Reconfigurable Intelligent Surfaces: Principles and

Opportunities
Yuanwei Liu, Xiao Liu, Xidong Mu, Tianwei Hou, Jiaqi Xu, Marco Di Renzo,
Naofal Al-Dhahir

To cite this version:

Yuanwei Liu, Xiao Liu, Xidong Mu, Tianwei Hou, Jiaqi Xu, et al.. Reconfigurable Intelligent Surfaces:
Principles and Opportunities. Communications Surveys and Tutorials, IEEE Communications Society,
2021, �10.1109/COMST.2021.3077737�. �hal-03357962�

HAL Id: hal-03357962

https://hal.science/hal-03357962
Submitted on 29 Sep 2021

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 1

Reconfigurable Intelligent Surfaces:

Principles and Opportunities
Yuanwei Liu, Senior Member, IEEE, Xiao Liu, Xidong Mu, Tianwei Hou,
Jiaqi Xu, Marco Di Renzo, Fellow, IEEE, and Naofal Al-Dhahir Fellow, IEEE
arXiv:2007.03435v3 [eess.SP] 4 May 2021

Abstract

Reconfigurable intelligent surfaces (RISs), also known as intelligent reflecting surfaces (IRSs),
or large intelligent surfaces (LISs)1 , have received significant attention for their potential to enhance
the capacity and coverage of wireless networks by smartly reconfiguring the wireless propagation
environment. Therefore, RISs are considered a promising technology for the sixth-generation (6G) of
communication networks. In this context, we provide a comprehensive overview of the state-of-the-art
on RISs, with focus on their operating principles, performance evaluation, beamforming design and
resource management, applications of machine learning to RIS-enhanced wireless networks, as well as
the integration of RISs with other emerging technologies. We describe the basic principles of RISs both
from physics and communications perspectives, based on which we present performance evaluation of
multi-antenna assisted RIS systems. In addition, we systematically survey existing designs for RIS-
enhanced wireless networks encompassing performance analysis, information theory, and performance
optimization perspectives. Furthermore, we survey existing research contributions that apply machine
learning for tackling challenges in dynamic scenarios, such as random fluctuations of wireless channels

Y. Liu, X. Liu, J. Xu are with the School of Electronic Engineering and Computer Science, Queen Mary University of London,
London E1 4NS, UK. (email: [email protected]; [email protected]; [email protected]).
X. Mu is with School of Artificial Intelligence and Key Laboratory of Universal Wireless Communications, Ministry of
Education, Beijing University of Posts and Telecommunications, Beijing, China (email: [email protected]).
T. Hou is with the School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing 100044, China
(email: [email protected]).
M. Di Renzo is with Université Paris-Saclay, CNRS, CentraleSupélec, Laboratoire des Signaux et Systèmes, 3 Rue Joliot-Curie,
91192 Gif-sur-Yvette, France. (emails: [email protected]).
M. Di Renzo’s work was supported in part by the European Commission through the H2020 ARIADNE project under grant
agreement number 871464 and through the H2020 RISE-6G project under grant agreement number 101017011.
N. Al-Dhahir is with the Department of Electrical and Computer Engineering, The University of Texas at Dallas, Richardson,
TX 75080 USA. (email: [email protected]).

1
Without loss of generality, we use the name of RIS in the remainder of this paper.
 2

and user mobility in RIS-enhanced wireless networks. Last but not least, we identify major issues
and research opportunities associated with the integration of RISs and other emerging technologies for
applications to next-generation networks.

Index Terms

6G, intelligent reflecting surfaces (IRSs), large intelligent surfaces (LISs), machine learning, per-
formance optimization, reconfigurable intelligent surfaces (RISs), wireless networks

I. I NTRODUCTION

The unprecedented demands for high quality and ubiquitous wireless services impose enor-
mous challenges to existing cellular networks. Applications like rate-centric enhanced mobile
broadband (eMBB), ultra-reliable, low latency communications (URLLC), and massive machine-
type communications (mMTC) services are the targets for designing the fifth-generation (5G) of
communication systems. However, the goals of the sixth-generation (6G) of wireless communi-
cation systems are expected to be transformative and revolutionary encompassing applications
like data-driven, instantaneous, ultra-massive, and ubiquitous wireless connectivity, as well as
connected intelligence [1], [2]. Therefore, new transmission technologies are needed in order
to support these new applications and services. Reconfigurable intelligent surfaces (RISs), also
called intelligent reflecting surfaces (IRSs) [3], [4] or large intelligent surfaces (LISs) [5], [6],
comprise an array of reflecting elements for reconfiguring the incident signals. Owing to their
capability of proactively modifying the wireless communication environment, RISs have become
a focal point of research in wireless communications to mitigate a wide range of challenges
encountered in diverse wireless networks [7], [8]. The advantages of RISs are listed as follows:

• Easy to deploy: RISs are nearly-passive devices, made of electromagnetic (EM) material.
As illustrated in Fig. 1, RISs can be deployed on several structures, including but not limited
to building facades, indoor walls [9], aerial platforms, roadside billboards, highway polls,
vehicle windows, as well as pedestrians’ clothes due to their low-cost.
• Spectral efficiency enhancement: RISs are capable of reconfiguring the wireless propaga-
tion environment by compensating for the power loss over long distances. Virtual line-
of-sight (LoS) links between base stations (BSs) and mobile users can be formed via
 3

passively reflecting the impinging radio signals. The throughput enhancement becomes
significant when the LoS link between BSs and users is blocked by obstacles, e.g., high-rise
buildings. Due to the intelligent deployment and design of RISs, a software-defined wireless
environment may be constructed, which, in turn, provides potential enhancements of the
received signal-to-interference-plus-noise ratio (SINR).
• Environment friendly: In contrast to conventional relaying systems, e.g., amplify-and-
forward (AF) and decode-and-forward (DF) [10], RISs are capable of shaping the incoming
signal by controlling the phase shift of each reflecting element instead of employing a power
amplifier [11], [12]. Thus, deploying RISs is more energy-efficient and environment friendly
than conventional AF and DF systems.
• Compatibility: RISs support full-duplex (FD) and full-band transmission due to the fact
that they only reflect the EM waves. Additionally, RIS-enhanced wireless networks are
compatible with the standards and hardware of existing wireless networks [13].
Due to the aforementioned attractive characteristics, RISs are recognized as an effective
solution for mitigating a wide range of challenges in commercial and civilian applications.
There have been many recent studies on RISs and their contributions focus on several application
scenarios under different assumptions. As a result, the system models proposed by these research
contributions tend to be different. Thus, there is an urgent need to categorize the existing research
contributions, which is one of the main goals of this paper.
Fig. 1 illustrates the applications of RISs in diverse wireless communication networks. In
Fig. 1(a), RIS-enhanced cellular networks are illustrated, where RISs are deployed for bypassing
the obstacles between BSs and users. Thus, the quality of service (QoS) in heterogeneous
networks and the latency performance in mobile edge computing (MEC) networks are im-
proved [14], [15]. On the other hand, RISs can act as a signal reflection hub to support massive
connectivity via interference mitigation in device-to-device (D2D) communication networks [16],
or RISs can cancel undesired signals by smartly designing the passive beamforming in the
context of physical layer security (PLS) [17]. Additionally, RISs can be deployed to strengthen
the received signal power of cell-edge users and mitigating the interference from neighbor
cells [18], and the power loss over long distances can be compensated in simultaneous wireless
information and power transfer (SWIPT) networks [19]–[22]. In Fig. 1(b), RIS-assisted indoor
 4

(a) RIS-enhanced cellular networks beyond 5G (b) RIS-assisted indoor communications

RISs in SWIPT networks/energy
RIS-enhanced mobile Wind energy
havesting networks
Solar energy
edge computing
RIS-enhanced mmWave
MEC sever RISs on pedestrianÿs
mmWave Communication
clothes
communication
networks

Station
RISs in heterogeneous RIS-enhanced LiFi
networks communication
networks
RIS-enhanced physical
layer security
Macrocell AP RIS-enhanced WiFi networks

RIS-enhanced NOMA
Eavesdropper networks

Femtocell AP
Legitimate
RISs on the wall

Power domain
WiFi
RIS-enhanced D2D
communications Frequency domain

UAV RISs in AUV-enabled

networks RISs in intelligent
factory RISs in intelligent wireless
sensor networks
RISs in cellular-
connected UAV
networks

AUV

Sensors

RISs in AI- RISs in intelligent

robotics team agriculture

(c) RISs in unmanned systems for smart city (d) RISs in intelligent IoT networks

Figure 1: RISs in wireless communication networks.

communications are illustrated, where RISs can be deployed on walls for enhancing the QoS
in some rate-hungry indoor scenarios, such as virtual reality (VR) applications. Additionally,
in order to guarantee no blind spots in the coverage area of some block-sensitive scenarios,
such as visible light communications [23] and wireless fidelity (WiFi) networks, a concatenated
virtual RIS-aided LoS link between the access points (APs) and the users can be formed with
the aid of RISs, which indicates that both the propagation links between the APs and the RISs,
as well as between the RISs and the users can be in LoS. In Fig. 1(c), RIS-enhanced unmanned
systems are illustrated. RISs can be leveraged for enhancing the performance of unmanned
aerial vehicle (UAV) enabled wireless networks [24], cellular-connected UAV networks [25],
autonomous vehicular networks, autonomous underwater vehicle (AUV) networks, and intelligent
robotic networks by fully reaping the aforementioned RIS benefits. For instance, in RIS-enhanced
 5

UAV-aided wireless networks, one can adjust the phase shifts of RISs instead of controlling the
movement of the UAVs in order to form concatenated virtual LoS links between the UAVs and
the users. Therefore, the UAVs can maintain the hovering status only when the concatenated
virtual LoS links cannot be formed even with the aid of RISs, which reduces the movement
manipulations and the energy consumption of UAVs. In Fig. 1(d), RIS-enhanced Internet of
Things (IoT) networks are illustrated, where RISs are exploited for assisting intelligent wireless
sensor networks [26], intelligent agriculture, and intelligent factory [27].
There are some short magazine papers [3], [10], [28], [29], surveys and tutorials [6], [7], [30]–
[33] in the literature that introduced RISs and their variants, but the focus of these papers is
different from our work. More specifically, Wu et al. [3] provided an overview of the applications
of RISs as reflectors in wireless communications, and identified some challenges and future
research opportunities for implementing RIS-assisted wireless networks. Liang et al. [6] presented
an overview of the reflective radio technology with a particular focus on the large intelligent
surface/antennas. In [7], Di Renzo et al. provided a comprehensive overview of employing RISs
for realizing smart radio environments in wireless networks, where an electromagnetic-based
communication-theoretic framework for analyzing and optimizing metamaterial-based RISs is
presented and a survey of recent research contributions on RISs is given. Huang et al. [10]
introduced the concept of holographic multiple-input and multiple-output (MIMO) surfaces
(HMIMOS), and discussed both active and passive RISs, encompassing the hardware archi-
tectures, operation modes, and applications in communications. In [28], Liaskos et al. presented
one kind of RIS prototype, namely the HyperSurface tile, for realizing programmable wireless
environments. Gacanin et al. [29] gave an overview of employing artificial intelligence (AI) tools
in RIS-assisted radio environments. Di Renzo et al. [30] introduced the concept of smart radio
environments empowered by RISs, and discussed recent research progresses and future poten-
tial challenges. Basar et al. [31] reviewed recent research efforts on RIS-empowered wireless
networks, identified the differences between RISs and other technologies, and presented future
research challenges and opportunities. Gong et al. [32] surveyed recent research works on RIS-
assisted wireless networks and discussed emerging applications and implementation challenges of
RISs. From the perspective of enhancing the communication performance, Wu et al. [33] gave
a tutorial on design issues in RIS-assisted wireless networks, including passive beamforming
 6

optimization, channel estimation, and deployment design.

Although the aforementioned magazines/surveys/tutorials presented either general concepts or
specific aspects of RISs (e.g., from a physics-based or a communication-based perspective), the
fundamental performance limits of RISs and some potential applications in wireless networks
are not covered. The comparison between widely employed mathematical tools for performance
evaluations and optimizations in RIS-enhanced wireless networks is also not discussed. Moreover,
a detailed framework based on machine learning (ML) tools for designing RIS-enhanced wireless
networks is missing, except for a short magazine paper [29]. Motivated by all the aforementioned
considerations, this paper provides a comprehensive discussion of RIS-enhanced wireless network
principles, from physics to wireless communications, and discusses research opportunities for
exploiting RISs in diverse applications, such as unmanned systems, non-orthogonal multiple
access (NOMA), and ML. Table I illustrates the comparison of this treatise with the existing
magazines/surveys/tutorials in the context of RISs.
Against the above observations, our main contributions are as follows.
Wu Liang Di Renzo Huang Liaskos Gacanin Di Renzo Basar Gong Wu This
— Classifications Key Contents et al. et al. et al. et al. et al. et al. et al. et al. et al. et al. work
[3] [6] [7] [10] [28] [29] [30] [31] [32] [33]
Distinguishing ray-optics & wave-optics perspective X X
Macroscopic description of metasurface-based RIS X X X
Surface equivalence theorems X X X
Electromagnetics
Distinguishing far-field & near-field X X
Formulating the reflection coefficient X X X X X X
Physics-based
Power conservation principle X
RIS control mechanism X X X X X X X X X X X
RIS Modeling
Typical tunable functions X X X X X
RIS hardware and prototypes X X X X X X X X X
Realizations
RIS synthesis methods X X X X X
Path loss models X X X X X X
Performance Stochastic geometry X X X X X
evaluation Discussion on stochastic analysis tools X X
Discussion on small-scale fading channels X
Information-theoretic capacity limits X
Passive beamforming optimization X X X X X X X X
RIS-aided Resource management X X
communication Comparison of employed mathematical approaches X
design Channel estimation X X X X X X X X
Deployment design X X X X X
Modulation X X X
Communication-based
Localization and sensing X X
Architecture X X
ML-empowered DL X X X X X X
RIS RL X X
Comparison of RL-based algorithms X
Supervised, unsupervised, and federated learning X X
NOMA-RIS X X
PLS-RIS X X X X X X X
SWIPT-RIS X X X X X X
Compatibility UAV-RIS X X X X
AV/CV-RIS X
MEC-RIS X X X X
mmWave-RIS X X X X X X

Table I: Comparison of this work with available magazines/surveys/tutorials. Here, “DL” refers
to “Deep Learning”, “RL” refers to “Reinforcement Learning”, “AV” refers to “Autonomous
Vehicle”, and “CV” refers to “Connected Vehicle”.

• We overview the fundamental principles that govern the operation of RISs and their inter-
action with the EM signals. We also survey typical RIS functions and their corresponding
 7

principles. Specifically, we focus on patch-array based implementation and compare the

ray-optics perspective with the wave-optics perspective.
• We develop performance evaluation techniques for multi-antenna assisted RIS systems.
Research contributions are also summarized along with their advantages and limitations.
• We investigate RISs from the information-theoretic perspective, based on which we review
the protocols and approaches for jointly designing beamforming and resource allocation
schemes with different optimization objectives. Additionally, the major open research prob-
lems are discussed.
• We discuss the need of amalgamating ML and RISs. After reviewing the most recent research
contributions, we propose a novel framework for optimizing RIS-enhanced intelligent wire-
less networks, where big data analytic and ML are leveraged for optimizing RIS-enhanced
wireless networks.
• We identify major research opportunities associated with the integration of RISs into other
emerging technologies and discuss potential solutions.

As illustrated in Fig. 2, this paper is structured as follows. Section II elaborates on the

fundamental operating principles of RIS-enhanced wireless networks. Section III focuses on
the performance evaluation of multi-antenna RIS-assisted systems and the main advantages of
using RISs in wireless networks. In Section IV, the latest research activities on the joint design
of beamforming and resource allocation are discussed. The framework of ML-empowered RIS-
enhanced intelligent wireless networks is presented in Section V. Finally, Section VI investigates
the integration of RISs with other emerging technologies towards the design and optimization
of 6G wireless networks.

II. RIS: F ROM P HYSICS TO W IRELESS C OMMUNICATIONS

An RIS is a two-dimensional (2D) material structure with programmable macroscopic physical

characteristics. The most important characteristic of an RIS is that its EM wave response can be
reconfigured. In contrast to conventional wireless communication networks, the channels between
the transmitters and the receivers can be controlled in RIS-aided networks. Thus, the strength of
the desired received signal can be enhanced at the terminal devices. In this section, we introduce
 8

Section I: Introduction

Section II: RIS: from physics to wireless communications

A. Different categories of RISs

B. Understanding metasurfaces from the view of physics

C. Analysis of RISs: ray optics perspective v.s. wave optics
perspective
D. Achieving tunability: patch array RIS

E. RIS operating principles F. Discussions and outlook

Section III: Performance analysis of multi-antenna assisted RIS systems

A. Channel models

B. Performance analysis

C. Benchmark schemes D. Discussions and outlook

Section IV: RIS beamforming and resource allocation

A. Information-theoretic perspective

B. Joint transmit and passive beamforming design

C. Resource management in RIS-enhanced networks

D. Discussions and outlook

Section V: Machine learning for RIS-enhanced communication systems

A. Motivations and architecture of integrating ML in RIS-
enhanced wireless networks
B. Deep learning for RIS-enhanced communication systems
C. Reinforcement learning for RIS-enhanced communication
systems

D. A novel architecture of ML-empowered RIS-enhanced

wireless networks

E. Other ML techniques for RIS-enhanced communication systems

F. Discussions and outlook

Section VI: Integrating RISs with other technologies towards 6G

A. NOMA and RIS B. PLS and RIS

C. SWIPT and RIS D. UAV and RIS

E. Autonomous Driving/Connected Vehicles and RIS

F. Discussions and outlook

Figure 2: Organization of the present paper.

the fundamental principles which govern the operation of RISs and their interaction with the
EM signals. We also survey typical RIS functions and their corresponding principles.

A. Different Categories of RISs

Considering their structures, RISs can be realized by using metamaterial or patch-array based
technologies. Metamaterial-based RISs are referred to as metasurfaces. Deployed at different
locations, RISs can be designed to work as reflecting/refracting surfaces between the BS and the
user or waveguide surfaces operating at the BS. Considering the tuning mechanisms, RISs can be
 9

Table II: LIST OF ACRONYMS

AF Amplify-and-Forward
AO Alternating Optimization
AUV Autonomous Underwater Vehicle
BC Broadcast Channel
BS Base Station
CSI Channel State Information
D2D Device-to-Device
DF Decode-and-Forward
DL Deep Learning
EE Energy Efficiency
EM Electromagnetic
FD Full-Duplex
HD Half-Duplex
IRS Intelligent Reflecting Surface
IoT Internet of Things
LIS Large Intelligent Surface
LoS Line-of-Sight
MEC Mobile Edge Computing
MIMO Multiple-Input and Multiple-Output
MISO Multiple-Input and Single-Output
ML Machine Learning
MOS Mean Opinion Score
NP Non-deterministic Polynomial-time
NOMA Non-Orthogonal Multiple Access
OFDM Orthogonal Frequency Division Multiplexing
OMA Orthogonal Multiple Access
PDF Probability Density Function
PLS Physical Layer Security
QoS Quality of Service
RIS Reconfigurable Intelligent Surface
RL Reinforcement Learning
SCA Successive Convex Approximation
SE Spectral Efficiency
SG Stochastic Geometry
SIC Successive Interference Cancelation
SIMO Single-Input and Multiple-Output
SINR Signal-to-Interference-plus-Noise Ratio
SISO Single-Input and Single-Output
SNR Signal-Noise Ratio
SWIPT Simultaneous Wireless Information and Power Transfer
UAV Unmanned Aerial Vehicle
VLC Visible Light Communication
VR Virtual Reality
WiFi Wireless Fidelity
5G Fifth-Generation
6G Sixth-Generation

reconfigured electrically, mechanically, or thermally. Depending on their energy consumptions,

RISs can be categorized as passive-lossy, passive-lossless, or active. The active or passive nature
of RISs determines their ultimate performance capabilities. It is worth mentioning, however, that
RISs cannot be completely passive because of their inherent property of being configurable.
Here, we discuss three important RIS working operations: waveguide [34], refraction [35], and
reflection [36]. With the aid of Love’s field equivalence principle [37], the reflected and refracted
EM field can be studied by introducing equivalent surface electric and magnetic currents [35].
 10

Metamaterial Based
Structure
Patch-array Based

Reflecting/Refracting RIS
Power Source
Waveguide RIS

RIS
Mechanical Stretching

Energy consumption Electrical Excitation

Thermal Excitation

Passive-lossy

Tuning Mechanisms Passive-lossless

Active

Figure 3: Different types of metasurfaces.

