PaperlayerView exact revision on arXiv

Source-linked AI summary

Rate-Splitting Multiple Access for Downlink Communication Systems: Bridging, Generalizing and Outperforming SDMA and NOMA

Yijie Mao, Bruno Clerckx, Victor O. K. Li

arXiv:1710.11018v4cs.IT

TL;DR

SDMA and NOMA rely on the opposing extremes of treating interference as noise or fully decoding it, creating limitations under heterogeneous loads, deployments, and imperfect CSIT. The paper proposes RSMA, which partially decodes interference and partially treats it as noise, and reports robust performance gains over both schemes with reduced scheduler and receiver complexity.

  • Problem

    SDMA and NOMA fundamentally rely on fully treating interference as noise or fully decoding it, limiting robustness across network loads, user deployments, and imperfect CSIT.

  • Method

    The paper proposes RSMA, a linearly precoded rate-splitting framework that transmits common symbols for different user groups alongside private symbols and uses SIC to partially decode interference.

  • Results

    RSMA softly bridges and outperforms SDMA and NOMA across user deployments, CSIT inaccuracies, and network loads, while 1-layer and 2-layer forms maintain robust performance with lower scheduler and receiver complexity.

  • Takeaways & Limitations

    A simple one-layer RS requires no user ordering, grouping, or dynamic switching and uses a single SIC layer while still significantly outperforming NOMA.

Abstract

from arXiv · show

Space-Division Multiple Access (SDMA) utilizes linear precoding to separate users in the spatial domain and relies on fully treating any residual multi-user interference as noise. Non-Orthogonal Multiple Access (NOMA) uses linearly precoded superposition coding with successive interference cancellation (SIC) and relies on user grouping and ordering to enforce some users to fully decode and cancel interference created by other users. In this paper, we argue that to efficiently cope with the high throughput, heterogeneity of Quality-of-Service (QoS), and massive connectivity requirements of future multi-antenna wireless networks, multiple access design needs to depart from SDMA and NOMA. We develop a novel multiple access framework, called Rate-Splitting Multiple Access (RSMA). RSMA is a more general and powerful multiple access for downlink multi-antenna systems that contains SDMA and NOMA as special cases. RSMA relies on linearly precoded rate-splitting with SIC to decode part of the interference and treat the remaining part of the interference as noise. This capability of RSMA to partially decode interference and partially treat interference as noise enables to softly bridge the two extremes of fully decoding interference and treating interference as noise, and provide room for rate and QoS enhancements, and complexity reduction. The three multiple access schemes are compared and extensive numerical results show that RSMA provides a smooth transition between SDMA and NOMA and outperforms them both in a wide range of network loads (underloaded and overloaded regimes) and user deployments (with a diversity of channel directions, channel strengths and qualities of Channel State Information at the Transmitter). Moreover, RSMA provides rate and QoS enhancements over NOMA at a lower computational complexity for the transmit scheduler and the receivers (number of SIC layers).

I. INTRODUCTION

Future downlink networks must support massive connectivity, heterogeneous QoS and CSIT conditions across underloaded and overloaded regimes. The paper proposes RSMA to address limitations of SDMA and NOMA by partially decoding interference while treating the remainder as noise.

  • I. INTRODUCTION: 5G and beyond networks must serve heterogeneous devices with differing capabilities, deployments, QoS demands and CSIT quality across varying network loads.
  • I. INTRODUCTION: The paper introduces RSMA as a multiple-access framework intended to retain SDMA and NOMA benefits while addressing their limitations.
  • A. SDMA and NOMA: The Extremes: SDMA performs well in underloaded settings but becomes vulnerable to overloaded networks, channel non-orthogonality, user-pairing requirements and imperfect CSIT.
  • A. SDMA and NOMA: The Extremes: Multi-antenna NOMA can handle overloaded deployments with aligned channels and diverse strengths, but suffers from inefficient spatial-dimension use, scheduling complexity and imperfect-CSIT limitations.
  • A. SDMA and NOMA: The Extremes: NOMA and SDMA rely on opposite interference strategies: fully decoding interference or fully treating residual interference as noise.
  • B. RSMA: Bridging the Extremes: RSMA uses linearly precoded rate-splitting and SIC to partially decode interference and partially treat it as noise, thereby bridging SDMA and NOMA.

