PaperlayerView exact revision on arXiv

Source-linked AI summary

Multiple Access Techniques for Intelligent and Multi-Functional 6G: Tutorial, Survey, and Outlook

Bruno Clerckx, Yijie Mao, Zhaohui Yang, Mingzhe Chen, Ahmed Alkhateeb, Liang Liu, Min Qiu, Jinhong Yuan, Vincent W. S. Wong, Juan Montojo

arXiv:2401.01433v1cs.ITeess.SPmath.IT

TL;DR

Future 6G networks are expected to integrate AI and multiple functions, prompting a rethinking of multiple access. The paper synthesizes MA mechanisms and research directions; it highlights RSMA’s unifying role while concluding that substantial work remains.

  • Problem

    The integration of AI and multiple functions in future wireless networks calls for rethinking MA techniques to serve users, devices, services, and training nodes efficiently.

    Source: §VIII · p. 43

  • Method

    The paper synthesizes MA approaches and research directions spanning RSMA, AI-enabled learning, sensing and localization, digital twins, and IoT random access.

    Source: E. Rate-Splitting Multiple Access · p. 8, 2) Deep Learning · p. 18, 4) Technology Outlook and Future Works · pp. 28, 30

  • Results

    RSMA provides a unified, conceptually simple understanding of SDMA, PD-NOMA, physical-layer multicasting, and MA based on space, power, and signal dimensions.

    Source: §VIII · p. 43

  • Takeaways & Limitations

    The paper concludes that 6G’s AI and multifunctionality shift will spur MA redesign, with substantial research still needed.

    Source: §VIII · p. 43

Abstract

from arXiv · show

Multiple access (MA) is a crucial part of any wireless system and refers to techniques that make use of the resource dimensions to serve multiple users/devices/machines/services, ideally in the most efficient way. Given the needs of multi-functional wireless networks for integrated communications, sensing, localization, computing, coupled with the surge of machine learning / artificial intelligence (AI) in wireless networks, MA techniques are expected to experience a paradigm shift in 6G and beyond. In this paper, we provide a tutorial, survey and outlook of past, emerging and future MA techniques and pay a particular attention to how wireless network intelligence and multi-functionality will lead to a re-thinking of those techniques. The paper starts with an overview of orthogonal, physical layer multicasting, space domain, power domain, ratesplitting, code domain MAs, and other domains, and highlight the importance of researching universal multiple access to shrink instead of grow the knowledge tree of MA schemes by providing a unified understanding of MA schemes across all resource dimensions. It then jumps into rethinking MA schemes in the era of wireless network intelligence, covering AI for MA such as AI-empowered resource allocation, optimization, channel estimation, receiver designs, user behavior predictions, and MA for AI such as federated learning/edge intelligence and over the air computation. We then discuss MA for network multi-functionality and the interplay between MA and integrated sensing, localization, and communications. We finish with studying MA for emerging intelligent applications before presenting a roadmap toward 6G standardization. We also point out numerous directions that are promising for future research.

I. INTRODUCTION

The paper frames 6G as a shift toward AI-enabled, multifunctional networks, where multiple access (MA) must serve broader users and services. It surveys MA families and motivates unified designs alongside research on AI, network multifunctionality, and emerging applications.

A. From Communication-centric 5G to Intelligent and Multi-functional 6G

  • Future networks face demands for higher throughput and efficiency, reliability, massive connectivity, broad coverage, and increasingly heterogeneous quality of service.The paper links these demands to applications including augmented and virtual reality, industrial automation, and the Internet of Things.Source: A. From Communication-centric 5G to… · p. 1

B. The Crucial Role of Multiple Access

  • MA allocates time, frequency, power, spatial, and signal resources across communication devices and other service targets, making its redesign central to multifunctional networks.The paper includes sensing targets, devices to be charged, and AI nodes among the entities networks may serve.Source: B. The Crucial Role of Multiple Access · p. 2

C. Objectives and Organization

  • The paper surveys OMA, SDMA, PD-NOMA, RSMA, code-domain MA, and schemes exploiting other domains, discussing their advantages, disadvantages, and relationships.It presents this broad overview beyond a classification focused on orthogonal versus non-orthogonal access.Source: C. Objectives and Organization · pp. 2–3
  • The paper motivates universal multiple access (UMA) to unify RSMA with code-domain and other MA dimensions, while noting that no UMA yet optimally suits all applications.The figure depicts scheme unification as moving from leaves toward a trunk.Source: C. Objectives and Organization · p. 3, Figure 1 · p. 3
  • The paper reviews AI for MA and MA for AI, then considers MA for sensing, localization, communications, emerging applications, and a 6G standardization roadmap.Its topics include AI-enabled resource allocation, channel estimation, receiver design, user behavior prediction, federated learning, and over-the-air computation.Source: C. Objectives and Organization · p. 3

II. AN OVERVIEW OF MULTIPLE ACCESS TECHNIQUES

  • OMA avoids multi-user interference by assigning users distinct, non-overlapping frequency bands, time slots, or codes.The overview also describes OFDMA as allocating groups of narrow subcarriers and time slots.Source: A. Orthogonal Multiple Access · pp. 3–4

A. Orthogonal Multiple Access

  • OMA is simple and interference-free for low-to-moderate user loads, but can waste spectrum and limits capacity to the available orthogonal resources.A low-rate user may occupy an entire resource block, and improving performance through scheduling can increase signaling overhead.Source: A. Orthogonal Multiple Access · p. 4

B. Physical Layer Multicasting

These sections describe physical-layer multicasting, spatial sharing through SDMA, and power-domain NOMA as distinct ways to serve multiple users, with different interference-management approaches and operating conditions.

