1

Intelligent Reflecting Surface: Practical Phase Shift

Model and Beamforming Optimization
Samith Abeywickrama, Rui Zhang, Fellow, IEEE, and Chau Yuen, Senior Member, IEEE

Abstract—Intelligent reflecting surface (IRS) that enables the leading to substantial energy loss and hence low reflection
control of the wireless propagation environment has been looked amplitude [5]. Furthermore, these losses mainly come from the
arXiv:1907.06002v1 [eess.SP] 13 Jul 2019

upon as a promising technology for boosting the spectrum and semiconductor devices, metals, and dielectric substrates used
energy efficiency in future wireless communication systems. Prior
works on IRS are mainly based on an ideal phase shift model in the IRS, and thus are not avoidable in practice. In fact, this is
with full reflection regardless of phase shift, which, however, is a long standing problem for reflection-based metasurfaces [6].
practically difficult to realize. In contrast, we propose in this letter In [7], amplifiers are integrated into the reflecting elements to
a practical phase shift model that captures the phase dependent compensate the energy loss, which is not suitable for passive
amplitude in the reflection coefficient. Applying this new model IRS and also practically costly.
to an IRS-aided wireless system, we formulate a problem to
maximize its achievable rate by jointly optimizing the transmit In [1]–[3], by assuming the ideal phase shift model, IRS
beamforming and the IRS reflect beamforming. The formulated reflection is designed to have the maximum phase alignment
problem is non-convex and difficult to be optimally solved in between the IRS-reflected and non-IRS-reflected signals at the
general, for which we propose a low-complexity suboptimal designated receivers. In contrast, when the amplitude depends
solution based on the alternating optimization (AO) technique. on the phase shift at each reflecting element, such an optimal
Simulation results unveil a substantial performance gain achieved
by the joint beamforming optimization based on the proposed reflection design is not feasible as each phase shift needs
phase shift model as compared to the conventional ideal model. to be properly chosen to have a better balance between the
amplitude and phase alignment. Therefore, if the IRS reflection
Index Terms—Intelligent reflecting surface, passive array,
beamforming optimization, phase shift model. is designed for a practical system based on the ideal phase shift
model, it inevitably causes certain performance degradation.
To the best of the authors’ knowledge, the practical phase shift
I. I NTRODUCTION model and corresponding beamforming optimization algorithm
Intelligent reflecting surface (IRS) assisted wireless com- design for IRS-aided wireless systems has not been reported
munication has recently emerged as a promising solution to in the literature yet.
enhance the capacity and energy efficiency for future wireless This thus motivates this letter, where we first propose a
systems. Specifically, an IRS is able to establish favourable practical phase shift model and verify its accuracy with the
channel responses by controlling the wireless propagation experimental results reported in literature. Next, based on this
environment through its reconfigurable passive reflecting el- model and considering an IRS-aided point-to-point communi-
ements (see e.g. [1]–[4] and the references therein). However, cation system, we formulate a new problem to maximize its
the existing works on IRS mostly assume an ideal phase shift achievable rate by jointly optimizing the transmit beamforming
model with full reflection, i.e., unity amplitude at each reflec- and the IRS reflect beamforming. As this problem is non-
tion element regardless of the phase shift, which, however, is convex, we propose a low-complexity algorithm to solve it
practically difficult to realize due to the hardware limitation. sub-optimally by leveraging the alternating optimization (AO)
The amplitude response of a typical passive reflecting technique. Simulation results are also presented to demonstrate
element is non-uniform with respect to its phase shift. In the performance gain by the joint beamforming optimization
particular, the amplitude exhibits its minimum value at the zero based on the proposed practical phase shift model over the
phase shift, but monotonically increases and asymptotically conventional ideal model.
approaches unity amplitude at the phase shift of π or −π. Notations: In this letter, scalars are denoted by italic letters,
This is due to the fact that when the phase shift approaches vectors and matrices are denoted by bold-face lower-case and
zero, the image currents, i.e., the currents of a virtual source upper-case letters, respectively. For a complex-valued vector v,
that accounts for the reflection, are in-phase with the reflecting kvk, vH , and diag(v) denote its ℓ2 -norm, conjugate transpose,
element currents, and thus the electric field and the current and a diagonal matrix with each diagonal element being the
flow in the element are enhanced. As a result, the dielectric corresponding element in v, respectively. Scalar vi denotes the
loss, metallic loss, and ohmic loss increase dramatically, i-th element of vector v. For a square matrix A, An,k denotes
its entry in the n-th row and k-th column. Cx×y denotes
S. Abeywickrama is with the Department of Electrical and Computer Engi-
neering, National University of Singapore, Singapore, and also with Singapore the space of x × y complex-valued matrices. j denotes the
University of Technology and Design, Singapore (e-mail: [email protected]). imaginary unit, i.e., j 2 = −1. For a complex-valued scalar
R. Zhang is with the Department of Electrical and Computer Engineering, v, |v|, arg(v), and v̄ denote its absolute value, phase, and
National University of Singapore, Singapore (e-mail: [email protected]).
C. Yuen is with Singapore University of Technology and Design, Singapore complex conjugate, respectively. E(·) denotes the statistical
(e-mail: [email protected]). expectation.
 2

