Rate-Splitting Multiple Access (RSMA)

• RSMA is a multiple access framework that splits each user’s message into common and private components, enabling flexible interference management.
• It employs linear precoding and successive interference cancellation to optimize throughput and energy efficiency in MISO broadcast channels.
• RSMA bridges the gap between SDMA and NOMA by adapting message splitting and power allocation, yielding significant gains in rate, latency, and robustness.

Rate-Splitting Multiple Access (RSMA) is a multiple access framework for downlink multi-antenna systems in which each user message is split into a common part and a private part, all common parts are jointly encoded into a stream decoded by all users, and the private parts are encoded into user-specific streams. In the formulation of downlink RSMA for MISO broadcast channels, RSMA relies on linearly precoded rate-splitting with successive interference cancellation (SIC) to decode part of the interference and treat the remaining part of the interference as noise. This enables it to softly bridge the two extremes of fully decoding interference and treating interference as noise, while containing SDMA and NOMA as special cases (Mao et al., 2017).

1. Intellectual lineage and core idea

A precursor of the modern term appears in the Gaussian multi-access channel literature, where Rate Splitting Multiple Access is described as a code division multiple-access technique that can achieve any base in the multi-access capacity polymatroid without high coding complexity or synchronization among the transmitting users. In that setting, each physical user can be split into virtual users with powers $P_{i,1}=\alpha_iP_i$ and $P_{i,2}=(1-\alpha_i)P_i$, and a single successive-decoding order over at most $2K-1$ codebooks can realize any boundary rate-tuple on the dominant face (Mao et al., 2016).

Modern downlink RSMA repurposes rate-splitting for precoded broadcast channels. Its defining idea is not simply superposition coding, but the deliberate partition of each message into a portion that is commonly decoded and a portion that remains private. This gives the transmitter an adjustable interference-management mechanism: interference can be partly decoded through the common layer and partly left as residual interference in the private layer. The 2017 downlink formulation argues that this flexibility is needed to cope with high throughput, heterogeneity of Quality-of-Service (QoS), and massive connectivity requirements in future multi-antenna wireless networks (Mao et al., 2017).

A recurring misconception is that RSMA is exhausted by the simplest one-common-stream architecture. The broader framework is more general. In CoMP joint transmission, for example, each user message can be split into sub-messages indexed by subsets of users, and streams of different orders can be superposed and decoded through a SIC hierarchy. This suggests that the widely used one-layer design is best understood as the simplest operating point of a larger RSMA design space rather than as the whole framework (Mao et al., 2018).

2. Canonical downlink model

The canonical formulation considers a MISO broadcast channel with one base station equipped with $N_t$ antennas serving $K$ single-antenna users indexed by $\mathcal K=\{1,\dots,K\}$. The transmitted signal is

$$\mathbf x \;=\; \mathbf P_c\,s_c \;+\sum_{k=1}^K\mathbf P_k\,s_k$$

subject to

$$\mathbb E\{\|\mathbf x\|^2\}\;=\;\mathrm{tr}\bigl(\mathbf P_c\mathbf P_c^H+\!\sum_{k=1}^K\mathbf P_k\mathbf P_k^H\bigr)\;\le\;P_{\rm tot}.$$

Here $s_c$ is the common symbol, $s_k$ is the private symbol for user $P_{i,2}=(1-\alpha_i)P_i$0, and user $P_{i,2}=(1-\alpha_i)P_i$1 observes

$P_{i,2}=(1-\alpha_i)P_i$2

Each message $P_{i,2}=(1-\alpha_i)P_i$3 is split as

$P_{i,2}=(1-\alpha_i)P_i$4

with all $P_{i,2}=(1-\alpha_i)P_i$5 jointly encoded into the common stream $P_{i,2}=(1-\alpha_i)P_i$6, and $P_{i,2}=(1-\alpha_i)P_i$7 encoded into $P_{i,2}=(1-\alpha_i)P_i$8 (Mao et al., 2017).

Decoding proceeds in two stages. User $P_{i,2}=(1-\alpha_i)P_i$9 first decodes $2K-1$0 while treating all private streams as noise, performs SIC to remove the common component, and then decodes $2K-1$1 while treating the remaining private streams as noise. The corresponding SINRs are

$2K-1$2

The common-stream rate at user $2K-1$3 is $2K-1$4, and the common stream must satisfy

$2K-1$5

If $2K-1$6 is the portion of the common-rate budget allocated to user $2K-1$7, with $2K-1$8, then the total rate of user $2K-1$9 is

$N_t$0

A standard design objective is weighted sum-rate maximization over $N_t$1, $N_t$2, and $N_t$3, optionally with per-user QoS constraints $N_t$4. This canonical formulation has been reused, with minor changes in notation and power constraints, across later studies on energy efficiency, finite blocklength, mobility-robust design, and integrated sensing-and-communication settings (Mao et al., 2017).