II. SYSTEM MODEL

The system models a MISO broadcast channel in which a multi-antenna base station linearly precodes users’ streams under power, rate, and QoS constraints. It formulates SDMA and NOMA baselines and evaluates their weighted-sum-rate optimization using WMMSE-based methods.

  • System model: The base station with Nt antennas serves K single-antenna users over a MISO broadcast channel under a total transmit-power constraint.The received signal includes each user’s channel and additive white Gaussian noise; perfect CSIT is assumed in the basic model.
  • Optimization objective: Beamforming is designed to maximize users’ weighted sum rate subject to the base station power constraint and individual QoS constraints.The weights prioritize users, while Rth_k represents an individual rate threshold.
  • SDMA baseline: SDMA independently encodes K user messages into K linearly precoded streams, with each user decoding only its desired stream and treating interference as noise.The transmit signal is x = Ps, with E{ssH} = I and tr(PPH) ≤ Pt.
  • Solution approach: WMMSE reformulates the WSR problems and uses alternating optimization, while rate regions are approximated over rate-weight vectors and decoding orders.For SDMA, the resulting rate-region points are enclosed by their convex hull; SC–SIC similarly evaluates orders separately.
  • NOMA baseline: NOMA uses superposition coding and successive interference cancellation, with SC–SIC decoding users according to an optimized order.For a fixed order, earlier users’ messages are decoded successively at later users, and all possible orders may be considered for WSR maximization.

2) SC–SIC per group:

The section describes grouped NOMA and rate-splitting transmission structures. Rate-splitting uses common streams carrying parts of private messages, allowing partial interference decoding and continuous bridging between linear precoding and SIC extremes.

  • 2) SC–SIC per group:: SC–SIC per group partitions users into disjoint groups and applies an independently ordered SIC process within each group.Inter-group signals remain interference when decoding messages inside a group, and WSR optimization considers grouping methods and decoding orders.
  • A. Two-user example: In the two-user RS structure, each message is split into common and private parts, with the common parts jointly encoded into s12 and private parts encoded into s1 and s2.The streams are linearly precoded and superposed under the total transmit-power constraint.
  • A. Two-user example: Both users decode s12 while treating private-stream interference as noise, then subtract it and decode their own private stream.Thus, each user decodes part of the other user’s interfering message before treating the remaining private interference as noise.
  • Rate allocation: Common-message rate portions are allocated among intended users to form total achievable user rates and optimize weighted sum rate.The two-user rate region is obtained by optimizing the common-rate vector over rate weights using WMMSE.
  • A. Two-user example: RS reduces to MU–LP when no power is allocated to the common stream and to SC–SIC when one user fully decodes the other’s message.Encoding both users’ messages into the common stream yields a physical-layer multicasting strategy.
  • Rate allocation: The RS common message carries parts of private messages rather than public information intended as a whole for every user.This distinguishes RS from a conventional unicast-multicast system.
  • Rate allocation: OMA is a subset of MU–LP obtained by allocating power exclusively to one private stream.This places single-user beamforming within the linear-precoding baseline family.

C. Generalized rate-splitting

Generalized RS assigns a stream to every user subset, carrying the corresponding users’ message portions while other users treat that stream as noise. Users decode intended streams through SIC across stream orders, and all decoding orders contribute to the achievable rate region.

  • C. Generalized rate-splitting: For every user subset A, the base station transmits a stream sA decoded by users in A and treated as noise by users outside A.The stream carries message portions from all users in A.
  • C. Generalized rate-splitting: The generalized framework contains one K-order common stream, K private 1-order streams, and distinct intermediate-order streams for user subsets.For K = 3, the 3-order stream is s123 and the 2-order vector is [s12, s13, s23]T.
  • C. Generalized rate-splitting: Each user decodes its intended streams by SIC, starting with the K-order stream and proceeding downward to the 1-order stream.Within each order, a decoding permutation determines which streams are decoded first.
  • C. Generalized rate-splitting: A shared stream’s achievable rate is limited by the users in its intended subset, and its common rate is divided among those users.The resulting user rate combines allocated common-rate portions with the private-stream rate.
  • C. Generalized rate-splitting: For fixed weights and decoding orders, the generalized WSR problem can be solved with WMMSE relationships established for all data streams.The rate region is obtained by evaluating rate weights for each order and taking the convex hull of the union over orders.