Fig. 2: Transmission line model of a unit reflecting element.

Fig. 1: An IRS-aided wireless system.

III. P RACTICAL P HASE S HIFT M ODEL

II. S YSTEM M ODEL
A. Equivalent Circuit Model
We consider a multiple-input single-output (MISO) wireless
An IRS is typically constructed as a printed circuit board
system where an IRS composed of N reflecting elements is
(PCB), where the reflecting elements are equally spaced in a
deployed to assist in the communication from an access point
two-dimensional plane. A unit reflecting element is composed
(AP) with M antennas to a single-antenna user, as illustrated
of a metal patch on the top layer of the PCB dielectric substrate
in Fig. 1. The IRS reflecting elements are programmable via
and a full metal sheet on the bottom layer [2]. Moreover,
an IRS controller. Furthermore, IRS controller communicates
a semiconductor device2, which can vary the impedance of
with the AP via a separate link for information exchange
the reflecting element by controlling its biasing voltage, is
between the IRS and the AP on the channel state information
embedded into the top layer metal patch so that the element
(CSI) and other information needed for implementing the joint
response can be dynamically tuned in real time without chang-
design of the AP transmission and IRS reflection. It is assumed
ing the geometrical parameters [10]. In other words, when the
that the signals that are reflected by the IRS more than once
geometrical parameters are fixed, the semiconductor device
have negligible power due to substantial path loss and thus
controls the phase shift and amplitude (absorption level).
are ignored. In addition, we consider a quasi-static flat-fading
As the physical length of a unit reflecting element is usually
model, where it is assumed that all the wireless channels
smaller than the wavelength of the desired incident signal,
remain constant over each transmission block.
its response can be accurately described by an equivalent
Let hd ∈ CM×1 , hr ∈ CN ×1 , and G ∈ CN ×M denote the
lumped circuit model regardless of the particular geometry of
baseband equivalent channels from the AP to user, from the
the element [11]. As such, the metallic parts in the reflecting
IRS to user, and from the AP to IRS, respectively. Without loss
element can be modeled as inductors as the high-frequency
of generality, let v ∈ CN ×1 denote the reflection coefficient
current flowing on it produces a quasi-static magnetic field. In
vector of the IRS, where |vn | ∈ [0, 1] and arg(vn ) ∈ [−π, π)
Fig. 2, the equivalent model for the n-th reflecting element is
are the amplitude and the phase shift on the combined incident
illustrated as a parallel resonant circuit and its impedance is
signal, respectively, for n ∈ {1, . . . , N } [2]. Note that for the
given by
ideal phase shift model considered in [1]–[3], |vn | = 1, ∀n,
1
regardless of the phase shift, arg(vn ). The transmit signal at jωL1 (jωL2 + jωCn + Rn )
the AP is given by x = ws, where w ∈ CM×1 denotes the Zn (Cn , Rn ) = 1 , (3)
jωL1 + (jωL2 + jωC + Rn )
beamforming vector and s denotes the transmit symbol, which n