3. Optimization and algorithmic design

The basic weighted sum-rate problem is non-convex in the common and private precoders. A standard solution is the weighted mean-square-error (WMMSE) reformulation. In this approach, receive equalizers and MSE weights are introduced for the common and private streams, the SINR–MSE equivalence is used to convert rate terms into augmented WMSE expressions, and alternating optimization is performed between closed-form MMSE receiver updates and a convex joint update of precoders and common-rate allocations via QCQP. In the canonical formulation, this alternating-optimization procedure converges to a stationary point of the original weighted sum-rate problem (Mao et al., 2017).

Energy-efficiency design leads to a non-convex fractional program. In that setting, a Successive Convex Approximation (SCA) method introduces auxiliary variables for the numerator and denominator of the energy-efficiency ratio, linearizes non-convex product and fractional SINR constraints by first-order Taylor expansions, and solves a convex program iteratively. The resulting SCA algorithm is guaranteed to converge to a stationary point with monotonically increasing energy efficiency, although global optimality is not assured (Mao et al., 2018).

Beyond WMMSE and SCA, alternative algorithmic routes have been proposed. For downlink MIMO RSMA, one line of work approximates the non-smooth $N_t$5 operator in the common-rate term via LogSumExp, reformulates the smoothed sum spectral-efficiency objective as a log-sum of Rayleigh quotients, and interprets the first-order optimality condition as an eigenvector-dependent nonlinear eigenvalue problem. This leads to a generalized power-iteration algorithm whose per-iteration complexity is $N_t$6, and whose leading eigenvector is a local optimal solution of the smoothed problem (Park et al., 2021).

The algorithmic repertoire expands further in specialized scenarios. In DFRC, an ADMM decomposition combines WMMSE on the communication side with Majorization–Minimization on the radar side under per-antenna constraints. In finite-blocklength RSMA, SCA is used to handle dispersion-penalized rate expressions and yields a sequence of convex QCQPs. In mobility-robust massive-MIMO settings, a closed-form power-splitting coefficient $N_t$7 is derived for a lower bound on ergodic sum-rate, reducing the need for full joint non-convex optimization (Xu et al., 2021).

4. Relationship to SDMA and NOMA

The foundational claim of the modern RSMA literature is structural: SDMA and NOMA are recovered as special cases of RSMA. SDMA is obtained when the common-stream precoder is deactivated, $N_t$8, so that each user treats all interference as noise. NOMA, in the SC–SIC sense, is recovered when one user places its entire message in the common stream while the others place their messages in private streams; with suitable power ordering, this becomes classical SC–SIC (Mao et al., 2017).

Scheme	Specialization within RSMA	Interference handling
SDMA	$N_t$9	Treat interference as noise
NOMA (SC–SIC)	Suitable message placement in common/private streams	Fully decode some interference
RSMA	General case	Partially decode and partially treat as noise

This “soft bridging” interpretation has two technical consequences. First, RSMA is not tied to the channel conditions favorable to only one extreme. The original downlink study compares SDMA, NOMA, and RSMA over underloaded and overloaded regimes, diverse channel directions and strengths, and varying CSIT quality, and presents RSMA as a smooth transition between the two classical strategies. Second, the simplest one-layer architecture is intentionally low complexity: it requires no per-user ordering or grouping at the base station, only one SIC per user, and often achieves more than $K$0 of full RS performance at very low complexity (Mao et al., 2017).

Complexity comparisons in later work refine this picture. For receivers, RSMA imposes one SIC stage per user, whereas NOMA may require up to $K$1 SIC stages in the worst case. For the scheduler, 1-layer RS removes the need for user grouping and ordering that characterizes many NOMA designs. At the same time, general multi-layer RSMA can be considerably richer than either SDMA or one-layer RS. In CoMP joint transmission, full RSMA and 1-layer RS both achieve the highest sum-rates, while 1-layer RS preserves the practical advantage of only one SIC per user (Mao et al., 2018).