In the three working conditions, the RIS converts and radiates a wave (either induced by an
incident wave or fed by a waveguide) into a desired propagating wave in free space. The
surface equivalence principles (SEPs) including Love’s field equivalence principle and Huygens’
principle are introduced in Section II-B.
1) Waveguide RIS: R. Smith et al. [34] presented a theoretical study of waveguide-fed
metasurfaces. The elements in the metasurface are modeled as uncoupled magnetic dipoles.
The magnitude of each dipole element is proportional to the product of the reference wave
and each element’s polarizability. By tuning the polarizability, the metasurface antenna can
perform beamforming. Each element on the metasurface serves as a micro-antenna. Compared
to conventional antenna arrays, the compact waveguide metasurface occupies less space and can
transmit towards wider angles.
2) Refracting RIS: Viktar S. et al. [38] proposed a theoretical design of a perfectly refracting
and reflecting metasurfaces. The authors used an equivalent impedance matrix model so that
the tangential field components at the two sides of the metasurface are appropriately optimized.
Moreover, three possible device realizations are discussed: self-oscillating teleportation metasur-
faces, non-local metasurfaces, and metasurfaces formed by only lossless components. The role of
omega-type bianisotropy in the design of lossless-component realizations of perfectly refractive
 11

surfaces is discussed.
3) Reflecting RIS: Dai et al. [36] designed a digital coding reflective metasurface. The
elements in the metasurface contain varactor diodes with a tunable biasing voltage. By pre-
designing several digitized biasing voltage levels, each element can apply discrete phase shifts
and achieve beamforming for the reflected wave.
The rest of this section is focused on the operating principles for RISs that operate as reflectors.

B. Understanding Metasurfaces From the View of Physics

A wireless signal is essentially an EM wave propagating in a three-dimensional space. Atten-

uation or reduction of the signal strength occurs as the EM wave propagates through the space
and interacts with the scattering objects. From basic principles of electromagnetism, the signal
power per unit area is proportional to the square of the electric field of the corresponding wave
in a given media. As far as reflective and refractive smart surfaces are concerned, this requires
the understanding of how the EM waves interact with the surrounding objects. The equivalence
principle, especially the SEP, is the building block for studying the EM wave transformations.
Some authors also call it Love’s field equivalence principle. The principle can be adopted for both
external problems (source-free region) and internal problems. Love’s field equivalence principle
states that the EM field outside or inside a close surface can be uniquely determined by the
electric and magnetic currents on the surface. As shown in Fig. 4, the equivalent problem for
the region I can be reformulated by placing equivalent currents on S that satisfy the boundary
condition for each particular case and filling the region II with the same medium of constitutive
parameters  and µ. Thus, the equivalent currents (−Js ,−Ms ), together with the original source
currents (J1 ,M1 ), radiate the correct fields in region I. The equivalent problem for region II can
be formulated similarly.
Love’s field equivalence principle is the theoretical foundation for analyzing the radiation
pattern of RISs. However, the SEP does not specify how to calculate the EM field produced by
the surface currents. To obtain the signal strength at an arbitrary field point, the Huygens-Fresnel
principle can be employed. The Huygens-Fresnel principle is a method of analysis applied to
problems of wave propagation, which states that every point on a wavefront is itself the source
of spherical wavelets, and the secondary wavelets emanating from different points mutually
 12

interfere. The sum of these spherical wavelets forms the wavefront. Based on the Huygens-
Fresnel principle, the EM field scattered by an RIS (in reflection and refraction) can be quantified
analytically.
As far as waveguide-based RISs are concerned, the operating principle can be summarized as
follows. In [39], the EM wave manipulation of the waveguide metasurface performs the coupling
between three-dimensional free space waves and two-dimensional surface waves. As a result,
the metasurface can be regarded as a hologram, which carries additional information about its
radiated signal propagating in the 3D space. After being excited by the source, this pre-designed
information is coupled into the radiated field. Fig. 5 conceptually represents a pre-designed
holographic waveguide-based RIS.

C. Analysis of RISs: Ray Optics Perspective v.s. Wave Optics Perspective

To characterize the interaction between an RIS and the impinging EM waves, one can adopt
approximations and tools either from the perspective of ray-optics or wave-optics. These two
perspectives have been used by physicists for a long time. Even though they are based on some
approximations, these two methods of analysis are useful in order to obtain important insights
into the interaction of light or radio waves with materials. In the RIS literature, both methods
of analysis are often employed. However, the assumptions behind their use and their physical
interpretations are intrinsically different. To shed light on their differences and similarities,
we compare the two methods in this subsection. As shown in Fig 6(a), from the ray-optics
perspective, an EM wave is modeled as a collection of geometrical rays with varying phases.
The phase of each ray increases linearly with the optical path length as its traverses through
the vacuum or other media. As a result, at each location of the i-th ray, a phase (denoted
by φi ) can be defined. When the ray interacts with a material, the phenomenon is studied by
determining the relationship between the change of the phase and the material refraction index.
The desirable reflected wave is obtained if the ensemble of rays obey the proper co-phase
condition. The wave-optics perspective is shown in Fig 6(b), where an EM wave is represented
by the corresponding electric field and magnetic field. At each position, each of these two
vector fields can be characterized by a time-varying complex-valued vector, with a direction, an
amplitude, and a phase. From the wave-optics perspective, the interaction between the wave and
 13

Original Fields M1 E1 , H 1 Null Field

M1

J1 J1
-Js,-Ms

Ɛ, μ0 Ɛ0 , μ0 Ɛ, μ Ɛ, μ
S S
Regionĉ RegionĊ Regionĉ RegionĊ

(a) Original problem (b) Equivalent problem for Regionĉ

Figure 4: Love’s equivalence internal problem (for region I).

desired radiation

surface currents
metasurface impedance surface

Figure 5: Conceptual illustration of the holographic impedance smart surface.

the material can be studied using the equivalent principles discussed in the previous subsection.
The points with equal phase values form a series of surfaces in space, which we refer to as
wavefronts. As a result, the desired scattered waves (reflected or refracted) are obtained if proper
wavefront transformations are performed by the RIS.
1) Comparing Ray-Optics and Wave-Optics Perspectives: Table III highlights some of the
differences between the two methods of analysis. Compared with wave-optics, the ray-optics
perspective is a stronger simplification of the real system. As a result, it is easier to be adopted

Ray-optics Wave-optics
Wave representation Geometrical rays Vector fields
Theoretical foundation Snell’s law Maxwell’s equations
Surface profile Phase discontinuity Surface impedance
Requirement of the reflected wave Co-phase condition Proper wavefront
Power flow Not accurate Accurate

Table III: Comparing different wave representation perspectives

 14

and can produce a quick prediction about the RIS design. However, ray-optics methods fail
when considering the RIS power flow. Wave-optics methods, on the other hand, predict the
power flow by using the Poynting vector, which enables us to study the local and overall RIS
power consumption. This is an important issue to consider when designing and manufacturing
RISs. For example, the authors of [38] and [40] point out that it is impossible to realize lossless
plane-wave beam steering with locally passive RISs. One has to adopt the wave-optics perspective
to study the power flow of the system. Moreover, the authors of [40] state that, according to their
simulation results, one can expect increasingly improved performance if the RISs are designed
based on wave-optics approximation in comparison to those designed based on the ray-optics
approximation. In conclusion, both perspectives have their advantages and limitations. However,
adopting the wave-optics perspective is the most appropriate choice for most cases. A study of
the differences, in terms of power flow and reflection coefficient, between ray optics and wave
optics methods of analysis is reported in [7].

D. Achieving Tunability: Patch Array RIS

The EM characteristics of an RIS, such as the phase discontinuity, can be reconfigured by

tuning the surface impedance, through various mechanisms. Apart from electrical voltage, other
mechanisms can be applied, including thermal excitation, optical pump, and physical stretching.
Among them, electrical control is the most convenient choice, since the electrical voltage is easier
to be quantized and controlled by field-programmable gate array (FPGA) chips. The choice of
RIS materials include semiconductors [41] and graphene [42].
Regardless of the tuning mechanisms, we focus our attention on patch-array smart surfaces in
the following text. The general geometry layout of this type of RIS can be modeled as a periodic
(or quasi-periodic in the most general case) collection of unit cells integrated on a substrate.
For ease of description, we limit our discussion to RISs that are based on a local design, in
which the cells do not interact with each other. A local design usually results in the design of
sub-optimal RISs. A comprehensive discussion about local and non-local designs can be found in
[7]. To characterize the tunability of the RIS, the method of equivalent lumped-element circuits
can be adopted. As shown in Fig. 7, the unit cell is equivalent to a lumped-element circuit with
a load impedance Zl . Particularly, the equivalent load impedance can be tuned by changing the
 15

cophase at proper wave-

target front towards
Einc target

¶ Eref

¶s(x2,y2)
Ms
¶ Js
surface currents
RIS ¶s(x1,y1) RIS

(a) Ray-optics design (b) Wave-optics design

Figure 6: Comparison between ray-optics and wave-optics perspective.

bias voltage of the varactor diode. When modeling patch-array RISs in wireless communication
systems, we can characterize, under a local design, each of its unit cells through an equivalent
reflection coefficient. For example, the reflection coefficient of the i-th cell can be modeled as
follows:
ri = βi · ejφi (1)

where βi and φi correspond to the amplitude response and the phase response, respectively.
As shown in [43], the equivalent reflection coefficient depends on the tuning impedance of the
lumped circuit that controls each unit cell, as well as the self and mutual impedances (if mutual
coupling cannot be ignored) at the ports of the RIS. In particular, as shown in [44] and [43],
βi and φi in (1) are usually not completely independent with each other, i.e., βi = f (φi ). In the
following sections, Φ(r~x ) refers to the phase discontinuity introduced by the RIS as a function
of the position on the RIS, and φmn refers to the phase discontinuity of the (m, n)-th element
of a patch-array RIS.
Existing designs of patch-array RISs can apply discrete phase control and, in some cases,
amplitude control. Arun et al. [45] designed RFocus, which is a two-dimensional surface with a
rectangular array of passive antennas. The size of each passive unit is λ/4 × λ/10 and the
EM waves are either reflected or refracted. The authors show that the RFocus surface can
be manufactured at a low cost, and that it can improve the median signal strength by 9.5
times. Welkie et al. [46] developed a low-cost device embedded in the walls of a building to
passively reflect or actively transmit radio signals. Dunna et al. [47] realized ScatterMIMO,
which uses a smart surface to increase the scattering in the environment. In their hardware
 16

design, each reflector unit uses a patch antenna connected to four open-ended transmission
lines. The transmission lines provide 0, π/2, π or 3π/2 phase shifts. Based on measurements, it
is shown that ScatterMIMO increases the throughput by factor of two and the signal-noise ratio
(SNR) by 4.5 dB as compared with baselines schemes. It is worth mentioning that tunability
can be achieved with implementations other than patch-array based surfaces. For example,
PIVOTAL COMMWARE proposed a new technique called Holographic Beam Forming (HBF).
The proposed holographic beamformer has a low cost and power consumption, as compared
with other transmission technologies, such as massive MIMO and phased arrays.

E. RIS Operating Principles

Considering single-beam reflection, a patch-array based RIS can be configured to serve a

terminal device in the far-field and near-field regions. Among the many operating functions and
configurations of RISs, anomalous reflection and beamforming are widely used in the context of
wireless communications. Adopting the wave-optics perspective, anomalous reflection is a wave-
front transformation from a plane wave to another plane wave, while beamforming is a wavefront
transformation from a plane wave to a desired wavefront. Adopting the ray-optics perspective,
we present the operating principles of these two configurations in the following text. As far as
anomalous reflection is concerned, the RIS is designed to reflect an incident beam to a far-field
terminal, following the generalized laws of reflection [48]. As far as beamforming (also called
focusing) is concerned, the incident wave is focused towards a targeted region, often referred
to as the focal point. The required RIS configuration follows the co-phase condition [49]. The
relation between these two operating principles is discussed in detail in [50]. Before presenting
these two different principles, the physical distinction between the near-field region and the
far-field region is clarified.
1) Near Field v.s. Far Field: In the spirit of dimensional analysis, the characteristics of a
system can be represented by dimensionless numbers. In order to separate the near-field region
from the far-field region, a proper dimensionless number is needed. Let L and RF denote the
antenna aperture size of the RIS and the focal distance, respectively. Assume that z is the distance
of a particular field point to the RIS. Theoretically, the far-field and near-field regimes can be
differentiated as follows: The distance of 2L2 /λ is a commonly used criterion to decide the
 17

C
varactor
L
R

Zl

RIS
Figure 7: Schemetic diagram of the varactor RIS.

boundary between the near-field and far-field regions (see [51], equation (1)). The position
corresponding to z = 2L2 /λ is the boundary between the near-field region and the far-field
region. This result comes from the inspection of the power density variation with the distance
between a field point and the RIS. Within the near field where z < 2L2 /λ, the power density
shows significant variations. The peak position of the power density in the near-field region,
namely RF , changes with different RIS configurations. Using proper co-phase conditions, beam
focusing can be achieved within the near-field of the RIS. It is worth mentioning that, in general,
the boundary between the near-field and far-field regimes depends on the specific configuration
of the RIS, as it was recently remarked in [50].
In general, the essential difference between the near-field region and the far-field region is
how the power density changes with distance. Consider, for example, that the RIS focuses the
wave within an area a. The total energy incident on the RIS is proportional to the solid angle,
Ω spanned by the surface area of the RIS with respect to the location of the transmitter. After
the reflection, the transmitted energy is spread over the area a. Thus, the power density around
the focal point is proportional to Ω/a. Moreover, according to [45], the area a is proportional
to λ2 (1 + 4z 2 /L2 ), as a result of the Abbe diffraction limit. In the far-field region, the second
term inside the brackets dominates and Ω/a is proportional to L2 Ω/z 2 , which is the typical
 18

spherical dissipation of the signal power with the distance. In the near-field region, the first term
dominates and the area a becomes very small. As a result, a high focusing gain can be achieved.
With the aid of an RIS, in general, the signal can be enhanced in both the near-field and far-
field regimes. However, the rationale of this enhancement is different. For near-field applications,
the RIS is expected to enhance the signal strength for users located at targeted locations with
respect to the RIS, while reducing the signal at other locations. For far-field applications, the
RIS is typically expected to enhance the signal strength for users located at targeted angles with
respect to the RIS.
In the following text, adopting the ray-optics perspective, we discuss the generalized laws of
refraction and reflection, as well as the corresponding co-phase condition.
2) The Generalized Laws of Refraction and Reflection: From a geometrical optics perspective,
anomalous reflection and refraction from an RIS can be described by using the generalized laws
of refraction and reflection [48], which is a natural derivation of both Fermat’s principle and the
boundary conditions governed by Maxwell’s equations.

Principle 1. Achieving anomalous reflection (Fig. 8(a))

Suppose that the phase discontinuity at the boundary is a function of the position along the
x direction Φ(r~x ), where r~x is the position vector on the boundary. Moreover, suppose that the
derivative of the phase discontinuity exists. Then, the angle of reflection (θ1 ) and the angle of
refraction (θ2 ) are [52]:
" #
−1 λ dΦ
θ1 = sin sinθi + , (2)
2πn1 dx
" #
n1 λ dΦ
θ2 = sin−1 sinθi + , (3)
n2 2πn2 dx

where θi is the angle of incidence, λ is the wavelength of the transmitted signal in vacuum, and
n1 , n2 are the refractive indexes, as shown in Fig. 9. 

There are other results related to the generalized laws of refraction and reflection, including the
critical angles for total internal reflection or refraction. The main result just presented here states
that, when a phase discontinuity is introduced at the boundary surface, the angles of reflection
 19

BS Source

User A User B User C User

(a) Anomalous reflection (b) Beamforming

(beam steering) (beam focusing)

Figure 8: Typical functions of reflecting surfaces.

Incident
z plane
Incident
beam
Reflected
beam
y 煭濼
煭濄

瀁濄 x

瀁濅
煭濅
metasurface

Refracted
beam

Figure 9: Illustration of the generalized laws of refraction and reflection.

and refraction depend not only on the angle of incidence but also on the wavelength, refractive
indexes, and the gradient of the phase discontinuity. This gives extra controllable parameters
to manipulate the reflected and refracted EM waves. As a result, anomalous reflection can be
achieved by tuning the phase gradient (dΦ/dx) based on (2) or, in the discrete patch-array
implementation, by tuning the length of the super-lattice. However, the assumption that the
derivative of the phase discontinuity is constant (dΦ/dx = const.) does not necessary hold if
different wave transformations are needed.
3) Co-phase condition: Focusing is usually implemented when the RIS is within the near-field
of the source or the terminal is close to the RIS. In these cases, the curvatures of the incident
 20

and reflected wavefront are non-negligible. The optimization of the surface aims to produce a
pencil-beam pointing towards the direction of the terminal. When the link between the source
and the RIS, as well as the link between the RIS and the terminal are in LoS, the following
co-phase condition [49] can be applied.

Principle 2. Achieving beamforming (focusing) (Fig. 8(b))

Let rmn denote the position of the (m, n)-th RIS element, rs denote the position of the source,
and û denote the direction of the observer with respect to the surface. As shown in Fig. 10, φmn
can be chosen as follows [53]:

− k0 (|~rmn − ~rs | − ~rmn · û) + φmn = 2π · t (4)

where t = 1, 2, 3... and k0 = 2π/λc . 

The two design principles just discussed provide guidance on how to configure the RIS phase
shift patterns for typical applications. In more complex wireless communication systems, the
RIS role is more intricate, thus cannot be categorized by the two working functions described
in Fig. 8. In these cases, to determine the RIS configuration, an optimization problem needs be
formulated. These issues are elaborated in the following sections.

F. Discussions and Outlook

In the design and configuration of RISs, both theoretical limitations and hardware implementa-
tion limitations affect the overall performance of the resulting system. Theoretical limitations are,
e.g., the result of considering simplified assumptions or adopting naive perspectives during the
modeling of the RIS and its interaction with the EM wave. Hardware implementation limitations
come, e.g., from the discretization of an ideally continuous RIS profile. In the following text,
we elaborate on three major points.
1) Hardware limitations: In practical application scenarios, many hardware parameters signif-
icantly affect the achievable performance of the system. For example, the number of quantization
levels of the RIS phase shifts, the maximum number of elements that is possible to integrate
on the substrate, and the percentage of scattering environment that can be coated by an RIS.
 21

source

rs

rmn
center of
coordinate
û

observation
direction

Figure 10: Coordinate representation of co-phase condition.

Existing research contributions studied the limitations and tradeoffs caused by these hardware
constraints by analyzing their effect on the channel distribution [54], the power scaling law [55],
and performance metrics such as the outage probability [56] and ergodic capacity [57]. At the
time of writing, the analysis of the impact of hardware limitations and how an RIS performs
compared to other available technologies are open research issues.
2) System design simplifications: The adoption of oversimplified models for RIS hardware
or channel models result in limitations on the system design. Because of the complex nature
of RISs and their interaction with the environment, initial research contributions have adopted
simple models. For example, using hardware models based on local designs, the ray-optics
perspective for channel modeling, and decoupling reflection amplitude and phase shift of the
RIS elements. Even though the ray-optics approximation can yield effective designs in some
cases (as shown in Section II-E), it is preferable to adopt the wave-optics design in practical
cases [50]. Further research efforts are needed to bridge complex physical models of different
RIS implementations with widely used communication models [43].
3) Optimization limitations: To reap the benefits of deploying RISs in wireless networks, the
RIS parameters (e.g., the reflection coefficient and deployment location) and network resource al-
 22

location (e.g., transmit beamforming and user scheduling) should be jointly optimized. However,
the resulting joint optimization problems are normally non-convex and involve highly-coupled
variables, which make it challenging to derive a globally optimal solution. Though some efficient
algorithms have been recently proposed to compute high-quality suboptimal solutions [58]–[60],
the performance limits of RISs remain unknown. To fully understand the attainable performance
limits, information-theoretic perspective investigations [61], [62] are important and sophisticated
mathematical tools [63] are expected to be employed. Further details are given in Sections VI
and V.

III. P ERFORMANCE A NALYSIS OF M ULTI -A NTENNA A SSISTED RIS S YSTEMS

In Section II, we have discussed the fundamental physical properties of the RISs. However,
how RISs affect the communication performance is still an open problem. To systematically
survey existing designs for RIS-enhanced networks, we discuss the following topics: (1) channel
models, (2) performance analysis, and (3) benchmark schemes.