B. Physical Layer Multicasting

  • Physical-layer multicasting jointly encodes multiple users’ messages into one stream decoded by all, enabling simultaneous decoding but making the weakest user determine the rate.Joint coding can increase reliability and reduce latency, while constraining stronger users and spectrum use.Source: B. Physical Layer Multicasting · pp. 4–5

C. Space Division Multiple Access

  • SDMA lets users share time-frequency resources through spatial separation; practical linear precoding treats multiuser interference as noise, unlike capacity-achieving but computationally demanding DPC.MU-LP is particularly useful with semi-orthogonal channels, comparable signal strengths, and perfect CSIT.Source: C. Space Division Multiple Access · p. 5
  • SDMA’s spectrum-efficiency and interference benefits rely on perfect CSIT and an underloaded network, while inaccurate CSIT, overload, or poorly aligned user channels degrade performance.Overloaded scenarios may use user grouping, at the cost of reduced QoS and increased latency; scheduling and signaling also add overhead.Source: C. Space Division Multiple Access · pp. 5–6

D. Power-domain Non-Orthogonal Multiple Access

  • PD-NOMA uses superposition coding and successive interference cancellation, and its SISO capacity results motivated MIMO extensions, but its 5G study did not yield convincing gains over SDMA/MU-MIMO.In MIMO NOMA, users are grouped; within-group interference is managed by SIC and between-group interference is treated as noise.Source: D. Power-domain Non-Orthogonal… · p. 6
  • PD-NOMA can improve spectrum efficiency in extremely overloaded networks with closely aligned channels and diverse channel strengths.Its stated operating advantage depends on those channel conditions.Source: D. Power-domain Non-Orthogonal… · p. 6
  • PD-NOMA can improve fairness in extremely overloaded networks when users have closely aligned channels and large channel-strength disparities.Power allocation favors weaker-channel users to enable successful interference cancellation.Source: D. Power-domain Non-Orthogonal… · p. 6
  • MIMO NOMA’s sum multiplexing gain is no greater than SDMA’s, while SIC, resource optimization, and signaling add complexity and overhead.Its performance is also sensitive to CSIT inaccuracy and network load.Source: D. Power-domain Non-Orthogonal… · p. 6

E. Rate-Splitting Multiple Access

Rate-splitting multiple access (RSMA) splits user messages into common and private parts to manage interference flexibly. Its variants span several established access schemes, with reported gains in spectral efficiency, fairness, latency, and other measures, alongside added optimization and signaling burdens.

E. Rate-Splitting Multiple Access

  • Its message splitting lets downlink users decode common and private streams, while uplink sub-messages can be decoded non-consecutively for partial interference decoding and partial treatment as noise.This gives RSMA a flexible interference-management strategy.Source: E. Rate-Splitting Multiple Access · p. 7
  • 1-layer RS uses a single common stream and individual private streams; HRS adds group-specific common streams and requires two SIC layers, while GRS explores all possible user groupings.1-layer RS requires one SIC layer; HRS requires two.Source: E. Rate-Splitting Multiple Access · p. 7
  • RSMA is a superset of multicasting, OMA, SDMA, and PD-NOMA, reducing to each under specific stream or power allocations.This follows because RSMA can turn off streams to recover those schemes.Source: E. Rate-Splitting Multiple Access · pp. 7–8
  • RSMA is reported to achieve the best sum-degree of freedom with perfect and imperfect CSIT, and its rate region approaches DPC’s capacity region with perfect CSIT.With imperfect CSIT, it can achieve a larger rate region than DPC.Source: E. Rate-Splitting Multiple Access · p. 9
  • RSMA is reported to improve energy efficiency and max-min fairness, and to achieve higher max-min rates and more pronounced spectral-efficiency gains under stringent user QoS constraints.Its benefits are also reported for cell-edge coverage through cooperative user relaying and for latency with finite-length codes.Source: E. Rate-Splitting Multiple Access · p. 9
  • RSMA’s costs include additional encoding and signaling, joint optimization of common and private precoders with common-rate allocation, and potentially increasing SIC burden in GRS.The expanded optimization space increases computational burden; increasing SIC layers can also burden receivers and introduce error propagation.Source: E. Rate-Splitting Multiple Access · p. 10

F. Code-Domain Multiple Access

Code-domain multiple access schemes let users share time-frequency resources through spreading sequences or codewords, while other schemes distinguish users using interleaving or resource patterns. Their reported trade-offs include spectral efficiency and overload support against detection, design, and synchronization challenges.

