PaperlayerView exact revision on arXiv

Source-linked AI summary

Mode Switching for MIMO Broadcast Channel Based on Delay and Channel Quantization

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

arXiv:0812.3120v1cs.ITmath.IT

TL;DR

Imperfect CSIT makes it uncertain when MU-MIMO outperforms SU-MIMO, since the modes are affected differently by delay and quantization. The paper derives rate approximations and uses them for mode switching; MU is active only under restricted channel conditions.

  • Problem

    With imperfect CSIT from delay and channel quantization, it is unclear when MU-MIMO can outperform SU-MIMO.

    Source: §I · p. 2

  • Method

    The paper derives closed-form approximations for SU- and MU-mode ergodic rates and proposes switching modes based on those rates.

    Source: C. SU/MU Mode Switching · p. 9, §VI · p. 20

  • Results

    MU-MIMO throughput saturates at high SNR due to residual inter-user interference, and MU mode is active only in the medium-SNR regime with small normalized Doppler and a large codebook.

    Source: §VI · p. 20

  • Takeaways & Limitations

    The operating conditions for MU mode are restricted: normalized Doppler must be small and the channel-quantization codebook large.

    Source: §VI · p. 20

Abstract

from arXiv · show

Imperfect channel state information degrades the performance of multiple-input multiple-output (MIMO) communications; its effect on single-user (SU) and multi-user (MU) MIMO transmissions are quite different. In particular, MU-MIMO suffers from residual inter-user interference due to imperfect channel state information while SU-MIMO only suffers from a power loss. This paper compares the throughput loss of both SU and MU MIMO on the downlink due to delay and channel quantization. Accurate closed-form approximations are derived for the achievable rates for both SU and MU MIMO. It is shown that SU-MIMO is relatively robust to delayed and quantized channel information, while MU MIMO with zero-forcing precoding loses spatial multiplexing gain with a fixed delay or fixed codebook size. Based on derived achievable rates, a mode switching algorithm is proposed that switches between SU and MU MIMO modes to improve the spectral efficiency, based on the average signal-to-noise ratio (SNR), the normalized Doppler frequency, and the channel quantization codebook size. The operating regions for SU and MU modes with different delays and codebook sizes are determined, which can be used to select the preferred mode. It is shown that the MU mode is active only when the normalized Doppler frequency is very small and the codebook size is large.

I. INTRODUCTION

With delayed or quantized channel state information at the transmitter (CSIT), it is unclear when MU-MIMO can outperform SU-MIMO. The paper compares their rates and proposes choosing a mode based on the estimated rates.

I. INTRODUCTION

  • The introduction frames MU-MIMO's promise as serving multiple mobiles simultaneously, while imperfect CSIT can degrade its throughput gain.The downlink requires CSIT to separate spatial channels for different users.Source: §I · p. 2, A. Related Work · p. 2

A. Related Work

  • The paper asks when MU-MIMO with imperfect CSIT can deliver higher throughput than SU-MIMO, addressing a comparison not covered by prior work focused mainly on spatial multiplexing gain.The authors identify prior rate-loss analysis as unable to accurately characterize throughput loss.Source: A. Related Work · p. 3

B. Contributions

  • SU-MIMO has a constant rate loss under imperfect CSIT, whereas MU-MIMO suffers residual inter-user interference; the paper derives closed-form ergodic-rate approximations for both.The analysis characterizes residual interference from delay and quantization and treats the two effects as equivalent.Source: B. Contributions · pp. 3–4
  • The proposed mode-switching algorithm uses ergodic sum rate as a function of average SNR, normalized Doppler frequency, and quantization codebook size.The mode-switching points are calculated from the derived ergodic-rate approximations.Source: B. Contributions · p. 4
  • With fixed delay and codebook size, SU is preferred at low and high SNR; MU can operate at medium SNR with very small normalized Doppler and a large codebook.The minimum feedback bits per user for activating MU grow approximately linearly with the number of transmit antennas.Source: B. Contributions · p. 4

II. SYSTEM MODEL

The system model considers a MIMO downlink with a multi-antenna base station and single-antenna users, with transmission to one or multiple users. Its CSI-delay model relates channel correlation to normalized Doppler frequency.

