as
single physical point which combines all the received signals

and then re-scatters the combined signal with an adjustable phase Nr Nt 
Ld
d p
shift [19]. Let Gp ∈ CNr ×Nt , T p ∈ CM ×Nt and Rp ∈ CNr ×M Gp = ˜l e−j2πfs τl P aMS (θld )aH d
BS (γl ),
Ld
respectively denote the BS-MS channel, the BS-RIS channel and l=1
the RIS-MS channel associated with the pth subcarrier. Also, 
Nt M 
Lt
t p
define Tp = α̃l e−j2πfs τl P aRIS (ϑa,l , ϑe,l )aH
BS (γl ),
  Lt
l=1
Φ  diag ejφ1 , . . . , ejφM , (3) 
Nr M 
Lr
r p
where φm ∈ [0, 2π] is the phase shift introduced by the mth Rp = β̃l e−j2πfs τl P aMS (θl )aH
RIS (χa,l , χe,l ),
Lr
passive element of the RIS. Hence, the channel matrix H p l=1
associated with the pth subcarrier can be expressed as (9)
where fs is the sampling frequency.
H p = Gp + Rp ΦT p . (4) In the above equations, the normalized array response vectors
are defined as follows. For ULA with N antennas, the normal-
Thus, the achievable spectral efficiency can be expressed as [33] ized array response vector is given by

1 
P
1 1
log det I + 2 (W BB,p W RF )† H p F RF F BB,p a(γ) = √ [1 · · · ej(n−1) λ sin(γ) · · · ej(N −1) λ sin(γ) ]T ,
2πd 2πd
R=
P p=1 2 σ N
 (10)
× F BB,p F RF H p (W RF W BB,p ) .
H H H
where d and λ are the antenna spacing and the signal wavelength.
(5) For UPA with M = My × Mz antennas, the normalized array

Authorized licensed use limited to: University of Leeds. Downloaded on December 04,2024 at 21:29:32 UTC from IEEE Xplore. Restrictions apply.
7630 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 6, JUNE 2023

response vector is defined as where V p and U p are obtained via the ordered SVD of the
pth subcarrier’s channel: H p = U p Σp V H p , in which Σp is
1
a(ϑa , ϑe ) = √ [1 · · · ej(My −1) λ sin(ϑa ) cos(ϑe ) ]T
2πd
a diagonal matrix of singular values, and Λ ∈ RNs ×Ns is a
M diagonal matrix whose diagonal elements are calculated via the
⊗[1 · · · ej(Mz −1) sin(ϑe ) T water-filling power allocation scheme and Λ satisfies tr(Λ) = ρ.
2πd
λ ] . (11)
Thanks to the massive array gain provided by the RIS and the
C. Problem Formulation large number of antennas at the BS, the effective signal-to-noise
ratio (SNR) is large, in which case an equal power allocation
Our objective is to jointly devise the active hybrid pre- scheme is near-optimal [40]. Therefore we can approximate F p
coder/combiner and the passive beamforming matrix Φ to max- as:
imize the achievable spectral efficiency. The problem can be 
ρ
formulated as F p ≈ V p [:, 1 : Ns ]. (15)
 Ns
P
1
 max  log2 det I + 2 (W BB,p W RF )† Substituting the optimal precoder/combiner back to (13), we
Φ,F RF ,{F BB,p } P
p=1
,
p=1
σ arrive at a problem that only concerns the optimization of the
W RF ,{W BB,p }P ,
p=1 passive reflection matrix Φ :

P  
× H p F RF F BB,p F H
BB,p F RF H p (W RF W BB,p )
H H ρ 2
max log2 det I + Σ̄
Φ
p=1
Ns σ 2 p
s.t. F RF F BB,p 2F ≤ ρ, ∀p = 1, . . . , P,
s.t. Φ = diag(ejφ1 , . . . , ejφM ). (16)
|F RF [i, j]| = |W RF [i, j]| = 1, ∀i, j,
where Σ̄p  Σp [1 : Ns , 1 : Ns ].
H p = Gp + Rp ΦT p , ∀p = 1, . . . , P,
In the following, we will first discuss how to optimize the
Φ = diag(ejφ1 , . . . , ejφM ). (12) passive reflection matrix Φ. After the passive reflection matrix
is determined, the effective channel matrices {H p }P p=1 can
Note that this work assumes the knowledge of the channel state be accordingly calculated, and then the optimal fully digital
information (CSI) for joint active and passive beamforming precoder/combiner can be determined via (14)–(15). Finally,
design. Channel estimation for RIS-assisted mmWave systems a manifold optimization-based technique can be employed to
is an important and challenging task that has been investigated search for the hybrid analog and digital precoder/combiner to
in some works, e.g. [35], [36], [37], [38], [39]. approximate the optimal fully digital precoder/combiner.
To simplify the above problem, we first ignore the constraint
imposed by the hybrid analog/digital structure and consider III. PROPOSED PASSIVE BEAMFORMING DESIGN METHOD
a fully digital precoder/combiner. Thus the problem can be
simplified as Note that the singular values in (16) cannot be expressed by Φ
in an explicit way. Therefore it is challenging to directly optimize

P  
1 (16). To address this difficulty, we exploit the inherent structure
max log2 det I + 2 W †p H p F p F H
p H H
p W p of the effective channels. First, define
{Φ,{F p },{W p }} σ
p=1 
Nr Nt
s.t. F p 2F ≤ ρ, ∀p = 1, . . . , P, l  ˜l (17)
Ld
H p = Gp + Rp ΦT p , ∀p = 1, . . . , P, 
Nt M
αl  α̃l (18)
Φ = diag(ejφ1 , . . . , ejφM ). (13) Lt

where F p ∈ CNt ×Ns and W p ∈ CNr ×Ns respectively denote Nr M
βl  β̃l (19)
the fully digital precoder and combiner associated with pth Lr
subcarrier. Once an optimal fully digital precoder/combiner is For simplicity, we assume that complex path gains are arranged
obtained, a manifold optimization-based method can be utilized in a decreasing order in terms of magnitude, i.e. |ρ1 | > · · · >
to devise the hybrid precoder (combiner) that approximates |ρLd | ({αl } and {βl } are arranged in a similar way). We write
the optimal fully digital precoder (combiner). Note that such the BS-RIS-MS channel as
a strategy has been numerically and theoretically justified by
previous studies [32], [33], [40], [41]. Rp ΦT p =
Given a fixed Φ, we now discuss how to obtain the optimal L
 r
r p
fully digital precoder and combiner. Suppose rank(H p ) ≥ Ns βm e−j2πfs τm P aMS (θm )aH
RIS (χa,m , χe,m ) Φ
holds valid for all subcarriers. According to [42], the optimal m=1
solution is given as L
 t
t p

