1

A Primer on Rate-Splitting Multiple Access:

Tutorial, Myths, and Frequently Asked Questions
Bruno Clerckx, Fellow, IEEE, Yijie Mao, Member, IEEE, Eduard A. Jorswieck, Fellow, IEEE, Jinhong Yuan,
Fellow, IEEE, David J. Love, Fellow, IEEE, Elza Erkip, Fellow, IEEE, and Dusit Niyato, Fellow, IEEE

(Invited Paper)
arXiv:2209.00491v2 [cs.IT] 10 Jan 2023

Abstract—Rate-Splitting Multiple Access (RSMA) has emerged multibeam satellite, space-air-ground integrated networks, un-
as a powerful multiple access, interference management, and manned aerial vehicles, integrated sensing and communications,
multi-user strategy for next generation communication systems. grant-free access, network slicing, cognitive radio, optical/visible
In this tutorial, we depart from the orthogonal multiple access light communications, mobile edge computing, machine/federated
(OMA) versus non-orthogonal multiple access (NOMA) discus- learning, etc. We finally address common myths and answer
sion held in 5G, and the conventional multi-user linear precoding frequently asked questions, opening the discussions to interesting
approach used in space-division multiple access (SDMA), multi- future research avenues. Supported by the numerous benefits and
user and massive MIMO in 4G and 5G, and show how multi-user applications, the tutorial concludes on the underpinning role
communications and multiple access design for 6G and beyond played by RSMA in next generation networks, which should
should be intimately related to the fundamental problem of inspire future research, development, and standardization of
interference management. We start from foundational principles RSMA-aided communication for 6G.
of interference management and rate-splitting, and progressively
delineate RSMA frameworks for downlink, uplink, and multi- Index Terms—Rate-Splitting, Rate-Splitting Multiple Access,
cell networks. We show that, in contrast to past generations of Next Generation Multiple Access, Non Orthogonal Multiple
multiple access techniques (OMA, NOMA, SDMA), RSMA offers Access, Space Division Multiple Access, Multi-user MIMO, In-
numerous benefits: 1) enhanced spectral, energy and computation terference Management, 6G
efficiency; 2) universality by unifying and generalizing OMA,
SDMA, NOMA, physical-layer multicasting, multi-user MIMO I. I NTRODUCTION
under a single framework that holds for any number of antennas
Ommunication systems are inherently multi-user sys-
at each node (SISO, SIMO, MISO, and MIMO settings); 3)
flexibility by coping with any interference levels (from very
weak to very strong), network loads (underloaded, overloaded),
C tems. Multiple access (MA) techniques play the crucial
role of deciding how to make use of the resources (e.g.,
services (unicast, multicast), traffic, user deployments (channel time, frequency, power, antenna, and code) to serve those
directions and strengths); 4) robustness to inaccurate channel
state information (CSI) and resilience to mixed-critical quality of
multiple users. Next generation communications systems, e.g.,
service; 5) reliability under short channel codes and low latency. 6G and beyond, will have to cope with increasing demands for
We then discuss how those benefits translate into numerous high throughput, reliability, heterogeneity of quality of service
opportunities for RSMA in over forty different applications and (QoS), and massive connectivity to satisfy the requirements
scenarios of 6G, e.g., multi-user MIMO with statistical/quantized of further-enhanced mobile broadband (FeMBB), extremely
CSI, FDD/TDD/cell-free massive MIMO, millimeter wave and
terahertz, cooperative relaying, physical layer security, reconfig-
ultra reliable and low-latency communication (eURLLC), ul-
urable intelligent surfaces, cloud-radio access network, internet- tra massive machine type communication (umMTC), mixture
of-things, massive access, joint communication and jamming, thereof, and of new services such as integrated sensing and
non-orthogonal unicast and multicast, multigroup multicast, communications (ISAC), integrated satellite-terrestrial, and ex-
Manuscript received August 25, 2022; accepted Dec 15, 2022. This work
tended reality. To that end, it is critical to understand how next
has been supported in part by the National Nature Science Foundation of generation MA can fulfill those demands and requirements
China under Grant 62201347. (Corresponding author: Yijie Mao.) by going beyond the conventional orthogonal versus non-
B. Clerckx is with the Department of Electrical and Electronic Engineering
at Imperial College London, London SW7 2AZ, UK and with Silicon Austria
orthogonal discussion held in 5G.
Labs (SAL), Graz A-8010, Austria (email: [email protected]). Table I details the main abbreviations used throughout this
Y. Mao is with the School of Information Science and Tech- work.
nology, ShanghaiTech University, Shanghai 201210, China (e-mail:
[email protected]).
E. A. Jorswieck is with the Institute for Communications Technology at A. Beyond Orthogonal versus Non-Orthogonal
TU Braunschweig, Brunswick, Germany (email: [email protected]).
J. Yuan is with the School of Electrical Engineering and Telecommunica- In the past decade, MA schemes have often been classified
tions, University of New South Wales, Sydney, NSW 2052, Australia (e-mail: into two categories, namely orthogonal (serving a single user
[email protected]). per resource) versus non-orthogonal (serving multiple users
D. J. Love is with the School of Electrical and Computer Engineering,
Purdue University, West Lafayette, IN, USA (email: [email protected]). per resource). This classification has triggered the question Is
E. Erkip is with the Electrical and Computer Engineering Department, New a non-orthogonal approach to MA better than an orthogonal
York University Tandon School of Engineering, Brooklyn, NY 11201, USA approach? and has led to the emergence of the rich literature
(email: [email protected]).
D. Niyato is with the School of Computer Engineering, Nanyang Techno- on non-orthogonal multiple access (NOMA) versus orthogonal
logical University, Singapore (email: [email protected]). multiple access (OMA) [1], [2]. This question was motivated
2

TABLE I
L IST OF ABBREVIATIONS .

BC Broadcast Channel MU–LP Multi-User Linear Precoding

CDMA Code Division Multiple Access MU-MIMO Multi-User Multiple-Input Multiple-Output
CoMP Coordinated Multi-Point NOMA Non-Orthogonal Multiple Access
CSI Channel State Information NOUM Non-Orthogonal Unicast and Multicast
CSIT/R Channel State Information at the Transmitter/Receiver OFDMA Orthogonal Frequency Division Multiple Access
C-RAN Cloud-Radio Access Networks OMA Orthogonal Multiple Access
DoF Degree-of-Freedom QoS Quality of Service
DPC Dirty Paper Coding RF Radio Frequency
DPCRS Dirty Paper Coded Rate-Splitting RIS Reconfigurable Intelligent Surfaces
D2D Device-to-Device RS Rate-Splitting
EE Energy Efficiency RSMA Rate-Splitting Multiple Access
(F)eMBB (further-)enhanced Mobile Broadband Service SC Superposition Coding
ER Ergodic Rate SDMA Space Division Multiple Access
ESR Ergodic Sum Rate SE Spectral Efficiency
FDD Frequency Division Duplex SIC Successive Interference Cancellation
FDMA Frequency Division Multiple Access SISO Single-Input Single-Output
F-RAN Fog-Radio Access Networks SNR Signal-to-Noise Ratio
HK Han and Kobayashi SWIPT Simultaneous Wireless Information and Power Transfer
IC Interference Channel SIMO Single-Input Multiple-Output
IRS Intelligent Reconfigurable Surface TDD Time Division Duplex
LLS Link-Level Simulation TDMA Time-Division Multiple Access
MA Multiple Access THz TeraHertz
MAC Multiple Access Channel UAV Unmanned Aerial Vehicles
MIMO Multiple-Input Multiple-Output (e)URLLC (extremely) Ultra-Reliable Low-Latency Communication
MISO Multiple-Input Single-Output VLC Visible Light Communication
MMF Max-Min Fairness V2X Vehicle-to-Everything
(u)mMTC (ultra) massive Machine-Type Communication WSR Weighted Sum Rate
mmWave millimeter-Wave ZFBF Zero-Forcing Beamforming

by the claim that prior generations of cellular communication seen as two extreme interference management strategies where
networks are based on OMA serving multiple users on orthog- the former treats interference as noise and the latter fully
onal resources using time division multiple access (TDMA), decodes interference [8]. An alternative interpretation of this
frequency division multiple access (FDMA), code division difference is to note that SDMA (and other forms of linearly
multiple access (CDMA), or orthogonal frequency division precoded and non-linearly precoded MU-MIMO schemes)
multiple access (OFDMA) [3]. relies on a transmit-side interference cancellation strategy
Classifying MA schemes into non-orthogonal vs orthogonal while NOMA can be seen as a receive-side interference
is however over-simplistic as it does not fully reflect modern cancellation strategy [9]. Such major differences have however
communication designs. Indeed, those systems are equipped not been captured and not been addressed when answering the
with multiple antennas, and spatial domain processing in the aforementioned question, though they lead to drastic perfor-
form of multi-user linear precoding (MU–LP), space division mance and complexity gaps between the two schemes [7].
multiple access (SDMA), multi-user multiple-input multiple- Hence, instead of contrasting orthogonal vs non-orthogonal, a
output (MU-MIMO), massive MIMO, is an integral part of different classification should be considered in next generation
4G and 5G for twenty years. Importantly, MU-LP/SDMA/MU- wireless networks. In this paper, we will show that the fun-
MIMO serves users in a non-orthogonal manner since multiple damental question behind MA design should instead be how
users are allocated different precoders/beams in the same time- to manage multi-user interference? Answering this question
frequency grid and interfere with each other in the same cell. will shed the light on the differences between non-orthogonal
Hence both 4G and 5G already use a combination of or- approaches to MA designs and on a new classification of MA
thogonal (in the time-frequency domains) and non-orthogonal schemes based on how the interference is managed. Even more
(in the spatial domain) approaches in the form of OFDMA importantly, this exercise will bring to light the powerful and
combined with SDMA/MU-MIMO/massive MIMO. It is well newly emerging Rate-Splitting Multiple Access (RSMA) for
documented that non-orthogonality can be beneficial in both downlink and uplink communications.
single and multi-antenna settings [4]–[6].
A major drawback of such classification is that it tends
to amalgamate many different MA schemes under the non- B. Toward Rate-Splitting Multiple Access
orthogonal umbrella without contrasting them or truly un-
derstanding the essence of those schemes. This has caused RSMA refers to a broad class of multi-user schemes whose
unnecessary confusions and misunderstandings in the past few commonality is to rely on the rate-splitting (RS) principle
years [7]. For instance, SDMA and power-domain NOMA1 [10]–[12]. RS consists in splitting a user message (e.g., infor-
are two different non-orthogonal approaches to MA but are mation bits) into two or multiple parts such that each of those
fundamentally different. Indeed, SDMA and NOMA can be parts can be decoded flexibly at one or multiple receivers.
A receiver would have to retrieve each part to reconstruct the
1 In the sequel, we simply use NOMA to refer to power-domain NOMA. original message. A key benefit of RS and its message splitting
 3

capability is to flexibly manage inter-user interference, as it Freedom, DoF2 ) for downlink multi-user mutliple-input single-
will appear clear throughout this paper. output (MISO) and MIMO transmissions with imperfect CSIT
[17]–[22].
Though the RS principle first appeared in the information This literature opens various doors that have a major impact
theoretic literature in the late 70’s and early 80’s, RS (and on MA designs. First, since imperfect CSIT is more general
consequently the emerging RSMA) has received a renewed than perfect CSIT3 , finding efficient schemes for imperfect
interest in the past decade in the broader communication com- CSIT leads to discovering a broader and general class of
munity. What triggered this renewed interest is a different line communication strategies that would subsume perfect CSIT
of research, seemingly unrelated to the MA literature at first, strategies as particular instances. Second, it gives communi-
on understanding the fundamental limits of robust interference cation engineers clear and fundamentally grounded guidelines
management, i.e., how to manage interference in a multi- on how to design robust schemes. Third, it provides refresh-
user multi-antenna communications system in the presence of ing and new thoughts about low complexity non-orthogonal
imperfect channel state information at the transmitter (CSIT) scheme that are applicable and beneficial even in perfect CSIT
[13]. settings. Fourth, it brings RS, originally developed for the two-
user single-antenna interference channel [10], [11], into MU-
MIMO, which was never investigated despite the rich literature
Conventional multi-user multi-antenna approaches such as
on MU-MIMO schemes in the past two decades [6]. Fifth, RS
SDMA/MU-MIMO heavily rely on timely and highly-accurate
serves users in a non-orthogonal manner by partially treating
CSIT. Unfortunately, in practice, CSIT is always imperfect
interference as noise and partially decoding interference. This
due to pilot reuse, channel estimation errors, pilot con-
highlights the usefulness to depart from the extremes of (fully)
tamination, limited and quantized feedback accuracy, de-
treat interference as noise (as in SDMA) and (fully) decode
lay/latency, mobility (due to ever-increasing speeds of ve-
interference (as in NOMA) in multi-antenna networks, but also
hicle/trains/satellite/flying objects and emerging applications
the usefulness to bridge and unify those two extremes [8].
as Vehicle-to-Everything), radio frequency (RF) impairments
Sixth, RS can be seen as a smart combination of transmit-side
(e.g., phase noise), inaccurate calibrations of RF chains, sub-
and receive-side interference cancellation strategy where the
band level estimation [14]. Consequently SDMA/MU-MIMO
contribution of the common stream is adjusted to benefit from
are inherently non-robust. The classical approach to dealing
the best balance between transmit and receive cancellation.
with this practical limitation takes a “robustification” stance –
This departs from the transmit-side only and receive-side only
techniques that have been developed under the assumption of
interference cancellation strategies of SDMA (and the rich
perfect CSIT are tweaked to account for imperfect CSIT [15],
literature on MU-MIMO) and NOMA, respectively [9].
[16].
The above design makes RS a fundamental building block
of a powerful MA framework for downlink and uplink com-
The challenge is that any CSIT inaccuracy results in a munications, namely, RSMA [8], [23].
residual multi-user interference that needs proper management In the downlink, RSMA uses linearly or non-linearly pre-
instead of “robustification”. Despite its significance in realistic coded RS at the transmitter to split each user message into
wireless deployments, the fundamental limits of multi-user one or multiple common messages and a private message.
multi-antenna communications with imperfect CSIT are still The common messages are combined and encoded into com-
an open problem, i.e., the capacity and capacity achieving mon streams for the intended users. Successive interference
schemes remain to be found. Consequently, this lack of un- cancellation (SIC) - or any other forms of joint decoding - is
derstanding of how to design robust communication schemes used at each user to sequentially decode the intended common
has led to modern systems like 4G and 5G designs to be streams (and therefore decode part of the interference). Such
fundamentally designed for perfect CSIT (using SDMA and linearly or non-linearly precoded generalized RSMA has been
treat interference as noise) instead of being designed from demonstrated to be a powerful framework to bridge, reconcile,
scratch to be truly robust to imperfect CSIT [9], [14]. and generalize SDMA, NOMA, OMA, and physical-layer
multicasting and further boost system spectral and energy
Interestingly, we are now in a much better position to design efficiencies for downlink transmissions with both perfect and
truly robust MIMO wireless networks accounting for imperfect partial CSIT [8], [24], [25].
CSIT and its resulting multi-user interference [9], [13], [14]. In the uplink, users split their message into multiple streams
It is indeed known that to benefit from imperfect CSIT and and allocate proper transmit power to each stream. The re-
tackle the multi-user interference, the transmitter should take 2 The Degrees-of-Freedom (DoF), or multiplexing gain, is a first-order
an RS approach that splits the messages into common and approximation of the rate at high signal-to-noise ratio (SNR). It can be
private parts, encodes the common parts into a common viewed as the pre-log factor of the rate at high SNR and be interpreted as the
stream, and private parts into private streams and superposes number or fraction of interference-free stream(s) that can be simultaneously
communicated to a user (or multiple users). The DoF achieved depends on the
in a non-orthogonal manner the common stream on top of communication strategy used. The larger the DoF, the faster the rate increases
all private streams [14]. The common stream is decodable with the SNR. Hence, ideally a communication strategy should achieve the
by all receivers, while the private streams are to be decoded highest DoF possible. Readers are referred to [7] for more details on various
definitions used to assess the DoF performance of multiple access schemes.
by their corresponding receivers only. Such an approach is 3 Perfect CSIT is obtained by setting the channel estimation error in the
optimal from an information theoretic perspective (Degrees-of- imperfect CSIT model to zero.
4

ceiver at the base station performs SIC to retrieve each stream RSMA uniquely appeared in recent years to fulfill all these
and reconstruct the original messages. By splitting messages, requirements [25], thanks to its inherent message splitting
inter-user interference can be dynamically managed without capability, which is not featured in any other MA schemes.
utilizing time sharing among users to achieve the capacity. This split capability provides several benefits at once as it
Here again, NOMA is a subset of RSMA and is obtained when allows to: 1) partially decode interference and partially treat
user messages are not split. This message splitting property interference as noise (hence its efficiency, flexibility, reliability,
can enable core services where users have intermittent access and resilience), 2) reconcile the two extreme strategies of
behaviour such as URLLC and mMTC, as well as enable interference management and multiple MA schemes into a
different combinations of core services to serve users with single framework (hence its generality/universality), 3) achieve
heterogeneous profiles [23], [26]. the optimal DoF in practical scenarios subject to imperfect
CSIT (hence its robustness). This contrasts with conventional
approaches like OMA, SDMA, and NOMA, that achieve only
C. Challenges and Opportunities for RSMA some of the aforementioned features [7]. Indeed, SDMA has
low complexity and works well in perfect CSIT conditions es-
In our view, MA schemes will have to address new chal- pecially in underloaded regimes, while downlink multi-antenna
lenges and requirements in next generation networks: NOMA is inefficient in general as it incurs a severe DoF loss
• Efficient: Time, frequency, power, spatial, and code- despite an increased receiver complexity due to an inefficient
domain resources should be used to serve users in the use of SIC receivers [7]. OMA, SDMA, and NOMA are not
most efficient way, from both spectral efficiency and general as they are suited for particular propagation conditions
energy efficiency perspectives, while enhancing QoS and [24]. Similarly they are not as flexible as RSMA since they
fairness across devices and services (e.g., unicast, mul- rely on eliminating interference by orthogonalization (OMA),
ticast) and making the best use of the available compu- treat interference as noise (SDMA), and decode interference
tational resources (transmitter and receiver complexity). (NOMA), which are suboptimal strategies in general settings
This calls for performance evaluations of MA schemes [8], [30]. Though NOMA can be made more robust than
beyond OMA vs NOMA [7]. SDMA in the presence of imperfect CSIT, the DoF of NOMA
• Universal: Simpler is better, i.e., a single unified and and SDMA are both suboptimal, therefore incurring a rate
general MA scheme would be easier to implement and loss compared to RSMA [7]. Similarly SDMA has also been
optimize than a combination of multiple MA schemes, shown to be less resilient than RSMA [31]. Finally, SDMA
each optimized for specific conditions. This calls for a and NOMA are less reliable with finite block code lengths
deeper understanding of non-orthogonality in MA designs [32]–[34].
and of how non-orthogonal approaches to MA schemes
relate to each other [7], [24].
D. Objectives and Contributions
• Flexible: Wireless networks are dynamic and MA
schemes should be flexible or versatile enough to cope It is worth to point out (again) that SDMA, NOMA, and
with various interference levels, network density and RSMA are all non-orthogonal approaches to MA design.
load (underloaded, overloaded), topology, services, user However, classifying them all as non-orthogonal does not give
distribution, channel directions and strengths in general us a clue as to where the difference in efficiency, universality,
deployments. This calls for MA schemes that depart from flexibility, robustness and resilience, reliability and low latency
those two extreme interference management strategies, comes from. Answering the fundamental question how to
namely fully treat interference as noise and fully decode manage multi-user interference? is crucial because it helps us
interference [8]. to design efficient MA schemes, to reveal how MA schemes
• Robust and Resilient: MA designs need to depart from the manage interference, and identify and predict under what
conventional aforementioned “robustification” stance and conditions a given MA scheme may be efficient, universal,
adopt a true robustness to practical scenarios subject to flexible, robust, resilient, reliable and have low latency.
imperfect CSIT [14]. Aside being robust, MA schemes In this paper, we provide a tutorial on RSMA and address
need to also become increasingly resilient to emerging common myths and frequently asked questions. To that end,
services involving mixed-criticality, i.e., services with we make the following contributions:
different priorities. This is particularly relevant for appli- • We start this tutorial by departing from the orthogonal vs
cations in safety-critical contexts such as driverless traffic non-orthogonal classification and rather classify schemes
[27]. as a function of how they manage interference. We go
• Low Latency and Reliable: Providing reliable and low- back to the basics of RS proposed for the two-user
latency communications for intelligent transportation and interference channel, and show step by step, how to build
industrial automation is key in next generation commu- downlink, uplink, and multi-cell RSMA and its various
nications. Critical sources of latency relate to link estab- schemes. We also show how interference management
lishment, packet re-transmissions, and data blocklength. and MA designs are closely related and how OMA,
Hence, an MA scheme needs to be able to perform SDMA, NOMA, and RSMA schemes differ in terms of
reliably under short channel codes and decrease the how to manage interference. We then explain that by
number of re-transmissions [28], [29]. understanding the interference management capability of
 5

a scheme, we can understand the drawbacks and benefits

of the MA schemes.
• We generalize the downlink, uplink, and multi-cell
RSMA frameworks to MIMO settings with multiple
antennas at all nodes. This builds upon recent efforts to
design MIMO RSMA [35], contrast with recent survey
and works [7], [25] that were primarily limited to MISO
settings, and confirm that the generality and universality
of RSMA are not limited to single-input single-output
Fig. 1. Two-user symmetric Gaussian interference channel.
(SISO), MISO, or single-input multiple-output (SIMO)
settings, but hold for general MIMO settings. Conse-
quently, recent MIMO NOMA schemes [36], [37] are for 6G. Section VI debunks some myths that have appeared in
shown particular instances of MIMO RSMA. the RSMA literature. Section VII presents answers to several
• We demonstrate that, in contrast to previous generation
common questions that are often asked about RSMA. Section
MA schemes (as OMA, SDMA, NOMA), RSMA is VIII concludes this tutorial overview.
spectrally and energy efficient, universal, flexible, robust Notations: Bold upper and lower case letters denote matrices
and resilient, reliable and low latency. Those unique and column vectors, respectively. (·)T , (·)H , | · |, k · k, E{·},
features of RSMA are discussed in over 40 applications and tr(·) represent the transpose, Hermitian, absolute value,
of RSMA in different systems, scenarios, and services. Euclidean norm, expectation, and trace operators, respectively.
We briefly describe how RSMA is beneficial for each of CN (0, σ 2 ) denotes the circularly symmetric complex Gaussian
those applications to demonstrate how powerful RSMA (CSCG) distribution with zero mean and variance σ 2 .
is for next generation networks.
• We discuss numerous myths/misunderstandings and fre-
quently asked questions about RSMA that we have per- II. K EY Q UESTION BEHIND M ULTIPLE ACCESS D ESIGN :
sonally witnessed. Myths and questions cover a wide H OW TO M ANAGE I NTERFERENCE ?
range of topics, ranging from RSMA relationship with In this section, we present fundamentals about how to
NOMA, gain over MU-MIMO and massive MIMO, role manage interference. To that end, we focus on an example of a
and content of common streams, standardization status of symmetric two-user interference channel, which is the simplest
RSMA, implementation complexity of RSMA, interplay scenario that enables to capture the essence of interference
with other techniques, role of machine learning in RSMA, management. We then use that simple example to draw major
etc. Those myths and frequently asked questions about lessons for MA design.
RSMA nicely complement the tutorial part.
This paper also differs from the recent survey paper on A. Interference Channel
RSMA [25] by providing a tutorial flavour instead of survey,
Let us consider a two-user symmetric Gaussian interference
by diving further into the key role of interference management,
channel (IC) of Fig. 1. We have four nodes, two transmitters
by addressing myths and frequently asked questions, and by
(Tx) and two receivers (Rx), each equipped with a single
expanding the framework to general MIMO settings. It also
antenna. Tx-1 (resp. Tx-2) wants to transmit a message W1
differs from [26], [38], [39] that give an overview of RSMA
(resp. W2 ) to Rx-1 (resp. Rx-2). To simplify the number of
and focus on the specific interplay between RSMA and ISAC
parameters and obtain some insight into the role of interfer-
and reconfigurable intelligent surfaces (RIS).
ence, we assume some symmetry in that the direct channel
from a Tx to its intended Rx (Tx-1 → Rx-1, Tx-2 → Rx-2)
E. Organization and Notations is hd and the cross channel (Tx-1 → Rx-2, Tx-2 → Rx-1) is
The remainder of this paper is organized as follows. Section hc . Hence, the stronger hc compared to hd , the stronger the
II introduces fundamentals about how to manage interfer- interference from Tx-k to Rx-j, j 6= k.
ence, delineates the RS strategy for the 2-user interference A few strategies come to mind to manage this interference:
channel, relates the problem of MA design with interference • Orthogonalize: Tx-1 and Tx-2 transmit on orthogonal
management, and draws important observations. Section III resources (e.g., time or frequency) so that that they do
builds upon important lessons drawn from previous section not interfere with each other at all. This is suboptimal
and introduces various architectures of 2-user RSMA for both since this strategy does not care about how weak the
downlink and uplink for general MIMO settings at all nodes. interference is.
Important lessons about the relationships between existing • Treat Interference as (additive white Gaussian) Noise:
MA schemes and RSMA are summarized. Section IV extends This is a natural approach whenever hc is very weak com-
the RSMA design to general K-user settings for downlink, pared to hd since the interference created by Tx-k would
uplink and multi-cell deployments. Section V shows how the be buried in the noise of Rx-j, j 6= k. The drawback
RSMA architectures find applications in 40 different scenarios of this strategy is that the interference actually carries
relevant in 6G. The breadth and depth of those applications and information and has a structure that could potentially be
scenarios showcase how powerful and underpinning RSMA is exploited in mitigating its effect [30].
6

by t = PPc the fraction of the transmit power allocated to the

common stream.
The received signals at both receivers can be written as
y1 = hd x1 + hc x2 + n1 , (2)
y2 = hc x1 + hd x2 + n2 , (3)
2
where nk ∼ CN (0, σn,k ) is the additive white Gaussian noise
(AWGN). Without loss of generality, we assume the variance
2
Fig. 2. RS for two-user SISO IC (HK scheme) [11]. σn,k = 1. We also assume perfect CSIT and perfect CSI at
the receivers (CSIR).
Each Rx decodes the common streams and its intended
• Decode Interference: Fully decode the interference at private stream as illustrated in Fig. 2. To that end, each Rx
the receivers is especially relevant and actually optimal jointly decodes the common streams into messages W cc,1 and
when the interference is strong enough that hc is large c
Wc,2 by treating the private streams as noise. Then Rx-k
compared to hd [40]. In that case, Rx-k has a better cancels the common streams from the received signal and
reception of Tx-j’s signal than the intended receiver Rx- decodes the intended private stream sk into message W cp,k
j, k 6= j. by treating the other private stream sj , j 6= k, as noise. Rx-k
However, we can do more than those seemingly three different finally reconstructs the original message by recombining W cp,k
strategies by adopting a rate-splitting (RS) approach [10], [11], c c
and Wc,k into Wk , which is the same as Wk in the absence of
[30]. RS splits the transmitted information into two parts, any decoding error. By doing so, Rx-k has partially decoded
namely a common part to be decoded by both receivers and the message of Rx-j, namely W cc,j , and therefore cancelled
a private part to be decoded only by the intended receiver. off part of the interference.
The key benefit of RS is that by enabling a receiver to decode Because of symmetry, the rates of the common streams sc,1
the common part, part of the interference is cancelled off, and sc,2 are the same and are simply denoted as Rc . Similarly,
while the remaining private part from the other transmitter is the rates of the private streams s1 and s2 are also the same and
treated as noise. This enables RS (by properly adjusting the are denoted as Rp . Assuming Gaussian signaling and infinite
power to the common and private parts) to partially decode blocklength, Rc must satisfy the following inequalities at Rx-1
interference and partially treat the remaining interference as (and equivalently at Rx-2 due to symmetry)
noise, therefore bridging, unifying, and actually outperforming !
2
the two extremes of treat interference as noise and decode tP |hd |
Rc ≤ log2 1 + , (4)
interference, as it will appear clearer in the sequel. 1+I
!
2
tP |hc |
Rc ≤ log2 1 + , (5)
B. Rate-Splitting for Two-User Interference Channel 1+I
 h i
We adopt an RS strategy4 to split each transmitter message 2 2
tP |hd | + |hc |
Wk into two parts denoted as a common part Wc,k and a 2Rc ≤ log2 1 + , (6)
private part Wp,k as per Fig. 2 [10], [11], [30]. On the one 1+I
hand, a codebook shared between both transmitters5 /5G New h i
2 2
Radio (NR) system, all codebooks are shared since all users where I = (1 − t) P |hd | + |hc | . Inequalities (4), (5), and
use the same family of modulation and coding schemes (MCS) (6) originate from the fact that the common streams are jointly
specified in the standard. is used to encode the common parts decoded first at each receiver by treating the private streams
Wc,k , k = 1, 2, into the common stream sc,k , k = 1, 2, such as noise. Those three inequalities are obtained by writing
that they are decodable by both receivers. The private parts the rate constraints of a two-user multiple access channel
Wp,k , k = 1, 2 are on the other hand encoded into private (MAC) formed by two virtual transmitters sending respectively
streams sk , k = 1, 2, constructed from independent codebooks. sc,1 and sc,2 to Rx-1 (or 2 due to symmetry) subject to the
Tx-k then transmits the superposition of its common stream additional noise power I created by the private streams s1
sc,k and private stream sk with proper power allocation to and s2 . Hence, (4) refers to the individual rate constraint of
both streams as decoding sc,1 at Rx-1 by treating s1 and s2 as noise. Similarly,
p p (5) refers to the individual rate constraint of decoding sc,2 at
xk = Pc sc,k + Pk sk . (1)
Rx-2 by treating s1 and s2 as noise. The sum-rate constraint
By defining sk = [sc,k , sk ]T and assuming that E[sk sH
k ] = I,
writes as (6).
the average transmit (sum) power constraint at each transmitter The rate Rp of the private streams in (7) is obtained by
is P with Pc and Pk = P −Pc being the power allocated to the noticing that a private stream is decoded after cancelling the
common and private stream respectively. Let us also denote common streams while treating the other interfering private
stream as noise. We can therefore write
4 In
the two-user IC, RS is also known as Han-Kobayashi strategy [11].
!
2
5 This
is not an issue in modern systems since, for example, in an 4G Long
(1 − t) P |hd |
Rp ≤ log2 1 + 2 , (7)
Term Evolution (LTE) 1 + (1 − t) P |hc |
 7

2
where the presence of (1 − t) P |hc | at the denominator By doing so, the interference caused by a private stream has
expresses the interference from the private stream of Tx-k little impact on the other receiver’s performance (compared
owed to Rx-j private stream. to the impairments already caused by the noise). At the same
Combining (4)-(6) and (7), the achievable (symmetric) rate time, it does not prevent each transmitter from experiencing a
in this two-user IC is written as the sum of the private and relatively large private rate as long as |hc | ≤ |hd |.