A. Channel Models

1) Path Loss Models

Some research contributions on the path loss for RISs are available in [7], [64] and [50], which
showed that the power scattered by an RIS is usually formulated in terms of an integral that
accounts, by leveraging the Huygens principle, for the impact of the entire surface in the free-
space scenario, where scattering, shadowing, and reflection are ignored. Closed-form expressions
of the integral are, on the other hand, difficult to obtain, except for some asymptotic regimes,
which correspond to viewing the RIS as electrically small and electrically large (with respect
to the wavelength and the transmission distances). It is worth mentioning, in addition, that the
path-loss model depends on the particular phase gradient applied by the RIS. Notably, the scaling
laws can be different if the RIS operates as an anomalous reflector and as a focusing lens. In
the following, we briefly discuss two scaling laws that have recently been reported for RISs
that operate as anomalous reflectors (described in the previous section). Further information and
details can be found in [7] and [50].
 23

• Electrically Small RISs: In this asymptotic regime, the RIS is assumed to be relatively
small in size compared with the transmission distances. In this regime, the RIS can be
approximated as a small-size scatterer. In general, the path-loss scales with the reciprocal
of the product of the distance between the transmitter and the center of the RIS and the
distance between the center of the RIS and the receiver. In addition, the received power
usually increases with the size of the RIS. The received power is usually maximized in the
direction of anomalous reflection, where the path loss through the RIS follows the “product
of distances” models, which can be formulated as:

L(dSR , dRU ) ≈ λS (dSR dRU )−1 , (5)

where λS denotes the coefficient of the electrically small scenario, dSR and dRU represent
the distance of source-RIS and RIS-user links, respectively. A detailed discussion is given
in [50, Secs. IV-A, IV-B, IV-C].
• Electrically Large RISs: In this asymptotic regime, the RIS is assumed to be large (ideally
infinitely large) in size compared with the transmission distances and the wavelength. In
this regime, the RIS can be approximated as a large flat mirror. Let us denote by x0 the
point of the RIS (if it exists) at which the first-order derivative of the total phase response
of the combined incident signal, reflected signal, and the surface reflection coefficient of
the RIS is equal to zero. In general, the path-loss asymptotically scales with the reciprocal
of a weighted sum of the distance between the transmitter and x0 and the distance between
x0 and the receiver. In addition, the received power is not dependent on the size of the RIS,
which is viewed as asymptotically infinite. This result substantiates the fact that the power
scaling law of the RIS is physically correct, since it does not grow to infinity as the size
of the RIS goes to infinity. This is because the scaling law and the behavior of the RIS are
different with respect to the electrically small regime. In this case, the path loss through
the RIS follows the “sum of distances” models, which can be approximated as:

L(dSR , dRU ) ≈ λL (dSR + dRU )−1 , (6)

where λL denotes the coefficient of the electrically large scenario. A detailed discussion is
 24

given in [50, Secs. IV-A, IV-B, IV-C].

2) Spatial Models
Stochastic geometry (SG) tools are capable of capturing the location randomness of users thus
enabling the derivation of computable or closed-form expressions of key performance metrics.
Specifically, several spatial processes exist for modeling the locations of users in different wireless
networks, i.e., the homogeneous Poisson point process (HPPP) [65], [66], the Poisson cluster
process (PCP) [67], [68], the Binomial point process (BPP), as well as the Hard core point
process (HCPP) [69]. We list some promising approaches for analyzing the performance of
RIS-enhanced networks by using SG in Table IV. In [70], the RIS elements are employed on
obstacles, and it is assumed that randomly distributed users are located in the serving area of
the RISs. In [71], the objects are modeled by a modified random line process of fixed length
and with random orientations and locations. Therein, the probability that a randomly distributed
object that is coated with an RIS acts as a reflector for a given pair of transmitter and receiver
was investigated. In [72], PCP was invoked in the RIS-enhanced large-scale networks, where the
angle of reflection is constrained by the angle of incidence. Therefore, the randomly distributed
users and BSs are located at the same side of the RIS.

Table IV: Summary of RIS-enhanced SG Networks

Approaches Advantages Disadvantages Ref.

RIS-enhanced HPPP Fairness-oriented design Restricted user distribution [5]
RIS-enhanced two layer HPPP Coverage-hole enhancement RISs are deployed at obstacles [70]
HPPP conditioned on angle Practical design Not tractable [73]
RIS-enhanced HPPP conditioned on angle Practical design for RIS-enhanced networks Complicated –

3) Small-Scale Fading Models

Currently, two main approaches have been used for analyzing the performance of RIS-aided
systems in small-scale fading channels: (1) the central-limit-theorem-based (CLT-based) distri-
bution and (2) the use of approximated distributions.

• CLT-based Distribution: Let us consider a single-antenna BS that communicates with a

single-antenna user with the aid of an RIS of N elements. If the two received signals, from
the BS and from the RIS, can be coherently combined, the effective channel power gain is
 25

given by
2
rH Φg + h ,

s.t. β1 , · · · , βN = 1 (7)

φ1 , · · · , φN ∈ [0, 2π) ,

where h ∈ C1×1 , g ∈ CM ×1 , r ∈ CM ×1 denote the channels of the BS-user, BS-RIS, and

RIS-user links, respectively. Φ = diag β1 ejφ1 , β2 ejφ2 , . . . , βN ejφN denotes the reflection-
coefficient matrix of the RIS, where {β1 , β2 . . . , βN } and {φ1 , φ2 . . . , φN } represent the
amplitude coefficients and phase shifts of the RIS elements, respectively. In this setup, the
CLT-based technique stands as an approximation tool for analyzing the performance in the
low-medium-SNR regime. This is due to the fact that the distribution of the probability
density function (PDF) in the range 0 to 0+ is not precise by using the CLT [74]. In
Rayleigh fading channels, the distribution of an RIS-enhanced link follows a modified
Bessel function [74]. Since both transmitter and RISs are part of the infrastructure, and
the RISs are typically positioned to exploit the LoS path with respect to the locations of
the transmitters and the receivers for increasing the received signal power, Zhang et al. [56]
studied Rician fading channels, and the analysis showed that the signal power follows a non-
central chi-squared distribution with two degrees of freedom. Ding and Poor [75] proposed
an RIS-enhanced network, where RISs are utilized for effectively aligning the directions of
the users’ channel gains. By utilizing the CLT-based technique, Cheng et al. [76] studied
the multi-RIS network, where the channel distributions were investigated with or without
BS-user links.
• Approximated Distribution: The exact distribution of the received SNR of the signal
reflected from an RIS is non-trivial to be obtained, and hence the use of approximated
distributions is often necessary. Qian et al. [55] proposed a simple approximated distribution
of the received SNR, and proved that the received SNR can be approximated by two (or
one) Gamma random variables and the sum of two scaled non-central chi-square random
variables. A prioritized signal enhancement design was proposed by Hou et al. [77], where
both the outage performance and ergodic rate of the user with the best channel gain were
calculated. Lyu and Zhang [78] proposed a single-input and single-output (SISO) network
 26

with multiple randomly deployed RISs, and showed that the exact distribution in terms of
received signal power can be approximated by a Gamma distribution. Makarfi et al. [26]
proposed an RIS-enhanced network, whose equivalent channel is modeled by the Fisher-
Snedecor F distribution.

Based on the above mentioned contributions [26], [56], [74], [75], [77], [78], where only
approximated channel distributions are obtained, the exact channel distribution of RIS-enhanced
networks is still an open problem. Based on recent research results, for example, a fundamental
limitation lies in the calculation of the diversity order of RIS-enhanced networks under ideal
operating conditions and in the presence of hardware limitations, e.g., quantized phase shifts.
1
For example, the diversity order obtained by using the CLT-based distribution [74] is 2
in the
N
high-SNR regime, whereas the diversity order is 3
if an approximation based on the Gamma
distribution is used, where N denotes the number of RIS elements. However, the CLT-based and
Gamma-based distributions are not exact, making the performance analysis of RIS-enhanced
networks an interesting problem for future research. Furthermore, since the exact distribution
contains higher-order components, which approach zero in the high-SNR regime, most of the
previous contributions [4], [79] adopt the approximated distribution method for modeling small-
scale fading channels, and the exact distribution of RIS-enhanced networks is still an open
problem. For example, recent exact results on the impact of phase noise on the diversity order
of RIS-enhanced transmission can be found in [80].

B. Performance Analysis

In this subsection, we briefly discuss currently available papers on the performance analysis
of RISs that are realized as large arrays of tiny and inexpensive antennas whose phase response
is locally optimized. By offering extra diversity in the spatial domain, multi-antenna techniques
are of significant importance. The application of multi-antenna enhanced RIS networks has
attracted substantial interest from academia [5], [77], [78] and industry [12], [81], [82]. Given
the increasing number of research contributions on RISs, its advantages are becoming more
clear, especially in terms of spectral efficiency (SE) and energy efficiency (EE) enhancement.
There are several key challenges for performance analysis in RIS-enhanced networks. One of
the main challenges is to evaluate the exact distributions of the cascade channels between the
 27

BS and users through RISs. Another challenge is evaluating the effective channel gain after
passive beamforming at the RIS. Table V summarizes the existing contributions on RISs with
multiple antennas and illustrates their comparison. RIS-enhanced single user networks have been
analyzed in Section III.A. Hence, we turn out attention to RIS-enhanced multi-user networks.
A prioritized signal-enhancement-based (SEB) was proposed by Hou et al. [77], where passive
beamforming is designed for the user with the best channel gain, and all the other users rely on
RIS-enhanced beamforming.

Table V: Important contributions on RIS-enhanced networks. “DL” and “UL” represent downlink
and uplink, respectively. The “sum-rate gain” implies that the gain brought by invoking RIS
technique

Ref. Scenarios Direction Users Main Objectives Techniques

[5] MIMO DL Multiple users OP and throughput Fairness SEB
[12] SISO DL Single user sum-rate gain Compare with relay
[26] SISO UL Single user OP and throughput Effective channel gain
[74] SISO DL Single user Effective channel gain Compare with random phase shifting
[56] SISO DL Single user Effective channel gain SEB
[75] SISO DL Single user OP SEB
[77] SISO DL Multiple users OP and throughput Prioritized SEB
[78] SISO DL Single user Effective channel gain SEB
[55] MIMO DL Single users Effective channel gain Random matrix theory and CLT
[83] MIMO DL Multiple users sum-rate gain SEB
[82] MIMO DL Multiple users Interference cancellation SCB and less constraint at RAs
[84] SISO UL Multiple users Sum-rate Minimum required finite resolution
[85] MISO DL Multiple users Sum-rate Multi-RIS distribution
[86] MISO DL Multiple users Sum-rate Discrete phase shifts

• RIS-enhanced Signal Enhancement Designs: By assuming that multiple waves are co-
phased at the users, the received signal can be significantly enhanced, which leads to the
following optimization problem:
2
max rH Φg + h

s.t. β1 , · · · , βN = 1 (8)

φ1 , · · · , φN ∈ [0, 2π) .

In order to further enhance the SE of RIS-enhanced networks, multiple antenna techniques

can be employed at both the BS and users. Yuan et al. [85] proposed a cognitive-radio-
based RIS-enhanced multiple-input and single-output (MISO) network, where both perfect
and imperfect channel state information (CSI) setups were considered. However, in many
research works, continuous amplitude coefficients and phase shifts are assumed at the RISs,
 28

whilst in practice the phase shifts of RISs may not be continuous. Thus, You et al. [86]
proposed a discrete phase shifts model for a MISO enhanced RIS network. Zhang et al. [84]
then evaluated the required number of bits for finite-resolution RISs in an uplink SISO
network. Hou et al. [5] investigated an RIS-enhanced MIMO network, where a fairness
oriented design was considered by applying SG tools for modeling the impact of the users’
locations.
• RIS-enhanced Signal Cancellation Designs: Another application of deploying RISs in
wireless networks is signal cancellation, where the reflected signals and the direct signals
can be destructively combined. The corresponding optimization problem can be formulated
as follows:
2
min rH Φg + hI

s.t. β1 , · · · , βN ≤ 1 (9)

φ1 , · · · , φN ∈ [0, 2π) .

where hI denotes the aggregate interference signals from other-cell BSs. By assuming that
both the inter-cell and intra-cell interferences are perfectly known, the optimal solution
to (9) is to adjust both the signal phase and amplitude coefficients of the BS-RIS-user
links to the opposite of the effective interference links with the same amplitude. By doing
so, some promising applications can be realized, e.g. RIS-enhanced PLS and interference
cancellation. On the one hand, by assuming that perfect CSI is available at the RIS controller,
the inter-cell and intra-cell interferences can be eliminated. On the other hand, considering
the PLS requirements, RISs also stand as a potential solution for cooperative jamming
techniques, i.e., RISs act as artificial noise sources. By adopting this approach, several
contributions have been made. Hou et al. [82] proposed an RIS-enhanced interference
cancellation technique in a MIMO network, where the inter-cluster interference can be
eliminated without active beamforming weights and detection vectors. Furthermore, this
work can be adopted for application to coordinated multi-point (CoMP) networks for inter-
cell interference cancellation in cellular networks. Shi et al. [87] investigated an RIS-
enhanced secure beamforming technique, where the secrecy rate of the legitimate user was
derived. Lyu et al. [88] investigated an RIS jamming scenario, where RISs act as jammers
 29

for attacking legitimate communications without using any internal energy.

C. Benchmark Schemes

In order to assess the advantages and limitations of RISs, two benchmark transmission tech-
nologies are usually considered: 1) surfaces with random phase shifts; and 2) relay networks.

• Random Phase Shifts: RISs are capable of shifting the phase of the incident signal, and
hence multiple signals can be boosted or eliminated at the user side or at the BS side.
Hence, a well-accepted benchmark scheme to quantify the performance enhancement by
RIS elements is given by a surface that is not configurable and that can ideally be modeled
by a surface with random phase shifts [74].
• Relay Networks: Generally speaking, relay-aided networks can be classified into two pairs
of classic relaying protocols, which are 1) FD and HD relay networks; and 2) AF and DF
relay networks. By assuming that the optimal power split strategies of both the AF and DF
relays are employed, the performance gain between RIS-enhanced and relay-aided networks
can be compared. Specifically, Bjornson et al. [12] compared the achievable data rate of
both RIS-enhanced and DF-relay-aided SISO network, where the BS-user links are blocked.
It was pointed out that when the number of tunable elements of the RISs is large enough,
an RIS-enhanced network is capable of outperforming a DF-relay-aided network. In an
effort to provide a comprehensive analysis for both RIS-enhanced and relay-aided networks,
Ntontin et al. [83] compared the system performance of classic maximal ratio transmission
(MRT) and maximal ratio combining (MRC) techniques. Fig. 11 illustrates the potential
benefits of RIS-enhanced networks compared with both HD-relay and FD-relay networks
in terms of network throughput [77]. Here, the performance of HD-relaying is obtained for
an equal time-split ratio. We can see that the network throughput gap between the RIS-
enhanced network and the other pair of relay aided networks becomes smaller, when the
number of RIS elements increases. For example, when the number of RIS elements N = 23
and the transmit power P = 25 dBm, the proposed RIS-enhanced network is capable of
outperforming both FD and HD relay aided networks, which indicates that the RIS-enhanced
network becomes more competitive, when the number of RIS elements is large enough.
 30

13
RIS-aided network
FD-relay network
12
HD-relay network
11

Spectral efficiency (BPCU)

10
P=30 dBm
9

6
P=25 dBm

3
5 10 15 20 25 30 35 40 45 50
Number of RISs

Figure 11: Spectral efficiency of RIS-enhanced, FD-relay as well as HD-relay networks versus
the number of RIS elements. Please refer to [77] for simulation parameters.

Table VI provides a comparison between RIS-enhanced and relay aided networks in terms
of advantages and limitations.

Table VI: Comparison of the RIS-enhanced and Benchmarks. The “Power” Implies Additional
Power Supply at the RISs or at the Relay. The “CT” Denotes Concurrent Transmission

Mode Pros Cons Delay Power Interference CT

RIS-enhanced High EE, simple device CSI must be perfectly known 7 7 7 X
HD relaying under AF protocol No decoding at relay Noise is amplified X X 7 7
HD relaying under DF protocol No self-interference Latency is high X X 7 7
FD relaying under AF protocol No decoding at relay Noise and interference are amplified 7 X Self-interference X
FD relaying under DF protocol No latency Rate ceiling occurs 7 X Self-interference X
MIMO relay High SE High cost, difficult to realize at mmWave 7 X X X

D. Discussions and Outlook

Although previous research contributions have analyzed the approximated performance of

RIS-enhanced networks, there are still three major open research problems.
1) Path loss experiments for outdoor scenarios: Since only [64] reported experimental measure-
ments of the path loss in free-space environments, the development of experimentally-validated
path-loss models in outdoor scenarios is an open research issue, especially in the presence of
reflecting and scattering objects.
2) Exact distributions: Current research methods for performance analysis are based on the
CLT-based distribution and approximated distribution [26], [76], which are, however, accurate
 31

only in the high-SNR regime. More advanced and accurate analytical models are needed for
analyzing the diversity order and the network performance in the low-SNR regime.
3) Integrated application scenarios: To serve the desired users in different scenarios, signal
enhancement and signal cancellation designs were well investigated in recent research works [77],
[82], [83]. However, the desired signals and interference signals can be simultaneously enhanced
and mitigated, which constitutes an important future direction.

IV. RIS B EAMFORMING AND R ESOURCE A LLOCATION

As described in the previous section, deploying RISs enables high performance enhancements
in wireless networks. Motivated by this benefit, how to jointly design the transmit and passive
beamforming as well as the optimal allocation of the wireless resources has become an important
task for RIS-enhanced wireless networks. In this section, we first discuss the information-theoretic
performance limits of RISs. Then, we review recent research contributions with a particular
focus on the joint beamforming optimization and the resource allocation design. Along the
literature review, representative mathematical tools for facilitating the two types of design are
also discussed along with their benefits and drawbacks.

A. Information-Theoretic Perspective

In order to unveil the fundamental performance limits of RISs, several research works [61],
[62], [89] have been devoted to investigating the RIS performance gains from an information-
theoretic perspective.

• Capacity-achieving design: In [61], Karasik et al. derived the capacity for an RIS-aided
single-input and multiple-output (SIMO) communication system. With finite input signal
constellations, it was proved that a joint information encoding scheme at both the transmitted
signals and the RIS configurations is necessary for achieving the channel capacity [61].
Based on this insight, the authors further proposed a practical transmission strategy by utiliz-
ing layered encoding and successive cancellation decoding techniques. Numerical examples
showed that the proposed joint encoding scheme outperforms the conventional max-SNR
scheme.
 32

7 7
NOMA,1-bit OMA, 1-bit
NOMA,2-bit OMA, 2-bit
MR = 32 NOMA, without-RIS MR = 32 OMA, continuous
6 Capaicty region 6
OMA, without-RIS
improvement
Rate region
by increasing MR
5 5
Rate at user 2 (bit/s/Hz)

Rate at user 2 (bit/s/Hz)

improvement
by increasing bits

4 Capacity region 4
improvement Rate region
by increasing bits improvement
3 3 by increasing MR

Capacity region Rate region

2 2
improvement improvement
with RIS with RIS
1 1
MR = 16 MR = 16

0 0
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
Rate at user 1 (bit/s/Hz) Rate at user 1 (bit/s/Hz)

(a) Capacity regions with RIS-NOMA. (b) Rate regions with RIS-OMA.
Figure 12: Illustration of the capacity and rate regions for a random channel realization with
different RIS phase resolutions. MR denotes the number of RIS reflection elements. The full
parameter settings can be found in [62].

• Capacity region characterization: The capacity region of the fading SISO broadcast channel
(BC) was proved to be achieved by invoking the superposition coding (SC) at the transmitter
and the successive interference cancellation (SIC) at multiple receivers [90], i.e., employing
the NOMA transmission. Inspired by these results, Mu et al. [62] investigated the capacity
and rate regions of RIS-enhanced multi-user wireless communication systems achieved by
NOMA and orthogonal multiple access (OMA), respectively. The Pareto boundary of each
region was characterized by solving a series of sum rate maximization problems via the
rate-profile technique. As shown in Fig. 12(a) and Fig. 12(b), by deploying an RIS, the
NOMA capacity region and the OMA rate region can be improved. The capacity/rate regions
are further enlarged by employing more precise phase resolutions and larger numbers of
reflection elements. Furthermore, Zhang et al. [89] investigated the capacity region of the
multiple access channel (MAC) with two users in both centralized and distributed RIS
deployment strategies. The results demonstrated that the centralized RIS deployment strategy
can achieve a higher capacity gain than the distributed strategy.
 33

RIS
controller

RIS
RI S

BS
Direct link User m
Reflect link
User n

Figure 13: Illustration of joint transmit and passive beamforming design.

B. Joint Transmit and Passive Beamforming Design

1) Optimization objectives: As shown in Fig. 13, an RIS is deployed to assist the transmission
between the BS and the users by passively reflecting the signals. The RIS reflection coefficients
can be adjusted by the BS through an RIS controller. Hence, the transmit beamforming at
the BS and the passive beamforming at the RIS have to be jointly designed to improve the
communication performance. In the following, we review the related research works in terms of
their considered optimization objectives.

