PaperlayerView exact revision on arXiv

Source-linked AI summary

Downlink Training Techniques for FDD Massive MIMO Systems: Open-Loop and Closed-Loop Training with Memory

Junil Choi, David J. Love, Patrick Bidigare

arXiv:1309.7712v2cs.IT

TL;DR

FDD massive MIMO makes conventional training and CSI acquisition costly as the base station antenna count grows. The paper proposes open-loop and closed-loop training with memory, using successive channel estimation and limited feedback. Closed-loop memory training can outperform open-loop training, particularly when SNR is low, training is short relative to N_t, or the block index is small.

  • Problem

    FDD massive MIMO requires substantial downlink training and CSI feedback because large antenna arrays make conventional training overhead costly.

  • Method

    The paper proposes open-loop and closed-loop training with memory, where users exploit channel statistics and prior observations, with closed-loop users feeding back selected training-signal indices.

  • Results

    Closed-loop training with memory is slightly better than open-loop/single-shot training with T = N_t when α = 0.9 and i = 9.

  • Takeaways & Limitations

    Small feedback can further reduce downlink training overhead even when the transmitter lacks channel-statistics side information.

Abstract

from arXiv · show

The concept of deploying a large number of antennas at the base station, often called massive multiple-input multiple-output (MIMO), has drawn considerable interest because of its potential ability to revolutionize current wireless communication systems. Most literature on massive MIMO systems assumes time division duplexing (TDD), although frequency division duplexing (FDD) dominates current cellular systems. Due to the large number of transmit antennas at the base station, currently standardized approaches would require a large percentage of the precious downlink and uplink resources in FDD massive MIMO be used for training signal transmissions and channel state information (CSI) feedback. To reduce the overhead of the downlink training phase, we propose practical open-loop and closed-loop training frameworks in this paper. We assume the base station and the user share a common set of training signals in advance. In open-loop training, the base station transmits training signals in a round-robin manner, and the user successively estimates the current channel using long-term channel statistics such as temporal and spatial correlations and previous channel estimates. In closed-loop training, the user feeds back the best training signal to be sent in the future based on channel prediction and the previously received training signals. With a small amount of feedback from the user to the base station, closed-loop training offers better performance in the data communication phase, especially when the signal-to-noise ratio is low, the number of transmit antennas is large, or prior channel estimates are not accurate at the beginning of the communication setup, all of which would be mostly beneficial for massive MIMO systems.

I. INTRODUCTION

FDD massive MIMO makes conventional downlink training costly because large antenna arrays require substantial training and CSI resources. The paper proposes open-loop and closed-loop training with memory to reduce this overhead through successive channel estimation and limited feedback.

  • Motivation: FDD dominates current cellular systems, but massive MIMO research largely assumes TDD with channel reciprocity.The paper therefore focuses on downlink training for FDD massive MIMO systems.
  • Prior Training: Single-shot training estimates each channel from only the current received training signal and discards past signals.This conventional design motivates retaining previous channel information.
  • Motivation: Large transmit arrays can make downlink training overhead overwhelm precious FDD resources.Earlier MIMO systems used few antennas, so training duration was often treated as negligible.
  • Proposed Frameworks: The paper proposes open-loop and closed-loop frameworks with memory that use successive channel estimation for FDD massive MIMO.The base station and user share a common training-signal set in advance.
  • Proposed Frameworks: Open-loop training sends shared training signals round-robin while the user predicts or estimates channels from previous estimates and long-term statistics.The assumed statistics include temporal and spatial correlations known to the user.
  • Proposed Frameworks: Closed-loop training lets the user select and feed back the best future training signal using prior channel knowledge and previously received signals.The framework uses limited feedback to improve channel estimation with less training overhead.

III. SINGLE-SHOT TRAINING AND THE CEILING EFFECT

The conventional single-shot framework discards prior training observations and estimates each channel from the current observation alone. Under practical channel conditions, its normalized received SNR does not keep increasing with the number of transmit antennas.

  • Conventional Single-Shot Training: Prior single-shot training discards previously received training signals and estimates the current channel from only the current signal.This is the defining limitation examined in the section.
  • Ceiling Effect: With an optimal single-shot training signal and fixed training length below the antenna count, normalized average received SNR quickly saturates as N_t increases.The resulting SNR ceiling means adding transmit antennas does not improve performance under most practical channel conditions.
  • Channel Estimation Model: The paper analyzes MMSE channel estimation at the user for a complex Gaussian channel with covariance R.The spatial correlation matrix R is assumed fixed because user location changes little at moderate velocities.

IT + XHRX

This section derives the optimal training structure for closed-loop single-shot MMSE estimation and relates performance to spatial correlation, training duration, and transmit power. The optimal signal uses dominant spatial directions, while the resulting bounds characterize when additional training helps.