F p = V p [:, 1 : Ns ]Λ1/2 , W p = U p [:, 1 : Ns ], (14) × αn e−j2πfs τn P aRIS (ϑa,n , ϑe,n )aH

BS (γn )
n=1

Authorized licensed use limited to: University of Leeds. Downloaded on December 04,2024 at 21:29:32 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: JOINT BEAMFORMING AND CHANNEL RECONFIGURATION 7631


Lr 
Lt
p
are the top K (in terms of magnitude) diagonal entries of D p and
βm αn e−j2πfs (τm +τn ) P aMS (θm )
r t
= (p) 

{dD,k }K k=1 that are the top K (in terms of magnitude) diagonal
m=1 n=1
entries of D dp , where we have K + K  = Ns and
× aH
RIS (χa,m , χe,m )ΦaRIS (ϑa,n , ϑe,n ) aBS (γn )
H
(p)
dmn
dR,k  D p [ik , ik ], ∀k = 1, . . . , K, (30)
(p)

Lr 
Lt
t p
dD,k  D dp [k, k], ∀k = 1, . . . , Ns − K. (31)
−j2πfs (τm
r
+τn )P
= βm αn e dmn aMS (θm )aH
BS (γn )
m=1 n=1 Note that the selection indices, {ik } and {k}, are the same for
different subcarriers as the magnitudes of diagonal entries of D p
= AMS D p AH
BS , (20)
and D dp ,
where
|D p [i, i]| = |v H pii ||βi αi |, ∀i = 1, . . . , min (Lt , Lr ), (32)
AMS  [aMS (θ1 ) · · · aMS (θLr )], (21)
|D dp [i, i]| = |i |, ∀i = 1, . . . , Ld , (33)
ABS  [aBS (γ1 ) · · · aBS (γLt )], (22)
are both independent of the subcarrier index p. Therefore, the
and D p is an Lr × Lt matrix with its (m, n)th entry given by
truncated singular value matrix of H p , Σ̄p , can be approximated
p
D p [m, n] = dmn βm αn e−j2πfs (τm +τn ) P
r t
(23) as

in which dmn is defined as ˆ = diag d(p) , d(p) , . . . , d(p) ,
Σ̄ p R,1 R,2 R,K
dmn  aH
RIS (χa,m , χe,m )ΦaRIS (ϑa,n , ϑe,n )

(p) (p) (p)
dD,1 , dD,2 , . . . , dD,Ns −K . (34)
= v H (a∗RIS (χa,m , χe,m ) ◦ aRIS (ϑa,n , ϑe,n ))
ˆ }P into the objective function of (16), we
Substituting {Σ̄
= v H pmn , (24) p p=1
have
Here v  diag(ΦH ) and pmn  aRIS (χa,m , χe,m ) ◦ aRIS  
P
ρ ˆ2
(ϑa,n , ϑe,n ). log2 det I + Σ̄
Thus, we can rewrite the overall effective channel H p as p=1
Ns σ 2 p
H p = Gp + Rp ΦT p K  
  ρ  (p) 2
P
  = log2 1+ d 
H Dp 0 H Ns σ 2 R,k
= ĀMS D̄ p ĀBS = ĀMS ĀBS , (25) p=1 k=1
0 D dp   
s −K
ρ  (p) 2
N
where + log2 1+ d 
Ns σ 2 D,k
k=1
ĀMS  [aMS (θ1 ) · · · aMS (θLr ) aMS (θ1d ) ··· aMS (θL
d
)], 
d
  
ρ|βik αik |2  H ik ik 2
K
(26)
=P log2 1 + v p
Ns σ 2
ĀBS  [aBS (γ1 ) · · · aBS (γLt ) aBS (γ1d ) · · · aBS (γLd d )], k=1
(27) s −K
N  
ρ|i |2
d p d p + log2 1 + . (35)
D dp  diag(1 e−j2πfs τl P , . . . , Ld e−j2πfs τl P ). (28) i=1
Ns σ 2
It can be easily verified that
Therefore the optimization (16) can be re-expressed as
|a(γi )H a(γj )| → 0, N → ∞, (29)

K
 
for any sin(γi ) = sin(γj ). Therefore for sufficiently large values max log2 1 + aik |v H pik ik |2 + RD (Ns − K)
v,{ik }K
of Nt and Nr , ĀMS and ĀBS can be treated as orthonormal k=1 k=1
matrices in which the column vectors form an orthonormal set. s.t. v = [ejφ1 · · · ejφM ]H , K ≤ Ns . (36)
Also, as shown in [25], if the passive beamforming vector v is
optimized to maximize {|dm,m |}, i.e. the magnitude of diagonal where ai  ρ|βi αi |2
and
Ns σ 2
entries of D p , the optimized v will be approximately orthogonal
to pmn , ∀m = n, thus leading to a quasi-diagonal matrix D p 
n
 
with negligible off-diagonal entries. Inspired by this fact, we can RD (n)  log2 1 + ρ|i |2 /(Ns σ 2 ) . (37)
H
treat ĀMS D̄ p ĀBS as an approximation of the truncated SVD of i=1

the effective channel H p . To simplify the problem, we first suppose K is pre-specified. In

Without loss of generality, we assume that the top Ns singular this case, the second term in the above optimization becomes
(p)
values of H p comprise of two subsets, namely, {dR,k }K k=1 that a constant. Hence we only need to jointly optimize v and the

Authorized licensed use limited to: University of Leeds. Downloaded on December 04,2024 at 21:29:32 UTC from IEEE Xplore. Restrictions apply.
7632 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 6, JUNE 2023

index set {ik }K

k=1 to maximize the spectral efficiency:
B. Manifold-Based Method for Passive Beamforming Design
Since the optimal index set {ik } = {1, . . . , K  } has been

K
 
max log2 1 + aik |v H pik ik | 2 determined, we see that RD (Ns − K  ) is a constant and the
v,{ik }K
k=1 optimization (36) can be simplified as
k=1

s.t. v = [ejφ1 · · · ejφM ]H (38) 