1) Low-Density Spreading (LDS) MA:

  • LDS-CDMA spreads user symbols with sparse sequences, reducing interference per chip and enabling low-density receiver graphs for message-passing or belief-propagation detection.LDS-OFDM extends the approach by spreading data across selected subcarriers; link-level PAPR and performance improvements over OFDM were reported.Source: 1) Low-Density Spreading (LDS) MA · p. 10
  • SCMA maps bits directly to sparse multidimensional codewords, combining modulation and spreading to gain spectral efficiency while retaining overloading and moderate detection complexity.Its overloading performance is limited by the number of sparse codewords and their sparsity.Source: 2) Sparse Code Multiple Access (SCMA) · p. 10

3) Multi-User Shared Access (MUSA):

  • MUSA uses short, dense, non-orthogonal sequences that users may select autonomously for grant-free IoT access, with SIC exploiting their low cross-correlation.A user may choose different sequences for different symbols to benefit from interference averaging.Source: 3) Multi-User Shared Access (MUSA) · p. 10

4) Successive Interference Cancellation Amenable Multiple Access (SAMA):

  • SAMA jointly designs its signature matrix and SIC-based message-passing algorithm, using frequency spreading with variable sparsity to obtain diversity gain.Across CD-NOMA, shared resources can improve spectral efficiency, but complex iterative detection, code design, channel estimation, and synchronization are challenges.Source: 4) Successive Interference… · p. 11

G. Multiple Access in Other Domains

1) Interleave-Division Multiple Access (IDMA):

  • IDMA distinguishes users through user-specific interleaving rather than distinct spreading sequences, and its chip-by-chip receiver complexity grows linearly with the number of users.Low-rate forward-error-correction codes can replace CDMA spreading to provide increased coding gain.Source: 1) Interleave-Division Multiple Access… · p. 11

2) Pattern Division Multiple Access (PDMA):

  • PDMA maps users onto shared time, frequency, or spatial resource groups using different patterns designed with disparate diversity order and sparsity for belief-propagation detection.A 500% gain in supported users over OFDMA was reported for the uplink at a 1% system packet drop rate.Source: 2) Pattern Division Multiple Access… · pp. 11–12

G. Multiple Access in Other Domains

  • Compute-and-forward relaying decodes noisy linear combinations of lattice codewords, and CFMA uses recovered sums as side information to decode other users’ codewords.The two-user Gaussian MAC capacity region was proved achievable under lattice decoding, provided both users’ SNR exceeds 1+…Source: 3) Compute-Forward Multiple Access… · p. 12

3) Compute-Forward Multiple Access (CFMA):

The section surveys multiple-access designs that use structured signaling, code coupling, layering, indices, or delay-Doppler resources. It also identifies shared design challenges and classifies MA dimensions to frame future development.

3) Compute-Forward Multiple Access (CFMA):

  • CFMA can achieve any rate point on the Gaussian MAC capacity region’s dominant face without time-sharing or rate-splitting, though practical implementations use binary LDPC codes and sum-product decoding.The cited proof used random lattice codes, which can be difficult to implement.Source: 3) Compute-Forward Multiple Access… · p. 12

4) Lattice-Partition-Based Multiple Access (LPBMA) without SIC:

  • LPBMA uses lattice-carved constellations whose superposition preserves lattice structure, enabling single-user TIN decoding and capacity-region performance within a constant gap.The gap is independent of the number of users and channel parameters.Source: 4) Lattice-Partition-Based Multiple… · p. 12

5) Spatially Coupled Multiple Access (SC-MA):

  • Spatially coupled MA repeats, permutes, and delays streams to create a sparse coupled graph; two-stage decoding asymptotically achieves Gaussian MAC capacity without decoder-demodulator feedback.Sliding-window decoding limits latency for streaming, but synchronization errors and fractional delay can be problematic.Source: 5) Spatially Coupled Multiple Access… · p. 12

6) Layered-Division Multiplexing (LDM):

  • LDM layers signals at different power levels and with potentially different coding, modulation, and antenna technologies, then decodes the layers using SIC.LDM is implemented on top of OFDM, with signal superposition before OFDM modulation.Source: 6) Layered-Division Multiplexing (LDM) · p. 13

7) Index Modulation Multiple Access (IMMA):

  • IMMA carries additional index bits through partial-resource activation patterns, but its joint maximum-likelihood detectors can be challenging to implement.Patterns may use time slots, subcarriers, antennas, spreading codes, or power levels.Source: 7) Index Modulation Multiple Access… · p. 13

8) Delay-Doppler Domain Multiple Access (DDMA):

  • ODDM achieves orthogonality at fine delay-Doppler resolutions with realizable pulses, while combining DD-domain modulation with SDMA or RSMA is proposed for high-mobility channels.ODDM could serve as a delay-Doppler-domain OMA scheme.Source: 8) Delay-Doppler Domain Multiple… · p. 13

H. Technology Outlook and Future Works

  • MA designs share challenges in optimization, channel estimation, and receiver complexity, as well as sensitivity to synchronization and RF impairments.Receiver complexity includes operations such as SIC.Source: 1) Common Challenges in MA designs · p. 13

1) Common Challenges in MA designs:

  • The section groups MA resources into time, frequency, power, space, and signal dimensions, with signal covering message splitting, coding, modulation, interleaving, and spreading.The mapping of schemes to these dimensions is non-exhaustive.Source: 2) MA dimensions · p. 13

2) MA dimensions:

The section maps MA schemes across resource dimensions and presents RSMA as a unifying step toward universal multiple access. It then surveys AI approaches to resource allocation, including traditional machine learning and neural methods.

