PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning for Distributed Channel Feedback and Multiuser Precoding in FDD Massive MIMO

Foad Sohrabi, Kareem M. Attiah, Wei Yu

arXiv:2007.06512v2cs.ITeess.SP

TL;DR

FDD massive MIMO needs effective channel feedback because downlink CSI must be estimated and returned without channel reciprocity. The paper formulates this as distributed source coding and jointly trains user- and BS-side DNNs for pilots, feedback, and precoding. The proposed strategy achieves strong performance with limited feedback, including about 80% of full-CSIT sum rate using about 7 feedback bits versus more than 20 for separate source coding.

  • Problem

    FDD massive MIMO requires downlink channel estimation and feedback, while conventional independent user CSI processing does not exploit the distributed precoding objective.

  • Method

    The paper jointly trains user-side and BS-side DNNs to design pilots, channel estimation, distributed compression, rate-limited feedback, and multiuser precoding.

  • Results

    At 12 bits/s/Hz sum rate, the DSC strategy requires about 7 feedback bits versus more than 20 bits for separate source coding; about 20 feedback bits achieves almost 90% of perfect-CSIT ZF sum rate.

  • Takeaways & Limitations

    The end-to-end DSC strategy can provide strong FDD precoding performance with short pilots and very limited feedback overhead.

Abstract

from arXiv · show

This paper shows that deep neural network (DNN) can be used for efficient and distributed channel estimation, quantization, feedback, and downlink multiuser precoding for a frequency-division duplex massive multiple-input multiple-output system in which a base station (BS) serves multiple mobile users, but with rate-limited feedback from the users to the BS. A key observation is that the multiuser channel estimation and feedback problem can be thought of as a distributed source coding problem. In contrast to the traditional approach where the channel state information (CSI) is estimated and quantized at each user independently, this paper shows that a joint design of pilots and a new DNN architecture, which maps the received pilots directly into feedback bits at the user side then maps the feedback bits from all the users directly into the precoding matrix at the BS, can significantly improve the overall performance. This paper further proposes robust design strategies with respect to channel parameters and also a generalizable DNN architecture for varying number of users and number of feedback bits. Numerical results show that the DNN-based approach with short pilot sequences and very limited feedback overhead can already approach the performance of conventional linear precoding schemes with full CSI.

I. INTRODUCTION

The paper frames limited-feedback FDD massive MIMO precoding as distributed source coding and uses DNNs to jointly optimize channel compression, pilots, feedback, and precoding. It targets efficient end-to-end design while addressing training and generalizability across system conditions.

  • FDD massive MIMO requires downlink channel estimation and feedback because uplink-downlink channel reciprocity cannot be assumed.
  • Optimized distributed channel compression and joint feedback–precoding design are reported to outperform conventional independent CSI estimation and feedback.
  • The end-to-end precoding system can be viewed as distributed source coding that computes a precoding matrix from users’ independent channels under finite feedback rates.
  • DNNs support end-to-end optimization of distributed source coding components, with parallelizable inference helping manage computational burden.
  • The proposed design jointly optimizes downlink pilots, user-side channel estimation and quantization, rate-limited feedback, and BS-side multiuser precoding.
  • The paper addresses generalizability by widening training parameter ranges and modifying architecture and training for varying system dimensions, including users and feedback rates.

B. Related Work

Prior limited-feedback approaches either compress each user’s CSI separately or use codebooks, while this paper formulates end-to-end multiuser FDD precoding as a distributed source-coding problem. The formulation jointly optimizes pilots, distributed feedback, and BS precoding to maximize spectral efficiency under limited feedback.

  • Prior approaches: Existing CSI-feedback schemes exploit spatial or temporal correlation, then reconstruct channels and apply conventional MRT or ZF precoding.
  • Prior approaches: Codebook-based methods select precoding directions from a codebook, but interference management requires sufficiently large feedback, whereas p=1 cannot manage interference.
  • Prior approaches: Recent DNN-based FDD methods mainly address single-user settings or BS-side CSI reconstruction assuming perfect user CSI.
  • Paper positioning: The paper addresses the multiuser case by jointly training pilots and system parameters while accounting for CSI estimation error to improve downlink spectral efficiency.
  • System formulation: The signal model uses a BS with M antennas serving K<M single-antenna users through linear precoding, with user rates determined by desired-signal power, interference, and noise.
  • System formulation: The optimization jointly selects downlink pilots, user feedback functions, and the BS precoding function, viewing distributed user processing and centralized precoding as DSC.