K 
 
max log2 1 + ai v H P i v
v
i=1
A. Determining the Index Set {ik }K
k=1 and the Value K

To gain insight into the above optimization problem, no- s.t. v = [ejφ1 · · · ejφM ]H . (43)
tice that the vector pij , which is defined as the Hadamard
where P i  pii (pii )H . To achieve a good balance between the
product of two steering vectors, i.e. pij  aRIS (χa,i , χe,i ) ◦
computational complexity and the performance, a manifold-
aRIS (ϑa,j , ϑe,j ), is asymptotically orthogonal to pmn for any
based method can be employed to solve the passive beamforming
i = m or j = n. This fact inspires us to construct an orthonormal
design problem.
matrix
√ P ∈ CM ×M , in which its first Lr Lt columns are {p̄ij 
Specifically, for the optimization (43), the search space can
M pij }L t ,Lr
i=1,j=1 and the rest M − Lr Lt columns, denoted by be regarded as a submanifold of CM , which is known as the
{q i }, form an orthonormal set that are √ orthogonal to the first
complex circle manifold (CCM) [43] and defined as
L L
√r t ij columns. Here the scaling factor M is used to ensure
M p has unit norm. We can now express v as M = S M  {u ∈ CM : |ui | = 1, i = 1, 2, · · · M }. (44)

Lt 
Lr M −L
t Lr
v = Pc = c(m−1)Lr +n p̄mn + cLt Lr +i q i The tangent space that consists of all tangent vectors to M at
m=1 n=1 i=1
the point v t can be expressed as
(39)
Tvt M = {z ∈ CM : R{z ◦ (v t )∗ } = 0}. (45)
Meanwhile, the following holds
Since the descent is performed on the manifold, we need to find

Lt 
Lr M −L
t Lr the direction of the greatest decrease from the tangent space,
M = vH v = |c(m−1)Lr +n |2 + |cLt Lr +i |2 which is known as the negative Riemannian gradient. It can be
m=1 n=1 i=1 readily verified thattheRiemannian gradient of the objective
(40) function f (v)  − K i=1 log2 (1 + ai v P i v) is given by [43]:
H

Substituting (39) into (38) and ignoring the unit-modulus  

∇fM (v t ) = ProjTvt M ∇f (v t )
constraint placed on v, we arrive at
= ∇f (v t ) − R{∇f (v t ) ◦ (v t )∗ } ◦ v t , (46)

K  ai 
max log2 1 + k |c(ik −1)Lr +ik |2
{ci },{ik }K
k=1
M where Proj(·) is the orthogonal projection operator and the
k=1
Euclidean gradient ∇f (v t ) is given as

M
s.t. |ci |2 ≤ M. (41) 
K
1

2ai P i v t
i=1 ∇f (v ) = −t
. (47)
i=1
ln 2 1 + ai (v t )H P i v t
Since we have a1 ≥ · · · ≥ aNs , choosing {ik } = {1, . . . , K}
can achieve a maximum objective function value, in which Next, the gradient decent method is utilized to update v t on
case cj should be set to zero except for those K entries the tangent space as
{c(i−1)Lr +i }K i=1 . In fact, due to the massive array gain provided
by the RIS, for those nonzero entries, an equal power allocation v̄ t = v t − t ∇fM (v t ). (48)
scheme is near-optimal, i.e., we have |c(i−1)Lr +i |2 = M K , ∀i =
1, . . . , K,in which case the above objective function value is where t is chosen as the Armijo step size [44]. Finally, the
equal to K retraction operator is performed to map the point on the tangent
i=1 log2 (1 + ai /K).
We now discuss how to determine the value of K. Based on space onto complex circle manifold M:
the above analysis, the problem (36) can be simplified as  T
1 1
 v t+1
= R(v̄ )  v̄ ◦
t t
,..., t . (49)
 ai  |v̄1 | |v̄M |
K t
max log2 1+ + RD (Ns − K)
K K
i=1 t
where v̄m denotes the mth element of v̄ t . Therefore, by iter-
s.t. K ≤ Ns . (42) atively updating (48) and (49) until the gap of the objective
function between two iterations attains a small value, an effective
Thus the optimal K  can be obtained via a simple exhaustive solution of v  can be obtained. For clarity, The above algorithm
search as the object function value can be easily calculated. is summarized in Algorithm 1.

Authorized licensed use limited to: University of Leeds. Downloaded on December 04,2024 at 21:29:32 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: JOINT BEAMFORMING AND CHANNEL RECONFIGURATION 7633

Algorithm 1: Proposed Passive Beamforming Design Al- Algorithm 2: Proposed Joint Beamforming Algorithm.
gorithm. Input: {Gp }P p=1 , {T p }p=1 , {Rp }p=1 , Rt , Rr , Ns , ρ, σ .
P P 2
1: Input channel state information, number of data streams 1: Optimize the RIS’s reflection coefficients
Ns ; Φ = diag((v  )H ) via Algorithm 1;
2: for K = 0, 1, . . . , Ns do 2: Compute the effective channel
3: Calculate the objective function value of (42); H p = Gp + Rp Φ T p , ∀p = 1, . . . , P ;
4: end for 3: Obtain optimal precoding and combining matrices, i.e.
5: Determine K  as the K that achieves the maximum F p and W p , via (14) for p = 1, . . . , P ;
objective function value; 4: Obtain hybrid precoding and combining matrices, i.e.
6: Initialize v 0 ∈ M and select a pre-defined threshold ; F RF , {F BB,p }Pp=1 , W RF and {W BB,p }p=1 via the
P
7: repeat manifold-based method discussed in Section IV.
8: Calculate the Euclidean gradient ∇f (v t ) via (47); Output: Φ , F RF , {F BB,p }P p=1 , W RF and {W BB,p }p=1 .
P
9: Calculate the Riemannian gradient ∇fM (v t ) via (46);
10: Choose an Armijo step size t as shown in [44];
11: Update v̄ t on the tangent space via (48); as
12: Update v t+1 via mapping v t onto the complex circle
manifold as shown in (49); 
P
min F p − F RF F BB,p 2F
13: until the gap of the objective function values of f (v) F RF ,{F BB,p }P
p=1 p=1
between two iterations is smaller than 
14: Output the optimal reflecting vector v  = v t+1 . s.t. |F RF [i, j]| = 1, ∀i, j. (51)