2) MA dimensions:

  • MA schemes exploit combinations of time, frequency, power, space, and signal dimensions; RSMA basic schemes use space, power, and message splitting, while some variants add time or frequency.Signal includes dimensions such as message split and combiner, coding, modulation, interleaving, and spreading.Source: 2) MA dimensions · p. 13, Table II · p. 14

3) RSMA - A First Step Toward Unification in 6G:

  • RSMA unifies PD-NOMA, SDMA, and physical-layer multicasting as a superset that can reduce to each and access operating points unavailable to them individually.The paper reports RSMA’s advantages across more than 40 applications, including integrated sensing and communications.Source: 3) RSMA - A First Step Toward… · p. 13

4) Toward Universal Multiple Access in Beyond 6G:

  • Universal multiple access (UMA) is proposed as a research vision, not an existing scheme, to unify RSMA with other dimensions and simplify multi-functional 6G resource use.Relationships among code-domain schemes and their interplay with RSMA remain largely unexplored; combining schemes can also increase receiver complexity.Source: 4) Toward Universal Multiple Access in… · pp. 14–15

III. ARTIFICIAL INTELLIGENCE FOR MULTIPLE ACCESS

  • The AI–MA survey covers AI for resource allocation, channel estimation, receiver design, and user-behavior prediction, as well as MA for federated learning, edge intelligence, and over-the-air computation.The paper characterizes this as a comprehensive review of the interplay between AI and MA.Source: §III · p. 15

A. AI-empowered MA Resource Allocation and Optimization

  • Resource-allocation choices depend on the MA scheme and system requirements, while conventional optimization can become computationally burdensome at 6G scale.AI methods can learn allocation from data or the environment without requiring complex mathematical models; the survey groups them into traditional ML, RL, and DL.Source: A. AI-empowered MA Resource Allocation… · p. 16

1) Machine learning:

  • Traditional ML methods support tasks such as spectrum sensing and user grouping, but can be computationally costly on large datasets or struggle with complex hierarchical relationships.The section also describes DL as useful for processing large, intricate datasets in resource-allocation tasks.Source: 1) Machine learning · p. 16, 2) Deep Learning · p. 16

2) Deep Learning:

The section surveys deep-learning architectures and reinforcement learning for wireless resource allocation, then weighs their potential benefits against data, training, and interpretability challenges. It describes neural models that approximate optimization methods and adapt allocation decisions across settings.

2) Deep Learning:

  • A three-layer DNN closely approximates WMMSE for resource allocation while reducing computational time by orders of magnitude.The cited work frames this as “learn to optimize.”Source: 2) Deep Learning · p. 17
  • CNNs reduce parameter counts and support feature extraction, RNNs capture temporal correlations, and GNNs model variable-sized wireless networks for resource allocation.The section also describes transformers, autoencoders, transfer learning, and algorithm-driven methods such as deep unfolding.Source: 2) Deep Learning · pp. 17–18
  • Meta-learning produces RSMA precoders and resource allocations with performance comparable to convex optimization-based approaches and significantly lower running time.The paper notes its relevance to large RSMA networks where WMMSE may require impractical computational time.Source: 2) Deep Learning · p. 18

3) Reinforcement Learning:

  • Deep reinforcement learning supports sequential resource-allocation decisions in dynamic, large-scale networks without explicit transmission-environment models.The paper contrasts it with Q-learning, whose Q-value storage becomes impractical as system complexity and hidden states increase.Source: 3) Reinforcement Learning · pp. 18–19

4) Advantages and Disadvantages:

  • AI resource allocation may offer efficient inference and flexible performance, but depends on representative data, can require substantial training and updates, and may lack interpretability.Biased or shifting channel data can impair generalization, while end-to-end neural approaches can be black boxes.Source: 4) Advantages and Disadvantages · pp. 19–20

B. AI-empowered MA Channel Estimation

The section surveys AI methods for channel estimation and CSI compression, receiver design, and user-behavior prediction. These approaches connect learned channel and user information to MA operation and resource configuration.

B. AI-empowered MA Channel Estimation

  • AI channel-estimation studies include joint pilot design and estimation, prior-aided GM-LAMP, and grant-free activity detection with channel estimation.For FDD massive MIMO, CsiNet and follow-on work use learned compression and feedback, including temporal modeling and quantized feedback –.Source: B. AI-empowered MA Channel Estimation · pp. 20–21

C. AI-empowered MA Receiver Design

  • Model-based receivers depend on accurate assumptions about noise, interference, and propagation, while conventional AI receivers require substantial training data, overhead, and periodic retraining.Model-based deep learning replaces selected computations with compact neural networks; cited NOMA and RSMA studies report better performance than conventional model-based SIC receivers.Source: C. AI-empowered MA Receiver Design · p. 21

D. AI for MA User Behavior Predictions

  • User-behavior prediction links historical user data to future QoS requirements and timely resource configuration, while different MA transmission schemes add design challenges.The surveyed prediction areas are user activity, wireless traffic, content popularity, and user mobility.Source: D. AI for MA User Behavior Predictions · pp. 21–22
  • For activity prediction, LSTM approximates the dynamic relationship between user activity and grouping in NOMA; the cited framework improves reliability and complexity performance [232].The survey also describes activity prediction for OMA cognitive-radio and mobile broadband settings using supervised learning and other ML methods.Source: 1) User Activity Prediction · p. 22

1) User Activity Prediction:

The section reviews ML approaches to predicting user activity, traffic, content popularity, and mobility, then identifies open directions for applying AI to more advanced and universal MA. It also introduces MA designed for AI applications.