is independent of w, and has zero-mean and unit variance (i.e., where L1 , L2 , Cn , Rn , and ω denote the bottom layer
E(|s|2 ) = 1). We have dropped the time index for notational inductance, top layer inductance, effective capacitance, effec-
simplicity. The received baseband signal at the user is thus tive resistance, and angular frequency of the incident signal,
given by respectively. Note that Rn determines the amount of power
dissipation due to the losses in the semiconductor devices,
y = (hH H
r diag(v)G + hd )x + z, (1) metals, and dielectrics, which cannot be zero in practice, and
where z denotes the additive white Gaussian noise (AWGN) Cn specifies the charge accumulation related to the element
at the receiver with zero mean and variance σ 2 . geometry and semiconductor device. As the transmission line
In this letter, we aim to maximize the achievable rate or diagram in Fig. 2 depicts, the reflection coefficient, i.e., vn in
spectrum efficiency (SE) in bits per second per Hertz (bps/Hz) (1), is the parameter that describes the fraction of the reflected
by jointly optimizing the AP beamforming vector w and the electromagnetic wave due to the impedance discontinuity
IRS reflection vector v. Accordingly, the achievable rate/SE between the free space impedance Z0 and element impedance
is given by1 Zn (Cn , Rn ) [12], which is given by
2 Zn (Cn , Rn ) − Z0
|(hH H
 
r diag(v)G + hd )w| vn = . (4)
RSE = log2 1 + . (2) Zn (Cn , Rn ) + Z0
σ2
1 Note that the considered system model can be also applied to wireless 2 In practice, a positive-intrinsic-negative (PIN) diode, a variable capaci-
power transfer (WPT) [2] as the harvested radio-frequency (RF) energy at the tance (varactor) diode, or a metal-oxide-semiconductor field-effect transistor
receiver is generally modeled as an increasing function of the received signal (MOSFET) can be used as the semiconductor device mentioned here [6], [9],
power [8], i.e., the term |(hH H 2
r diag(v)G + hd )w| given in (2). [10].
 3

(a) Phase shift and amplitude versus Cn and Rn . (a) The phase shift model with different parameters.

(b) Amplitude versus phase shift. (b) Simulation results for the proposed phase shift
model.
Fig. 3: Reflection coefficient of a unit reflecting element.
Fig. 4: The proposed phase shift model.

Since vn is a function of Cn and Rn , the reflected elec- 3 are in accordance with the experimental results reported in
tromagnetic waves can be manipulated in a controllable and literature (see [5] and Fig. 5 (b) in [6]), indicating that the
programmable manner by varying Cn ’s and Rn ’s. circuit model given by (3) and (4) accurately captures the
To demonstrate this, Fig. 3 illustrates the behaviour of the physics of a reflecting element in practice.
amplitude and the phase shift, i.e., |vn | and arg(vn ), respec- It is also worth noting that to obtain an ideal phase shift
tively, for different values of Cn and Rn . Note that to align control, where |vn | = 1, ∀ arg(vn ) ∈ [−π, π), each reflecting
with the experimental results in [6], Cn is varied from 0.47 pF element should exhibit zero energy dissipation. However, for
to 2.35 pF when L1 = 0.7 nH, L2 = 2.5 nH, Z0 = 377 Ω, and practical hardware, energy dissipation is unavoidable3 and the
ω = 2π × 2.4 × 109 . It is observed that a reflecting element typical behaviour of the reflection amplitude is similar to Fig.
is capable of achieving almost 2π full phase tuning, while 3. Therefore, incorporating a practical phase shift model to
the phase shift and amplitude both vary with Cn and Rn in design beamforming algorithms is essential to optimize the
general. It is also observed that the minimum amplitude occurs performance of IRS-aided wireless systems.
near zero phase shift and approaches unity (the maximum) at
the phase shift of π or −π, which is explained as follows.
When the phase shift is around π or −π, the reflective currents B. Proposed Phase Shift Model
(also termed as image currents) are out-of-phase with the In order to characterize the fundamental relationship be-
element currents, and thus the electric field and the current tween the reflection amplitude and phase shift for designing
flow in the element are both diminished, thus resulting in IRS-aided wireless systems, we propose in this subsection
minimum energy loss and highest reflection amplitude. In an analytical model for the phase shift which is generally
contrast, when the phase shift is around zero, the reflective applicable to a variety of semiconductor devices used for
currents are in-phase with the element currents, and thus the implementing the IRS. Let vn = βn (θn )ejθn with θn ∈
electric field and the current flow in the element are both [−π, π) and βn (θn ) ∈ [0, 1] respectively denote the phase shift
enhanced. As a result, the dielectric loss, metallic loss, and
3 In [6], R = 2.5 Ω in each reflecting element due to the diode junction
ohmic loss increase dramatically, leading to substantial energy n
resistance, while in [5], although the reflecting element does not contain any
dissipation and thus lowest reflection amplitude. Furthermore, semiconductor device, its amplitude response follows a similar shape to Fig.
it is worth noting that the numerical results illustrated in Fig. 3 due to the metallic loss and dielectric loss.
 4

