Intelligent Reflecting Surface Enhanced Wideband

MIMO-OFDM Communications: From Practical

Model to Reflection Optimization
Hongyu Li, Student Member, IEEE, Wenhao Cai, Yang Liu, Member, IEEE, Ming Li, Senior Member, IEEE, and
Qian Liu, Member, IEEE

Abstract—Intelligent reflecting surface (IRS) is envisioned as quality of service (QoS) requirements for future wireless
arXiv:2007.14243v1 [eess.SP] 26 Jul 2020

a revolutionary technology for future wireless communication communications.

systems since it can intelligently change radio environment and In the current paradigm of wireless communication op-
integrate it into wireless communication optimization. However,
most recent investigation utilized an ideal IRS reflection model, timization, the radio environment and wireless propagation
which is impractical and can cause significant performance medium remain an uncontrollable factor, which cannot be
degradation in realistic wideband systems. In this work, we first included in the optimization formulations. Thus, channel fad-
study the amplitude-frequency-phase relationship of reflected ing effect due to the randomness in the radio environment is
signals and present a simplified practical IRS reflection model generally a major challenge for the maximization of EE/SE
for wideband signals. Then, an IRS enhanced wideband mul-
tiuser multi-input single-output orthogonal frequency division performance of wireless communications. Recently, an inno-
multiplexing (MU-MISO-OFDM) system is investigated. We aim vative concept of intelligent reflecting surface (IRS) has been
to jointly design the transmit beamformer and IRS reflection introduced in the wireless communication research community
to maximize the average sum-rate over all subcarriers. With as a revolutionary technology, which can realize controllable
the aid of the relationship between sum-rate maximization and radio environment and combat stochastic wireless propagation
mean square error (MSE) minimization, the original problem
is equivalently transformed into a multi-block/variable problem, medium [4]-[12].
which can be solved by classic block coordinate descent (BCD) The IRS consists of a large number of nearly passive
method. Complexity and convergence for both cases are analyzed elements with ultra-low power consumption. Particularly, each
or illustrated. Simulation results demonstrate that the proposed element of IRS is composed of configurable electromagnetic
algorithm can offer significant average sum-rate enhancement (EM) internals, which are capable of controlling the phase shift
compared to that achieved using the ideal reflection model, which
confirms the importance of the use of the practical model for the and amplitude of the incident EM wave in a programmable
design of wideband systems. manner. Adaptively adjusting elements of IRS can collabo-
ratively achieve reflection beamforming and shape the propa-
Index terms— Intelligent reflecting surface (IRS), multi- gation environment suitable for wireless communications. The
user multi-input single-output (MU-MISO), orthogonal fre- channel/beamforming gain can be effectively improved and the
quency division multiplexing (OFDM), beamforming. communication quality can be enhanced. Free of containing
radio frequency (RF) chains, large-scale IRS can be deployed
I. I NTRODUCTION
in different communication situations with lower power con-
The continuous popularizing of intelligent devices and the sumption and cost. Therefore, IRS is envisioned to revolu-
rapid development of emerging wireless services have caused tionize the current communication optimization paradigm by
the exponential increase of the demand for wireless network integrating the smart radio environment and expected to play
traffic. This motivates the research on key enabling technolo- an important role in future wireless communications.
gies, such as massive multi-input multi-output (MIMO), ultra- Attracted by the sheer advantage of IRS, the investigation
dense network, and the use of millimeter wave (mmWave) of IRS for improving the performance of various wireless
bands [1]-[3], for the fifth-generation (5G) and beyond net- communication systems is a thriving research area in the last
works. However, the above technologies still inevitably face few years. A majority of recent research efforts have been
challenges mainly due to high cost and power consump- devoted to the IRS designs with focus on power allocation
tions when employing multiple antennas, cells (base stations and/or beamformer for both single-user systems [13]-[15]
(BSs)), and/or hardware components (e.g. radio frequency and multi-user systems [16]-[20] using different metrics, e.g.
(RF) chains). Therefore, researchers have never stopped their power minimization [14], [19], max-min fairness [19], [20],
efforts to seek spectral- and energy-efficiency (SE/EE) solu- SE maximization [13]-[15], [18], and EE maximization [16].
tions to accommodate the demanding data rate and diverse In some recent works [21]-[23], practical IRS implementation
H. Li,W. Cai, Y. Liu, and M. Li are with the School of Infor- with finite/low-resolution phase shifts are considered. In order
mation and Communication Engineering, Dalian University of Technol- to further highlight the flexibility of the IRS employment,
ogy, Dalian 116024, China, (e-mail: [email protected], wenhao- many researchers also studied the coordination of multiple
[email protected], [email protected], [email protected]).
Q. Liu is with the School of Computer Science and Technology, Dalian Uni- IRSs [24]-[26]. Moreover, IRS technique has also been em-
versity of Technology, Dalian 116024, China (e-mail: [email protected]). ployed in other applications, e.g. physical layer security [27]-
[29], cognitive radio [30]-[32], as well as index modulation
Z(C, f)
[33], [34], etc. L1
It is worth noting that the IRS-assisted wireless commu- Incident Reflected
Signal of f Signal of f
nication systems mentioned above are restricted to narrow-
band channels. When considering more general wideband
L2 R C
frequency-selective channels, the problem will be quite dif-
ferent and more difficult to be solved since the common IRS A Reflecting Element
should be designed for all subcarriers, while the conventional
digital beamformers can be individually optimized for each Fig. 1. The equivalent circuit of an IRS element.
subcarrier. Some work has studied the IRS-enhanced wideband
orthogonal frequency division multiplexing (OFDM) system
for both simple single-input single-output (SISO) case [35]- tween sum-rate maximization and mean square error
[37] and more typical multi-user MISO case [38]. (MSE) minimization, the problem is converted to a multi-
The aforementioned work assumes that IRS have an ideal block/variable optimization, which can be solved by the
model with perfect signal reflection, i.e. each element has con- classical block coordinate descent (BCD) method.
stant magnitude, variable phase shift, and the same response • Finally, we evaluate our proposed design. We analyze the

for wideband signals. The design of IRS with such an ideal complexity and illustrate the convergence. Moreover, the
reflection model can be easily implemented using classical op- performance of the proposed algorithm is validated by
timization tools, e.g. semidefinite relaxation (SDR), manifold extensive simulation studies, which confirm the effective-
optimization, majorization minimization (MM), etc. However, ness of the design with the practical model compared to
it is extremely difficult to implement an IRS having such an that with the ideal one.
ideal reflection model due to the hardware circuit limitation Notations: Boldface lower-case and upper-case letters indi-
[39], [40]. Therefore, these “ideal” designs will cause non- cate column vectors and matrices, respectively. C and R+ de-
negligible performance loss in realistic systems since the ideal note the set of complex and positive real numbers, respectively.
model cannot precisely describe the response of a practical (·)∗ , (·)T , (·)H , and (·)−1 denote the conjugate, transpose,
IRS. Therefore, it is important and necessary to analyze the conjugate-transpose operations, and inversion, respectively.
response characteristic of a practical IRS and establish an E{·} represents statistical expectation. <{·} denotes the real
accurate and practical IRS reflection model. The authors in part of a complex number. IL indicates an L × L identity
[41] have illustrated the fundamental relationship between re- matrix. kAkF denotes the Frobenius norm of matrix A. kak2
flection amplitude and phase shift under a narrowband scenario denotes the `2 norm of vector a. ⊗ denotes the Kronecker
and demonstrated the performance enhancement with their product. blkdiag(·) denotes a block matrix such that the main-
proposed practical model compared to that with the ideal one. diagonal blocks are matrices and all off-diagonal blocks are
When expanding to wideband communications, unfortunately, zero matrices. Finally, A(i, :), A(:, j), and A(i, j) denote the
the above two-dimensional amplitude-phase relationship can- i-th row, the j-th column, and the (i, j)-th element of matrix
not accurately describe the response of the practical IRS, A, respectively. a(i) denotes the i-th element of vector a.
which will vary with the frequencies of incident signals. In
our previous work [42], we have analyzed this issue and II. P RACTICAL IRS M ODELING
established a three-dimensional amplitude-frequency-phase re-
The hardware construction of IRS is usually based on
lationship to precisely describe the response of practical IRS
the printed circuit board (PCB) with uniformly distributed
in wideband systems. Nevertheless, this practical model is so
reflecting elements on a planar surface. A typical IRS generally
complicated that it will cause great difficulties in the IRS
consists of three layers: i) An outer layer with a large number
reflection design. This motivates us to further simplify the
of metal elements printed on the PCB dielectric substrate; ii) a
practical IRS model in order to facilitate the reflection design
copper plate to avoid the leakage of signal energy; iii) a control
without significant accuracy loss.
circuit board for IRS control [4]. A semiconductor device, such
In this paper, we consider an IRS-enhanced wideband MU-
as the positive-intrinsic-negative (PIN) diode, is embedded into
MISO-OFDM communication system. Specifically, we present
each metal element in the outer layer to tune the reflecting
a simplified practical reflection model of IRS and take it into
response, e.g. phase shift and amplitude. The response of each
consideration for the reflection design. Our main contributions
reflecting element can be equivalently modeled as a parallel
are summarized as follows:
resonance circuit as shown in Fig. 1. Thus, the impedance of
• We re-analyze the characteristic of IRS elements, i.e.
an IRS element for the signal of frequency f can be written
phase and amplitude variations of IRS elements when as
responding to signals with different frequencies. Based 1
j2πf L1 (j2πf L2 + j2πf C + R)
on our previous work, we present a leaner practical model Z(C, f ) = 1 , (1)
j2πf L1 + j2πf L2 + j2πf C + R
of IRS reflection, which is applicable to the designs of
typical communication scenarios. where L1 , L2 , C, and R denote the metal plate inductance,
• Then, we aim to jointly design the beamformer and outer layer inductance, effective capacitance, and the loss
the reflection of IRS to achieve maximum average sum- resistance, respectively. The reflection coefficient of each IRS
rate over all subcarriers. Based on the equivalence be- element, denoted as φ, is fundamentally the ratio of the power
 1 1
3
f = 2.35GHz = 11 /12
0.9 0.95 f = 2.4GHz Fitted, = 11 /12
f = 2.45GHz
Amplitude