common rates Rsym = Rp + Rc in (8). The terminology rate- Strong (|hd |2 ≤ |hc |2 ≤ |hd |2 1 + P |hd |2 ): As the
splitting appears clearly here where we note that Rsym is split strength of hc further increases relatively to hd and enters
into two parts, namely, the private rate Rp and the common the regime where |hc | is larger than |hd |, each receiver is
rate Rc . The benefit of this RS architecture is the ability to able to decode both the interfering signal and the desired
adjust the power allocation to the common and private streams signal by performing for instance SIC or joint decoding. In
through parameter t (and therefore the amount of information other words, Rsym is maximized by choosing t = 1 and
to be carried by the common and private streams) to maximize messages Wk , k = 1, 2, are entirely encoded in the common
Rsym depending on the channel strengths of hd and hc . streams sc,k , respectively, so that they are both decoded by
Remark 1: The terminology common and private is taken both receivers (i.e., there are no private streams in this regime
from the information theory literature. In simple words, com- and I = 0). With |hc | ≥ |hd |, inequality 
(5) is inactive
2 2
mon (also sometimes called public) simply refers to a message and Rsym =  21 log2 1 + P
|hd | + |hc | as long as
1 2 2 2

or stream that is to be decoded by multiple users though the log 1 + P |h d | + |h c | ≤ log 1 + P |h d | , i.e.,
2

2 2 2
content of the message is not necessarily intended to all those 2 2 2
P |hd | ≤ P |hc | ≤ P |hd | 1+P |hd | . In other words, the
users. Indeed, sc,k , k = 1, 2, are decoded by both receivers sum-rate constraint (6) is active and forces each
transmitter to
though the content of sc,k is only intended to Rx-k. Common 2
transmit at a rate smaller than log2 1 + P |hd | , i.e., smaller
also contrasts with a multicast message that is a message than the rate achievable without any interference. It is worth

2 2 
2
decoded by multiple users but that is genuinely intended to noting that 21 log2 1+P |hd | +|hc | ≤ log2 1+P |hd |
all those users. Private on the other hand refers to a message P |hc |2

can equivalently be written as log2 1+ 1+P |h |2 ≤ log2 1+
or stream that is only decoded by its intended receiver and is 2

d

treated as noise at other receivers. Indeed, sk is decoded only P |hd | , which expresses that when Rx-k performs SIC to
by Rx-k and is treated as noise by Rx-j with j 6= k. Messages decode the interfering signal sc,j (j 6= k) before decoding
Wk , k = 1, 2, are unicast messages (each intended to a single sc,k , the decodability of sc,j at Rx-k puts a constraint on the
user), but they are split into a common and private parts for transmission rate of sc,j at Tx-j.
2 2 2

interference management benefits. Very strong (|hc | ≥ |hd | 1 + P |hd | ): In this regime,
the interference link is even stronger such
that 12 log2 1 +
2 2 
2
P |hd | + |hc | ≥ log2 1 + P |hd | and the sum-
C. Interference Regimes rate constraint (6) becomes inactive. In other words, when
We can identify four interference regimes depending on the performing SIC, the decodability of sc,j at Rx-k does not
relative strengths6 of hd and hc : restrict the transmission rate of sc,j at transmitter j. Instead the
Very weak (|hc | ≪ |hd |): Whenever |hc | is much smaller rates are only limited by the direct links and each transmitter
2

than |hd |, the maximization of Rsym ends up allocating all can transmit at a rate Rsym = log2 1 + P |hd | equal to the
the transmit power
 to the private
 stream such that t = 0 and one achievable without any interference. The strong and very
P |hd |2
Rsym = log2 1 + 1+P 2 . In such a regime, message Wk strong interference regimes correspond to the regimes where
|hc |
is entirely encoded in the private stream sk (i.e., no splitting the Decode Interference strategy is optimal.
of the messages occurs) such that any interference from Tx-k What distinguishes the weak interference regime from the
to Rx-j is treated as noise. This corresponds to the regime strong (and very strong) regime is that the interference in
where the Treat Interference as Noise strategy is optimal. the former is not strong enough so that only a part of the
Weak (|hc | ≤ |hd |): As the strength of hc increases relatively interference (through the common message) can be decoded by
to hd but as long as |hc | remains smaller than |hd |, one enters a receiver while the interference in the latter is so strong that
the weak interference regime where splitting the messages be- the receiver can fully decode the interference. Consequently,
comes necessary to maximize Rsym . In that regime, inequality moving from the very weak interference regime to the strong
(4) is inactive (since r.h.s. of (5) < r.h.s. of (4)) and t > 0 is to interference regime, the messages are respectively encoded
be chosen, i.e., a non-zero power is allocated to the common into private streams only, a mixture of common and private
stream. There exists a tradeoff between achieving a large streams, and into common streams only as further summarized
private rate and minimizing the interference caused to the other in Table II.
receiver. A qualitative and insightful way of allocating power Though the capacity of the two-user IC remains unknown,
to the common stream is to choose t so that the interference RS can achieve a rate that is within a single bit per second
level caused by the private stream has roughly the same level per hertz (bit/s/Hz) of the capacity for all values of the
as the other receiver’s noise level, i.e., from (7) choose t such channel parameters (even in the non-symmetric case) and
2 P |hc |2 −1 is information-theoretically optimal at asymptotic high SNR
that (1 − t) P |hc | ≈ 1 or equivalently t ≈ P |h |2 [30].
c
regimes [30].
6 In the asymmetric IC, the interference regimes take a more complicated Fig. 3 illustrates the symmetric rate of the various inter-
form [30]. ference management strategies using a concrete numerical
8

 h i 
tP |hd |2 + |hc |2
! ! ! 
(1 − t) P |hd |2  tP |hd |2 tP |hc |2 1 
Rsym = log2 1+ + min log2 1+ , log2 1+ , log2 1 +  (8)
1 + (1 − t) P |hc |2  1+I 1+I 2 1+I 

TABLE II TABLE III

M ESSAGES - TO - STREAMS MAPPING IN THE TWO - USER IC. C OMPARISON OF PREFERABLE INTERFERENCE LEVELS OF DIFFERENT
INTERFERENCE MANAGEMENT STRATEGIES .
sk sc,k
Very weak W1 – Interference levels Very weak Weak Strong Very strong
Weak Wp,k Wc,k Orthogonalize × × × ×
Strong/Very strong – W2 √
Treat interf. as noise × ×
√ ×
√
decoded by its intended Rx-k and decoded by Decode interference × ×
treated as noise by Rx-j both Rx √ √ √ √
Rate-Splitting
√
Notations: : Suited. ×: Not well suited.

9
RS D. Lessons Learned
8 Orthogonalization
Treat interference as noise • Conventional interference management strategies rely
Decode interference
Symmetric rate Rsym [bits/s/Hz]

7 on orthogonalization, treat interference as noise, or

6 decode interference.
• Orthogonalize the resources so as to completely elimi-
5
nate multi-user interference is clearly suboptimal.
4 • Treat interference as noise is an efficient strategy when-
ever the interference level is very weak but is inefficient
3
whenever the interference level grows to weak, strong
2 or very strong.
• Decode interference is an efficient strategy whenever
1 very weak weak strong the interference level is strong and very strong but is
0 inefficient whenever the interference level is very weak
10-3 10-2 10-1 100 101
or weak.
INR/SNR
• RS splits messages into common and private parts so
Fig. 3. Symmetric rate versus INR/SNR. INR=P |hc |2 , SNR=P |hd |2 =1000. as to partially decode interference and partially treat
interference as noise. This allows RS to bridge, unify,
and generalize the two extremes of treat interference as
noise and decode interference and being efficient in all
example. The x-axis is the ratio between the interference-to- four interference regimes.
noise ratio (INR) and SNR; hence the higher INR/SNR, the • RS is a superset of treat interference as noise and
stronger the interference. The different interference regimes decode interference strategies, and can specialize to
are clearly visible. We note how RS can softly bridge all strate- each of them depending on how messages are mapped
gies and outperform them in the weak interference regime7. to streams.
In contrast, the treat interference as noise strategy (resp.
decode interference) quickly becomes suboptimal as INR/SNR Those lessons are summarized in Table III and Fig. 4.
increases (resp. decreases).
Remark 2: If each transmitter is equipped with a directive
antenna, each transmitter could increase the direct channel hd
and decrease the cross channel hc by suitably pointing the
directive antenna. If this operation can be done accurately,
the interference level is more likely to shift toward weaker
regimes. This suggests that it would make more sense to
combine multi-antenna processing (enabling directive beams)
with an interference management strategy designed for weaker
interference regimes (treat interference as noise and RS) rather
than stronger interference regimes (decode interference).
Fig. 4. The relationship between RS, treat interference as noise and decode
interference strategies. Each set illustrates the optimization space of the
7 Note that if the two-user IC is used to represent a two-cell network, the corresponding interference management strategy. The optimization space of
inter-cell interference level is unlikely to be in the strong regime since a user RS is larger such that other strategies are just subsets.
is associated with its closest base station.
 9

Fig. 5. Two-user MISO downlink (Gaussian MISO broadcast channel).

Fig. 6. Two-user downlink MISO RSMA without message combiner.

III. T WO -U SER R ATE -S PLITTING M ULTIPLE ACCESS

Building upon the RS scheme for two-user IC, we can antennas, precoding/beamforming can be performed across the
obtain some insight into how to design RSMA for two- two antennas and the transmit signal model can be written as
X
user downlink (broadcast channel) and uplink (multiple ac- x= pc,k sc,k + pk sk , (10)
cess channel) in the next two sub-sections. We can also k=1,2
relate the above discussion to MA design and show a direct
relationship/analogy between interference management and where pc,k is the precoder of common stream k and pk is the
conventional MA designs such as OMA, SDMA, and NOMA. precoder of private stream k. Defining s = [sc,1 , s1 , sc,2 , s2 ]T
and assuming that E[ssH ] = I, the P average transmit sum
power constraint8 at the transmitter is k=1,2 Pc,k + Pk ≤ P
A. Two-User Downlink Rate-Splitting Multiple Access 2 2
with Pc,k = kpc,k k and Pk = kpk k the power allocated to
Let us consider a single transmitter potentially equipped common stream k and private stream k, respectively. Uniquely,
with M antennas serving two users. To that end, suppose that since sc,1 and sc,2 are transmitted from the same transmitter
the two transmit antennas in Fig. 1 cooperate by exchanging and are decoded by both users in this downlink setting, one
CSI and messages such that they effectively belong to the can choose pc,1 = pc,2 = pc so that we only P have a single
same transmitter as per Fig. 5. The two-user IC then becomes common precoder and x = pc (sc,1 + sc,2 ) + k=1,2 pk sk .
a two-user MISO broadcast channel (BC) with M = 2 User-k can now decode using SIC or joint decoding common
transmit antennas where the base station transmits to two streams sc,1 and sc,2 and then private stream sk . Using SIC,
active receiving users whose channels are givenby h1 and this architecture would require each receiver k to be equipped
h2 , respectively.
  example of Fig. 5, h1 = hd hc
In the with two SIC layers so as to decode three streams sc,j ,
and h2 = hc hd . Considering a transmit signal vector x sc,k , and sk , j 6= k. In summary, from two messages, this
spanning across the transmit antennas, the received signal at architecture creates four streams, and two SIC are required
user-k can be written as to recover the original messages. This RSMA architecture is
illustrated9 in Fig. 6.
y k = hk x + n k , (9) A second MISO RSMA architecture is obtained by noting
2 that an additional benefit of the downlink is that instead of
where nk ∼ CN (0, σn,k ). Such a multi-antenna BC is a basic
encoding each common part into a common stream Wc,k →
building block of modern downlink communication systems.
sc,k , we can first combine the common parts into a common
Assuming perfect CSIT and CSIR, the capacity of this channel
message Wc = {Wc,1 , Wc,2 } that is then encoded into a single
is known and is achieved by Dirty Paper Coding (DPC) [41]–
common stream Wc → sc such that
[43]. DPC is nevertheless complex to implement due to its X
inherent nonlinear encoding/precoding mechanism. RSMA has x = pc s c + pk s k . (11)
appeared in the past few years as an appealing strategy to k=1,2
achieve close performance to DPC while maintaining the low
complexity of linear precoding [8]. In the presence of im- By defining s = [sc , s1 , s2 ] and assuming that E[ssH ] = I,
T

perfect CSIT, the capacity and the capacity-achieving strategy the average
P transmit (sum) power constraint at the transmitter
2
are unknown. It is nevertheless known that RS plays a central is Pc + k=1,2 Pk ≤ P with Pc = kpc,k k the power allocated
role to achieve the optimal DoF (and generalized DoF) [17]– to the unique common stream. The received signal at user-k,
[22], and RSMA can outperform DPC [9]. In the sequel, we k = 1, 2, j 6= k, is written as
delineate progressively the key design principles and schemes y k = hk pc s c + hk pk s k + hk pj s j + n k . (12)
of RSMA in the two-user downlink scenario.
1) MISO RSMA: Inspired by the IC, the first MISO RSMA 8 We assume a sum power constraint but a per-antenna power constraint

architecture is obtained by splitting the two messages W1 could also be considered.

9 Only the receiver of user-1 is detailed in Fig. 6. The receiver of user-2
and W2 into common and private parts and encoding each
follows the same principle as that of user-1 where two SIC layers are needed
part into a corresponding stream such that Wc,k → sc,k and to decode and cancel the common streams before retrieving user-2’s private
Wp,k → sp,k , k = 1, 2. Given the presence of multiple message.
10

Fig. 7. Two-user downlink MISO RSMA with message combiner (1-layer

RS).

Fig. 8. The relationship between OMA, NOMA, SDMA, physical-layer

This is the so-called 1-layer RS architecture of RSMA because multicasting, and RSMA in the downlink two-user case. Each set illustrates
it relies on a single common stream and therefore a single the optimization space of the corresponding communication strategy. The
SIC layer at each receiver, hence simplifying the encoding optimization space of RSMA is larger such that SDMA, NOMA, and physical-
layer multicasting are just subsets.
complexity (three streams used instead of four) and the de-
coding complexity (one SIC layer instead of two) compared
to (10) and (1). Indeed both users decode the single common
stream sc into Wcc by treating the interference from all private
streams as noise. Each user-k then retrieves the estimate Wcc,k
c c
from Wc . Using SIC, Wc is then re-encoded, precoded, and
subtracted from the received signal such that user-k decodes
its private stream sk into W cp,k by treating the remaining
interference from the other private streams as noise. Finally
user-k recombines W cc,k and Wcp,k into the message Wck which
is the same as the original message Wk if no decoding error
occurs. This architecture is illustrated in Fig. 7.
2) Revisiting the Interference Regimes: It is worth relating
(12) to the lessons learned in Section II. In (12), after pre-
coding, the two-user MISO BC can effectively be seen as a
two-user IC where user-k decodes a common stream sc and a
private stream sk subject to interference from private stream
sj , j 6= k. The interference regime experienced by user-k
in this effective two-user IC is therefore determined by the
strength of the precoded channels hk pk and hk pj (with hk pk
and hk pj taking the role of hd and hc , respectively):
• |hk pj | ≪ |hk pk |: very weak interference regime and
interference should be treated as noise, i.e., pc = 0 Fig. 9. A beam representation of RSMA and its sub-schemes SDMA, NOMA,
OMA, and physical-layer multicasting.
and only private streams are used. Decoding interference
would perform badly. This scenario can typically occur
when the channels are close to being orthogonal. schemes such as SDMA and NOMA are specifically tailored
• |hk pj | ≤ |hk pk |: weak interference regime and non- for one specific interference management strategy (e.g., treat
zero power should be allocated to all common and interference as noise, decode interference), one specific in-
private streams. This would typically occur whenever the terference regime (e.g., very weak, strong) and one type of
channels are neither orthogonal nor aligned. channel conditions (e.g., orthogonal, aligned), and how RSMA
• |hk pj | ≥ |hk pk |: strong interference regime and inter- can unify them all.
ference should be decoded. Treat interference as noise 3) Unifying OMA, SDMA, NOMA, and Multicasting: In the
would perform badly. This scenario can typically occur simple two-user case, OMA, SDMA, NOMA, and physical-
when the channels are aligned. layer multicasting are particular instances of RSMA, as illus-
This shows that depending on the propagation conditions trated by Fig. 8, the message to stream mapping in Table IV
in multi-antenna settings (e.g., angle between user channel [24], and Fig. 9.
directions), the interference regime can change from very weak SDMA is a special case of RSMA by forcing pc = 0.
to strong. Consequently, downlink MA schemes therefore In this way, Wk is directly encoded into sk and the system
P
need the ability to softly evolve from the extreme of decode model writes as x = k=1,2 pk sk . By doing so, each stream
interference to treat interference as noise. RSMA has the is decoded by its intended user by treating any residual
flexibility to cope with all those interference regimes and interference from the other stream as noise. Recalling Tables II
propagation conditions. and III, from an interference management strategy, SDMA is
It will now appear clear in the sequel how existing MA reminiscent of and builds upon the Treat Interference as Noise
 11

TABLE IV TABLE VI
M ESSAGES - TO - STREAMS MAPPING IN THE TWO - USER MISO BC [24]. C OMPARISON OF DIFFERENT MA SCHEMES IN TERMS OF TRANSMIT- SIDE
VS RECEIVE - SIDE INTERFERENCE CANCELLATION .
s1 s2 sc
SDMA W1 W2 – Interf. cancel. transmit-side receive-side both sides
√
NOMA W1 – W2 SDMA/DPC × ×
√
OMA W1 – – NOMA × ×
√
Multicasting – – W1 , W2 RSMA × ×
RSMA Wp,1 Wp,2 Wc,1 , Wc,2 √
Notations: : Relevant. ×: Not relevant.
decoded by its intended user and decoded by
treated as noise by the other user both users
the performance of those MA schemes as a function of the
TABLE V propagation conditions [24].
C OMPARISON OF PREFERABLE RESIDUAL MULTI - USER INTERFERENCE
LEVELS OF DIFFERENT MA SCHEMES .
RSMA can also be seen as a smart combination of transmit-
side and receive-side interference cancellation strategy where
Interference levels Very weak Weak (Very) Strong the contribution of the common stream is adjusted according
√
SDMA × ×
√ to the level of interference that needs to be canceled by
NOMA ×
√ ×
√ √
RSMA the receiver, therefore departing from the transmit-side only
√ and receive-side only interference cancellation strategies of
Notations: : Suited. ×: Not well suited.
SDMA (and DPC) and NOMA, respectively. This is further
summarized in Table VI.
strategy which would be efficient only if the residual multi- Example 1: To further illustrate the split of the mes-
user interference |hk pj | is sufficiently weak as in orthogonal sages and the flexibility of RSMA, let us imagine that
channels [24]. the message of user-1 W1 = (a1 a2 a3 a4 ) ∈ W1 =
NOMA is a special case of RSMA by forcing the encoding {0000, 0001, 0010, . . ., 1111}, where |W1 | = 16. Similarly,
of message W2 entirely into sc (i.e., Wc = W2 ) and W1 the message of user-2 is W2 = (b1 b2 b3 ) ∈ W2 =
into s1 while turning off s2 (P2 = 0). By doing so, user-1 {000, 001, 010, . . . , 111}, where |W2 | = 8. In SDMA, W1
fully decodes the message of user-2 (and therefore interference would be encoded into s1 and W2 into s2 . Assuming uncoded
created by user-2 stream) and 1-layer RS system model transmission for simplicity, s1 and s2 would then be a 16-
becomes the NOMA system model x = pc sc +p1 s1 . Note that QAM symbol and a 8-PSK symbol, respectively. In NOMA,
NOMA also utilizes a common stream since the message of W1 would be encoded into s1 and W2 into sc , also using
one of the two users, namely W2 in this example, is decoded a 16-QAM symbol and a 8-PSK symbol, respectively. In
by both users. Connecting back to Tables II and III, from RS, we split user-1’s message in, e.g., Wc,1 = (a1 a2 ),
an interference management strategy, NOMA is reminiscent Wp,1 = (a3 a4 ), and user-2’s message in, e.g., Wc,2 = (b1 ),
of and builds upon the Decode Interference strategy which Wp,2 = (b2 b3 ). The common message is then constructed as
would be efficient only if the multi-user interference level is Wc = (Wc,1 Wc,2 ) = (a1 a2 b1 ), which is then encoded into
sufficiently strong as in aligned channels [24]. sc using a 8-PSK symbol. Wp,1 and Wp,2 are encoded into
OMA is a special case of RSMA by forcing only one user s1 and s2 using QPSK symbols, respectively. The interested
2 2
to be scheduled, e.g., user-1 (i.e., kpc k = kp2 k = 0). reader is referred to [25] for more examples.
Physical-layer multicasting is a special case of RSMA 4) Rate Analysis: Under the assumption of Gaussian sig-
obtained when messages W1 , W2 are both encoded into sc naling and infinite blocklength, and perfect CSIT and CSIR,
(i.e., Wc = {W1 , W2 }) and the private streams are turned off the instantaneous rates for decoding the common and private
(P1 = P2 = 0). streams at user-k are given as
Drawing an analogy with Table III, we can conclude that !
|hk pc |2
MA schemes operate in some preferred residual multi-user Rc,k = log2 1 + 2 2 ,
interference regimes as shown in Table V. The above clearly |hk p1 | + |hk p2 | + 1
! (13)
shows how three different non-orthogonal approaches to MA |hk pk |2
designs, namely SDMA, NOMA, and RSMA, fundamentally Rk = log2 1 + 2 ,
|hk pj | + 1
differ based on how multi-user interference is managed.
2
NOMA is such that at least one user is forced to fully where the noise variance was normalized σn,k = 1 without
decode the message(s) of other co-scheduled user. SDMA loss of generality. To ensure that sc is successfully decoded
and RSMA do not follow this approach since they both do by both users, its rate cannot exceed
not force a user to fully decode the messages of another co-
Rc = min {Rc,1 , Rc,2 } . (14)
scheduled user. SDMA actually treats any residual interference
as noise, and RSMA is built upon the principle of splitting As sc contains sub-messages Wc,1 , Wc,2 of the two users,
the messages so as to partially treat interference as noise and the rate distribution of Rc among the users is adapted to the
partially decode the remaining interference. Consequently, this amount of sub-messages that each user contributed. Let Ck
difference in managing interference has deep consequences denote the portion of rate Rc allocated to user-k for Wc,k .
on the universality of MA schemes, with RSMA being a Then, we have
superset of SDMA and NOMA as per Fig. 8, but also on C1 + C2 = Rc . (15)
12

10
The overall achievable rate of user-k is first channel, i.e., H2 ≥st H1 , is derived for which the
capacity region is given by
Rk,tot = Ck + Rk . (16)
Here again, the terminology rate-splitting appears clearly in R1 ≤ I(V ; Y1 |H1 ),
(16) where we note that the rate of each user is split into two R2 ≤ I(X; Y2 |V, H2 ), (22)
parts, namely, the rate of sk (i.e., the private rate) and part of
the rate of sc (i.e., the common rate). with Markov chain V −X −Y2 −Y1 . Even though the capacity
Common metrics to design the systems include 1) weighted expressions in (22) look familiar to the rates achievable with
sum rate (WSR) u1 R1,tot + u2 R2,tot where u1 and u2 are SC and SIC (and NOMA), the optimal coding and decoding
weights set to account for fairness among users (for instance strategy to achieve (22) remains unknown because the same
when conducting proportional fair scheduling); 2) max-min marginal property of the BC is applied to transform the joint
fair (MMF) that aims at maximizing mink=1,2 Rk,tot ; 3) distribution of the underlying channel to fully coupled.
R +R2,tot 5) Precoder Design and Power Allocation: Recall from
energy efficiency (EE) 1 (P 1,tot where η ∈ (0, 1]
1 +P2 +Pc )+Pcir
η
Section II-C that a qualitative and insightful way of allocating
and Pcir are respectively the power amplifier efficiency and
power to the common stream is so that the interference level
the circuit power consumption. All three metrics could also be
caused by the private stream has roughly the same level as
subject to a QoS constraint Rk,tot ≥ Rkth with Rkth a minimum
the other receiver’s noise level, i.e., (1 − t) P |hc |2 ≈ 1. In
rate threshold to be achieved by user-k.
the downlink scenario, looking at (13), we can think in a
Remark 3: We can wonder what happens if M = 1, namely
similar way and choose the power allocation P1 and P2 to the
downlink SISO. RSMA strategy (11) yields 2 2
p p p two private streams such that |h2 p1 | ≈ 1 and |h1 p2 | ≈ 1,
x = Pc sc + P1 s1 + P2 s2 (17) i.e., instead of allocating all the transmit power P to the
private streams (as done in SDMA), we allocate a fraction
and the common and private rates write as
! of the total available power to them such that the signal-to-
|hk |2 Pc interference-plus-noise ratios (SINRs) of the private streams
Rc,k = log2 1 + 2 2 , are not interference limited. If we were allocating a higher
|hk | P1 + |hk | P2 + 1 2
! (18) power to the private streams with |hk pj | ≫ 1, the SINR
|hk |2 Pk |hk pk |2 |hk pk |2
≈ |h p |2 would saturate and the private rates
Rk = log2 1+ 2 , |hk pj |2 +1 k j
|hk | Pj + 1 would not increase further. Hence, the key is for the private
streams not to enter the interference limited regime, and
for k = 1, 2. Without loss of generality, we consider |h2 | ≤
allocate any remaining power to the common stream whose
|h1 | so that
! SINR will keep increasing linearly as Pc increases. Doing
2 so, R1 + R2 would be roughly equivalent to the sum-rate
|h2 | Pc
Rc = log2 1 + 2 2 . (19) achieved by SDMA, but the common rate Rc would provide
|h2 | P1 + |h2 | P2 + 1
an additional rate increase over what SDMA can offer.
Since the capacity of the Gaussian SISO BC with perfect CSIT The above is particularly insightful when the CSIT is
and CSIR is obtained by performing superposition coding (SC) imperfect and the transmitter only has an estimate ĥk of user-
with SIC (SISO NOMA), optimizing P1 , P2 , Pc to maximize k channel h = ĥk + h̃k with h̃k the channel acquisition
the WSR would lead to choosing P2 = 0 so that the message error. In such setting, residual multi-user interference scaling
of the weaker user (user-2) W2 is entirely encoded sc and
in √ 2
√ as h̃k pj would be unavoidable regardless of how the
SISO RSMA becomes SISO NOMA x = Pc sc + P1 s1 precoder pj is designed. Following such an approach, we can
and its achievable rates are demonstrate that RSMA is information theoretically optimal
 
R1 = log2 1 + |h1 |2 P1 , (20) from the DoF perspective in the presence of imperfect CSIT
! [18], [19], [46] while SDMA and NOMA are not [7]. Because
|h2 |2 Pc of its inherent robustness to imperfect CSIT, RSMA can also
Rc = log2 1 + 2 . (21)
|h2 | P1 + 1 afford smaller feedback overhead than conventional SDMA
[47].
This shows how RSMA can act as NOMA in Gaussian SISO More systematic methods to design the precoder and power
downlink. allocation can be obtained using either closed form low com-
Remark 4: The capacity region of a Gaussian SISO BC is plexity suboptimal techniques or using optimization techniques
known for both cases with fixed and fading channels when detailed in [25]. Extensive results have demonstrated that the
there are perfect CSIT and CSIR. However, the problem is power allocation to the common stream vs private stream
open in general when there is imperfect CSIT and only limited depends on a number of factors, including the angle between
cases are known [44]. The fading SISO BC with only perfect user channels [8], [24], the disparity of channel strengths [8],
CSIR but imperfect CSIT lacks the degraded structure in [24], the objective function to maximize [18], [48], the quality
general for arbitrary fading distributions. In these cases, SC of CSIT [18], [48], the QoS, the network load, etc.
and SIC (and therefore NOMA) can not achieve the capacity
region. In [45, Theorem 1], the sufficient condition that the 10 Stochastic order H ≥
2 st H1 means that P(H2 ≥ s) ≥ P(H1 ≥ s) for
second channel is stochastically larger than or equal to the all s from the support of H1 , H2 .
 13

TABLE VIII
T WO DIFFERENT CSIT PATTERNS .