Algorithm 1 Alternating Optimization (AO) for Solving (P1) B. Proposed Algorithm

1: Initialize: {θn }N
n=1 We propose an AO algorithm to find an approximate solu-
2: Output: v tion to (P1), by iteratively optimizing the phase shift of one
3: repeat of the N reflecting elements with those of the others being
4: for n = 1 to N do fixed at each time, and repeatedly doing this procedure for
5: Find θn∗ as the solution to (P1-AO). all N elements until the objective value in (10) converges.
6: end for The convergence is guaranteed as the optimal value of (P1) is
∗
7: Obtain vn = βn (θn∗ )ejθn , ∀n. upper-bounded by a finite value. To this end, the problem for
8: until the objective value of (P1) with the obtained v optimizing the reflection of the n-th element is formulated as
reaches convergence;  
(P1 − AO) : max βn (θn ) 2|ϕn | cos(θn + arg(ϕn )) + Ψn,n
θn
(11)
and the corresponding amplitude. Specifically, βn (θn ) can be
s.t. − π ≤ θn ≤ π, (12)
expressed as
H H H
 k where Ψ = diag(h
P r )GG diag(h  r ), ĥd = diag(hr )Ghd ,
sin(θn − φ) + 1 N
βn (θn ) = (1 − βmin ) + βmin , (5) and ϕn = m6=n Ψn,m v̄m + ĥd,n . Note that (11) is
2
obtained by taking the terms associated with βn (θn ) and θn in
where βmin ≥ 0, φ ≥ 0, and k ≥ 0 are the constants related to the expansion of (10), while the derivation is omitted due to
the specific circuit implementation. As depicted in Fig. 4 (a), the space limitation. The problem (P1-AO) is a single-variable
βmin is the minimum amplitude, φ is the horizontal distance non-convex optimization problem, where its approximate solu-
between −π/2 and βmin , and k controls the steepness of the tion can be obtained numerically via a one-dimensional search
function curve. Note that for k = 0, (5) is equivalent to the over [−π, π). The overall iterative algorithm to solve (P1) is
ideal phase shift model, i.e., βn (θn ) = 1, ∀n. In practice, IRS given in Algorithm 1.
circuits are fixed once they are fabricated and these parameters Remark 1: Note that for the ideal phase shift model consid-
can be easily found by a standard curve fitting tool. ered in [1]–[3], βn (θn ) and θn can be designed to maximize
Fig. 4 (b) illustrates that the proposed phase shift model
(11) by setting βn∗ (θn ) = 1 and θn∗ = − arg(ϕn ), ∀n.
closely matches the simulation results presented in Section However, such an optimal reflection design is not feasible for
III-A for a practical reflecting element. In the sequel, we adopt
a practical IRS due to the dependency of βn (θn ) on θn as
the model in (5) for beamforming design in IRS-aided wireless depicted in Fig. 3 (b). For instance, if arg(ϕn ) = 0, θn∗ = 0
communication. Moreover, we assume that the circuits of the may not be a favourable phase design as it yields the lowest
reflecting elements are all identical, and thus the same model
reflection amplitude. In this case, θn∗ needs to be properly
parameters, i.e., βmin , φ, and k, apply to each of the elements. chosen to have a better balance between βn (θn ) and φn .