B. Channel Model and Conventional Approaches

The paper considers sparse mmWave FDD massive MIMO channels and argues that conventional independent channel quantization can be improved through distributed source coding and end-to-end neural design.

  • Channel model: The considered FDD massive MIMO system operates in a limited-scattering mmWave environment where channel sparsity can support feedback design.The channel is modeled using a limited number of propagation paths, with path gains, angles of departure, and array responses describing the channel.
  • Conventional approach: Conventional feedback estimates each user’s sparse channel parameters before quantizing and transmitting them to the base station.The process uses sparse recovery to estimate channel parameters, followed by quantization and base-station processing.
  • Motivation: The paper targets excellent performance with short training sequences and small amounts of limited feedback information.It identifies independent per-user quantization as an opportunity for improvement because the base station jointly uses users’ estimated channels for precoding.
  • Distributed source coding: Distributed source coding can improve on independent channel quantization because users’ channel information is jointly used to compute the downlink precoding matrix.The proposed framework maps each user’s received pilots directly to feedback bits and maps all users’ feedback bits directly to the precoding matrix.
  • Neural design: The proposed neural architecture also incorporates downlink pilot design by modeling the pilot matrix as a linear neural-network layer.The resulting architecture represents an end-to-end FDD downlink system.

III. FDD DOWNLINK PRECODING SYSTEM DESIGN USING DEEP LEARNING

The paper models the FDD downlink system as a neural network spanning training, feedback, and data transmission, enabling joint design of pilots, user feedback, and base-station precoding.

  • System design: The proposed design uses neural networks to represent downlink training, uplink feedback, and downlink data transmission in an FDD system.These are organized into downlink training and uplink feedback, followed by downlink data transmission.
  • System design: The training objective jointly designs the downlink pilots, each user’s feedback function, and the base station’s precoding function.The paper denotes these components by the pilot matrix, user-side mappings, and downlink precoding scheme.

A. DNN Representation of DSC Strategy

The architecture represents the FDD downlink precoding pipeline as neural-network operations, beginning with pilot transmission and user observations before feedback generation.

  • Architecture: The neural representation models downlink pilot transmission, user operations during training, and base-station operations during data transmission.This modeling connects the physical FDD downlink system to an end-to-end neural architecture.
  • Downlink training: The base station sends a pilot matrix over L downlink transmissions, and user k receives a channel-dependent noisy observation.The received signal is represented as the channel response to the pilot matrix plus additive noise.
  • Downlink training: The pilot matrix is modeled as the weight matrix of a fully connected linear layer followed by additive zero-mean noise.This gives the neural representation a direct correspondence to pilot transmission.
  • Downlink training: The pilot matrix is constrained so that each transmission satisfies the base station’s per-transmission power limit P.The constraint is used to model the physical transmit-power constraint rather than only as a machine-learning regularizer.
  • User operation: Each user’s objective is to summarize its received observation into B information bits for downlink precoding.The user-side mapping is implemented by an R-layer fully connected DNN.

2) Uplink Feedback:

The uplink feedback and precoding stages use neural mappings from user observations to bipolar feedback bits and from all users’ bits to a normalized precoding matrix.

  • Uplink Feedback: The user-side DNN generates bipolar feedback bits using trainable layers, ReLU hidden activations, and a sign activation at the output.The real representation of each received observation is used as the DNN input.
  • Uplink Feedback: The user-side network’s final layer has B outputs, while hidden-layer dimensions are determined by the numbers of neurons ℓr.The listed weight and bias dimensions specify the layer shapes from the real-valued input through the feedback output.
  • Downlink Precoding Design: The base station maps the collected feedback bits from all K users through a T-layer dense DNN to produce the precoding vectors.The aggregate feedback vector q is the input to the base-station network.
  • Downlink Precoding Design: A normalization layer is used at the base-station network’s final layer to satisfy the total power constraint.The other base-station layers use ReLU activations.
  • Architecture: The overall architecture comprises downlink pilot training, uplink feedback, and downlink precoding design in an end-to-end K-user FDD system.The trainable parameters include the pilot matrix and the user-side and base-station DNN parameters.
  • Architecture: The architecture assumes no explicit channel-parameter information at the base station, which instead learns channel statistics from training data.Potential inputs such as sparsity, angular information, and pathloss are not used in this setup.