D. Structured and low-complexity rate-splitting

Generalized RS offers rate and QoS enhancements through additional SIC layers, while 1-layer RS and 2-layer HRS reduce implementation complexity to one and two SIC layers, respectively.

  • Generalized RS provides more room for rate and QoS enhancements at the cost of more receiver SIC layers.
  • 1-layer RS and 2-layer HRS are introduced as low-complexity strategies for K users.They require one and two SIC layers at each receiver, respectively.

1) 1-layer RS:

1-layer RS uses one common stream decoded by all users followed by private streams, while 2-layer HRS adds group-common streams and two-stage SIC.

  • 1) 1-layer RS:: 1-layer RS transmits one K-order common stream and K private streams using linear precoding.Users decode the common stream first, subtract it, then decode their private streams while treating other private streams as noise.
  • 1) 1-layer RS:: 1-layer RS requires no transmitter-side user ordering or grouping because all users decode the common message with one SIC layer.
  • 1) 1-layer RS:: 1-layer RS is a sub-scheme of generalized RS and a super-scheme of MU–LP when common-message power is set to zero.
  • 2) 2-layer HRS:: 2-layer HRS adds one group-common stream for each user group alongside the all-user common stream and private streams.In the K = 4, G = 2 example, groups are K1 = {1, 2} and K2 = {3, 4}.
  • 2) 2-layer HRS:: 2-layer HRS requires two SIC layers per user, compared with |Kg|−1 layers for SC–SIC per group, and avoids user ordering.Streams with higher order are decoded before streams with lower order.

E. Encompassing existing NOMA and SDMA

Generalized RS contains SDMA and NOMA as special cases and balances interference management flexibility with implementation complexity.

  • RSMA partially decodes interference and partially treats remaining interference as noise through common and private messages.
  • Appropriate stream, message, and precoder choices reduce the proposed RS scheme to multi-antenna NOMA, making NOMA a particular RS case.
  • SDMA and NOMA are special instances of generalized RS, while 1-layer RS and 2-layer HRS provide intermediate architectures.
  • Setting all higher-order stream powers to zero reduces generalized RS to SDMA, while distinct stream orders yield SC–SIC.
  • RSMA balances performance and complexity across channel gain differences and channel angles, with multilayer RS offering greater flexibility.
  • The WSR optimization is transformed into a non-convex WMMSE problem solved by alternating optimization over equalizers, weights, common-rate variables, and precoders.With fixed equalizers and weights, the remaining precoder-rate subproblem is a convex QCQP.

V. NUMERICAL RESULTS

The numerical evaluation compares SDMA, NOMA, and RSMA across network loads and diverse user deployments.

  • The evaluation covers underloaded and overloaded regimes, diverse channel directions and strengths, and varying CSIT quality.It reports rate-region results for two users and WSR comparisons for three, four, and ten users.

A. Underloaded two-user deployment with perfect CSIT

The section defines the rate-region comparison procedure for two users under perfect CSIT and describes initialization choices for the non-convex WSR optimizations.

  • For K = 2, rate-region boundaries are generated by varying user weights while setting individual rate constraints to zero.User-1 has fixed weight u1 = 1, and user-2 uses the specified logarithmic weight grid.
  • DPC provides the capacity-region benchmark in the perfect-CSIT scenario.The DPC region is generated using the algorithm in [46].
  • WSR non-convexity makes precoder initialization important to the final result.The described initialization uses MRT with SVD for RS and specified SVD/MRT choices for SC–SIC.
  • The evaluated setup includes two-user rate-region comparisons with transmit-antenna and channel-variance settings represented in the figures.The supplied figure passages include underloaded two-user comparisons and parameter cases with Nt = 4 and differing σ2 values.