II. SYSTEM MODEL

  • The downlink has a base station with Nt antennas and single-antenna users, and serves either one user in SU mode or multiple users in MU mode.The transmission period is shorter than the channel coherence time, and the channel is assumed constant during that period.Source: §II · p. 4
  • The channel model assumes spatially white Gaussian channels with entries distributed as CN(0, 1), independently across users.The authors note that results differ across channel models and leave investigation of other models to future work.Source: §II · p. 5

A. CSI Delay Model

  • The CSI-delay model uses a one-symbol delay and relates channel correlation to normalized Doppler frequency.The correlation coefficient is given by ρ = J0(2πfdTs); the authors state that extension to multiple-symbol delay is straightforward.Source: A. CSI Delay Model · pp. 5–6

B. Channel Quantization Model

The paper models limited feedback using quantized channel directions and uses conventional SU and MU transmission techniques as the basis for its analysis. Under perfect CSIT, SU beamforming follows the channel direction.

B. Channel Quantization Model

  • With limited feedback, each user selects a channel-direction codeword from a shared codebook and feeds back its index using B bits.The codebook contains unit-norm vectors, and the analysis uses random vector quantization averaged over random codebooks and fading.Source: B. Channel Quantization Model · pp. 6–7

III. TRANSMISSION TECHNIQUES

  • The analysis uses conventional transmission techniques and ergodic rate, with one user for SU-MIMO and Nt users for MU-MIMO.Designing precoders that account for imperfect CSIT is left to future work.Source: §III · p. 7

A. SU-MIMO System

  • With perfect CSIT, SU-MIMO beamforms along the channel direction, and its ergodic capacity matches that of a maximal-ratio-combining diversity system.The paper calls this transmission choice eigen-beamforming.Source: A. SU-MIMO System · p. 8

B. MU-MIMO System

The paper uses zero-forcing (ZF) as a tractable MU-MIMO precoder and proposes choosing between SU and MU modes using estimated achievable rates. The switching strategy is intended for settings with limited feedback or strict delay constraints, but handling rate mismatch is outside the paper's scope.

B. MU-MIMO System

  • ZF precoding is used for its simple structure, which supports closed-form analysis, and its full spatial multiplexing gain with only a power offset versus optimal DPC.The paper also states that ZF is optimal among linear precoders at asymptotically high SNR.Source: B. MU-MIMO System · p. 8
  • With perfect CSIT, ZF precoders are designed to eliminate inter-user interference; under imperfect CSIT, users can estimate and feed back their received SINRs.The analysis assumes equal power allocation.Source: B. MU-MIMO System · pp. 8–9

C. SU/MU Mode Switching

  • The proposed strategy estimates both modes' achievable ergodic rates, selects the higher-rate mode, and serves one user in SU mode or Nt users in MU mode.The transmitter is assumed to know the delay and quantization conditions.Source: C. SU/MU Mode Switching · p. 9
  • The low-complexity strategy requires instantaneous channel feedback only from selected users, suiting systems with limited feedback bits, few feedback users, or strict delay constraints.It can be combined with random user selection or scheduling based on round-robin or queue length rather than channel status.Source: C. SU/MU Mode Switching · pp. 9–10
  • The paper assumes perfect transmitter knowledge of active users' actual received SINRs, although estimation error and feedback delay can cause rate mismatch and outage.Investigating how to handle those events is explicitly outside the paper's scope.Source: C. SU/MU Mode Switching · p. 10

IV. PERFORMANCE ANALYSIS AND MODE SWITCHING

The analysis derives achievable-rate approximations for SU and MU modes, using received SNR and residual interference to characterize imperfect CSIT. For SU beamforming, it provides average-SNR and ergodic-rate results under channel quantization and delay.