User-1
√ User-2
√
subband A
subband B × ×
Non-alternating CSIT

User-1 User-2
√
subband A ×
√
subband B ×
Fig. 10. Cooperative downlink RSMA with user relaying [52]. Alternating CSIT

√
TABLE VII Notations: : Good CSIT quality. ×: Poor CSIT quality.
M ESSAGES - TO - STREAMS MAPPING IN THE TWO - USER COOPERATIVE MISO
DOWNLINK [52].
of CSIT in that subband (good on subband A and bad on
s1 s2 sc θ
SDMA W1 W2 – θ =1 subband B). One could also apply the same strategy to the
NOMA W1 – W2 θ =1 alternating CSIT pattern of Table VIII. However, because of
OMA W1 – – θ =1 the alternating feature, simply doing the spatial domain RSMA
Multic. – – W1 , W2 θ =1
RSMA Wp,1 Wp,2 Wc,1 , Wc,2 θ =1
transmission (11) would not be as efficient as doing an RSMA
C-NOMA W1 – W2 θ <1 transmission across the two subbands. In other words, RSMA
DF – – W2 θ <1 can benefit from a multi-channel transmission in space-time
C-RSMA Wp,1 Wp,2 Wc,1 , Wc,2 θ <1 (ST) or in space-frequency (SF) depending on the CSIT pattern
decoded by its intended user and decoded by
treated as noise by the other user both users [47], [54], [55]. Such a multi-channel RSMA transmission is
particularly helpful in the presence of alternating CSIT, i.e.,
Notations: θ = 1 refers to non-cooperative schemes as in Table IV. θ < 1
refers to cooperative RSMA schemes. whenever the CSIT changes across time or frequency in an
alternating user-specific manner. The alternating CSIT pattern
of Table VIII is a typical (and practical) scenario where the
6) Cooperative RSMA: An interesting extension is achieved transmitter wants to serve two users but the CSIT of user-1
by considering that the transmitter can opportunistically ask (resp. user-2) is better on time/frequency B (resp. A) and worse
one of the users to act as a relay. This is known as a coopera- on time/frequency A (resp. B). In such scenario, an ST/SF
tive relay BC in information theory [49]–[51]. Specifically, in RSMA scheme can further increase the DoF over conventional
our setting, the common stream is decoded by both users and RSMA [54], [55].
the transmitter can ask the relay user to forward the decoded Compared to the RSMA scheme of (11), ST/SF-RSMA
common message/stream to the other user to efficiently cope scheme transmits an additional common stream (obtained from
with a wide range of propagation conditions (disparity of a further split of the messages), i.e., s0 , across the two channel
user channel strengths and directions) and compensate for uses. Specifically, considering the alternating CSIT pattern of
the performance degradation due to deep fading [52], [53], Table VIII, the transmitted signals in subbands A and B write
as illustrated in Fig. 10. The parameter 0 < θ ≤ 1 refers to can be expressed as follows:
the time split between the direct transmission phase and the
(A)
X (A) (A)
relaying phase, i.e., θ = 1 indicates that no time is allocated x(A) = p0 s0 + p(A) c sc
(A)
+ pk s k , (23)
to relaying, hence boiling down to the conventional downlink k=1,2
(B)
X (B) (B)
(B)
transmission of Fig. 5. x = p0 s 0 + p(B)
c sc
(B)
+ pk s k , (24)
The user relaying feature of cooperative RSMA enables to k=1,2
enlarge the pool of possible schemes within the RSMA frame-
work as shown by the messages-to-streams mapping in the where the superscript (i) refers to the subband. We note the
two-user cooperative MISO downlink of Table VII. We note addition of the new common stream s0 that has been repeated
(i)
how the mapping of Table IV has been extended to include across the two channel uses (precoded by p0 in channel use
cooperation where conventional decode and forward (DF) and i = 1, 2). If the CSIT quality becomes non-alternating, the
cooperative NOMA (C-NOMA) are particular instances of common stream s0 becomes useless (zero power is allocated
cooperative RSMA (C-RSMA). to s0 ) and SF-RSMA boils down to (11) in each subband. The
7) Space-Time / Space-Frequency RSMA: Most RSMA receiver at both users is more complicated since two common
literature deal with a single channel use (be it in time or streams have to be decoded at each user before decoding the
frequency). This means the signal model as expressed in (11) respective private stream.
is applied in a given time slot or given frequency subband. The decoding works as follows. Let us focus on user-1 for
(A)
For instance, in the non-alternating CSIT pattern of Table simplicity. First, user-1 decodes sc and s0 sequentially using
VIII, (11) would be applied separately on subbands A and subband A observation by treating the private streams as noise.
B, i.e., each subband determining the split of the message, Secondly, after removing s0 from subband B observation,
(B)
power allocation and precoders based on the available quality user-1 recovers sc by treating the private streams as noise.
14

(A)
Thirdly, by removing all common streams, user-1 decodes s1
(B)
and s1 by treating the other private stream as noise.
More details on the CSIT pattern conditions needed for the
optimality of separate RSMA in each subband can be found
from [56]. Another interesting use of ST-RSMA, where RS is
combined with space-time block coding to address the lack of
knowledge of the channel phase information at the transmitter,
has been developed in [57].
8) MIMO RSMA: Let us now consider a two-user MIMO
downlink with M transmit antennas and N receive antennas Fig. 11. Downlink MIMO RSMA [35].
at each user. The transmitter wants to transmit two N -
dimensional vectors of messages w1 and w2 to user-1 and
user-2, respectively. To that end, for each user, each of those RSMA again is a superset of all schemes. We note that SDMA
N messages is split into a common part and a private part. is replaced by MU-MIMO where a vector of private streams
Common parts of both users are combined and encoded into a is transmitted to each user. NOMA in the form of having
N -dimensional vector of common streams sc and private parts one user (user-1) decoding all the messages of the other user
are encoded into two N -dimensional vectors of private streams (user-2) is illustrated, along with OMA. Additionally, other
sk . M × N precoding are performed on the common and subschemes of RSMA are listed for the sake of illustration.
private stream vectors such that the transmit signal is written For instance, in subscheme 1, both users decode the 2nd
as X messages and, in subscheme 2, both users decode the 1st
x = Pc sc + Pk sk . (25) messages, i.e., each user decodes one message of the other
k=1,2 user. Subscheme 3 is when both users decode the 2nd message
of user-2. Note that the MIMO NOMA scheme proposed
Let us consider an example with N = 2 to illustrate
in [36], [37] is also a subscheme of MIMO RSMA, and it
the universality of the framework. We have four messages
(1) (2) switches among MU-MIMO, NOMA, or the three additional
to transmit (two for each user) w1 = [W1 , W1 ] and
(1) (2) (j) subschemes of RSMA when different numbers of transmit
w2 = [W2 , W2 ], where Wi refers to the jth message antennas and receive antennas are considered.
of user-i, i = 1, 2, j = 1, 2. We split the four messages into
Optimization of MIMO RSMA with both perfect and im-
common and private parts such that
n o perfect CSIT can be performed as in [35]. Similar to the
(1) (1)
W1 = Wc,1 , Wp,1 ,
(1)
(26) MISO case, RSMA outperforms MU-MIMO and NOMA in
n o the MIMO case. Note that RSMA is information theoretically
(2) (2) (2)
W1 = Wc,1 , Wp,1 , (27) optimal from a DoF perspective in MIMO BC with imperfect
n o CSIT [21], [22]. In the asymmetric case where the number of
(1) (1) (1)
W2 = Wc,2 , Wp,2 , (28) receive antennas is not the same at each user, RSMA with a
n o multi-channel transmission (similar to ST/SF RSMA) needs to
(2) (2) (2)
W2 = Wc,2 , Wp,2 . (29)
be used to achieve optimality. The reader is referred to [22]
(1) (2) for more details on such a ST RSMA scheme for asymmetric
Common parts are then combined into Wc and Wc and
(1) (2) MIMO BC.
then respectively encoded into common stream sc and sc
n o
(1) (1)
Wc(1) = Wc,1 , Wc,2 → s(1)c , (30)
n o B. Two-User Uplink Rate-Splitting Multiple Access
(2) (2)
Wc(2) = Wc,1 , Wc,2 → s(2)c , (31)
In the uplink, we consider two single-antenna users simul-
to formthe 2-dimensional common stream vector sc = taneously transmitting their messages to a receiver equipped
so as
(1) (2) T
sc sc . Private parts are encoded into private streams with M antennas. When M = 1, such two-user MAC can be
considered as a special case of the two-user IC in Fig. 1 where
(1) (1) (2) (2)
Wp,1 → s1 , Wp,1 → s1 , (32) the two receivers are colocated and cooperatively decode the
(1) (1) (2) (2) messages of the two users.
Wp,2 → s2 , Wp,2 → s2 , (33)
1) Two-User Architectures: Inspired by the RSMA design
so as to  create two 2-dimensional private stream
 vectors in IC, the uplink RSMA architecture has been proposed in
(1) (2) T (1) (2) T
s1 = s1 s1 and s2 = s2 s2 . At the [23] by splitting the message of one user into two parts.
receivers, both users decode the vector of common streams Without loss of generality, we assume that user-1’s message
first using the two receive antennas, perform SIC, and then W1 is split into W1,1 and W1,2 . By independently encoding the
decode their respective private stream vector while treating two parts into s1,1 , s1,2 , respectively allocating transmit power
the co-scheduled user’s private stream vector as noise. Fig. 11 P1,1 , P1,2 , and superposing the two streams, the transmit signal
illustrates the MIMO RSMA architecture. at user-1 is given by
The mapping of messages to streams of Table IV is now p p
further extended to account for MIMO RSMA in Table IX. x1 = P1,1 s1,1 + P1,2 s1,2 . (34)
 15

TABLE IX
M ESSAGES - TO - STREAMS MAPPING IN THE TWO - USER MIMO BC.

s1 s2 sc
(1) (2) (1) (2) (1) (2)
s1 s1 s2 s2 sc sc
(1) (2) (1) (2)
MU-MIMO W1 W1 W2 W2 – –
(1) (2) (1) (2)
NOMA W1 W1 – – W2 W2
(1) (2)
OMA W1 W1 – – – –
(1) (1) (2) (2)
Multicasting – – – – W1 , W2 W1 , W2
(1) (2) (1) (2) (1) (1) (2) (2)
RSMA Wp,1 Wp,1 Wp,2 Wp,2 Wc,1 , Wc,2 Wc,1 , Wc,2
(1) (1) (2) (2)
subscheme 1 W1 – W2 – W1 W2
(2) (2) (1) (1)
subscheme 2 – W1 – W2 W1 W2
(1) (2) (1) (2)
subscheme 3 W1 W1 W2 – – W2
decoded by its intended user and treated as noise by the other user decoded by both users

TABLE X
M ESSAGES - TO - STREAMS MAPPING IN THE TWO - USER MULTIPLE ACCESS
CHANNEL .

User-1 User-2
s1,1 s1,2 s2
NOMA W1 – W2
OMA W1 – –
RSMA W1,1 W1,2 W2

Fig. 12. Two-user uplink SIMO RSMA. NOMA is a subscheme of RSMA by forcing the encoding of
message W1 entirely into s1,1 and W2 into s2 while turning
off s1,2 (P1,2 = 0). There is no message splitting at user-
1 and one layer of SIC is required at the receiver to decode
and remove the entire message of one user (i.e., user-1) before
decoding the message of the other user (i.e., user-2). Note that,
though we here refer to this strategy as (uplink) NOMA, this
strategy is nothing else than the traditional SIC at the receiver
to achieve the corner points of the MAC capacity region.
Fig. 13. The relationship between OMA, NOMA, and RSMA in the uplink
two-user case. Each set illustrates the optimization space of the corresponding
OMA is also a subscheme of RSMA by forcing one user to
communication strategy. The optimization space of RSMA is larger than OMA be scheduled, i.e., W1 is directly encoded into s1,1 while W2
and NOMA. is turned off.
3) Uplink MIMO RSMA: An extension to the two-user
At user-2, the message W2 is directly encoded into s2 . By uplink SIMO RSMA in Fig. 12 is the two-user uplink
MIMO RSMA with N transmit antennas at each user and
√ certain power P2 , the transmit signal at user-2 is
allocating
x2 = P2 s2 . M receive antennas at the receiver. In this case, each user-
The signal received at the receiver is k transmits a N -dimensional vector of messages wk =
(1) (N )
[Wk , . . . , Wk ] to the receiver. At user-1, each of its N
y = h1 x1 + h2 x2 + n (j) (j)
p p p (35) messages W1 , j ∈ {1, . . . , N } is split into two parts, W1,1
= P1,1 h1 s1,1 + P1,2 h1 s1,2 + P2 h2 s2 + n, (j) (j)
and W1,2 , and encoded independently into s1,1 and s1,2 .
(j)

(1) (N )
where h1 , h2 ∈ CM×1 are the channel vectors and n ∼ The two N -dimensional vectors s1,1 = [s1,1 , . . . , s1,1 ]T and
CN (0, IM ) is the AWGN vector. The receiver can employ (1) (N )
s1,2 = [s1,2 , . . . , s1,2 ]T are linearly precoded by the precoders
SIC or joint decoding to decode the three streams s1,1 , s1,2 , s2 . P1,1 , P1,2 ∈ C M×N
and superposed such that the transmit
If using SIC, two layers of SIC are required at the receiver. signal at user-1 is
In summary, two-user uplink RSMA creates two virtual users
from the user that splits its message into two parts and in
x1 = P1,1 s1,1 + P1,2 s1,2 . (36)
total three streams are sent to the receiver. Two SIC layers
are required to recover the original messages of the two users.
This uplink RSMA architecture is illustrated in Fig. 12. At user-2, the message vector w2 is encoded into a N -
2) Unifying OMA and NOMA: In the two-user MAC, both dimensional stream vector s2 and linearly precoded by P2 .
OMA and NOMA are subschemes of RSMA, as per Fig. 13. The transmit signal at user-2 is x = P2 s2 .
The corresponding messages to streams mapping is illustrated An example of the two-user uplink MIMO RSMA when
in Table X. N = 2 is illustrated in Fig. 14. Each message in the message
16

IV. K -U SER RSMA

The schemes developed for two-user in the previous sections
can all be extended to K-user and conclusions derived for
two-user hold for K-user as well. Nevertheless, the K-user
scenario also opens the door to other RSMA schemes with
a variable number of SIC layers. In this section, the state-of-
the-art K-user RSMA schemes are delineated in the downlink,
uplink, and multi-cell MIMO settings where each node is
equipped with multiple antennas.

Fig. 14. Two-user uplink MIMO RSMA.

A. Downlink
TABLE XI We first consider a downlink symmetric MIMO BC where
M ESSAGES - TO - STREAMS MAPPING IN THE TWO - USER MIMO MULTIPLE a transmitter equipped with M antennas serves K users, each
ACCESS CHANNEL .
equipped with N receive antennas. The users are indexed by
User-1 User-2 K = {1, . . . , K}. Without loss of generality, we assume a
s1,1 s1,2 s2 Qk -dimensional vector of messages is transmitted to user-k,
(1) (2) (1) (2) (1) (2) (1) (Q )
s1,1 s1,1 s1,2 s1,2 s2 s2 i.e., wk = [Wk , . . . , Wk k ]T , where Qk ≤ min(M, N ).
(1) (2) (1) (2)
NOMA W1 W1 – – W2 W2 Depending on the different RSMA schemes adopted, user
(1) (2)
OMA W1 W1 – – – – messages {wk , k ∈ K} are split, combined, and encoded
(1) (2) (1) (2) (1) (2)
RSMA W1,1 W1,1 W1,2 W1,2 W2 W2 into different stream vectors, which are then mapped to the
transmit antennas using linear or non-linear precoders and
forms the transmit signal x. The signal is transmitted through
vector w1 of user-1 is split into two parts as
n o a MIMO BC and the receive signal at user-k is given by
(1) (1) (1)
W1 = W1,1 , W1,2 , yk = HH k x + nk , where Hk ∈ C
M×N
is the channel matrix
n o 2
(2) (2) (2)
(37) between the base station and user-k, nk ∼ CN (0, σn,k IN ) is
W1 = W1,1 , W1,2 . the AWGN vector at user-k. Without loss of generality, the
noise variances of all users are assumed to be equal to 1, i.e.,
The four submessages are respectively encoded into four 2
σn,k = 1, ∀k ∈ K. Next, we detail the transceiver architectures
streams and create two 2-dimensional stream vectors s1,1 =
 (1) 
(2) T
 
(2) T for different RSMA schemes in such symmetric MIMO BC.
s1,1 s1,1 and s1,2 = s(1) 1,2 s1,2 . The two stream
1) 1-layer RS: The system model of 1-layer RS for MIMO
vectors are linearly precoded and superposed to form the
BC is first proposed in [35]. As illustrated in Fig. 15, each
transmit signal x1 . At user-2, the two messages
 in w2 are
 (i)
(2) T message Wk in wk for user-k is split into one com-
directly encoded into two streams in s2 = s(1) 2 s2 (i)
mon sub-message and one private sub-message as Wk =
and linearly precoded to form the transmit signal x2 . The (i) (i)
mapping of messages to streams for this example is illustrated {Wc,k , Wp,k }, ∀i ∈ {1, . . . , Qk }, resulting in one common
(1) (Q )
in Table XI. Again, MIMO NOMA and OMA are subschemes message vector wc,k = [Wc,k , . . . , Wc,k k ]T and one private
(1) (Q )
of MIMO RSMA in the uplink. Though we here did not split message vector wp,k = [Wp,k , . . . , Wp,kk ]T for user-k. The
user-2 messages, other variants of uplink RSMA schemes can common messages vectors of all users {wc,1 , . . . , wc,K } are
also be derived by also splitting w2 . combined into Qc (Qc ≤ min(M, N )) common messages
wc ∈ CQc ×1 and encoded into one common stream vector
(1) (Q )
C. Lessons Learned sc = [sc , . . . , sc c ]T , which is decoded by all users.
• OMA, SDMA, and NOMA are respectively based on The private message vector wp,k of user-k is independently
(1) (Q )
the orthogonalization, treat interference as noise, and encoded into the private stream vector sk = [sk , . . . , sk k ]T
decode interference strategy. This limits the application to be decoded by user-k only. The K + 1 data stream
of each of those MA schemes to a specific interference vectors sc , s1 , . . . , sK are linearly precoded by the precoders
regime. Pc , P1 , . . . , PK and superposed at the transmitter. The result-
• RSMA relies on the RS interference strategy and is ing transmit signal is
efficient in all interference regimes. K
X
• RSMA bridges, unifies, and generalizes OMA, SDMA x = Pc sc + Pk sk . (38)
and NOMA. k=1
• RSMA is a superset of OMA, SDMA and NOMA,
The signal received at user-k is
and can specialize to each of them depending on how
messages are mapped to streams. K
X
• RSMA is applicable to both downlink and uplink, to y = HH
k Pc sc + HH
k Pj sj + nk . (39)
SISO, MISO and MIMO settings. j=1

• RSMA can be extended to the cooperative relay setting User-k employs SIC or joint decoding to decode two stream
and to space-time/frequency transmission. vectors sc and sk . Only one single layer of SIC is required
 17

(1) (Q )
[sK , . . . , sK c ]T , which is decoded by all users. For group-
g, g ∈ G, the inner-group common message vectors of users
K
in group-g {wk g , ∀k ∈ Kg } are combined into Qcg (Qcg ≤
min(M, N )) common messages wKg ∈ CQcg ×1 and encoded
(1) (Q )
into one common stream vector sKg = [sKg , . . . , sKgcg ]T
to be decoded by all users in group-g. The private message
vector wkk of user-k is independently encoded into the private
(1) (Q )
stream vector sk = [sk , . . . , sk k ]T to be decoded by
user-k only. The overall K + G + 1 data stream vectors
sK , sK1 , . . . , sKG , s1 , . . . , sK are linearly precoded by the cor-
responding precoders PK , PK1 , . . . , PKG , P1 , . . . , PK . The
resulting transmit signal is
Fig. 15. K-user downlink 1-layer MIMO RSMA.
G
X K
X
x = PK sK + PKg sKg + Pk sk . (42)
at each user if employing SIC to decode the data streams. g=1 k=1
Assuming Gaussian signalling, the instantaneous rates of de- The signal received at user-k is
coding the common and private stream vectors based on SIC G K
are given by X X
  y = HH
k PK sK + HH
k PKg sKg + HH
k Pj sj + nk . (43)
−1
Rc,k = log2 det I + PH
c Hk (Rc,k ) HHk Pc ,
g=1 j=1
  (40) In contrast to 1-layer RS where each user only decodes two
−1
Rk = log2 det I + PHk Hk (Rk ) HHk Pk , streams, in HRS, each user-k (k ∈ Kg ) employs SICs or
joint decoding to decode three stream vectors sK , sKg , and
where Rc,k and Rk are noise plus interference covariance
sk . Two layers of SIC are required at each user if employing
matrices defined as
SIC. Assuming Gaussian signalling, the instantaneous rates of
K
X decoding sK , sKg , and sk based on SIC are given by
Rc,k = I + HH H
k Pj Pj Hk ,  
−1
j=1
(41) RK,k = log2 det I + PH K Hk (RK,k ) HHk PK ,
K
X 
−1 H 
Rk = I + HH H
k Pj Pj Hk . RKg ,k = log2 det I + PH Kg H k R K g ,k H k P K g , (44)
j=1,j6=k  
−1
Rk = log2 det I + PH k Hk (Rk ) HH k Pk ,
To guarantee the common stream vector is successfully
decoded by all users, it is achievable rate is Rc = where the noise plus interference covariance matrices RK,k ,
mink {Rc,1 , . . . P
, Rc,K }. As sc contains sub-messages of K RKg ,k , and Rk are defined as
users, we have k∈K Ck = Rc , where Ck denotes the portion G K
of Rc allocated to user-k for the transmission of wc,k . The X X
RK,k = I + HH H
k PKg PKg Hk + HH H
k Pj Pj Hk ,
overall achievable rate of user-k is Rk,tot = Ck + Rk . g=1 j=1
2) Hierarchical RS: Hierarchical RS (HRS) is proposed in G K
X X
[58] for MISO BC where the transmitter is equipped with RKg ,k = I + HH H
HH H
k PKi PKi Hk + k Pj Pj Hk ,
multiple antennas and each receiver is equipped with a single i=1,i6=g j=1
antenna. In this subsection, we extend the system model G K
X X
to MIMO BC where each node has multiple antennas. In Rk = I + HH H
HH H
k PKi PKi Hk + k Pj Pj Hk .
HRS, users are clustered into G separate groups indexed by i=1,i6=g j=1,j6=k
G = {1, . . . , G} according to the similarity of their channel (45)
covariance matrices.
S Each group-g contains a subset of users To guarantee the successful decoding of the inter-group and
Kg such that g∈G Kg = K. In contrast to 1-layer RS inner-group common stream vectors, the achievable rates fol-
where each user message is split into two parts, in HRS, low RK = mink {RK,k , k ∈ K} and RKg = mink {RKg ,k , k ∈
(i)
each user message Wk , i ∈ {1, . . . , Qk } in wk for user- Kg }. As sK and sKg contain sub-messages of multiple users,
P P K
k (assuming user-k is in group g) is split into three sub- we have k∈K CkK = RK and k∈Kg Ck g = RKg , where
(i) (i) (i) (i)
messages as Wk = {WK,k , WKg ,k , Wk,k }, resulting in three K
CkK and Ck g respectively denote the portion of RK and RKg
sub-message vectors, namely, an inter-group common message K
(1) (Q ) allocated to user-k for the transmission of wkK and wk g . The
vector wkK = {WK,k , . . . , WK,kk }, an inner-group common K
K (1) (Qk ) overall achievable rate of user-k is Rk,tot = CkK + Ck g + Rk .
message vector wk g = {WKg ,k , . . . , WKg ,k }, and a private It is easy to observe that HRS is a more general framework
(1) (Q )
message vector wkk = {Wk,k , . . . , Wk,kk }. The inter-group than 1-layer RS for MIMO BC. HRS boils down to 1-layer RS
common messages of all users {w1K , . . . , wK
K
} are combined when the inner-group common stream vectors {sKg , ∀g ∈ G}
into Qc (Qc ≤ min(M, N )) common messages wK ∈ are turned off, i.e., PKg = 0, ∀g ∈ G. In this case, sK for
CQc ×1 and encoded into one common stream vector sK = HRS is equivalent to sc for 1-layer RS.
18

Fig. 17. 3-user downlink MIMO generalized RS.

Following the decoding order from the K-order stream vector

down to the 1-order stream vector as well as a certain decoding
order πl,k to decode the l-order stream vectors in Sl,k (i.e.,
sπl,k (i) is decoded before sπl,k (j) if i < j), we obtain the
Fig. 16. 4-user downlink MIMO HRS.
rate of decoding l-order stream vector sπl,k (i) at user-k under
Gaussian signalling and SIC decoding as
In Fig. 16, a four-user two-group HRS example is illustrated Rπl,k (i),k =
with user-1 and user-2 in group-1, user-3 and user-4 in group- 
−1 H 
log2 det I + PH πl,k (i) Hk Rπl,k (i),k Hk Pπl,k (i) ,
2. The user sets K = {1, 2, 3, 4}, K1 = {1, 2}, K2 = {3, 4}
are simply denoted as 1234, 12, 34, respectively. User-1 and (47)
user-2 are required to decode the inter-group common stream where Rπl,k (i),k is given as
vector s1234 and the inner-group common stream vector s12 X
Rπl,k (i),k = I + HH H
k Pπl,k (j) Pπl,k (j) Hk
while user-3 and user-4 are required to decode the s1234 and j>i
s34 .
l−1 |S
X X l′ ,k |
3) Generalized RS: Generalized RS (GRS) is proposed in
+ HH H
k Pπl′ ,k (j) Pπl′ ,k (j) Hk (48)
[8] for MISO BC to further enhance the spectral efficiency
l′ =1 j=1
by splitting and encoding multiple common streams, which X
are required to be decoded by different subsets of users. It + HH H
k PA′ PA′ Hk .
is a generalized transmission framework that embraces 1- A′ ⊆K,k∈A
/ ′
layer RS, HRS, linearly precoded MU-MIMO, and MIMO To guarantee the successful decoding of each stream vector sA ,
NOMA as special cases. We here extend it to MIMO BC. At the achievable rates follow RA = mink {RA,k , k P ∈ A}. As sA
the transmitter, each user message is split into 2K−1 parts, contains sub-messages of users in A, we have k∈A CkK =
(i) (i)
i.e., Wk = {WA′ ,k | A′ ∈ K, k ∈ A′ }. The message RA , where CkA denotes the portion of RA allocated to user-
vector wk is therefore split into 2K−1 sub-message vectors k for the transmission A
P of wkA′. The overall achievable rate of
′ ′ (1) (Qk )
wkA , ∀A′ ∈ K, k ∈ A′ } with wkA = {WA′ ,k , . . . , WA′ ,k }. user-k is Rk,tot = k∈A′ Ck + Rk,k . GRS is a more general
A framework than 1-layer RS and HRS for MIMO BC. It boils
The sub-message vectors {wk | k ∈ A} for all users in
a given subset A ∈ K are combined and encoded into one down to 1-layer RS when only the K-order stream vector sK
stream vector sA of dimension QcA × 1 which is only decoded and the 1-order stream vectors sk , k ∈ K are active. sK for
by users in A and treated as noise by the remaining users. GRS is equivalent to sc for 1-layer RS in such case. It reduces
Following the concept of stream order introduced in GRS for to HRS when only sK , sk , k ∈ K, and sKg , g ∈ G are active.
MISO BC [8], we denote the stream vectors to be decoded A toy 3-user GRS example is illustrated in Fig. 17. The
by l number of users as l-order stream vectors. All l-order user message vector of each user is split into 4 sub-message
stream vectors {sA′ | A′ ∈ K, |A′ | =Pl} are wrapped up into vectors. For user-1, w1 is split into {w1123 , w112 , w113 ,
a larger dimensional vector sl ∈ C A′ ∈K,|A′ |=l QcA′ ×1 and w11 }, each is required to be decoded by different groups of
linearly precoded by the precoding matrix Pl composed by users. {w1123 , w2123 , w3123 } are combined into w123 , which is
{PA′ | A′ ∈ K, |A′ | = l}. The resulting transmit signal is then encoded into the 3-order stream vector s123 . Similarly,
K K
{w112 , w212 } are combined into w12 and encoded into the 2-
X X X order stream vector s12 . Following this method, we obtain in
x= Pl sl = PA′ sA′ . (46)
total 7 transmit stream vectors. 3 layers of SIC is adopted
l=1 l=1 A′ ⊆K,|A′ |=l
at each user to decode the intended 4 stream vectors. In this
At user-k, 2K−1 − 1 layers of SIC or joint decoding is example, user-1 decodes s123 , s12 , s13 , s1 sequentially with
employed to decode the required streams vectors. The set sπ2,1 (1) = s12 and sπ2,1 (2) = s13 .
of l-order stream vectors required to be decoded at user-k 4) Dirty paper coded RS: Dirty paper coded RS (DPCRS)
is denoted by Sl,k = {sA′ | A′ ⊆ K, |A′ | = l, k ∈ A′ }. is a non-linearly precoded RS framework built upon DPC. It is
 19

proposed in [9] for MISO BC and is extended to MIMO BC in

this subsection. We first illustrate the transmission architecture
of the simplest DPCRS model 1-layer Dirty Paper Coded RS
(1-DPCRS), and then briefly discuss its extension to multi-
layer dirty paper coded rate-splitting (M-DPCRS).
Fig. 18 illustrates the proposed 1-layer MIMO DPCRS.
It is an extension of 1-layer RS in Section IV-A1 by en-
abling dirty paper encoded private streams. Following 1-layer
RS, the K user messages are split and combined into one
common message vector wc and K private message vectors
wp,1 , . . . , wp,K . wc is encoded into a common stream vector
Fig. 18. K-user downlink 1-layer MIMO DPCRS.
sc using a codebook shared by all users while wp,1 , . . . , wp,K
are encoded and precoded by DPC for a certain encoding
order π into the private stream vectors sπ(1) , . . . , sπ(K) . The
resulting transmit signal is
K
X
x = Pc sc + Pπ(k) sπ(k) , (49)
k=1

where Pc , Pπ(1) , . . . , Pπ(K) are the precoders for the corre-

sponding streams sc , sπ(1) , . . . , sπ(K) .
Each user-π(k) decodes sc and sπ(k) based on SIC or joint
decoding. Assuming Gaussian signalling, the instantaneous
rates of decoding the common and private stream vectors based
on SIC are given by

−1 H 
Rc,π(k) = log2 det I + PH c H π(k) R c,π(k) H π(k) c ,
P Fig. 19. 3-user downlink multi-layer MIMO DPCRS.


−1
Rπ(k) = log2 det I + PH π(k) Hπ(k) Rπ(k) HH
π(k) Pπ(k) ,
(50) Fig. 20 illustrates the proposed K-user uplink MIMO
where Rc,π(k) and Rπ(k) are noise plus interference covari- RSMA based on SC at the transmitters and SIC at the receiver.
(1) (Q )
ance matrices defined as A Qk -dimensional message vector wk = [Wk , . . . , Wk k ]T
is transmitted from user-k to the receiver, where Qk ≤
K
X (i)
min(M, N ). At user-k, each user message Wk is split
Rc,π(k) = I + HH H
π(k) Pπ(j) Pπ(j) Hπ(k) , (i) (i)
j=1
into two sub-messages Wk,1 and Wk,2 , resulting in two
(51) sub-message vectors wk,1 and wk,2 . Note that splitting
K
X PK
HH k=1 Qk − 1 messages is sufficient to achieve the capacity
H
Rπ(k) = I + π(k) Pπ(j) Pπ(j) Hπ(k) .
j>k region without time sharing in this setting. For illustration
simplicity, we here consider a more general model in which
The calculation of the overall achievable rate for each user the user messages of all users are split. At user-k, the sub-
follows 1-layer RS, which is not detailed here. Besides 1- message vectors wk,1 and wk,2 are independently encoded
layer MIMO DPCRS, we can also integrate DPC and different into sk,1 and sk,2 , precoded by Pk,1 and Pk,2 , and superposed
RS schemes to enhance the system performance especially at the transmitter. The resulting transmit signal is
in imperfect CSIT. For instance, Fig. 19 illustrates another
DPCRS scheme, namely, multi-layer MIMO DPCRS, which xk = Pk,1 sk,1 + Pk,2 sk,2 , ∀k ∈ K. (52)
marries DPC and GRS. In contrast to GRS in Fig. 17 which
relies on linear precoding to precode all streams, multi-layer The signals of all users are transmitted via MIMO MAC,
MIMO DPCRS considers dirty paper coded private streams. the receive signal is
X
y= Hk xk + n, (53)
B. Uplink k∈K

M×N
Uplink RSMA is first introduced in [23] for SISO MAC and where Hk ∈ C is the channel vector between user-k and
further extended to SIMO MAC in [25]. The main benefit of the receiver, and n ∼ CN (0, IM ) is the AWGN vector. The
uplink RSMA discovered in existing works is that the capacity receiver then employs the receive filters Wk,1 , Wk,2 , k ∈ K
region of the Gaussian MAC can be achieved by uplink to detect the stream vectors for all users. Denote the decod-
RSMA without time sharing among users. In the following, ing order of the 2K received stream vectors {sk,i | k ∈
we introduce a generic uplink RSMA transmission framework {1, . . . , K}, i ∈ {1, 2}} by π such that πk,i < πk′ ,i′ if sk,i
for MIMO MAC, which is simply denoted by uplink MIMO is decoded before sk′ ,i′ . 2K − 1 layers of SIC is needed to
RSMA. decode all stream vectors. Assuming Gaussian signaling and
20

Fig. 20. K-user uplink MIMO RSMA. (a) Coordinated transmission (b) Cooperative transmission

Fig. 21. Multi-cell RSMA-enabled transmission [25].

infinite blocklength, the rate of decoding ski at the receiver is

given as follows where HH kj ∈ C
N ×M
is the channel between transmitter-j and
  user-k and nk ∼ CN (0, IN ) is the AWGN vector at user-k.
−1
Rk,i = log2 det I + PH H H
k,i Hk Wk,i (Rk,i ) Wk,i Hk Pk,i , Each user-k is required to decode all common stream vectors
(54) sc,1 , . . . , sc,K and the intended private stream sp,k based on
where Rk,i is defined as SIC or joint decoding. For a certain decoding order πk such
X that sc,πk (i) is decoded before sc,πk (j) at user-k if i < j.
Rk,i = I + Wk,i Hk′ Pk′ ,i′ PH H H
k′ ,i′ Hk′ Wk,i .
πk′ ,i′ >πk,i
Assuming Gaussian signaling and infinite blocklength, the rate
(55) of decoding sc,πk (i) and sp,k at user-k are obtained as follows
The proposed uplink MIMO RSMA model is generic. It c
Rk,π k (i)
=
embraces uplink SISO RSMA in [23] and uplink SIMO RSMA 
−1 H 
in [25] as two sub-schemes. The study of uplink MIMO log2 det I + PH c,πk (i) Hkπk (i) Rc,πk (i) Hkπk (i) Pc,πk (i) ,
 