• Transmit power minimization or EE maximization: In [58], Wu et al. minimized the transmit

power for an RIS-enhanced MISO system in both single-user and multi-user scenarios.
Alternating optimization (AO) based algorithms were developed to find locally-optimal
solutions. The passive beamforming was designed by invoking the semidefinite relaxation
(SDR) approach. It was revealed that an RIS can simultaneously enhance the desired signal
strength and mitigate the interference for the multi-user scenario. The same problem was
further investigated in [91] by taking discrete RIS phase shifts into consideration. The
optimal solutions were derived by applying the branch-and-bound method and exhaustive
search for single-user and multi-user scenarios, respectively. To reduce the computational
complexity, efficient successive refinement algorithms were further designed. It was shown
that the proposed low complexity algorithms are capable of achieving near-optimal perfor-
mances. Han et al. [92] investigated physical-layer broadcasting in an RIS-aided network,
where the total transmit power for satisfying QoS requirements of all users was minimized.
 34

Furthermore, Fu et al. [93] focused on an RIS-enhanced MISO downlink NOMA system,

where the transmit power was minimized by jointly optimizing the transmit and passive
beamforming vectors as well as user decoding orders. In order to overcome the drawbacks
of the SDR approach, an alternating difference-of-convex (DC) method was proposed for
handling the non-convex rank-one constraint. Zhu et al. [94] proposed an improved quasi-
degradation condition for the RIS-enhanced MISO NOMA system to minimize the transmit
power. Under this condition, NOMA is shown to be able to outperform the zero-forcing
beamforming scheme. Zheng et al. [95] compared the minimum transmit power performance
between OMA and NOMA in a discrete phase shift RIS-enhanced SISO system. A near-
optimal solution was obtained by applying the linear approximation initialization and the AO
method. The results showed that NOMA may perform worse than TDMA when the users
have symmetric deployments and rate requirements, which revealed the importance of user
pairing in the RIS-assisted NOMA system. Huang et al. [59] solved the EE maximization
problem in an RIS-enhanced multi-user MISO system, where a realistic RIS power con-
sumption model was proposed in terms of the number of reflection elements and the phase
resolutions at the RIS. An AO-based algorithm was designed for addressing the formulated
problem, where the RIS phase shifts and the BS transmit power allocation were optimized by
invoking the gradient descent method and the fractional programming method. The results
demonstrated that the RIS achieves significantly better EE performance than the traditional
active relay-assisted communication. In contrast to the aforementioned works based on the
perfect CSI assumption, Zhou et al. [96] investigated the robust beamforming design for an
RIS-enhanced multi-user MISO system with imperfect CSI assumptions. The transmit power
was minimized while satisfying QoS requirements of all users under all possible channel
error realizations. The formulated non-convex problem was transformed into a sequence of
semidefinite programming (SDP) subproblems, where the CSI uncertainties and the non-
convex unit-modulus constraints were handled by applying approximation transformations
and the convex-concave procedure [97], respectively. In [98], the robust beamforming design
was further studied under two channel error models, namely the bounded CSI error model
and the statistical CSI error model. The S-procedure and the Bernstein-Type inequality
were applied in each model. The results unveiled that the RIS may degrade the system
 35

performance when the channel error is high. Zappone et al. [99] modeled the overhead for
carrying out the channel estimation and adjusting the RIS. Based on the proposed overhead
model, the EE of an RIS-empowered MIMO communication network was maximized by
jointly optimizing the RIS phase shifts as well as the transmitted and received filters.
• SE or capacity maximization: Yu et al. [100] considered the SE maximization problem in
an RIS-enhanced MISO system. Since the SDR approach only provides an approximate
solution [58], two efficient algorithms were proposed by invoking the fixed-point iteration
method and the manifold optimization method for the passive beamforming design. It was
demonstrated that the proposed algorithms can achieve a higher performance and consume a
lower complexity than the SDR approach. To solve the same problem, a branch-and-bound
algorithm was further proposed by Yu et al. [101], which is capable of obtaining a globally
optimal solution. Though suffering from an extremely high computational complexity, the
proposed branch-and-bound algorithm serves as a performance benchmark to verify the
effectiveness of existing suboptimal algorithms. In [102], Ning et al. focused on an RIS-
enhanced downlink MIMO system to maximize the SE. The passive beamforming was
designed by using the sum of gains maximization principle and by utilizing the alternating
direction method of multipliers. In [103], Ying et al. considered an RIS-enhanced mmWave
hybrid MIMO system, where the phase shifts at the RIS were designed by leveraging the
angle information of the LoS BS-RIS channel. Moreover, Perović et al. [9] investigated
RIS-assisted indoor mmWave communications, where two schemes were developed to
maximize the channel capacity. Zhang et al. [104] characterized the fundamental capacity
limit of RIS-aided point-to-point MIMO communication systems, by jointly optimizing
the RIS reflection coefficients and the MIMO transmit covariance matrix. The communi-
cation capacity was maximized in both the narrowband transmission with frequency-flat
fading channels and the broadband orthogonal frequency division multiplexing (OFDM)
transmission over frequency-selective fading channels. Yang et al. [8] proposed a practical
transmission protocol by considering the channel estimation for an RIS-enhanced OFDM
system under frequency-selective channels. To reduce the required training overhead, the
RIS reflection elements were divided into multiple groups and only the combined channel
of each group has to be estimated. Based on the proposed grouping scheme, the achievable
 36

rate was maximized by jointly optimizing the power allocation at the transmitter and the
phase shifts at the RIS with the AO method. In [105], You et al. designed a transmission
protocol by considering the channel estimation with discrete phase shifts at the RIS. To
reduce the channel estimation errors, a low complexity discrete fourier transform (DFT)-
Hadamard based reflection pattern scheme was developed. The achievable data rate was
further maximized based on the estimated channel by designing the RIS phase shifts using
the proposed successive refinement algorithm.
• Sum rate maximization: In [106], Huang et al. maximized the sum rate in RIS-enhanced
multi-user MISO downlink communications. By employing the zero-forcing precoding at
the BS, the RIS reflection matrix and the power allocation were alternately optimized
with the aid of the majorization-minimization approach. Moreover, the weighted sum rate
maximization problem was investigated by Guo et al. [107]. Under the AO framework, the
transmit beamforming was obtained using the fractional programming method, and three
iterative algorithms were designed for optimizing the reflection coefficients in terms of
different types of RIS reflection elements. In [108], the asymptotic optimal discrete passive
beamforming solution was derived and a modulation scheme was proposed to maximize
the achievable sum rate for the RIS-enhanced multi-user MISO transmission. To further
enhance the performance, Jung et al. [108] designed a joint user scheduling and transmit
power control scheme, which can strike the tradeoff between the rate fairness and the
maximum sum rate among the users. Mu et al. [60] focused their attention on the sum
rate maximization problem in an RIS-enhanced MISO NOMA system with both ideal and
non-ideal assumptions of RIS elements. The non-convex rank-one constraint of the passive
beamforming design was handled by invoking the sequential rank-one constraint relaxation
approach, which is guaranteed to obtain a locally optimal rank-one solution. Instead of
optimizing the passive beamforming with the instantaneous CSI, Zhao et al. proposed a
two-timescale transmission protocol for maximizing the achievable average sum rate in
an RIS-enhanced multi-user system [109]. To reduce the channel training overhead and
complexity, the RIS phase shifts were firstly optimized with the statistical CSI. Then, the
transmit beamforming was designed with the instantaneous CSI and the optimized RIS
phase shifts.
 37

• User fairness: Nadeem et al. [110] maximized the minimum SINR of an RIS-enhanced
MISO system, where the BS-RIS-user link was assumed to be a LoS channel. A determin-
istic approximations was developed for the minimum SINR performance under the optimal
linear precoder by employing the random matrix theory. As a result, the RIS phase shifts
can be optimized using the channel’s large-scale statistics, which can significantly reduce
the overhead of the signal exchange [110]. Yang et al. [111] investigated the max-min
rate problem in an RIS-enhanced NOMA system in both single-antenna and multi-antenna
cases. A combined-channel-strength based user ordering scheme was proposed for achieving
a near-optimal performance.

Table VII: Contributions on joint transmit and passive beamforming design

Ref. Scenarios Phase shifts CSI Main Objectives Techniques/Characteristics

[58] SU/MU-DL-MISO Continuous Perfect Transmit power AO-SDR-based algorithm and two stage algorithm
[91] SU/MU-DL-MISO Discrete Perfect Transmit power Near-optimal ZF-based successive refinement method
[92] MU-DL-MISO Continuous Perfect Transmit power Physical-layer broadcasting
[93] MU-DL-MISO NOMA Continuous Perfect Transmit power Alternating DC method
[94] MU-DL-MISO NOMA Continuous Perfect Transmit power Improved quasi-degradation condition
[95] MU-DL-SISO NOMA Discrete Perfect Transmit power Asymmetric and symmetric user pairing
[59] MU-DL-MISO Continuous Perfect EE RIS power consumption model
[96] MU-DL-MISO Continuous Imperfect Transmit power The worst-case robust beamforming design
[98] MU-DL-MISO Continuous Imperfect Transmit power Imperfect cascaded channels at the transmitter
[99] SU-DL-MIMO Continuous Estimated EE Overhead model for channel estimation and RIS configuration
[100] SU-DL-MISO Continuous Perfect SE Fixed point iteration and manifold optimization methods
[101] SU-DL-MISO Continuous Perfect SE Branch-and-bound algorithm
[102] SU-DL-MIMO Continuous Perfect SE Sum of gains principle
[103] SU-DL-MIMO mmWave Continuous Perfect SE Broadband hybrid beamforming
[9] SU-DL-MIMO mmWave Continuous Perfect Channel capacity RIS-assisted indoor mmWave environments
[104] SU-DL-MIMO Continuous Perfect Channel capacity Frequency-flat fading and frequency-flat selective channels
[8] SU-DL-SISO OFDMA Continuous Estimated Achievable rate RIS element grouping scheme
[105] SU-UL-SISO Discrete Estimated Achievable rate Near-orthogonal DFT-Hadamard based reflection patterns
[106] MU-DL-MISO Continuous Perfect Sum rate Majorization-minimization approach
[107] MU-DL-MISO Continuous/Discrete Perfect Weighted sum rate Iterative algorithms with closed-form expressions
[108] MU-DL-MISO Discrete Perfect Sum rate Interference-free modulation scheme
[60] MU-DL-MISO NOMA Continuous/Discrete Perfect Sum rate Sequential rank-one constraint relaxation approach
[109] MU-DL-MISO Discrete Statistical CSI Sum rate Low channel training overhead
[110] MU-DL-MISO Continuous Statistical CSI Max-min SINR Signal exchange overhead reduction
[111] MU-DL-SISO/MISO NOMA Continuous Perfect Max-min rate Near-optimal NOMA user ordering scheme

All the aforementioned research contributions on the joint transmit and passive beamforming
design are summarized in Table VII. “SU” and “MU” represent single-user and multi-user,
respectively. “DL” and “UL” represent downlink and uplink, respectively.
 38

2) Approaches for passive beamforming design: An example of the joint transmit and passive
beamforming design problem can be formulated as follows

min / max f (w, θ|H) (10a)

w,θ

s.t. w ∈ T , (10b)

θ ∈ P, (10c)

where H denotes the set of given CSI, w and θ denote the transmit beamforming vector and
passive beamforming vector, receptively, T and P denote the corresponding feasible set for w
and θ, respectively. Here, f (w, θ|H) denotes the objective function that depends on w and θ
given the CSI. Let θi = βi ejφi be the ith element of the passive beamforming vector θ. Depending
on the specific implementation of the RIS, three case studies can be considered [60].

• Continuous amplitude and phase shift: In this case, it is assumed that the amplitude and
phase shift of each RIS element can be adjusted continuously, which results in the following
feasible set.

P1 = {βi , φi |βi ∈ [0, 1] , φi ∈ [0, 2π)} . (11)

• Constant amplitude and continuous phase shift: In this case, it is assumed that the
amplitude and phase shift of each RIS element are fixed, e.g., βi = 1, and can be adjusted
continuously, respectively. The corresponding feasible set is given by

P2 = {βi , φi |βi = 1, φi ∈ [0, 2π)} . (12)

• Constant amplitude and discrete phase shift: In this case, it is assumed that the amplitude
and phase shift of each RIS element are fixed, e.g., βi = 1, and can be adjusted based on
a discrete set of values, respectively. The feasible set can be expressed as

P3 = {βi , φi |βi = 1, φi ∈ D} , (13)

where D = 0, 2π , · · · , (N − 1) 2π

N N
and N denotes the number of candidate phase shifts.

It is worth noting that the first two case studies are difficult to realize in practice. Due to the
 39

hardware constraints in fact, it is quite challenging to realize continuous amplitude and phase
shift control. However, the first two case studies can be used to characterize the theoretical
performance upper bounds of RISs.
By inspection of the three case studies just considered, we evince that the joint beamforming
optimization problem is generally a non-convex problem since w and θ are coupled together. The
existing algorithms for solving the non-convex joint beamforming optimization in RIS-assisted
wireless networks are mainly based on the AO method. The advantage of this approach is that,
given the passive beamforming vector, the transmit beamforming design becomes a conventional
problem, which has been extensively investigated. However, the passive beamforming design
under given transmit beamforming vectors is still a non-trivial task to tackle. The main challenges
to solve the problem include the unit modulus constraint and the discrete nature of the feasible set.
In the following, we list the approaches employed in current research contributions for optimizing
the passive beamforming. Table VIII summarizes the characteristics of those approaches.

Table VIII: Summary of approaches to passive beamforming design

Approaches Phase shifts Advantages Disadvantages Ref.

SDR Continuous Relax to convex problem Require rank-one solution construction [58], [94], etc.
Quantization method Discrete Easy to implement Substantial performance loss [111]
Branch-and-bound Continuous/Discrete Optimal solution Relatively high complexity [91], [101]
Iterative algorithms Continuous/Discrete Good complexity-performance tradeoff Performance depends on initialization [91], [93], etc.

• SDR: A common method for handling the non-convex unit-modulus constraint is to trans-
form the passive beamforming vector into a rank-one and positive semi-definite matrix. By
applying the SDR approach, which ignores the non-convex rank-one constraint, the original
non-convex problem becomes a convex SDP problem that can be solved by using many
efficient convex optimization tools. If the obtained matrix solution is not rank-one, Gaussian
randomization methods are usually used. However, the constructed rank-one solution is, in
general, suboptimal and it may be even infeasible for the original passive beamforming
design problem, which not only causes some performance degradation but cannot guarantee
the convergence of the AO-based iterative algorithm either.
• Quantization method: Under the assumption of finite resolution phase shifts, one method
is to relax each discrete phase shift variable φi ∈ 0, 2π , · · · , (N − 1) 2π

N N
into a continuous
 40

variable φi ∈ [0, 2π). After solving the relaxed problem, the obtained continuous solutions
are quantized to their nearest discrete values. However, the quantization method may lead
to a substantial performance loss, especially for low-resolution phase shifts. Additionally, it
is worth mentioning that the non-convex unit-modulus constraints still exist after applying
the continuous relaxations.
• Branch-and-bound: Due to the non-convex nature of the passive beamforming design
problem, it is challenging to obtain the optimal solution with standard convex optimization
techniques. The branch-and-bound approach has been applied for solving polynomial-time
(NP)-hard discrete and combinatorial optimization problems and some specific continuous
optimization problems. For example, the branch-and-bound approach was adopted for deriv-
ing the optimal solution of the discrete passive beamforming design [91] and the continuous
passive beamforming design [101].
• Iterative algorithms: The main idea of iterative algorithms is to obtain a locally optimal or
a high-quality suboptimal solution for the original problem at an acceptable computational
complexity. Some iterative algorithms were developed for the passive beamforming design,
such as the successive refinement algorithm [91], [104], [105], [107], the alternating DC
algorithm [93], the conjugate gradient search [59], the fixed point iteration method and
the manifold optimization method [100], and the sequential rank-one constraint relaxation
approach [60]. It was shown that these proposed iterative algorithms can achieve a good
tradeoff between performance and computational complexity.

C. Resource Management in RIS-enhanced Networks

1) Resource allocation problems: In Fig. 14, a large-scale RIS-assisted transmission scenario

is considered, where multiple BSs serve multiple users with the aid of multiple RISs. In this
context, several key issues need to be discussed.

• Subchannel assignment: The bandwidth efficiency can be improved by properly allocating

users to different subchannels. If the RIS elements are not frequency-selective, one single
common RIS reflection matrix needs to be shared among all subchannels, which makes
the resulting optimization problems challenging to solve. To address this difficulty, Yang et
al. [112] proposed a dynamic passive beamforming scheme, where the resource blocks are
 41

Power
Pow
User 6

User 2 User 5

User 1
Frequency
(i) subchannel assignment

BS 1
RIS 2 BS 3
RIS 1
User 1
User 2 User 5 User 6
User 4
BS 2
User 3
User index

User index
U2 U4 U6 U3 U5
U1 U3 U5 U2 U4
BS 1 BS 2 BS 3 BS index RIS 1 RIS 2 RIS index
(ii) user-BS association (iii) user-RIS association

Figure 14: Illustration of resource management in large-scale RIS-assisted networks.

dynamically allocated to different user groups with varied RIS phase shifts over different
time slots. In [113], Zuo et al. investigated the joint subchannel assignment, power allo-
cation, and passive beamforming design problem in a multi-channel downlink RIS-NOMA
network.
• User-RIS association: In multi-RIS assisted multi-user communications, how to associate
the users to different RISs is an interesting problem. The user-RIS association schemes
in general determine the overall network performance. Considering a multi-RIS assisted
massive MIMO system, Li et al. [114] found that the automatic interference cancelation
property holds for RISs with infinitely large sizes. Then, the considered max-min SINR
problem can be transformed into a user-RIS association problem, which was efficiently
solved by the proposed greedy search algorithm.
• Multi-cell RIS association: In multi-cell scenarios, the optimization problem becomes much
sophisticated for jointly considering the user-BS association, the user-RIS association, and
the subchannel assignment. In some initial works considering multi-cell scenarios [18],
[115], RISs were deployed to enhance the performance of cell-edge users for simultaneously
improving the received desired signal power and mitigating the received interference from
other cells.
 42

2) Approaches to resource allocation problems: Scheduling different users with different

subchannels/RISs/BSs is an NP-hard problem. Though the optimal solution can be obtained
by exhaustively searching over all possible association combinations, it requires a prohibitively
high computational complexity, especially for the large-scale networks in Fig. 14. Therefore,
low complexity and efficient algorithms have to be developed for striking a performance-versus-
complexity tradeoff. In the following, we present some promising approaches, and discuss their
advantages and disadvantages. Table IX summarizes the characteristics of those approaches.

Table IX: Summary of approaches to resource allocation problems

Approaches Advantages Disadvantages Ref.

Binary relaxation Relax to convex feasible set Existence of performance gap -
Matching theory Achieve near-optimal performance Require predefined preference list [113]
Heuristic algorithms Flexible complexity-performance tradeoff Unstable performance [114]

• Binary relaxation: One idea is to relax the binary variable α ∈ {0, 1} into the continuous
variable α ∈ [0, 1], where α represents the user association state. By doing so, the non-
convex integer constraint is relaxed to be convex, and conventional convex optimization
techniques can be applied for solving the relaxed problem. It is worth pointing out that the
relaxed problem might still be non-convex especially when the optimization variables are
highly-coupled. Additional efforts, such as utilizing the successive convex approximation
(SCA) method, are required to obtain an approximate solution. Moreover, this kind of re-
laxation may result in a considerable performance loss between the original integer problem
and the relaxed one.
• Matching theory: Matching theory is a powerful method developed for solving the combi-
natorial user association problems. The user association combinational optimization prob-
lem in RIS-enhanced networks can be modeled as a high dimensional Users-BSs-RISs-
Subchannels matching problem. Though high dimensional matching is NP-hard, it can
be decomposed into several 2D matching subproblems, which can be efficiently solved.
For example, Zuo et al. [113] applied many-to-one matching theory for the subchannel
assignment in an RIS-enhanced NOMA system, which can achieve a near-optimal perfor-
mance. However, leveraging matching theory requires to establish a predefined preference
list for both users and resources. As the channel conditions always fluctuate in RIS-enhanced
 43

networks, these preference lists may need to be dynamically updated, which needs further
investigations.
• Heuristic algorithms: For solving computationally complex problems, one commonly
employed method is to develop heuristic algorithms, where approximate solutions for the
original optimization problem can be obtained with an acceptable computational complexity.
In the aforementioned works, the greedy search based heuristic algorithm was designed for
solving the user-RIS association problem [114]. However, the performance of heuristic
algorithms is sensitive to the designed strategies, which is not always stable.

D. Discussions and Outlook

With the growing number of research contributions on RIS-enhanced communications, the

advantages of RISs have been verified in terms of SE, EE, and user fairness. However, most of
the existing treatises solved the non-convex joint beamforming optimization problem employing
the AO method, which decouples the joint transmit and passive beamforming design into two
subproblems. Though a high-quality suboptimal solution can be obtained, advanced optimization
techniques are required for solving this problem to characterize the optimal performance gain
introduced by RISs, which also provides an important benchmark for verifying the optimality of
any other low complexity suboptimal algorithms. Furthermore, in RIS-enhanced communication
systems, it is known that obtaining accurate CSI is rather challenging due to the nearly passive
working mode of RISs. Investigating the robust joint beamforming design and resource alloca-
tion [96], [98] constitutes an important research direction for practical RIS implementations.
Besides passive beamforming, RISs also introduce the following DoFs, which can be further
exploited to reap the benefits of RISs in the future work.