5. Performance, robustness, and physical-layer evidence

The original downlink numerical results are organized around finite-SNR performance across underloaded and overloaded regimes, perfect and imperfect CSIT, random i.i.d. Rayleigh and deterministic geometric channels, and heterogeneous QoS constraints. The headline findings are that RSMA consistently outperforms both SDMA and NOMA in sum-rate and rate-region; in perfect CSIT with orthogonal users it is approximately identical to SDMA and optimal DPC; in severe CSIT imperfection it retains DoF gains that SDMA loses; and in overloaded or heterogeneous-QoS scenarios it exceeds NOMA and SDMA by up to $K$2–$K$3 in sum-rate (Mao et al., 2017).

Energy efficiency is treated explicitly in later work. Using a power-consumption model with power-amplifier inefficiency and circuit power, RSMA is reported to be more energy-efficient than SDMA and NOMA in a wide range of user deployments with a diversity of channel directions and channel strengths, with representative gains of up to $K$4–$K$5 higher energy efficiency at the same power-amplifier and circuit-power parameters (Mao et al., 2018).

Finite-blocklength studies extend the comparison from asymptotic spectral efficiency to latency-constrained operation. In a two-user $K$6 scenario at $K$7 dB, RSMA requires $K$8 to achieve sum-rate $K$9 bit/s/Hz for $\mathcal K=\{1,\dots,K\}$0, whereas SDMA needs $\mathcal K=\{1,\dots,K\}$1. In overloaded $\mathcal K=\{1,\dots,K\}$2-user, $\mathcal K=\{1,\dots,K\}$3 simulations, to reach $\mathcal K=\{1,\dots,K\}$4 bit/s/Hz SDMA requires $\mathcal K=\{1,\dots,K\}$5, while RSMA requires only $\mathcal K=\{1,\dots,K\}$6. The interpretation in that literature is direct: smaller blocklength implies lower latency, so RSMA can achieve the same transmission rate as SDMA and NOMA with shorter blocklengths, especially in overloaded networks (Xu et al., 2021).

Mobility-robust design studies reach a similar conclusion under stale CSIT. In a $\mathcal K=\{1,\dots,K\}$7, $\mathcal K=\{1,\dots,K\}$8 OFDM system over a 3GPP Urban Macro channel, RSMA with a closed-form power split sustains sum-rates and throughputs up to vehicular speeds of $\mathcal K=\{1,\dots,K\}$9 km/h, whereas SDMA throughput collapses and saturates at moderate SNR and speed; gains of $$\mathbf x \;=\; \mathbf P_c\,s_c \;+\sum_{k=1}^K\mathbf P_k\,s_k$$0–$$\mathbf x \;=\; \mathbf P_c\,s_c \;+\sum_{k=1}^K\mathbf P_k\,s_k$$1 are reported across SNRs of $$\mathbf x \;=\; \mathbf P_c\,s_c \;+\sum_{k=1}^K\mathbf P_k\,s_k$$2–$$\mathbf x \;=\; \mathbf P_c\,s_c \;+\sum_{k=1}^K\mathbf P_k\,s_k$$3 dB (Dizdar et al., 2021).

Physical-layer and link-level evidence has progressively moved RSMA beyond rate expressions. Polar-coded BICM, practical QAM constellations, AMC, and one-stage SIC were incorporated in downlink RSMA PHY design, and link-level evaluations confirmed throughput benefits over SDMA and NOMA; in one two-user imperfect-CSIT scenario at $$\mathbf x \;=\; \mathbf P_c\,s_c \;+\sum_{k=1}^K\mathbf P_k\,s_k$$4 dB, RSMA achieves approximately $$\mathbf x \;=\; \mathbf P_c\,s_c \;+\sum_{k=1}^K\mathbf P_k\,s_k$$5 bps/Hz versus approximately $$\mathbf x \;=\; \mathbf P_c\,s_c \;+\sum_{k=1}^K\mathbf P_k\,s_k$$6 for SDMA and approximately $$\mathbf x \;=\; \mathbf P_c\,s_c \;+\sum_{k=1}^K\mathbf P_k\,s_k$$7 for NOMA (Dizdar et al., 2020). More recent link and system level studies using 5G-NR PDSCH, ray-tracing, QuaDRiGa, and 3GPP TR 38.901 environments report roughly $$\mathbf x \;=\; \mathbf P_c\,s_c \;+\sum_{k=1}^K\mathbf P_k\,s_k$$8 dB BLER gain at $$\mathbf x \;=\; \mathbf P_c\,s_c \;+\sum_{k=1}^K\mathbf P_k\,s_k$$9 BLER relative to SDMA, and large percentile gains in user rate and weakest-user rate, including entries marked “$$\mathbb E\{\|\mathbf x\|^2\}\;=\;\mathrm{tr}\bigl(\mathbf P_c\mathbf P_c^H+\!\sum_{k=1}^K\mathbf P_k\mathbf P_k^H\bigr)\;\le\;P_{\rm tot}.$$0” for $$\mathbb E\{\|\mathbf x\|^2\}\;=\;\mathrm{tr}\bigl(\mathbf P_c\mathbf P_c^H+\!\sum_{k=1}^K\mathbf P_k\mathbf P_k^H\bigr)\;\le\;P_{\rm tot}.$$1 with ZF in both UMi and InH settings (Aditya et al., 13 Feb 2025).