B. DNN Training with a Hidden Binary Layer

The proposed DNN is trained end to end to maximize sum rate while jointly learning pilots, user feedback schemes, and BS precoding. Because binary thresholding blocks ordinary back-propagation, training uses a sigmoid-adjusted straight-through estimator with annealing.

  • Training objective: The training objective maximizes expected sum rate over channel and downlink-training noise distributions.The trainable parameter space includes the pilot matrix, user feedback schemes, and BS precoding scheme.
  • Training procedure: Stochastic gradient descent approximates the expectation with empirical training samples and updates all trainable parameters through back-propagation.The partial derivatives are computed through the chain rule in the directed computation graph.
  • Binary layer: Binary thresholding prevents direct conventional back-propagation because its derivative is zero almost everywhere.The issue occurs for neural layers preceding the binary layer.
  • Binary layer: The method replaces the binary-layer derivative during back-propagation with a sigmoid-adjusted straight-through estimator.The sigmoid-adjusted estimator uses the gradient of 2 sigm(u) − 1.
  • Binary layer: Slope annealing gradually increases the sigmoid slope during training, and the paper applies this estimator to the user-side binary layer.The annealing factor is increased across epochs during back-propagation.

IV. GENERALIZABILITY

The paper examines whether its DNN remains useful when channel and system parameters differ between training and testing. It also addresses varying feedback rates by changing quantization levels without retraining the network architecture.

  • Generalizability: Generalizability is evaluated by training under one set of system parameters and testing under another.The paper distinguishes parameters that change the input distribution from other system parameters.
  • Feedback-rate variation: Varying the feedback-rate limit B would otherwise require a separate DNN because it changes the user-output and BS-input dimensions.This dependence is identified as impractical when feedback may vary.
  • Feedback-rate variation: The BS-side DNN maps K×S quantized signals to the precoding matrix while maximizing average sum rate.Each user sends Q-bit quantized versions of S soft binary signals, so B = S × Q.

C. Towards Generalizability for K

For varying user counts, the paper reuses a common user-side encoder when user channels are i.i.d. and retrains only the BS-side DNN. The implementation trains the end-to-end design with SGD-based optimization, validation monitoring, and fixed channel and noise distributions.

  • Towards Generalizability for K: When users’ channel distributions are i.i.d., one common user-side DNN can serve systems with different numbers of users.The encoding parameters are first learned in a single-user system and then shared across users.
  • Towards Generalizability for K: Only the BS-side DNN must be separately trained and stored for different total numbers of users.This avoids training different user-side DNNs for each network size.
  • Implementation: The proposed network is implemented with TensorFlow and Keras and trained according to the paper’s SGD-based procedure.The implementation section provides the architecture and training details.
  • Implementation: The default architecture uses four fully connected layers at both user and BS sides, with batch normalization before each dense layer.User-side hidden widths are [1024, 512, 256, B], while BS-side widths end at 2MK.
  • Pilot design: The pilot matrix is optimized as a trainable TensorFlow variable while enforcing a per-pilot power constraint.Its initialization uses an i.i.d. complex Gaussian distribution, followed by normalization during optimization.
  • Training procedure: Training uses mini-batches, gradient updates with annealed sigmoid-adjusted straight-through approximation, and validation-based stopping.The procedure tracks the best validation rate and updates annealing and learning rates over epochs.
  • Training assumptions: The experiments fix channel and downlink-training noise distributions to generate arbitrarily many training samples.The minimum training-set size needed for reasonable performance is left for future work.
  • Training procedure: Generalization is monitored on a holdout validation set of Nv = 10^4 samples, and the best-performing parameters are retained.Training terminates after no improvement for a specified number of epochs, such as Nep,max = 300.

VI. NUMERICAL RESULTS

The numerical study evaluates the proposed deep learning precoding framework for FDD systems with limited feedback against conventional MRT and ZF precoding. The baselines span settings with perfect CSI, limited feedback, and unavailable prior CSI.