• RIS deployment design: The reflection link via the RIS experiences a severer path loss
than the direct link. Therefore, the deployment location of the RIS has to be carefully
designed to achieve considerable performance enhancements. How to jointly optimize the
passive beamforming and the deployment location at the RIS as well as the wireless
resource allocation at the AP/BS is a non-trivial task, which deserves further research
efforts. In particular, one of prominent challenges is that the deployment location of the
RIS determines both the path loss and the LoS components of the reflection channels,
 44

which causes the optimization variables to be highly-coupled. Therefore, efficient algorithms

need to be designed. An initial study [116] has investigated the optimal RIS deployment
strategy for both NOMA and OMA transmissions, which showed that asymmetric and
symmetric RIS deployment locations among users are preferable for NOMA and OMA,
respectively. Additionally, the RIS is usually deployed to avoid signal blockage to achieve
a LoS dominated channel, thus having a small path loss. However, such a LoS channel
based RIS deployment strategy may be ineffective, especially for RIS-assisted multi-user
communications. This is because the LoS dominated channels are low-rank, the resulting ill-
condition channel matrices significantly limit the achievable capacity even with a relatively
small path loss. How to deploy the RIS to strike a balance between the path losses and the
Non-LoS (NLoS) components of channels is another interesting problem, which deserves
further research interests.
• Dynamical RIS configuration: In [61], [62], the authors revealed that, from an information-
theoretic perspective, the capacity-achieving transmission schemes need to dynamically
adjusting the RIS. However, most of the existing research contributions assumed that the RIS
reflection coefficients can be only adjusted once for each channel coherence duration. Note
that one of the most typical application scenarios of RISs is to assist transmission of the
users, who are static or moving slowly in the vicinity of RISs. In this case, the duration of
one channel coherence block is usually tens of milliseconds, which is much larger than the
time duration for adjusting the RIS (e.g., 220 microseconds in [45]). Therefore, adjusting RIS
multiple times in one channel coherence duration, namely the dynamic RIS configuration,
is practically valid. This unique characteristic opens up new research opportunities, such
as dynamical passive beamforming and resource allocation schemes, which merit further
investigations.

V. M ACHINE L EARNING FOR RIS- ENHANCED C OMMUNICATION S YSTEMS

ML techniques have gained remarkable interests in wireless communications due to their

learning capability and large search-space [29], [117], [118]. We survey existing research contri-
butions, which apply ML techniques for tackling challenges in RIS-enhanced wireless networks.
 45

Finally, potential research challenges and opportunities of ML-empowered RIS systems are
presented.

A. Motivations and Architecture for Integrating ML in RIS-enhanced Wireless Networks

In this subsection, we first present the challenges of the conventional RIS-enhanced wireless
networks and the motivations for integrating ML in these networks, followed by the system
architecture of ML-empowered RIS-enhanced wireless networks.
To effectively exploit RISs for optimizing wireless networks, preliminary research contribu-
tions have studied a number of technical challenges that include channel estimation/modeling,
joint transmit and passive beamforming design, as well as resource allocation from the BS
to the users. Powerful optimization techniques, such as convex optimization [107], iterative
algorithm [58], gradient descent approach [59], and alternating optimization algorithm [119] have
been adopted for addressing the aforementioned fundamental challenges. Although important
insights have been gained by these research contributions. the following limitations still exist in
conventional RIS-enhanced wireless networks:

• The users are generally assumed to be static for simplicity, i.e., the dynamic mobility of users
is typically ignored. Another limitation in the existing literature is that the communication
environment is assumed to be perfectly known, the differentiation of users’ demand is always
ignored as well.
• The RIS/BS are not capable of learning from the unknown environment or from the limited
feedback of the users. In practical applications of RISs in wireless networks, the system
parameters are treated as random variables, which naturally leads itself to the derivation of
insightful joint probability distributions conditioned on the users’ tele-traffic demand and
mobility. However, this is a highly dynamic stochastic environment, which is difficult for
employing conventional optimization approaches. Additionally, the feedback from the users
is usually resource-hungry and limited, which aggravates the challenges for the conventional
RIS-enhanced wireless networks.
• Finally, instantaneous CSI of all the channels are assumed to be available at the BS.
However, CSI acquisition in RIS-enhanced wireless networks becomes more challenging
 46

than that in the conventional relay systems due to the passive nature of RISs, which also
aggravates the challenge imposed on the conventional RIS-enhanced wireless networks.

B. Deep Learning for RIS-enhanced Communication Systems

Deep learning (DL) has shown great potentials to revolutionize communication systems. It
can be applied in diverse areas of RIS-enhanced wireless networks due to its powerful learning
capabilities [120]–[122].
The acquisition of timely and accurate CSI plays a pivotal role in wireless systems, especially
in MIMO networks. However, CSI acquisition becomes more challenging due to the large number
of antennas in massive MIMO systems [123]. In order to tackle this challenge, a number of
research contributions have adopted DL for estimating the CSI, especially for exploiting CSI
structures beyond linear correlations.
In contrast to the conventional AF relay-aided wireless networks, in RIS-enhanced wireless
networks, the RIS is a passive device, which is not capable of performing active transmis-
sion/reception and signal processing [124]. In an effort to estimate a large number of unknown
parameters caused by RISs, Taha et al. [125] exploited the DL method for learning the RIS
reflection matrices directly from the sampled channel knowledge without any knowledge of
the RIS array geometry. Liu et al. [126] proposed a deep denoising neural network assisted
compressive channel estimation for RIS-assisted mmWave systems with a low training overhead.
Elbir et al. [127] presented a DL framework for channel estimation in the RIS-enhanced MIMO
system. It was shown that the proposed convolutional neural networks (CNNs)-based approach
achieves lower normalized mean-square-error (NMSE) and more robust performance than other
benchmarks.
The data-driven DL approach has the advantage of model-free representation or function
learning such that no explicit models of the complicated wireless channels are needed, at the
expense of requiring large amounts of training data and corresponding computational power.
Thus, the DL method can be adopted for estimating the CSI of RIS-enhanced wireless networks.
Apart from the aforementioned applications of DL in RIS-enhanced wireless networks, Huang
et al. [128] leveraged a deep neural network (DNN)-based approach in the indoor communication
environment for estimating the mapping between a user’s position and the configuration of the
 47

RIS to maximize the received SNR. Additionally, DL can also be applied for learning the optimal
RIS phase shift configuration. Gao et al. [129] proposed a DL-based algorithm for optimally
designing the phase shift of the RIS by training the DL offline. It can be observed that the
proposed unsupervised learning mechanism outperforms the conventional optimization approach
in terms of computational complexity. Khan et al. [130] investigated the signal estimation
and detection in the RIS-enhanced wireless networks. A DL-based approach was proposed for
estimating channels and phase angles from a reflected signal received by an RIS. With the aid
of DL, the bit-error-rate (BER) performance of the system was improved.

C. Reinforcement Learning for RIS-enhanced Communication Systems

Reinforcement learning (RL) is a powerful AI paradigm that can be used to empower agents
by interacting with the environment. More explicitly, by exploiting the learning capability (e.g.,
learning from the environment, learning from the feedback of users, and learning from its
mistakes) of the RL model, the challenges encountered in the conventional RIS-enhanced wireless
networks may be mitigated, thus leading to improved performance.
The core idea of employing RL techniques in the RIS-enhanced wireless networks is that they
allow the BS/RIS to improve their service quality by learning from the environment, from their
historical experience, and from the feedback of the users [131]. More explicitly, RL models can be
used for supporting the BS/RIS (agents) in their interactions with the environment (states), whilst
finding the optimal behavior (actions) of the BS/RIS. Furthermore, the RL model can incorporate
farsighted system evolution (long-term benefits) instead of only focusing on current states. Thus,
it is applied for solving challenging problems in the RIS-enhanced wireless networks.
As illustrated in Fig. 15, the RL algorithms can be divided into three categories, namely,
value-based algorithms, policy-based algorithms, and actor-critic algorithms. Both advantages
and disadvantages exist in the RL algorithms. Since RISs have discrete phase shifts, the DQN
algorithm is more suitable for tackling the corresponding phase shift design problem.
To fully reap the benefits of deploying RISs in wireless networks, the joint transmit and passive
beamforming design of the RIS-enhanced system has been considered in MISO systems [142],
[143], OFDM-based systems [144], wireless security systems [145] and millimeter wave sys-
tems [146] with the aid of RL algorithms. In contrast to the AO method, which alternately
 48

RL algorithm Key features Pros Cons

Utilize the Q-table to Can find the optimal policy Only suitable for scenarios with
Q-learning train Q-value offline without requiring knowledge small state space and action
about the environment space
Allows the agent Allow the agent to choose Only suitable for scenarios with
SARSA to approach the optimal optimal actions at each time step small state space and action
policy online in a real-time fashion space
Invoke NN to as an Can be invoked in scenarios with
DQN approximator of Q- large state space and action space Over estimation on action values
function
Invoke two Q-function to Can solve the problem of over Only discrete state space and
Double DQN select and evaluate Q- estimation action space
values, respectively
Value-based
(only be invoked in scenarios Invoke two NNs to Have a faster convergence rate
with discrete state space and Dueling DQN estimate the action value than DQN model High complexity
action space) and state value

Noisy DQN Adding Gaussian noise Enhance the exploring Only discrete state space and
layer in the DQN model performance action space
Invoke a distribution Higher accuracy than DQN in
Distributional DQN function instead of terms of evaluating Q-function Require the distribution
Expectation to update Q- information of reward function
function
Asynchronous DQN Multi-agent can train Have a faster learning speed than
parallelly DQN High complexity
Reinforcement Learning

Return-based off-policy Only suitable for scenarios with

Retrace algorithm More efficient than Q-learning small state space and action
space
Approximate a stochastic
policy directly using an Sensitive to the parameter
PG independent function Easy for application; setting
approximator with its
own parameters
Policy-based Invoke a loss function to Have a faster convergence rate Can only be invoked in special
(can be invoked in scenarios with TRPO find the optimized than value-based approaches limiting cases; Have a low data
continuous state space and action parameters efficiency
space)
Alternate between Easier and more general than Have a tradeoff between sample
PPO sampling data through TRPO complexity, simplicity and wall-
interaction with the time
environment
Utilize critic to update the
value function parameters, Have the advantages of both
AC while utilize actor to value-based algorithm and Have a lower convergence rate
update the policy policy-based algorithm
parameters
Multi-agent can train Changes in the approximated
Actor-critic A3C parallelly function get propagated much Have a higher variance
(can be invoked in scenarios with more quickly
continuous state space and action Combination of AC and Have the advantages of both Higher complexity; Can not be
space) DDPG DQN policy gradient and DQN invoked in a random scenario
Invoke two critics to Overcome the problem of over
TD3 update the value function estimation Higher complexity
parameters
Based on maximum More stable; Enhance the
SAC entropy exploring performance Higher complexity

Figure 15: Key features, pros and cons of RL algorithms [131]–[141]. (PG represents Policy
Gradient, TRPO denotes Trust Region Policy Optimization, PPO represents Proximal Policy
Optimization, AC denotes Actor-Critic, A3C represents Asynchronous Advantage Actor-Critic,
DDPG denotes Deep Deterministic Policy Gradient, TD3 represents Twin Delayed DDPG, SAC
denotes Soft Actor-Critic)

optimizes the transmit beamforming at the BS and the passive beamforming at the RIS, the
RL-based solution is capable of simultaneously designing them. More explicitly, Huang et
 49

al. [142] applied a deep deterministic policy gradient (DDPG) based algorithm for maximizing
the throughput by utilizing the sum rate as instant rewards for training the DDPG model. In the
proposed model, the continuous transmit beamforming and RIS phase shift were jointly optimized
with low complexity. Taha et al. [144] proposed a deep reinforcement learning (DRL) based
algorithm for maximizing the achievable communication rate by directly optimizing interaction
matrices from the sampled channel knowledge. In the proposed DRL model, only one beam was
utilized for each training episode. Thus, the training overhead was avoided, while the dataset
collection phase was not required. Zhang et al. [146] presented a DRL based algorithm for
maximizing the throughput with both perfect and imperfect CSI. A quantile regression method
was applied for modeling a return distribution for each state-action pair, which modeled the
intrinsic randomness in the MDP interaction between the RIS and communication environment.
Helin et al. [145] considered the application of RISs to PLS. The system secrecy rate was
maximized with the aid of the DRL model by jointly optimizing the beamforming and phase
shift matrices under different users’ QoS requirements and time-varying channel conditions.
Additionally, post-decision state and prioritized experience replay schemes were utilized to
enhance the learning efficiency and secrecy performance.

D. A Novel Architecture of ML-empowered RIS-enhanced Wireless Networks

As a benefit of the ML-based framework, many challenges in conventional wireless commu-

nication networks have been circumvented, leading to enhanced network performance, improved
reliability, and agile adaptivity [117]. Fig. 16 illustrates a novel ML-empowered architecture for
RIS-enhanced wireless networks. As shown in this figure, RISs are installed on the facade
of a building for enhancing the wireless performance [3], [30]. The RIS is linked with a
controller, which controls the reflecting elements for hosting the functionality of phase-shifting
and amplitude absorption. A two-step approach is applied in the proposed ML-empowered RIS-
enhanced wireless networks.

• As illustrated in the data collection, data processing, and feature extraction parts in Fig. 16.
The associated user information (e.g., device type, position, data rate demand, mobility,
caching demand, and computing ability) is collected, stored, and processed. Thus, the users’
behaviors and requirements can be predicted for efficiently deploying and operating the RIS.
 50

Decision making
Data collection Data processing Data warehouse
State&Reward
Phase shift policy
Deployment policy
Action
Resource allocation policy
Reinforcement learning

Data integration
n user 3
User mobility
information from OSN and clearning base

Power domain
station
user 4
user 2
X
user 1 user 3
blocked
user 2
Frequency domain
Dynamic radio resource allocation user 1
User data demand
information from
telecom operator Feature extraction
update
Online
Feature Data
refineme
extraction modeling
nt

Channel state
information from 3D
radio map
Deep Learning

Figure 16: Architecture of ML-empowered RIS-enhanced wireless networks.

Meanwhile, the predicted information can be modified online with the currently collected
data as the input.
• Given the extracted features, adaptive schemes are leveraged for controlling the RISs,
designing the phase shifts, resource allocation, and interference cancellation.

In the proposed ML-empowered RIS-enhanced wireless networks, RISs are capable of rapidly
adapting to the dynamic environment by learning both from the environment and from the
feedback of the users.
Research on the RIS deployment is fundamental and essential. However, there is a paucity of
research on the problem of RIS position determination. Additionally, current research contribu-
tions mainly consider the performance optimization for both single-user and multi-user scenarios
by optimizing the phase shift and/or precoding solutions of the RIS-enhanced communication
systems [6], [18], [110], [147], [148].
Considering the RIS deployments based on the users’ mobility information and particular data
demand implicitly assumes that the long-term movement information and tele-traffic requirement
of users are capable of being learned/predicted. With this proviso, the deployment and control
method of RISs may be designed periodically for maximizing the long-term benefits and hence
reducing the control overhead. By considering the long-term mobility and data demand of users,
 51

RIS-enhanced wireless networks become highly dynamic systems. Meanwhile, in an effort

to maximize the service quality in an unknown environment, RISs are supposed to learn by
interacting with the environment and adapting the control/deployment policy based on the limited
feedback of the users to overcome the uncertainty of the environment.
In this subsection, an RL-based model is presented to jointly design the deployment policy
and phase shift policy of RISs while considering the time-varying data demand of users. As
illustrated in Fig. 17, in the RL-based model, the BS acts as an agent. Since a controller is
installed, the BS can control both resource allocation policy for users and the RIS’s position
and phase shifts. At each timeslot, the BS periodically observes the state of the RIS-enhanced
system. The state space consists of the RIS phase shifts, the allocated power to each user, as
well as coordinates of both the RIS and users. An action is carried out by the BS for selecting
the optimal control policy. The actions contain changing positions and varying phase shifts of
the RIS, as well as varying the allocated power. The key underlying principle of the decision
policy is carrying out an action that makes the DQN model obtain the maximum Q-value at
each time slot. Following each action, the BS receives a penalty/reward, rt , determined by the
formulated objective function.
• State of the RL model. The state space consists of four parts: 1) the current phase shift
of each reflecting element at the RIS; 2) the current 3D position of the RIS; 3) the current
2D position of each user; 4) the current power allocated from the BS to each user.
• Action of the RL model. The action space consists of three parts: 1) the variable quantity
of the n-th reflecting element’s phase shift; 2) the moving direction and distance of the RIS;
3) the variable quantity of the k-th user’s transmit power.
• Reward of the RL model. The reward function is decided by the EE of the system. When
action taken by the BS improves EE, the BS obtains a reward. Otherwise, when a reduction
occurs in EE, the BS receives a penalty.
Fig. 18 characterizes the EE of the system in networks both with and without the assistance
of an RIS. The EE is defined as the ratio between the system achievable sum mean opinion
score (MOS) and the sum energy dissipation in Joule. It was shown that the EE of the system is
enhanced by employing an RIS. The RIS-barycenter line indicates that the RIS is placed at the
barycenter of all users. The RIS-random line indicates that the RIS is randomly deployed, while
 52

Environment

User
mobility Agent
Predicting
Phase shift
Action of RISs
. Q-Value
Execution . New state State st+1 Reward Reward rt+1

Experience
.
Position of Deep neural network
RISs
Experience Replay
User data Memory
demand Training

Action
Section

Figure 17: DRL model for RIS networks.

65

60 RIS-optimal
RIS-barycenter
55 RIS-random
Energy Efficiency [MoS/Joule]

No-RIS
50

45

40

35

30

25

20

15

10
2 4 6 8 10 12 14 16 18 20
Transmit power (dBm)

Figure 18: EE with and without RIS [63].

the RIS-optimal line indicates that the RIS is deployed at the optimal position derived from the
proposed decaying double deep Q-network (D3 QN) algorithm. The results of Fig. 18 confirm
that there exists an optimal position for the RIS as far as the EE of the RIS-enhanced system is
concerned. The performance of the RIS-enhanced system is improved by deploying the RIS at
the optimal position compared to the random deployment strategy and the strategy of placing it
at the barycenter.
 53

E. Other ML Techniques for RIS-enhanced Communication Systems

Besides DL and RL techniques, a range of supervised learning and unsupervised learning

algorithms have been applied in the current generation wireless networks. Thus, these approaches
can also be adopted for tackling challenges in the RIS-enhanced wireless systems.
1) Supervised Learning techniques for RIS-enhanced Communication Systems: As one of the
key branches of ML, powerful supervised learning techniques, such as regression, decision tree
and random forest, K-nearest neighbors (KNN), support vector machines (SVM), and Bayes
classification, have been adopted in diverse scenarios for tackling challenges such as spectrum
sensing [149], traffic/QoE prediction [150], channel/antenna selection [151], and networking
association [152]. In the RIS-enhanced wireless networks, supervised learning algorithms can also
be applied for solving the related problems with sufficient training data due to their advantages
of low complexity and fast convergence speed.
2) Unsupervised Learning techniques for RIS-enhanced Communication Systems: In contrast
to the supervised learning techniques, unsupervised learning methods do not rely on prior
knowledge, which is not data-hungry. Thus, the unsupervised learning algorithms [153] such
as K-means clustering, expectation-maximization, principal component analysis (PCA), and
independent component analysis (ICA) can be applied in the RIS-enhanced wireless networks for
tackling challenges such as BS deployment, user clustering/association [154], channel/network
state detection [155], data aggregation [156], and interference cancellation [157].
3) Federated Learning techniques for RIS-enhanced Communication Systems: Federated learn-
ing, which explores training statistical models directly on remote devices, has become a focal
point in the area of large-scale ML and distributed optimization [158]. Since the federated
learning algorithm is trained at the edge in distributed networks, the inaccessibility of private
data is no longer a problem. Due to its privacy-preserving nature, federated learning algorithms
can be applied for the deployment and design of multiple RISs, where each RIS can act as a
distributed learner, trains its generated data and transfers its local model parameters instead of
the raw training dataset to an aggregating unit. Thus, the deployment and design policy can be
learned in a decentralized manner.
 54

F. Discussions and Outlook

By exploiting ML learning capabilities, the aforementioned challenges encountered in RIS-

enhanced wireless networks may be mitigated. This is due to the reason that RISs can learn by
interacting with the environment and adapt the control/deployment policy based on the feedback
of users to overcome the dynamic/uncertainty of the environment. It can be learned that the RL
model can incorporate farsighted system evolution instead of only focusing on current states,
which can reap long-term benefits for RIS-enhanced wireless networks. However, ML models
will also pose some new challenges, including the layer design for the DL model, the state-
action construction, and the reward function design for the RL model. In addition, simultaneously
employing multiple RISs becomes more challenging due to the cooperation amongst RISs. Hence,
intelligent deployment and design for multi-RIS enhanced wireless networks is highly desired.
Finally, in the current ML models, either discrete or continuous state space is modeled for
optimizing parameters in RIS-enhanced wireless networks, while joint discrete and continuous
parameters exist in the networks. Hence, the joint discrete and continuous state space design in
ML-enabled RIS-enhanced wireless networks is still challenging and constitutes an interesting
topic.

