PaperlayerView exact revision on arXiv

Source-linked AI summary

Multi-mode Transmission for the MIMO Broadcast Channel with Imperfect Channel State Information

Jun Zhang, Marios Kountouris, Jeffrey G. Andrews, Robert W. Heath Jr

arXiv:0903.5108v2cs.ITmath.IT

TL;DR

Imperfect CSIT makes MIMO broadcast-channel transmission vulnerable to residual inter-user interference. The paper adapts the number of active users to balance the relevant gains and interference, and reports improved spectral efficiency over SU-MIMO at medium SNR alongside throughput close to instantaneous-CSI feedback under a feedback constraint.

  • Problem

    Imperfect CSIT causes residual inter-user interference in the MIMO broadcast channel.

    Source: §I · p. 2

  • Method

    The proposed multi-mode strategy adapts the number of active users based on average achievable rate, accounting for transmit array gain, spatial multiplexing gain, and residual inter-user interference.

    Source: §VI · p. 17

  • Results

    Multi-mode transmission improves spectral efficiency over SU-MIMO at medium SNR and provides throughput close to instantaneous-CSI feedback under a total feedback constraint.

    Source: §VI · p. 17

  • Takeaways & Limitations

    Combined with round-robin scheduling, the strategy serves a large number of users while reducing feedback requirements.

    Source: §VI · p. 17

Abstract

from arXiv · show

This paper proposes an adaptive multi-mode transmission strategy to improve the spectral efficiency achieved in the multiple-input multiple-output (MIMO) broadcast channel with delayed and quantized channel state information. The adaptive strategy adjusts the number of active users, denoted as the transmission mode, to balance transmit array gain, spatial division multiplexing gain, and residual inter-user interference. Accurate closed-form approximations are derived for the achievable rates for different modes, which help identify the active mode that maximizes the average sum throughput for given feedback delay and channel quantization error. The proposed transmission strategy is combined with round-robin scheduling, and is shown to provide throughput gain over single-user MIMO at moderate signal-to-noise ratio. It only requires feedback of instantaneous channel state information from a small number of users. With a feedback load constraint, the proposed algorithm provides performance close to that achieved by opportunistic scheduling with instantaneous feedback from a large number of users.

I. INTRODUCTION

MU-MIMO's spatial-gain promise is limited by imperfect CSIT and residual interference; the introduction reviews mode-switching work whose choices or assumptions leave active-user trade-offs unresolved.

I. INTRODUCTION

  • With single-antenna mobiles, MU-MIMO's spatial division multiplexing gain is proportional to the BS antenna count, and sum throughput grows linearly with N_t at high SNR.Source: §I · pp. 1–2
  • CSIT is needed to separate users' spatial channels, but it is difficult to obtain perfectly, and its imperfections cause residual inter-user interference that degrades MIMO-BC performance.Source: §I · p. 2
  • Limited feedback supplies partial CSIT, while obtaining full spatial division multiplexing gain requires carefully designed feedback whose rate grows linearly with SNR in dB and transmit-antenna count.Source: §I · p. 2

A. Related Work

  • Prior mode-switching algorithms mostly rely on instantaneous CSIT, and no explicit switching point tied to key system parameters had been obtained.Source: A. Related Work · p. 3
  • In, asymptotic analysis optimized active-user count to maximize spectral efficiency based on average SNR and channel-quantization codebook size; in, the delayed, quantized-CSIT ZF system’s SU/MU switching point was explicitly derived based on average SNR and normalized Doppler.Source: A. Related Work · p. 3
  • The described switching approach considers only SU mode or the mode serving as many users as transmit antennas, neglecting how active-user count affects gains and residual interference.Source: A. Related Work · p. 3

B. Contributions

The paper proposes multi-mode transmission (MMT), selecting the number of active users to balance array gain, spatial multiplexing gain, and residual interference. It derives rate approximations and reports throughput benefits, alongside design guidance for imperfect CSIT and feedback constraints.

