Source-linked AI summary
Noncoherent Trellis Coded Quantization: A Practical Limited Feedback Technique for Massive MIMO Systems
Junil Choi, Zachary Chance, David J. Love, Upamanyu Madhow
TL;DR
Massive MIMO needs accurate CSI, but FDD limited-feedback codebooks become impractical as antenna dimensions grow. The paper proposes NTCQ, which maps Grassmannian quantization to noncoherent detection and uses trellis-based encoding; simulations report near-optimal performance with moderate complexity and feedback overhead.
Problem
Accurate CSI is needed for massive-MIMO beamforming and multiplexing, but FDD codebook quantization does not scale efficiently with many transmit antennas.
Method
NTCQ maps Grassmannian channel quantization to noncoherent sequence detection and implements it through TCQ and Viterbi-based encoding.
Results
NTCQ and its extensions achieve near-optimal performance with moderate complexity and feedback overhead, while encoding complexity scales linearly with transmit antennas.
Takeaways & Limitations
NTCQ provides a flexible and scalable limited-feedback approach that can exploit temporal and spatial channel correlations.
Takeaways & Limitations
Feedback-overhead limits generally prevent scaling NTCQ to indefinitely many antennas, although overhead may be reasonable for initial 32-to-64-antenna deployments.
Abstract
from arXiv · showhide
Accurate channel state information (CSI) is essential for attaining beamforming gains in single-user (SU) multiple-input multiple-output (MIMO) and multiplexing gains in multi-user (MU) MIMO wireless communication systems. State-of-the-art limited feedback schemes, which rely on pre-defined codebooks for channel quantization, are only appropriate for a small number of transmit antennas and low feedback overhead. In order to scale informed transmitter schemes to emerging massive MIMO systems with a large number of transmit antennas at the base station, one common approach is to employ time division duplexing (TDD) and to exploit the implicit feedback obtained from channel reciprocity. However, most existing cellular deployments are based on frequency division duplexing (FDD), hence it is of great interest to explore backwards compatible massive MIMO upgrades of such systems. For a fixed feedback rate per antenna, the number of codewords for quantizing the channel grows exponentially with the number of antennas, hence generating feedback based on look-up from a standard vector quantized codebook does not scale. In this paper, we propose noncoherent trellis-coded quantization (NTCQ), whose encoding complexity scales linearly with the number of antennas. The approach exploits the duality between source encoding in a Grassmannian manifold and noncoherent sequence detection. Furthermore, since noncoherent detection can be realized near-optimally using a bank of coherent detectors, we obtain a low-complexity implementation of NTCQ encoding using an off-the-shelf Viterbi algorithm applied to standard trellis coded quantization. We also develop advanced NTCQ schemes which utilize various channel properties such as temporal/spatial correlations. Simulation results show the proposed NTCQ and its extensions can achieve near-optimal performance with moderate complexity and feedback overhead.
I. INTRODUCTION
Massive MIMO needs accurate CSI, but FDD feedback schemes based on vector-quantized codebooks become computationally impractical as antenna dimensions grow. The paper proposes NTCQ, mapping Grassmannian quantization to noncoherent detection and using trellis-based encoding to achieve scalable CSI quantization with adaptive extensions.
- Accurate CSI provides beamforming gains in SU-MIMO and multiplexing gains in MU-MIMO, while inaccurate CSI causes massive MU-MIMO sum-rate saturation.
- FDD massive MIMO requires scalable CSI feedback because TDD reciprocity-based approaches do not directly address predominantly FDD cellular deployments.
- Fixed-feedback-rate vector quantization requires exponentially many codewords as antenna count grows, making conventional codebook search computationally infeasible.
- NTCQ connects Grassmannian source coding with noncoherent sequence detection and uses TCQ with Viterbi decoding to avoid look-up-based codebook bottlenecks.
- NTCQ achieves near-optimal beamforming gain with encoding complexity linear in channel dimension, while adaptive variants exploit temporal and spatial channel correlations.
- NTCQ can adapt antenna dimension, CSI accuracy, and feedback overhead by changing code length, constellation size, and coded modulation.
B. Feedback Overhead
Feedback overhead must scale with antenna dimension to preserve beamforming or multiplexing performance, but codebook-based search then becomes computationally infeasible. The paper motivates NTCQ as a framework combining manageable overhead with efficient encoding.
- For SU-MISO RVQ, maintaining a fixed normalized beamforming-gain loss requires feedback overhead proportional to the number of transmit antennas.
- For MU-MIMO zero-forcing beamforming, achieving full multiplexing gain requires per-user feedback scaling linearly with SNR in dB and the number of transmit antennas.
- Although linear feedback growth can be acceptable, VQ search complexity grows exponentially with antenna count as O(Mt2BMt), making it infeasible for massive MIMO.
- NTCQ uses the analogy between beamforming-vector search and noncoherent sequence detection to obtain near-optimal CSI quantization with efficient computation.