= 2 /3
0.8 =0 Fitted Curve 2
0.9 Fitted, = 2 /3
= - /3 = /3
0.7 = - 2 /3
Fitted, = /3
= /3 0.85 1 =0

Phase shift (rad)

0.6 = 2 /3
Fitted, = 0

Amplitude
2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3
0.8 = - /3
0 Fitted, = - /3
Frequency (GHz)
0.75 = - 2 /3
4 Fitted, = - 2 /3
=0 = -11 /12
0.7 -1
2 = - /3 Fitted, = -11 /12
Phase (rad)

= - 2 /3
= /3 0.65
0 = 2 /3 -2
0.6
-2

0.55 -3
-4
2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 -3 -2 -1 0 1 2 3 2.35 2.36 2.37 2.38 2.39 2.4 2.41 2.42 2.43 2.44 2.45
Frequency (GHz) Phase Shift (rad) Frequency (GHz)

(a) (b) (c)

Fig. 2. (a) The illustration of the dual phase- and amplitude-squint [42]. With a certain phase shift θ for the carrier frequency fc = 2.4GHz, the amplitudes
and phase shifts for other frequencies vary. With different phase shift θ, the variation range and trend of the amplitudes and phase shifts for other frequencies
also vary. (b) The relationship between the amplitude and the phase shift for corresponding frequencies. (c) The phase shift as a function of frequency.

of the reflected signal to that of the incident one, which is TABLE I

therefore given by VALUES OF PARAMETERS FOR THE SIMPLIFIED IRS MODEL .

Z(C, f ) − Z0 i=1 i=2 i=3 i=4 i=5

, φ= (2)
Z(C, f ) + Z0 ai 0.06 11.27 10.88 89.64 26.11
where Z0 denotes the free space impedance. Here, we should bi 0.02 0.008996 0.9799 0.01268 0.9796
ci 0.5736 -1.897 -1.471 0.2899 1.673
emphasize that the reflection of the IRS element is a function
of C and f . When each element is controlled by selecting
an appropriate capacitance C, the response of each element Fig. 2(b) that, the relationship between amplitude and phase
is also associated with the frequency of the incident signals. shift for different frequencies can be viewed as a quadratic
Our previous work [42] has demonstrated that the same IRS function. Moreover, within the bandwidth B = 100MHz, the
element actually exhibits different responses (i.e. different am- curves for different frequencies do not have obvious difference,
plitudes and phase shifts) to signals with different frequencies, which motivates us to use a unified fit function for different
which is referred to as dual phase- and amplitude-squint effect frequencies. Then in Fig. 2(c), we can find that the phase
in this paper. Fig. 2(a) illustrates an example of the amplitude shift as a function of frequency can be approximately fitted as
and phase shift variations of an IRS element as a function of a straight line. When the BPS θ of one IRS element varies,
frequency. Let us name the phase shift θ for signal of central the slope and intercept for the phase shift-frequency line will
carrier frequency fc as the basic phase shift (BPS) for clear be different.
and concise description. We can observe from Fig. 2(a) that, Motivated by the above findings, the simplified amplitude
if we change the BPS θ, the phase shifts and amplitudes for F(θ, f ) and phase shift G(θ, f ) of one certain IRS element
other frequencies will be quite different, which illustrates the corresponding to the incident signal of frequency f can be
severe beam deviations due to the dual phase- and amplitude- modeled as
squint. It is worth noting that this kind of dual phase- and
amplitude-squint is an intrinsic phenomenon depending on F(θ, f ) = a1 G2 (θ, f ) + b1 G(θ, f ) + c1 , (3a)
the practical IRS circuit implementation, which cannot be
G(θ, f ) = K(θ)f + B(θ), (3b)
simply ignored in realistic IRS-enhanced wideband systems.
Therefore, it is necessary to consider the phase- and amplitude- K(θ) = a2 sin(b2 θ + c2 ) + b3 sin(b3 θ + c3 ), (3c)
squint into account by developing an accurate reflection model B(θ) = a4 sin(b4 θ + c4 ) + b5 sin(b5 θ + c5 ), (3d)
of each IRS element, which is crucial for the following joint
beamforming and reflecting design. where the functions K(θm ) and B(θm ) denote the slope and
In [42], we have established an accurate three-dimensional intercept for the phase shift-frequency line, respectively. The
amplitude-phase-frequency model to describe the dual phase- central frequency of each subcarrier fi (GHz) is defined as
and amplitude-squint. Unfortunately, this model is so com- fi , fc + (i − N2+1 ) N B
, ∀i ∈ N . Parameters {ai , bi , ci }5i=1
plicated that it may significantly increase the difficulty and are related to specific circuit implementation. For practical
complexity of IRS design. To effectively simplify this model examples showing in Fig. 2(a), the values of these parameters
while maintaining its accuracy, we consider a more practical are given in Table I. Specifically, the fitted results are shown as
wideband situation that the relative bandwidth, i.e. the ratio of dash lines in Figs. 2(b) and 2(c), which illustrate the accuracy
bandwidth and the carrier frequency B/fc , is less than 5%. of the proposed simplified model. In the next section, we
Take the case that the carrier frequency fc = 2.4 GHz and attempt to utilize this practical model in an IRS-enhanced
bandwidth B = 100 MHz as example. It can be observed from wideband MU-MISO-OFDM system and develop an effective
 IRS
BS

G
IDFT & Add CP & RF Chain

Nt
K Precoder
hr
IDFT & Add CP & RF Chain

hd
User1
RF Chain & Del CP & DFT
BS-IRS-User Link
BS-User Link UserK
RF Chain & Del CP & DFT
IRS Control Link
IRS Controller

Fig. 3. The illustration of an IRS-enhanced MU-MISO-OFDM system.