RSMA is still in its infancy. Its SE/EE performance and −1
Rkp = log2 det I + PH p,k Hkk (Rk ) HH kk Pp,k .
applications in different services such as eMBB, URLLC, and
mMTC or their hybrid services are worth more investigations. (58)
where Rc,πk (i) and Rk are defined as
X
C. Multi-cell Rc,πk (i) = I + HH H
kπk (j) Pc,πk (j) Pc,πk (j) Hkπk (j)
j>i
In multi-cell networks, the transmission from multiple trans-
K
X
mitters to multiple receivers can be categorized into “coordi-
nated transmission" and “cooperative transmission" depending + HH H
kj Pp,j Pp,j Hkj , (59)
j=1
on whether the data is shared among the transmitters [6]. X
RSMA has been investigated in both transmissions [59]– Rk = I + HH H
kj Pp,j Pp,j Hkj .
[62]. It is shown to be a promising strategy to enhance j∈K,j6=k
SE by providing a powerful inter-cell and intra-cell inter- The achievable rate of user-k is
ference management capability. In the following subsections,  c
Rk,tot = min R1,k c
, . . . , RK,k + Rkp . (60)
the transmission models of MIMO RSMA for “coordinated
transmission" and “cooperative transmission" are respectively As discussed in [59], The performance of coordinated
delineated. MIMO RSMA could be further boosted by splitting the
1) Coordinated transmission: A K-cell coordinated MIMO message of each user into N > 2 parts. Each part is decoded
is illustrated in Fig.21(a). The message of each user is sent by a different group of users with the same instantaneous CSIT
from the serving cell (with one RSMA-enabled transmitter) quality.
while the resource allocation and user scheduling are coordi- 2) Cooperative transmission: As discussed in [25], coop-
nated among cells. At each transmitter-k, the message vector erative MIMO requires all transmitters to be connected as a
wk for user-k is split into two sub-message vectors wc,k and virtual giant transmitter to serve all users in the downlink.
wp,k , independently encoded into two stream vectors sc,k and All downlink transmission frameworks of RSMA discussed
sp,k , and linearly precoded by the precoding matrices Pc,k in IV-A therefore can be applied for cooperative MIMO.
and Pp,k . The transmit signal at transmitter-k is given as Compared with single-cell MIMO BC which is subject to
a sum-power constraint, multi-cell cooperative MIMO BC
xk = Pc,k sc,k + Pp,k sp,k . (56) requires per-cell transmit power constraint and the fronthaul
The signal of all transmitters are transmitted to all users. capacity constraints [62].
The signal received at user-k is
D. Numerical Results
K
X
yk = HH Next, we show numerically the benefits of RSMA in terms
kj (Pc,j sc,j + Pp,j sp,j ) + nk , (57)
j=1
of the significant wireless communication performance met-
 21

15 8 10
DPCRS 5
DPC
RS
ER User-2 (bps/Hz)

ER User-2 (bps/Hz)
6 8
NOMA 4
10 SDMA

ER2,tot (bit/s/Hz)

ER2,tot (bit/s/Hz)
4 6 3

5
RS RS 4
2 2
MU-MIMO MU-MIMO
NOMA NOMA DPCRS
DPC DPC DPC
2 1 RS
0 0
0 5 10 15 0 5 10 15 NOMA
ER User-1 (bps/Hz) ER User-1 (bps/Hz) SDMA
0 0
0 2 4 6 8 10 0 2 4 6 8 10
(a) σ12 = 1, σ22 = 1 (b) σ12 = 1, σ22 = 0.09 ER1,tot (bit/s/Hz) ER1,tot (bit/s/Hz)

(a) σ12 = 1, σ22 = 1 (b) σ12 = 1, σ22 = 0.09

Fig. 22. Ergodic rate region comparison in MIMO BC with perfect CSIT,
SNR = 20 dB, M = 4, K = 2, N = 2, and Qc = Q1 = Q2 = 2 [35].
Fig. 23. Ergodic rate region comparison in MISO BC with imperfect CSIT,
SNR = 20 dB, α = 0.6, M = 4, K = 2 [9].

rics, namely, spectral efficiency, energy efficiency, generality,

7
flexibility, robustness, reliability and latency. 8

Energy Efficiency (bits/J/Hz)

Energy Efficiency (bits/J/Hz)
Unless stated otherwise, the elements of the channel matrix 7 6

Hk are generated as i.i.d. complex Gaussian random variables 6

5

CN (0, σk2 ) and the noise variances are normalized. The imper- 5

fect CSIT model is Hk = H bk + H e k , where both the channel 4

4

b
estimate Hk and the channel error H e k have i.i.d. complex 3 RS 3 RS
NOMA NOMA

Gaussian random entries drawn from CN (0, σk2 − σe,k 2

) and 2
SDMA SDMA

4
6
8
CN (0, σe,k ), respectively. All results are averaged over 100

10
12
14
16
18
20
22
24
26
28
4
6
8
10
12
14
16
18
20
22
24
26
28
30
P (dBm) P (dBm)

channel instances. (a) σ12 = 1, σ22 =1 (b) σ12 = 1, σ22 = 0.09

1) Spectrally and Energy Efficient: Consider a downlink
Fig. 24. Energy efficiency versus transmit power budget comparison in MISO
MU-MIMO with one transmitter equipped with M = 4 BC with perfect CSIT.
antennas and K = 2 users, each is equipped with N = 2
antennas. Each user requires Q1 = Q2 = 2 streams, which
are split, combined, and encoded into a common stream vector In Fig. 23, we further compare SE of different strategies
of length Qc = 2 and two private stream vectors of length 2. in imperfect CSIT. Consider a downlink MU-MISO with
Readers are referred to [35] for more details on the simulation M = 4 and K = 2. The power of the channel error is
setting and the rate region optimization algorithm. Fig. 22 defined as σe,k2
= σk2 P α , where P = tr(PPH ) and α
illustrates the ergodic rate region of RS, MU-MIMO, NOMA, is fixed to 0.6. More details of the simulation setting and
and DPC in perfect CSIT with equal channel variances or the rate region optimization algorithm can be found in [9].
10 dB channel variance disparity between the users. Note Besides linearly precoded RS, the ergodic rate region of 1-
that in the two-user case, HRS and GRS are equivalent to layer DPCRS is also illustrated. In both subfigures, the ergodic
1-layer RS, which are therefore simply denoted by “RS". rate region of DPC drops significantly as it is sensitive to
We observe that in both subfigures, the rate region of RS CSIT uncertainty. Surprisingly, linearly precoded RS achieves
is larger than that of MU-MIMO and NOMA, and is much a larger rate region than DPC while maintaining a much lower
closer to the capacity region of MIMO BC achieved by transceiver complexity. As RS has been proved to achieve the
DPC. When the users have equal channel variances as in Fig. optimal multiplexting gain/DoF of MISO BC with imperfect
22(a), NOMA achieves the worst performance as no channel CSIT [19], such gain is reflected in its spectral efficiency
strength disparity can be leveraged to manage interference. gain over MU-MIMO and NOMA. DPCRS, which marries the
Although MU-MIMO is capable of exploiting the full DoF advantages of DPC and RS, achieves the largest achievable rate
in this setting (i.e., underloaded and perfect CSIT), there is region in both subfigures at the sacrifice of a higher transceiver
still a rate region gap between RS and MU-MIMO. As MU- complexity than DPC and RS.
MIMO always treats interference as noise, it only works well
Fig. 24 compares EE of different strategies for two-user
when user channels are (semi-)orthogonal [8]. When the users
MU-MISO
PK with perfect CSIT, M = 2, N = 2. EE is defined
have a 10 dB channel variance disparity, the rate region gap
k=1 Rk,tot
between NOMA and RS reduces. Interestingly, the rate region as 1 tr(PP H )+P [63], [64], where η = 0.35 is the power
η cir

of NOMA and MU-MIMO outperform each other in part while amplifier efficiency, Pcir = M Pdyn + Psta is the static circit
RS always outperforms both. Although NOMA can utilize power consumption with Pdyn = 27 dBm and Psta = 1 mW.
the channel variance disparity to improve its rate region, it There is a 1 bit/s/Hz minimum rate threshold for each user.
incurs a theoretical DoF loss and therefore rate loss due to RS achieves a larger EE than NOMA and SDMA in both
its interference management principle of forcing one user to subfigures. Again, NOMA achieves the EE worst performance
fully decode the streams of all users [7]. Owing to its powerful when there is no channel strength disparity while it slightly
interference management capability of partially decoding the outperforms SDMA when there is a 10 dB channel strength
interference and partially treating the interference as noise, RS disparity and low transmit power constraint.
enhances the spectral efficiency in various user deployments. Fig. 22–24 show the SE and EE gains of RSMA in the
22

[dB]
[dB]
0 0

5.4

dB
dB
MULTICAST RS SDMA
-5 -5

channel strength disparity

channel strength disparity
NOMA RS SDMA
5.2
-10 -10
Weighted Sum Rate (Mbits/s)

5 -15 -15
OMA

NOMA
4.8 -20 -20
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

4.6
(a) SNR= 10 dB (b) SNR= 30 dB
4.4
Fig. 26. Operation regions for RS, SDMA, NOMA, OMA and Multicast with
perfect CSIT [24].
4.2
RSMA
NOMA
4
FDMA
TDMA NOMA with three user groups (G = 3) which clusters 6
3.8
0.5 1 1.5 2 users into 3 user groups and inner-group users are served
Transmit power budget (dBm) by SC and SIC. Readers are referred to [7] for more details
on the parameter settings and baseline schemes of MISO
Fig. 25. Sum rate versus the transmit power budget of each user in SISO
MAC with perfect CSIT [65]. NOMA. When the number of transmit antennas is M = 3,
the network load is extremely overloaded. In this case, MISO
NOMA (G = 1) achieves a better MMF rate than SDMA and
downlink for different user deployments and CSIT conditions. MISO NOMA (G = 3). However, as the number of transmit
In the uplink, the authors in [65] have also shown that RSMA antennas increases, the performance of MISO NOMA (G = 1)
achieves a higher spectral efficiency than NOMA and other decreases due to a significant loss in DoF. MU–LP outperforms
OMA strategies, as illustrated in Fig. 25 thanks to its capability the two MISO NOMA when the network is underloaded
of achieving each point at the boundary of the capacity region. (M = 6). In all subfigures, 1-layer RS, which only uses a
In contrast, NOMA can only reach the corner points at the single layer of SIC at each user, achieves the best MMF rate
boundary of the capacity region and time sharing is needed no matter the network is underloaded or overloaded. As per
to achieve the points along the line segment between the two Fig. 22–27, RSMA is a flexible MA scheme which is suited
corner points. for different user deployments (diverse channel directions and
2) General and Unified: As widely acknowledged and strength), network loads (underloaded and overloaded), CSIT
discussed in the literature of RSMA [8], [9], [24], [25], condition (perfect and imperfect CSIT), and SNR regimes
[29], RSMA is a general and universal MA scheme, interfer- (low, medium, and high SNR regimes). The root of such
ence management strategy, and non-orthogonal transmission flexibility is the powerful interference management ability
framework that unifies SDMA, NOMA, OMA, physical-layer of RSMA introduced by the common stream (vectors). By
multicasting and treats them as sub-schemes or special cases. adjusting the sub-messages encapsulated in each common
Such universality is reflected in the SE and EE gain over stream as well as the power allocation, RSMA dynamically
existing MA schemes as per Fig. 22–25, as well as in its alters the portion of interference to be pre-canceled at the
operational region illustrated in Fig. 26. In Fig. 26, a specific transmitter and decoded at the receivers.
channel realization h1 = √12 [1, 1]H and h2 = √γ2 [1, ejθ ]H is 4) Robust: Robustness is one of the most important benefits
|hH h |2 of RSMA discovered in recent years [9], [18], [19], [67].
considered. The x-axis is ρ = 1 − kh1 k12 kh2 2 k2 , which indicates
the channel angle between the two users (ρ = 0 and ρ = 1 The robustness of RSMA is grounded in deep information
corresponding to aligned and orthogonal channel directions, theoretic results, where RSMA is shown to achieve the optimal
respectively). The y-axis is the channel disparity between the DoF in MISO BC when CSIT is imperfect [19]. Motivated
two users in dB, which is given as γdB = 20 log10 (γ). The by such discovery, different sources that impair CSIT are
parameters and the precoder design are detailed in [24]. RS studied with RSMA such as pilot contamination [68], channel
automatically reduces to the existing MA schemes in some estimation errors [18], user mobility [66], etc. RSMA is shown
regimes while it outperforms all existing schemes in the re- a significant SE gain over existing MA baselines for all sources
maining regimes especially when the SNR is high. Therefore, of CSIT impairments. Taking user mobility as an example
instead of using OMA, NOMA, SDMA and optimize them in Fig. 28, we have shown in [66] that RSMA achieves a
for each propagation deployment, one can use a single and significantly higher user speed (i.e., 40 km/h) than SDMA
universal RS scheme in wireless communication networks. (i.e., 5 km/h) for a given QoS rate constraint (i.e., 8 bps/Hz)
3) Flexible: Fig. 27 illustrates the impact of network load thanks to its powerful interference management ability.
to the system performance for a MISO BC with K = 6 5) Reliable and Low Latency: : As shown in [32], [34]
users and imperfect CSIT with α = 0.5. The variances for downlink and in [69] for uplink, RSMA is more reliable
of user channels are randomly selected from [0.1, 1]. Two than SDMA and NOMA under finite blocklength. This is
MISO NOMA schemes are considered as baseline schemes, illustrated in Fig. 29 for underloaded and overloaded downlink
namely, MISO NOMA with a single user group (G = 1) settings where we note that RSMA can achieve a given
which is motivated by SC and SIC in SISO BC, and MISO MMF rate at a lower blocklength, therefore enabling lower
 23

2 3
1-layer RS
E. Lessons Learned
1-layer RS
MISO NOMA (G=3) MISO NOMA (G=3)

1.5
MISO NOMA (G=1)
SDMA
2.5 MISO NOMA (G=1)
SDMA
• RSMA framework is a superset of OMA, SDMA and
MMF Rate (bit/s/Hz)

MMF Rate (bit/s/Hz)

2 NOMA, and can specialize to each of them depending
1 1.5 on how messages are mapped to streams. This holds for
1 both uplink and downlink, as well as in SIMO, MISO
0.5
0.5 and MIMO settings.
0 0
• All instances of RSMA are supersets of SDMA and
0 5 10 15 20 25 30 0 5 10 15 20 25 30
SNR (dB) SNR (dB) OMA, but not all, such as 1-layer RS scheme, are
(a) Overloaded M = 3 (b) Overloaded M = 4 supersets of NOMA. This enables RSMA schemes with
better performance than NOMA at a much lower receiver
4
1-layer RS
6
1-layer RS complexity (e.g., only one SIC layer).
3.5 MISO NOMA (G=3) MISO NOMA (G=3)

3
MISO NOMA (G=1)
SDMA
5 MISO NOMA (G=1)
SDMA
• RSMA can be linearly precoded or non-linearly precoded.
MMF Rate (bit/s/Hz)

MMF Rate (bit/s/Hz)

2.5
4 • RSMA is applicable to all major settings of a cellular
2 3 network, namely downlink, uplink, and multicell, for
1.5
2 general MIMO deployments (with SISO, SIMO, MISO
1
1
as special cases).
0.5

0 0
• RSMA is spectrally and energy efficient, general and
0 5 10 15 20 25 30 0 5 10 15 20 25 30
SNR (dB) SNR (dB) unified, flexible, robust, reliable and has lower latency.
(c) Overloaded M = 5 (d) Underloaded M = 6

Fig. 27. Max-min rate vs. SNR comparison in MISO BC with imperfect V. N UMEROUS A PPLICATIONS FOR RSMA
CSIT, α = 0.5, K = 6, σk2 ∈ [0.1, 1] [7].
Given its fundamental communication theoretic principles,
and its unique features (efficiency, universality, flexibility,
robustness and resilience, reliability and low latency), RSMA
22
15

SDMA
20
SDMA, SNR=25dB
SDMA, SNR=35dB
RSMA, SNR=25dB
finds applications in all modern multi-user scenarios. We here
18
RSMA
16
RSMA, SNR=35dB
provide a description of over forty promising scenarios and ap-
Throughput [bps/Hz]

10
Throughput [bps/Hz]

14
12 plications of RSMA and briefly explain how and why RSMA
10

5
8 is beneficial11 . A subset of those scenarios and applications is
6
4
2
illustrated in Fig. 30. All those applications demonstrate the
0
5 10 15 20 25
User Speed (km/h)
30 35 40
0
3 7 11 15 19 23 27 31 35 39 42
suitability of RSMA for FeMBB, eURLLC, umMTC, and new
User Speed (km/h)
wireless services in 6G [28], [29]. Importantly, the benefits and
(a) M = 32, K = 8, N = 1, Qc = (b) M = 64, K = 8, Nr =
Qp = 1 4, Qc = Qp = 3 applications of RSMA have not only been explored from an
academic perspective with the performance superiority over
Fig. 28. Throughput vs. user speed comparison in (massive) MIMO BC with other MA techniques confirmed via stochastic analysis [70]–
outdated CSIT employing OFDM and a 3GPP channel model, and 10 ms CSI
feedback delay [66]. [73] but have also been confirmed using realistic link-level
simulations over 5G compliant channel models [33], [35],
[74]–[77]. Interested readers are also encouraged to consult
[25], [26], [38], [39] for some more descriptions of interesting
4.8 1.4
applications and future works.
MMF Rate (bit/s/Hz)
MMF Rate (bit/s/Hz)

4.6 1.2
4.4 1) Downlink SISO: RSMA can be used for both degraded
1
4.2 or non-degraded BC. In the degraded BC, RSMA boils down
0.8
4 to NOMA (since NOMA is capacity achieving for degraded
0.6
3.8 BC) but RSMA can be used to decrease the receiver complex-
3.6 0.4
0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500 ity and still come close to the capacity region with a reduced
Blocklength (bit) Blocklength (bit)
number of SIC layers compared to NOMA [8]. In the non-
(a) Underloaded M = 4, K = 2 (b) Overloaded (M=4, K=8) degraded BC, RSMA achieves a strictly larger rate region than
Fig. 29. MMF rate versus blocklength (solid lines) of downlink SDMA NOMA (see Remark 4).
(red), NOMA (yellow), RSMA (blue) at SNR=20dB [34]. Dashed line is 2) Downlink MISO: RSMA can be used in various MISO
the upperbound achieved with infinite blocklength. settings with perfect and imperfect CSIT. The design and pre-
coder optimization of RSMA would depend on the objective
value, e.g., WSR, MMF, EE, and the modeling of the CSIT
imperfection, e.g., unbounded vs bounded error [8], [18], [48],
latency communications, compared to other MA schemes. It
[64], [78]–[81]. One challenge with imperfect CSIT is to make
is important to recall here again that 1-layer RS scheme of
sure that the rate optimized are achievable. 1-layer RS is the
RSMA is used in Fig. 29, and therefore requires a single
most common scheme but other schemes based on multi-layer
SIC. This is in contract with NOMA that achieves lower
performance at the expense of the need for 7 SIC layers at 11 Other promising scenarios and applications are also discussed in Section
the receivers in Fig. 29(b). VII on frequently asked questions.
24

Fig. 30. A subset of promising scenarios and applications of RSMA.

 25

RSMA can also be used [8]. A strong benefit of RSMA is its ination due to multiple users sharing the same12 pilot sequence
inherent robustness to imperfect CSIT. when performing uplink channel sounding and channel aging
3) Downlink MIMO: RSMA can be used in settings with due to mobility and latency between the channel sounding
multiple receive antennas at the users. MIMO RSMA en- phase and the data transmission phase. Those two issues
ables to transmit vectors of common and private streams, effectively lead to imperfect CSIT that need to be tackled.
which requires a special design [22], [35], [82]–[86]. Various RSMA has been shown robust to mitigate both problems [66],
CSIT/CSIR assumptions and objective values can also be [68], [99]
considered in the MIMO setting. 11) Hardware Impairments: Another impairment in Mas-
4) Uplink: Uplink RSMA avoids the need of time sharing sive MIMO comes from the hardware, e.g., phase noise,
to achieve the capacity region, which finds new applications digital-to-analogue (DAC) and analogue-to-digital (ADC) con-
in modern systems subject to latency constraints [23], [65], verters. Phase noise can effectively lead to a scenario where
[69], [87]–[89]. Much attention has been brought to the uplink the CSIT is imperfect and RSMA has been shown to be robust
SISO RSMA, but uplink multi-antenna RSMA is also possible against hardware impairments such as phase noise [100].
as shown in Section IV-B. Quantization errors due to finite resolution DAC and ADC
5) Unifying OMA, SDMA, NOMA, Multicasting: RSMA also lead to multi-user interference that can be efficiently and
is universal and general in the sense that it encompasses flexibly managed using RSMA [101]–[103].
many different schemes (such as OMA, SDMA, NOMA, 12) Cell-Free Massive MIMO: Cell-free refers to dis-
physical-layer multicasting) as particular instances [7], [8], tributed antennas systems, which is subject to the same pilot
[24], [90]. It is interesting to study and understand under what contamination problem as conventional TDD massive MIMO.
propagation conditions, e.g., disparity of channel strengths and The disparity of path loss nevertheless makes the analysis
angle between user channels, RSMA boils down to each of and performance gap between schemes different. RSMA was
those schemes [24]. found robust to pilot contamination in various cell-free topolo-
gies [104].
6) Statistical CSIT: RSMA is helpful in scenarios where
13) Millimeter Wave and TeraHertz Systems: Higher fre-
only statistical CSIT (i.e., the distribution information of
quency bands are subject to their own set of challenges such
the user channels) is available, This is a particular instance
as high path loss, blocking and expensive RF chains. RSMA
of imperfect CSIT, but important in practice as it leads to
can be used and designed in conjunction with hybrid analogue-
low feedback overhead. This can be for instance a channel
digital precoding to outperform conventional SDMA strategies
covariance matrix calculated by averaging over frequency and
at high frequencies [91], [105], [106]. RSMA in combination
time resources and reported once in a while to the base
with cooperative communications can also be used to further
station. RSMA has been found robust to scenarios where only
combat the path loss and blocking issues [107].
statistical CSIT is available [58], [91]–[93]
14) Non-Linear Precoding and DPC: RSMA is commonly
7) Quantized Feedback: Quantized feedback is another designed using linear precoding due to its practical use in
popular CSI feedback mechanism that relies on a codebook to real systems. Nevertheless, non-linear precoding could also
quantize the channel. Due to quantization error, the CSIT is be used to precode the private streams, using for instance
imperfect and RSMA has been found robust to the quantization DPC, Tomlinson-Harashima precoding, vector perturbation,
error [47], [94]. Interestingly the number of feedback bits etc. Recent works have demonstrated the benefits of non-linear
needed to achieve a given sum-rate performance is decreased precoded RSMA [9], [108]–[110]. The combination of RSMA
with RSMA compared to conventional SDMA/MU-MIMO, and DPC (so-called DPCRS strategy as per Fig. 18 and Fig.
therefore enabling an overhead reduction [47], [95]. 19) is particularly suited to imperfect CSIT settings since we
8) Imperfect CSIT and CSIR: Aside imperfect CSIT, for know that DPC is capacity achieving with perfect CSIT. As the
which we know RSMA is robust, imperfect CSIR is also CSIT quality improves, the power allocation to the common
an important problem that has received less attention [96]. stream decreases, and DPCRS progressively boils down to
The rate achievability and the SIC error propagation are conventional DPC for multi-antenna BC. It is of interest to
two important issues to consider when studying RSMA with investigate how we could further enhance the performance in
imperfect CSIR. imperfect CSIT beyond that achieved by DPCRS strategy [9].
9) Frequency Division Duplex (FDD) Massive MIMO: One 15) Cooperative and Relaying Systems: Cooperative com-
popular strategy in FDD massive MIMO is to rely on two-tier munications and RSMA, i.e., cooperative RSMA as in Fig.
precoders where the first tier precoder is based on channel 10, form a happy marriage as shown in [52], [53] since a
statistics to cancel inter-cluster interference and the second tier user can decode the common stream and its private stream
based on instantaneous CSI feedback to manage intra-cluster and forward the common stream to help improving the de-
interference [97], [98]. Unfortunately if the space spanned by codability of the common stream at a cell edge user. By
the covariance matrices of the clusters overlap or if the channel doing so, RSMA can efficiently cope with a wide range of
statistics and/or the CSIT is imperfect, the performance of such propagation conditions (disparity of user channel strengths
approaches degrade and RSMA can be used to mitigate those
12 In massive machine-type communications, due to the large number of
issues [58].
devices, their sporadic access behaviour and limited coherence interval, active
10) Time Division Duplex (TDD) Massive MIMO: TDD devices have a higher chances to utilize the same pilot for uplink channel
massive MIMO also has its impairments such as pilot contam- estimation.
26

and directions), compensate for the performance degradation and hardware complexity because RSMA superiority means
due to large path loss, extend the coverage, and outperforms RSMA can afford operating with less complex RIS architec-
SDMA, NOMA, and cooperative NOMA. Those observations tures while maintaining the same overall performance [39].
also hold in the imperfect CSIT setting [111]. RSMA can 20) Multi-Cell Networks: The concept of RS for the 2-user
also find other interesting applications in multi-user relaying IC can be extended to more than 2 users, or to multiple cells
systems [112]–[115]. It would be worth further exploring how, in cellular network context [139]. Furthermore each node in
in the classical relay channel [116], as well as the cooperative each cell can be equipped with multiple transmit antennas and
communication works [117], [118], RS inherently plays a role precoders jointly with the message split need to be optimized
in the decode-and-forward protocol and its variations. across all cells to maximize the objective function accounting
16) Full Duplex: Self interference is a major problem for any potential imperfection in the CSIT [22], [59], [60],
in full duplex systems. Thanks to its flexible interference [140]–[142]. Of particular interest, [59] showed that in a multi-
management ability, RSMA can increase the range of self- cell multi-antenna setting with imperfect CSIT, the transmitters
interference over which full duplex outperforms half duplex should adopt a so called topological RS (TRS) strategy that
[119]. consists in a multi-layer structure and in transmitting multiple
17) Physical Layer Security: Security at the physical layer common messages to be decoded by groups of users rather
has attracted a lot of attention in the past decade [120]. Two in- than all users. Though TRS was studied from a communication
teresting scenarios occur for RSMA: first, the eavesdropper is and information theoretic point of view, it remains to be
not one of the users served by RSMA (hence, the eavesdropper studied how TRS could be optimized using optimization tools
is not the intended recipient of any message transmitted by the so as to maximize its performance at finite SNR.
transmitter, but only intercepts confidential messages sent to 21) Coordinated Multi-Point (CoMP): RSMA can be used
other authorized users), and second, the eavesdropper is one of for CoMP joint transmission where all base stations collabo-
the users and can therefore decode the common stream [121], rate by sharing CSI and data. In such a setting, all antennas
[122]. In the former case, the common stream can be used to at the base station act together to form a giant BC or MAC
manage the interference between users and therefore increase (depending on whether the focus is on downlink or uplink)
the sum-rate of those users but also to act as an artificial and precoders and power across streams can be optimized as a
noise (AN) to confuse the eavesdroppers. In the latter case, function of the CSI and path loss disparity subject to the per-
an eavesdropper could decode a common stream and could cell power constraints. It was shown in [61] that whenever
therefore reconstruct the original message if it can decode there is little inter-user channel strength disparity but large
the corresponding private. This leads to a secrecy constraint inter-cell channel disparity, SDMA was a suitable option. On
for the private stream and a tradeoff between sum-rate and the other hand, whenever there is a large inter-user channel
secrecy rate [123]–[125]. RSMA achieves better WSR and is strength disparity but little inter-cell channel disparity, NOMA
more robust to channel errors than SDMA while ensuring all was a better option. In comparison, RSMA always bridges,
users’ security requirements. In contrast, because the entirety generalizes, and outperforms existing SDMA and NOMA
of the message of a user is mapped to the common stream strategies. It was shown to be suited to any deployment with
in NOMA, NOMA cannot guarantee all users’ secrecy rate any inter-user and inter-cell channel disparities. Applications
constraints [123]. Other interesting analysis of RSMA in the of RSMA to uplink CoMP would be worth investigating.
presence of untrusted users has recently appeared in [126]. 22) Cloud and Fog-Radio Access Network (C/F-RAN): C-
18) Energy Efficient Networks: EE is an increasingly im- RAN consists of multiple remote radio heads (RRH) connected
portant metric. In general, the superiority of RSMA in terms of through fronthaul links to a baseband unit that performs central
spectral efficiency translates into an EE gain over conventional processing. CoMP joint transmission can be implemented
MA baselines such as OMA, SDMA and NOMA [63], [127]. across all those RRH while accounting for the challenging
Nevertheless, SE maximization and EE maximization are two feature that the fronthaul links have a limited capacity. Various
conflicting objectives in the moderate and high SNR regimes. RSMA approaches have been proposed for C-RAN: 1) RSMA
This calls for the study of the tradeoff between the two criteria. transmit signal is compressed before being transmitted over the
In [128], the performance of RSMA is shown to be superior fronthaul [143]; 2) the common and private streams of RSMA
to or equal to SDMA and NOMA in terms of SE, EE, and to be transmitted over each fronthaul are wisely selected
their tradeoff. so as to satisfy the fronthaul constraints [62], [144]–[152].
19) Reconfigurable Intelligent Surfaces (RIS): RIS is For a given fronthaul capacity constraint, both approaches
equipped with a large number of passive elements placed in the have demosntrated that RSMA is more spectrally and energy
environment that can be adjusted so as to provide a passive efficient than SDMA and NOMA.
beamforming gain and make the channel propagation more 23) Dynamic Resource Management and Cross-Layer Op-
favourable. The interplay between RIS and RSMA is very timization: In RSMA-aided networks, network resource man-
attractive and has been shown very promising by several works agement has received more and more attention. Many of these
[39], [129]–[138]. The advantages of RIS-aided RSMA are works, however, only consider the short-term optimization of
higher spectral efficiency, coverage extension and beam con- system resources without taking into account the long-term
trol flexibility thanks to the presence of the common stream, network operation constraints and objectives. Furthermore,
robustness to CSI imperfection which is welcome given the they do not consider random traffic arrivals. To meet the
channel acquisition challenge in RIS, and lower computational explosive access demands of mobile devices, the multicast
 27