VI. I NTEGRATING RIS S WITH OTHER T ECHNOLOGIES T OWARDS 6G

Current research contributions have proved that RIS-enhanced wireless networks are capable
of obtaining tuned channel gains, improved QoS, enhanced coverage range, and reduced energy
dissipation. These significant performance enhancements can be applied to diverse wireless
communication networks. In this section, we identify the major issues and research opportunities
on the path to 6G associated with the integration of RISs and other emerging technologies, such
as NOMA, PLS, SWIPT, UAV-enabled wireless networks, and autonomous driving networks.

A. NOMA and RIS

In an effort to improve the SE and user connectivity of RIS-enhanced wireless networks,

power-domain NOMA technology is adopted, whose key idea is to superimpose the signals of
two users at different powers for exploiting the spectrum more efficiently by opportunistically
 55

exploring the users’ different channel conditions [66], [159]. Li et al. [160] considered a MISO-
NOMA downlink communication network for minimizing the total transmit power by jointly
designing the transmit precoding vectors and the reflecting coefficient vector. In [8], Yang et al.
jointly optimized the phase shifts matrix of the RIS, as well as the power allocation from the
BS to the users. Thus, the minimum decoding SINR of all users was maximized for optimizing
the throughput of the system by considering user fairness. Ding et al. [75] proposed a novel
design of RIS assisted NOMA networks. It can be observed in [75] that, the directions of users’
channel vectors can be aligned with the aid of the RIS, which emphasizes the importance of
implementing NOMA technology. For an RIS-NOMA system, the core challenge is that the
decoding order is dynamically changed due to the configuration of phase shifts of the RIS.
Mu et al. [60] proposed an RIS-enhanced multiple-antenna NOMA transmission framework to
maximize the throughput of the system by considering the NOMA SIC decoding order condition.
The SCA technique and sequential rank-one constraint relaxation based algorithm were applied
to obtain a locally optimal solution. Ni et al. [161] proposed a resource allocation framework in
multi-cell RIS-NOMA networks, where the achievable sum rate was maximized by solving the
joint optimization problem of user association, sub-channel assignment, power allocation, phase
shifts design, and decoding order determination.
In contrast to the conventional MIMO-NOMA systems, RIS-NOMA technology can overcome
the challenges of the dynamic environment such as random fluctuation of wireless channels,
blocking, and user mobility in an energy-efficient manner. The NOMA system can obtain tuned
channel gains, improved fair resource allocation, enhanced coverage range, and high EE with
the aid of RISs [162]. However, NOMA also gives rise to new challenges when integrated with
RISs. For multi-antenna NOMA transmission, the decoding order is not determined by the users’
channel gains order, since additional decoding rate conditions need to be satisfied to guarantee
successful SIC [60]. Additionally, both the active beamforming and passive phase shift design
affect the decoding order among users and user clustering, which makes the decoding order
design, user clustering, and joint beamforming design highly-coupled in RIS-NOMA networks.
 56

B. PLS and RIS

It has been shown that RISs are capable of simultaneously enhancing the desired signal
power at the intended user and mitigating the interference power at other unintended users [58].
Inspired by this result, several researchers explored the potential performance gain in the context
of PLS by applying the RIS [119], [163]–[169]. Yu et al. [163] considered an RIS-enhanced
multiple-input single-output single eavesdropper (MISOSE) channel, where the eavesdropper
is equipped with a single antenna. The secrecy rate was maximized by jointly optimizing the
transmit beamforming and the RIS phase shift matrix by using an AO-based algorithm. It was
demonstrated that the secrecy performance can be significantly improved by deploying the RIS.
Cui et al. [164] focused on the scenario where the eavesdropper has a better direct channel
condition than that of the legitimate receiver and they are also highly correlated in space, where
the achievable secrecy rate is rather limited in conventional communications. However, it was
shown that the direct signals and the reflected signals can be destructively combined at the
eavesdropper with the aid of RISs, thus significantly improving the secrecy rate. The same
problem was further investigated in [119], [166] by considering a multi-antenna eavesdropper or
legitimate receiver. Chu et al. [165] minimized the transmit power while satisfying the secrecy
rate requirement in the RIS-enhanced MISOSE system. Chen et al. [167] studied the minimum
secrecy rate maximization problem in the RIS-enhanced multi-user multiple-input single-output
multiple eavesdropper (MISOME) system by considering both the continuous and discrete RIS
phase shifts. Injecting artificial noise (AN) is an effective technique to enhance the secrecy
rate [170]. Motivated by this result, Guan et al. [168] examined the effectiveness of employing
AN in an RIS-enhanced MISOME system. The achievable secrecy rate was maximized by jointly
optimizing the transmit beamforming, the passive beamforming, and AN. The results verified
the necessity of using AN, especially for systems with a large number of eavesdroppers. Yu et
al. [169] considered an RIS-enhanced multi-user MISOME system under imperfect CSI with
the aim of maximizing the sum rate, subject to the maximum information leakage constraint.
An efficient AO-based algorithm was developed to optimize the transmit beamforming, the AN
covariance matrix, and the RIS phase shifts. Numerical results showed that significant secrecy
performance gains can be achieved by the RIS.
 57

One critical issue of the RIS-enhanced PLS is that the joint design of transmit and passive
beamforming requires the CSI of both AP-eavesdropper and RIS-eavesdropper links, which is
quite challenging to obtain. This is because besides the nearly passive working mode of RISs, in
practice, eavesdroppers usually stay almost silent to hide their positions and only detect signals in
the air. Therefore, robust joint beamforming designs under the imperfect CSI of the eavesdropper
are essential to guarantee secure transmission. Moreover, given the uncertainty of eavesdroppers,
deploying the RIS may increase the probability of information leakage since the eavesdropper
can receive not only the direct signals from the AP but also the reflected signals from the RIS.
The situation may become even worse when there are multiple cooperative eavesdroppers. In this
case, setting a protected zone to establish an eavesdropper-exclusion area with carefully deployed
RISs would help to enhance the secrecy performance, which deserve further investigations.

C. SWIPT and RIS

SWIPT is an attractive technique for future IoT networks. However, the low EE at the
energy receivers is the main bottleneck in practical SWIPT systems. To overcome this limitation,
deploying the RIS is a promising solution and the RIS-assisted SWIPT has been investigated
in [19]–[22]. In [19], Wu et al. investigated an RIS-assisted SWIPT system, subject to individual
SINR requirements of information receivers. The weighted sum power received by energy
receivers was maximized by jointly optimizing the transmit and passive beamforming with the
proposed AO-based algorithm. Moreover, Tang et al. [20] maximized the minimum received
power among energy receivers. The results in [19] and [20] showed that deploying an RIS
can improve the energy harvesting efficiency. Pan et al. [21] studied the weighted sum rate
maximization problem in the RIS-assisted SWIPT MIMO system, subject to the energy harvesting
requirement of each energy receiver. A block coordinate descent (BCD)-based algorithm was
designed to find a Karush-Kuhn-Tucker (KKT) stationary point of the original optimization
problem. Wu et al. [22] extended the RIS-assisted SWIPT system into a multi-RIS case, where
the transmit power was minimized while satisfying the different QoS constraints at information
users and energy users. It was shown that the RIS enlarges the wireless power transfer range
and reduces the number of required energy beams.
 58

Note that the above research contributions studied performance gain of deploying RISs for
SWIPT mainly from the communication perspective and ignored the EM characteristic of RISs.
As discussed in previous sections, there are substantial differences between the near-field region
and the far-field region of RISs. Therefore, sophisticated EM-based wireless power transfer
models are required for fully reaping the benefits of RISs, which need to be investigated in
future work.

D. UAV and RIS

RISs can be applied in UAV-enabled wireless networks, where UAVs are employed to com-
plement and/or support the existing terrestrial cellular networks [171], [172]. An RIS enhances
the UAV coverage and service quality by compensating for the power loss over long distances,
as well as forming virtual LoS links between UAVs and mobile users via passively reflecting
their received signals. Li et al. [24] jointly optimized the UAV trajectory and the RIS phase
shifts in an iterative manner. It was shown in [24] that, the average achievable rates of the
users were significantly improved with the aid of RISs. On the other hand, RISs can also be
applied in UAV-aided wireless relay networks for enhancing performance. Zhang et al. [173]
considered the effective placement of a single UAV, which was equipped with an RIS to assist the
mmWave downlink transmission while considering user mobility. By jointly designing the UAV
trajectory and the RIS reflection parameters, a virtual LoS connection between the BS and users
was guaranteed. Thus, both the average data rate and the achievable downlink LoS probability
were improved. Yang et al. [174] derived the analytical expressions of outage probability, BER,
and average capacity by approximating the PDF of the instantaneous SNR in RIS-assisted UAV
relaying systems. Mu et al. [175] proposed a novel RIS-aided multi-UAV NOMA transmission
framework, where an RIS was deployed to enhance the desired signal strength between UAVs
with their served ground users while mitigating the inter-UAV interference. Liu et al. [176]
integrated UAVs in RIS-enhanced wireless networks for enhancing the service quality of the
UAV. With the aid of RISs, the energy consumption of the UAV was significantly reduced.
Due to the fact that UAVs are battery-powered, how to reduce their energy consumption is
one of the key challenges. The limited flight-time of UAVs (usually under 30 minutes) hampers
the wide commercial roll-out of UAV-aided networking. By deploying RISs, one can adjust the
 59

RIS phase shift instead of controlling the UAV movement for forming virtual LoS links between
the UAV and the users. Therefore, the UAV can maintain hovering status only when the virtual
LoS links can not be formed even with the aid of the RIS. By invoking the aforementioned
protocol, the total energy consumption of the UAV is minimized, which in turn, maximizes the
UAV endurance. Additionally, by mounting a compact distributed laser charging (DLC) receiver
or wireless power transmission (WPT) receiver antenna inside the UAVs, while a DLC/WPT
transmitter is deployed on the ground or the building roof, the UAVs can be charged as long as
they are flying within the coverage range of the DLC/WPT transmitter [131], [177]. However,
the LoS connection between UAVs and the charging stations/vehicles have to be guaranteed,
which is challenging in the urban scenario when the LoS link between UAVs and charging
stations/vehicles are blocked by high-rise buildings with a high probability. RISs are capable
of smartly reconfiguring the wireless propagation environment by forming virtual LoS links
between UAVs and the charging stations/vehicles via passively reflecting their received signals.
Thus, the quality of charging service is enhanced with the aid of the RIS.
The RIS-enhanced UAV communication scenario is naturally a highly dynamic one, which
falls into the field of ML. When considering both the trajectory design of UAVs and the phase
shift design of RISs, the former one can be formed as a continuous state space while the latter
one is usually formed as a discrete one. Hence, how to simultaneously deal with both continuous
and discrete state space is challenging in ML-empowered RIS-enhanced UAV networks.

E. Autonomous Driving/Connected Vehicles and RIS

RISs can also be deployed in vehicle-to-infrastructure (V2I) assisted autonomous driving

systems, where V2I components are employed to complement the costly onboard units (OBUs).
V2I networks enable autonomous vehicles (AVs) or connected vehicles (CVs) to receive reliable
real-time traffic information from BSs, the information is collected by roadside base stations
(RBSs) and transmitted from RBSs to BSs, which facilitates the interaction among AVs/CVs and
road users, hence enhancing their safety and traffic efficiency [178]–[181]. Since AVs/CVs quality
and reliability are non-negotiable, the AVs/CVs system must be real-time, while the transmission
is supposed to be 100% reliable. However, the service quality of current V2I communication
systems cannot be guaranteed due to the complex channel terrain in the urban environment and
 60

the complexity of road conditions, such as bad weather. Makarfi et al. [182] and Wang [183]
proved that the performance of vehicular networks can be significantly improved with the aid
of RISs. Since the RISs are made of EM material, which can be installed on key surfaces, such
as building facades, highway polls, advertising panels, vehicle windows, and even pedestrians’
clothes. With the massive deployment of RISs, a virtual LoS connection between the BSs and
AVs, as well as between the RSUs and BSs will be guaranteed, which enhances the reliability
of V2I communications.
In RIS-enhanced autonomous driving systems, the driving safety of AVs is the primary
consideration. In terms of safety, collisions have to be avoided, while the traffic rules also need
to obey. Additionally, in RIS-enhanced V2I-assisted autonomous driving systems, the wireless
service quality for AVs has to be guaranteed at each timeslot. Hence, how to improve the
reliability of RIS-enhanced autonomous driving systems is an open and challenging problem.

F. Discussions and Outlook

The studies of RISs have unveiled promising research opportunities, such as NOMA, PLS,
SWIPT, UAV-enabled wireless networks, and autonomous driving networks. Recent research
contributions have proved that RIS-enhanced wireless networks can achieve tuned channel gains,
improved QoS, enhanced coverage range, and reduced energy dissipation. However, the network
and beamforming designs are highly coupled due to dynamic control of the RIS phase shifts,
which brings challenges to these new research directions.

VII. C ONCLUSIONS , C HALLENGES , AND P OTENTIAL S OLUTIONS

A. Concluding remarks

In this paper, recent research works on RIS-enhanced wireless networks proposed for applica-
tions to next-generation networks have been surveyed with an emphasis on the following aspects:
operating principles of RISs, performance evaluation of multi-antenna assisted RIS systems, joint
beamforming design and resource allocation for RISs, ML in RIS-enhanced wireless networks,
and their integration with other key 6G technologies. We have highlighted the advantages and
limitations of employing RISs for communication applications. Further research efforts are
needed to bridge the complex physical models of the different RISs implementations with widely
 61

used communication models. We have considered the performance evaluation of multi-antenna

assisted RIS systems by systematically surveying existing designs for RIS-enhanced wireless
networks from the views of performance analysis, information theory, and optimization. In
addition, we have discussed existing research contributions that applying ML tools for tackling
the dynamic essence of the wireless environment such as random fluctuations of wireless channels
and user mobility. Design guidelines for ML-empowered RIS-enhanced wireless networks have
also been discussed. The benefits of integrating RISs with NOMA, UAV-terrestrial networks,
PLS, SWIPT, and AVs/CVs have been discussed. However, the research of RIS-enhanced wireless
networks is still at a very early stage and there are ample opportunities for important contributions
and advances in this field. Some of them are listed as follows.

B. Challenges and Potential Solutions

1) CSI Acquisition: The acquisition of timely and accurate CSI plays a pivotal role in RIS-
enhanced wireless systems, especially in MIMO-RIS and MISO-RIS networks. The majority
of current research contributions assume perfect CSI available at the BS, RISs controllers, as
well as the users. However, obtaining CSI in RIS-enhanced wireless networks is a non-trivial
task, which requires a non-negligible training overhead. Additionally, in RIS-assisted NOMA
networks, users in each cluster have to share the CSI with each other for implementing SIC.
However, due to the passive characteristic of RISs, the CSI acquisition and exchanging are
non-trivial. Potential solutions can be developed by employing DL methods for exploiting CSI
structures beyond linear correlations.
2) Pareto-Optimization for Satisfying Multiple Objectives: In contrast to the conventional
wireless networks, RIS-enhanced wireless networks are characterized by more rapidly fluctuating
network topologies and more vulnerable communication links. Furthermore, RISs are more likely
to be deployed in an environment with heterogeneous mobility profiles. Hence, the networks
operate in a complex time-variant hybrid environment, where the classic mathematical models
have limited accuracy. Additionally, the challenging optimization problems encountered in RIS-
enhanced wireless networks usually have to satisfy multiple objectives (e.g., delay, throughput,
BER, and power) in order to arrive at an attractive solution. To elaborate, by definition it is
only possible to improve any of the metrics considered at the cost of degrading at least one of
 62

the others. The collection of Pareto-optimal points is referred to as the Pareto front. However,
determining the entire Pareto-front of optimal solutions is still challenging. Potential solutions
may be investigating near-real-time ML-aided Pareto-optimization for tackling the high-dynamic
adaptation of RIS-enhanced wireless networks.

R EFERENCES

[1] K. B. Letaief, W. Chen, Y. Shi, J. Zhang, and Y.-J. A. Zhang, “The roadmap to 6G: AI empowered wireless networks,”

IEEE Commun. Mag., vol. 57, no. 8, pp. 84–90, 2019.

[2] W. Saad, M. Bennis, and M. Chen, “A vision of 6G wireless systems: Applications, trends, technologies, and open

research problems,” IEEE network, vol. 34, no. 3, pp. 134–142, 2020.

[3] Q. Wu and R. Zhang, “Towards smart and reconfigurable environment: Intelligent reflecting surface aided wireless

network,” IEEE Commun. Mag., vol. 58, no. 1, pp. 106–112, 2020.

[4] Y. Cheng, K. H. Li, Y. Liu, K. C. Teh, and H. Vincent Poor, “Downlink and uplink intelligent reflecting surface aided

networks: NOMA and OMA,” IEEE Trans. Wireless Commun., doi: 10.1109/TWC.2021.3054841, 2021.

[5] T. Hou, Y. Liu, Z. Song, X. Sun, Y. Chen, and L. Hanzo, “MIMO assisted networks relying on intelligent reflective

surfaces,” arXiv:1910.00959, 2019.

[6] Y.-C. Liang, R. Long, Q. Zhang, J. Chen, H. V. Cheng, and H. Guo, “Large intelligent surface/antennas (LISA): Making

reflective radios smart,” J. Commun. Inf. Netw., vol. 4, no. 2, pp. 40–50, 2019.

[7] M. Di Renzo, A. Zappone, M. Debbah, M. S. Alouini, C. Yuen, J. de Rosny, and S. Tretyakov, “Smart radio environments

empowered by reconfigurable intelligent surfaces: How it works, state of research, and the road ahead,” IEEE J. Sel. Areas

Commun., vol. 38, no. 11, pp. 2450–2525, 2020.

[8] Y. Yang, B. Zheng, S. Zhang, and R. Zhang, “Intelligent reflecting surface meets OFDM: Protocol design and rate

maximization,” IEEE Trans. Commun., vol. 68, no. 7, pp. 4522–4535, 2020.

[9] N. S. Perović, M. D. Renzo, and M. F. Flanagan, “Channel capacity optimization using reconfigurable intelligent surfaces

in indoor mmwave environments,” in IEEE Proc. of International Commun. Conf. (ICC), 2020, pp. 1–7.

[10] C. Huang, S. Hu, G. C. Alexandropoulos, A. Zappone, C. Yuen, R. Zhang, M. D. Renzo, and M. Debbah, “Holographic

MIMO surfaces for 6G wireless networks: Opportunities, challenges, and trends,” IEEE Wireless Commun., vol. 27, no. 5,

pp. 118–125, 2020.

[11] M. Di Renzo, K. Ntontin, J. Song, F. H. Danufane, X. Qian, F. Lazarakis, J. de Rosny, D. . Phan-Huy, O. Simeone,

R. Zhang, M. Debbah, G. Lerosey, M. Fink, S. Tretyakov, and S. Shamai, “Reconfigurable intelligent surfaces vs. relaying:

Differences, similarities, and performance comparison,” IEEE Open J. Commun. Soc., vol. 1, pp. 798–807, 2020.

[12] E. Björnson, Ö. Özdogan, and E. G. Larsson, “Intelligent reflecting surface versus decode-and-forward: How large surfaces

are needed to beat relaying?” IEEE Wireless Commun. Lett., vol. 9, no. 2, pp. 244–248, 2020.

[13] S. Zhou, W. Xu, K. Wang, M. Di Renzo, and M.-S. Alouini, “Spectral and energy efficiency of IRS-assisted MISO

communication with hardware impairments,” IEEE Wireless Commun. Lett., vol. 9, no. 9, pp. 1366–1369, 2020.
 63

[14] Y. Cao and T. Lv, “Intelligent reflecting surface enhanced resilient design for MEC offloading over millimeter wave

links,” arXiv:1912.06361, 2019.

[15] T. Bai, C. Pan, Y. Deng, M. Elkashlan, A. Nallanathan, and L. Hanzo, “Latency minimization for intelligent reflecting

surface aided mobile edge computing,” IEEE J. Sel. Areas Commun., vol. 38, no. 11, pp. 2666–2682, 2020.