IV. PERFORMANCE ANALYSIS AND MODE SWITCHING

  • The analysis first examines average received SNR for beamforming and average residual interference for ZF, then derives closed-form achievable-rate approximations for both modes.These approximations are used to select the active mode.Source: §IV · p. 10

A. SU Mode–Eigen-Beamforming

  • Theorem 1 characterizes average received SNR for SU beamforming with channel quantization and CSI delay, separating the degradations associated with each impairment.The text identifies Δ(Q)_BF and Δ(D)_BF as the respective impacts of quantization and delay.Source: A. SU Mode–Eigen-Beamforming · p. 11
  • With fixed codebook size and delay, imperfect CSIT causes a constant SNR degradation factor independent of transmit power and a constant rate loss at high SNR.The passage attributes the SNR decrease to the quantization error term.Source: A. SU Mode–Eigen-Beamforming · p. 11
  • Theorem 2 gives an approximation for the ergodic achievable rate of SU beamforming with delay; the analysis also treats the delayed channel direction as the beamforming vector.For the combined delay-and-quantization case, the beamformer uses the quantized channel direction with delay.Source: A. SU Mode–Eigen-Beamforming · pp. 10–13

B. Zero-Forcing

For ZF, imperfect CSIT produces residual inter-user interference from channel quantization and delay. With fixed error or codebook size, that interference limits spatial multiplexing; maintaining a constant rate loss requires improving CSIT as SNR rises.

1) Average Residual Interference:

  • Average residual interference increases with the number of interferers and their transmit power, and worsens as the user's CSIT becomes less accurate.A larger delay or smaller codebook size yields higher residual interference.Source: 1) Average Residual Interference · p. 14
  • With fixed delay or codebook size, ZF becomes interference-limited and cannot achieve full spatial multiplexing gain.The paper notes that power allocation could improve performance, but leaves it for future work.Source: 1) Average Residual Interference · p. 14
  • To keep a constant rate loss per user and sustain spatial multiplexing gain, channel error from delay and quantization must decrease as SNR increases.The codebook-size and CSI-delay condition is stated for a constant rate loss of log2 δ0 bps/Hz per user.Source: 1) Average Residual Interference · p. 14
  • For a given ρ², the feedback bits per user must scale with SNR; in the no-delay case, the condition reduces to the limited-feedback requirement.The paper says this no-delay result agrees with the limited-feedback-only result.Source: 1) Average Residual Interference · p. 14

2) Achievable Rate:

The paper approximates MU ZF achievable rates from the residual interference caused by delay and channel quantization. It treats those two impairment terms as equivalent in distribution up to their means, and reports that its rate approximation is accurate at low to medium SNR.

2) Achievable Rate:

  • The ZF achievable-rate approximation is reported as accurate at low to medium SNR, while the cited lower bound is described as very loose.The paper introduces a more accurate approximation after noting the lower bound's looseness.Source: 2) Achievable Rate · p. 15
  • Under the quantization-cell approximation, the residual interference term due to quantization is exponential with mean δ.The paper uses this distribution to derive the achievable-rate approximation.Source: 2) Achievable Rate · p. 16
  • Residual interference from delay and quantization has equivalent effects in the analysis, differing only in their means.The paper says the resulting approximation also applies to the limited-feedback system and verifies it by simulation.Source: 2) Achievable Rate · pp. 16–17
  • Theorem 4 approximates each MU user's ergodic achievable rate with both CSI delay and channel quantization.The analysis derives the approximation from the distributions of the residual interference terms.Source: 2) Achievable Rate · p. 16

C. Mode Switching

The numerical results assess rate approximations and map when SU/BF or MU/ZF is preferred. The MU operating region narrows with greater Doppler, smaller codebooks, and more transmit antennas.

C. Mode Switching

  • For B = 10, v = 20 km/hr, fc = 2 GHz, and Ts = 1 msec, the lower bound is very loose while the approximation is accurate, especially for Nt = 2.The approximation is reported to be a lower bound, and imperfect CSIT reduces sum rate as Nt increases.Source: C. Mode Switching · p. 17
  • The ZF approximation is accurate at low-to-medium SNR and becomes a lower bound at high SNR, with about 0.7 bps/Hz total error or 0.175 bps/Hz per user.The comparisons use B = 18, fc = 2 GHz, v = 10 km/hr, and Ts = 1 msec.Source: C. Mode Switching · p. 17

V. NUMERICAL RESULTS

  • Numerical results plot operating regions across normalized Doppler, feedback bits, and transmit-antenna counts, and demonstrate an extension from ZF to MMSE precoding.The operating-region plots are used to identify the active mode for given parameters.Source: §V · p. 18, A. Operating Regions · p. 18