communication of a satellite and aerial-integrated network 25) Overloaded Cellular Internet of Things and Massive
with RSMA is studied in [153]. Specifically, the unmanned Access: Cellular networks will have to cope with extensive
aerial vehicle (UAV) sub-network uses RSMA to support IoT devices, and consequently serve simultaneously a large
massive access of internet of things (IoT) devices in a content number of devices with heterogeneous demands and CSIT
delivery scenario. However, in practice, users may require qualities. In [160], [161], RSMA is used to tackle such a
different types of data traffic, such as high-quality video scenario by considering an overloaded MISO downlink with
streaming, voice, video phone, online games, network broad- two groups of CSIT qualities, namely, one group of users
casting, and so on. This necessitates the cross-layer design of (representative of high-end devices) for which the transmitter
RSMA-aided system to cope with traffic exposure rate and has partial knowledge of the CSI, the other group of users
long-term system constraints. It is recommended that joint (representative of IoT devices) for which the transmitter only
adaptive source encoding and cross-layer resource manage- has knowledge of the statistical CSI. RSMA is shown to
ment schemes be studied. For example, using a supervised be DoF-optimal in such a setting with heterogeneous CSIT
learning-based approach for cross-layer resource allocation qualities, more efficient than various MA baselines, robust to
in a NOMA-aided system [154], the execution time can be CSIT inaccuracy, and flexible to cope with heterogeneous QoS
reduced by 98% while ensuring that each user has at least one rate constraints of all high-end and low-end users.
subcarrier. Given the outstanding characteristics of RSMA, 26) Joint Communication and Jamming: Thanks to its
more cross-layer (across physical - PHY, medium access flexibility and robustness, RSMA can be used to efficiently
control - MAC, and application - APPL - layers) dynamic communicate to information users (IUs) and simultaneously
resource management solutions and optimizations are expected jam adversarial users (AUs) to disrupt their communications
to bring further improvement to the system performance. [162], [163]. The precoders and power allocation to common
Though most of the RSMA optimization works focus on PHY and private streams can be optimized based on imperfect CSIT
and MAC layers (beamforming and power allocation, time and for IUs and statistical CSIT for AUs to maximize the sum-
frequency resource allocation), some works apply RSMA for rate under jamming power constraints on the pilot subcarriers
the wireless streaming of video and consider joint optimization of AUs (as jamming pilot subcarriers is known to be very
across PHY, MAC, and APPL (hence, with the optimization effective disruptive method). RSMA is shown to outperform
of the encoding rate adaptation for video) [155]. This is an significantly conventional MA schemes.
important research direction for RSMA, especially for mobile 27) Non-Orthogonal Unicast and Multicast (NOUM):
internet (where video takes more than 70% of traffic) and NOUM refers to mixed traffic services where a multicast
emerging applications such as 360 video, autonomous driving, (genuinely intended to multiple users, i.e., not user-specific)
metaverse for 6G and beyond. message is transmitted to multiple users and additionally
one unicast (user-specific) message is transmitted per user.
24) Wireless Caching: The small storage capacity of edge Conventionally OMA is used to transmit unicast and multicast
devices combined with the high energy consumption of active services on different resources. This is suboptimal and a better
caching shorten the standby period of user devices. By posi- approach is to superpose the multicast message on top of the
tioning caching devices adjacent to the user terminal, wireless unicast messages and use one SIC at each receiver to decode
caching networks (WCN) can reduce recurrent file transfers the multicast message first and then the intended unicast
to achieve low power consumption caching. Furthermore, the message. RSMA can do better by making a more efficient
spectrum efficiency can be increased by combining RSMA use of the SIC. Indeed, by splitting the unicast message into
with WCN. Two caching policies can be examined, i.e., the common and private parts and encoding jointly the common
most popular content (MPC) and the intelligent coded caching parts and the multicast message into a common stream to be
(CC) policy. In MPC policy, multiple popular files can be decoded by all users, the SIC can be efficiently exploited
superimposed in the power domain to form a mixed file based for the dual purpose of separating multicast from unicast
on RSMA, which is broadcasted to cache devices by the base but also better manage interference between unicast streams
station [156]–[158]. Another approach of cache placement is [127], [164]–[166]. Taking the 2-user architecture of Fig. 7
the coded caching (CC), where partitions of the files are stored as an example, RSMA in NOUM is obtained by encoding
instead [159]. Multiple cache-enabled receivers can be served a multicast message W0 (genuinely intended to both users)
using both caching policies within the help of RSMA [156]– along with common parts Wc,1 and Wc,2 into a common
[158]. It is shown that the caching gains can be improved stream. User-k then decodes the common stream to retrieve
significantly, the mutual benefit can be achieved through W0 and Wc,k , before decoding its private stream. RSMA has
collaborative design of caching placements and RSMA [156]– been shown to outperform SDMA and NOMA counterparts in
[158]. In a way, RSMA for caching-aided multi-antenna BC NOUM [127], [164], [165].
expands upon, and inherits the main features of, RSMA used 28) Multigroup Multicast: This scenario considers K users
for the classical multi-antenna BC with imperfect CSIT. Addi- grouped into G < K groups and a transmitter that delivers
tionally, using RSMA, the caching device can provide services on the downlink one multicast message per group, i.e., all
to numerous users in the same time-frequency resource block. users in the same group are interested in the same multicast
The quantity of caching files, their level of popularity, and message. Conventional BC is a subset of that setting where
their network resources all affect the key indicator of wireless there is only one user per group. Such a scenario can occur
caching networks, e.g., the system hit probability. in broadcasting services, caching settings, satellite communi-
28

cations, multi-view video, virtual reality (360) video, online 32) Constructive Interference Exploitation/Symbol-Level
gaming, metaverse, etc. The challenge of multigroup multicast Precoding: RSMA is conventionally studied with Gaussian
is that the number of users is often large compared to the inputs, but in practice finite constellations need to be used.
number of transmit antennas, i.e., overloaded network, which Constructive interference (CI), also called symbol-level pre-
creates severe multigroup interference issues. RSMA, thanks coding, exploits the finite constellation such the information
to its flexibility, is able to tackle that multigroup interference symbols are used, along with the CSI, in order to exploit the
efficiently and outperform both SDMA and NOMA schemes multi-user interference to increase the useful signal received
significantly [67], [167]–[172] and cope with various other power. In other words, CI designs the transmit precoders such
multicast scenarios relevant to 6G [173]. that the resulting interference is constructive to the desired
29) Multibeam Satellite Communications: Multibeam satel- symbol, i.e., the interference signal pushes/moves the received
lite is often modeled by a multigroup multicast where each symbols away from the decision thresholds of the constellation
beam can be thought of one group. This comes from the towards the direction of the desired symbol. Interestingly,
superframe/frame-based precoding assumption of multibeam RSMA and CI techniques can be combined to further enhance
satellite communication which is that one codeword encodes the sum-rate achieved by RSMA with finite input alphabet
all users’ messages in one beam. This is used in practice to [186], [187].
increase the efficiency of the error correcting codes. Other 33) Reliability and Low Latency : There are many reli-
challenges in satellite systems include the frequency reuse ability and latency-sensitive applications, such as industrial
across multiple spot beams creating high levels of interference, automation, smart grid and intelligent transportation. In order
the per-feed power constraints, the satellite channel subject to reduce the transmission latency, shortpackets with finite
to line-of-sight and large path loss, the imperfect CSIT and blocklength codes are typically adopted. This brings a stringent
latency in CSI acquisition due to large round trip delay, the latency requirement to the physical layer. Interestingly, the ef-
high doppler for low earth orbit satellite, and the overloaded ficiency, robustness, flexibility benefits of RSMA over SDMA
settings (many terminals in each beam). RSMA has been found and NOMA in the infinite blocklength regime were also
to be quite suited to tackle all those challenges thanks to its confirmed in the finite blocklength regime for both downlink
efficiency, flexibility, and robustness [168], [174]–[177]. and uplink, offering therefore the additional reliability and the
low latency needed to enable those applications [32], [34],
30) Unmanned Aerial Vehicles (UAV)-Assisted Networks: [69]. In other words, RSMA can achieve a given performance
Due to the great mobility and flexibility, UAVs are able to requirement with a smaller blocklength than that needed by
provide services to users when they are outside of the base SDMA and NOMA. The significant performance gains of
station’s coverage area or when the user channel conditions RSMA over SDMA and NOMA were also confirmed in link-
are unfavorable. However, a UAV relies on its limited capacity level evaluations with practical codes and finite blocklengths
battery to fly, hover and communicate, which leads to a limited [33], [75].
endurance. Fortunately, RSMA can reduce the communication 34) Integrated Radar Sensing and Communications: Inte-
energy consumption of UAV. By acting as an aerial base sta- grated (radar) sensing and communications (ISAC) merges
tion, the UAV can serve multiple ground users simultaneously wireless communications and remote sensing into a single
using RSMA [178]–[181]. RSMA for UAV leads to new joint system, where both functionalities are combined via shared
cross-layer optimization problem of UAV location, transmit use of the spectrum, the hardware platform, and a joint
power allocation, bandwidth, and RSMA, accounting for CSI signal processing framework. It also enables sensing capabil-
availability and traffic. RSMA and UAV locations can also be ities of the network to help communications and inversely.
designed along with methods used for predicting the cellular The challenge is that a transmitter not only has to serve
traffic according to the analysis of previous data, therefore multiple users simultaneously but also has to satisfy radar
further saving transmit power consumption [182]. performance requirements, lwhich leads to a tradeoff between
31) Space-Air-Ground/Satellite-Terrestrial Integrated Net- communication and radar performance and calls for agile
works: In satellite-terrestrial (or in space-air-ground) inte- and versatile MA schemes. The flexibility and robustness of
grated network, the satellite sub-network shares the same RSMA becomes particularly handy in this setting as it was
RF band with the terrestrial sub-network. A higher spectrum shown that RSMA can provide a better communication-radar
efficiency and throughput is achieved via dynamic spectrum tradeoff than SDMA and NOMA schemes for a wide range
access sharing to enhance spectrum utilization. However, se- of propagation conditions and radar metrics [38], [76], [101],
vere interference in and between the sub-networks is induced [102], [188]–[191].
by an aggressive frequency reuse, which calls for the use of 35) Grant-based, Grant-Free and Semi-Grant Free Trans-
efficient multi-antenna [183] and multiple access strategies mission and Massive Random Access: Grant-based transmis-
[153], [184], [185]. In this context, RSMA has been shown to sion has been a conventional approach to access the net-
exhibit significant performance gains compared with various work when the network load is small. With the emergence
traditional transmission strategies such as SDMA and NOMA of IoT devices and mMTC, hybrid grant-based (GB) and
in various settings where only CSI is shared among sub- grant-free (GF) transmissions are needed to reduce latency.
networks (to enable coordination of precoders) or where CSI Grant-free access however leads to collisions that need to be
and data are shared (to enable cooperation as in CoMP joint managed. In such scenario, GF users meet opportunities to
transmission across all antennas of both sub-networks) [184]. share wireless resources with GB users. In [192], RSMA was
 29

used in this setting where the GF users split their messages 40) Wireless Information and Power Transfer (WIPT):
to realize distributed contentions and utilize transmit power WIPT has the similarity with ISAC that both systems need
most effectively for robust transmissions, meanwhile keeping to use the spectrum to deliver two services: either sensing and
themselves transparent to the GB user. RSMA was shown communications, or power and communications. Interestingly,
to significantly decrease outage probability and achieve full similarly to ISAC, the common stream in RSMA can be
multi-user diversity gain without restricting the GB and GF helpful to manage multi-user interference and at the same time
users’ target rates. boost the performance of the other service, i.e., sensing or
36) Network Slicing: To guarantee the performance of power. Consequently, RSMA has been found more efficient
heterogeneous services involving FeMBB, eURLLC, umMTC than SDMA and NOMA in WIPT [201], [202].
in 6G, network slicing is needed to allocate resources to 41) Vortex Wave Communications/Orbital Angular Momen-
different services. Network slicing can be done in an OMA tum: RSMA can be combined with orbital angular momentum
fashion, which means that different services are isolated and (OAM) to benefit from the flexibility and robustness of RSMA
allocated orthogonal (non-interfering) resources. However, as and the additional degrees of freedom of OAM [203].
the number of users grows, OMA-based slicing may not be the 42) Mobile Edge Computing (MEC): Equipped with pow-
optimal scheme for all scenarios, and a non-orthogonal scheme erful computing and storage capabilities, MEC can cope with
may achieve a better performance. Thanks to its message the challenges of providing superior and latency-critical com-
splitting and corresponding flexibility, RSMA has emerged puting by enabling edge users to offload their tasks for nearby
as a superior MA scheme for network slicing, outperforming processing, therefore reducing the backhaul bottlenecks, net-
OMA and NOMA in many scenarios [193], [194]. work delays, and transmission costs. Since RSMA can achieve
the full rate boundary of the MAC, while NOMA can achieve
37) Cognitive Radio: Cognitive radio may increase spectral
efficiency through secondary spectrum sharing / dynamic spec- only several separated points on the rate boundary, RSMA
can be used more efficiently than NOMA to aid MEC where
trum access, transmitting under the interference temperature.
multiple users can offload their tasks while maintaining the
Various setups can be considered where primary transmitter-
receiver links communicate simultaneously with a group of QoS of each user [204].
cognitive secondary users that form one of classical multi- 43) Mixed Criticality: Thanks to the efficiency of RSMA,
user channels such as MAC, IC, or BC [195]. In the BC, QoS is enhanced [8], [25]. Thanks to its flexibility, RSMA can
RSMA can for instance be employed at the secondary trans- also deliver QoS enhancements in applications where, due to
mitter to communicate with secondary users while limiting the the diversity of users, services and applications in 6G, mixed
interference to primary users [163], [196]. Since the primary criticality QoS levels are assigned to those users and services
link remains oblivious to the secondary system operation in [27], [31].
cognitive radio, another potential benefit of RSMA is whether
the primary system does not need to time-share the channel VI. M YTHS
with the secondary users, unlike the conventional spectral-gap Myth 1: RSMA is a special (power-domain) NOMA
filling approaches [195].
It is actually the opposite with (power-domain) NOMA
38) Optical and Visible Light Wireless Communications: being a special RSMA technique. In the same way as decoding
RSMA is commonly studied for RF communications, but it interference is a particular instance of RS (and has been known
can also be applied to other communication systems such as to be so since the 80s and the seminal works on RSMA [11],
optical and visible light, though the constraints from those [23]), NOMA is a particular instance of RSMA, as illustrated
signals and systems need to be captured in RSMA design and in Fig. 4, 8 and 13. However, one needs to check more
optimization, e.g., visible light signals have peak and average carefully at how the schemes have been built as NOMA is
optical power constraints (limited for eye safety and practical not a special case of all RSMA schemes.
illumination requirement), are non-negative and real due to In uplink, NOMA simply relies on SIC. In other words,
the intensity modulation and direct detection technique [197]– there is nothing special to NOMA as it is just an SIC receiver
[199]. Similarly to RF systems, RSMA outperforms SDMA (similarly to the SIC used in spatial multiplexing for point to
and NOMA in optical and visible light communications [197]– point MIMO). RSMA on the hand not only relies on SIC but
[199]. also on splitting of the messages at the users. As shown in
39) Multi-carrier: Frequency domain using multi-carrier Section III-B, by adjusting the split and the power allocation
transmission (e.g., OFDM/OFDMA) can be combined with to the resulting streams, uplink RSMA boils down to uplink
RSMA in the same way SDMA and OFDM/OFDMA work NOMA.
together in 4G and 5G, namely the spectrum is divided into In downlink SISO and MISO, all power-domain NOMA
subbands (made of contiguous or distributed subcarriers) and schemes are characterized by having at least one user being
multiple users are paired together on one or multiple subbands forced to fully decode the message(s) of other co-scheduled
using RSMA [162], [163], [171], [200]. Resource allocation, user(s) [7]. In the two-user, NOMA (as well as SDMA, OMA,
including user pairing per subband and power allocation to and physical-layer multicasting) is a subset of RSMA as shown
common and private streams, needs to be optimized. RSMA- in Section III-A. In the general K-user MISO case as it would
OFDMA inherits the same benefits as SDMA-OFDMA since depend on the specific RSMA scheme used. 1-layer RS is a
RSMA builds upon SDMA/MU-MIMO. superset of SDMA since by turning off (i.e., allocating no
30

10 10

Sum Rate (bit/s/Hz)

Sum Rate (bit/s/Hz)

8
8

6
6
4

4
NOMA 2 NOMA
1-layer RS 1-layer RS
2 0
5 10 15 20 25 30 5 10 15 20 25 30
SNR (dB) SNR (dB)

(a) Rth
k =0 (b) Rth
k = 0.2 bit/s/Hz
Fig. 31. Relationship between existing MA strategies and the K-user RSMA
framework.
Fig. 32. Sum-rate vs. SNR comparison in three-user SISO BC with perfect
CSIT, K = 3, σ12 = 1, σ22 = 0.3, σ32 = 0.1.

power to) the common stream, 1-layer RS boils down to MU–

LP. On the other hand, 1-layer RS is not a superset of NOMA. Other types of non-linear precoded RSMA schemes are left
1-layer RS and NOMA are particular instances/schemes of for further investigations.
the RSMA framework based on the generalized RS relying on This is completely different from NOMA. NOMA does not
multiple layers of SIC at each receiver, as discussed in Section build upon SDMA/MU-MIMO. With G groups, K-user MISO
IV and [7], [25]. NOMA can boil down to G-user SDMA by turning off the
In downlink MIMO, there is less research on RSMA. power to the weaker users in each group, but K-user MISO
Nevertheless, RSMA is shown in [7], [35] to outperform NOMA can mathematically never boil down to K-user SDMA
and be a superset of NOMA whenever at least one user [7]. The rate/DoF of K-user NOMA can therefore be worse
is forced to fully decode the multiple streams of other co- than that of K-user SDMA [7].
scheduled users. More generally, as it appeared in Table IX Myth 3: RSMA is only beneficial for multi-antenna
and related discussion, RSMA in a MIMO setting will always downlink
be a superset than NOMA because RSMA has the message RSMA is definitely beneficial in multi-antenna downlink,
splitting capability for each message (and therefore the related but RSMA is also beneficial in single-antenna downlink,
interference management capability), which does not feature single/multi-antenna uplink, in multi-cell, and in relaying.
in NOMA schemes. This implies that the optimization space In single-antenna downlink (SISO BC), the benefits of
of RSMA will be larger than that of NOMA. RSMA depends on whether the BC is degraded or non-
The relationship between SDMA, NOMA, 1-layer RS, 2- degraded. In the degraded BC, RSMA boils down to NOMA
layer (hierarchical) RS (as introduced first in [58] for FDD (since NOMA is capacity achieving for degraded BC) but
massive MIMO), and RSMA is further illustrated in Fig. 31. RSMA can be used to decrease the receiver complexity and
Myth 2: RSMA cannot outperform MU-MIMO still come close to the capacity region with a reduced number
MU-MIMO schemes can rely on linear or nonlinear precod- of SIC layers compared to NOMA [8]. We illustrate in Fig.
ing schemes [6]. For both types of precoders, we can design 32 a three-user example to compare the sum rate of NOMA
linear precoded or non-linear precoded RSMA that are always and 1-layer RS when users’ channel variances are σ12 =
a superset of MU-MIMO, and would therefore always achieve 1, σ22 = 0.3, σ32 = 0.1 and users are with and without QoS
at least the same performance as MU-MIMO. rate constraints. In the two subfigures, 1-layer RS respectively
Taking K-user 1-layer RS for simplicity (but the discus- achieves 99.84% and 97.65% rate of NOMA while only a
sion holds for other RSMA architectures), by decreasing the single layer of SIC is required at each user (instead of 2 for
amount of power allocated to the common stream, 1-layer RS NOMA). In the non-degraded BC, RSMA achieves a strictly
progressively converges to K-user SDMA/MU-MIMO and in larger rate region than NOMA (see Remark 4). Examples for
the limit where no power is allocated to the common stream, non-degraded SISO BC comprise (but are not limited to) multi-
K-user 1-layer RS swiftly boils down to K-user SDMA/MU- cell BC or BC with interference [205], BC with imperfect
MIMO. Hence, 1-layer RS really builds upon SDMA/MU- CSIT [45], RIS-assisted BC [206], IC with moderate or low
MIMO and SDMA/MU-MIMO is a subscheme of 1-layer RS, interference [11]. Intuitively the underlying feature of these
which provides a guarantee to 1-layer RS that its rate and examples is, that the channels to the receivers cannot be
DoF are always the same or better than those of SDMA/MU- ordered. For all examples above, there exists reasons, why the
MIMO. The same observation holds for non-linear precoded channels, either because there are some carriers where one
RSMA as discussed in [108], [109] for Tomlinson-Harashima channel is better than the other and vice versa, or because
precoded RSMA and in [9], [165] for dirty paper coded on some channel realizations one user receives more or less
RSMA. Note that since dirty paper coding achieves the ca- interference than the other, or because the channel state is not
pacity of MIMO Gaussian BC with perfect CSIT, applying perfectly known at the transmitter and therefore, the encoder
dirty paper coded RSMA to a perfect CSIT setting would end cannot determine the optimal pre-coding order. In practice, we
up allocating zero power to the common streams, however do not have perfect CSIT. Therefore, we will always have to
non-zero power would be allocated to common streams to operate on non-degraded SISO BCs.
boost the performance in imperfect CSIT settings [9], [165]. In uplink (MAC), RSMA achieves every point at the bound-
 31

ary of the capacity region without time sharing [23]. As always required to be split. Whether the message of a single
illustrated in Fig. 25, RSMA can also enhance the spectral user or the messages of multiple users are split depends on the
efficiency when proportional user fairness is considered [65]. objective function. For instance in Example 1, both messages
In multi-cell (IC), RSMA outperforms SC and SIC in the W1 and W2 are split, but it could happen that only W1 is split
weak interference regime. It enables an enhanced interference or only W2 is split.
management of the intra-cell and inter-cell interference as Splitting the message of a single user (as in [18]) or more
the amount of intra/inter-cell interference to be canceled at users (as in [8]) at the transmitter has no impact on the
the transmitters or decoded at the receivers can be flexibly performance if the objective is to maximize the WSR or EE
adjusted. RSMA therefore achieves improved spectral and (defined by sum rate dividing the sum transmit power) subject
energy efficiency over the conventional coordinated schemes to transmit power constraint. In such case, the major question
without RSMA [60], [141]. is whether RS is helpful or not, and how much of the total
In relaying and cooperative systems, by enabling the users (sum) information should be carried by the common message
with strong channel strength to relay and forward the common regardless of how the common message is split.
stream to the users with weaker channel strength, not only can However, if more user fairness is considered, i.e., when
cooperative RSMA improves spectral and energy efficiency the objective is to maximize the minimum rate among users
[53], [207], but also offering substantial benefits in terms of or/and subject to QoS rate constraints for each user, the choice
coverage extension [25]. of which users to split the messages at the transmitter will
Myth 4: The common stream in RSMA is a multicast influence the final performance. The best method is to leave
stream required for multiple users the possibility to split the messages of all users so as to provide
A common stream in RSMA is multicasted at the physical rooms for allocating the rate of the common stream among
layer since it is to be decoded by multiple users. However the users.
content of the common stream is not necessarily intended to
Myth 7: RSMA has to sacrifice a higher receiver
those users. This is a difference from multicasting and broad-
complexity in order to outperform (power-domain) NOMA
casting where a message is genuinely intended to multiple
Because RSMA does not enforce a given stream to be
users, and therefore decoded by multiple users. In RSMA,
fully decoded or to be fully treated as noise, but rather split
the common stream is created for interference management
one or multiple messages such that a message is partially
purpose, not because the content of the common stream is
decoded by another users, RSMA can explore a wider space of
intended to multiple users.
communication schemes. This consequently leads to relatively
It is also possible to do multicasting on top of RSMA. This
simple schemes like 1-layer RS that relies on a single SIC to
is the case where K users want to receive unicast messages
outperform multi-SIC NOMA schemes. It was for instance
W1 , . . . , WK (one for each user), but additionally a multicast
shown in [7] how the DoF, and therefore rate, of 1-layer RS
message W0 is transmitted and intended to all K users. In that
can be significantly larger than that of complicated NOMA
case, RSMA can be used to encode in a common stream the
schemes. Also RSMA builds upon SDMA/MU-MIMO and
multicast message W0 along with parts of the unicast messages
the addition of SIC layers comes with a performance enhance-
W1 , . . . , WK , as discussed in NOUM subsection V-27 [127],
ment. This contrasts with NOMA where the performance (DoF
[165].
and rate) can degrade as we increase the number of SIC, as
Recall that NOMA also has a common message/stream,
a consequence of the restrictive design philosophy of NOMA
though commonly not denoted using such terminology in
[7]. The above discussion is further illustrated in Fig. 27.
the NOMA literature. Hence, the common message is not a
message that is originally intended for all users. It is required Myth 8: With more data streams to send from the trans-
to be decoded by all users but is not necessarily intended for mitter, beamforming design and power control become
all users. very complicated in RSMA
Myth 5: As the common stream needs to be decoded by For any MA scheme, the larger the number of users and
multiple users, it causes privacy/security issues streams to serve, the higher the complexity for beamforming
Note that decoding the common stream at the physical and power control. Hence this is not an issue specific to
layer does not imply the sharing of data as encryption is RSMA, but would hold for SDMA, NOMA, etc. Nevertheless,
commonly implemented at higher layers and decryption is with RSMA, low complexity beamforming and power control
performed using user-specific codes. Same would go for other can be perofrmed. For instance, in 1-layer RS, private streams
schemes relying on SIC and interference decoding such as could be precoded using zero-forcing beamforming (ZFBF)
NOMA. There is therefore no privacy/security issues as long and have the power allocated uniformly (that would be much
as higher layer encryption is performed. However, from a compliant with the way MU-MIMO is implemented in practice
physical layer security/secrecy perspective [208], [209], the in 4G and 5G). The precoder of the common streams can be
problem is different and RSMA can be designed to maximize designed using low complexity techniques [24], [25], [58],
the secrecy rate as discussed in Subsection V-17. [66]. What remains to be designed is the power allocated to
Myth 6: The message of each user is required to be split the common stream that could depend on the network load,
into one or multiple common parts and a private part propagation conditions, and metric [24], [58], [66].
In RSMA, the message of each user could be split into Myth 9: RSMA requires high SNR to achieve perfor-
one or multiple common parts and a private part, but is not mance gain over NOMA or SDMA
32

in massive MIMO and to maintain multi-user connectivity

6
1-layer RS in mobility conditions despite the delayed CSIT. In [58],
MISO NOMA (G=3) [91], RSMA was applied to resolve the imperfect CSIT
5 MISO NOMA (G=1) problem in FDD massive MIMO and was shown to provide
SDMA
performance gain over conventional massive MIMO based on
MMF Rate (bit/s/Hz)

4
SDMA processing. In [68], [104], RSMA was investigated
in TDD massive MIMO and cell-free massive MIMO and
3
was shown to be robust to pilot contamination and to provide
significant gains whenever there is a likelihood of users being
2
allocated the same uplink pilot sequence. This is motivated
1
by mMTC where the probability of multiple users sharing the
same pilot is very high. In [100], RSMA was used to enhance
0
the robustness of massive MIMO in the presence of hardware
5 10 15 20 25 30 35 40 impairments and in particular phase noise. RSMA was shown
SNR (dB) to outperform conventional massive MIMO in the presence of
phase noise. In [119], RSMA was finally shown to be robust
Fig. 33. Max-min fairness rate vs. SNR comparison in MISO BC with
imperfect CSIT, α = 0.5, M = 5, K = 6, σk2 = 1, ∀k ∈ K. [7]. in multi-pair massive MIMO relay systems.

VII. F REQUENTLY A SKED Q UESTIONS

The gain of RSMA over SDMA and NOMA depends We here classify frequently asked questions about RSMA
on many parameters, including network load, propagation in three different categories, namely principles and benefits,
conditions, objective function and QoS constraints, and CSIT standardization and implementation, and applications and in-
quality. The more we favour fairness (e.g., MMF objective terplay with other technologies.
function, WSR, QoS constraints), the higher the gain of
RSMA over a wide range of SNR. The lowest gain of RSMA
would be experienced in massive MIMO regime with sum- A. Principles and Benefits of RSMA
rate maximization and accurate CSIT. In such scenario, RSMA Question 1: What is the design principle of RSMA?
would boil down to SDMA or allocate a very small amount In the downlink, the design principle of RSMA is much
of power to the common streams only at high SNR. We here different to SDMA and NOMA.
illustrate a MMF rate comparison result when the network is The NOMA design philosophy is based on having a stream
overloaded and the CSIT is imperfect, as per Fig. 33. The to be fully decoded by another user. For instance, in the MISO
relative rate gain of 1-layer RS over MISO NOMA (G = 3) case, NOMA forces one user to fully decode all streams in a
is 35% when SNR is 15 dB, and 21% when SNR is 10 dB, group, i.e., its intended stream and the co-scheduled streams
which is still non-negligible. Importantly, recall that this gain in the group. This leads to the strong constraint that the entire
is achieved with a single SIC layer, while NOMA (G = 1) message of one of the users is mapped onto a common stream,
and NOMA (G = 3) require 5 and 1 SIC layers, respectively, e.g., W2 mapped to sc decoded by both user-1 and user-2 in
and achieve much worse MMF rate. Table IV.
Myth 10: RSMA only works when instantaneous channel This is radically different from SDMA design philoso-
state information is available at the transmitter phy where messages are independently encoded into private
RSMA can operate on any form of CSIT, such as instan- streams and each receiver decodes its intended stream treating
taneous CSIT with high or low accuracy and various CSI any residual multi-user interference as noise (even when the
acquisition mechanisms [18], [68], [104], delayed CSIT [66], interference level is not weak enough to be treated as noise),
or statistical CSIT based on for instance second order statistics as per Table IV. Interference is never decoded at the receivers
of the channel, e.g., spatial covariance matrix [58], [91], [92]. due to the absence of common stream(s).
A detailed summary of the imperfect CSIT models that have In RSMA, a message of a given user is not forced to be
been adopted in the existing works of RSMA can be found in treated as noise or be decoded by a user. Instead we have full
[25]. flexibility on how to encode it and can evolve in the grey zone
Myth 11: The gain of RSMA in massive MIMO is in-between. Hence, similarly to SDMA, in MISO downlink,
marginal K private streams are enabled, but in contrast to SDMA, each
The gain of RSMA in massive MIMO with perfect CSIT is user can partially decode the message of another user thanks to
negligible or inexistent since the transmitter can form pencil the presence of the common streams. In contrast to NOMA, no
beams that would provide high beamforming gain and si- user is forced to fully decode the stream(s) of a co-scheduled
multaneously eliminate multi-user interference. Hence private user since all private streams are encoded independently and
streams are sufficient and RSMA boils down to conventional each receiver decodes its intended private stream treating any
SDMA-based massive MIMO. residual interference from the other private streams as noise.
As we depart from perfect CSIT and considers practical This comes with huge benefits as RSMA builds non-
imperfection in CSIT acquisition, RSMA starts providing orthogonal transmission strategies upon SDMA (and therefore
gains. In [66], RSMA was shown to provide significant gains MU-MIMO) so that the performance benefits of SDMA are
 33