  • Optimal Training Structure: The optimal closed-loop single-shot training signal with unlimited feedback minimizes the channel-estimation MSE for a given training design.The formulation assumes the user can inform the base station of the best training signal.
  • Optimal Training Structure: The optimal training signal transmits along the first T dominant eigen-directions of R.This structure follows from decomposing the spatial correlation matrix as R = UΛU^H.
  • Performance Relations: If one correlation spectrum majorizes another, the optimal orthogonal training MSE for the more spatially correlated channel is no larger.The comparison is expressed as MSE(X_H) ≤ MSE(X_L).
  • Performance Relations: More spatial correlation, longer training, or higher transmit power can yield lower channel-estimation MSE.The stated conclusions are derived for the optimal signal and also supported numerically for a general training signal.

B. Ceiling effect of single-shot training

Single-shot training has a received-SNR ceiling that does not generally grow linearly with the number of transmit antennas. The bound depends on channel correlation structure and training duration, while increasing training length reduces data-communication resources.

  • Perfect feedback of the optimal training signal and estimated channel is assumed to isolate the effect of training.The resulting bound is therefore evaluated under idealized feedback conditions.
  • The upper bound on Γ_ss,opt is non-trivial in general but becomes trivial when the channel-correlation matrix has rank one.The rank-one condition is tr(R) = λ_1 = N_t.
  • Single-shot training’s normalized average received SNR is not a linearly increasing function of N_t.For i.i.d. Rayleigh fading, it remains fixed for given T and ρ even as N_t approaches infinity.
  • With the exponential spatial-correlation model, the maximum Γ_ss,opt is a function of T and a, not N_t.Here, a controls spatial correlation, with larger a indicating stronger correlation.
  • Even with optimal single-shot training, Γ_ss,opt saturates in highly correlated channels.The figure compares simulations with upper bounds at ρ = 20 dB and T = 4; increasing T can raise Γ_ss,opt but reduces the T − L channel uses available for data.
  • Exploiting temporal channel correlation is identified as a way to reduce the single-shot ceiling effect.The paper argues that training should leverage temporal correlation to benefit from many antennas.

IV. PROPOSED TRAINING FRAMEWORKS

The proposed frameworks use shared training signals and channel memory to reduce FDD massive-MIMO downlink-training overhead. Open-loop training cycles through signals without feedback, while successive estimation uses prior observations and channel statistics.

  • B. Closed-loop training with memory: The closed-loop framework feeds back the preferred training signal, and the paper derives its performance upper bound under perfect feedback.The section also presents training-signal-set design and preferable system parameters relative to open-loop training.
  • The frameworks share a common training-signal set between the base station and user.The open-loop set is indexed with B bits and contains 2^B signals.
  • A. Open-loop training with memory: The user’s successive MMSE estimate combines the current training observation with previous observations and channel statistics.A Kalman filter, or a more advanced particle filter, can track channel evolution using temporal and spatial information.
  • The framework can also use a time-varying P, analogous to differential codebooks, for better performance.This extension is stated as compatible with the proposed framework rather than evaluated here.

B. Closed-loop training with memory

Closed-loop training with memory lets the user select future training signals using channel prediction, prior training observations, and limited feedback. Selection can target either channel-estimation MSE or normalized average received SNR.

  • Training-signal selection uses channel statistics and previously received training signals to track the predicted channel at the current fading block.
  • The predefined training set is shared in advance, while the base station lacks direct access to the user’s channel statistics and prior received training signals.
  • The user selects the best training signal from a predefined set using channel prediction and feeds its B-bit index back to the base station.The base station uses the selected signal in the next fading block.
  • The user can select the training signal by minimizing channel-estimation MSE or maximizing normalized average received SNR for data communication.
  • Maximizing the SNR-based objective is equivalent to minimizing MSE augmented with q(P_i), whose impact is non-negligible for moderately large N_t, highly spatially correlated channels, and low SNR.

C. Closed-loop training with memory with full feedback to minimize MSE

With unlimited feedback, the MSE-optimal closed-loop training signal follows the dominant eigen-directions of the predicted channel covariance. This full-feedback solution provides an MSE lower bound and shows why memory improves estimation over single-shot training in temporally correlated channels.

  • The full-feedback optimum is possible only with unlimited feedback and therefore provides an MSE lower bound for closed-loop training with memory.
  • The optimal training signal uses the first T dominant eigen-directions of the prediction matrix R_i|i−1.
  • The dominant eigenvectors vary with the fading block, so full-feedback training scans among the eigen-directions of the original spatial correlation matrix.
  • The MSE for the optimal signal is given by the paper’s full-feedback lower-bound expression derived using the Kalman-filter update.
  • For i > 0 in temporally correlated channels with η ≈ 1, memory-based closed-loop training has lower MSE than closed-loop single-shot training and can reduce its ceiling effect.