{hed , . . . , h
ed ed Nt
algorithm to jointly design the transmit beamforming and IRS k,0 k,D−1 }, where hk,d ∈ C , d ∈ D , {0, . . . , D−
reflecting. 1}, ∀k ∈ K, is the impulse response corresponding to the d-
th delay tap. Similarly, the wideband channel from the BS
to the IRS is given by {G e d ∈ CM ×Nt ,
e D−1 } with G
e 0, . . . , G
III. S YSTEM M ODEL AND P ROBLEM F ORMULATION
∀d ∈ D. The wideband channel from the IRS to user k is given
A. System Model by {her , . . . , h
er er M
k,0 k,D−1 } with hk,d ∈ C , ∀d ∈ D, ∀k ∈ K.
We consider a wideband MU-MISO-OFDM system with Receiver: After propagating through the wideband channels
N subcarriers, as shown in Fig. 3. The BS employs Nt of both the BS-user link and the BS-IRS-user link, the signal
antennas to communicate with K single-antenna users. This s is corrupted by additive Gaussion white noise (AGWN).
e
wireless transmission is assisted by an IRS of M passive Down-converting to the baseband and removing the CP, we
elements. Denote N = {1, . . . , N }, Nt = {1, . . . , Nt }, obtain the time-domain received signal for user k given as
K = {1, . . . , K}, and M = {1, . . . , M } as the set of the follows
indices of subcarriers, transmit antennas, users, and elements ed +H
e r ΦG)(F H
y
ek = (Hk k
e ⊗ INt )Ws + n
e k , ∀k, (6)
of the IRS, respectively. The phase shifts of IRS elements are
individually adjusted via a controller. In this paper, exact and where the block cyclic channel matrix H e d ∈ CN ×N Nt of the
k
instantaneous channel state information (CSI) is assumed to BS-user link is defined as
be available at the BS, which can be obtained via the efficient 
e d )H
(h 0TNt ... e d )H
(h

k,0 k,1
channel estimation approaches proposed by the recent works  .. .. .. 
[36], [43], [44]. Next, we will describe the communication  . (he d )H . . 
 k,0 
process in detail. 
 ed H
.
.. . . . (h

d (h ) ed )H 
Transmitter: Let si , [s1,i , . . . , sK,i ]T ∈ CK be the trans- He = k,D−1 k,D−1
,
 
k . ..
mit symbols for all users associated with the i-th subcarrier
 T
0Nt (he d
)H . . .

 k,D−1 
with E{si sHi } = IK , ∀i ∈ N . The vector si is first digitally .. .. ..
 
0TNt
 
precoded by a precoder matrix Wi = [w1,i , . . . , wK,i ] ∈  . . . 
T T d H
CNt ×K , ∀i ∈ N , in the frequency domain and then converted 0Nt 0Nt ... (hk,0 )
e
to the time domain by the inverse discrete Fourier transform e H, . . . , G
e H , 0N ×M ,
∀k ∈ K. Similarly, we define [G 0 D−1
(IDFT), which yields the overall time-domain signal e s as H
t
. . . , 0Nt ×M ] as the first block column of the block cyclic
s = (FH ⊗ INt )Ws,
e (4) channel matrix G e ∈ CM N ×N Nt of the BS-IRS link and
[hk,0 , . . . , hk,D−1 , 0M , . . . , 0M ]H as the first block column
e r e r
where F ∈ CN ×N is the normalized discrete Fourier of the block cyclic channel matrix H e r ∈ CN ×N M of the IRS-
k
transform (DFT) matrix and is defined as F(m, n) , user link. The reflection matrix Φ of IRS is defined as Φ =
−j2π(m−1)(n−1)
√1 e N , ∀m, n ∈ N . The overall precoding matrix
N blkdiag(Φ1 , . . . , ΦN ), where Φi , diag(φi,1 , . . . , φi,M ),
W is given by W , blkdiag(W1 , . . . , WN ), and the overall ∀i ∈ N . Here, φi,m denotes the reflection coefficient of the
transmit symbol vector s can be written as s , [sT1 , . . . , sTN ]T . m-th IRS element for the i-th subcarrier. Different from the
After adding the cyclic prefix (CP) of size Ncp , the signal is ideal model that each element exhibits the same reflection
up-converted to the RF domain via Nt RF chains. coefficient for different frequencies (i.e. |φi,m | = 1, and
Channel: In the considered wideband MU-MISO-OFDM ∠φ1,m = . . . = ∠φN,m , ∀i ∈ N , ∀m ∈ M), we adopt the
system, the wideband channel from the BS to user k is mod- practical model presented in the previous section. In particular,
eled by a D-tap (D ≤ Ncp ) finite-duration impulse response the amplitude and phase shift of φi,m actually vary with
 ed +H
Hk =F(H e r ΦG)(F
e H
⊗ INt ) (5a)
k k
(a)
=F(He d Γ1 ΓT + H e r Γ2 ΓT ΦΓ2 ΓT GΓ e 1 ΓT )(FH ⊗ IN )Γ1 ΓT (5b)
k 1 k 2 2 1 t 1
e d Γ1 + H
=F(H e r Γ2 × ΓT ΦΓ2 × ΓT GΓ e 1 ) × ΓT (FH ⊗ IN )Γ1 ×ΓT (5c)
k k
| {z } | {z } | 2 {z } | 2 {z } |1 {z t } 1
 
G
e 1,1 . . . G e 1,N
t
(b)
ed ,...,H e d ] + [H er ,...,H e r ]Φ e × .. .. .. H T
=F([H ) × (INt ⊗ F )Γ1 (5d)

k,1 k,Nt k,1 k,M  . . .
G
e M,1 . . . G e M,N
t
XM XM
d H r H d H er Φ H T
=[FH
e F +F
k,1 H
e
k,m Φm Gm,1 F , . . . , FHk,Nt F + F
e e e Hk,m m Gm,Nt F ]Γ1
e e (5e)
m=1 m=1
(c) XM XM
=[Λdk,1 + Λrk,m Φ
e m Ξm,1 , . . . , Λd +
k,Nt Λrk,m Φe m Ξm,N ]ΓT
t 1 (5f)
m=1 m=1
(d)
=diag[(hdk,1 )H + (hrk,1 )H Φ1 G1 , . . . , (hdk,N )H + (hrk,N )H ΦN GN ], ∀k. (5g)

the BPS θm and follow the relationship given in (3), i.e. on the i-th subcarrier for user k as
|φi,m | = F(θm , fi ), ∠φi,m = G(θm , fi ), ∀i ∈ N , ∀m ∈ M. yk,i =[(hdk,i )H + (hrk,i )H Φi Gi ]Wi si + nk,i
e k ∈ CN (0, σ 2 IN ), ∀k ∈ K, is the AGWN. After DFT, the
n
received signal in the frequency domain can be written as =[(hdk,i )H + (hrk,i )H Φi Gi ]wk,i sk,i + [(hdk,i )H +
K (10)
X
ed +H
e r ΦG)(F H (hrk,i )H Φi Gi ] wp,i sp,i + nk,i , ∀k, ∀i,
yk = F(H k k
e ⊗ INt )Ws + nk ,
(7) p=1,p6=k
= Hk Ws + nk , ∀k,
where nk,i denotes the i-th element of nk .

where nk , Fe nk , ∀k ∈ K. The equivalent frequency-domain

IV. J OINT T RANSMIT B EAMFORMER AND IRS
channel Hk for user k is given by (5) on the top of this page,
R EFLECTION DESIGN
where (a) holds by introducing two column permutation square
matrices Γ1 and Γ2 with Γ1 ΓT1 = IN Nt , Γ2 ΓT2 = IN M , A. Problem Formulation
which convert a block cyclic matrix to several cyclic matrices With the received signal given in (10), the signal-to-
arranged in rows [45]. In this way, the block cyclic channels interference-plus-noise ratio (SINR) on the i-th subcarrier for
are rearranged as a sequence of cyclic matrices. Specifically, user k can be calculated as
e d ∈ CN ×N ,
(b) holds by defining cyclic channel matrices H k,n |[(hdk,i )H + (hrk,i )H Φi Gi ]wk,i |2
Ge m,n ∈ C N ×N
, and He r
∈ C N ×N
, ∀m ∈ M, ∀n ∈ γk,i = P , ∀k, ∀i.
k,m
p6=k |[(hdk,i )H + (hrk,i )H Φi Gi ]wp,i |2 + σ 2
Nt , ∀k ∈ K, as (11)
In this paper, our goal is to jointly design the transmit
e d (:, i) , H
H e d (:, n + (i − 1)Nt ), beamformer W and the BPS matrix Θ , diag(θ1 , . . . , θM ),
k,n k