6. Extensions across network architectures and services

RSMA has been specialized to a wide set of architectures in which interference has more structure than in the canonical single-cell downlink. In coordinated multi-point joint transmission, RSMA is optimized under per-base-station power constraints and QoS constraints, and the reported behavior is that SDMA is more suited to little inter-user disparity and large inter-cell disparity, NOMA is more suited to large inter-user disparity and little inter-cell disparity, while RSMA is suited to any deployment; in the three-cell, three-user Wyner setting, full RSMA and 1-layer RS achieve up to $$\mathbb E\{\|\mathbf x\|^2\}\;=\;\mathrm{tr}\bigl(\mathbf P_c\mathbf P_c^H+\!\sum_{k=1}^K\mathbf P_k\mathbf P_k^H\bigr)\;\le\;P_{\rm tot}.$$2 gain over SDMA/NOMA (Mao et al., 2018).

In overloaded cellular IoT with heterogeneous CSIT, two strategies are introduced: Time Partitioning-RSMA and Power Partitioning-RSMA. The high-SNR result is that PP-RSMA achieves the optimum Degrees-of-Freedom in an overloaded MISO BC with heterogeneous CSIT qualities, while finite-SNR analysis shows explicit sum-rate gain over TP-RSMA and baseline schemes, together with robustness to CSIT inaccuracy and flexibility under QoS constraints (Mao et al., 2020).

Integrated sensing and communication has generated several RSMA variants. In multi-antenna DFRC, the common stream fulfills the triple function of managing interference among communication users, managing interference between communication and radar, and beampattern approximation; RSMA-assisted DFRC achieves a better tradeoff between weighted sum-rate and beampattern approximation than SDMA-assisted DFRC and orthogonal radar-communication strategies (Xu et al., 2021). In mono-static ISAC with multiple moving targets, RSMA is used to jointly maximize the max-min fairness rate and minimize the largest eigenvalue of the CRB matrix, and the reported outcome is a strictly larger MMF-versus-CRB trade-off region than SDMA (Chen et al., 2023).

Other extensions follow the same principle of inserting a partially decoded layer where conventional designs only switch between orthogonality and treating interference as noise. In multi-carrier joint communications and jamming, 1-layer RSMA with OFDM and artificial noise achieves up to $$\mathbb E\{\|\mathbf x\|^2\}\;=\;\mathrm{tr}\bigl(\mathbf P_c\mathbf P_c^H+\!\sum_{k=1}^K\mathbf P_k\mathbf P_k^H\bigr)\;\le\;P_{\rm tot}.$$3 sum-rate gain over SDMA overall, and still yields a $$\mathbb E\{\|\mathbf x\|^2\}\;=\;\mathrm{tr}\bigl(\mathbf P_c\mathbf P_c^H+\!\sum_{k=1}^K\mathbf P_k\mathbf P_k^H\bigr)\;\le\;P_{\rm tot}.$$4–$$\mathbb E\{\|\mathbf x\|^2\}\;=\;\mathrm{tr}\bigl(\mathbf P_c\mathbf P_c^H+\!\sum_{k=1}^K\mathbf P_k\mathbf P_k^H\bigr)\;\le\;P_{\rm tot}.$$5 gain under realistic frequency-selective fading generated by QuaDRiGa’s 3GPP urban-macro channel (Dizdar et al., 2021). In integrated satellite-terrestrial networks, a super-common message is placed above the usual common/private layers so that satellite users and cellular users can decode part of the inter-network interference; in low-Earth orbit conditions this coordinated RSMA framework gains up to $$\mathbb E\{\|\mathbf x\|^2\}\;=\;\mathrm{tr}\bigl(\mathbf P_c\mathbf P_c^H+\!\sum_{k=1}^K\mathbf P_k\mathbf P_k^H\bigr)\;\le\;P_{\rm tot}.$$6 over RSMA without the super-common message, while the optimal super-common power tends to zero as altitude increases and inter-network interference weakens (Lee et al., 2023).