guaranteed but extra performance is observed by the use of SIC Considering a downlink without QoS constraints, the rate
receivers. Indeed, a performance gain over SDMA is expected region of RSMA comes much closer to the capacity region
from a more complex receiver architecture. To achieve this, (achieved by DPC) than SDMA and NOMA when CSIT is
one should enable the versatility at the transmitter to encode perfect [8]. When CSIT becomes imperfect CSIT, linearly
messages such that parts of them can be decoded by all precoded RSMA is able to achieve a larger rate region than
users using SIC while the remaining parts are decoded by complex DPC (and SDMA and NOMA) [9]. As RSMA
their intended receivers and treated as noise by non-intended achieves the optimal DoF in both perfect and imperfect CSIT
receivers. This is uniquely achieved with RSMA by providing [19], it optimally exploits the spatial dimensions and the avail-
flexibility in the message-to-streams mapping, as demonstrated ability of CSIT. This contrasts with SDMA and NOMA that
in Tables IV,VII, and IX. Indeed, instead of keeping a rigid are suboptimal [7]. Considering an uplink, RSMA outperforms
mapping of a message into a predefined stream (as in SDMA NOMA without time sharing [23], [69].
and NOMA), RSMA allows each user to send part of a Energy efficient: Thanks to its flexibility and universality,
message in one or multiple common stream(s) and the rest in the EE of RSMA is also larger than or equal to that of existing
one of the K private streams. By adjusting the power levels of MA techniques (OMA, SDMA, NOMA) in a wide range of
the common and private streams, one can adjust the amount of user deployments [63], [127], [128].
interference that occurs on the private stream, so that its level Enhancing QoS and fairness: RSMA exhibits an even larger
can be weak enough to be treated as noise. This enables RSMA performance gain over other MA techniques whenever each
to manage multi-user interference by partially decoding the user is subject to a QoS rate constraint or whenever a higher
interference and treating the remaining interference as noise. weight is allocated to the user with a weaker channel condition
In the uplink, the design philosophy of RSMA and NOMA [7], [8], [25]. Therefore, the ability of a wireless network
is similar at the receivers since both aim at using SIC to decode architecture to partially decode interference and partially treat
all incoming streams. However, for a fixed number of users, interference as noise leads to enhanced QoS and user fairness.
the number of streams to decode at the receivers is different Reducing complexity: RSMA has the double benefit of
because of the specific features in the encoding operation at the simultaneously boosting the performance and decreasing the
transmitters between RSMA and NOMA. NOMA philosophy complexity (at the transmitter and the receiver) compared
is to map a message into a stream such that in a two-user with multi-antenna NOMA. Recall indeed that multi-antenna
system, NOMA decodes user-1 before or after user-2. On NOMA that requires user grouping, ordering, switching (be-
the other hand, RSMA splits a message (say of user-1) into tween NOMA and SDMA) at the transmit scheduler and
sub-messages, encodes them into independent streams, and multiple layers of SIC at the receivers. On the other hand,
superposes them. This split enables the receiver to decode part 1-layer RS without any user ordering, grouping or dynamic
of user-1’s message before user-2 and the other part of user-1’s switching at the transmit scheduler and with a single layer
message after user-2. of SIC at each receiver is capable of achieving significant
Question 2: What are the major benefits of RSMA? performance gain over NOMA [7]. In contrast to SDMA
Universal: RSMA is a more general multiple access frame- that requires user pairing to pair users with semi-orthogonal
work that unifies and generalizes (and consequently outper- channels, RSMA is suited to all channel conditions and it
forms) OMA, SDMA (and multi-antenna) NOMA. OMA, does not require complex user scheduling and pairing [58].
SDMA, and NOMA, (and other schemes as physical-layer Moreover, RSMA is capable of further reducing CSI feedback
multicasting) are all particular instances of RSMA [8], [24]. overhead in the presence of quantized feedback [47]. RSMA
Flexible: RSMA is flexible to cope with all user de- nevertheless incurs a higher receiver complexity than SDMA
ployments (with a diversity of channel directions, chan- due to the use of SIC.
nel strengths), network loads (underloaded and overloaded Reducing latency and improving reliability: Reducing the
regimes), and interference levels (weak, medium, strong). transmit packet size is widely known as one major approach
RSMA automatically reduces to other MA techniques ac- of achieving low-latency communications, also known as,
cording to the channel conditions, i.e., it reduces to SDMA short-packet transmission [210]. RSMA for short packet trans-
when user channels are orthogonal in the underloaded MISO mission has been investigated in [32], [34], [69] for both
BC with perfect CSIT. When the channels are aligned with downlink and uplink, and RSMA has been shown to use
certain channel strength disparities, it reduces to power-domain shorter blocklength and therefore lower latency to achieve
NOMA. For other channel conditions, RSMA takes advantage equal sum rate and MMF rate with SDMA and NOMA. It
of the common streams and more efficiently manages multi- therefore has a great potential to enhance URLLC services in
user interference by partially decoding the interference and 6G [28].
partially treating the remaining interference as noise [8], [24]. Question 3: What does the common message contain?
Robust: RSMA is robust to any CSIT inaccuracy [14], From Example 1, we note that the common message contains
[18]. This is very relevant in modern downlink multi-antenna information bits from the original unicast (user-specific) mes-
deployments. While OMA, NOMA, SDMA, all incur a DoF sages of user-1 and user-2. In that example, both messages
loss in the presence of imperfect CSIT, RSMA is DoF-optimal W1 and W2 are split to create the common message; hence
and therefore, less sensitive to CSIT inaccuracy. the common message carries some bits from user-1 and
Spectrally efficient: The spectral efficiency of RSMA is from user-2. But we could have instances where only one
always larger than or equal to that of existing MA techniques. of the users message is split, say W1 only, and in such
34

case, the common message carries some bits of user-1 only. the larger the disparity of channel gains, the larger the benefits
The common message, whether it carries bits of one user or of using RSMA schemes with multiple SIC. In other words,
multiple users, is always decoded by multiple users. given a cell size, the disparity of channel strengths among
Question 4: Why do we combine the common parts of users could be statistically obtained using the conventional
user messages into a single common message? path loss model, and the designer can then decide how many
Transmitter and receiver design is greatly simplified by SIC would be worth given the complexity that can be afforded.
combining the common parts of multiple users into a single Nevertheless, 1-layer RSMA already brings significant gain
common message, as shown with the two RSMA strategies of even for realistic channel strength disparities. For instance,
the 2-user downlink in Section III-A1. Moving to a K-user in Fig. 27, with 10dB path loss difference and additional
scenario, if each user’s message is split into two parts without Rayleigh fading, 1-layer RS with a single layer of SIC
combining the messages, the sender would have to encode the outperforms NOMA with five SIC layers. Further gains could
2K message and design 2K precoders. Each user needs more be obtained by using RSMA scheme with say two SIC, but
SIC layers and the decoding order needs to be optimized at this shows how powerful, RSMA is to efficiently make use of
the transmitter. In contrast, the 1-layer RS transmitter only the SIC architecture, and therefore reduce receiver complexity
encodes and precodes K + 1 streams, and only one layer of [7].
SIC is needed for each user, hence no worry about decoding In the uplink, NOMA is heavily dependent on time sharing
order. This greatly reduces the complexity of 1-layer RS. to achieve good performance and attain capacity. The gain
Question 5: How does the optimized rate allocation for of RSMA over NOMA is explicit whenever we cannot af-
the common stream guide the practical message split at ford time sharing among users. RSMA achieves the capacity
the transmitter and the modulation and coding scheme? without the use of time sharing, which finds applications
The reader is invited to consult [25] for detailed examples of in scenarios where communication overhead and stringent
practical message splits and modulation and coding schemes. synchronization requirements due to the coordination of the
Question 6: In what scenarios can RSMA achieve transmissions of all users is not possible [23]. This occurs
explicit performance gain over SDMA and NOMA? for instance in services requiring grant-free access, which
Let us start with the downlink. From a DoF perspective, allow collisions to reduce the access latency stemming from
RSMA achieves the same or higher DoF than SDMA and than the channel grant procedure [25], [88], [192]. Other uplink
NOMA with a lower number of SIC layers [7]. This means scenarios where RSMA outperforms NOMA is in uplink with
that in the high SNR regime, RSMA will always outperform finite blocklength [69] or in network slicing [193].
those two schemes and the gains would be larger as the The reader is also invited to consult the many applications
objective functions accounts for fairness, the system gets more and scenarios discussed in Section V where references have
overloaded, or the quality of the CSIT degrades. A DoF gain demonstrated that RSMA outperforms SDMA and NOMA.
also translates to a rate gain at finite SNR though the exact Question 7: What can we learn from information theory
gain depends on the disparity of channel strengths among users about RSMA? How can information theory guide the
and the angle between user channels. modern study of RSMA?
In the low to medium SNR regime, the gains of RSMA As re-visited in Section II, the roots of RS can be dated
over SDMA and NOMA will depend on the user channel back to [10], [11], where the coding scheme was introduced
orthogonality and the disparity of channel strength, as shown for the SISO IC in the weak interference regime. At that time,
for a 2-user scenario in Fig. 26. SDMA favors orthogonal the MAC capacity region was characterized completely by
channels, accurate CSIT and similar channel strengths among separate random coding and SIC at the receiver. The capacity
users. NOMA favours aligned channels in each group and a region of the degraded BC was also known to be achieved
large disparity of channel strengths. As we depart from those by SC and SIC. Initial results on the IC showed that capacity
extremes, or as the CSIT quality degrades, RSMA provides region can be achieved by the same techniques in the case
explicit gains over SDMA and NOMA even more when of very strong [40] and strong [211] interference. The basic
constraints on fairness, QoS, or minimum rate are imposed idea to develop a coding scheme which allows the receives
by the network [8], [24], [25]. to decode part of the interference, to bridge the two extreme
RSMA was shown to reduce to SDMA in the presence of cases treat interference as noise (TIN) and decode interference,
orthogonal channels (or close to orthogonal) and outperform led to the development of RS. A very important ingredient in
SDMA otherwise. RSMA was shown to reduce to NOMA the development of the best achievable rate region for the IC
and achieve the same rate performance whenever: 1) the SNR is time sharing [212]. Due to the remaining interference, the
is low, 2) the channels are closely aligned, 3) there is a achievable rate region is non-convex. Only time sharing be-
sufficiently large disparity of channel gains, and 4) the CSIT tween different coding and decoding strategies including TIN,
is perfect. If all four conditions are met, NOMA, RSMA, FDMA, and RS can guarantee the best achievable rate region.
and DPC schemes achieve very similar performance (if not It must be stressed that there is a gap to the outer bound of
the same performance). As we depart from those conditions, the capacity of IC with weak interference. A characterization
NOMA incurs a loss over RSMA. Same observations also hold of the capacity region within a finite number of bits is derived
in more general K-user settings [9]. in [30]. The achievable scheme in this work is based on a
When it comes to the role played by the channel gain simplified RS approach.
disparity among users, it is important to note that with RSMA, In the BC and IC, the RS approach was introduced in recent
 35

works in the context for multi-antenna settings (MISO and introduced in the standard, would address numerous issues
MIMO) with imperfect CSIT . Though the Gaussian MIMO and therefore boost the performance of all those work items.
BC with perfect CSIT is known and achieved by DPC [42], NOMA was heavily investigated in 5G but not so well
the capacity and capacity-achieving schemes of those channels received at the end. From a theoretical point of view, it is
with imperfect CSIT are still unknown, but RS is known to predictible that (power-domain) NOMA would not fly given
play a crucial role in achieving optimality in DoF [18]–[22], its deficiencies and the lack of clear gains (and even loss)
[46], generalized DoF [17], [213], [214], and constant gap over MU-MIMO [7], [221]. In contrast, RSMA does not suffer
[215], [216]. from those issues as it really builds upon SDMA/MU-MIMO.
In the MAC, the RS approach was first introduced in [23]. Hence 6G could envision a single transmission mode based on
The motivation behind this development was to apply single- RSMA as a replacement or an enhancement of the MU-MIMO
user coding without requiring synchronization among users. transmission mode used in 4G and 5G, but also play numerous
There the term RSMA was coined. Already in [23], the other roles in the entire air interface, such as enabling efficient
important cases with fading and interference were considered simultaneous unicast-multicast-broadcast transmissions, with
to bring the proposed coding and decoding scheme to practical numerous new applications in automotive driving, VR, 360
applications. video, metaverse, etc.
In multihop-multiflow communication, e.g., in a 2 × 2 × 2 Question 9: Does RSMA create more complications for
setting comprised of 2 sources, 2 relays, and 2 destinations, implementation?
the combination of RS with decode-and-forward and amplify- The 1-layer RS strategy and its benefits in terms of im-
and-forward schemes plays a crucial role in achieving the plementation and complexity over SDMA and NOMA have
fundamental limits [217]. been already discussed above. There are nevertheless other
RS is also a crucial ingredient in the theoretical and practical challenges to overcome to make RSMA practical. In most of
principles of the broadcast approach to communication over the RSMA works, its transmitter-side design assumes Gaussian
state-dependent channels and networks [218]. This is relevant inputs, and it can be tricky to fit a Gaussian-optimized RSMA-
in scenarios where the transmitters have access to only the based precoder into a real physical layer, where it often
probabilistic description of the time-varying states while re- deals with the finite blocklength and finite constellations, or
maining unaware of their instantaneous realizations [218]. with pre-standardized MCS. Interestingly, recent efforts have
Later the term multiple access was used also for scenarios been made to make the 1-layer RS strategy work with the
in which several users share a link including BC. The term state-of-the-art channel codes and modulation and significant
(power-domain) NOMA was introduced mainly for the down- throughput performance have been observed using realistic
link transmission corresponding to BC setup. It corresponds link-level simulations [33], [35], [74]–[77], though more work
to SC SIC [219] and achieves the capacity region in degraded is needed in this area.
BC [220]. The term RSMA was also applied to the BC in [8] Moving to other RSMA schemes, like the generalized
where it can provide significant achievable rate gains. RS strategy [8], can be complex to implement especially
Network information theory [12] provides the solid basis for for a large number of users. Nevertheless, as a generalized
system modeling, a taxonomy of known results with achiev- framework of RSMA, it embraces SDMA, NOMA, physical-
able encoding and decoding schemes with unique taxonomy, layer multicasting, OMA as special cases, and suggests a novel
and a toolbox of methods and schemes for the modern study of method to softly bridge existing MA techniques without using
multi-user communication systems. All currently standardized naive hard switching.
and developed transceiver schemes have their roots in the Moreover, the generalized RS is applicable to the scenar-
fundamental information theoretic results - even if the name ios with relatively small K and it achieves non-negligible
and the terminology might have changed. performance gain over existing MA techniques. Hence, the
transmitter could schedule a small number of users in each
resource block.
B. Standardization and Implementation of RSMA Another use of the generalized RS is to act as a benchmark
Question 8: What is the status of RSMA standardiza- to demonstrate the performance of other low-complexity RS
tion? Why would RSMA succeed in 6G when NOMA was strategies such as 1-layer RS and 2-layer HRS. It enables
not well received in 5G? to identify which common streams are effective or ineffec-
RSMA is still very new in 3GPP and has not been discussed tive, and consequently trim the generalized RS into a low
by standard bodies yet. The machinery required for RSMA complexity RS scheme that would rely on a subset of the
is nevertheless already partially being studied, discussed and common streams. From our experience, in many applications,
developed in 3GPP. Indeed, past 3GPP study/work items low-complexity RS strategies achieve performance reasonably
such as MU-MIMO, full-dimensional MIMO, coordinated close to that of generalized RS while their complexities are
multi-point (CoMP), multi-user superposition transmission much lower than the generalized RS and NOMA strategies.
(MUST), NOMA, network-assisted interference cancellation This demonstrates that by departing from the extremes of
and suppression (NAICS), multicast and broadcast services, treating interference as noise and fully decoding interference,
can be leveraged to design RSMA. However the key novelty one can find alternative MA strategies that are spectrally and
of RSMA, namely relying on message split, has not been energy efficient and simultaneously computationally efficient
discussed in standardization bodies. Inversely, RSMA, once (relatively low complexity and small number of SIC layers).
36

It helps us draw the conclusion that 1-layer RS is a good resource allocation for optimizing achievable rate and energy
alternative to the generalized RS in many practical scenarios. efficiency.
Another benefit of generalized RS is the ability to come It should also be noted that the RSMA framework can be
up with a set of schemes whose performance improve as expanded in the time or frequency domains to get a space-
the number of SIC layers increases. This is helpful to figure time or space-frequency RSMA framework as in [22], [47], as
out the performance gap vs complexity tradeoff between low- discussed in Section III-A7. This is particularly relevant when
complexity RS strategies and generalized RS and understand the CSIT quality changes across users and time or frequency
whether the addition of one or multiple common streams is or when we deal with asymmetric downlink or multi-cell
worth the complexity increase. This contrasts with NOMA framework where the receivers have a different number of
where a larger number of SIC layers can lead to a lower receive antennas.
performance [7]. Question 11: Can RSMA be integrated with emerging
Therefore, the generalized RS is a significant strategy in the waveform, e.g., orthogonal time frequency space (OTFS),
framework of RSMA. orthogonal delay-Doppler division multiplexing (ODDM)?
OTFS was recently proposed as a new two-dimensional
modulation scheme [226] [227], which multiplexes informa-
C. Applications and Interplay between RSMA and other Wire-
tion symbols in the delay-Doppler (DD) plane (or domain)
less Technologies
rather than the time-frequency (TF) domain as for conventional
Question 10: Can RSMA be integrated with other multi-carrier modulation or OFDM schemes in the current
MA techniques such as OFDMA, SDMA, power-domain 4G/5G cellular and WiFi networks. The DD domain symbol
NOMA, and code-domain NOMA? multiplexing enables a direct coupling of the transmitted sym-
RSMA can definitely be combined with OFDMA in the bols with the channel’s delay-Doppler spread function, which
same way as it is done with SDMA/MU-MIMO in 4G and 5G, has nice properties, such as quasi-static, compact, and sparse,
namely a group of users are paired and served using RSMA to be exploited to achieve a low channel estimation overhead
on a given resource block or subband. Though much remains and full channel diversity with low complexity receivers [228]
to be investigated in OFDMA-RSMA design, some research [229] [230] [231] [232]. OTFS has also stimulated additional
on the user grouping and power allocation optimization has research on delay-Doppler plane modulation with orthogonal
been initiated in [171], [172]. pulses, such as orthogonal delay-Doppler division multiplexing
Interplay between SDMA or power-domain NOMA and (ODDM) [233] [234], which can achieve orthogonality with
RSMA would not bring benefits, since SDMA is always respect to the channels’ delay and Doppler resolutions that
part of any RSMA scheme, namely when it comes to the are generally much smaller than the symbol duration and
transmission and reception of the private streams, and power- subcarrier spacing in conventional OFDM.
domain NOMA is part of the generalized RS architecture of ODDM, or general delay-Doppler plane multi-carrier mod-
RSMA. ulation, is a promising waveform for future wireless sys-
On the other hand, there is no effort so far on the interplay tems, particularly on doubly-selective channels. This type
between code-domain NOMA and RSMA and this is an area of schemes not only provides robust performance in high-
of interest to see whether we can further enhance RSMA mobility channels, they can also be a viable choice for the
performance by bringing the code-domain dimension. Code- future ISAC, due to that their transmit/receive pulses and
domain NOMA (CD-NOMA) employs carefully designed in- the corresponding ambiguity function satisfy the orthogonality
terleavers and/or code sequences to multiplex users. The idea property with respect to the delay-Doppler resolutions. Similar
was inspired by the traditional CDMA [222] or interleaver- to OFDMA, ODDM itself can be employed as an orthogonal
division multiple access (IDMA) [223]. Some well-known multiple access scheme.
examples of CD-NOMA include sparse code multiple access With this property, it is natural to have RSMA combined
(SCMA) [224] and non-orthogonal coded access (NOCA) with OTFS/ODDM waveform in MISO or MIMO systems.
[225]. In SCMA, each user is assigned a sparse codeword The combination of RSMA and OTFS/ODDM can provide
according to its message. To exploit the sparsity introduced high system design flexibility in terms of its resource alloca-
by the codeword, the receiver adopts message passing al- tion and optimization, in the mean time having a potential to
gorithm for multi-user detection. In contrast, NOCA assigns offering high spectral efficiency, signal localisation, integrated
each user a dense spreading signature to fully utilize the sensing and communication capabilities.
available time-frequency resources. The receiver exploits the Question 12: Can RSMA use discrete signaling without
low cross-correlation properties of spreading sequences for SIC?
interference mitigation/suppression. Nevertheless, it is possible To analyze the performance of RSMA, such as achievable
to incorporate the design principle of CD-NOMA into RSMA. rate and energy efficiency, Gaussian signaling is often as-
Notice that the split common and private messages in RSMA sumed. As we know, Gaussian signaling is the optimal signal-
can be seen as virtual users. Each of the virtual user can ing in many channels, e.g., point-to-point Gaussian channels,
be assigned a dedicated sequence for enabling code-domain Gaussian multiple access channels, and Gaussian broadcast
multiplexing. Some interesting research directions on code- channels. Hence, assuming Gaussian signaling becomes nat-
domain RSMA can include but not limited to: design of ural in RSMA. Under this assumption, RSMA can handle
sparse/dense codewords, detection and decoding architectures, interference with different strengths effectively.
 37

In practical communication systems, discrete signaling, i.e., Different from the centralized ML method, FL uploads
coded modulation, is used. In addition, for some applications trained model parameters rather than raw data [240]. However,
with stringent requirements on decoding complexity and la- when training involves wireless edge devices at the edge
tency, e.g., downlink URLLC services, TIN is more favorable network, communications could become a significant problem.
than interference decoding. If Gaussian signaling is assumed, In conventional FL, TDMA is typically used for uplink trans-
low-complexity TIN is only optimal when the interference is mission. However, the central cloud has to wait until it receives
very weak, e.g., see Table III. Indeed, despite being the optimal information from all the user. RSMA enables multiple users to
input distribution for many channels, Gaussian signaling is upload information by sharing the uplink channel at the same
also the worst noise. On the other hand, discrete signaling can time. Thus, by incorporating RSMA in the FL framework, it
behave differently from Gaussian signaling when being treated is expected that the aggregation latency can be reduced while
as noise. When the interference strength is not very weak, it maintaining model training quality. For example, multiple IoT
is possible for discrete signaling with TIN to achieve a strictly devices in fog radio access networks can cooperate to perform
larger achievable rate region than that for Gaussian signaling a FL task by repeatedly uploading locally updated models to
with TIN. a cloud server. To overcome the performance limitations due
Discrete signaling with TIN, i.e., without SIC, has been to finite capacity front-haul links, a rate-splitting transmission
investigated in recent works, e.g., [235]–[238]. In [235], a scheme at IoT devices can be used [241]. With flexible hybrid
lattice partition based NOMA scheme was proposed for the edge and cloud decoding strategy achieved by RSMA, we can
single antenna K-user Gaussian broadcast channel. It was reduce the completion time of FL while maintaining a specific
rigorously proved that the scheme based on discrete signaling target global accuracy.
and TIN is capable of achieving the whole capacity region to Not only should the PHY layer benefit from the integration
within a constant gap independent of the number of users and of RSMA and ML, but the focus should also be on cross-layer
channel parameters. The same results hold true for the afore- design such as aiding network orchestration, to truly assist
mentioned channel model with only statistical CSI available at wireless communication for intelligent 6G [26]. For instance,
the transmitter side [236]. By further exploiting the algebraic ML has been used for UAV deployment in combination with
properties of lattices, the full diversity of the broadcast channel RSMA transmission [182]. RSMA was shown to require less
with block fading and close-to-perfect SIC error performance power compared to other MA schemes.
can be attained for each user with TIN decoding [237]. Finally, Question 14: How would RSMA work in Millimeter-
in [238], it was shown in the first time that the capacity wave and Terahertz networks?
region of the Gaussian (asymmetric) interference channel can Early research on millimeter-wave communication focused
be achieved to within a constant gap by discrete signaling and on the strong directionality benefits that could be attained
TIN. using beamforming. The general idea was to communicate
To unleash the full potential of RSMA in practical commu- with each user using a narrow spatial beam. It was thought
nication systems, it is important to exploit the properties of that the small beamwidths would in turn generate minimal
discrete inputs with low-complexity TIN in RSMA. In light interference between users, which would allow for extremely
of the above works [235]–[238], discrete signaling with TIN efficient SDMA designs.
should also benefit RSMA to achieve a better tradeoff between The results of millimeter-wave 5G deployments have begun
performance and complexity. to change this thinking. The millimeter-wave benefits to 5G
Question 13: How can machine/federated learning help NR has been relatively disappointing, with few devices being
RSMA? What is the role of machine/federated learning in scheduled on the large, under-utilized frequency bands. It is
RSMA design? thought that this problem is a result of the beam blockage
Machine learning (ML) and federated learning (FL) can be issues, which appear to be a much bigger problem in a large-
used on multiple fronts. scale deployment than was initially thought.
ML can be used at the receiver of RSMA. In [77], a To make millimeter-wave and higher frequencies practical
model-based deep learning (MBDL) method was used to for future broadband wireless access, it is likely that many
propose new and practical RSMA receiver designs exploiting of the tight spatial directivity design philosophies will need
the conventional SIC receiver and the robustness and model to change. Beamwidths may need to increase, and multi-beam
agnosticism of deep learning techniques. Thanks to its ability communication may become a necessity. This will likely result
to generate on demand non-linear symbol detection boundaries in interference becoming the dominant limiting factor as it is
in a pure data-driven manner, the MBDL receiver was shown in sub-6 GHz frequencies.
to significantly outperform conventional SIC receiver with Even in scenarios where blockage is unlikely, the ever-
imperfect CSIR. increasing number of antennas will likely not lead to
ML can be used at the transmitter of RSMA to optimize beamwidths predicted by theory. Ideally, the number of feed-
the resource allocation, power allocation, task offloading (as back bits should scale linearly with the number of users [95],
in MEC), and beamforming design as an alternative to con- [242]–[244]. From a standardization perspective, this is almost
ventional convex optimization methods. In [239], the power impossible to maintain in the long-term. This will result in
allocation for each transmit stream was designed using a deep a broadening of beams and a mismatch between the beam-
reinforcement learning algorithm. It was shown that RSMA formers/precoders used and the CSI. This effect will cause
achieves a significant performance gain over SDMA. millimeter-wave and terahertz systems to have interference
38

challenges. [5] A. Goldsmith, Wireless Communications. Cambridge University Press,

There has already been some initial work on the applica- 2005.
[6] B. Clerckx and C. Oestges, MIMO wireless networks: Channels,
tion of RSMA at millimeter-wave frequency [91]. This work techniques and standards for multi-antenna, multi-user and multi-cell
focused on the CSI feedback mismatch issue and showed a systems. Academic Press, 2013.
number of benefits that could simplify implementations in [7] B. Clerckx, Y. Mao, R. Schober, E. A. Jorswieck, D. J. Love, J. Yuan,
L. Hanzo, G. Y. Li, E. G. Larsson, and G. Caire, “Is NOMA efficient
practice. Recently, RSMA has been shown to expand the in multi-antenna networks? A critical look at next generation multiple
coverage of terahertz systems [107]. access techniques,” IEEE Open J. Commun. Soc., vol. 2, pp. 1310–
1343, June 2021.
[8] Y. Mao, B. Clerckx, and V. O. K. Li, “Rate-splitting multiple access
VIII. C ONCLUSIONS for downlink communication systems: Bridging, generalizing, and
outperforming SDMA and NOMA,” EURASIP J. Wireless Commun.
This paper has provided a tutorial on RSMA. To demon- Netw., vol. 2018, no. 1, p. 133, May 2018.
strate how powerful RSMA is, the tutorial has departed from [9] Y. Mao and B. Clerckx, “Beyond dirty paper coding for multi-antenna
the oversimplisitc OMA vs NOMA discussion held in 5G and broadcast channel with partial CSIT: A rate-splitting approach,” IEEE
Trans. Commun., vol. 68, no. 11, pp. 6775–6791, Nov. 2020.
has re-centered the discussion and design of multiple access [10] A. Carleial, “Interference channels,” IEEE Trans. Inf. Theory, vol. 24,
techniques around the key role of interference management. no. 1, pp. 60–70, Jan. 1978.
Building upon first principles, the tutorial has demonstrated [11] T. Han and K. Kobayashi, “A new achievable rate region for the
interference channel,” IEEE Trans. Inf. Theory, vol. 27, no. 1, pp. 49–
how existing multiple access techniques are fundamentally 60, Jan. 1981.
limited by their inherent interference management strategy, [12] A. El Gamal and Y.-H. Kim, Network information theory. Cambridge
namely orthogonalization in OMA, treat interference as noise university press, 2011.
[13] A. G. Davoodi and S. A. Jafar, “Aligned image sets under channel
in SDMA, and decode interference in NOMA. In contrast uncertainty: Settling conjectures on the collapse of degrees of freedom
RSMA schemes build upon the RS principle which enables to under finite precision CSIT,” IEEE Trans. Inf. Theory, vol. 62, no. 10,
partially decode interference and partially treat interference as pp. 5603–5618, Oct. 2016.
[14] B. Clerckx, H. Joudeh, C. Hao, M. Dai, and B. Rassouli, “Rate splitting
noise. Consequently RSMA has been shown to provide unique for MIMO wireless networks: A promising PHY-layer strategy for LTE
benefits, including enhanced spectral, energy and computation evolution,” IEEE Commun. Mag., vol. 54, no. 5, pp. 98–105, May 2016.
efficiency; universality by unifying and generalizing OMA, [15] M. B. Shenouda and T. N. Davidson, “On the design of linear
transceivers for multiuser systems with channel uncertainty,” IEEE J.
SDMA, NOMA, physical-layer multicasting; flexibility by Sel. Areas Commun., vol. 26, no. 6, pp. 1015–1024, 2008.
coping with any interference levels, network loads, services, [16] N. Vucic, H. Boche, and S. Shi, “Robust transceiver optimization in
traffic, user deployments; robustness to inaccurate channel downlink multiuser mimo systems with channel uncertainty,” in Proc.
IEEE Int. Conf. Commun. (ICC), 2008, pp. 3516–3520.
state information (CSI) and resilience to mixed-critical quality [17] A. Gholami Davoodi, B. Yuan, and S. A. Jafar, “GDoF region of the
of service; reliability under short channel codes and low MISO BC: Bridging the gap between finite precision and perfect CSIT,”
latency. IEEE Trans. Inf. Theory, vol. 64, no. 11, pp. 7208–7217, 2018.
[18] H. Joudeh and B. Clerckx, “Sum-rate maximization for linearly pre-
Future systems will see a growing demand for spectrum coded downlink multiuser MISO systems with partial CSIT: A rate-
intensive applications and for integrated wireless systems. splitting approach,” IEEE Trans. Commun., vol. 64, no. 11, pp. 4847–
Moreover they will have to face growing concerns for con- 4861, Nov. 2016.
[19] E. Piovano and B. Clerckx, “Optimal DoF region of the K-user MISO
gested and contested electromagnetic environments (in both BC with partial CSIT,” IEEE Commun. Lett., vol. 21, no. 11, pp. 2368–
civilian and defense networks). These will push network de- 2371, Nov. 2017.
signers to adopt sophisticated interference management, multi- [20] H. Joudeh and B. Clerckx, “DoF region of the MISO BC with partial
CSIT: Proof by inductive Fourier-Motzkin elimination,” in Proc. IEEE
user, and multiple access techniques. Thanks to its deep root Int. Workshop Signal Process. Adv. Wireless Commun. (SPAWC), 2019.
in information theory, numerous benefits and applications, [21] A. Gholami Davoodi and S. Jafar, “Degrees of freedom region of
and its superiority over previous generation multiple access the (m, n1 , n2 ) MIMO broadcast channel with partial CSIT: An
application of sum-set inequalities based on aligned image sets,” IEEE
techniques (OMA, SDMA, NOMA), RSMA will play a grow- Trans. Inf. Theory, vol. 66, no. 10, pp. 6256–6279, Oct. 2020.
ing and underpinning role in next generation communication [22] C. Hao, B. Rassouli, and B. Clerckx, “Achievable DoF regions of
systems. It is hoped that the RSMA techniques presented in MIMO networks with imperfect CSIT,” IEEE Trans. Inf. Theory,
vol. 63, no. 10, pp. 6587–6606, Oct. 2017.
this article will help inspiring future research in this exciting [23] B. Rimoldi and R. Urbanke, “A rate-splitting approach to the Gaussian
new area and pave the way for designing and implementing multiple-access channel,” IEEE Trans. Inf. Theory, vol. 42, no. 2, pp.
RSMA in 6G and beyond. 364–375, Mar. 1996.
[24] B. Clerckx, Y. Mao, R. Schober, and H. V. Poor, “Rate-splitting
unifying SDMA, OMA, NOMA, and multicasting in MISO broadcast
R EFERENCES channel: A simple two-user rate analysis,” IEEE Wireless Commun.
Lett., vol. 9, no. 3, pp. 349–353, Mar. 2020.
[1] Y. Saito, A. Benjebbour, Y. Kishiyama, and T. Nakamura, “System- [25] Y. Mao, O. Dizdar, B. Clerckx, R. Schober, P. Popovski, and H. V.
level performance evaluation of downlink non-orthogonal multiple Poor, “Rate-splitting multiple access: Fundamentals, survey, and future
access (NOMA),” in Proc. IEEE Annu. Symp. Pers. Indoor Mobile research trends,” IEEE Commun. Surv. Tutor., vol. 24, no. 4, pp. 2073–
Radio Commun. (PIMRC), 2013. 2126, 2022.
[2] L. Dai, B. Wang, Y. Yuan, S. Han, I. Chih-lin, and Z. Wang, “Non- [26] A. Mishra, Y. Mao, O. Dizdar, and B. Clerckx, “Rate-splitting multiple
orthogonal multiple access for 5G: Solutions, challenges, opportunities, access for 6G—part I: Principles, applications and future works,” IEEE
and future research trends,” IEEE Commun. Mag., vol. 53, no. 9, pp. Commun. Lett., vol. 26, no. 10, pp. 2232–2236, Oct. 2022.
74–81, Sept. 2015. [27] R.-J. Reifert, S. Roth, A. A. Ahmad, and A. Sezgin, “Comeback kid:
[3] Y. Liu, Z. Qin, M. Elkashlan, Z. Ding, A. Nallanathan, and L. Hanzo, Resilience for mixed-critical wireless network resource management,”
“Nonorthogonal multiple access for 5G and beyond,” Proc. IEEE, vol. arXiv preprint arXiv:2204.11878, 2022.
105, no. 12, pp. 2347–2381, Dec. 2017. [28] O. Dizdar, Y. Mao, W. Han, and B. Clerckx, “Rate-splitting multiple
[4] D. Tse and P. Viswanath, Fundamentals of Wireless Communication. access: A new frontier for the PHY layer of 6G,” in Proc. IEEE 92nd
Cambridge University Press, 2005. Veh. Technol. Conf. (VTC Fall), 2020.
 39