P
min W p − W RF W BB,p 2F
W RF ,{W BB,p }P
C. Convergence Analysis p=1 p=1

To analyze the convergence of Algorithm 1, we refer to the s.t. |W RF [i, j]| = 1, ∀i, j. (52)
following theorem given in [45].
Let
Theorem 1: For an optimization problem defined on compact
Riemannian manifold M, i.e. F   [F 1 , . . . , F P ],

min f (v), (50) W   [W 1 , . . . , W P ],

v∈M
F BB  [F BB,1 , . . . , F BB,P ],
if the objective function is bounded below on M, i.e. f (v) >
−∞, ∀v ∈ M, and f (v) has locally Lipschitz continuous gra- W BB  [W BB,1 , . . . , W BB,P ], (53)
dient in Euclidean space, the Riemannian gradient descent
The problem can be formulated as
method with backtracking Armijo line-search can sufficiently
returns a point v (t+1) that satisfies f (v (t+1) ) ≤ f (v (t) ) with min F  − F RF F BB 2F ,
f ∈Mf ,F BB
∇fM (v t ) ≤ 0 in O(1/20 ) iterations.
It can be easily verified that for our optimization problem, min W  − W RF W BB 2F , (54)
its objective function is bounded below on M. Meanwhile, the w∈Mw ,W BB

norm of the Hessian matrix is obviously locally bounded and where f  vec(F RF ), w  vec(W RF ) and the search space
therefore the objective function has locally Lipschitz continu- are complex circle manifold defined as Mf  S Nt Rt , Mw 
ous gradient in Euclidean space. Thus, the convergence of the S Nr Rr . Similarly, a manifold search algorithm [46] can be
proposed passive beamforming algorithm can be guaranteed. utilized to solve (54), where we optimize F RF (W RF ) and F BB
(W BB ) in an alternating manner. The detailed procedure can be
IV. HYBRID PRECODING/COMBINING DESIGN found in [46] and thus omitted here. For clarity, the proposed
algorithm is summarized in Algorithm 2.
After the passive beamforming matrix Φ = diag((v  )H ) is
optimized, the effective channel associated with each subcarrier,
i.e. H p = Gp + Rp ΦT p , ∀p can be accordingly determined. V. COMPUTATIONAL COMPLEXITY ANALYSIS
Hence the optimal fully-digital precoder and combiner, F p and We now analyze the computational complexity of our pro-
W p , associated with the pth subcarrier can be obtained via a posed joint beamforming design algorithm as summarized in
truncated SVD of the effective channel H p , as showed in (14). Algorithm 2. Specifically, our proposed method proceeds in two
Since a hybrid beamforming/combining structure is employed, steps. The first step is to solve the passive beamforming problem
we need to design a common analog precoding (combining) whose objective is to optimize the RIS’s reflection coefficients.
matrix F RF (W RF ) and a set of baseband precoding (combin- In this step, a manifold-based method is employed to solve the
ing) matrices for different subcarriers {F BB,p } ({W BB,p }) to passive beamforming problem. The dominant operation for the
approximate {F p }({W p }). Such a problem can be formulated manifold-based algorithm is to calculate the Riemannian and

Authorized licensed use limited to: University of Leeds. Downloaded on December 04,2024 at 21:29:32 UTC from IEEE Xplore. Restrictions apply.
7634 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 6, JUNE 2023

TABLE I
COMPUTATIONAL COMPLEXITY COMPARISON

Euclidean gradient (46)–(47), which has a computational com-

plexity at the order of O(M ). Let I1 denotes the number of itera-
tions required to converge, the first step involves a computational
complexity of O(M I1 ). Given fixed reflection coefficients, the
second step aims to search for the optimal hybrid precoding
and combining matrices used by the transceiver. To obtain the
optimal precoder/combiner, we need to compute the SVD of the
Fig. 2. An illustration of simulation setup.
effective channel for each subcarrier, which has a computational
complexity in the order of O(P Nr Nt min(Nr , Nt )). After that,
a manifold-based algorithm is employed to search for a common Slightly different from our previous definition, the effective
analog precoding (combining) matrix and subcarrier-dependent channel is modeled as follows by considering the transmitter
baseband precoding (combining) matrices to approximate the and receiver antenna gains [47]
optimal precoder (combiner). It can be readily verified that
this task involves a computational complexity in the order of H p = Gt Gr (Gp + Rp ΦT p ) , (56)
O(P Nr I2 + P Nt I2 ), where I2 denotes the number of itera-
tions required to converge. Therefore, the overall computa- where Gt and Gr represent the BS’s and MS’s antenna gains, re-
tional complexity of our proposed algorithm is in the order of spectively. For the BS-RIS channel T p and the RIS-MS channel
O(M I1 + P Nr Nt min(Nr , Nt ) + P Nr I2 + P Nt I2 ). Rp , the gain of the LOS path is assumed to follow a complex
For comparison, we consider two state-of-the-art methods, the Gaussian distribution α̃1 (β̃1 ) ∼ CN (0, −10−0.1κ ), in which κ
alternating optimization (AO)-based algorithm [29], [30] which is calculated according to (55) and we set a = 61.4, b = 2, and
alternatively optimizes the precoding matrix and RIS’s reflecting σξ = 5.8dB as suggested by LOS real-world channel measure-
coefficients, and the simple joint design (SJD) method [31] ments [48]. On the other hand, the gain of the NLOS paths is
whose basic idea is to set the RIS coefficients to align the assumed to follow α̃i (β̃i ) ∼ CN (0, −10−0.1(κ+μ) ), where μ is
dominant path of the BS-RIS-MS channel. To reduce the com- the Rician factor [49]. As for the BS-MS channel Gp , the path
putational complexity of the AO algorithm, the majorization- gain is assumed to follow ˜i ∼ CN (0, −10−0.1(κ+δ) ), in which
minimization (MM) technique proposed in [30] is considered κ is calculated by setting a = 72, b = 2.92, and σξ = 8.7dB, as
in our experiments, and the method is termed as “AO-MM”. suggested by NLOS real-world channel measurements [48], and
Let IAO and IMM denote the number of iterations required for each path is assumed to pass through tintedglass walls with an
the AO scheme and the MM technique, respectively. To provide additional penetration loss δ = 40.1 dB [50]. The delay spread τ
a better comparison of these methods, we suppose Nt ≥ Nr , for each path is independently generated and follows a uniform
Nr  Rr = Rt ≥ Ns , and M  Ld = Lt = Lr , which is a distribution.
reasonable setup for practical mmWave systems. The compu- Unless otherwise specified, we assume that the RIS employs a
tational complexity of respective algorithms is summarized in UPA with M = My × Mz = 16 × 16 passive elements, the BS
Table I. We see that our proposed method has a lower complexity and the MS adopt ULAs with Nt = 16 and Nr = 16 antennas.
than the AO-based method. Other parameters are set as follows: P = 16, Rt = Rr = 8,
Ns = 6, Lt = Lr = Ld = 8, Gt = 24.5dBi, Gr = 0dBi, fs =
VI. SIMULATION RESULTS 2 GHz. The central carrier frequency is set to 28 GHz, the
bandwidth of subcarrier is set to 250 MHz and thus the noise
We now present simulation results to illustrate the perfor- power is σ 2 = −174 + 10 log10 B = −90 dBm.
mance of our proposed algorithm. In our simulations, we con- We compare our proposed method with the following two
sider a three-dimensional setup as shown in Fig. 2. The coor- algorithms, namely, the alternating optimization (AO)-based
dinates of the BS, the RIS and the MS are respectively set as algorithm [29], [30] which alternatively optimizes the precoding
(xBS , 0, zBS ), (0, yRIS , zRIS ), and (xMS , yMS , zMS ), where we set matrix and RIS’s reflecting coefficients, and the simple joint
xBS = 2 m, zBS = 10 m, yRIS = 148 m, zRIS = 10 m, xMS = 5 m, design (SJD) method [31] whose basic idea is to set the RIS co-
yMS = 150 m, and zMS = 1.5 m. The distance-dependent path efficients to align the dominant path of the BS-RIS-MS channel.
loss is modeled as To reduce the computational complexity of the AO algorithm, the
˜ + ξ,
κ = a + 10b log10 (d) (55) majorization-minimization (MM) technique proposed in [30]
is considered in our experiments, and the method is termed as
where d˜ denotes the distance between the transmitter and the “AO-MM”. Besides, we also consider the case when there is no
receiver, and ξ ∼ N (0, σξ2 ). RIS in the system, which is labeled as “No RIS”.