Gm,n (p, q) , G(m + (p − 1)M, n + (q − 1)Nt ),

e e (8) which essentially control the IRS reflection of wideband
e r (:, i) , H
e r (:, m + (i − 1)M ), ∀i, p, q ∈ N . signals, to maximize the average sum-rate for the MU-MISO-
H k,m k OFDM system, subject to the constraints of the phase shift
matrix and the transmit power constraint. Therefore, the joint
Additionally, the rearranged reflection matrix Φ e transmit beamformer and IRS reflection design problem can
is given by Φ e , blkdiag(Φ1 , . . . , ΦM ), where
e e be formulated as
e m , diag(φ1,m , . . . , φN,m ), ∀m ∈ M. Then (c) holds
Φ N K
since the DFT matrix can diagonalize the cyclic matrix. 1 XX
max log2 (1 + γk,i ) (12a)
W,Θ N
Here Λdk,n , Λrk,m , and Ξm,n are diagonal matrices whose i=1 k=1
diagonal elements are the corresponding eigenvalues of H ed , s.t. |φi,m | = F(θm , fi ), ∀i, m, (12b)
k,n
r
Hk,m , and Gm,n , respectively. Finally, (d) holds by defining
e e ∠φi,m = G(θm , fi ), ∀i, m, (12c)
frequency-domain channels hdk,i ∈ CNt , hrk,i ∈ CM , and θm ∈ [−π, π], ∀m, (12d)
Gi ∈ CM ×Nt , ∀k ∈ K, ∀i ∈ N , as N
X
kWi k2F ≤ P, (12e)
hdk,i (n) , (Λdk,n (i, i))∗ , hrk,i (m) , (Λrk,m (i, i))∗ , i=1
(9)
Gi (m, n) , Ξm,n (i, i), ∀m ∈ M, ∀n ∈ Nt . where P is the total transmit power.
Problem (12) is difficult to solve due to the complex form
Substituting (5g) into (7), we can obtain the received signal of the objective and the non-convex constraints of the BPS
matrix. Additionally, it is worth noting that the amplitude and 2) Auxiliary variable $: When the beamformers Wi , ∀i ∈
phase shift of each IRS element will change with different N , the BPS matrix Θ, and weighting parameters ρk,i , ∀k ∈
frequencies when considering practical IRS responses for K, ∀i ∈ N , are all fixed, the sub-problem with respect to the
wideband signals. In other words, we focus on the design of auxiliary variable $k,i can be formulated as
BPS matrix Θ, but the response of practical IRS for signals
with different subcarriers varies, i.e. reflection matrix Φi , min ρk,i MSEk,i , ∀k, ∀i, (17)
$k,i
∀i ∈ N , are different at each subcarrier. This fact will further
perplex the problem. To deal with these issues, in the next which is an unconstrained convex problem. Thus, problem (17)
section, we attempt to first transform problem (12) into a more can be solved by setting the partial derivative of the objective
tractable multi-variable/block optimization and then iteratively in (17) with respect to $k,i to zero, which yields the optimal
cope with each block. value of $k,i as

?
[(hdk,i )H + (hrk,i )H Φi Gi ]wk,i
B. Problem Reformulation $k,i = PK , ∀k, ∀i.
P p=1 |[(hdk,i )H + (hrk,i )H Φi Gi ]wp,i |2 + σ 2
To tackle the difficulty rising from the log(·) function (18)
and the fractional form of “SINRs” in problem (12), we first 3) Beamformer W: With weighting parameters ρk,i , aux-
reformulate the original sum-rate maximization problem as a iliary variables $k,i , ∀k ∈ K, ∀i ∈ N , and the BPS matrix
modified MSE minimization problem [46]. Let us first define Θ given, the sub-problem with respect to the beamformer
the modified MSE function for user k on the i-th subcarrier Wi , ∀i ∈ N , can be written as
as
∗ ∗ N K K
MSEk,i =E{($k,i yk,i − sk,i )($k,i yk,i − sk,i )∗ } 1 XX X
∗
min ρk,i |$k,i [(hdk,i )H + (hrk,i )H Φi Gi ]wp,i |2
K W N i=1 p=1
X
∗ k=1
= |$k,i [(hdk,i )H + (hrk,i )H Φi Gi ]wp,i |2 
∗
p=1 (13) − 2<{$k,i [(hk,i ) + (hrk,i )H Φi Gi ]wk,i }
d H
(19a)
∗
− 2<{$k,i [(hdk,i )H + (hrk,i )H Φi Gi ]wk,i } N
X
+ |$k,i |2 σ 2 + 1, ∀k, ∀i, s.t. kWi k2F ≤ P. (19b)
i=1
where $k,i ∈ C, ∀k ∈ K, ∀i ∈ N , are auxiliary variables. By
introducing weighting parameters ρk,i ∈ R+ , ∀k ∈ K, ∀i ∈ N , For convenience, we define the equivalent channel hk,i ,

H
∗
problem (12) can be equivalently transformed into the follow- $k,i ((hdk,i )H + (hrk,i )H Φi Gi ) , ∀k ∈ K, ∀i ∈ N . Then,
ing form [46]: problem (19) can be concisely rewritten as
N K N K K
1 XX 1 XXX 
max (log2 ρk,i − ρk,i MSEk,i + 1) (14a) ρp,i |hH 2 H
W,Θ,ρ,$ N i=1 min p,i wk,i | − 2ρk,i <{hk,i wk,i }
k=1 W N i=1 p=1
k=1
s.t. (12b)-(12e), (14b) (20a)
N
where ρ and $ denote the sets of variables ρk,i and $k,i , X
∀k ∈ K, ∀i ∈ N , respectively. Now, the newly formulated s.t. kWi k2F ≤ P. (20b)
i=1
problem (14) is more tractable than the original problem after
removing the complex fractional term (i.e. SINRs) from the Since the objective and constraint of problem (20) are all
log(·) function. In particular, problem (14) is a typical multi- convex, this problem can be optimally solved using the classic
variable/block problem, which can be solved using classical Lagrange multiplier optimization. To be specific, by introduc-
block coordinate descent (BCD) iterative algorithms [47]. In ing a multiplier µ ≥ 0 corresponding to the power constraint
the following subsection, we will decompose problem (14) (20b), problem (20) can be transformed into an unconstrained
into four block optimizations and discuss the solution for each Lagrangian optimization:
block in details. N X
K X
K
X 
min ρp,i |hH 2 H
p,i wk,i | − 2ρk,i <{hk,i wk,i }
C. Block Update W,µ
i=1 k=1 p=1
1) Weighting parameter ρ: Fixing beamformers Wi , ∀i ∈ N
!
X
N , the BPS matrix Θ, and auxiliary variables $k,i , ∀k ∈ +µ kWi k2F −P (21a)
K, ∀i ∈ N , the sub-problem with respect to the weighting i=1
N X
K  K
parameter ρk,i is given by X
H
X
= min wk,i ρp,i hp,i hH
p,i wk,i
max log2 ρk,i − ρk,i MSEk,i , ∀k, ∀i, (15) W,µ
i=1 k=1 p=1
ρk,i 
and the optimal solution can be easily obtained by checking − 2ρk,i <{hH
k,i wk,i }
H
+ µwk,i wk,i − µP. (21b)
the first-order optimality condition of problem (15), i.e.
Similar to the solution of the previous two blocks, this
ρ?k,i = MSE−1
k,i = 1 + γk,i , ∀k, ∀i. (16) unconstrained convex problem can be solved by checking
the first-order optimality condition, which yields the optimal Towards this end, we first split the objective (24b) as
beamforming vector as N
1 X H
(φ Ai φi − 2<{φH
i bi })
N i=1 i
K
X −1
? N M M
wk,i = ρp,i hp,i hH
p,i + µINt ρk,i hk,i , ∀k, ∀i, (22) 1 X X X 
= Ai (m, n)φ∗i,m φi,n − 2<{φ∗i,m bi (m)}
p=1 N i=1 m=1 n=1
N M
1 XXX
where the optimal multiplier µ is associated with the total = (Ai (m, n)φ∗i,m φi,n + Ai (n, m)φ∗i,n φi,m )
N i=1 m=1
power constraint and can be easily determined n6=m
PN using ? 2
a bisec- 
tion search over the set Sµ , {µ ≥ 0 | i=1 kWi kF ≤ P }. + Ai (m, m)|φi,m |2 − 2<{φ∗i,m bi (m)}
4) BPS matrix Θ: Given weighting parameters ρk,i , aux- N M
(a)1 X X  n X  o
iliary variables $k,i , and beamfomers Wi , ∀i ∈ N , ∀k ∈ K, = 2< Ai (m, n)φi,n − bi (m) φ∗i,m
the sub-problem with respect to the BPS matrix Θ can be N i=1 m=1
n6=m
presented as 2