B. Contributions

  • The proposed algorithm adaptively selects the number of active users, balancing transmit array gain, spatial multiplexing gain, and residual inter-user interference.The paper combines MMT with channel-independent scheduling such as round-robin to serve a large user population with low complexity and reduced feedback.Source: B. Contributions · p. 3
  • Accurate closed-form achievable-rate approximations for different modes support selecting the mode that provides the highest throughput.The paper states that its analysis can identify which source dominates CSIT imperfection, guiding codebook sizing for a given delay-mobility setting.Source: B. Contributions · pp. 3–4
  • MMT provides throughput gain over SU-MIMO and, under a feedback overhead constraint, performance close to opportunistic scheduling with instantaneous feedback from many users.The proposed strategy does not provide multiuser diversity gain because it does not exploit instantaneous CSIT.Source: B. Contributions · p. 3
  • The SU mode is recommended at low and high SNR, while MU mode can be used at medium SNR with its active-user count adjusted; the full Nt-user mode is normally not activated.The full MU mode has the highest residual interference and no array gain; practical MU operation is recommended only with low mobility, with more feedback bits or advanced feedback techniques.Source: B. Contributions · p. 4
  • For a fixed transmission mode, MMSE precoding outperforms ZF, but under imperfect CSIT both benefit from MMT and provide close performance.The paper prefers ZF with MMT because it requires less feedback and its preferred mode can be easily calculated.Source: B. Contributions · p. 4
  • With a total feedback-bit constraint, high-quality feedback from fewer than Nt users plus mode selection provides good performance relative to instantaneous feedback from many users.The paper also reports that ZF with MMT requires less feedback than MMSE and that its preferred mode can be easily calculated.Source: B. Contributions · p. 4
  • Compared with the earlier SU/MU switching study, this paper considers the active-user count, extends MMT to larger user populations, and adds practical design insights.The earlier study considered only the full MU mode with Nt users and emphasized analysis.Source: B. Contributions · p. 4

II. SYSTEM MODEL

The system model considers a MIMO broadcast channel with limited-feedback CSIT that may be delayed, and compares SU and MU transmission modes. Its delay model assumes a one-symbol CSI delay and uses a Gauss-Markov fit for numerical analysis.

II. SYSTEM MODEL

  • The model has Nt transmit antennas and U single-antenna users, with the active mode M representing the number of served users; it assumes equal power allocation and normalized noise.Eigen-beamforming is used for SU mode, while ZF precoding is used for MU mode.Source: §II · p. 5
  • The analysis assumes spatially white Gaussian channels, perfect CSI at receivers, and transmitter CSI obtained through limited feedback with delay.The investigation of other channel models is left to future work.Source: §II · p. 5

A. CSI Delay Model

  • The delay model assumes a one-symbol CSI delay, with channel correlation tied to normalized Doppler frequency; the numerical analysis uses a Gauss-Markov fit of Clarke's isotropic-scattering model.The authors state that the analysis can extend to multiple-symbol delays and to estimation or prediction error.Source: A. CSI Delay Model · pp. 5–6

B. Channel Quantization

Channel directions are quantized using codebooks known to transmitter and receiver, and performance is averaged over fading and random codebooks. The paper analyzes mode throughput and describes how mode count trades spatial multiplexing gain against array gain.

B. Channel Quantization

  • Each user's channel direction is quantized to its closest codeword from a user-specific codebook of size 2^B, with codewords of unit norm.For analysis, quantization vectors are independently drawn from the isotropic distribution on the Nt-dimensional unit sphere, and results average over codebooks and fading.Source: B. Channel Quantization · p. 6

III. THROUGHPUT ANALYSIS AND MODE SELECTION

  • The active-user count that maximizes sum throughput is closely related to transmit array gain, spatial multiplexing gain, and residual inter-user interference.MMT adaptively selects the mode to balance these effects, and selected users may be chosen with channel-independent scheduling such as round-robin.Source: §III · p. 7

A. Perfect CSIT

  • In perfect-CSIT mode M, the system has spatial multiplexing gain M and array gain Nt − M + 1 per user, so increasing M raises multiplexing gain while reducing array gain.For SU mode, eigen-beamforming provides transmit array gain Nt.Source: A. Perfect CSIT · p. 7, 3) Mode Selection · p. 8

3) Mode Selection:

Mode selection chooses among transmission modes to trade spatial multiplexing gain against array gain. With delayed and quantized CSIT, the paper derives achievable-rate approximations to support that choice.