D. Design of training signal set P

The finite training-signal codebook is designed to approximate the MSE-optimal eigen-direction scans, while parameter analysis identifies when closed-loop memory provides the greatest gain. Gains are emphasized at low SNR, early estimation, and with limited training dimensions.

  • D. Design of training signal set P: The training set should contain nearly orthogonal signals, and Grassmannian subspace packing is used to generate such a set numerically.
  • D. Design of training signal set P: The adopted GSP training set is selected from candidate sets by maximizing their minimum chordal distance and is used for performance evaluation.
  • Variation with SNR: Closed-loop training is expected to be more beneficial in the low-SNR regime because the selected training subspace matters more there than in the high-SNR regime.
  • Variation with length of training phase: When T = N_t, there is no preferable training direction; for 1 < T < N_t, the advantage of the dominant-eigenvector subspace shrinks as T increases.
  • Variation with fading block index i: Closed-loop training with memory gains more when prior channel estimates are inaccurate at the beginning of channel estimation.

V. SIMULATION RESULTS

Monte Carlo simulations evaluate successive channel estimation for correlated FDD massive MIMO channels. The proposed open- and closed-loop memory frameworks outperform open-loop/single-shot training, with gains depending on correlation, feedback, SNR, training length, and channel variation.

  • 10,000-iteration Monte Carlo simulations use 10 temporally and spatially correlated fading blocks per iteration.The evaluation uses Jakes’ temporal correlation model and a normalized average received SNR metric.
  • PGSP codebook gains increase with B for Nt = 16 and 64, while fewer than 10 bits appear sufficient for a notable gain.The gain from larger B is more prominent when Nt is large, and B = 6 is used in later simulations.
  • Both proposed memory-based frameworks outperform open-loop/single-shot training with the same training length T.The comparison includes open-loop memory and closed-loop memory using MSE- and SNR-based selection.
  • Closed-loop memory gains are larger at low ρ, small T relative to Nt, and early fading blocks.SNR-based closed-loop memory can also provide a non-negligible gain over MSE-based closed-loop memory for moderately large Nt and low ρ in highly correlated channels.
  • As spatial correlation increases, performance improves for all schemes.The simulations vary the spatial correlation parameter a, with larger a corresponding to more highly correlated channels.
  • At Nt = 16, MSE-based closed-loop memory has lower MSE than SNR-based closed-loop memory.This result demonstrates a tradeoff between the SNR and MSE metrics used for closed-loop training.
  • Open-loop/single-shot received SNR quickly saturates as Nt increases, whereas memory-based training reduces this ceiling with small T, especially for large a.The comparison includes closed-loop/single-shot training with full feedback, which also exhibits the ceiling effect.
  • High user velocity causes almost 1.4dB received-SNR loss in the saturation regime.The paper recommends more frequent open-loop sounding at high velocity and notes that memory-based open-loop training remains applicable.

VI. CONCLUSION

The paper proposes successive channel prediction and estimation for open- and closed-loop FDD massive MIMO training. Using prior channel information and limited feedback improves estimation with short per-block training and can further reduce downlink training overhead.

  • Successive channel prediction and estimation form the basis of the proposed open- and closed-loop FDD massive MIMO training frameworks.
  • Prior channel statistics and previous received training signals improve channel estimation with short training signals in each fading block.
  • Small feedback identifying the next training signal can further reduce downlink training overhead without transmitter-side channel statistics.

APPENDIX A PROOF OF LEMMA 1

The proof derives the single-shot MSE optimization from the channel covariance structure. Majorization and Schur-convexity establish that stronger spatial eigenvalue concentration yields no larger MSE under the stated ordering.

  • With fixed R, minimizing MSE is transformed into an optimization problem using trace identities and the eigendecomposition R = UΛU^H.
  • The transformed optimization uses eX^H eX = ρI_T and a block generalized Rayleigh quotient to obtain the optimal single-shot training solution.
  • The proof applies Schur-convexity of a symmetric function f(x) to compare training-related eigenvalue vectors through majorization.
  • When λ(R_H) ≻ λ(R_L), the resulting MSE ordering is MSE(X_H) ≤ MSE(X_L).

APPENDIX C PROOF OF LEMMA 3

The proof decomposes the channel into an MMSE estimate and an independent residual, then propagates residual covariance through successive training blocks. The block-one result is generalized recursively to later blocks.

  • The channel is decomposed as h = b̂h + r, with independent residual r due to MMSE-estimator orthogonality.
  • The residual covariance is represented through its eigenvectors and eigenvalues as R_r = U_rΛ_rU_r^H.
  • The proof evaluates expectations over the residual or noise and the channel, then bounds terms using the largest residual eigenvalue.
  • For the first fading block, the optimal training structure uses the ordered eigenvectors of the prior covariance, and its MSE is given by the derived block-one expression.
  • The block-one MSE expression is extended to i > 1 by recursive derivation.
Loading 1309.7712v2…