[16] Y. Cao and T. Lv, “Sum rate maximization for reconfigurable intelligent surface assisted device-to-device communications,”

arXiv:2001.03344, 2020.

[17] L. Yang, J. Yang, W. Xie, M. O. Hasna, T. Tsiftsis, and M. Di Renzo, “Secrecy performance analysis of RIS-aided

wireless communication systems,” IEEE Trans. Veh. Technol., vol. 69, no. 10, pp. 12 296–12 300, 2020.

[18] C. Pan, H. Ren, K. Wang, W. Xu, M. Elkashlan, A. Nallanathan, and L. Hanzo, “Multicell MIMO communications

relying on intelligent reflecting surfaces,” IEEE Trans. Wireless Commun., vol. 19, no. 8, pp. 5218–5233, 2020.

[19] Q. Wu and R. Zhang, “Weighted sum power maximization for intelligent reflecting surface aided SWIPT,” IEEE Wireless

Commun. Lett., vol. 9, no. 5, pp. 586–590, 2019.

[20] Y. Tang, G. Ma, H. Xie, J. Xu, and X. Han, “Joint transmit and reflective beamforming design for IRS-assisted multiuser

miso swipt systems,” in IEEE Proc. of International Commun. Conf. (ICC), 2020, pp. 1–6.

[21] C. Pan, H. Ren, K. Wang, M. Elkashlan, A. Nallanathan, J. Wang, and L. Hanzo, “Intelligent reflecting surface enhanced

MIMO broadcasting for simultaneous wireless information and power transfer,” IEEE J. Sel. Areas Commun., vol. 38,

no. 8, pp. 1719–1734, 2020.

[22] Q. Wu and R. Zhang, “Joint active and passive beamforming optimization for intelligent reflecting surface assisted SWIPT

under QoS constraints,” IEEE J. Sel. Areas Commun., vol. 38, no. 8, pp. 1735–1748, 2020.

[23] H. Wang, Z. Zhang, B. Zhu, J. Dang, L. Wu, L. Wang, K. Zhang, and Y. Zhang, “Performance of wireless optical

communication with reconfigurable intelligent surfaces and random obstacles,” arXiv:2001.05715, 2020.

[24] S. Li, B. Duo, X. Yuan, Y.-C. Liang, M. Di Renzo et al., “Reconfigurable intelligent surface assisted UAV communication:

Joint trajectory design and passive beamforming,” IEEE Wireless Commun. Lett., vol. 9, no. 5, pp. 716–720, 2020.

[25] D. Ma, M. Ding, and M. Hassan, “Enhancing cellular communications for UAVs via intelligent reflective surface,” in

Proc. IEEE Wireless Commun. Netw. Conf. (WCNC), 2020, pp. 1–6.

[26] A. U. Makarfi, K. M. Rabie, O. Kaiwartya, O. S. Badarneh, X. Li, and R. Kharel, “Reconfigurable intelligent surface

enabled iot networks in generalized fading channels,” in IEEE Proc. of International Commun. Conf. (ICC), 2020, pp.

1–6.

[27] X. Mu, Y. Liu, L. Guo, J. Lin, and R. Schober, “Intelligent reflecting surface enhanced indoor robot path planning: A

radio map-based approach,” IEEE Trans. Wireless Commun., doi: 10.1109/TWC.2021.3062089, 2021.

[28] C. Liaskos, S. Nie, A. Tsioliaridou, A. Pitsillides, S. Ioannidis, and I. Akyildiz, “A new wireless communication paradigm

through software-controlled metasurfaces,” IEEE Commun. Mag., vol. 56, no. 9, pp. 162–169, 2018.

[29] H. Gacanin and M. Di Renzo, “Wireless 2.0: Toward an intelligent radio environment empowered by reconfigurable

meta-surfaces and artificial intelligence,” IEEE Veh. Technol. Mag., vol. 15, no. 4, pp. 74–82, 2020.

[30] M. Di Renzo, M. Debbah, D.-T. Phan-Huy, A. Zappone, M.-S. Alouini, C. Yuen, V. Sciancalepore, G. C. Alexandropoulos,
 64

J. Hoydis, H. Gacanin et al., “Smart radio environments empowered by AI reconfigurable meta-surfaces: An idea whose

time has come,” EURASIP J. Wireless Commun., 2019.

[31] E. Basar, M. Di Renzo, J. De Rosny, M. Debbah, M.-S. Alouini, and R. Zhang, “Wireless communications through

reconfigurable intelligent surfaces,” IEEE Access, vol. 7, pp. 116 753–116 773, 2019.

[32] S. Gong, X. Lu, D. T. Hoang, D. Niyato, L. Shu, D. I. Kim, and Y. C. Liang, “Towards smart wireless communications

via intelligent reflecting surfaces: A contemporary survey,” IEEE Commun. Surv. Tut., vol. 22, no. 4, pp. 2283–2314,

2020.

[33] Q. Wu, S. Zhang, B. Zheng, C. You, and R. Zhang, “Intelligent reflecting surface aided wireless communications: A

tutorial,” IEEE Trans. Commun., doi: 10.1109/TCOMM.2021.3051897, 2021.

[34] D. R. Smith, O. Yurduseven, L. P. Mancera, P. Bowen, and N. B. Kundtz, “Analysis of a waveguide-fed metasurface

antenna,” Physical Review Applied, vol. 8, no. 5, p. 054048, 2017.

[35] B. O. Zhu, K. Chen, N. Jia, L. Sun, J. Zhao, T. Jiang, and Y. Feng, “Dynamic control of electromagnetic wave propagation

with the equivalent principle inspired tunable metasurface,” Scientific reports, vol. 4, no. 1, pp. 1–7, 2014.

[36] J. Y. Dai, J. Zhao, Q. Cheng, and T. J. Cui, “Independent control of harmonic amplitudes and phases via a time-domain

digital coding metasurface,” Light: Science & Applications, vol. 7, no. 1, p. 90, 2018.

[37] S. R. Rengarajan and Y. Rahmat-Samii, “The field equivalence principle: Illustration of the establishment of the non-

intuitive null fields,” vol. 42, no. 4, pp. 122–128, 2000.

[38] V. S. Asadchy, M. Albooyeh, S. N. Tcvetkova, A. Dı́az-Rubio, Y. Ra’di, and S. Tretyakov, “Perfect control of reflection

and refraction using spatially dispersive metasurfaces,” Physical Review B, vol. 94, no. 7, p. 075142, 2016.

[39] B. H. Fong, J. S. Colburn, J. J. Ottusch, J. L. Visher, and D. F. Sievenpiper, “Scalar and tensor holographic artificial

impedance surfaces,” vol. 58, no. 10, pp. 3212–3221, 2010.

[40] N. M. Estakhri and A. Alù, “Wave-front transformation with gradient metasurfaces,” Physical Review X, vol. 6, no. 4, p.

041008, 2016.

[41] B. O. Zhu, J. Zhao, and Y. Feng, “Active impedance metasurface with full 360 reflection phase tuning,” Scientific reports,

vol. 3, p. 3059, 2013.

[42] N. K. Emani, A. V. Kildishev, V. M. Shalaev, and A. Boltasseva, “Graphene: a dynamic platform for electrical control

of plasmonic resonance,” Nanophotonics, vol. 4, no. 1, pp. 214–223, 2015.

[43] G. Gradoni and M. Di Renzo, “End-to-end mutual-coupling-aware communication model for reconfigurable intelligent

surfaces: An electromagnetic-compliant approach based on mutual impedances,” IEEE Wireless Commun. Lett., doi:

10.1109/LWC.2021.3050826, 2021.

[44] S. Abeywickrama, R. Zhang, Q. Wu, and C. Yuen, “Intelligent reflecting surface: Practical phase shift model and

beamforming optimization,” IEEE Trans. Commun., vol. 68, no. 9, pp. 5849–5863, 2020.

[45] V. Arun and H. Balakrishnan, “Rfocus: Beamforming using thousands of passive antennas,” in 17th USENIX Symposium

on Networked Systems Design and Implementation (NSDI 20), 2020, pp. 1047–1061.

[46] A. Welkie, L. Shangguan, J. Gummeson, W. Hu, and K. Jamieson, “Programmable radio environments for smart spaces,”

in Proceedings of the 16th ACM Workshop on Hot Topics in Networks, 2017, pp. 36–42.
 65

[47] M. Dunna, C. Zhang, D. Sievenpiper, and D. Bharadia, “ScatterMIMO: enabling virtual MIMO with smart surfaces,” in

Proceedings of the 26th Annual International Conference on Mobile Computing and Networking, 2020, pp. 1–14.

[48] R. J. Bell, K. R. Armstrong, C. S. Nichols, and R. W. Bradley, “Generalized laws of refraction and reflection,” JOSA,

vol. 59, no. 2, pp. 187–189, 1969.

[49] J. Huang and J. A. Encinar, “Reflectarray antennas, a john wiley & sons,” Inc., Publication, 2008.

[50] M. Fadil H. Danufaneand Di Renzo, J. de Rosny, and S. Tretyakov, “On the path-loss of reconfigurable intelligent surfaces:

An approach based on green’s theorem applied to vector fields,” arXiv:2007.13158, 2020.

[51] R. C. Johnson, H. A. Ecker, and J. S. Hollis, “Determination of far-field antenna patterns from near-field measurements,”

Proc. IEEE, vol. 61, no. 12, pp. 1668–1694, 1973.

[52] H.-T. Chen, A. J. Taylor, and N. Yu, “A review of metasurfaces: physics and applications,” Reports on progress in physics,

vol. 79, no. 7, p. 076401, 2016.

[53] J. Huang and R. J. Pogorzelski, “A ka-band microstrip reflectarray with elements having variable rotation angles,” IEEE

Trans. Antennas Propagat., vol. 46, no. 5, pp. 650–656, 1998.

[54] J. Xu and Y. Liu, “A novel physics-based channel model for reconfigurable intelligent surface-assisted multi-user

communication systems,” arXiv:2008.00619, 2020.

[55] X. Qian, M. Di Renzo, J. Liu, A. Kammoun, and M. . S. Alouini, “Beamforming through reconfigurable intelligent

surfaces in single-user MIMO systems: SNR distribution and scaling laws in the presence of channel fading and phase

noise,” IEEE Wireless Commun. Lett., vol. 10, no. 1, pp. 77–81, 2021.

[56] Z. Zhang, Y. Cui, F. Yang, and L. Ding, “Analysis and optimization of outage probability in multi-intelligent reflecting

surface-assisted systems,” arXiv:1909.02193, 2019.

[57] A. A. A. Boulogeorgos and A. Alexiou, “Ergodic capacity analysis of reconfigurable intelligent surface assisted wireless

systems,” in 2020 IEEE 3rd 5G World Forum (5GWF), 2020, pp. 395–400.

[58] Q. Wu and R. Zhang, “Intelligent reflecting surface enhanced wireless network via joint active and passive beamforming,”

IEEE Trans. Wireless Commun., vol. 18, no. 11, pp. 5394–5409, 2019.

[59] C. Huang, A. Zappone, G. C. Alexandropoulos, M. Debbah, and C. Yuen, “Reconfigurable intelligent surfaces for energy

efficiency in wireless communication,” IEEE Trans. Wireless Commun., vol. 18, no. 8, pp. 4157–4170, 2019.

[60] X. Mu, Y. Liu, L. Guo, J. Lin, and N. Al-Dhahir, “Exploiting intelligent reflecting surfaces in NOMA networks: Joint

beamforming optimization,” IEEE Trans. Wireless Commun., vol. 19, no. 10, pp. 6884–6898, 2020.

[61] R. Karasik, O. Simeone, M. Di Renzo, and S. Shamai Shitz, “Beyond max-SNR: Joint encoding for reconfigurable

intelligent surfaces,” in 2020 IEEE International Symposium on Information Theory (ISIT), 2020, pp. 2965–2970.

[62] X. Mu, Y. Liu, L. Guo, J. Lin, and N. Al-Dhahir, “Capacity and optimal resource allocation for IRS-assisted multi-user

communication systems,” IEEE Trans. Commun., doi: 10.1109/TCOMM.2021.3062651, 2021.

[63] X. Liu, Y. Liu, Y. Chen, and H. V. Poor, “RIS enhanced massive non-orthogonal multiple access networks: Deployment

and passive beamforming design,” IEEE J. Sel. Areas Commun., vol. 39, no. 4, pp. 1057–1071, 2021.

[64] W. Tang, M. Z. Chen, X. Chen, J. Y. Dai, Y. Han, M. Di Renzo, Y. Zeng, S. Jin, Q. Cheng, and T. J. Cui, “Wireless
 66

communications with reconfigurable intelligent surface: Path loss modeling and experimental measurement,” IEEE Trans.

Wireless Commun., vol. 20, no. 1, pp. 421–439, 2021.

[65] Y. Liu, Z. Ding, M. Elkashlan, and H. V. Poor, “Cooperative non-orthogonal multiple access with simultaneous wireless

information and power transfer,” IEEE J. Sel. Areas Commun., vol. 34, no. 4, pp. 938–953, 2016.

[66] Y. Liu, Z. Qin, M. Elkashlan, Y. Gao, and L. Hanzo, “Enhancing the physical layer security of non-orthogonal multiple

access in large-scale networks,” IEEE Trans. Wireless Commun., vol. 16, no. 3, pp. 1656–1672, 2017.

[67] W. Yi, Y. Liu, and A. Nallanathan, “Modeling and analysis of D2D millimeter-wave networks with poisson cluster

processes,” IEEE Trans. Commun., vol. 65, no. 12, pp. 5574–5588, 2017.

[68] T. Hou, Y. Liu, Z. Song, X. Sun, and Y. Chen, “Exploiting NOMA for UAV communications in large-scale cellular

networks,” IEEE Trans. Commun., vol. 67, no. 10, pp. 6897–6911, 2019.

[69] H. ElSawy, E. Hossain, and M. Haenggi, “Stochastic geometry for modeling, analysis, and design of multi-tier and

cognitive cellular wireless networks: A survey,” IEEE Commun. Surveys Tutorials, vol. 15, no. 3, pp. 996–1019, 2013.

[70] M. A. Kishk and M. Alouini, “Exploiting randomly-located blockages for large-scale deployment of intelligent surfaces,”

IEEE J. Sel. Areas Commun., vol. 39, no. 4, pp. 1043–1056, 2021.

[71] M. Di Renzo and J. Song, “Reflection probability in wireless networks with metasurface-coated environmental objects:

An approach based on random spatial processes,” EURASIP J. Wireless Commun., no. 99, 2019.

[72] C. Zhang, W. Yi, and Y. Liu, “Reconfigurable intelligent surfaces aided multi-cell NOMA networks: A stochastic geometry

model,” arXiv:2008.08457, 2020.

[73] K. S. Ali, M. Haenggi, H. ElSawy, A. Chaaban, and M.-S. Alouini, “Downlink non-orthogonal multiple access (NOMA)

in poisson networks,” IEEE Trans. Commun., vol. 67, no. 2, pp. 1613–1628, 2018.

[74] Z. Ding, R. Schober, and H. V. Poor, “On the impact of phase shifting designs on IRS-NOMA,” IEEE Wireless Commun.

Lett., vol. 9, no. 10, pp. 1596–1600, 2020.

[75] Z. Ding and H. Vincent Poor, “A simple design of IRS-NOMA transmission,” IEEE Commun. Lett., vol. 24, no. 5, pp.

1119–1123, 2020.

[76] Y. Cheng, K. H. Li, Y. Liu, K. C. Teh, and G. K. Karagiannidis, “Non-orthogonal multiple access (NOMA) with multiple

intelligent reflecting surfaces,” arXiv:2011.00211, 2020.

[77] T. Hou, Y. Liu, Z. Song, X. Sun, Y. Chen, and L. Hanzo, “Reconfigurable intelligent surface aided NOMA networks,”

IEEE J. Sel. Areas Commun., vol. 38, no. 11, pp. 2575–2588, 2020.

[78] J. Lyu and R. Zhang, “Spatial throughput characterization for intelligent reflecting surface aided multiuser system,” IEEE

Wireless Commun. Lett., vol. 9, no. 6, pp. 834–838, 2020.

[79] Z. Tang, T. Hou, Y. Liu, J. Zhang, and L. Hanzo, “Physical layer security of intelligent reflective surface aided NOMA

networks,” arXiv:2011.03417, 2020.

[80] P. Xu, G. Chen, Z. Yang, and M. D. Renzo, “Reconfigurable intelligent surfaces-assisted communications with discrete

phase shifts: How many quantization levels are required to achieve full diversity?” IEEE Wireless Commun. Lett., vol. 10,

no. 2, pp. 358–362, 2021.

[81] L. Dai, B. Wang, M. Wang, X. Yang, J. Tan, S. Bi, S. Xu, F. Yang, Z. Chen, M. Di Renzo et al., “Reconfigurable
 67

intelligent surface-based wireless communications: Antenna design, prototyping, and experimental results,” IEEE Access,

vol. 8, pp. 45 913–45 923, 2020.

[82] T. Hou, Y. Liu, Z. Song, X. Sun, and Y. Chen, “MIMO-NOMA networks relying on reconfigurable intelligent surface:

A signal cancellation-based design,” IEEE Trans. Commun., vol. 68, no. 11, pp. 6932–6944, 2020.

[83] K. Ntontin, J. Song, and M. Di Renzo, “Multi-antenna relaying and reconfigurable intelligent surfaces: End-to-end SNR

and achievable rate,” arXiv:1908.07967, 2019.

[84] H. Zhang, B. Di, L. Song, and Z. Han, “Reconfigurable intelligent surfaces assisted communications with limited phase

shifts: How many phase shifts are enough?” IEEE Trans. Veh. Technol., vol. 69, no. 4, pp. 4498–4502, 2020.

[85] J. Yuan, Y. C. Liang, J. Joung, G. Feng, and E. G. Larsson, “Intelligent reflecting surface-assisted cognitive radio system,”

IEEE Trans. Commun., vol. 69, no. 1, pp. 675–687, 2021.

[86] C. You, B. Zheng, and R. Zhang, “Channel estimation and passive beamforming for intelligent reflecting surface: Discrete

phase shift and progressive refinement,” IEEE J. Sel. Areas Commun., vol. 38, no. 11, pp. 2604–2620, 2020.

[87] W. Shi, X. Zhou, L. Jia, Y. Wu, F. Shu, and J. Wang, “Enhanced secure wireless information and power transfer via

intelligent reflecting surface,” vol. 25, no. 4, pp. 1084–1088, 2021.

[88] B. Lyu, D. T. Hoang, S. Gong, D. Niyato, and D. I. Kim, “IRS-based wireless jamming attacks: When jammers can

attack without power,” IEEE Wireless Commun. Lett., vol. 9, no. 10, pp. 1663–1667, 2020.

[89] S. Zhang and R. Zhang, “Intelligent reflecting surface aided multiple access: Capacity region and deployment strategy,”

in 2020 IEEE 21st International Workshop on Signal Processing Advances in Wireless Communications (SPAWC), 2020,

pp. 1–5.

[90] L. Li and A. J. Goldsmith, “Capacity and optimal resource allocation for fading broadcast channels .I. ergodic capacity,”

IEEE Trans. Inf. Theory, vol. 47, no. 3, pp. 1083–1102, 2001.

[91] 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, 2020.

[92] H. Han, J. Zhao, D. Niyato, M. D. Renzo, and Q. Pham, “Intelligent reflecting surface aided network: Power control for

physical-layer broadcasting,” in IEEE Proc. of International Commun. Conf. (ICC), 2020, pp. 1–7.

[93] M. Fu, Y. Zhou, and Y. Shi, “Reconfigurable intelligent surface empowered downlink non-orthogonal multiple access,”

IEEE Trans. Commun., doi: 10.1109/TCOMM.2021.3066587, 2021.

[94] J. Zhu, Y. Huang, J. Wang, K. Navaie, and Z. Ding, “Power efficient IRS-assisted NOMA,” IEEE Trans. Commun.,

vol. 69, no. 2, pp. 900–913, 2021.

[95] B. Zheng, Q. Wu, and R. Zhang, “Intelligent reflecting surface-assisted multiple access with user pairing: NOMA or

OMA?” IEEE Commun. Lett., vol. 24, no. 4, pp. 753–757, 2020.

[96] G. Zhou, C. Pan, H. Ren, K. Wang, M. Di Renzo, and A. Nallanathan, “Robust beamforming design for intelligent

reflecting surface aided miso communication systems,” IEEE Wireless Commun. Lett., vol. 9, no. 10, pp. 1658–1662,

2020.

[97] T. Lipp and S. Boyd, “Variations and extension of the convex–concave procedure,” Optimization and Engineering, vol. 17,

no. 2, pp. 263–287, 2016.

 68

[98] G. Zhou, C. Pan, H. Ren, K. Wang, and A. Nallanathan, “A framework of robust transmission design for IRS-aided MISO

communications with imperfect cascaded channels,” IEEE Trans. Signal Process., vol. 68, pp. 5092–5106, 2020.