1) User Activity Prediction:

  • In NOMA, dynamic and coupled user activity and grouping make activity prediction difficult; an LSTM-based mapping improves reported reliability and complexity performance.The survey also notes ML approaches for activity prediction in OMA cognitive-radio settings.Source: 1) User Activity Prediction · p. 22

2) Wireless Traffic Prediction:

  • For predicted UAV locations, RSMA achieves the lowest total transmit power among TDMA, FDMA, NOMA, and RSMA in the cited comparison.The passage attributes this result to RSMA achieving all points in the capacity region.Source: 2) Wireless Traffic Prediction · p. 22

3) Content Popularity Prediction:

  • Content popularity prediction supports caching popular content near intended users to improve delay and energy performance; surveyed methods include attention for OMA and autoencoders for NOMA,.Source: 3) Content Popularity Prediction · pp. 22–23

4) User Mobility Prediction:

  • Mobility prediction uses movement history and, in some approaches, time-varying graph or physical-environment relationships to characterize user movement –.Source: 4) User Mobility Prediction · p. 23

E. Technology Outlook and Future Works

  • AI for resource allocation, optimization, channel estimation, receivers, and behavior prediction remains limited for advanced schemes such as RSMA and CD-NOMA, while UMA adds optimization complexity across five dimensions.The outlook proposes AI as a complement to communication and information theory for identifying UMA design ingredients and architectures.Source: E. Technology Outlook and Future Works · p. 23

IV. MULTIPLE ACCESS FOR ARTIFICIAL INTELLIGENCE

  • The paper turns from AI applied to MA toward MA tailored to support federated learning, edge intelligence, and over-the-air computation.Source: §IV · p. 23

A. MA for Federated Learning and Edge Intelligence

The section surveys how OMA, spatial and power-domain access, RSMA, and over-the-air computation support wireless federated learning. It contrasts their resource use, trade-offs, and approaches to aggregating model updates.

A. MA for Federated Learning and Edge Intelligence

1) OMA for FL:

  • OMA assigns separate spectrum resources to devices, and its resource needs or latency grow linearly with transmitter count, limiting scalability –.Surveyed work optimizes allocation, scheduling, and related techniques to improve convergence, energy use, or accuracy.Source: 1) OMA for FL · pp. 23–24

2) SDMA for FL:

  • MIMO and spatial-domain methods let devices transmit FL parameters on shared time-frequency resources, with studies optimizing convergence time and communication efficiency –.Source: 2) SDMA for FL · p. 24

3) PD-NOMA for FL:

  • PD-NOMA groups devices and manages within-group interference using SC-SIC; surveyed work jointly optimizes transmission, power, scheduling, computation, or energy and latency –.The passage notes that decoding other users’ interference limits flexibility and that NOMA does not adapt well to imperfect CSI.Source: 3) PD-NOMA for FL · p. 24, 4) RSMA for FL · p. 25

4) RSMA for FL:

  • RSMA can balance signal and interference under imperfect CSI; cited work reports lower latency and power than TDMA and NOMA for clustered FL.In fog radio access networks, splitting messages between edge and cloud decoding trades fronthaul overhead against learning accuracy.Source: 4) RSMA for FL · p. 25

B. Over The Air Computation

  • AirComp aggregates simultaneous device transmissions through waveform superposition, making per-iteration FL transmission delay independent of device count.Unlike conventional MA, it uses interference among cooperating devices to compute a function of their transmissions.Source: B. Over The Air Computation · p. 25

1) Analog AirComp:

  • Analog AirComp uses linear modulation and power control for weighted-sum aggregation, but synchronization and phase offsets can cause aggregation errors; digital AirComp uses coding and modulation to handle phase misalignment.AirComp requires transmission timing alignment; digital receivers decode the aggregate rather than each device’s individual FL parameters.Source: 1) Analog AirComp · pp. 25–26, 2) Digital AirComp · p. 26

2) Digital AirComp:

This section surveys multiple access for integrated sensing and communications (ISAC), digital over-the-air aggregation, and multimodal sensing-aided communications. It reviews how these approaches use wireless resources and identifies unresolved design challenges.

2) Digital AirComp:

  • Digital AirComp designs include quantized and modulated model transmission, non-orthogonal OFDM aggregation, and one-bit gradient majority-voting schemes.The selection of an MA scheme for federated learning depends on participating devices, wireless and computational resources, hardware, model architecture, and available setting information.Source: 2) Digital AirComp · p. 26, C. Technology Outlook and Future Works · p. 26

A. Joint Sensing and Communications

  • ISAC allocates shared time, frequency, space, code, and power resources and designs waveforms for both communication and sensing objectives.Sharing spectrum can provide access to more bands, but requires coordination and joint processing.Source: A. Joint Sensing and Communications · p. 27
  • ISAC based on SDMA can offer a communication-sensing performance trade-off, but may underperform in overloaded scenarios and remains insufficiently studied under imperfect channel estimation.Joint MIMO radar-communication designs use shared waveforms or jointly precoded radar and communication waveforms.Source: 1) SDMA-Assisted ISAC · p. 27

2) PD-NOMA Assisted ISAC:

  • PD-NOMA ISAC can improve the communication-sensing trade-off over OMA, but SIC decoding scales with user count and adds latency and complexity.Fully decoding interference may also lose multiplexing gain with multiple antennas.Source: 2) PD-NOMA Assisted ISAC · p. 27
  • RSMA-assisted DFRC combines radar probing with common and private communication streams, outperforming SDMA and NOMA counterparts while imposing demanding transmitter-side joint optimization.The framework is reported to manage multi-user and inter-beam interference while meeting radar mean-square-error constraints.Source: 3) RSMA-Assisted ISAC · p. 27, Figure 16 · p. 27

B. Multimodal Sensing-aided Communications

  • Large arrays and high-frequency bands increase channel-estimation and beam-training overhead, while blockage sensitivity challenges reliability and latency.Multimodal sensing data have been used for beam selection, tracking, blockage prediction, channel-statistics prediction, and precoding.Source: B. Multimodal Sensing-aided… · p. 28
  • Multimodal sensing can inform user scheduling, handover prediction, and resource allocation using user positions, mobility, and environmental dynamics.For example, well-separated users may share time-frequency resources with less interference.Source: B. Multimodal Sensing-aided… · p. 28, Figure 17 · p. 28
  • Practical multimodal sensing-aided access requires user identification and differentiation, and addressing sensing-communication coverage mismatch through distributed sensing.Initial user-identification approaches augment sensing with sparse channel or beam-power measurements.Source: B. Multimodal Sensing-aided… · p. 29

C. From Multimodal Sensing to Digital Twin-Assisted Communications

The section connects multimodal sensing and digital twins with communication and localization resource decisions, then reviews semantic communication and its performance indicators. It highlights open constraints in digital-twin design, localization, and semantic communication.

C. From Multimodal Sensing to Digital Twin-Assisted Communications

  • Digital twins fuse multimodal sensing with 3D maps and ray tracing to build real-time or near-real-time representations of communication environments.When calibrated to site-specific measurements, these twins can support model pre-training, network-procedure optimization, or predictions for the real system.Source: C. From Multimodal Sensing to Digital… · p. 29
  • Digital-twin-specific multiple access must allocate resources across sensors and modalities to meet fidelity and freshness objectives while scaling across varying task requirements.Different sensor channels may have different QoS requirements conditioned on the final digital-twin construction objectives.Source: C. From Multimodal Sensing to Digital… · p. 29

D. Communication-aided sensing/localization systems

2) PD-NOMA-Assisted Localization:

  • RSMA-assisted cooperative ISAC jointly supports user communications and target localization, and is reported to achieve the best trade-off among the considered baselines.The work formulates communication-centric and localization-centric Pareto optimizations; Fig. 18 illustrates cooperative base stations and common/private message splitting.Source: 3) RSMA-Assisted Localization · p. 30, Figure 18 · p. 30

3) RSMA-Assisted Localization:

  • MA-assisted localization studies commonly assume static users, use the Cramér–Rao bound without practical estimators that attain it, and may incur SIC decoding latency.The section also calls for studying movement, synchronization errors, other MA dimensions, and ultimately universal MA for localization.Source: 4) Technology Outlook and Future Works · p. 30

A. Semantic Communications

  • Semantic communication transmits meaningful information using shared knowledge, and its MA design needs suitable performance indicators because it typically lacks a well-defined metric such as Shannon capacity.The section distinguishes simulation-based KPIs fitted from semantic-rate-versus-SINR results and analysis-based KPIs using a shared probability graph; semantic compression ratio is extracted semantic bits divided by original bits.Source: A. Semantic Communications · pp. 30–31

1) OMA-aided SeCom:

The section surveys OMA, PD-NOMA, SDMA, RSMA, and model division multiple access for semantic communication, then reviews OMA and NOMA for Metaverse applications. It describes the different access mechanisms and their reported trade-offs.

1) OMA-aided SeCom:

  • In OMA-aided semantic communication, channel assignment and semantic-symbol count can be optimized for spectral efficiency, while energy-focused designs include latency constraints and computation resources.OMA is easy to deploy, but its served-user count is limited by wireless resources.Source: 1) OMA-aided SeCom · p. 32

2) PD-NOMA-aided SeCom:

  • PD-NOMA-aided semantic communication can transmit semantic information to multiple users simultaneously and support coexistence between semantic-information and conventional bit-interested users.The reviewed framework also considers joint conventional and semantic transmissions.Source: 2) PD-NOMA-aided SeCom · p. 32

3) SDMA-aided SeCom:

  • SDMA-aided semantic communication uses precoding in multi-antenna systems to transmit semantic information to multiple users through spatial-domain characteristics.Reviewed work also considers multimodal SDMA-aided semantic communication.Source: 3) SDMA-aided SeCom · p. 32

4) RSMA-aided SeCom:

  • RSMA-aided semantic communication sends shared knowledge by multicast and individual data by unicast, reducing common-knowledge resource use relative to NOMA and SDMA.The cited work reports better overall performance than NOMA- and SDMA-aided semantic communication.Source: 4) RSMA-aided SeCom · p. 32

5) MDMA-aided SeCom:

  • Model division multiple access serves users with distinct model-information resources, recovering their information in the model domain and improving performance over FDMA and NOMA in cited work.A point-cloud semantic-communication application and a learning-based DeepMA realization are also described.Source: 5) MDMA-aided SeCom · p. 32