[29] O. Dizdar, Y. Mao, Y. Xu, P. Zhu, and B. Clerckx, “Rate-splitting channel with user relaying,” IEEE Trans. Wireless Commun., vol. 19,
multiple access for enhanced URLLC and eMBB in 6G,” in Proc. no. 10, pp. 6362–6376, Oct. 2020.
IEEE Int. Symp. Wireless Commun. Syst. (ISWCS), 2021. [54] R. Tandon, S. A. Jafar, S. Shamai, and H. V. Poor, “On the synergistic
[30] R. H. Etkin, D. N. C. Tse, and H. Wang, “Gaussian interference channel benefits of alternating CSIT for the MISO broadcast channel,” IEEE
capacity to within one bit,” IEEE Trans. Inf. Theory, vol. 54, no. 12, Trans. Inf. Theory, vol. 59, no. 7, pp. 4106–4128, 2013.
pp. 5534–5562, Dec. 2008. [55] J. Chen and P. Elia, “Optimal DoF region of the two-user MISO-BC
[31] R.-J. Reifert, S. Roth, A. A. Ahmad, and A. Sezgin, “Energy efficiency with general alternating CSIT,” in Proc. 47th Asilomar Conf. Signals,
in rate-splitting multiple access with mixed criticality,” in Proc. IEEE Syst. Comput., 2013, pp. 1860–1864.
Int. Conf. Commun. (ICC) Workshop, 2022, pp. 681–686. [56] H. Joudeh and B. Clerckx, “On the separability of parallel MISO
[32] Y. Xu, Y. Mao, O. Dizdar, and B. Clerckx, “Rate-splitting multiple broadcast channels under partial CSIT: A degrees of freedom region
access with finite blocklength for short-packet and low-latency down- perspective,” IEEE Trans. Inf. Theory, vol. 66, no. 7, pp. 4513–4529,
link communications,” IEEE Trans. Veh. Technol., vol. 71, no. 11, pp. July 2020.
12 333–12 337, 2022. [57] C. Mosquera, N. Noels, T. Ramírez, M. Caus, and A. Pastore, “Space-
[33] O. Dizdar, Y. Mao, W. Han, and B. Clerckx, “Rate-splitting multiple time rate splitting for the MISO BC with magnitude CSIT,” IEEE
access for downlink multi-antenna communications: Physical layer Trans. Commun., vol. 69, no. 7, pp. 4417–4432, July 2021.
design and link-level simulations,” in Proc. IEEE Annu. Symp. Pers. [58] M. Dai, B. Clerckx, D. Gesbert, and G. Caire, “A rate splitting strategy
Indoor Mobile Radio Commun. (PIMRC), 2020. for massive MIMO with imperfect CSIT,” IEEE Trans. Wireless
[34] Y. Xu, Y. Mao, O. Dizdar, and B. Clerckx, “Max-min fairness of Commun., vol. 15, no. 7, pp. 4611–4624, July 2016.
rate-splitting multiple access with finite blocklength communications,” [59] C. Hao and B. Clerckx, “MISO networks with imperfect CSIT: A
IEEE Trans. Veh. Technol., pp. 1–6, 2022. topological rate-splitting approach,” IEEE Trans. Commun., vol. 65,
[35] A. Mishra, Y. Mao, O. Dizdar, and B. Clerckx, “Rate-splitting multiple no. 5, pp. 2164–2179, May 2017.
access for downlink multiuser MIMO: Precoder optimization and PHY- [60] N. Ha, W. Shin, M. Vaezi, and H. V. Poor, “Coordinated rate splitting
layer design,” IEEE Trans. Commun., vol. 70, no. 2, pp. 874–890, 2022. multiple access for multi-cell downlink networks,” in Proc. 54th
[36] A. Krishnamoorthy, Z. Ding, and R. Schober, “Precoder design and Asilomar Conf. Signals, Syst. Comput., 2020.
statistical power allocation for MIMO-NOMA via user-assisted simul- [61] Y. Mao, B. Clerckx, and V. O. K. Li, “Rate-splitting multiple access for
taneous diagonalization,” IEEE Trans. Commun., vol. 69, no. 2, pp. coordinated multi-point joint transmission,” in Proc. IEEE Int. Conf.
929–945, Feb. 2021. Commun. (ICC) Workshop, 2019.
[37] A. Krishnamoorthy and R. Schober, “Uplink and downlink MIMO-
[62] A. A. Ahmad, Y. Mao, A. Sezgin, and B. Clerckx, “Rate splitting
NOMA with simultaneous triangularization,” IEEE Trans. Wireless
multiple access in C-RAN: A scalable and robust design,” IEEE Trans.
Commun., vol. 20, no. 6, pp. 3381–3396, June 2021. Commun., pp. 1–1, Sept. 2021.
[38] L. Yin, Y. Mao, O. Dizdar, and B. Clerckx, “Rate-splitting multiple
[63] Y. Mao, B. Clerckx, and V. O. K. Li, “Energy efficiency of rate-splitting
access for 6G—part II: Interplay with integrated sensing and commu-
multiple access, and performance benefits over SDMA and NOMA,”
nications,” IEEE Commun. Lett., vol. 26, no. 10, pp. 2237–2241, Oct.
in Proc. IEEE Int. Symp. Wireless Commun. Syst. (ISWCS), 2018.
2022.
[64] B. Matthiesen, Y. Mao, P. Popovski, and B. Clerckx, “Globally optimal
[39] H. Li, Y. Mao, O. Dizdar, and B. Clerckx, “Rate-splitting multiple
beamforming for rate splitting multiple access,” in Proc. IEEE Int.
access for 6G—part III: Interplay with reconfigurable intelligent sur-
Conf. Acoust. Speech Signal Process. (ICASSP), 2021.
faces,” IEEE Commun. Lett., vol. 26, no. 10, pp. 2242–2246, Oct. 2022.
[40] A. Carleial, “A case where interference does not reduce capacity [65] Z. Yang, M. Chen, W. Saad, W. Xu, and M. Shikh-Bahaei, “Sum-
(corresp.),” IEEE Trans. Inf. Theory, vol. 21, no. 5, pp. 569–570, 1975. rate maximization of uplink rate splitting multiple access (RSMA)
communication,” IEEE Trans. Mob. Comput., vol. 21, no. 7, pp. 2596–
[41] M. Costa, “Writing on dirty paper (corresp.),” IEEE Trans. Inf. Theory,
2609, July 2022.
vol. 29, no. 3, pp. 439–441, May 1983.
[42] H. Weingarten, Y. Steinberg, and S. S. Shamai, “The capacity region of [66] O. Dizdar, Y. Mao, and B. Clerckx, “Rate-splitting multiple access to
the Gaussian multiple-input multiple-output broadcast channel,” IEEE mitigate the curse of mobility in (massive) MIMO networks,” IEEE
Trans. Inf. Theory, vol. 52, no. 9, pp. 3936–3964, Sept. 2006. Trans. Commun., vol. 69, no. 10, pp. 6765–6780, Oct. 2021.
[43] G. Caire and S. Shamai, “On the achievable throughput of a multi- [67] H. Joudeh and B. Clerckx, “Rate-splitting for max-min fair multigroup
antenna Gaussian broadcast channel,” IEEE Trans. Inf. Theory, vol. 49, multicast beamforming in overloaded systems,” IEEE Trans. Wireless
no. 7, pp. 1691–1706, July 2003. Commun., vol. 16, no. 11, pp. 7276–7289, Nov. 2017.
[44] D. N. C. Tse and R. D. Yates, “Fading broadcast channels with state [68] A. Mishra, Y. Mao, C. K. Thomas, L. Sanguinetti, and B. Clerckx,
information at the receivers,” IEEE Trans. Inf. Theory, vol. 58, no. 6, “Mitigating intra-cell pilot contamination in massive MIMO: A rate
pp. 3453–3471, June 2012. splitting approach,” IEEE Trans. Wireless Commun., 2022.
[45] P.-H. Lin, E. A. Jorswieck, R. F. Schaefer, M. Mittelbach, and C. R. [69] J. Xu, O. Dizdar, and B. Clerckx, “Rate-splitting multiple access for
Janda, “New capacity results for fading Gaussian multiuser channels short-packet uplink communications: A finite blocklength analysis,”
with statistical CSIT,” IEEE Trans. Commun., vol. 68, no. 11, pp. 6761– IEEE Commun. Lett., pp. 1–1, 2022.
6774, Nov. 2020. [70] Q. Zhu, Z. Qian, B. Clerckx, and X. Wang, “Rate-splitting multiple
[46] S. Yang, M. Kobayashi, D. Gesbert, and X. Yi, “Degrees of freedom access in multi-cell dense networks: A stochastic geometry approach,”
of time correlated MISO broadcast channel with delayed CSIT,” IEEE arXiv preprint arXiv:2207.11430, 2022.
Trans. Inf. Theory, vol. 59, no. 1, pp. 315–328, Jan. 2013. [71] E. Demarchou, C. Psomas, and I. Krikidis, “On the sum rate of
[47] C. Hao, Y. Wu, and B. Clerckx, “Rate analysis of two-receiver MISO miso rate splitting with spatial randomness,” in Proc. IEEE Int. Symp.
broadcast channel with finite rate feedback: A rate-splitting approach,” Wireless Commun. Syst. (ISWCS), 2021.
IEEE Trans. Commun., vol. 63, no. 9, pp. 3232–3246, Sept. 2015. [72] ——, “Channel statistics-based rate splitting with spatial randomness,”
[48] H. Joudeh and B. Clerckx, “Robust transmission in downlink mul- in Proc. IEEE Int. Conf. Commun. (ICC) Workshop, 2020.
tiuser MISO systems: A rate-splitting approach,” IEEE Trans. Signal [73] S. A. Tegos, P. D. Diamantoulakis, and G. K. Karagiannidis, “On the
Process., vol. 64, no. 23, pp. 6227–6242, Dec. 2016. performance of uplink rate-splitting multiple access,” IEEE Commun.
[49] Y. Liang and V. V. Veeravalli, “Cooperative relay broadcast channels,” Lett., vol. 26, no. 3, pp. 523–527, Mar. 2022.
IEEE Trans. Inf. Theory, vol. 53, no. 3, pp. 900–928, Mar. 2007. [74] H. Chen, D. Mi, Z. Chu, P. Xiao, Y. Xu, and D. He, “Link-level
[50] Y. Liang and G. Kramer, “Rate regions for relay broadcast channels,” performance of rate-splitting based downlink multiuser MISO systems,”
IEEE Trans. Inf. Theory, vol. 53, no. 10, pp. 3517–3535, Oct. 2007. in Proc. IEEE Annu. Symp. Pers. Indoor Mobile Radio Commun.
[51] R. Dabora and S. D. Servetto, “Broadcast channels with cooperating (PIMRC), 2020.
decoders,” IEEE Trans. Inf. Theory, vol. 52, no. 12, pp. 5438–5454, [75] L. Yin, O. Dizdar, and B. Clerckx, “Rate-splitting multiple access for
Dec. 2006. multigroup multicast cellular and satellite communications: PHY layer
[52] J. Zhang, B. Clerckx, J. Ge, and Y. Mao, “Cooperative rate splitting for design and link-level simulations,” in Proc. IEEE Int. Conf. Commun.
MISO broadcast channel with user relaying, and performance benefits (ICC) Workshop, 2021.
over cooperative NOMA,” IEEE Signal Process. Lett., vol. 26, no. 11, [76] R. Cerna-Loli, O. Dizdar, and B. Clerckx, “Rate-splitting multiple
pp. 1678–1682, Nov. 2019. access for multi-antenna joint radar and communications with partial
[53] Y. Mao, B. Clerckx, J. Zhang, V. O. K. Li, and M. A. Arafah, “Max- CSIT: Precoder optimization and link-level simulations,” arXiv preprint
min fairness of K-user cooperative rate-splitting in MISO broadcast arXiv:2201.10621, 2021.
40

[77] R. C. Loli, O. Dizdar, B. Clerckx, and C. Ling, “Model-based [101] O. Dizdar, A. Kaushik, B. Clerckx, and C. Masouros, “Rate-splitting
deep learning receiver design for rate-splitting multiple access,” arXiv multiple access for joint radar-communications with low-resolution
preprint arXiv:2205.00849, 2022. DACs,” in Proc. IEEE Int. Conf. Commun. (ICC) Workshop, 2021.
[78] M. Medra and T. N. Davidson, “Robust downlink transmission: An [102] ——, “Energy efficient dual-functional radar-communication: Rate-
offset-based single-rate-splitting approach,” in Proc. IEEE Int. Work- splitting multiple access, low-resolution DACs, and RF chain selec-
shop Signal Process. Adv. Wireless Commun. (SPAWC), 2018. tion,” IEEE Open J. Commun. Soc., vol. 3, pp. 986–1006, 2022.
[79] B. Matthiesen, Y. Mao, A. Dekorsy, P. Popovski, and B. Clerckx, [103] S. Park, J. Choi, J. Park, W. Shin, and B. Clerckx, “Rate-splitting
“Globally optimal spectrum- and energy-efficient beamforming for rate multiple access for quantized multiuser MIMO communications,” arXiv
splitting multiple access,” IEEE Trans. Signal Process., vol. 70, pp. preprint arXiv:2208.00643, 2022.
5025–5040, Oct. 2022. [104] A. Mishra, Y. Mao, L. Sanguinetti, and B. Clerckx, “Rate-splitting
[80] Z. Li, C. Ye, Y. Cui, S. Yang, and S. Shamai, “Rate splitting for multi- assisted massive machine-type communications in cell-free massive
antenna downlink: Precoder design and practical implementation,” mimo,” IEEE Commun. Lett., vol. 26, no. 6, pp. 1358–1362, June 2022.
IEEE J. Sel. Areas Commun., vol. 38, no. 8, pp. 1910–1924, Aug. [105] O. Kolawole, A. Panazafeironoulos, and T. Ratnarajah, “A rate-splitting
2020. strategy for multi-user millimeter-wave systems with imperfect CSI,”
[81] J. Park, J. Choi, N. Lee, W. Shin, and H. V. Poor, “Sum spectral in Proc. IEEE Int. Workshop Signal Process. Adv. Wireless Commun.
efficiency optimization for rate splitting in downlink MU-MISO: A (SPAWC), 2018.
generalized power iteration approach,” in Proc. IEEE Wireless Com- [106] Z. Li, S. Yang, and T. Clessienne, “A general rate splitting scheme
mun. Netw. Conf. (WCNC), 2021. for hybrid precoding in mmwave systems,” in Proc. IEEE Int. Conf.
[82] A. R. Flores, R. C. De Lamare, and B. Clerckx, “Linear precoding Commun. (ICC), 2019.
and stream combining for rate splitting in multiuser MIMO systems,” [107] H. Cho, B. Ko, B. Clerckx, and J. Choi, “Coverage increase at thz
IEEE Commun. Lett., vol. 24, no. 4, pp. 890–894, Apr. 2020. frequencies: A cooperative rate-splitting approach,” arXiv preprint
[83] Z. Li, S. Yang, and S. Shamai, “On linearly precoded rate splitting for arXiv:2208.04541, 2022.
Gaussian MIMO broadcast channels,” IEEE Trans. Inf. Theory, vol. 67, [108] A. R. Flores, B. Clerckx, and R. C. de Lamare, “Tomlinson-harashima
no. 7, pp. 4693–4709, July 2021. precoded rate-splitting for multiuser multiple-antenna systems,” in
[84] A. Krishnamoorthy and R. Schober, “Successive null-space precoder Proc. IEEE Int. Symp. Wireless Commun. Syst. (ISWCS), 2018.
design for downlink MU-MIMO with rate splitting and single-stage [109] A. R. Flores, R. C. De Lamare, and B. Clerckx, “Tomlinson-Harashima
SIC,” in Proc. IEEE Int. Conf. Commun. (ICC) Workshop, 2021. precoded rate-splitting with stream combiners for MU-MIMO sys-
[85] H. D. Tuan, A. A. Nasir, M.-N. Nguyen, and M. Masood, tems,” IEEE Trans. Commun., vol. 69, no. 6, pp. 3833–3845, June
“Han–kobayashi signaling in MIMO broadcasting,” IEEE Commun. 2021.
Lett., vol. 23, no. 5, pp. 855–858, May 2019.
[110] A. R. Flores, R. C. de Lamare, and B. Clerckx, “Multi-branch
[86] C. Kaulich, M. Joham, and W. Utschick, “Efficient rate splitting method tomlinson-harashima precoding for rate splitting based systems with
using successive beamforming techniques,” in Proc. IEEE Int. Symp. multiple antennas,” in Proc. IEEE Int. Conf. Acoust. Speech Signal
Pers. Indoor Mobile Radio Commun. (PIMRC), 2020, pp. 1–6. Process. (ICASSP), 2021, pp. 4885–4889.
[87] A. Grant, B. Rimoldi, R. Urbanke, and P. Whiting, “Rate-splitting
[111] J. Zhang and J. Ge, “Cooperative rate-splitting for downlink multiuser
multiple access for discrete memoryless channels,” IEEE Trans. Inf.
MISO systems with partial CSIT,” IEEE Trans. Commun., vol. 69,
Theory, vol. 47, no. 3, pp. 873–890, Mar. 2001.
no. 11, pp. 7504–7519, 2021.
[88] H. Liu, T. A. Tsiftsis, K. J. Kim, K. S. Kwak, and H. V. Poor,
[112] B. Matthiesen and E. A. Jorswieck, “Optimal resource allocation for
“Rate splitting for uplink NOMA with enhanced fairness and outage
non-regenerative multiway relaying with rate splitting,” in Proc. IEEE
performance,” IEEE Trans. Wireless Commun., vol. 19, no. 7, pp. 4657–
Int. Workshop Signal Process. Adv. Wireless Commun. (SPAWC), 2018.
4670, July 2020.
[89] J. Cao and E. M. Yeh, “Asymptotically optimal multiple-access com- [113] A. Salern and L. Musavian, “Rate splitting in multi-pair energy har-
munication via distributed rate splitting,” IEEE Trans. Inf. Theory, vesting relaying systems,” in Proc. IEEE Int. Symp. Wireless Commun.
vol. 53, no. 1, pp. 304–319, Jan. 2007. Syst. (ISWCS), 2018.
[90] Y. Mao and B. Clerckx, Multiple Access Techniques. Cham: Springer [114] B. Zheng, X. Wang, M. Wen, and F. Chen, “NOMA-based multi-pair
International Publishing, 2021, pp. 63–100. two-way relay networks with rate splitting and group decoding,” IEEE
[91] M. Dai and B. Clerckx, “Multiuser millimeter wave beamforming J. Sel. Areas Commun., vol. 35, no. 10, pp. 2328–2341, Oct. 2017.
strategies with quantized and statistical CSIT,” IEEE Trans. Wireless [115] N. H. Khattab and S. S. Soliman, “On data rate maximization using
Commun., vol. 16, no. 11, pp. 7025–7038, Nov. 2017. hybrid rate-splitting/noma network diamond relaying,” in International
[92] L. Yin, B. Clerckx, and Y. Mao, “Rate-splitting multiple access for Conference on Factory Communication Systems (WFCS), 2022, pp. 1–
multi-antenna broadcast channels with statistical CSIT,” in Proc. IEEE 4.
Wireless Commun. Netw. Conf. (WCNC), 2021. [116] T. Cover and A. Gamal, “Capacity theorems for the relay channel,”
[93] D. B. Amor, M. Joham, and W. Utschick, “Bilinear precoding for IEEE Trans. Inf. Theory, vol. 25, no. 5, pp. 572–584, 1979.
fdd massive mimo system with imperfect covariance matrices,” in [117] A. Sendonaris, E. Erkip, and B. Aazhang, “User cooperation diversity.
International ITG Workshop on Smart Antennas, 2020, pp. 1–6. part I. system description,” IEEE Trans. Commun., vol. 51, no. 11, pp.
[94] G. Lu, L. Li, H. Tian, and F. Qian, “MMSE-based precoding for rate 1927–1938, 2003.
splitting systems with finite feedback,” IEEE Commun. Lett., vol. 22, [118] J. Laneman, D. Tse, and G. Wornell, “Cooperative diversity in wireless
no. 3, pp. 642–645, Mar. 2018. networks: Efficient protocols and outage behavior,” IEEE Trans. Inf.
[95] N. Jindal, “MIMO broadcast channels with finite-rate feedback,” IEEE Theory, vol. 50, no. 12, pp. 3062–3080, 2004.
Trans. Inf. Theory, vol. 52, no. 11, pp. 5045–5060, Nov. 2006. [119] A. Papazafeiropoulos and T. Ratnarajah, “Rate-splitting robustness
[96] J. An, O. Dizdar, B. Clerckx, and W. Shin, “Rate-splitting multiple in multi-pair massive MIMO relay systems,” IEEE Trans. Wireless
access for multi-antenna broadcast channel with imperfect CSIT and Commun., vol. 17, no. 8, pp. 5623–5636, Aug. 2018.
CSIR,” in Proc. IEEE Annu. Symp. Pers. Indoor Mobile Radio Com- [120] R. Bassily, E. Ekrem, X. He, E. Tekin, J. Xie, M. R. Bloch, S. Ulukus,
mun. (PIMRC), 2020. and A. Yener, “Cooperative security at the physical layer: A summary
[97] A. Adhikary, J. Nam, J.-Y. Ahn, and G. Caire, “Joint spatial division of recent advances,” IEEE Signal Process. Mag., vol. 30, no. 5, pp.
and multiplexing—the large-scale array regime,” IEEE Trans. Inf. 16–28, 2013.
Theory, vol. 59, no. 10, pp. 6441–6463, Oct. 2013. [121] H. Fu, S. Feng, W. Tang, and D. W. K. Ng, “Robust secure beamform-
[98] J. Chen and V. K. N. Lau, “Two-tier precoding for FDD multi-cell ing design for two-user downlink MISO rate-splitting systems,” IEEE
massive MIMO time-varying interference networks,” IEEE J. Sel. Areas Trans. Wireless Commun., vol. 19, no. 12, pp. 8351–8365, Dec. 2020.
Commun., vol. 32, no. 6, pp. 1230–1238, 2014. [122] P. Li, M. Chen, Y. Mao, Z. Yang, B. Clerckx, and M. Shikh-Bahaei,
[99] C. K. Thomas, B. Clerckx, L. Sanguinetti, and D. Slock, “A rate “Cooperative rate-splitting for secrecy sum-rate enhancement in multi-
splitting strategy for mitigating intra-cell pilot contamination in massive antenna broadcast channels,” in Proc. IEEE Annu. Symp. Pers. Indoor
MIMO,” in Proc. IEEE Int. Conf. Commun. (ICC) Workshop, 2020. Mobile Radio Commun. (PIMRC), 2020.
[100] A. Papazafeiropoulos, B. Clerckx, and T. Ratnarajah, “Rate-splitting to [123] H. Xia, Y. Mao, X. Zhou, B. Clerckx, and C. Li, “Secure beamforming
mitigate residual transceiver hardware impairments in massive MIMO design for rate-splitting multiple access in multi-antenna broadcast
systems,” IEEE Trans. Veh. Technol., vol. 66, no. 9, pp. 8196–8211, channel with confidential messages,” arXiv preprint arXiv:2202.07328,
Sept. 2017. 2022.
 41

[124] H. Xia, Y. Mao, B. Clerckx, X. Zhou, S. Han, and C. Li, “Weighted message decoding in cloud radio access networks,” IEEE Access, vol. 7,
sum-rate maximization for rate-splitting multiple access based secure pp. 80 350–80 365, June 2019.
communication,” Proc. IEEE Wireless Commun. Netw. Conf. (WCNC), [146] A. A. Ahmad, J. Kakar, R. Reifert, and A. Sezgin, “UAV-assisted C-
2022. RAN with rate splitting under base station breakdown scenarios,” in
[125] A. Salem, C. Masouros, and B. Clerckx, “Secure rate-splitting multiple Proc. IEEE Int. Conf. Commun. (ICC) Workshop, 2019.
access: How much of the split signal to reveal?” Proc. IEEE Int. Conf. [147] A. A. Ahmad, B. Matthiesen, A. Sezgin, and E. Jorswieck, “Energy
Commun. (ICC) Workshop, 2022. efficiency in C-RAN using rate splitting and common message decod-
[126] M. Abolpour, S. Aïssa, L. Musavian, and A. Bhowal, “Rate splitting ing,” in Proc. IEEE Int. Conf. Commun. (ICC) Workshop, 2020.
in the presence of untrusted users: Outage and secrecy outage perfor- [148] A. A. Ahmad, Y. Mao, A. Sezgin, and B. Clerckx, “Rate splitting
mances,” IEEE Open J. Commun. Soc., vol. 3, pp. 921–935, 2022. multiple access in C-RAN,” in Proc. IEEE Annu. Symp. Pers. Indoor
[127] Y. Mao, B. Clerckx, and V. O. K. Li, “Rate-splitting for multi-antenna Mobile Radio Commun. (PIMRC), 2020.
non-orthogonal unicast and multicast transmission: Spectral and energy [149] A. A. Ahmad, H. Dahrouj, A. Chaaban, A. Sezgin, T. Y. Al-Naffouri,
efficiency analysis,” IEEE Trans. Commun., vol. 67, no. 12, pp. 8754– and M.-S. Alouini, “Power minimization via rate splitting in downlink
8770, Dec. 2019. cloud-radio access networks,” in Proc. IEEE Int. Conf. Commun. (ICC)
[128] G. Zhou, Y. Mao, and B. Clerckx, “Rate-splitting multiple access for Workshop, 2020.
multi-antenna downlink communication systems: Spectral and energy [150] M. Z. Hassan, M. J. Hossain, J. Cheng, and V. C. M. Leung,
efficiency tradeoff,” IEEE Trans. Wireless Commun., vol. 21, no. 7, pp. “Device-clustering and rate-splitting enabled device-to-device coop-
4816–4828, 2022. eration framework in fog radio access network,” IEEE Trans. Green
[129] K. Weinberger, A. A. Ahmad, and A. Sezgin, “On synergistic benefits Commun. Netw., vol. 5, no. 3, pp. 1482–1501, Sept. 2021.
of rate splitting in IRS-assisted cloud radio access networks,” in Proc. [151] ——, “Energy-spectrum efficient content distribution in Fog-RAN
IEEE Int. Conf. Commun. (ICC) Workshop, 2021. using rate-splitting, common message decoding, and 3D-resource
[130] Z. Yang, J. Shi, Z. Li, M. Chen, W. Xu, and M. Shikh-Bahaei, “Energy matching,” IEEE Trans. Wireless Commun., vol. 20, no. 8, pp. 4929–
efficient rate splitting multiple access (RSMA) with reconfigurable 4946, Aug. 2021.
intelligent surface,” in Proc. IEEE Int. Conf. Commun. (ICC) Workshop, [152] R.-J. Reifert, A. A. Ahmad, Y. Mao, A. Sezgin, and B. Clerckx, “Rate-
2020. splitting multiple access in cache-aided cloud-radio access networks,”
[131] A. Bansal, K. Singh, and C.-P. Li, “Analysis of hierarchical rate Frontiers in Commun. and Netw., 2021.
splitting for intelligent reflecting surfaces-aided downlink multiuser [153] Z. Lin, M. Lin, T. de Cola, J.-B. Wang, W.-P. Zhu, and J. Cheng,
MISO communications,” IEEE Open J. Commun. Soc, vol. 2, pp. 785– “Supporting IoT with rate-splitting multiple access in satellite and
798, Apr. 2021. aerial integrated networks,” IEEE Internet Things J., vol. 8, no. 14,
[132] A. Jolly, S. Biswas, and K. Singh, “An analysis on rate-splitting pp. 11 123–11 134, July 2021.
multiple access for IRS aided 6G communication,” arXiv preprint [154] S.-M. Tseng, G.-Y. Chen, and H.-C. Chan, “Cross-layer resource
arXiv:2106.04418, 2021. management for downlink BF-NOMA-OFDMA video transmission
[133] A. Bansal, K. Singh, B. Clerckx, C.-P. Li, and M.-S. Alouini, “Rate- systems and supervised/unsupervised learning based approach,” IEEE
splitting multiple access for intelligent reflecting surface aided multi- Trans. Veh. Technol., vol. 71, no. 10, pp. 10 744–10 753, 2022.
user communications,” IEEE Trans. Veh. Technol., vol. 70, no. 9, pp. [155] L. Zhao, Y. Cui, Z. Liu, Y. Zhang, and S. Yang, “Adaptive streaming of
9217–9229, Sept. 2021. 360 videos with perfect, imperfect, and unknown FoV viewing prob-
[134] K. Weinberger, A. A. Ahmad, A. Sezgin, and A. Zappone, “Synergistic abilities in wireless networks,” IEEE Trans. Image Process., vol. 30,
benefits in IRS- and RS-enabled C-RAN with energy-efficient cluster- pp. 7744–7759, Aug. 2021.
ing,” IEEE Trans. Wireless Commun., vol. 21, no. 10, pp. 8459–8475, [156] E. Piovano, H. Joudeh, and B. Clerckx, “On coded caching in the
2022. overloaded miso broadcast channel,” in Proc. IEEE Int. Symp. Inf.
[135] H. Fu, S. Feng, and D. W. K. Ng, “Resource allocation design for Theory (ISIT), 2017.
IRS-aided downlink MU-MISO RSMA systems,” Proc. IEEE Int. Conf. [157] E. Piovano, H. Joudeh, and B. Clerckx, “Generalized degrees of
Commun. (ICC) Workshop, 2021. freedom of the symmetric cache-aided MISO broadcast channel with
[136] T. Fang, Y. Mao, S. Shen, Z. Zhu, and B. Clerckx, “Fully connected partial CSIT,” IEEE Trans. Inf. Theory, vol. 65, no. 9, pp. 5799–5815,
reconfigurable intelligent surface aided rate-splitting multiple access 2019.
for multi-user multi-antenna transmission,” in Proc. IEEE Int. Conf. [158] E. Demarchou, C. Psomas, and I. Krikidis, “Rate splitting with wireless
Commun. (ICC) Workshop, 2022, pp. 675–680. edge caching: A system-level-based co-design,” IEEE Trans. Commun.,
[137] A. S. de Sena, P. H. J. Nardelli, D. B. da Costa, P. Popovski, and vol. 70, no. 1, pp. 664–679, Jan. 2022.
C. B. Papadias, “Rate-splitting multiple access and its interplay with [159] J. Hachem, N. Karamchandani, S. Diggavi, and S. Moharir,
intelligent reflecting surfaces,” IEEE Commun. Mag., pp. 1–12, 2022. “Coded caching for heterogeneous wireless networks,” arXiv preprint
[138] K. Zhao, Y. Mao, Z. Yang, L. Lian, and B. Clerckx, “Reconfigurable arXiv:2006.01025, 2020.
intelligent surfaces empowered cooperative rate splitting with user [160] E. Piovano, H. Joudeh, and B. Clerckx, “Overloaded multiuser MISO
relaying,” in Proc. IEEE Int. Symp. Wireless Commun. Syst. (ISWCS), transmission with imperfect CSIT,” in Proc. 50th Asilomar Conf.
2022, pp. 1–6. Signals, Syst. Comput., 2016.
[139] G. Zhou, W. Xu, and G. Bauch, “Interference mitigation with rate [161] Y. Mao, E. Piovano, and B. Clerckx, “Rate-splitting multiple access for
splitting in multi-cell wireless networks,” in Proc. IEEE Int. Conf. overloaded cellular internet of things,” IEEE Trans. Commun., vol. 69,
Wireless and Mobile Comput., Netw. and Commun. (WiMob), 2012. no. 7, pp. 4504–4519, July 2021.
[140] H. Dahrouj and W. Yu, “Multicell interference mitigation with joint [162] O. Dizdar and B. Clerckx, “Rate splitting multiple access for multi-
beamforming and common message decoding,” IEEE Trans. Commun., antenna multi-carrier joint communications and jamming,” in Int. Conf.
vol. 59, no. 8, pp. 2264–2273, Aug. 2011. in Sens. Signal Process. for Defence (SSPD), 2021.
[141] J. Zhang, J. Zhang, Y. Zhou, H. Ji, J. Sun, and N. Al-Dhahir, “Energy [163] ——, “Rate-splitting multiple access for communications and jamming
and spectral efficiency tradeoff via rate splitting and common beam- in multi-antenna multi-carrier cognitive radio systems,” IEEE Trans.
forming coordination in multicell networks,” IEEE Trans. Commun., Inf. Forensics Secur., vol. 17, pp. 628–643, Feb. 2022.
vol. 68, no. 12, pp. 7719–7731, Dec. 2020. [164] Y. Mao, B. Clerckx, and V. O. K. Li, “Rate-splitting for multi-antenna
[142] X. Su, Y. Yuan, and Q. Wang, “Performance analysis of rate splitting non-orthogonal unicast and multicast transmission,” in Proc. IEEE Int.
in K-user interference channel under imperfect CSIT: Average sum Workshop Signal Process. Adv. Wireless Commun. (SPAWC), 2018.
rate, outage probability and SER,” IEEE Access, vol. 8, pp. 136 930– [165] Y. Mao and B. Clerckx, “Dirty paper coded rate-splitting for non-
136 946, July 2020. orthogonal unicast and multicast transmission with partial CSIT,” in
[143] D. Yu, J. Kim, and S. Park, “An efficient rate-splitting multiple access Proc. 54th Asilomar Conf. Signals, Syst. Comput., 2020.
scheme for the downlink of C-RAN systems,” IEEE Wireless Commun. [166] L. F. Abanto-Leon, M. Hollick, B. Clerckx, and G. H. A. Sim,
Lett., vol. 8, no. 6, pp. 1555–1558, Dec. 2019. “Sequential parametric optimization for rate-splitting precoding in non-
[144] A. A. Ahmad, H. Dahrouj, A. Chaaban, A. Sezgin, and M. Alouini, “In- orthogonal unicast and multicast transmissions,” Proc. IEEE Int. Conf.
terference mitigation via rate-splitting in cloud radio access networks,” Commun. (ICC), 2022.
in Proc. IEEE Int. Workshop Signal Process. Adv. Wireless Commun. [167] O. Tervo, L. Trant, S. Chatzinotas, B. Ottersten, and M. Juntti,
(SPAWC), 2018. “Multigroup multicast beamforming and antenna selection with rate-
[145] A. Alameer Ahmad, H. Dahrouj, A. Chaaban, A. Sezgin, and splitting in multicell systems,” in Proc. IEEE Int. Workshop Signal
M. Alouini, “Interference mitigation via rate-splitting and common Process. Adv. Wireless Commun. (SPAWC), 2018.
42