3) Mode Selection:

  • As M increases, spatial multiplexing gain rises while array gain falls, creating a tradeoff that determines the throughput-maximizing mode.The mode-selection procedure compares the achievable rates for the candidate modes.Source: 3) Mode Selection · p. 8

B. Imperfect CSIT

  • Because exact achievable rates are difficult to derive under delayed and quantized CSIT, the paper provides accurate closed-form approximations for selecting a mode.The analysis includes delay and channel quantization, and the SU-mode average achievable rate is cited from.Source: B. Imperfect CSIT · p. 8

1) Signal term:

The section derives achievable-rate approximations for transmission modes under imperfect channel state information and analyzes how residual interference shapes mode selection. The analysis balances array gain, spatial multiplexing gain, and interference to maximize throughput.

1) Signal term:

2) Interference Term:

  • Residual interference increases with delay, transmit power, and the number of served users, but decreases as codebook size increases.Its growth with transmit power makes the system interference-limited at high SNR.Source: 2) Interference Term · p. 10
  • Closed-form achievable-rate approximations are derived for MU modes with both channel delay and quantization, including a high-SNR approximation.Source: 2) Interference Term · pp. 10–11
  • The approximation is very accurate at low to medium SNR; at high SNR it becomes a lower bound, with accuracy decreasing as M increases.Source: 2) Interference Term · p. 12
  • Mode M = 3 provides higher throughput than full MU mode M = 4 and gains around 2 bps/Hz over dual-mode switching at medium SNR.Full mode has the highest residual interference and no array gain, making it desirable to serve fewer than Nt users.Source: 2) Interference Term · p. 12
  • For B = 10 bits, M = 2 has the highest throughput throughout the considered fdTs range; for B = 20 bits, M = 3 has the highest rate at high SNR when fdTs is small, while M = 4 has lower throughput; M = 4 leads only for B = 30 bits and very small fdTs (< 10^-2).Source: 2) Interference Term · pp. 12–13

IV. MULTI-MODE TRANSMISSION WITH ROUND-ROBIN SCHEDULING

The proposed scheduler combines multi-mode transmission with round-robin user selection. It serves users in circular order while selecting the number of simultaneous users based on the transmission mode.

IV. MULTI-MODE TRANSMISSION WITH ROUND-ROBIN SCHEDULING

  • The scheduling algorithm is low complexity, provides temporal fairness, and suits delay-sensitive services or settings with feedback from only a few users.Source: §IV · p. 13
  • The scheduler serves users in circular order, moving each scheduled user to the queue tail, while allowing multiple users to be served in one slot.Source: §IV · p. 13
  • In each slot, MMT determines the number of selected users from the throughput analysis; for homogeneous users, a fixed mode M⋆ selects users from the queue head.Source: §IV · p. 13
  • Only the selected M users, with M ≪ U, need to provide instantaneous CSI for precoder design, and the algorithm performs close to opportunistic scheduling using instantaneous feedback from many users.Source: §IV · p. 13

V. NUMERICAL RESULTS

The numerical evaluation focuses on a homogeneous network with independent, identically distributed user channels. It uses four base-station transmit antennas, a value noted as implemented in broadband wireless standards such as 3GPP LTE.

V. NUMERICAL RESULTS

  • The evaluation considers a homogeneous network in which users have i.i.d. channels.Source: §V · p. 14
  • The base station is modeled with Nt = 4 transmit antennas.Source: §V · p. 14
  • The paper notes that four transmit antennas are implemented in broadband wireless standards such as 3GPP LTE.Source: §V · p. 14

A. Operating Region and Throughput Gain of Multi-mode Transmission

The operating regions show how the preferred transmission mode varies with SNR, user mobility, and feedback-bit count. Under the considered parameters, the full MU mode is not active, while single-user and lower MU modes are preferred in different conditions.

1) Operating Regions:

  • The operating-region plots mark the active mode across mobility and feedback-bit settings.Source: 1) Operating Regions · p. 14
  • The SU mode is active at both low and high SNR for given mobility and feedback bits, due to its array gain and robustness to imperfect CSIT, respectively.Source: 1) Operating Regions · p. 14
  • MU modes require low mobility and more feedback bits; at B = 15 bits, M = 2 requires v ≤ 12 km/hr and M = 3 requires v ≤ 6 km/hr.Source: 1) Operating Regions · p. 14
  • At v = 5 km/hr, activating M = 3 requires B ≥ 14 bits.Source: 1) Operating Regions · p. 14
  • The full MU mode M = Nt is not active under the considered parameters because it has the highest residual interference and provides no array gain.Source: 1) Operating Regions · p. 14

2) SU-MIMO vs. MMT:

MMT-ZF’s advantage over single-user beamforming depends on feedback quality and operating conditions. The results show how mobility, quantization error, and precoder choice shape mode performance.