RSMA has also been adapted to cognitive-radio MEC, multigroup multicast cellular and satellite systems, and cache-aided C-RAN. In the cognitive-radio MEC uplink considered in the cited work, dynamic rate splitting protects the primary user while maximizing the secondary user rate and yields a higher successful computation probability than conventional NOMA (Liu et al., 2022). In multigroup multicast, RSMA combined with finite-length polar coding, practical QAM, AMC, and SIC yields up to $$\mathbb E\{\|\mathbf x\|^2\}\;=\;\mathrm{tr}\bigl(\mathbf P_c\mathbf P_c^H+\!\sum_{k=1}^K\mathbf P_k\mathbf P_k^H\bigr)\;\le\;P_{\rm tot}.$$7 dB gain in terrestrial cellular and $$\mathbb E\{\|\mathbf x\|^2\}\;=\;\mathrm{tr}\bigl(\mathbf P_c\mathbf P_c^H+\!\sum_{k=1}^K\mathbf P_k\mathbf P_k^H\bigr)\;\le\;P_{\rm tot}.$$8 dB gain in multibeam satellite MMF throughput relative to SDMA (Yin et al., 2021). In cache-aided C-RAN with statistical CSIT, RSMA combined with heuristic clustering, SAA, and WMMSE provides explicit max-min rate gain over trivial interference-processing methods (Reifert et al., 2021).

A more recent multi-carrier development changes the splitter itself. Instead of a fixed common/private ratio, a channel-dependent splitter replicates private symbols that would otherwise traverse deep-faded subcarriers into the common stream. In the reported evaluations, the distributed version with RCI precoding achieves up to an order-of-magnitude lower BER at $$\mathbb E\{\|\mathbf x\|^2\}\;=\;\mathrm{tr}\bigl(\mathbf P_c\mathbf P_c^H+\!\sum_{k=1}^K\mathbf P_k\mathbf P_k^H\bigr)\;\le\;P_{\rm tot}.$$9 dB SNR than conventional RSMA and MU-MIMO, moving from $s_c$0 to $s_c$1 subcarriers yields a further $s_c$2–$s_c$3 rate boost, and average packet delay is reduced by up to $s_c$4 relative to a HARQ-based RSMA retransmission scheme at $s_c$5 dB (Ali et al., 16 Apr 2025).

7. Prototyping and 6G outlook

RSMA has increasingly been positioned as a candidate 6G physical-layer technique. Recent work characterizes it as a powerful and versatile physical layer multiple access technique that generalizes and has better interference management capabilities than 5G-based SDMA, and presents two independent SDR prototypes, one at Imperial College London and one at VIAVI. Both are based on $s_c$6 MHz OFDM, two base-station antennas, and SIC receivers, but differ in waveform stack, coding, and precoder design (Aditya et al., 13 Feb 2025).

The prototype results reinforce the simulation literature. In eMBB unicast, RSMA achieves higher sum throughput than SDMA and NOMA at any fairness target in the Imperial testbed. In an overloaded $s_c$7 VIAVI scenario, SDMA with scheduling can only serve two users per slot, whereas RSMA serves all users simultaneously by placing two users’ messages wholly in the common stream, yielding a higher and constant min-throughput over time. In an ISAC prototype with two users and one radar target, RSMA strictly enlarges the measured radar-SNR versus sum-rate envelope over SDMA, with the maximum gains in the high-interference, high-overlap scenario (Aditya et al., 13 Feb 2025).

Standardization remains prospective rather than settled. The cited 2025 overview states that 6G work in 3GPP is slated to begin in Release 20, that RSMA was proposed to 3GPP RAN in 2023 but was not studied in 5G, and that ETSI ISG MAT explicitly lists RSMA among candidate downlink methods for improving spectral efficiency, fairness, and latency in 6G (Aditya et al., 13 Feb 2025). A plausible implication is that the main open questions are no longer limited to rate optimality. They increasingly concern scaling to larger antenna arrays and user counts, receiver simplification through non-SIC or partial-SIC designs, hybrid beamforming at mmWave and THz, and cross-layer interaction with scheduling, HARQ, network slicing, and edge computing.