  • Numerical evaluation: The experiments assess the proposed deep learning-based precoding method for FDD systems with limited feedback.The evaluation compares its performance with conventional MRT and ZF precoding.
  • Baselines: MRT and ZF are evaluated under four CSI and feedback configurations, including perfect CSI and limited-feedback links.The settings also include no prior CSI with either infinite-capacity or limited-feedback links.
  • Baselines: With full CSI at the BS, conventional MRT and ZF use precoding matrices determined by the instantaneous channel H.The constants γMRT and γZF enforce the transmit-power constraint.

2) MRT/ZF with Full CSI at Receiver (CSIR) and Finite-Capacity Feedback:

The proposed DNN jointly learns channel feedback and precoding, outperforming separate channel-recovery baselines under limited pilot and feedback resources. Its sum-rate performance approaches ZF with perfect CSIT as pilots and feedback increase.

  • System design: The proposed system jointly maps downlink pilots and distributed feedback into a precoder, rather than separately estimating and quantizing each channel.The baseline separately estimates channel parameters, quantizes them, reconstructs channels at the BS, and then computes MRT/ZF precoders.
  • Short-pilot performance: With L = 8 and M = 64, the proposed method outperforms perfect-CSI MRT using only a 15-bit feedback link.MRT maximizes useful-signal power but does not explicitly manage interference.
  • Short-pilot performance: For L = 8, direct sum-rate optimization outperforms DNN channel recovery followed by MRT/ZF at the same feedback capacity.It also outperforms conventional OMP-based sparse recovery and limited-feedback MRT/ZF baselines.
  • Feedback efficiency: At 12 bits/s/Hz, the DSC strategy requires about 7 feedback bits versus more than 20 bits for DNN-based separate source coding.The 12 bits/s/Hz operating point is about 80% of the sum rate achieved with full CSIT.
  • Long-pilot performance: With L = M = 64, the proposed method outperforms other limited-feedback baselines and approaches perfect-CSIT ZF at B ≥50.The longer pilot enables near-perfect channel recovery at the user side.

C. Towards Generalizability in Lp

The paper evaluates robustness to mismatch between training and test channel-path distributions. Training across a wider range of path counts improves robustness, especially when pilot resources are limited.

  • Distribution mismatch: Testing a DNN trained only with Lp = 2 on other path counts degrades performance because training and test distributions mismatch.The degradation is more severe for the short-pilot setting L = 8.
  • Distribution mismatch: When L = 64 is not pilot-limited, the trained DNN can potentially handle a wider range of channel distributions.This contrasts with the more severe mismatch degradation observed at L = 8.
  • Robust training: Training with Lp ∈{2, 3, . . . , 8} produces more robust performance across channel-path distributions.The proposed DNN still outperforms limited-feedback baselines under training–testing mismatch.

D. Towards Generalizability in B

The paper develops a common DNN that can operate across feedback capacities and evaluates the proposed approach as the number of served users increases. These extensions improve generalizability but introduce constraints and greater training cost.

  • Quantization design: The proposed architecture uses a modified neural network whose tanh-layer outputs are quantized into regions with corresponding representation points.The illustrated system uses M = 64, K = 2, L = 8, and an optimal 3-bit quantizer.
  • Varying feedback capacity: A common DNN can operate at different feedback capacities, avoiding separate training for each value of B.The two-step training approach improves generalizability with respect to feedback capacity, but supports only integer values of S.
  • Varying number of users: For larger user counts, the BS decoding DNN must process an input of dimension KB and may require increased capacity.The experiment uses L = 8 and B = 30, with AoDs uniformly distributed over (−60°, +60°).
  • Varying number of users: With limited downlink training resources, the proposed framework achieves higher sum rate than the conventional baselines as the number of users increases.The comparison is reported for a K-user FDD system with M = 64, B = 30, and L = 8.
  • Overall design: The end-to-end system jointly designs downlink training, user-side channel feedback, and BS-side multiuser precoding for limited-feedback FDD massive MIMO.The approach bypasses explicit channel estimation and is especially effective when pilot length or feedback capacity is very limited.
Loading 2007.06512v2…