A. Operating Regions

  • ZF activation requires small normalized Doppler (fdTs < 0.055 at B = 20 and < 0.046 at B = 16); its operating region shrinks as Doppler rises and codebook size decreases.The mode should be active only when B ≥ 12 at v = 10 km/hr and B ≥ 14 at v = 20 km/hr.Source: A. Operating Regions · p. 18
  • For fixed fdTs and B, SU is active at both low and high SNR, while MU can occupy the intervening medium-SNR region.The reported explanation is SU array gain at low SNR and robustness to imperfect CSIT at high SNR.Source: A. Operating Regions · p. 18

B. ZF vs. MMSE Precoding

The paper extends its mode-switching analysis from ZF to MMSE precoding and concludes that imperfect-CSIT MU throughput saturates at high SNR. Its conclusions remain centered on ZF, whose sensitivity motivates further precoding work.

B. ZF vs. MMSE Precoding

  • The authors apply their ZF mode-switching results to MMSE, using the ZF system’s higher-SNR switching point for MMSE.At high SNR, MMSE precoders converge to ZF precoders, so the ZF derivations also apply there.Source: B. ZF vs. MMSE Precoding · p. 19
  • MMSE precoding outperforms ZF at low-to-medium SNR, converges to ZF at high SNR, and has the same rate ceiling as ZF.MMSE crosses the BF curve roughly at the same point, after which the SU mode is selected.Source: B. ZF vs. MMSE Precoding · p. 19
  • The MMSE and ZF switching points are very close, while calculated switching points are roughly 2.5–3 dB lower than simulation points.The comparison uses the second switching point for ZF and the sole switching point for MMSE.Source: B. ZF vs. MMSE Precoding · p. 19

VI. CONCLUSIONS

  • With fixed delay or feedback bits, MU-MIMO throughput saturates at high SNR because of residual inter-user interference.The conclusion describes the analysis as comparing SU and MU transmissions under delayed and quantized CSIT and deriving closed-form ergodic-rate approximations for both.Source: §VI · pp. 19–20
  • The study uses zero-forcing MU precoding, and the authors identify robust precoding and nonlinear precoding under imperfect CSIT as future work.They also propose investigating efficient power control and methods to improve CSIT quality with a fixed codebook size.Source: §VI · p. 20

APPENDIX

The appendix assembles integral identities and lemmas used in the rate derivations, then applies distributional and independence assumptions in proofs for the rate results.

A. Useful Results for Rate Analysis

  • The useful-results appendix introduces integral results and a lemma used repeatedly in achievable-rate derivations; the lemma is proved by integration by parts.It also identifies E1(x) as the first-order exponential-integral function.Source: A. Useful Results for Rate Analysis · pp. 20–21

B. Proof of Theorem 1; C. Proof of Theorem 2

D. Proof of Lemma 1

  • The Lemma 1 proof models quantization interference as the product z = xy, with x gamma-distributed and y beta-distributed and independent.The proof then derives the cumulative distribution of z.Source: D. Proof of Lemma 1 · p. 22

E. Proof of Theorem 4

Theorem 4’s proof approximates an ergodic achievable rate using a distribution for summed interference terms. The supplied figures identify rate-comparison and mode-switching visuals, but provide no further plotted outcomes here.

E. Proof of Theorem 4

  • Theorem 4’s proof models y as a weighted sum of two independent chi-square variables and approximates the user’s ergodic achievable rate.Its derivation assumes interference terms are mutually independent and independent of the signal term.Source: E. Proof of Theorem 4 · pp. 23–24

SYSTEM PARAMETERS

  • The appendix includes a system-parameter table, but its contents are not stated in the supplied passage.Only the table label is provided.Source: Table II · p. 27

MODE SWITCHING POINTS

  • The figures are identified as comparisons of ZF approximations and simulations, mode switching under delay and quantization, and operating regions across antenna counts.The supplied captions specify Nt = U = 4 for Fig. 1 and different Nt for Fig. 5; no plotted outcomes are given in these captions.Source: Figure 1 · p. 27, Figure 2 · p. 27, Figure 3 · p. 28
Loading 0812.3120v1…