C. Efficient Grassmannian Encoding using Euclidean Metrics
The paper shows that noncoherent Grassmannian channel quantization is equivalent to noncoherent sequence detection, enabling efficient Euclidean-metric encoding. Euclidean codebooks can attain asymptotically optimal beamforming performance under suitable coding conditions.
- Grassmannian encoding: Noncoherent sequence detection and MISO beamforming quantization are equivalent after optimizing over unknown amplitude and phase.The optimal codeword maximizes normalized channel alignment, while noncoherent detection minimizes an equivalent distance metric.
- Grassmannian encoding: A bank of coherent demodulators over discrete amplitude and phase values provides near-optimal noncoherent detection with complexity linear in Mt.Optimal noncoherent detection has quadratic complexity in Mt, whereas a small number of parallel coherent demodulators is sufficient for excellent quantization performance.
- Asymptotic performance: Euclidean distance quantization with a rich rotationally invariant Gaussian codebook is asymptotically optimal for i.i.d. Rayleigh channels at fixed bits per antenna.Its normalized beamforming-gain loss is asymptotically the same as that of the RVQ benchmark.
- Asymptotic performance: Nontrivial channel codes are required to approach optimal quantization performance; uncoded constellations do not achieve it.Constellation expansion in NTCQ is therefore needed for near-optimal performance.
- Finite-dimensional codebooks: Euclidean codebooks have favorable chordal-distance behavior despite optimizing a different metric from Grassmannian line-packing codebooks.For Mt = 8, simulations show the optimized Euclidean codebook has larger minimum chordal distance than averaged RVQ codebooks for all Btot values.
A. Euclidean Distance Codebook Design
NTCQ builds its Euclidean codebook from trellis-coded modulation, using a trellis search to select quantized channel entries efficiently. The resulting code structure supports Euclidean optimization while preserving useful chordal-distance properties.
- Codebook construction: NTCQ uses trellis-coded quantization to map source encoding onto trellis-coded modulation designed for coherent AWGN communication.TCM combines convolutional coding and modulation to maximize minimum Euclidean distance between codewords.
- Codebook construction: The trellis-generated codebook is optimized in Euclidean distance and is therefore expected to have favorable chordal-distance properties.The paper uses this Euclidean code structure as the basis for Grassmannian channel quantization.
- TCQ operation: TCQ quantizes a source vector by minimizing Euclidean distortion over trellis paths and uses the Viterbi algorithm for efficient search.The selected binary path is reconstructed through a convolutional encoder into a quantized constellation vector.
- NTCQ encoding: NTCQ performs Kα · Kθ parallel Viterbi searches for discrete amplitude and phase values, then selects the path with the best metric.For PSK constellations, α can be fixed to 1 because all candidate points have equal magnitude.
B. NTCQ with 8PSK (2 bits/entry)
The 8PSK NTCQ implementation uses an eight-state trellis and parallel phase searches to quantize two bits per channel entry. The best trellis path determines the feedback sequence and reconstructed beamforming vector.
- 8PSK implementation: The 8PSK implementation uses a rate 2/3 convolutional code and normalizes constellation points by the number of transmit antennas.Each trellis stage selects an entry of a candidate beamforming vector.
- Trellis search: Each trellis path corresponds to a unique candidate beamforming vector, and the selected path minimizes the Euclidean path metric.The output feedback is the binary sequence associated with the best path.
- Per-entry quantization: With 8PSK, each channel entry is effectively quantized using one of two QPSK subconstellations, yielding 2 bits per entry.The source constellation remains 8PSK, but trellis transitions restrict the per-entry choices.
- Phase search: The Viterbi path metric is computed recursively and in parallel for each candidate phase θ.The phase and path jointly minimizing the metric determine the output beamforming vector.
- Feedback overhead: Minimizing over phase increases encoding complexity but not feedback overhead because θbest is unnecessary for beamforming reconstruction.NTCQ nevertheless adds log2 S feedback bits to identify the trellis starting state.
C. NTCQ with 16QAM (3 bits/entry)
The 16QAM implementation follows the 8PSK procedure but incorporates α into path-metric computation. Its constellation points are labeled with binary sequences, and starting-state information requires additional feedback.
- The 16QAM constellation points used in NTCQ are labeled with binary sequences.
- 16QAM NTCQ follows the 8PSK procedure, with α included in path-metric computation.
- Additional log2 S feedback bits indicate the starting state of the best trellis path.
D. Complexity
NTCQ uses Viterbi trellis search, with complexity scaling linearly in the number of transmit antennas. Several variations and correlation-aware extensions trade feedback overhead or performance against implementation choices.
- Complexity: NTCQ quantizes one channel entry per trellis transition using one of 2^B constellation points across S states and M_t transitions.The resulting Viterbi complexity is O(2^B S M_t).
- Complexity: The overall NTCQ complexity is O(K_θK_α2^B S M_t), while small K_θ and K_α cause minimal performance degradation.As M_t grows, Theorem 1 permits K_θ → 1 and K_α → 1 without performance loss.
- Variations: Fixing the trellis starting state removes log2 S feedback bits, yielding exactly B M_t feedback bits at a small performance cost.The loss becomes negligible as M_t increases.