+ Ai (m, m)|φi,m | ,
(26)
N K
1 XX XK where (a) holds since Ai = AH i , ∀i ∈ N . Then, the sub-
∗
min ρk,i |$k,i [(hdk,i )H + (hrk,i )H Φi Gi ]wp,i |2 problem with respect to the m-th BPS element θm while fixing
Θ N i=1
k=1 p=1
 other elements can be formulated as
∗
− 2<{$k,i [(hk,i ) + (hrk,i )H Φi Gi ]wk,i }
d H
(23a) N 
X n X  o
min 2< Ai (m, n)φi,n − bi (m) φ∗i,m
s.t. (12b)-(12d). (23b) θm
i=1 n6=m

+ Ai (m, m)|θi,m |2 (27a)
T
By defining φi , [φi,1 , . . . , φi,M ] , k,p,i , hd (hdk,i )H wp,i , s.t. (12b)-(12d). (27b)
and vk,p,i , [(hrk,i )H diag(Gi wp,i )]H , ∀k, p ∈ K, ∀i ∈ N , P
problem (23) can be concisely rearranged as We further define χi,m , n6=m Ai (m, n)φi,n − bi (m), ∀i ∈
N , ∀m ∈ M, and substitute the constraints (12b), (12c) into
the objective (27a). Then, sub-problem (27) can be reformu-
N K K lated as
1 XX X
∗ H
min ρk,i |$k,i (hd k,p,i + vk,p,i φi )|2 N
Θ N X
i=1 k=1 p=1 min 2|χi,m |F(θm , fi ) cos(∠χi,m − G(θm , fi ))
 θm
∗ H
− 2<{$k,i (hd k,k,i + vk,k,i φi )} (24a) i=1
+ Ai (m, m)F2 (θm , fi )


(28a)
N
1 X H s.t. θm ∈ [−π, π]. (28b)
= min (φ Ai φi − 2<{φH
i bi }), (24b)
Θ N i=1 i
The objective of problem (28) is a summation of N com-
s.t. (12b)-(12d), (24c) plicated functions involving both trigonometric and quadratic
terms, which is difficult to deal with. The computational com-
plexity will be quite high when the numbers of IRS elements
where and/or subcarriers become large, which is the case for practical
communication systems. To reduce the calculation complexity,
K
X K
X we propose to further divide the whole bandwidth into Ns
Ai , ρk,i |$k,i |2 H
vk,p,i vk,p,i , ∀i, (25a) sub-bands, each of which comprises S , N/Ns subcarriers.
k=1 p=1 By approximating each sub-band as a “narrowband” channel
K
X  K
X  which has identical reflection coefficient configuration, prob-
bi , ρk,i $k,i vk,k,i − |$k,i |2 vk,p,i hd k,p,i , ∀i. lem (28) can be further simplified as the optimization of a
k=1 p=1 summation of much smaller number of functions, i.e.
(25b)
min g(θm )
θm
s.t. θm ∈ [−π, π], (29a)
Problem (24) is still difficult to solve since the BPS matrix
Θ to be optimized is embedded into a summation of N where the objective g(θm ) is defined as
complicated functions. To simplify the design, one feasible Ns
X
solution is to decompose the joint optimization of the entire g(θm ) = 2|χi,m |F(θm , fs,i ) cos(∠χi,m − G(θm , fs,i ))
matrix Θ into sub-problems, each of which deals with only i=1
one entry of Θ while fixing others. This alternative update of + αi,m F2 (θm , fs,i ) ,


Θ is conducted iteratively until the objective value converges. (30)
 PS
with fs,i , fc + (i − Ns2+1 ) NBs , χi,m , S1 j=1 χ(i−1)S+j,m , 10-3
S
and αi,m , S1 j=1 A(i−1)S+j (m, m), ∀i = 1, . . . , Ns . 5
P
Unfortunately, the above problem is still difficult to solve
since we cannot easily calculate the derivative of the objective

)
Phase 1

m
0

g(
and obtain the close-form solution. To tackle this difficulty,
Phase 2
we first try to explore the characteristic of the objective
(30) with the aid of numerical experiments. After numerous
-5
simulations (more than 5000 times), we find that objective -3 -2 -1 0 1 2 3
(30) has only one minimum point within the range [−π, π]. Phase Range

More concretely, objective (30) behaves like a kind of smooth (a)

double-peak-trough curve, whose minimum is achieved either
10 -3
at the minimum point or at two border points. Some of
examples are shown in Fig. 4. Motivated by this finding, 3
2
we propose a three-phase one-dimensional search method to
1
efficiently find optimal solutions, which is summarized as

g( m )
Phase 1
0
follows:
-1
Phase 1: Narrow the search range by a success-failure -2
Phase 2
method: Initialize a starting point θ0 as well as a -3
step size h > 0. If g(θ0 + h) < g(θ0 ), enlarge the -3 -2 -1 0 1 2 3
step size and search forward until the objective rises; Phase Range

otherwise, search reversely until the objective rises. (b)

Phase 2: Find the minimum point θ̄ by a golden section
10 -3
method: Successively section the search range which
includes the minimum point in the golden ratio until 1
reaching a predefined threshold.
0 Phase 1
g( m )

Phase 3: Determine the minimum value: Compare the values

of g(θ), g(−π), and g(π) to determine the minimal -1
Phase 2
value as well as its corresponding phase shift.
-2
The detail of the three-phase search algorithm is summarized
in Algorithm 1. Furthermore, red points marked in Fig. 4 are -3 -2 -1 0 1 2 3
search results, which illustrate the accuracy of the proposed Phase Range

algorithm. (c)
In realistic applications, the IRS is usually realized by finite-
or even low-resolution phase shifters to effectively reduce the Fig. 4. Examples of the objective (30) as a function of the BPS within the
range [−π, π].
hardware consumption. Therefore, we also consider the case
that the BPS θm for IRS has discrete phases controlled by b
bits, which are uniformly spaced within the range [−π, π), i.e. 5) Summary: Having approaches to solve the above four
2π sub-problems with respect to ρk,i , $k,i , wk,i , ∀k ∈ K, ∀i ∈
θm ∈ F , { i − π|i = 0, 1, . . . , 2b }, ∀m. (31) N , and Θ, the overall procedure for the joint beamformer and
2b
IRS design is finally straightforward. Given appropriate initial
In this case, the IRS design sub-problem is given by
values of wk,i , ∀k ∈ K, ∀i ∈ N , and Θ, we iteratively update
N
X the above four blocks alternatively order until convergence.
min 2|χi,m |F(θm , fi ) cos(∠χi,m − G(θm , fi )) The proposed joint beamformer and IRS design algorithm is
θm
i=1 therefore summarized in Algorithm 2.
+ Ai (m, m)F2 (θm , fi )