1) Random channel realizations:

Random channel realizations show that strategy performance depends on antenna surplus, channel-strength disparity, and whether users’ channels are degraded or non-degraded.

  • Random channel realizations:: When Nt exceeds the number of users, MU–LP performs well because user precoders tend to become more orthogonal.The comparison considers two or four transmit antennas serving two single-antenna users over random channel realizations.
  • Random channel realizations:: Equal average channel strengths make SC–SIC perform poorly, whereas asymmetric strengths favor SC–SIC when channels are closely aligned.SC–SIC is motivated by exploiting channel-strength differences among users.
  • Random channel realizations:: SC–SIC often loses performance in the general non-degraded MISO-BC.This limitation follows from applying a strategy motivated by degraded broadcast channels to a non-degraded multi-antenna channel.
  • Random channel realizations:: With a 5 dB channel-strength difference and Nt = 2, RS gains relative to MU–LP increase while its gap to SC–SIC decreases.The same passage reports that RS still achieves a larger rate region than both schemes and is closer to the DPC region.

2) Specific channel realizations:

Specific channel realizations vary user-channel angle, strength, antenna count, and SNR to compare RS with SC–SIC, MU–LP, and DPC.

  • 2) Specific channel realizations:: γ controls user-2’s relative channel strength, while θ controls the angle between the two user channels.γ = 1 represents equal strengths, γ = 0.3 represents a 5 dB difference, and θ ranges from aligned to orthogonal channels.
  • 2) Specific channel realizations:: When channels are sufficiently aligned, SC–SIC is more suitable; when they are sufficiently orthogonal, MU–LP is more suitable.The comparison identifies different favorable regimes for the two strategies based on channel geometry.
  • 2) Specific channel realizations:: For γ = 1 and Nt = 4, RS achieves a rate region equal to or larger than those of SC–SIC and MU–LP.RS has a clear improvement when channels nearly coincide, while MU–LP improves as channels become more orthogonal.
  • 2) Specific channel realizations:: For γ = 1 and Nt = 2, RS outperforms MU–LP and SC–SIC for all investigated channel angles.The RS–MU–LP gap enlarges relative to Nt = 4 because fewer antennas make orthogonal precoder design more difficult.
  • 2) Specific channel realizations:: With a 5 dB strength difference and Nt = 4, RS and SC–SIC lie much closer to DPC than in the equal-strength setting.SC–SIC and MU–LP can each outperform the other over parts of the rate region.
  • 2) Specific channel realizations:: RS achieves explicit gains over SC–SIC in most investigated scenarios despite the same one-layer SIC receiver complexity.The comparison reports RS as suited to different channel angles and channel-strength differences.
  • 2) Specific channel realizations:: As SNR increases, rate-region gaps widen and RS exhibits further performance benefits over MU–LP and SC–SIC.The section reports RS outperforming both alternatives in all investigated scenarios.

B. Underloaded two-user deployment with imperfect CSIT

Under imperfect CSIT, averaged rate regions are compared across channel angles and strengths; residual interference particularly affects MU–LP, while RS remains competitive across settings.

  • B. Underloaded two-user deployment with imperfect CSIT: The imperfect-CSIT model assumes perfect user channel estimation but imperfect instantaneous channel estimates at the base station.Estimated channels are perturbed by user-specific Gaussian estimation errors, and rates are averaged over 1000 error samples.
  • B. Underloaded two-user deployment with imperfect CSIT: When channels become sufficiently orthogonal, RS and MU–LP have almost identical rate regions, while aligned channels with sufficient strength disparity favor SC–SIC.These patterns hold for the γ = 1 and γ = 0.3 comparisons.
  • B. Underloaded two-user deployment with imperfect CSIT: Imperfect CSIT increases the RS–MU–LP rate-region gap because distorted interference nulling leaves residual interference at the receiver.The comparison is made against the corresponding perfect-CSIT results.
  • B. Underloaded two-user deployment with imperfect CSIT: SC–SIC is less sensitive to CSIT inaccuracy than MU–LP, but RS retains an evident rate-region advantage over SC–SIC.RS dynamically adjusts how much interference both users decode through its common stream.
  • B. Underloaded two-user deployment with imperfect CSIT: RS achieves equal or better performance than MU–LP and SC–SIC across the simulated channels, antennas, deployments, and CSIT inaccuracies.Additional results are reported for varied SNR, Nt, and γ.