Authorized licensed use limited to: University of Leeds. Downloaded on December 04,2024 at 21:29:32 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: JOINT BEAMFORMING AND CHANNEL RECONFIGURATION 7635

Fig. 3. Spectral efficiency versus the transmit power ρ. (a) μ = 10 dB. (b) μ = 1 dB.

Fig. 4. Spectral efficiency versus the number of data streams Ns . (a) μ = 10 dB. (b) μ = 1 dB.

Fig. 3 plots the spectral efficiency of respective algorithms does not improve as significantly as the other two methods as the
versus the transmit power ρ with different Rician factors μ. We number of data streams grows, which is due to the fact that the
see that the RIS-assisted system achieves a significant perfor- SJD method only utilizes the dominant path of the BS-RIS-MS
mance improvement over the system without RIS. Also, our channel and fails to fully utilize the spatial diversity rendered by
proposed method presents a clear performance advantage over the reflection channel for multiple stream transmission.
the other two methods as the transmit power increases. This In Fig. 5, we report the spectral efficiency of respective
is probably because our proposed method can fully utilize the algorithms versus the number of reflecting elements M , where
spatial diversity rendered by the BS-RIS-MS channel. Due to we set the transmit power ρ = 30 dBm. In Fig. 5, we see that for
the water-filling power allocation, the spatial diversity can be all algorithms, the spectral efficiency increases as the number of
better utilized when the transmit power is large. In addition, we reflection elements increases, and RIS-assisted systems provide
see that the performance improvement of our proposed method a substantial advantage over the no-RIS communication system.
over the other two methods is more pronounced when the Rician Meanwhile, our proposed method outperforms the AO-MM and
factor is set to μ = 1 dB. This is because the spatial diversity can the SJD method by a big margin. Note that, compared with
be more efficiently utilized for channels with a smaller Rician the AO-MM and the SJD, the performance advantage of our
factor. proposed method is more pronounced as M increases. This
Fig. 4 depicts the spectral efficiency of respective algorithms is because more reflecting elements can improve the gains of
as a function of the number of data streams Ns , where we set ρ = the reflected paths, thus providing a more favorable channel
30dBm. We can observe that our proposed method outperforms condition in which the spatial diversity can be more efficiently
the AO-MM method and the SJD method. Also, the SJD method exploited.

Authorized licensed use limited to: University of Leeds. Downloaded on December 04,2024 at 21:29:32 UTC from IEEE Xplore. Restrictions apply.
7636 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 6, JUNE 2023

Fig. 5. Spectral efficiency versus the number of RIS elements M . (a) μ = 10 dB. (b) μ = 1 dB.

Fig. 6. Average run time versus the number of RIS elements M . Fig. 7. Truncated condition number versus number of RIS elements M .

In Fig. 6, the average runtimes of respective algorithms are

depicts the truncated condition number of respective algorithms
depicted. It can be observed that our proposed method is com-
as a function of the number of reflecting elements M , where
putationally much more efficient than the AO-MM method. The
we set ρ = 30 dBm and μ = 10 dB. It can be observed that the
SJD method has the lowest complexity as its passive beam-
truncated condition number of the SJD method grows rapidly
forming vector, simply determined by the dominant reflection
as M increases. This is because the SJD method optimizes the
path, has a simple analytical solution. Nevertheless, as indicated
RIS’s reflection coefficients only to enhance the dominant path
earlier, the SJD method cannot fully exploit the spatial diversity
of the BS-RIS-MS channel, with other reflected paths neglected.
rendered by the reflection channel, and thus suffers a substantial
In contrast, the truncated condition number of our proposed
performance loss compared to our proposed method. We also
method increases slightly with an increasing M , thus creating
observe that the complexity of our proposed method increases
a more favorable channel for exploiting spatial diversity. We
slowly as M grows. Overall, our proposed method can achieve
see that the AO-MM method also achieves a condition number
a good balance between the performance and the computational
similar to our proposed method. Nevertheless, it is observed
complexity.
that the singular values of the effective channel obtained by
To gain an insight into how the RIS helps realize a favorable
the AO-MM are generally smaller than those of our proposed
propagation environment, we use the truncated condition num-
method, which prevents the AO-MM from achieving a higher
ber as a metric to evaluate the capability of respective algorithms
spectral efficiency.
in reconfiguring the wireless channel. The average truncated
 |Σ(p) [1,1]|
condition number is defined as P1 P p=1 |Σ(p) [N ,N ]| , where
s s
VII. CONCLUSION
Σ(p) [i, i] denotes the ith largest singular value of the effective
channel H p . It is noted that a smaller truncated condition num- This paper considered the problem of joint beamforming for
ber indicates a more favorable propagation environment since point-to-point RIS-assisted mmWave MIMO-OFDM systems.
we prefer the power of the channel to be uniformly distributed By exploiting the inherent structure of the mmWave effective
over singular values to support multi-stream transmission. Fig. 7 channel, we established a mathematical relationship between the