(32a)
s.t. θm ∈ F. (32b) D. Complexity Analysis
Similarly, we simplify this problem by dividing the whole In this subsection, we provide an analysis of the com-
bandwidth into several sub-bands, which yields the following plexity for the proposed joint beamformer and IRS design
problem: algorithm. In each iteration, updating the weighting param-
eter ρ has a complexity of O(N K 2 Nt M 2 ) approximately;
min g(θm ) (33a) updating the auxiliary variable $ requires O(N K(K +
θm
1)Nt M 2 ) operations; updating beamformer W requires about
s.t. θm ∈ F. (33b)
O(I1 N Nt K(3M 2 + Nt2 )) operations, where the parameter I1
Thanks to the employment of low-resolution phase shifters, denotes the iterations of bisection search. Finally, the order
(i.e. b ≤ 3 bit) to realize the IRS, it is possible to perform of complexity for updating BPS matrix Θ for continuous
a one-dimensional quick exhaustive search over the set F to phases is about O((5M Nt + M 3 )N K 2 + I2 Ns M (I3 + I4 )),
?
find the optimal BPS element θm . where I3 and I4 denotes the iterations for success-failure
Algorithm 1 Three-Phase One-Dimensional Search method and golden section method, respectively, and that for
Input: fs,i , χi,m , αi,m , ∀i ∈ N . discrete phases is O((5M Nt +M 3 )N K 2 +I5 Ns M 2b ), where
?
Output: θm . parameters I2 and I5 denote the numbers of iterations for
1: Phase 1: Success-failure method calculating Θ. Therefore, the total complexity of the proposed
2: Initialize θ0 , h > 0, θ1 = θ0 , θ2 = θ1 + h. algorithm is given by
3: if g(θ2 ) < g(θ1 ) then
4: θ3 = θ2 + h. Cc =O(Ic (N K 2 Nt M 2 + N K(K + 1)Nt M 2
5: if g(θ2 ) ≤ g(θ3 ) then + I1 N Nt K(3M 2 + Nt2 ) + (5M Nt + M 3 )N K 2
6: Obtain the narrowed range [θl , θr ] as θl = + I2 Ns M (I3 + I4 ))) (34a)
min{θ1 , θ3 }, θr = max{θ1 , θ3 }, and stop. (a) 2 3 2
7: else ≈O(Ic (N K M + 3I1 N Nt KM + I2 Ns M (I3 + I4 )))
8: h = 2h, θ1 = θ2 , θ2 = θ3 , θ3 = θ2 + h. (34b)
9: Goto step 5. Cd =O(Ic (N K 2 Nt M 2 + N K(K + 1)Nt M 2
10: end if + I1 N Nt K(3M 2 + Nt2 ) + (5M Nt + M 3 )N K 2
11: else
+ I5 Ns M 2b )) (34c)
12: h = −h, θ3 = θ1 , θ1 = θ2 , θ2 = θ3 , θ3 = θ2 + h.
(a)
13: Goto step 5. ≈O(Id (N K 2 M 3 + 3I1 N Nt KM 2 + I2 Ns M I5 Ns M 2b )),
14: end if (34d)
15: Phase 2: Golden section method
16: Set θ l = θl + 0.382(θr − θl ), θ r = θl + 0.618(θr − θl ), .
where (a) holds under assumptions M  Nt , M  K.
17: while θr − θl >  do
Parameters Ic (for continuous phases) and Id (for discrete
18: if g(θl ) ≤ g(θr ) then phases) are the numbers of iterations for Algorithm 2. Sim-
19: θr = θr , θr = θl , θl = θl + 0.382(θr − θl ). ulation results in the next section show that, under different
20: else settings, the proposed algorithm for both continuous and dis-
21: θl = θl , θl = θr , θr = θl + 0.618(θr − θl ). crete scenarios can converge within limited iterations, which
22: end if demonstrates the efficiency of the proposed algorithm.
23: end while
? V. S IMULATION R ESULTS
24: Obtain θm = (θl + θr )/2.
25: Phase 3: Determine θm ? A. Simulation Settings
26: if g(π) ≤ g(θm ) and g(π) ≤ g(−π) then In this section, we present simulation results to demon-
?
27: θm = π. strate the performance of the IRS-enhanced wideband MU-
28: else if g(−π) ≤ g(θm ) and g(−π) ≤ g(π) then MISO-OFDM system by showing the average sum-rate of the
?
29: θm = −π. proposed joint beamformer and IRS design. In the consid-
30: end if ered IRS-enhanced MU-MISO-OFDM system, we assume the
?
31: Return θm . number of subcarriers is N = 64. The number of taps is
set as D = 16 with half non-zero taps modeled as circularly
Algorithm 2 Joint Transmit Beamformer and IRS Reflection symmetric complex Gaussian (CSCG) random values. The CP
Design length is set to be Ncp = 16. The carrier frequency and
bandwidth is given by fc = 2.4GHz and B = 100MHz,
Input: hdk,i , hrk,i , Gi , ∀k ∈ K, ∀i ∈ N , P , B.
? respectively. The signal attenuation is set as ζ0 = 30 dB at a
Output: wk,i , ∀k ∈ K, ∀i ∈ N , Θ? .
reference distance 1 m for all channels. The path loss exponent
1: Initialize wk,i , ∀k ∈ K, ∀i ∈ N , Θ.
of the BS-IRS channel, the IRS-user channel, and the BS-user
2: while no convergence of objective (14a) do
channel is set as εBI = 2.8, εIU = 2.5, and εBU = 3.7,
3: Update ρk,i , ∀k ∈ K, ∀i ∈ N by (16).
respectively. The noise power at each user is set as σ 2 = −70
4: Update $k,i , ∀k ∈ K, ∀i ∈ N by (18).
dBm.
5: Update wk,i , ∀k ∈ K, ∀i ∈ N by (22). In the following simulation results, we assume a three
6: while no convergence of Θ do dimensional (3D) coordinate system is considered as shown
7: for m = 1 : M do in Fig. 5, where a uniform linear array (ULA) with antenna
8: Update θm by Algorithm 1 for continuous phases spacing dA = 0.3 m at the BS and a uniform planar array
or by an exhaustive search for low-resolution (UPA) with element-spacing dI = 0.03 m at the IRS and are
phases. located in y-z plane and x-y plane, respectively. The distance
9: end for between the reference antenna of the BS and the reference
10: end while element of the IRS is given by dBI . K users are randomly
11: end while
? located in x-z plane with the same distance dIU = 3 m as
12: Return wk,i , ∀k ∈ K, ∀i ∈ N , Θ? . well as random phase ϕk between the reference element of
the IRS and the k-th user. Based on the relative position given
in Fig. 5, the distances between the (p, q)-th IRS element and
the k-th user dp,q,k n,k
IU , the n-th antenna and the k-th user dBU , as
 z dBI
IRS
o dI x dA o φ o x
dIU
dI

BS dBI
IRS dBU
BS
User
y
y
(a) (b) z (c)

Fig. 5. An illustration of the relative position among the BS, IRS, and users.

well as the n-th antenna and the (p, q)-th IRS element dn,p,qBI ,
0.9 2.4

are given by
0.85 2.2
q
dp,q,k
IU = (pdI − dIU cos ϕk )2 + q 2 d2I + d2IU sin2 ϕk , 0.8 2
q

Average sum-rate

Average sum-rate
dn,k
BU = (dBI − dIU sin ϕk )2 + n2 d2A + d2IU cos2 ϕk , 0.75 1.8
q (35)
n,p,q 2 2 2 2
dBI = (qdI − ndA ) + p dI + dBI , 0.7 1.6
√
∀n ∈ Nt , ∀p, q = 1, . . . , M , ∀k ∈ K. 0.65
b=
b=1 b=
1.4
b=2 b=1
Then the fading component for the BS-IRS link, the BS-User b=2

0.6 1.2
link, and the IRS-User link is given by
q q 0.55
n,p,q n,p,q −εBI 1
ξBI = ζ0 (dBI ) , ξBU = ζ0 (dn,k
n,k
BU )
−εBU , 10 20 30
Number of iterations
40 50 10 20 30 40 50
Number of iterations
q (36)
p,q,k
ξIU = ζ0 (dp,q,k
IU ) −εIU , ∀n, ∀p, q, ∀k. (a) Nt = 6, M = 64 (b) Nt = 8, M = 256

Thus, the channels for three links are given by Fig. 6. Average sum-rate versus the number of iterations (K = 3, N = 64,
Ns = 8, P = −5 dB).
b r (m) = ξ p,q,k hr (m),
h b d (n) = ξ n,k hd (n),
h
k,i IU k,i k,i BU k,i
b i (m, n) = ξ n,p,q Gi (m, n),
G (37) 1
BI
√
∀n, ∀p, q, ∀k, ∀i ∈ N , ∀m = (p − 1) M + q. 0.95

0.9
B. System Performance
Average sum-rate

We start with presenting the convergence of the proposed 0.85

joint beamformer and IRS design by plotting the average sum- b=

0.8 LowRes
rate versus the number of iterations in Fig. 6. Simulation
results illustrate that the proposed algorithm can converge 0.75
within 30 iterations when using continuous phase shifters and
within 20 iterations when using low-resolution phase shifters 0.7

to realize the IRS. When the numbers of antennas and IRS

0.65
elements increase, the proposed algorithm can still converge
within limited iterations. Next in Fig. 7, we plot the average 0.6
sum-rate as a function of the resolution b (LowRes) of each 1 2 3 4 5 6 7 8
Number of iterations
IRS element. Fig. 7 shows that b = 4 is a sufficiently precise
resolution level and the performance improvement is marginal
Fig. 7. Average sum-rate versus the number of iterations (Nt = 6, M = 64,
when b is larger than 4. Moreover, considering the both results K = 3, N = 64, Ns = 8, P = −5 dB).
of the convergence speed as illustrated in Fig. 6 and the
influence of resolution b as shown in Fig. 7, it is more practical
and efficient to realize the IRS using low-resolution phase Fig. 8 shows the average sum-rate among all subcarriers
shifters in realistic systems. versus the transmit power P with the proposed algorithm for
 2.5 2.2
w/ IRS, Proposed, b = w/ IRS, Proposed, b =
w/ IRS, Ideal, b = 2 w/ IRS, Ideal, b =
w/ IRS, Proposed, b = 2 w/ IRS, Proposed, b = 2
2 w/ IRS, Ideal, b = 2 1.8 w/ IRS, Ideal, b = 2
w/ IRS, Proposed b = 1 w/ IRS, Proposed b = 1
w/ IRS, Ideal, b = 1 w/ IRS, Ideal, b = 1
1.6
w/ IRS, Random, b = w/ IRS, Random, b =
Average sum-rate