C. Underloaded three-user deployment with perfect CSIT

The underloaded three-user evaluation compares RS, 1-layer RS, SC–SIC, MU–LP, and related strategies using weighted sum rate under perfect CSIT. RS consistently outperforms MU–LP and SC–SIC, while decoding order and channel deployment affect NOMA performance.

  • Underloaded three-user comparison: RS always outperforms MU–LP and SC–SIC across the evaluated underloaded three-user scenarios and SNRs.The comparisons use weighted sum rate with different user weights and channel settings.
  • Underloaded three-user comparison: SC–SIC performs poorly in underloaded deployments because one user decodes all messages, reducing its sum DoF to 1.This sacrifices the available spatial multiplexing gains when Nt > K.
  • Underloaded three-user comparison: The RS weighted-sum-rate improvement is more explicit for u = [0.2, 0.3, 0.5] than for u = [0.4, 0.3, 0.3].The paper links this difference to improved system throughput and user fairness.
  • Decoding-order sensitivity: SC–SIC decoding order must be optimized jointly with the precoder, with six possible orders in the three-user case.The evaluated orders enumerate all permutations of the three user streams.
  • Decoding-order sensitivity: Order 3 is optimal in the reported deployment, decoding user-1 first despite user-3 having the weakest channel gain.The paper attributes this result to user-1 receiving the smallest weight, u1 = 0.2.
  • Robustness and complexity: 1-layer RS achieves equal or better performance than SC–SIC and MU–LP in most reported perfect-CSIT figures and all reported imperfect-CSIT figures.The simplified scheme also retains low scheduler and receiver complexity.
  • Overloaded comparison: In the overloaded three-user comparison, RS shows a clear weighted-sum-rate gain over SC–SIC, SC–SIC per group, and MU–LP.The setting uses two transmit antennas, three users, QoS constraints, and perfect CSIT.

2) Single transmit antenna deployment:

The single-antenna and more heavily overloaded evaluations examine whether simplified RS schemes can retain performance while reducing SIC complexity. 1-layer RS is close to SC–SIC in the SISO case, while RS variants outperform competing strategies in the overloaded multi-user settings.

  • Single transmit antenna deployment: 1-layer RS requires one SIC layer for all users, compared with two SIC layers for SC–SIC in the three-user SISO broadcast channel.The paper presents this as a substantial receiver-complexity reduction.
  • Single transmit antenna deployment: 1-layer RS achieves very close weighted sum rate to SC–SIC in the three-user SISO broadcast channel.The comparison averages performance over 10 random channel realizations.
  • Four-user overloaded deployment: In the four-user overloaded deployment, 2-layer HRS and 1-layer RS outperform the other compared schemes in both reported figures.The two RS schemes jointly mitigate inter-group and intra-group interference using one common message in these deployments.
  • Four-user overloaded deployment: 2-layer HRS, 1-layer RS, and 1-layer RS per group achieve equal or better performance than SC–SIC per group and MU–LP across the reported four-user channel conditions.These additional results use perfect CSIT and no channel-gain difference.
  • Ten-user overloaded deployment: 1-layer RS achieves DoF 2, whereas MU–LP and SC–SIC achieve DoF 1 in the extremely overloaded ten-user setting with QoS constraints.The two transmit antennas limit the maximum DoF of the deployment to 2.
  • Ten-user overloaded deployment: RS uses a common message to pack traffic from eight users while two private streams serve two users in the ten-user example.The paper connects this configuration to massive IoT and MTC services, where many devices decode only the common message.
  • Ten-user overloaded deployment: Increasing rate thresholds causes MU–LP's DoF to drop from 2 to 1 in the overloaded ten-user evaluation.The paper reports that RS maintains a weighted-sum-rate advantage in the extremely overloaded scenario.