Authorized licensed use limited to: University of Leeds. Downloaded on December 04,2024 at 21:29:32 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: JOINT BEAMFORMING AND CHANNEL RECONFIGURATION 7637

singular values of the effective channel and the reflection coef- [20] M. Cui, G. Zhang, and R. Zhang, “Secure wireless communication via
ficients, based on which we developed a SVD approximation- intelligent reflecting surface,” IEEE Wireless Commun. Lett., vol. 8, no. 5,
pp. 1410–1414, Oct. 2019.
based method for joint beamforming. An important advantage [21] B. Zheng and R. Zhang, “Intelligent reflecting surface-enhanced OFDM:
of the proposed method is that it can help achieve a desired Channel estimation and reflection optimization,” IEEE Wireless Commun.
channel by directly manipulating the singular values of the Lett., vol. 9, no. 4, pp. 518–522, Apr. 2020.
[22] Y. Yang, B. Zheng, S. Zhang, and R. Zhang, “Intelligent reflecting surface
effective channel. Simulation results shows that our proposed meets OFDM: Protocol design and rate maximization,” IEEE Trans.
method can realize a favorable propagation channel with a small Commun., vol. 68, no. 7, pp. 4522–4535, Jul. 2020.
condition number, and achieve a clear performance advantage [23] K. Feng, X. Li, Y. Han, and Y. Chen, “Joint beamforming optimization for
reconfigurable intelligent surface-enabled MISO-OFDM systems,” China
over existing methods. Commun., vol. 18, no. 3, pp. 63–79, Mar. 2021.
[24] B. Ning, Z. Chen, W. Chen, and J. Fang, “Beamforming optimization for
intelligent reflecting surface assisted MIMO: A sum-path-gain maximiza-
tion approach,” IEEE Wireless Commun. Lett., vol. 9, no. 7, pp. 1105–1109,
REFERENCES Jul. 2020.
[25] P. Wang, J. Fang, L. Dai, and H. Li, “Joint transceiver and large intelli-
[1] T. S. Rappaport, J. N. Murdock, and F. Gutierrez, “State of the art in
gent surface design for massive MIMO mmWave systems,” IEEE Trans.
60-GHz integrated circuits and systems for wireless communications,”
Wireless Commun., vol. 20, no. 2, pp. 1052–1064, Feb. 2021.
Proc. IEEE, vol. 99, no. 8, pp. 1390–1436, Aug. 2011.
[26] W. Chen, C.-K. Wen, X. Li, and S. Jin, “Channel customization for joint
[2] S. Rangan, T. S. Rappaport, and E. Erkip, “Millimeter-wave cellular
Tx-RISs-Rx design in hybrid mmWave systems,” Sep. 2021. [Online].
wireless networks: Potentials and challenges,” Proc. IEEE, vol. 102, no. 3,
Available: https://arxiv.org/abs/2109.13058
pp. 366–385, Mar. 2014.
[27] R. Li, B. Guo, M. Tao, Y.-F. Liu, and W. Yu, “Joint design of hybrid
[3] A. Ghosh et al., “Millimeter-wave enhanced local area systems: A high-
beamforming and reflection coefficients in RIS-aided mmWave MIMO
data-rate approach for future wireless networks,” IEEE J. Sel. Areas
systems,” IEEE Trans. Commun., vol. 70, no. 4, pp. 2404–2416, Apr. 2022.
Commun., vol. 32, no. 6, pp. 1152–1163, Jun. 2014.
[28] C. Feng, W. Shen, J. An, and L. Hanzo, “Joint hybrid and passive RIS-
[4] A. L. Swindlehurst, E. Ayanoglu, P. Heydari, and F. Capolino, “Millimeter-
assisted beamforming for mmWave MIMO systems relying on dynam-
wave massive MIMO: The next wireless revolution?,” IEEE Commun.
ically configured subarrays,” IEEE Internet of Things J., vol. 9, no. 15,
Mag., vol. 52, no. 9, pp. 56–62, Sep. 2014.
pp. 13913–13926, Aug. 2022.
[5] O. Abari, “Enabling high-quality untethered virtual reality,” in Proc. 1st
[29] S. Zhang and R. Zhang, “Capacity characterization for intelligent reflecting
ACM Workshop Millim.-Wave Netw. Sens. Syst., 2017, Art. no. 49.
surface aided MIMO communication,” IEEE J. Sel. Areas Commun.,
[6] G. R. MacCartney, T. S. Rappaport, and S. Rangan, “Rapid fading due to
vol. 38, no. 8, pp. 1823–1838, Aug. 2020.
human blockage in pedestrian crowds at 5G millimeter-wave frequencies,”
[30] W. Jiang, B. Chen, J. Zhao, Z. Xiong, and Z. Ding, “Joint active
in Proc. IEEE Glob. Commun. Conf., 2017, pp. 1–7.
and passive beamforming design for the IRS-assisted MIMOME-OFDM
[7] X. Tan, Z. Sun, D. Koutsonikolas, and J. M. Jornet, “Enabling indoor
secure communications,” IEEE Trans. Veh. Technol., vol. 70, no. 10,
mobile millimeter-wave networks based on smart reflect-arrays,” in Proc.
pp. 10369–10381, Oct. 2021.
IEEE Conf. Comput. Commun., 2018, pp. 270–278.
[31] S. H. Hong, J. Park, S.-J. Kim, and J. Choi, “Hybrid beamforming for
[8] P. Wang, J. Fang, X. Yuan, Z. Chen, and H. Li, “Intelligent reflect-
intelligent reflecting surface aided millimeter wave MIMO systems,” IEEE
ing surface-assisted millimeter wave communications: Joint active and
Trans. Wireless Commun., vol. 21, no. 9, pp. 7343–7357, Sep. 2022.