Average sum-rate
w/o IRS w/o IRS
1.5 1.4

1.2

1 1

0.8

0.5 0.6

0.4

0 0.2
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 50 100 150 200 250
P(dBW) Number of IRS elements
(a) Nt = 6, M = 64
Fig. 9. Average sum-rate versus the number of IRS elements M (Nt = 8,
4.5 K = 3, N = 64, Ns = 8, P = −5 dB).
w/ IRS, Proposed, b =
4 w/ IRS, Ideal, b =
w/ IRS, Proposed, b = 2
w/ IRS, Ideal, b = 2 1.5
3.5 w/ IRS, Proposed, b = 1
w/ IRS, Ideal, b = 1
3 w/ IRS, Random, b =
Average sum-rate

w/o IRS
2.5
1
Average sum-rate

1.5

1 w/ IRS, Proposed, b =
0.5 w/ IRS, Ideal, b =
0.5 w/ IRS, Proposed, b = 2
w/ IRS, Ideal, b = 2
w/ IRS, Proposed b = 1
0
w/ IRS, Ideal, b = 1
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0
w/ IRS, Random, b =
P(dBW) w/o IRS
0
(b) Nt = 8, M = 256 4 6 8 10 12 14 16
Number of transmit antennas
Fig. 8. Average sum-rate versus transmit power P (K = 3, N = 64,
Ns = 8).
Fig. 10. Average sum-rate versus the number of transmit antennas Nt (M =
64, N = 64, Ns = 8, K = 3, P = −5 dB).