B. Metaverse

  • Metaverse systems exchange sensed physical-world information through edge servers and the cloud, and use communications to update the digital world and transmit simulated information back to devices.The applications described include education, military, healthcare, real estate, and manufacturing.Source: B. Metaverse · pp. 32–33

1) OMA-aided Metaverse:

  • OMA-aided Metaverse work includes blockchain-based ubiquitous access control, learning-based VR systems, and covert communication.Source: 1) OMA-aided Metaverse · p. 33

2) NOMA-aided Metaverse:

  • NOMA-aided Metaverse can serve more physical devices than OMA with limited wireless resources, and a two-step allocation scheme is reported to improve average QoE utility.The scheme combines one-to-many matching for NOMA channel allocation with hedonic coalition formation for service selection.Source: 2) NOMA-aided Metaverse · p. 33

3) SDMA-aided Metaverse:

This section reviews RSMA for Metaverse systems and OMA, PD-NOMA, and RSMA for virtual-reality streaming. It compares access capacity and receiver demands and identifies design challenges for immersive applications.

3) SDMA-aided Metaverse:

  • RSMA-aided Metaverse is described as robust to channel-state-information inaccuracy and user mobility, supporting applications in high-mobility scenarios.The section also contrasts NOMA's greater device support than OMA with additional decoding complexity, and identifies cross-layer and multi-slot design as future directions.Source: 4) RSMA-aided Metaverse · p. 33, 5) Technology Outlook and Future Works · p. 33

C. Virtual Reality

  • VR streaming requires high bit rates and low motion-to-photon delay; delays above 20 milliseconds can cause motion sickness.The section says radio-resource allocation, MA, and network infrastructure must be optimized and integrated.Source: C. Virtual Reality · p. 33

1) OMA-aided VR:

  • OMA gives VR users simple receivers that can support lighter head-mounted displays and longer battery life, but orthogonal resources limit simultaneous unicast users.Reviewed OFDMA work spans multicast, resource allocation, QoE, and fairness objectives.Source: 1) OMA-aided VR · pp. 33–34

2) PD-NOMA-aided VR:

  • NOMA can transmit VR tiles to multiple users on the same radio resource, supporting more users than OMA, but SIC increases receiver complexity as user count grows.User grouping and decoding order affect performance, and SIC can put physical-layer privacy at risk.Source: 2) PD-NOMA-aided VR · pp. 34–35

3) RSMA-aided VR:

  • RSMA manages VR-user interference by decoding part as a common message and treating the remainder as noise; reviewed studies report benefits over SDMA and PD-NOMA.A hierarchical design groups users by overlapping fields of view and splits messages into super-common, group-common, and private streams.Source: 3) RSMA-aided VR · p. 35, Figure 21 · p. 35
  • VR access designs still need low-complexity receivers that preserve privacy and account for head-mounted displays’ limited battery, computation, weight, and possible multiple antennas.Source: 4) Technology Outlook and Future Works · p. 35

4) Technology Outlook and Future Works:

The outlook identifies open design questions for RSMA in VR streaming and surveys how RIS architectures interact with OMA, SDMA, PD-NOMA, and RSMA. The RIS section describes distinct optimization challenges and reported advantages, especially for RIS-assisted RSMA.

4) Technology Outlook and Future Works:

  • VR streaming research must jointly consider multi-timescale cache placement and delivery, UAV mobility, imperfect CSI, correlated or distinct tile requests, and untested RSMA integration with other access schemes.Whether RSMA can meet VR users’ hybrid requirements for 6G eMBB+ remains an open research direction.Source: 4) Technology Outlook and Future Works · pp. 35–36

D. Smart Radio and Reconfigurable Intelligent Surfaces

  • RIS architectures range from diagonal D-RIS to interconnected BD-RIS, whose hybrid and multi-sector modes support reflection and transmission across multiple sides or sectors.Hybrid mode is also called IOS or STAR-RIS.Source: D. Smart Radio and Reconfigurable… · p. 36

1) RIS-assisted OMA:

  • RIS-assisted OMA research spans beamforming, channel modeling and estimation, theoretical limits, and prototypes, with potential to improve spectrum efficiency and coverage without added energy consumption.The surveyed OMA studies initially focus on point-to-point communication using one or more resource blocks.Source: 1) RIS-assisted OMA · p. 36

2) RIS-assisted SDMA:

  • RIS-assisted SDMA jointly optimizes RIS passive and transmitter active beamforming, while multi-user channel estimation must address inter-user interference; multi-user prototyping remains early-stage.A blind beamforming strategy was demonstrated in real-world trials, and another prototype reported improved spectrum efficiency, [424].Source: 2) RIS-assisted SDMA · pp. 36–37

3) RIS-assisted PD-NOMA:

  • In RIS-assisted PD-NOMA, reflection or transmission links can increase channel-strength disparity for SIC, but beamforming also affects user grouping and decoding order; RIS can make TDMA require less power than NOMA.RIS-assisted NOMA has also been studied for physical-layer covertness.Source: 3) RIS-assisted PD-NOMA · pp. 37–38

4) RIS-assisted RSMA:

  • RIS-assisted RSMA studies report advantages over RIS-aided SDMA, NOMA, and OMA across spectrum and energy efficiency, fairness, and outage probability, while BD-RIS variants extend coverage –[444].RSMA’s reported resilience to CSI inaccuracies and user mobility makes RIS-assisted RSMA a proposed way to address channel-estimation limitations in RIS links.Source: 4) RIS-assisted RSMA · p. 38

5) Technology Outlook and Future Works:

The outlook presents RIS as a way to improve multiple-access efficiency, complexity, energy use, and coverage. For massive IoT access, it contrasts grant-based contention with grant-free approaches, including compressed-sensing activity detection whose asymptotic performance is theoretically characterized.