passive precoding design,” IEEE Trans. Veh. Technol., vol. 69, no. 12,
[32] A. Alkhateeb and R. W. Heath, “Frequency selective hybrid precoding
pp. 14960–14973, Dec. 2020.
for limited feedback millimeter wave systems,” IEEE Trans. Wireless
[9] N. S. Perović, M. D. Renzo, and M. F. Flanagan, “Channel capacity
Commun., vol. 64, no. 5, pp. 1801–1818, May 2016.
optimization using reconfigurable intelligent surfaces in indoor mmWave
[33] X. Yu, J.-C. Shen, J. Zhang, and K. B. Letaief, “Alternating minimization
environments,” in Proc. IEEE Int. Conf. Commun., 2020, pp. 1–7.
algorithms for hybrid precoding in millimeter wave MIMO systems,” IEEE
[10] X. Yang, C.-K. Wen, and S. Jin, “MIMO detection for reconfigurable
J. Sel. Areas Commun., vol. 10, no. 3, pp. 485–500, Apr. 2016.
intelligent surface-assisted millimeter wave systems,” IEEE J. Sel. Areas
[34] Z. Gao, C. Hu, L. Dai, and Z. Wang, “Channel estimation for millimeter-
Commun., vol. 38, no. 8, pp. 1777–1792, Aug. 2020.
wave massive MIMO with hybrid precoding over frequency-selective
[11] C. Liaskos, S. Nie, A. Tsioliaridou, A. Pitsillides, S. Ioannidis,
fading channels,” IEEE Commun. Lett., vol. 20, no. 6, pp. 1259–1262,
and I. Akyildiz, “A new wireless communication paradigm through
Jun. 2016.
software-controlled metasurfaces,” IEEE Commun. Mag., vol. 56, no. 9,
[35] P. Wang, J. Fang, H. Duan, and H. Li, “Compressed channel estimation
pp. 162–169, Sep. 2018.
for intelligent reflecting surface-assisted millimeter wave systems,” IEEE
[12] Y.-C. Liang, R. Long, Q. Zhang, J. Chen, H. V. Cheng, and H. Guo, “Large
Signal Process. Lett., vol. 27, pp. 905–909, 2020.
intelligent surface/antennas (LISA): Making reflective radios smart,” J.
[36] J. He, H. Wymeersch, and M. Juntti, “Channel estimation for RIS-aided
Commun. Inf. Netw., vol. 4, no. 2, pp. 40–50, Jun. 2019.
mmWave MIMO systems via atomic norm minimization,” IEEE Trans.
[13] S. Hu, F. Rusek, and O. Edfors, “Beyond massive MIMO: The potential
Wireless Commun., vol. 20, no. 9, pp. 5786–5797, Sep. 2021.
of data transmission with large intelligent surfaces,” IEEE Trans. Signal
[37] Y. Liu, S. Zhang, F. Gao, J. Tang, and O. A. Dobre, “Cascaded channel esti-
Process., vol. 66, no. 10, pp. 2746–2758, May 2018.
mation for RIS assisted mmWave MIMO transmissions,” IEEE Commun.
[14] E. Basar, M. Di Renzo, J. De Rosny, M. Debbah, M.-S. Alouini, and
Lett., vol. 10, no. 9, pp. 2065–2069, Sep. 2021.
R. Zhang, “Wireless communications through reconfigurable intelligent
[38] X. Wei, D. Shen, and L. Dai, “Channel estimation for RIS assisted wireless
surfaces,” IEEE Access, vol. 7, pp. 116753–116773, 2019.
communications—Part I: Fundamentals, solutions, and future opportuni-
[15] Q. Wu and R. Zhang, “Towards smart and reconfigurable environment:
ties,” IEEE Commun. Lett., vol. 25, no. 5, pp. 1398–1402, May 2021.
Intelligent reflecting surface aided wireless network,” IEEE Commun.
[39] X. Zheng, P. Wang, J. Fang, and H. Li, “Compressed channel estimation
Mag., vol. 58, no. 1, pp. 106–112, Jan. 2020.
for IRS-assisted millimeter wave OFDM systems: A low-rank tensor
[16] Y. Han, W. Tang, S. Jin, C.-K. Wen, and X. Ma, “Large intelligent surface-
decomposition-based approach,” IEEE Wireless Commun. Lett., vol. 11,
assisted wireless communication exploiting statistical CSI,” IEEE Trans.
no. 6, pp. 1258–1262, Jun. 2022.
Veh. Technol., vol. 68, no. 8, pp. 8238–8242, Aug. 2019.
[40] O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath, “Spa-
[17] Q. Wu and R. Zhang, “Intelligent reflecting surface enhanced wireless
tially sparse precoding in millimeter wave MIMO systems,” IEEE Trans.
network: Joint active and passive beamforming design,” in Proc. IEEE
Wireless Commun., vol. 13, no. 3, pp. 1499–1513, Mar. 2014.
Glob. Commun. Conf., 2018, pp. 1–6.
[41] X. Gao, L. Dai, S. Han, C.-L. I, and R. W. Heath, “Energy-efficient hybrid
[18] C. Huang, A. Zappone, G. C. Alexandropoulos, M. Debbah, and C.
analog and digital precoding for mmWave MIMO systems with large
Yuen, “Reconfigurable intelligent surfaces for energy efficiency in wire-
antenna arrays,” IEEE J. Sel. Areas Commun., vol. 34, no. 4, pp. 998–1009,
less communication,” IEEE Trans. Wireless Commun., vol. 18, no. 8,
Apr. 2016.
pp. 4157–4170, Aug. 2019.
[42] A. Goldsmith, S. Jafar, N. Jindal, and S. Vishwanath, “Capacity lim-
[19] Q. Wu and R. Zhang, “Intelligent reflecting surface enhanced wireless
its of MIMO channels,” IEEE J. Sel. Areas Commun., vol. 21, no. 5,
network via joint active and passive beamforming,” IEEE Trans. Wireless
pp. 684–702, Jun. 2003.
Commun., vol. 18, no. 11, pp. 5394–5409, Nov. 2019.