the cases of using continuous and low-resolution (i.e. b =

1, 2-bit) phase shifters with different settings (e.g. number of lower bounds for all transmit power ranges, which illustrates
antennas and/or IRS elements). For fair comparison, we also the advantages of employing IRS in wireless communications.
plot the average sum-rate for the following schemes: Moreover, the proposed algorithm also outperform the “w/
• The average sum-rate designed by our proposed simpli- IRS, Ideal” scheme, which demonstrates the importance of
fied IRS model in this paper and testified by the same precisely modeling the reflection characteristics of the practi-
IRS model, which is marked as “w/ IRS, Proposed”. cal IRS.
• The average sum-rate designed by the ideal IRS model To illustrate the advantage of employing IRS in enhancing
in [38] but testified by the proposed IRS model, which is wideband wireless communications, in Fig. 9 we plot the
marked as “w/ IRS, Ideal”. average sum-rate versus different numbers of IRS elements
• Lower bound I: The system with an IRS whose BPSs are M . A similar conclusion can be drawn from Fig. 9 that the
randomly selected within the range [−π, π] and calculated proposed algorithm can always achieve better performance
by the proposed IRS model, which is marked as “w/ IRS, compared with its competitors. Moreover, with the number
Random”. of IRS elements growing, the performance gap between the
• Lower bound II: The system with direct link only, which “w/ IRS” scheme and the “w/o IRS” one is becoming larger.
is marked as “w/o, IRS”. Finally, the average sum-rate as a function of the number of
It can be observed from Fig. 8 that the proposed algorithm can transmit antennas is illustrated in Fig. 10. A similar conclusion
achieve significantly better performance compared with two can be obtained from the above simulation results. More
importantly, the performance gap between the proposed “w/ [13] X. Yu, D. Xu, and R. Scholar, “MISO wireless communication systems
IRS, Proposed” scheme and the “w/ IRS, Ideal” scheme via intelligent reflecting surface,” in Proc. IEEE Int. Conf. Commun.
China (ICCC), Changchun, China, Dec. 2019.
becomes smaller with the increasing number of the transmit [14] Q. Wu and R. Zhang, “Intelligent reflecting surface enhanced wireless
antennas. This trend can be explained as follows: when the network via joint active and passive beamforming,” IEEE Trans. Wireless
number of transmit antennas grows, the channel gain achieved Commun., vol. 18, no. 11, pp. 5394-5409, Nov. 2019.
[15] Y. Han, W. Tang, S. Jin, C. Wen, and X. Ma, “Large intelligent surface-
by the direct link will gradually dominate the effective gain assisted wireless communication exploiting statistical CSI,” IEEE Trans.
of the entire channels, which, to some extent, weakens the Veh. Technol., vol. 68, no. 8, pp. 8238-8242, Aug. 2019.
influence of the IRS. [16] C. Huang, A. Zappone, G. C. Alexandropoulos, M. Debbah, and
C. Yuen, “Reconfigurable intelligent surfaces for energy efficiency in
VI. C ONCLUSIONS wireless communication,” IEEE Trans. Wireless Commun., to appear.
[17] C. Huang, R. Mo, and C. Yuen, “Reconfigurable intelligent surface as-
In this paper, we first simplified the practical IRS model sisted multiuser MISO systems exploiting deep reinforcement learning,”
and validated the accuracy of the proposed model based on IEEE J. Sel. Area Commun. (JSAC), to appear.
[18] H. Guo, Y.-C. Liang, J. Chen, and E. G. Larsson, “Weighted sum-rate op-
numerical simulations. With the simplified practical model, timization for intelligent reflecting surface enhanced wireless networks,”
we considered the problem of joint beamformer and IRS in Proc. IEEE Global Commun. Conf. (GLOBECOM), Waikoloa, HI,
design with both continuous and low-resolution phase shifters Dec. 2019.
[19] J. Zhao, “Optimizations with intelligent reflecting surfaces (IRSs) in
to maximize the average sum-rate of a wideband MU-MISO- 6G wireless networks: Power control, quality of service, max-min fair
OFDM system. We proposed a sub-optimal iterative algorithm beamforming for unicast, broadcast, and multicast with multi-antenna
with the aid of the equivalence between sum-rate maximization mobile users and multiple IRSs,” Aug. 2019. [Online]. Avaliable:
https://arXiv.org/abs/1908.03965
and MSE minimization. Simulation results demonstrated the
[20] Y. Liu, J. Zhao, M. Li, and Q. Wu, “Intelligent reflecting surface aided
significance of modeling the imperfect response characteristics MISO uplink communication network: Feasibility and SINR optimiza-
of IRS reflecting elements and its associated configuration tion,” July 2020. [Online]. Avaliable: https://arXiv.org/abs/2007.01482
design. With the tremendous difference between the ideal [21] B. Di, H. Zhang, L. Song, Y. Li, Z. Han, and H. V. Poor, “Hybrid
beamforming for reconfigurable intelligent surface based multi-user
reflection model and the practical reflection model, there are communications: Achievable rates with limited discrete phase shifts,”
many issues worthy to be studied and investigated, such as IEEE J. Sel. Area Commun., to appear.
IRS deployment, resource allocation, user scheduling, etc. [22] Q. Wu and R. Zhang, “Beamforming optimization for wireless network
aided by intelligent reflecting surface with discrete phase shifts,” IEEE
Trans. Commun., vol. 68, no. 3, pp. 1838-1851, Mar. 2020.
R EFERENCES [23] J. Xu, W. Xu, and A. L. Swindlehurst, “Discrete phase shift design for
[1] A. L. Swindlehurst, E. Ayanoglu, P. Heydari, and F. Capolino, practical large intelligent surface communication,” in Proc. IEEE Pacific
“Millimeter-wave massive MIMO: The next wireless revolution?” IEEE Rim Conf. on Commun., Computers and Signal Process. (PACRIM),
Commun. Mag., vol. 52, no. 9, pp. 56-62, Sept. 2014. Victoria, Canada, Aug. 2019.
[2] S. Zhang, Q. Wu, S. Xu, and G. Y. Li, “Fundamental green tradeoffs: [24] J. He, K. Yu, and Y. Shi, “Coordinated passive beamforming for
Progress, challenges, and impacts on 5G networks,” IEEE Commun. distributed intelligent reflecting surfaces network”, Feb. 2020. [Online].
Surveys & Tutorials, vol. 19, no. 1, pp. 33-56, First Quarter 2017. Avaliable: https://arxiv.org/abs/2002.05915
[3] Q. Wu, G. Y. Li, W. Chen, D. W. K. Ng, and R. Schober, “An overview [25] Z. Li, M. Hua, Q. Wang, and Q. Song, “Weighted sum-rate maximization
of sustainable green 5G networks,” IEEE Wireless Commun., vol. 24, for multi-IRS aided cooperative transmission”, IEEE Wireless Commun.
no. 4, pp. 72-80, Aug. 2017. Lett., to appear.
[4] Q. Wu and R. Zhang, “Towards smart and reconfigurable environment: [26] D-W. Yue, H. H. Nguyen, and Y. Sun “MmWave doubly-massive-MIMO
Intelligent reflecting surface aided wireless network,” IEEE Commun. communications enhanced with an intelligent reflecting surface”, Mar.
Mag., vol. 58, no. 1, pp. 106-112, Jan. 2020. 2020. [Online]. Avaliable: https://arxiv.org/abs/2003.00282
[5] C. Liaskos, S. Nie, A. Tsioliaridou, A. Pitsillides, S. Ioannidis, and I. [27] M. Cui, G. Zhang, and R. Zhang, “Secure wireless communication via
Akyildiz, “A new wireless communication paradigm through software- intelligent reflecting surface,” IEEE Wireless Commun. Lett., vol. 8, no.
controlled metasurfaces,” IEEE Commun. Mag., vol. 56, no. 9, pp. 162- 5, pp. 1410-1414, Oct. 2019.
169, Sep. 2018. [28] X. Guan, Q. Wu, and R. Zhang, “Intelligent reflecting surface assisted
[6] M. Di Renzo, A. Zappone, M. Debbah, M.-S. Alouini, C. Yuen, secrecy communication: Is artificial noise helpful or not?” IEEE Wireless
J. de Rosny, and S. Tretyakov, “Smart radio environments em- Commun. Lett., to appear.
powered by reconfigurable intelligent surfaces: How it works, state
[29] D. Xu, X. Yu, Y. Sun, D. W. K. Ng, and R. Schober, “Resource allocation
of research, and road ahead,” June 2020. [Online]. Available:
for secure IRS-assisted multiuser MISO systems,” in Proc. IEEE Global
https://arxiv.org/abs/2004.09352
Commun. Conf. (GLOBECOM), Waikoloa, HI, Dec. 2019.
[7] E. Bjornson, Ö. Özdogan, and E. G. Larsson, “Intelligent reflecting
[30] L. Zhang, C. Pan, Y. Wang, H. Ren, K. Wang, and A. Nallanathan,
surface vs. decode-and-forward: How large surfaces are needed to beat
“Robust beamforming design for intelligent reflecting surface aided
relaying?” IEEE Wireless Commun. Lett., vol. 9, no. 2, pp. 244-248,
cognitive radio systems with imperfect cascaded CSI,” Apr. 2020.
Feb. 2019.
[Online]. Avaliable: https://arxiv.org/abs/2004.04095
[8] S. Gong, et al. “Towards smart radio environment for wireless com-
munications via intelligent reflecting surface: A comprehensive survey,” [31] D. Xu, X. Yu, and R. Schober, “Resource allocation for intelligent
IEEE Commun. Surveys & Tutorials, to appear. reflecting surface-assisted cognitive radio networks,” Jan. 2020. [Online].
[9] K-K. Wong, K-F. Tong, Z. Chu, and Y. Zhang, “A vision to smart radio Avaliable: https://arxiv.org/abs/2001.11729
environment: Surface wave communication superhighways,” May 2020. [32] X. Guan, Q. Wu, and R. Zhang, “Joint power control and passive
[Online]. Available: https://arxiv.org/abs/2005.14082 beamforming in IRS-assisted spectrum sharing,” IEEE Commun. Lett.,
[10] Q. Wu, S. Zhang, B. Zheng, C. You, and R. Zhang, “Intelligent reflecting to appear.
surface aided wireless communications: A tutorial,” July 2020. [Online]. [33] A. Khaleel and E. Basar, “Reconfigurable intelligent surface-
Available: https://arxiv.org/abs/2007.02759 empowered MIMO systems,” Apr. 2020. [Online]. Avaliable:
[11] C. Huang, S. Hu, G. C. Alexandropoulos, A. Zappone, C. Yuen, R. https://arxiv.org/abs/2004.02238
Zhang, M. D. Renzo, M. Debbah, “Holographic MIMO surfaces for 6G [34] E. Basar, “Reconfigurable intelligent surface-based index modulation:
wireless networks: Opportunities, challenges, and trends,” IEEE Wireless A new beyond MIMO paradigim for 6G,” IEEE Trans. Commun., to
Communications Magazine, to appear. appear.
[12] Y. Liu, X. Liu, X. Mu, T. Hou, J. Xu, Z. Qin, M. D. Renzo, and N. [35] Y. Yang, B. Zheng, S. Zhang, and R. Zhang, “Intelligent reflecting
Al-Dhahir, “Reconfigurable intelligent surfaces: Principles and opportu- surface meets OFDM: Protocal design and rate maximization,” Nov.
nities,” July 2020. [Online]. Available: https://arxiv.org/abs/2007.03435 2019. [Online]. Available: https://arxiv.org/abs/1906.09956
[36] B. Zheng and R. Zhang, “Intelligent reflecting surface-enhanced OFDM:
Channel estimation and reflection optimization,” IEEE Wireless Com-
mun. Lett., vol. 9, no. 4, pp. 518-522, Apr. 2020.
[37] T. Bai, C. Pan, H. Ren, Y. Deng, M. Elkashlan, and A. Nallanathan,
“Resource allocation for intelligent reflecting surface aided wireless
powered mobile edge computing in OFDM systems,” Mar. 2020. [On-
line]. Available: https://arxiv.org/abs/2003.05511
[38] H. Li, R. Liu, M. Li, Q. Liu, and X. Li, “IRS-enhanced widebadn
MU-MISO-OFDM communication systems,” in Proc. IEEE Wireless
Commun. Networking Conf. (WCNC), Seoul, South Korea, May 2020.
[39] H. Rajagopalan and Y. Rahmat-Samii, “Loss quantification for mi-
crostrip reflectarray: Issue of high fields and currents,” in Proc. IEEE
Antennas and Propag. Society Int. Symposium, San Diego, CA, July
2008.
[40] W. Tang et al., “MIMO transmission through reconfigurable intelligent
surface: System design, analysis, and implementation,” Dec. 2019,
[Online]. Available: https://arxiv.org/abs/1912.09955
[41] S. Abeywickrama, R. Zhang, Q. Wu, and C. Yuen, “Intelligent reflecting
surface: Practical phase shift model and beamforming optimization,”
IEEE Trans. Commun., to appear.
[42] W. Cai, H. Li, M. Li, and Q. Liu, “Practical modeling and beamforming
for intelligent reflecting surface aided wideband systems,” IEEE Com-
mun. Lett., vol. 24, no. 7, pp. 1568-1571, July 2020.
[43] A. Taha, M. Alrabeiah, and A. Alkhateeb, “Enabling large intelligent
surfaces with compressive sensing and deep learning,” Apr. 2019.
[Online]. Avaliable: https://arxiv.org/abs/1904.10136
[44] B. Zheng, C. You, and R. Zhang, “Intelligent reflecting surface assisted
multi-user OFDMA: Channel estimation and training design,” Mar.
2020. [Online]. Avaliable: https://arxiv.org/abs/2003.00648
[45] Y. Kwon, J. Chung, and Y. Sung, “Hybrid beamformer design for
mmWave wideband multi-user MIMO-OFDM systems,” in Proc. IEEE
Int. Workshop on Signal Process. Advances in Wireless Commun.
(SPAWC), Sapporo, Japan, July 2017.
[46] Q. Shi, M. Razaviyayn, Z. Q. Luo, and C. He, “An iteratively weighted
MMSE approach to distributed sum-utility maximization for a MIMO
interfering broadcast channel,” IEEE Trans. Signal Process., vol. 59, no.
9, pp. 4331-4340, Sept. 2011.
[47] D. Bertsekas, Nonlinear Programming, 2nd ed. Belmont, MA, USA:
Athena Scientific, 1999.