[168] L. Yin and B. Clerckx, “Rate-splitting multiple access for multigroup in Proc. IEEE Int. Workshop Signal Process. Adv. Wireless Commun.
multicast and multibeam satellite systems,” IEEE Trans. Commun., (SPAWC), 2021.
vol. 69, no. 2, pp. 976–990, Feb. 2021. [191] L. Yin and B. Clerckx, “Rate-splitting multiple access for dual-
[169] A. Z. Yalcin, M. Yuksel, and B. Clerckx, “Rate splitting for multi-group functional radar-communication satellite systems,” in Proc. IEEE Wire-
multicasting with a common message,” IEEE Trans. Veh. Technol., less Commun. Netw. Conf. (WCNC), 2022.
vol. 69, no. 10, pp. 12 281–12 285, Oct. 2020. [192] H. Liu, T. A. Tsiftsis, B. Clerckx, K. J. Kim, K. S. Kwak, and H. V.
[170] H. Chen, D. Mi, Z. Liu, P. Xiao, and R. Tafazolli, “Rate-splitting Poor, “Rate splitting multiple access for semi-grant-free transmissions,”
for overloaded multigroup multicast: Error performance evaluation,” arXiv preprint arXiv:2110.02127, 2021.
in Proc. IEEE Int. Conf. Commun. (ICC) Workshop, 2020. [193] E. J. D. Santos, R. D. Souza, and J. L. Rebelatto, “Rate-splitting
[171] H. Chen, D. Mi, T. Wang, Z. Chu, Y. Xu, D. He, and P. Xiao, “Rate- multiple access for URLLC uplink in physical layer network slicing
splitting for multicarrier multigroup multicast: Precoder design and with eMBB,” IEEE Access, vol. 9, pp. 163 178–163 187, 2021.
error performance,” IEEE Trans. Broadcast., vol. 67, no. 3, pp. 619– [194] Y. Liu, B. Clerckx, and P. Popovski, “Network slicing for eMBB,
630, Sept. 2021. URLLC, and mMTC: An uplink rate-splitting multiple access ap-
[172] H. Chen, D. Mi, B. Clerckx, Z. Chu, J. Shi, and P. Xiao, “Joint power proach,” arXiv preprint arXiv:2208.10841, 2022.
and subcarrier allocation optimization for multigroup multicast systems [195] N. Devroye and P. Popovski, “Receiver-side opportunism in cognitive
with rate splitting,” IEEE Trans. Veh. Technol., vol. 69, no. 2, pp. 2306– networks,” in Proc. ICST CROWNCOM Conf., 2011.
2310, 2020. [196] M. R. Camana Acosta, C. E. G. Moreta, and I. Koo, “Joint power al-
[173] L. Zhao, Y. Cui, S. Yang, and S. Shamai, “An optimization framework location and power splitting for MISO-RSMA cognitive radio systems
for general rate splitting for general multicast,” IEEE Trans. Wireless with SWIPT and information decoder users,” IEEE Syst. J., vol. 15,
Commun., pp. 1–1, 2022. no. 4, pp. 5289–5300, Dec. 2021.
[174] M. Caus, A. Pastore, M. Navarro, T. Ramirez, C. Mosquera, N. Noels, [197] S. Naser, P. C. Sofotasios, L. Bariah, W. Jaafar, S. Muhaidat, M. Al-
N. Alagha, and A. I. Perez-Neira, “Exploratory analysis of superposi- Qutayri, and O. A. Dobre, “Rate-splitting multiple access: Unifying
tion coding and rate splitting for multibeam satellite systems,” in Proc. NOMA and SDMA in MISO VLC channels,” IEEE Open J. of Veh.
IEEE Int. Symp. Wireless Commun. Syst. (ISWCS), 2018. Technol., vol. 1, pp. 393–413, Oct. 2020.
[175] L. Yin and B. Clerckx, “Rate-splitting multiple access for multibeam [198] S. Tao, H. Yu, Q. Li, Y. Tang, and D. Zhang, “One-layer rate-splitting
satellite communications,” in Proc. IEEE Int. Conf. Commun. (ICC) multiple access with benefits over power-domain NOMA in indoor
Workshop, 2020. multi-cell visible light communication networks,” in Proc. IEEE Int.
[176] Z. W. Si, L. Yin, and B. Clerckx, “Rate-splitting multiple access for Conf. Commun. (ICC) Workshop, 2020.
multigateway multibeam satellite systems with feeder link interfer- [199] S. Ma, H. Zhou, Y. Mao, X. Liu, Y. Wu, B. Clerckx, Y. Wang, and S. Li,
ence,” IEEE Trans. Commun., vol. 70, no. 3, pp. 2147–2162, 2022. “Robust beamforming design for rate splitting multiple access-aided
[177] M. Vazquez, M. Caus, and A. Perez-Neira, “Rate splitting for MIMO MISO visible light communications,” arXiv preprint arXiv:2108.07014,
multibeam satellite systems,” in International ITG Workshop on Smart 2021.
Antennas, 2018. [200] L. Li, K. Chai, J. Li, and X. Li, “Resource allocation for multicar-
[178] S. K. Singh, K. Agrawal, K. Singh, and C.-P. Li, “Outage probability rier rate-splitting multiple access system,” IEEE Access, vol. 8, pp.
and throughput analysis of UAV-assisted rate-splitting multiple access,” 174 222–174 232, Sept. 2020.
IEEE Wireless Commun. Lett., vol. 10, no. 11, pp. 2528–2532, Nov. [201] Y. Mao, B. Clerckx, and V. O. K. Li, “Rate-splitting for multi-user
2021. multi-antenna wireless information and power transfer,” in Proc. IEEE
[179] A. Rahmati, Y. Yapici, N. Rupasinghe, I. Guvenc, H. Dai, and Int. Workshop Signal Process. Adv. Wireless Commun. (SPAWC), 2019.
A. Bhuyan, “Energy efficiency of RSMA and NOMA in cellular- [202] X. Su, L. Li, H. Yin, and P. Zhang, “Robust power- and rate-splitting-
connected mmWave UAV networks,” in Proc. IEEE Int. Conf. Com- based transceiver design in K-user MISO SWIPT interference channel
mun. (ICC) Workshop, 2019. under imperfect CSIT,” IEEE Commun. Lett., vol. 23, no. 3, pp. 514–
[180] W. Jaafar, S. Naser, S. Muhaidat, P. C. Sofotasios, and 517, Mar. 2019.
H. Yanikomeroglu, “Multiple access in aerial networks: From [203] R. Chen, F. Cheng, J. Lin, L. Liang, and Y. Sun, “Performance analysis
orthogonal and non-orthogonal to rate-splitting,” IEEE Open J. of Veh. of rate splitting multiple access based vortex wave communications,”
Technol., vol. 1, pp. 372–392, Oct. 2020. IEEE Wireless Commun. Lett., pp. 1–1, 2022.
[181] ——, “On the downlink performance of RSMA-based UAV communi- [204] H. Liu, Y. Ye, Z. Bai, K. J. Kim, and T. A. Tsiftsis, “Rate splitting mul-
cations,” IEEE Trans. Veh. Technol., vol. 69, no. 12, pp. 16 258–16 263, tiple access aided mobile edge computing in cognitive radio networks,”
Dec. 2020. arXiv preprint arXiv:2204.06208, 2022.
[182] L. Lu, Y. Hu, Y. Zhang, G. Jia, J. Nie, and M. Shikh-Bahaei, “Machine [205] S. Rezvani, E. A. Jorswieck, N. M. Yamchi, and M. R. Javan, “Optimal
learning for predictive deployment of uavs with rate splitting multiple SIC ordering and power allocation in downlink multi-cell NOMA
access,” in Proc. IEEE Glob. Commun. Conf. (GLOBECOM), 2020. systems,” IEEE Trans. Wireless Commun., vol. 21, no. 6, pp. 3553–
[183] Z. Xiao, Z. Han, A. Nallanathan, O. A. Dobre, B. Clerckx, J. Choi, 3569, 2022.
C. He, and W. Tong, “Antenna array enabled space/air/ground com- [206] A. Bansal, K. Singh, B. Clerckx, C.-P. Li, and M.-S. Alouini, “Rate-
munications and networking for 6G,” IEEE J. Sel. Areas Commun., splitting multiple access for intelligent reflecting surface aided multi-
vol. 40, no. 10, pp. 2773–2804, 2022. user communications,” IEEE Trans. Veh. Technol., vol. 70, no. 9, pp.
[184] L. Yin and B. Clerckx, “Rate-splitting multiple access for satellite- 9217–9229, 2021.
terrestrial integrated networks: Benefits of coordination and coopera- [207] H. Bastami, M. Moradikia, A. Abdelhadi, H. Behroozi, B. Clerckx, and
tion,” IEEE Trans. Wireless Commun., pp. 1–1, 2022. L. Hanzo, “Maximizing the secrecy energy efficiency of the cooperative
[185] Z. Lin, M. Lin, B. Champagne, W.-P. Zhu, and N. Al-Dhahir, “Secure rate-splitting aided downlink in multi-carrier UAV networks,” IEEE
and energy efficient transmission for RSMA-based cognitive satellite- Trans. Veh. Technol., pp. 1–17, 2022.
terrestrial networks,” IEEE Wireless Commun. Lett., vol. 10, no. 2, pp. [208] A. D. Wyner, “The wire-tap channel,” The Bell System Technical
251–255, Feb. 2021. Journal, vol. 54, no. 8, pp. 1355–1387, 1975.
[186] A. Salem, C. Masouros, and B. Clerckx, “Rate splitting with finite [209] M. Bloch, J. Barros, M. R. D. Rodrigues, and S. W. McLaughlin,
constellations: The benefits of interference exploitation vs suppression,” “Wireless information-theoretic security,” IEEE Trans. Inf. Theory,
IEEE Open J. Commun. Soc., vol. 2, pp. 1541–1557, June 2021. vol. 54, no. 6, pp. 2515–2534, 2008.
[187] A. Salem and C. Masouros, “Rate splitting approach under PSK [210] G. Durisi, T. Koch, and P. Popovski, “Toward massive, ultra-reliable,
signaling using constructive interference precoding technique,” in Proc. and low-latency wireless communication with short packets,” Proc.
IEEE Wireless Commun. Netw. Conf. (WCNC), 2019. IEEE, vol. 104, no. 9, pp. 1711–1726, 2016.
[188] C. Xu, B. Clerckx, S. Chen, Y. Mao, and J. Zhang, “Rate-splitting [211] H. Sato, “The capacity of the Gaussian interference channel under
multiple access for multi-antenna joint communication and radar trans- strong interference (corresp.),” IEEE Trans. Inf. Theory, vol. 27, no. 6,
missions,” in Proc. IEEE Int. Conf. Commun. (ICC) Workshop, 2020. pp. 786–788, 1981.
[189] ——, “Rate-splitting multiple access for multi-antenna joint radar and [212] I. Sason, “On achievable rate regions for the Gaussian interference
communications,” IEEE J. Sel. Top. Signal Process., vol. 15, no. 6, pp. channel,” IEEE Trans. Inf. Theory, vol. 50, no. 6, pp. 1345–1356, 2004.
1332–1347, Nov. 2021. [213] A. G. Davoodi and S. A. Jafar, “Transmitter cooperation under finite
[190] R. Cerna-Loli, O. Dizdar, and B. Clerckx, “A rate-splitting strategy precision CSIT: A GDoF perspective,” IEEE Trans. Inf. Theory, vol. 63,
to enable joint radar sensing and communication with partial CSIT,” no. 9, pp. 6020–6030, Sept. 2017.
 43

[214] ——, “Sum-set inequalities from aligned image sets: Instruments for [238] ——, “Discrete signaling and treating interference as noise for the
robust GDoF bounds,” IEEE Trans. Inf. Theory, vol. 66, no. 10, pp. Gaussian interference channel,” IEEE Trans. Inf. Theory, vol. 67,
6458–6487, 2020. no. 11, pp. 7253–7284, Nov. 2021.
[215] S. Yang and Z. Lit, “A constant-gap result on the multi-antenna [239] N. Q. Hieu, D. T. Hoang, D. Niyato, and D. I. Kim, “Optimal power
broadcast channels with linearly precoded rate splitting,” in Proc. IEEE allocation for rate splitting communications with deep reinforcement
Int. Workshop Signal Process. Adv. Wireless Commun. (SPAWC), 2018. learning,” IEEE Wireless Commun. Lett., vol. 10, no. 12, pp. 2820–
[216] Z. Li and S. Yang, “A linearly precoded rate splitting approach and its 2823, 2021.
optimality for mimo broadcast channels,” in Proc. IEEE Inf. Theory [240] T. Li, A. K. Sahu, A. Talwalkar, and V. Smith, “Federated learning:
Workshop (ITW), 2018. Challenges, methods, and future directions,” IEEE Trans. Signal Pro-
[217] J. Wang and S. A. Jafar, “Sum-GDoF of symmetric multi-hop interfer- cess., vol. 37, no. 3, pp. 50–60, Mar. 2020.
ence channel under finite precision CSIT using aligned-images sum-set [241] S.-H. Park and H. Lee, “Completion time minimization of fog-RAN-
inequalities,” IEEE Trans. Inf. Theory, vol. 68, no. 7, pp. 4470–4490, assisted federated learning with rate-splitting transmission,” IEEE
2022. Trans. Veh. Technol., to appear, 2022.
[218] A. Tajer, A. Steiner, and S. Shamai (Shitz), “The broadcast approach [242] C. K. Au-Yeung and D. J. Love, “On the performance of random vector
in communication networks,” Entropy, vol. 23, no. 1, 2021. [Online]. quantization limited feedback beamforming in a MISO system,” IEEE
Available: https://www.mdpi.com/1099-4300/23/1/120 Trans. Wireless Commun., vol. 6, no. 2, pp. 458–462, Feb. 2007.
[219] M. Vaezi and H. Vincent Poor, NOMA: An Information-Theoretic [243] P. Ding, D. J. Love, and M. D. Zoltowski, “Multiple antenna broadcast
Perspective. Cham: Springer International Publishing, 2019, pp. 167– channels with shape feedback and limited feedback,” IEEE Trans.
193. Signal Process., vol. 55, no. 7, pp. 3417–3428, July 2007.
[220] T. Cover, “Broadcast channels,” IEEE Trans. Inf. Theory, vol. 18, no. 1, [244] C. K. Au Yeung, D. J. Love, and S. Sanayei, “Trellis coded line
pp. 2–14, Jan. 1972. packing: Large dimensional beamforming vector quantization and
[221] B. Makki, K. Chitti, A. Behravan, and M.-S. Alouini, “A survey of feedback transmission,” IEEE Trans. Wireless Commun., vol. 10, no. 6,
NOMA: Current status and open research challenges,” IEEE Open J. pp. 1844–1853, June 2011.
Commun. Soc., vol. 1, pp. 179–189, 2020.
[222] J. G. Proakis, Digital Communications. New York, NY, USA:
McGraw-Hill, 2008.
[223] L. Ping, L. Liu, K. Wu, and W. Leung, “Interleave division multiple-
access,” IEEE Trans. Wireless Commun., vol. 5, no. 4, pp. 938–947,
2006.
[224] H. Nikopour and H. Baligh, “Sparse code multiple access,” in Proc.
IEEE Int. Symp. Pers. Indoor Mobile Radio Commun. (PIMRC), Sep.
2013, pp. 332–336.
[225] Nokia and Alcatel-Lucent Shanghai Bell, “Non-orthogonal multiple
access for new radio,” 3rd Generation Partnership Project (3GPP),
document R1-165019, 3GPP TSG RAN WG1 Meeting #85, May 2016. Bruno Clerckx is a Professor, the Head of the
[226] R. Hadani, S. Rakib, M. Tsatsanis, A. Monk, A. J. Goldsmith, Wireless Communications and Signal Processing
A. F. Molisch, and R. Calderbank, “Orthogonal time frequency space Lab, and the Deputy Head of the Communications
modulation,” in Proc. IEEE Wireless Commun. Netw. Conf. (WCNC), and Signal Processing Group, within the Electrical
2017. and Electronic Engineering Department, Imperial
[227] R. Hadani, S. Rakib, S. Kons, M. Tsatsanis, A. Monk, C. Ibars, College London, London, U.K. He is also the Chief
J. Delfeld, Y. Hebron, A. J. Goldsmith, A. F. Molisch, and Technology Officer (CTO) of Silicon Austria Labs
R. Calderbank, “Orthogonal time frequency space modulation,” 2018, (SAL) where he is responsible for all research areas
arXiv:1808.00519. of Austria’s top research center for electronic based
[228] P. Raviteja, K. T. Phan, and Y. Hong, “Embedded pilot-aided channel systems. He received the MSc and Ph.D. degrees
estimation for OTFS in delay–doppler channels,” IEEE Trans. Veh. in Electrical Engineering from Université Catholique
Technol., vol. 68, no. 5, pp. 4906–4917, 2019. de Louvain, Belgium, and the Doctor of Science (DSc) degree from Imperial
[229] Z. Wei, W. Yuan, S. Li, J. Yuan, G. Bharatula, R. Hadani, and College London, UK. He has authored two books on “MIMO Wireless
L. Hanzo, “Orthogonal time-frequency space modulation: A promising Communications” and “MIMO Wireless Networks”, 250 peer-reviewed in-
next-generation waveform,” IEEE Wireless Commun., vol. 28, no. 4, ternational research papers, and 150 standards contributions, and is the
pp. 136–144, 2021. inventor of 80 issued or pending patents among which 15 have been adopted
[230] P. Raviteja, K. T. Phan, Y. Hong, and E. Viterbo, “Interference in the specifications of 4G standards and are used by billions of devices
cancellation and iterative detection for orthogonal time frequency space worldwide. His research spans the general area of wireless communications
modulation,” IEEE Trans. Wireless Commun., vol. 17, no. 10, pp. 6501– and signal processing for wireless networks. He served as an editor or guest
6515, 2018. editor for the IEEE TRANSACTIONS ON COMMUNICATIONS, the IEEE
[231] T. Thaj and E. Viterbo, “Low complexity iterative rake decision TRANSACTIONS ON WIRELESS COMMUNICATIONS, and the IEEE
feedback equalizer for zero-padded OTFS systems,” IEEE Trans. Veh. TRANSACTIONS ON SIGNAL PROCESSING, EURASIP Journal on Wire-
Technol., vol. 69, no. 12, pp. 15 606–15 622, 2020. less Communications and Networking, IEEE ACCESS, the IEEE JOURNAL
[232] Q. Li, J. Yuan, and H. Lin, “Iterative mmse detection for orthogonal ON SELECTED AREAS IN COMMUNICATIONS, the IEEE JOURNAL
time frequency space modulation,” in Proc. IEEE Int. Conf. Commun. OF SELECTED TOPICS IN SIGNAL PROCESSING, the PROCEEDINGS
(ICC), 2022, pp. 1–6. OF THE IEEE, and the IEEE Open Journal of the Communications Society.
[233] H. Lin and J. Yuan, “Orthogonal delay-Doppler division multiplexing He was an Editor for the 3GPP LTE-Advanced Standard Technical Report
modulation,” IEEE Trans. Wireless Commun., vol. 1, no. 10, pp. 6501– on CoMP. He received the prestigious Blondel Medal 2021 from France
6515, 2022. for exceptional work contributing to the progress of Science and Electrical
[234] H. Lin and J. Yuan, “Multicarrier modulation on delay-Doppler plane: and Electronic Industries, the 2021 Adolphe Wetrems Prize in mathematical
Achieving orthogonality with fine resolutions,” in Proc. IEEE Int. Conf. and physical sciences from Royal Academy of Belgium, multiple awards
Commun. (ICC), 2022, pp. 1–6. from Samsung, IEEE best student paper award, and the EURASIP (European
[235] M. Qiu, Y.-C. Huang, S.-L. Shieh, and J. Yuan, “A lattice-partition Association for Signal Processing) best paper award 2022. He is a Fellow of
framework of downlink non-orthogonal multiple access without SIC,” the IEEE and the IET, and an IEEE Communications Society Distinguished
IEEE Trans. Commun., vol. 66, no. 6, pp. 2532–2546, Jun. 2018. Lecturer 2021-2023.
[236] M. Qiu, Y.-C. Huang, J. Yuan, and C.-L. Wang, “Lattice-partition-based
downlink non-orthogonal multiple access without SIC for slow fading
channels,” IEEE Trans. Commun., vol. 67, no. 2, pp. 1166–1181, Feb.
2019.
[237] M. Qiu, Y.-C. Huang, and J. Yuan, “Downlink non-orthogonal multiple
access without SIC for block fading channels: An algebraic rotation
approach,” IEEE Trans. Wireless Commun., vol. 18, no. 8, pp. 3903–
3918, Aug. 2019.
44

Yijie Mao is an Assistant Professor at the School David J. Love (S’98 - M’05 - SM’09 - F’15)
of Information Science and Technology, Shang- is the Nick Trbovich Professor of Electrical and
haiTech University (Shanghai, China). She received Computer Engineering at Purdue University. His
the B.Eng. degree from the Beijing University of research interests are in the design and analysis of
Posts and Telecommunications (Beijing, China), the broadband wireless communication systems, beyond
B.Eng. degree (Hons.) from the Queen Mary Uni- 5G wireless systems, multiple-input multiple-output
versity of London (London, United Kingdom) in (MIMO) communications, millimeter wave wireless,
2014, and the Ph.D. degree from the Electrical and software defined radios and wireless networks, cod-
Electronic Engineering Department, the University ing theory, and MIMO array processing. He holds
of Hong Kong (Hong Kong, China) in 2018. She was 32 issued US patents. He served as a Senior Editor
a Postdoctoral Research Fellow at the University of for IEEE Signal Processing Magazine, Editor for
Hong Kong from 2018 to 2019 and a postdoctoral research associate with the the IEEE Transactions on Communications, Associate Editor for the IEEE
Communications and Signal Processing Group, Department of the Electrical Transactions on Signal Processing, and Guest Editor for special issues of
and Electronic Engineering at the Imperial College London (London, United the IEEE Journal on Selected Areas in Communications and the EURASIP
Kingdom) from 2019 to 2021. Her research interests include the design of Journal on Wireless Communications and Networking. Since 2022, he is
future wireless communications and artificial intelligence-empowered wireless a Fellow of the American Association for the Advancement of Science
networks. Dr. Mao receives the Best Paper Award of EURASIP Journal on (AAAS). Along with his co-authors, he won best paper awards from the IEEE
Wireless Communications and Networking 2022 and the Exemplary Reviewer Communications Society (2016 Stephen O. Rice Prize and 2020 Fred Ellersick
for IEEE Transactions on Communications 2021. She is currently serving Prize), the IEEE Signal Processing Society (2015 IEEE Signal Processing
as a guest editor for special issues of IEEE Journal on Selected Areas in Society Best Paper Award), and the IEEE Vehicular Technology Society (2010
Communications and IEEE Open Journal of the Communications Society. She Jack Neubauer Memorial Award).
has been a vice-chair of IEEE ComSoc WTC SIG on rate-splitting multiple
access (RSMA) and a workshop co-chair for 2020-2022 IEEE ICC, 2021-
2022 IEEE WCNC, 2020-2022 IEEE PIMRC, and 2022 IEEE SECON, and
she has been a Technical Program Committee member of many symposia on
wireless communication for several leading international IEEE conferences.

Eduard A. Jorswieck is managing director of the Elza Erkip is an Institute Professor in the Electri-
Institute of Communications Technology and the cal and Computer Engineering Department at New
head of the Chair for Information Theory and Com- York University Tandon School of Engineering. She
munications Systems and Full Professor at Technis- received the B.S. degree in Electrical and Electronics
che Universität Braunschweig, Brunswick, Germany. Engineering from Middle East Technical University,
From 2008 until 2019, he was the head of the Ankara, Turkey, and the M.S. and Ph.D. degrees
Chair of Communications Theory and Full Professor in Electrical Engineering from Stanford University,
at TU Dresden, Germany. Eduard’s main research Stanford, CA, USA. Her research interests are in
interests are in the broad area of communications. information theory, communication theory, and wire-
He has co-authored some 160 journal papers, 15 less communications.
book chapters, 4 monographs, and more than 300 Dr. Erkip is a member of the Science Academy of
conference papers. Since 2017, he serves as Editor-in-Chief of the Springer Turkey and is a Fellow of the IEEE. She received the NSF CAREER award
EURASIP Journal on Wireless Communications and Networking. He currently in 2001, the IEEE Communications Society WICE Outstanding Achievement
serves as editor for IEEE Transactions on Communications. He has served Award in 2016, the IEEE Communications Society Communication Theory
on the editorial boards for IEEE Transactions on Signal Processing, IEEE Technical Committee (CTTC) Technical Achievement Award in 2018, and
Transactions on Wireless Communications, IEEE Signal Processing Letters, the IEEE Communications Society Edwin Howard Armstrong Achievement
and IEEE Transactions on Information Forensics and Security. In 2006, he Award in 2021. She was the Padovani Lecturer of the IEEE Information
received the IEEE Signal Processing Society Best Paper Award. Theory Society in 2022. Her paper awards include the IEEE Communications
Society Stephen O. Rice Paper Prize in 2004, the IEEE Communications
Society Award for Advances in Communication in 2013 and the IEEE
Communications Society Best Tutorial Paper Award in 2019. She was a
member of the Board of Governors of the IEEE Information Theory Society
2012-2020, where she was the President in 2018. She was a Distinguished
Jinhong Yuan (M’02–SM’11–F’16) is a Professor Lecturer of the IEEE Information Theory Society from 2013 to 2014.
and Head of Telecommunication Group with the
School of Electrical Engineering and Telecommuni-
cations, The university of New South Wales, Sydney,
Australia. He has published two books, five book
chapters, over 300 papers in telecommunications
journals and conference proceedings, and 50 indus-
trial reports. He is a co-inventor of one patent on
MIMO systems and four patents on low-density-
parity-check codes. He has co-authored four Best Dusit Niyato is the President’s Chair Professor in
Paper Awards and one Best Poster Award, including the School of Computer Science and Engineering,
the Best Paper Award from the IEEE International Conference on Communi- at Nanyang Technological University, Singapore. He
cations, Kansas City, USA, in 2018, the Best Paper Award from IEEE Wireless received B.Eng. from King Mongkuts Institute of
Communications and Networking Conference, Cancun, Mexico, in 2011, and Technology Ladkrabang (KMITL), Thailand in 1999
the Best Paper Award from the IEEE International Symposium on Wireless and Ph.D. in Electrical and Computer Engineering
Communications Systems, Trondheim, Norway, in 2007. He is an IEEE Fellow from the University of Manitoba, Canada in 2008.
and currently serving as an Associate Editor for the IEEE Transactions on His research interests are in the areas of sustainabil-
Wireless Communications and IEEE Transactions on Communications. He ity, edge intelligence, decentralized machine learn-
served as the IEEE NSW Chapter Chair of Joint Communications/Signal ing, and incentive mechanism design.
Processions/Ocean Engineering Chapter during 2011-2014 and served as
an Associate Editor for the IEEE Transactions on Communications during
2012-2017. His current research interests include error control coding and
information theory, communication theory, and wireless communications.