Authorized licensed use limited to: University of Leeds. Downloaded on December 04,2024 at 21:29:32 UTC from IEEE Xplore. Restrictions apply.
7638 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 6, JUNE 2023

[43] K. Alhujaili, V. Monga, and M. Rangaswamy, “Transmit MIMO radar Hongbin Li (Fellow, IEEE) received the B.S. and
beampattern design via optimization on the complex circle manifold,” M.S. degrees in electrical engineering from the Uni-
IEEE Trans. Signal Process., vol. 67, no. 13, pp. 3561–3575, Jul. 2019. versity of Electronic Science and Technology of
[44] P.-A. Absil, R. Mahony, and R. Sepulchre, Optim. Algorithms Matrix China, Chengdu, China, in 1991 and 1994, respec-
Manifolds. Princeton, NJ, USA:Princeton Univ. Press, 2007. tively, and the Ph.D. degree in electrical engineering
[45] N. Boumal, P.-A. Absil, and C. Cartis, “Global rates of convergence for from the University of Florida, Gainesville, FL, USA,
nonconvex optimization on manifolds,” IMA J. Numer. Anal., vol. 39, no. 1, in 1999. From July 1996 to May 1999, he was a
pp. 1–33, 2019. Research Assistant with the Department of Electrical
[46] H. Kasai, “Fast optimization algorithm on complex oblique manifold for and Computer Engineering, University of Florida.
hybrid precoding in millimeter wave MIMO systems,” in Proc. IEEE Glob. Since July 1999, he has been with the Department
Conf. Signal Inf. Process., 2018, pp. 1266–1270. of Electrical and Computer Engineering, Stevens In-
[47] B. Ning, Z. Chen, W. Chen, Y. Du, and J. Fang, “Terahertz multi-user mas- stitute of Technology, Hoboken, NJ, USA, where he is currently the Charles
sive MIMO with intelligent reflecting surface: Beam training and hybrid and Rosanna Batchelor Memorial Chair Professor. He was a Summer Visiting
beamforming,” IEEE Trans. Veh. Technol., vol. 70, no. 2, pp. 1376–1393, Faculty Member with the Air Force Research Laboratory in the summers of
Feb. 2021. 2003, 2004 and 2009. His general research interests include statistical signal
[48] M. R. Akdeniz et al., “Millimeter wave channel modeling and cellu- processing, wireless communications, and radars. Dr. Li was the recipient of
lar capacity evaluation,” IEEE J. Sel. Areas Commun., vol. 32, no. 6, the IEEE Jack Neubauer Memorial Award in 2013 from the IEEE Vehicular
pp. 1164–1179, Jun. 2014. Technology Society, Outstanding Paper Award from the IEEE AFICON Con-
[49] M. K. Samimi, G. R. MacCartney, S. Sun, and T. S. Rappaport, “28 ference in 2011, Provost’s Award for Research Excellence in 2019, Harvey N.
GHz millimeter-wave ultrawideband small-scale fading models in wireless Davis Teaching Award in 2003, and Jess H. Davis Memorial Research Award
channels,” in Proc. IEEE 83rd Veh. Technol. Conf., 2016, pp. 1–6. in 2001 from Stevens Institute of Technology, and Sigma Xi Graduate Research
[50] H. Zhao et al., “28 GHz millimeter wave cellular communication measure- Award from the University of Florida in 1999. He has been a Member of the
ments for reflection and penetration loss in and around buildings in New IEEE SPS Signal Processing Theory and Methods Technical Committee (TC)
York city,” in Proc. IEEE Int. Conf. Commun., 2013, pp. 5163–5167. and the IEEE SPS Sensor Array and Multichannel TC, an Associate Editor for
Signal Processing (Elsevier), IEEE TRANSACTIONS ON SIGNAL PROCESSING,
IEEE SIGNAL PROCESSING LETTERS, and IEEE TRANSACTIONS ON WIRELESS
Hanyu Wang (Student Member, IEEE) received the COMMUNICATIONS, and the Guest Editor of the IEEE JOURNAL OF SELECTED
B.Sc. degree in 2016 from the University of electronic TOPICS IN SIGNAL PROCESSING and EURASIP Journal on Applied Signal Pro-
Science and Technology of China, Chengdu, China, cessing. He has been involved in various conference organization activities,
where he has been working toward the Ph.D. degree including the General Co-Chair of the 7th IEEE Sensor Array and Multichannel
since September 2018. His research interests include Signal Processing Workshop, Hoboken, NJ, USA, June 17-20, 2012. Dr. Li is a
compressed sensing, millimeter wave, and massive Member of Tau Beta Pi and Phi Kappa Phi.
multiple-input and multiple-output (MIMO) commu-
nications.

Jun Fang (Senior Member, IEEE) received the B.S.

and M.S. degrees in electrical engineering from Xid-
ian University, Xi’an, China in 1998 and 2001, respec-
tively, and the Ph.D. degree in electrical engineering
from the National University of Singapore, Singa-
pore, in 2006. In 2006, he was a Postdoctoral Research
Associate with the Department of Electrical and Com-
puter Engineering, Duke University, Durham, NC,
USA. From January 2007 to December 2010, he was
a Research Associate with the Department of Elec-
trical and Computer Engineering, Stevens Institute of
Technology, Hoboken, NJ, USA. Since 2011, he has been with the University
of Electronic of Science and Technology of China, Chengdu, China, where
he is currently a Professor. His research interests include compressed sensing
and sparse theory, massive MIMO/mmWave communications, and statistical
inference. Dr. Fang was the recipient of the IEEE Jack Neubauer Memorial
Award in 2013 for the Best Systems Paper published in the IEEE TRANSACTIONS
ON VEHICULAR TECHNOLOGY. He was an Associate Technical Editor for IEEE
Communications Magazine from 2012 to 2020. He is also a Member of the
IEEE SPS Sensor Array and Multichannel TC, and the Senior Associate Editor
of IEEE SIGNAL PROCESSING LETTERS.

Authorized licensed use limited to: University of Leeds. Downloaded on December 04,2024 at 21:29:32 UTC from IEEE Xplore. Restrictions apply.