[99] A. Zappone, M. Di Renzo, F. Shams, X. Qian, and M. Debbah, “Overhead-aware design of reconfigurable intelligent

surfaces in smart radio environments,” IEEE Trans. Wireless Commun., vol. 20, no. 1, pp. 126–141, 2021.

[100] X. Yu, D. Xu, and R. Schober, “MISO wireless communication systems via intelligent reflecting surfaces,” in Proc.

IEEE/CIC Int. Conf. Commun. China (ICCC), 2019, pp. 735–740.

[101] ——, “Optimal beamforming for MISO communications via intelligent reflecting surfaces,” in 2020 IEEE 21st

International Workshop on Signal Processing Advances in Wireless Communications (SPAWC), 2020, pp. 1–5.

[102] B. Ning, Z. Chen, W. Chen, and J. Fang, “Beamforming optimization for intelligent reflecting surface assisted mimo: A

sum-path-gain maximization approach,” IEEE Wireless Commun. Lett., vol. 9, no. 7, pp. 1105–1109, 2020.

[103] K. Ying, Z. Gao, S. Lyu, Y. Wu, H. Wang, and M. Alouini, “GMD-based hybrid beamforming for large reconfigurable

intelligent surface assisted millimeter-wave massive MIMO,” IEEE Access, vol. 8, pp. 19 530–19 539, 2020.

[104] S. Zhang and R. Zhang, “Capacity characterization for intelligent reflecting surface aided MIMO communication,” IEEE

J. Sel. Areas Commun., vol. 38, no. 8, pp. 1823–1838, 2020.

[105] C. You, B. Zheng, and R. Zhang, “Intelligent reflecting surface with discrete phase shifts: Channel estimation and passive

beamforming,” in IEEE Proc. of International Commun. Conf. (ICC), 2020, pp. 1–6.

[106] C. Huang, A. Zappone, M. Debbah, and C. Yuen, “Achievable rate maximization by passive intelligent mirrors,” in Proc.

IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP), 2018, pp. 3714–3718.

[107] H. Guo, Y.-C. Liang, J. Chen, and E. G. Larsson, “Weighted sum-rate optimization for intelligent reflecting surface

enhanced wireless networks,” arXiv:1905.07920, 2019.

[108] M. Jung, W. Saad, M. Debbah, and C. S. Hong, “On the optimality of reconfigurable intelligent surfaces (RISs): Passive

beamforming, modulation, and resource allocation,” IEEE Trans. Wireless Commun., doi: 10.1109/TWC.2021.3058366,

2021.

[109] M. M. Zhao, Q. Wu, M. J. Zhao, and R. Zhang, “Intelligent reflecting surface enhanced wireless networks: Two-timescale

beamforming optimization,” IEEE Trans. Wireless Commun., vol. 20, no. 1, pp. 2–17, 2021.

[110] A. Kammoun, A. Chaaban, M. Debbah, M.-S. Alouini et al., “Asymptotic max-min SINR analysis of reconfigurable

intelligent surface assisted MISO systems,” IEEE Trans. Wireless Commun., vol. 19, no. 12, pp. 7748–7764, 2020.

[111] G. Yang, X. Xu, and Y.-C. Liang, “Intelligent reflecting surface assisted non-orthogonal multiple access,”

arXiv:1907.03133, 2019.

[112] Y. Yang, S. Zhang, and R. Zhang, “IRS-enhanced OFDMA: Joint resource allocation and passive beamforming

optimization,” IEEE Wireless Commun. Lett., vol. 9, no. 6, pp. 760–764, 2020.

[113] J. Zuo, Y. Liu, Z. Qin, and N. Al-Dhahir, “Resource allocation in intelligent reflecting surface assisted NOMA systems,”

IEEE Trans. Commun., vol. 68, no. 11, pp. 7170–7183, 2020.

[114] X. Li, J. Fang, F. Gao, and H. Li, “Joint active and passive beamforming for intelligent reflecting surface-assisted massive

MIMO systems,” arXiv:1912.00728, 2019.

 69

[115] H. Xie, J. Xu, and Y. F. Liu, “Max-min fairness in IRS-aided multi-cell MISO systems with joint transmit and reflective

beamforming,” IEEE Trans. Wireless Commun., vol. 20, no. 2, pp. 1379–1393, 2021.

[116] X. Mu, Y. Liu, L. Guo, J. Lin, and R. Schober, “Joint deployment and multiple access design for intelligent reflecting

surface assisted networks,” IEEE Trans. Wireless Commun., doi: 10.1109/TWC.2021.3075885, 2020.

[117] X. Liu, M. Chen, Y. Liu, Y. Chen, S. Cui, and L. Hanzo, “Artificial intelligence aided next-generation networks relying

on UAVs,” IEEE Wireless Commun., vol. 28, no. 1, pp. 120–127, 2021.

[118] C.-X. Wang, M. Di Renzo, S. Stanczak, S. Wang, and E. G. Larsson, “Artificial intelligence enabled wireless networking

for 5G and beyond: Recent advances and future challenges,” IEEE Wireless Commun., vol. 27, no. 1, pp. 16–23, 2020.

[119] H. Shen, W. Xu, S. Gong, Z. He, and C. Zhao, “Secrecy rate maximization for intelligent reflecting surface assisted

multi-antenna communications,” IEEE Commun. Lett., vol. 23, no. 9, pp. 1488–1492, 2019.

[120] A. Zappone, M. Di Renzo, M. Debbah, T. T. Lam, and X. Qian, “Model-aided wireless artificial intelligence: Embedding

expert knowledge in deep neural networks for wireless system optimization,” IEEE Veh. Technol. Mag., vol. 14, no. 3,

pp. 60–69, 2019.

[121] Z. Qin, H. Ye, G. Y. Li, and B. F. Juang, “Deep learning in physical layer communications,” IEEE Wireless Commun.,

vol. 26, no. 2, pp. 93–99, 2019.

[122] A. Zappone, M. Di Renzo, and M. Debbah, “Wireless networks design in the era of deep learning: Model-based, AI-based,

or both?” IEEE Trans. Commun., vol. 67, no. 10, pp. 7331–7376, 2019.

[123] C.-K. Wen, W.-T. Shih, and S. Jin, “Deep learning for massive MIMO CSI feedback,” IEEE Wireless Commun. Lett.,

vol. 7, no. 5, pp. 748–751, 2018.

[124] J. Chen, Y.-C. Liang, H. V. Cheng, and W. Yu, “Channel estimation for reconfigurable intelligent surface aided multi-user

MIMO systems,” arXiv:1912.03619, 2019.

[125] A. Taha, M. Alrabeiah, and A. Alkhateeb, “Enabling large intelligent surfaces with compressive sensing and deep learning,”

IEEE Access, vol. 9, pp. 44 304–44 321, 2021.

[126] S. Liu, Z. Gao, J. Zhang, M. Di Renzo, and M. Alouini, “Deep denoising neural network assisted compressive channel

estimation for mmWave intelligent reflecting surfaces,” IEEE Trans. Veh. Technol., vol. 69, no. 8, pp. 9223–9228, 2020.

[127] A. M. Elbir, A. Papazafeiropoulos, P. Kourtessis, and S. Chatzinotas, “Deep channel learning for large intelligent surfaces

aided mm-Wave massive MIMO systems,” IEEE Wireless Commun. Lett., vol. 9, no. 9, pp. 1447–1451, 2020.

[128] C. Huang, G. C. Alexandropoulos, C. Yuen, and M. Debbah, “Indoor signal focusing with deep learning designed

reconfigurable intelligent surfaces,” in 2019 IEEE 20th International Workshop on Signal Processing Advances in Wireless

Communications (SPAWC), 2019, pp. 1–5.

[129] J. Gao, C. Zhong, X. Chen, H. Lin, and Z. Zhang, “Unsupervised learning for passive beamforming,” IEEE Commun.

Lett., vol. 24, no. 5, pp. 1052–1056, 2020.

[130] S. Khan and S. Y. Shin, “Deep-learning-aided detection for reconfigurable intelligent surfaces,” arXiv:1910.09136, 2019.

[131] X. Liu, Y. Liu, Y. Chen, and L. Hanzo, “Trajectory design and power control for multi-UAV assisted wireless networks:

A machine learning approach,” IEEE Trans. Veh. Technol., vol. 68, no. 8, pp. 7957–7969, 2019.
 70

[132] H. Van Hasselt, A. Guez, and D. Silver, “Deep reinforcement learning with double Q-learning,” in Thirtieth AAAI

Conference on Artificial Intelligence, 2016.

[133] Z. Wang, T. Schaul, M. Hessel, H. Van Hasselt, M. Lanctot, and N. De Freitas, “Dueling network architectures for deep

reinforcement learning,” arXiv:1511.06581, 2015.

[134] M. Fortunato, M. G. Azar, B. Piot, J. Menick, I. Osband, A. Graves, V. Mnih, R. Munos, D. Hassabis, O. Pietquin et al.,

“Noisy networks for exploration,” arXiv:1706.10295, 2017.

[135] W. Dabney, M. Rowland, M. G. Bellemare, and R. Munos, “Distributional reinforcement learning with quantile regression,”

in Thirty-Second AAAI Conference on Artificial Intelligence, 2018.

[136] V. Mnih, A. P. Badia, M. Mirza, A. Graves, T. Lillicrap, T. Harley, D. Silver, and K. Kavukcuoglu, “Asynchronous

methods for deep reinforcement learning,” in International conference on machine learning, 2016, pp. 1928–1937.

[137] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,”

arXiv:1707.06347, 2017.

[138] P. Hämäläinen, A. Babadi, X. Ma, and J. Lehtinen, “PPO-CMA: Proximal policy optimization with covariance matrix

adaptation,” arXiv:1810.02541, 2018.

[139] T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver, and D. Wierstra, “Continuous control with

deep reinforcement learning,” arXiv:1509.02971, 2015.

[140] T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, “Soft actor-critic: Off-policy maximum entropy deep reinforcement

learning with a stochastic actor,” arXiv:1801.01290, 2018.

[141] N. C. Luong, D. T. Hoang, S. Gong, D. Niyato, P. Wang, Y.-C. Liang, and D. I. Kim, “Applications of deep reinforcement

learning in communications and networking: A survey,” IEEE Commun. Surv. Tut., vol. 21, no. 4, pp. 3133–3174, 2019.

[142] C. Huang, R. Mo, and C. Yuen, “Reconfigurable intelligent surface assisted multiuser MISO systems exploiting deep

reinforcement learning,” IEEE J. Sel. Areas Commun., vol. 38, no. 8, pp. 1839–1850, 2020.

[143] K. Feng, Q. Wang, X. Li, and C.-K. Wen, “Deep reinforcement learning based intelligent reflecting surface optimization

for MISO communication systems,” IEEE Wireless Commun. Lett., vol. 9, no. 5, pp. 745–749, 2020.

[144] A. Taha, Y. Zhang, F. B. Mismar, and A. Alkhateeb, “Deep reinforcement learning for intelligent reflecting surfaces:

Towards standalone operation,” in 2020 IEEE 21st International Workshop on Signal Processing Advances in Wireless

Communications (SPAWC), 2020, pp. 1–5.

[145] H. Yang, Z. Xiong, J. Zhao, D. Niyato, L. Xiao, and Q. Wu, “Deep reinforcement learning based intelligent reflecting

surface for secure wireless communications,” IEEE Trans. Wireless Commun., doi: 10.1109/TWC.2020.3024860, 2020.

[146] Q. Zhang, W. Saad, and M. Bennis, “Millimeter wave communications with an intelligent reflector: Performance

optimization and distributional reinforcement learning,” arXiv:2002.10572, 2020.

[147] J. Ye, S. Guo, and M. Alouini, “Joint reflecting and precoding designs for SER minimization in reconfigurable intelligent

surfaces assisted MIMO systems,” IEEE Trans. Wireless Commun., vol. 19, no. 8, pp. 5561–5574, 2020.

[148] M. Jung, W. Saad, and G. Kong, “Performance analysis of large intelligent surfaces (LISs): Uplink spectral efficiency

and pilot training,” arXiv:1904.00453, 2019.

 71

[149] K. Umebayashi, M. Kobayashi, and M. López-Benı́tez, “Efficient time domain deterministic-stochastic model of spectrum

usage,” IEEE Trans. Wireless Commun., vol. 17, no. 3, pp. 1518–1527, 2017.

[150] Z. Feng, X. Li, Q. Zhang, and W. Li, “Proactive radio resource optimization with margin prediction: A data mining

approach,” IEEE Trans. Veh. Technol., vol. 66, no. 10, pp. 9050–9060, 2017.

[151] K. G. M. Thilina, E. Hossain, and D. I. Kim, “DCCC-MAC: A dynamic common-control-channel-based MAC protocol

for cellular cognitive radio networks,” IEEE Trans. Veh. Technol., vol. 65, no. 5, pp. 3597–3613, 2015.

[152] P. Abouzar, K. Shafiee, D. G. Michelson, and V. C. Leung, “Action-based scheduling technique for 802.15. 4/zigbee

wireless body area networks,” in IEEE International Symposium on Personal, Indoor and Mobile Radio Communications,

2011, pp. 2188–2192.

[153] J. Wang, C. Jiang, H. Zhang, Y. Ren, K.-C. Chen, and L. Hanzo, “Thirty years of machine learning: the road to pareto-

optimal wireless networks,” IEEE Commun. Surv. Tut., vol. 22, no. 3, pp. 1472–1514, 2020.

[154] X. Liu, Y. Liu, and Y. Chen, “Reinforcement learning in multiple-UAV networks: Deployment and movement design,”

IEEE Trans. Veh. Technol., vol. 68, no. 8, pp. 8036–8049, 2019.

[155] A. Assra, J. Yang, and B. Champagne, “An EM approach for cooperative spectrum sensing in multiantenna CR networks,”

IEEE Trans. Veh. Technol., vol. 65, no. 3, pp. 1229–1243, 2015.

[156] A. Morell, A. Correa, M. Barceló, and J. L. Vicario, “Data aggregation and principal component analysis in WSNs,”

IEEE Trans. Wireless Commun., vol. 15, no. 6, pp. 3908–3919, 2016.

[157] J. Li, H. Zhang, and M. Fan, “Digital self-interference cancellation based on independent component analysis for co-time

co-frequency full-duplex communication systems,” IEEE Access, vol. 5, pp. 10 222–10 231, 2017.

[158] S. Niknam, H. S. Dhillon, and J. H. Reed, “Federated learning for wireless communications: Motivation, opportunities,

and challenges,” IEEE Commun. Mag., vol. 58, no. 6, pp. 46–51, 2020.

[159] Y. Liu, Z. Qin, M. Elkashlan, Z. Ding, A. Nallanathan, and L. Hanzo, “Non-orthogonal multiple access for 5G and

beyond,” Proc. IEEE, vol. 105, no. 12, pp. 2347–2381, 2017.

[160] Y. Li, M. Jiang, Q. Zhang, and J. Qin, “Joint beamforming design in multi-cluster MISO NOMA intelligent reflecting

surface-aided downlink communication networks,” arXiv:1909.06972, 2019.

[161] W. Ni, X. Liu, Y. Liu, H. Tian, and Y. Chen, “Resource allocation for multi-cell IRS-aided NOMA networks,” IEEE

Trans. Wireless Commun., doi: 10.1109/TWC.2021.3057232, 2021.

[162] A. S. d. Sena, D. Carrillo, F. Fang, P. H. J. Nardelli, D. B. d. Costa, U. S. Dias, Z. Ding, C. B. Papadias, and W. Saad, “What

role do intelligent reflecting surfaces play in multi-antenna non-orthogonal multiple access?” IEEE Wireless Commun.,

vol. 27, no. 5, pp. 24–31, 2020.

[163] X. Yu, D. Xu, and R. Schober, “Enabling secure wireless communications via intelligent reflecting surfaces,” in IEEE

Proc. of Global Commun. Conf. (GLOBECOM), 2019, pp. 1–6.

[164] M. Cui, G. Zhang, and R. Zhang, “Secure wireless communication via intelligent reflecting surface,” IEEE Wireless

Commun. Lett., vol. 8, no. 5, pp. 1410–1414, 2019.

[165] Z. Chu, W. Hao, P. Xiao, and J. Shi, “Intelligent reflecting surface aided multi-antenna secure transmission,” IEEE Wireless

Commun. Lett., vol. 9, no. 1, pp. 108–112, 2020.

 72

[166] L. Dong and H. Wang, “Secure MIMO transmission via intelligent reflecting surface,” IEEE Wireless Commun. Lett.,

vol. 9, no. 6, pp. 787–790, 2020.

[167] J. Chen, Y. Liang, Y. Pei, and H. Guo, “Intelligent reflecting surface: A programmable wireless environment for physical

layer security,” IEEE Access, vol. 7, pp. 82 599–82 612, 2019.

[168] X. Guan, Q. Wu, and R. Zhang, “Intelligent reflecting surface assisted secrecy communication: Is artificial noise helpful

or not?” IEEE Wireless Commun. Lett., vol. 9, no. 6, pp. 778–782, 2020.

[169] X. Yu, D. Xu, Y. Sun, D. W. K. Ng, and R. Schober, “Robust and secure wireless communications via intelligent reflecting

surfaces,” IEEE J. Sel. Areas Commun., vol. 38, no. 11, pp. 2637–2652, 2020.

[170] S. Goel and R. Negi, “Guaranteeing secrecy using artificial noise,” IEEE Trans. Wireless Commun., vol. 7, no. 6, pp.

2180–2189, 2008.

[171] Q. Wang, Z. Chen, H. Li, and S. Li, “Joint power and trajectory design for physical-layer secrecy in the UAV-aided

mobile relaying system,” IEEE Access, vol. 6, pp. 62 849–62 855, 2018.

[172] A. Osseiran, F. Boccardi, V. Braun, Kusume et al., “Scenarios for 5G mobile and wireless communications: the vision

of the METIS project,” IEEE Commun. Mag., vol. 52, no. 5, pp. 26–35, 2014.

[173] Q. Zhang, W. Saad, and M. Bennis, “Reflections in the sky: Millimeter wave communication with UAV-carried intelligent

reflectors,” in IEEE Proc. of Global Commun. Conf. (GLOBECOM), 2019, pp. 1–6.

[174] L. Yang, F. Meng, J. Zhang, M. O. Hasna, and M. Di Renzo, “On the performance of RIS-assisted dual-hop UAV

communication systems,” IEEE Trans. Veh. Technol., vol. 69, no. 10, pp. 12 296–12 300, 2020.

[175] X. Mu, Y. Liu, L. Guo, J. Lin, and H. V. Poor, “Intelligent reflecting surface enhanced multi-UAV NOMA networks,”

arXiv:2101.09145, 2021.

[176] X. Liu, Y. Liu, and Y. Chen, “Machine learning empowered trajectory and passive beamforming design in UAV-RIS

wireless networks,” IEEE J. Sel. Areas Commun., doi:10.1109/JSAC.2020.3041401, 2020.

[177] Q. Liu, J. Wu, P. Xia, S. Zhao, W. Chen, Y. Yang, and L. Hanzo, “Charging unplugged: Will distributed laser charging

for mobile wireless power transfer work?” IEEE Veh. Technol. Mag., vol. 11, no. 4, pp. 36–45, 2016.

[178] L. Yao, J. Wang, X. Wang, A. Chen, and Y. Wang, “V2X routing in a VANET based on the hidden Markov model,”

IEEE Trans. Intell. Transport. Syst, vol. 19, no. 3, pp. 889–899, 2018.

[179] R. P. D. Vivacqua, M. Bertozzi, P. Cerri, F. N. Martins, and R. F. Vassallo, “Self-localization based on visual lane marking

maps: An accurate low-cost approach for autonomous driving,” IEEE Trans. Intell. Transport. Syst, vol. 19, no. 2, pp.

582–597, 2018.

[180] X. Liu, Y. Liu, Y. Chen, and L. Hanzo, “Enhancing the fuel-economy of V2I-assisted autonomous driving: A reinforcement

learning approach,” IEEE Trans. Veh. Technol., vol. 69, no. 8, pp. 8329–8342, 2020.

[181] G. Williams, P. Drews, B. Goldfain, J. M. Rehg, and E. A. Theodorou, “Information-theoretic model predictive control:

Theory and applications to autonomous driving,” IEEE Trans. Robot, vol. 34, no. 6, pp. 1603–1622, 2018.

[182] A. Makarf, K. M. Rabie, O. Kaiwartya, K. Adhikari, X. Li, M. Quiroz-Castellanos, and R. Kharel, “Reconfigurable

intelligent surfaces-enabled vehicular networks: A physical layer security perspective,” arXiv:2004.11288, 2020.
 73

[183] J. Wang, W. H. Zhang, X. Bao, T. Song, and C. Pan, “Outage analysis for intelligent reflecting surface assisted vehicular

communication networks,” arXiv:2005.00996, 2020.