5) Technology Outlook and Future Works:

  • RIS can enhance spectrum efficiency across OMA, SDMA, NOMA, and RSMA through signal-strength improvement, interference mitigation, channel-strength disparity, and RSMA’s combination of access benefits.For SDMA, the described mechanism redirects intended signals and mitigates multi-user interference; for RSMA, it combines OMA, SDMA, and NOMA benefits.Source: 5) Technology Outlook and Future Works · p. 38
  • RIS beamforming can reduce computational and hardware complexity, improve energy efficiency, and extend coverage by redirecting signals around obstacles rather than relying on additional energy-consuming relay infrastructure.With sufficiently many RIS elements, classical RZF can achieve performance comparable to optimized precoding; RIS also enables 1-layer RS performance comparable to HRS/GRS without RIS,.Source: 5) Technology Outlook and Future Works · p. 38

E. Internet-of-Things Networks, Massive Connectivity, and Random Access

  • Massive IoT access must coordinate small packets reliably and with low latency despite collision risk from many devices; the surveyed alternatives are grant-based and grant-free access.Grant-based schemes rely on base-station grants, while grant-free access allows devices to transmit without prior resource requests,.Source: E. Internet-of-Things Networks,… · p. 38

2) Grant-Free Sourced Random Access based on Compressed Sensing:

3) Grant-Free Sourced Random Access based on AI:

This section surveys AI-based pilot design and activity detection for grant-free sourced access, then identifies unsourced-access assumptions and synchronization and activity-correlation challenges. It also places multiple-access evolution within the 3GPP standardization timeline.

3) Grant-Free Sourced Random Access based on AI:

  • AI-based pilot design has improved device detection over Gaussian pilots and Reed-Muller pilots, while deep learning approaches use either data-driven networks or model-driven versions of algorithms such as AMP and group LASSO –.Pilot design remains challenging because the relation between sensing matrices and detection performance is unknown, –.Source: 3) Grant-Free Sourced Random Access… · p. 40

4) Grant-Free Unsourced Random Access:

  • Unsourced random access uses a common codebook and evaluates per-user error probability; its receiver seeks messages, while user identity recovery is left to higher protocol layers –.The surveyed achievable schemes include concatenated coding, ALOHA, and random spreading.Source: 4) Grant-Free Unsourced Random Access · p. 40

5) Technology Outlook and Future Works:

  • IoT access research must address imperfect synchronization among low-cost devices and exploit temporal or spatial correlations in user activity to improve activity detection.The passage also raises questions about collision resolution, compressed-sensing formulations with synchronization errors, and whether sample covariance remains a sufficient statistic.Source: 5) Technology Outlook and Future Works · p. 40

VII. ROADMAP TO 6G STANDARDIZATION

  • The paper traces 3GPP’s progression from TDMA in 2G to CDMA in 3G and (O)FDMA in 4G and 5G, using earlier generations to frame potential 6G access considerations.The stated timeline anticipates a 6G study phase in Release 20 and initial specifications in Release 21, with completion expected between late 2028 and late 2029.Source: §VII · p. 40

A. MA in 3G and 4G

  • In 3G and 4G, access evolved from 3G uplink CDMA toward 4G DFT-S-OFDM, with scheduled packetized transmission and orthogonal frequency division supporting uplink multiplexing.The historical account also describes LTE random access collisions and spatial multiplexing with orthogonalized reference signals.Source: A. MA in 3G and 4G · p. 41

B. MA in 5G and 6G

The paper describes a non-conclusive 5G NR NOMA study and frames 6G access as an open choice among specialized schemes or a unified approach. It also surveys AI/ML use cases for NR and argues for universal multiple access across resource dimensions.

B. MA in 5G and 6G

  • The Release 16 NOMA study across mMTC, URLLC, and eMBB ended without a follow-up work item because the proliferating schemes had no clear winner or single normative scope.A two-step random-access improvement was approved instead of a normative NOMA specification.Source: B. MA in 5G and 6G · p. 42

C. AI/ML in 5G and 6G

  • Release 18’s NR AI/ML study selected CSI enhancement, beam management, and positioning as pilot use cases, with CSI compression using a two-sided model and other listed sub-use cases using single-sided models,.The paper anticipates specification-impact work in Release 19 and notes that mobility enhancement may also be investigated.Source: C. AI/ML in 5G and 6G · p. 42

VIII. CONCLUSIONS

  • The conclusion calls for rethinking access for AI-native, multi-functionality-native networks and presents RSMA as unifying SDMA, PD-NOMA, multicasting, and selected resource dimensions, while universal multiple access remains undeveloped.The proposed next step is to unify RSMA with other dimensions, including code-domain access.Source: §VIII · p. 43
Loading 2401.01433v1…