- Variations: Other NTCQ variations fix the first c_opt entry or use a tail-biting convolutional code.These are alternative trellis constructions described alongside the fixed-starting-state variation.
- Correlation-aware extensions: Advanced NTCQ schemes exploit temporal or spatial channel correlations to improve performance or reduce feedback overhead.Temporal correlation is modeled using a first-order Gauss-Markov process, with η representing correlation between consecutive channel entries.
B. Adaptive Scheme for Spatially Correlated Channels
The spatially adaptive NTCQ scheme addresses spatially correlated channels by decorrelating the channel before quantization and reconstructing it afterward. This separates correlation exploitation from quantization while retaining the performance gain of codebook skewing.
- Channel model: Closely spaced transmit antennas can produce spatially correlated channels modeled through a full-rank correlation matrix R.Prior codebook-skewing methods rotate and normalize codewords with respect to R to quantize the dominant eigenspace.
- Adaptive quantization: The transmitter and receiver are assumed to know R, enabling the receiver to decorrelate h[k] with R^-1.The decorrelated channel is denoted h_w[k].
- Adaptive quantization: The receiver quantizes h_w[k] with NTCQ, feeds back its quantized version, and the transmitter reconstructs the beamforming vector.This applies NTCQ after spatial decorrelation rather than directly to the correlated channel.
- Result: The procedure decouples spatial-correlation exploitation from quantization while providing the same performance gain as standard fixed-codebook skewing.The comparison is stated for the same spatially correlated-channel setting.
V. PERFORMANCE EVALUATION
The evaluation uses Monte Carlo simulations to assess NTCQ and its variants across independent, temporally correlated, and spatially correlated channels. Average beamforming gain in dB is the stated performance metric.
- Simulation setup: Monte Carlo simulations evaluate NTCQ, differential NTCQ, and spatially adaptive NTCQ.The schemes are tested in i.i.d., temporally correlated, and spatially correlated channel scenarios.
- Simulation setup: The evaluation covers i.i.d. channels, temporally correlated channels, and spatially correlated channels.These scenarios correspond to the original and correlation-aware NTCQ schemes.
- Metric: Average beamforming gain in dB is used as the performance metric, with expectation taken over h.The metric is defined for the simulation evaluation.
A. i.i.d. Rayleigh fading Channels
In i.i.d. Rayleigh fading, NTCQ and its variation approach benchmark beamforming performance while retaining favorable feedback and complexity properties. Differential NTCQ can also approach unquantized performance with low per-entry quantization and limited delay sensitivity.
- i.i.d. Rayleigh fading performance: The variation of NTCQ gives strictly lower JdB avg than original NTCQ when quantization levels for θk and αk are varied.With 1 bit/entry, Kθ = 4 is sufficient for near-maximal performance; fixing αk = 1 causes no loss at 3 bits/entry.
- i.i.d. Rayleigh fading performance: NTCQ closely matches the RVQ benchmark for all tested feedback levels and antenna counts, demonstrating near-optimal quantization performance.The theorem-based benchmark also closely approximates NTCQ’s JdB avg.
- i.i.d. Rayleigh fading performance: The variation of NTCQ achieves better JdB avg than PSK-SVQ across feedback levels and antenna counts, with a 0.25 to 1dB coding gain.NTCQ uses 2B+1 constellation points, whereas PSK-SVQ uses 2B points with B bits per entry.
- Differential feedback: Differential NTCQ with 1 bit/entry achieves almost the same performance as unquantized beamforming in the tested temporally correlated setting.This enables reducing feedback from 2 or 3 bits/entry when feedback overhead can adapt over time.
C. Spatially Correlated Channels
Spatially adaptive NTCQ exploits channel correlation to approach unquantized beamforming as spatial correlation increases, while preserving the original feedback overhead. The broader NTCQ design uses efficient encoding and supports channel-dependent adaptations.
- Spatially adaptive quantization: Spatially adaptive NTCQ performance becomes closer to unquantized beamforming as the dominant eigenvalue λ1 increases.The spatial correlation is generated with a Kronecker model, where λ1 = 1 corresponds to i.i.d. channels and larger λ1 indicates stronger correlation.
- Spatially adaptive quantization: The adaptive spatial scheme demonstrates effectiveness with the same feedback overhead as original NTCQ.The comparison varies λ1 for Mt = 20 in spatially correlated channels.
- NTCQ design: NTCQ maps Grassmann-manifold quantization to noncoherent sequence detection and implements encoding through Viterbi-based Euclidean-space quantization.A bank of coherent detectors provides a near-optimal implementation.
- NTCQ design: NTCQ has linear encoding complexity in the number of transmit antennas and can accommodate additional channel coefficients by increasing blocklength.The method can also be modified for temporal and spatial channel correlations.
- Scope and limitations: Feedback overhead would typically prevent scaling NTCQ to an indefinitely large number of antennas, although 32 to 64 antennas may be reasonable for initial deployments.The paper identifies scalable channel-sounding schemes as another requirement for practical FDD massive MIMO.