IV. B EAMFORMING O PTIMIZATION

C. Initialization Method
For the proposed AO algorithm, it is desirable to have the
A. Problem Formulation
initial value of {θn }Nn=1 to be close to the optimal solution
We aim to jointly optimize w and v such that the achiev- of (P1). To this end, we initialize {θn }N
n=1 in this letter using
able rate, RSE given in (2), is maximized. The problem is the solution of the following problem, which assumes the ideal
formulated as phase shift control, and its solution is given in [1].
2
(P0) : max |(hH H
r diag(v)G + hd )w| (6) (P2) : max k(hH H 2
w,v,{θn } r diag(ṽ)G + hd )k (13)
ṽ
s.t. kwk22 ≤ PT , (7) 2
s.t. |ṽn | = 1, ∀n = 1, . . . , N. (14)
jθn
vn = βn (θn )e , ∀n = 1, . . . , N, (8)
After solving (P2), we set θn = arg(ṽn ), ∀n, as the initial
− π ≤ θn ≤ π, ∀n = 1, . . . , N, (9) values for Algorithm 1.
where PT denotes the maximum transmit power constraint at
the AP. Following similar steps as in [13], problem (P0) can V. S IMULATION R ESULTS
be reformulated as We consider a MISO downlink wireless system consisting
(P1) : max k(hH diag(v)G + hH 2 of an AP with M = 2 antennas and a single-antenna user. It
r d )k (10)
v,{θn } is assumed that an IRS composed of N reflecting elements
s.t. (8) − (9), is deployed in the vicinity of the user while the AP and IRS
√ (hH diag(v)G+hH )H are assumed to be located 500 meters (m) apart. Rayleigh
with w∗ = PT k(hrH diag(v)G+hdH )k . Although simplified, fading is assumed for all the channels involved, and the
r d
problem (P1) is non-convex and difficult to be optimally solved signal attenuation at a reference distance of 1 m is set as
in general. In the next subsection, we solve (P1) by applying 40 dB. The path loss exponents are set to 2.2, 2.8, and 3.8
the AO technique. for the channels between AP-IRS, IRS-user, and AP-user,
 5

Fig. 5: Achievable rate versus the AP-user horizontal distance when N = 40. Fig. 6: Achievable rate versus number of reflecting elements at d = 498 m.