2) SU-MIMO vs. MMT:

  • At B=8, MMT-ZF provides throughput gain over SU-BF for SNR 0–18 dB; at B=12, it provides throughput gain across −5–25 dB, exceeding 20% for SNR 10–20 dB.Source: 2) SU-MIMO vs. MMT · p. 15

3) Delay vs. Quantization Error:

  • Increasing B cannot eliminate delay-mobility effects: quantization error dominates in the examined scenarios, but delay-mobility starts to dominate when B is too large.One example requires B=13 to reach the stated quantization-error level; further rate improvement stops once B>20.Source: 3) Delay vs. Quantization Error · p. 15
  • For a given delay and feedback-bit count, MU-MIMO should be used only at low mobility (≤12 km/hr); LTE’s 1 msec frame is expected to make it more applicable than WiMAX’s 5 msec frame.The results also state that LTE’s B=4 is insufficient for MU-MIMO and recommend B≥8.Source: 3) Delay vs. Quantization Error · p. 15

B. ZF vs. MMSE Precoding

  • For perfect CSIT, MMT-ZF and MMSE sum rates are very close; with imperfect CSIT, mode switching is also required for MMSE to improve spectral efficiency.Applying MMT to both systems brings ZF performance close to MMSE performance.Source: B. ZF vs. MMSE Precoding · pp. 15–16
  • With MMT, ZF performance approaches MMSE performance, while ZF is preferred because its instantaneous-CSI feedback is needed from fewer users and its preferred mode is easier to determine.MMSE can increase throughput at low SNR, but determining its preferred mode is difficult to analyze.Source: B. ZF vs. MMSE Precoding · p. 16

C. MMT Under a Feedback Overload Constraint

Under a fixed feedback budget, MMT-ZF almost always outperforms PU2RC and SU-BF and performs close to opportunistic ZF scheduling. Its conclusions also describe when mode switching and additional CSIT accuracy offer limited gains.

C. MMT Under a Feedback Overload Constraint

  • With a total feedback-bit constraint, the section compares MMT-ZF and US-ZF; it focuses on limited feedback and does not consider delay.The analysis determines the number of feedback bits per user and the number of active users.Source: C. MMT Under a Feedback Overload… · p. 16
  • MMT-ZF almost always outperforms PU2RC and SU-BF and performs close to US-ZF; at low SNR, PU2RC is slightly higher because of multiuser-diversity power gain.As the total feedback budget BT increases, US-ZF throughput increases, while MMT-ZF saturates when full mode M=Nt is active.Source: C. MMT Under a Feedback Overload… · p. 16

VI. CONCLUSIONS

  • The conclusion describes MMT as selecting the active-user count from average achievable rates while balancing transmit-array gain, spatial multiplexing gain, and residual interference.The full mode M=Nt is normally not activated because it has the highest residual interference and no array gain.Source: §VI · p. 17
  • Combined with round-robin scheduling, MMT serves many users and reduces feedback while providing throughput close to instantaneous-CSI feedback under a total feedback constraint.The scheduling description identifies average SNR, codebook size, and normalized Doppler frequency as selection factors.Source: §VI · p. 17

APPENDIX

The appendix derives rate approximations by modeling received powers and SINR distributions, then using them to obtain expected rates. It also gives modulo-based steps for the scheduling algorithm.

A. Proof of Theorem 2

  • Theorem 2’s proof assumes independent interference terms and derives the CDF of a random variable before evaluating the expectation of ln(1+x) by integration by parts.The proof forms a weighted sum of independent chi-square variables in the derivation.Source: A. Proof of Theorem 2 · pp. 17–18

B. Proof of Corollary 1

  • For delayed and limited-feedback systems, Corollary 1’s proof approximates received signal power and SINR, omitting a small quantization term for large B.It then uses the resulting SINR approximation to recover the stated result.Source: B. Proof of Corollary 1 · pp. 18–19

C. Proof of Theorem 3

  • Theorem 3’s proof approximates high-SNR SINR, derives the CDF of the resulting random variable, and follows the earlier expectation steps to obtain the result.The derivation specifies independent random variables and uses parameters for delay and quantization effects.Source: C. Proof of Theorem 3 · p. 19

SCHEDULING ALGORITHM

Loading 0903.5108v2…