APPENDIX

Across rate-region and weighted-sum-rate evaluations, RS generally matches or outperforms MU–LP and SC–SIC, with gains becoming more apparent as SNR increases. The comparisons also examine channel-strength differences, antenna counts, CSIT quality, and overloaded deployments.

  • Two-user rate regions: As SNR increases, the rate-region gaps among RS, SC–SIC, and MU–LP grow, with RS improvement becoming more obvious.In one comparison, RS encompasses the convex hull of the SC–SIC and MU–LP rate regions.
  • Two-user rate regions: When γ = 1 and Nt = 2, the three rate regions are very close at 10 dB, whereas RS explicitly improves over MU–LP and SC–SIC at 20 dB.These comparisons use imperfect CSIT.
  • Two-user rate regions: With γ = 0.3 and imperfect CSIT, the rate-region gap between RS and SC–SIC decreases, while SC–SIC is less sensitive to CSIT inaccuracy than MU–LP.The comparison covers SNR values of 10 dB and 20 dB.
  • Underloaded three-user deployment: In all evaluated underloaded three-user cases, RS achieves weighted sum rate equal to or better than MU–LP and SC–SIC.The cases vary user weights, channel-strength factors, and SNR under perfect CSIT.
  • Underloaded three-user deployment: RS becomes closer to SC–SIC as channel-gain differences increase, while RS and MU–LP are almost identical for u = [0.4, 0.3, 0.3].In sufficiently favorable channel directions, RS and MU–LP overlap with the optimal WSR achieved by DPC.
  • Overloaded three-user deployment: In overloaded three-user cases, RS has a clear WSR gain over SC–SIC, SC–SIC per group, and MU–LP; 1-layer RS outperforms them in most figures with reduced complexity.In some deployments, 1-layer RS achieves the same WSR as RS.
  • Average rate regions: Average-rate-region experiments over 10 random channel realizations identify 1-layer RS as an attractive alternative to SC–SIC.The compared settings include equal and unequal user error variances.

E. Underloaded three-user deployment with imperfect CSIT

The imperfect-CSIT experiments average rates over channel-error realizations while designing precoders from estimated channels. RS retains strong WSR performance across underloaded and overloaded settings, and 1-layer RS offers lower-complexity performance with robustness to CSIT errors.

  • Imperfect-CSIT model: Precoding uses estimated channels, and each rate-region point averages rates over 1000 generated channel-error realizations per user.The channel realization is modeled as the estimated channel plus an error term.
  • Underloaded three-user deployment: Under imperfect CSIT, the WSR gap between RS and MU–LP increases, while the gap between RS and 1-layer RS decreases.This compares the imperfect-CSIT results with perfect-CSIT simulations.
  • Underloaded three-user deployment: 1-layer RS achieves equal or better WSR than SC–SIC, SC–SIC per group, and MU–LP in all evaluated imperfect-CSIT figures.The authors characterize all RS forms as robust to imperfect CSIT.
  • Overloaded three-user deployment: In overloaded three-user imperfect-CSIT experiments, the WSR gaps between RS and SC–SIC per group and MU–LP increase dramatically, while the gap to SC–SIC decreases.The experiments use Nt = 2 and rth = [0.02, 0.08, 0.19, 0.3, 0.4, 0.4, 0.4] bit/s/Hz.
  • Overloaded three-user deployment: 1-layer RS per group always achieves equal or better WSR than SC–SIC per group and supports partial interference decoding within each group.It is described as more general than SC–SIC per group.
  • Overloaded deployments: At 0 dB or 5 dB with zero rate thresholds, MU–LP approaches RS because two transmit antennas can deliver two interference-free streams, whereas SC–SIC remains limited to DoF 1.This comparison concerns an overloaded deployment.
  • Overloaded deployments: In an extremely overloaded setting with diverse channel strengths, the WSR gap between RS and SC–SIC remains large, while 1-layer RS performs well with low scheduler and receiver complexity.SC–SIC is reported to achieve DoF 1 in this setting.
Loading 1710.11018v4…