respectively, according to [1]. The total transmit power at the the proposed practical model. In future work, it is worth
AP is PT = 36 dBm and σ 2 = −94 dBm. investigating such performance difference in more general
The user is assumed to lie on a horizontal line that is in IRS-aided wireless communication setups, such as multi-user
parallel to that connecting the AP and IRS, with the vertical systems [1], [3], OFDM-based system [14], physical layer
distance between these two lines equal to 2 m. By varying security system [15], and so on.
the horizontal distance between the AP and user, denoted by
d, in Fig. 5, the achievable rate averaged over 1000 channel R EFERENCES
realizations is shown for the following schemes: i) Upper [1] Q. Wu and R. Zhang, “Intelligent reflecting surface enhanced wireless
bound: solving (P2) by assuming the ideal phase shift model network via joint active and passive beamforming,” [Online]. Available:
https://arxiv.org/abs/1810.03961.
[1]; ii) Beamforming optimization by the AO algorithm under [2] Q. Wu and R. Zhang, “Towards smart and reconfigurable environment:
the proposed practical phase shift model with βmin = 0.2, Intelligent reflecting surface aided wireless networks,” [Online]. Avail-
k = 1.6, and φ = 0.43π; iii) Beamforming optimization able: https://arxiv.org/abs/1905.00152.
[3] C. Huang, A. Zappone, G. C. Alexandropoulos, M. Debbah, and
assuming the ideal phase shift model [1], while the practical C. Yuen, “Large intelligent surfaces for energy efficiency in wireless
phase shift model is used for computing the achievable rate. communication,” [Online]. Available: https://arxiv.org/abs/1810.06934.
Note that this scheme also corresponds to using the initial [4] M. D. Renzo et al., “Smart radio environments empowered by reconfig-
urable AI meta-surfaces: An idea whose time has come,” EURASIP J.
phase shifts of the AO algorithm only; and iv) Lower √ bound: on Wireless Commun. and Networking, vol. 2019, no. 1, pp. 129–148,
hd
the system without using an IRS by setting w∗ = PT kh dk
. May 2019.
It is observed from Fig. 5 that when the user moves closer [5] H. Rajagopalan and Y. Rahmat-Samii, “Loss quantification for mi-
crostrip reflectarray: Issue of high fields and currents,” in Proc. IEEE
to the IRS, the performance gap between ii) and iii) increases. Antennas and Propag. Society Int. Symposium, Jul. 2008, pp. 1–4.
This is due to the fact that the user benefits from the stronger [6] B. O. Zhu, J. Zhao, and Y. Feng, “Active impedance metasurface with
reflecting channel via IRS (hr ), and therefore accurate reflec- full 360 reflection phase tuning,” Scientific reports, vol. 3, pp. 3059–
3064, Oct. 2013.
tion design at the IRS becomes more crucial. In contrast, when [7] M. E. Bialkowski, A. W. Robinson, and H. J. Song, “Design, develop-
the user moves toward the AP, the performance gap between ment, and testing of X-band amplifying reflectarrays,” IEEE Trans. on
ii) and iii) decreases as the AP-user direct channel (hd ) Antennas and Propag., vol. 50, no. 8, pp. 1065–1076, Aug. 2002.
[8] Y. Zeng, B. Clerckx, and R. Zhang, “Communications and signals design
becomes dominant and the effect of IRS reflection becomes for wireless power transmission,” IEEE Trans. Commun., vol. 65, no. 5,
less significant. Moreover, by fixing the user at d = 498 m pp. 2264–2290, May 2017.
and varying the number of reflecting elements, N , in Fig. 6, [9] W. Tang, X. Li, J. Y. Dai, S. Jin, Y. Zeng, Q. Cheng, and
T. J. Cui, “Wireless communications with programmable metasur-
we plot the average achievable rate. It is also observed that face: Transceiver design and experimental results,” [Online]. Available:
the performance gap between ii) and iii) increases with N as https://arxiv.org/abs/1811.08119.
the IRS reflecting channel becomes stronger. [10] F. Liu et al., “Intelligent metasurfaces with continuously tunable local
surface impedance for multiple reconfigurable functions,” Phys. Rev.
Applied, vol. 11, no. 4, pp. 44 024–44 033, Apr. 2019.
VI. C ONCLUSION [11] S. Koziel and L. Leifsson, Surrogate-based modeling and optimization.
Springer, 2013.
In this letter, we proposed a practical IRS phase shift [12] D. M. Pozar, Microwave Engineering (3th Edition). New York John
model. Based on this new model and considering an IRS-aided Wiley & Sons, 2005.
MISO system, we formulated and solved a joint transmit and [13] Q. Wu and R. Zhang, “Beamforming optimization for intelligent reflect-
ing surface with discrete phase shifts,” in Proc. ICASSP, May 2019, pp.
reflect beamforming optimization problem to maximize the 7830–7833.
achievable rate, by applying the AO technique. Our simulation [14] Y. Yang, B. Zheng, S. Zhang, and R. Zhang, “Intelligent reflecting
results validated our proposed analytical model and showed surface meets OFDM: Protocol design and rate maximization,” [Online].
Available: https://arxiv.org/abs/1906.09956.
that beamforming optimization based on the conventional ideal [15] M. Cui, G. Zhang, and R. Zhang, “Secure wireless communication via
phase shift model, which has been widely used in the literature, intelligent reflecting surface,” IEEE Wireless Commun. Letters, 2019,
may lead to significant performance loss as compared to Early Access.