PaperlayerView exact revision on arXiv

Source-linked AI summary

Model-Driven Deep Learning Based Channel Estimation and Feedback for Millimeter-Wave Massive Hybrid MIMO Systems

Xisuo Ma, Zhen Gao, Feifei Gao, Marco Di Renzo

arXiv:2104.11052v3cs.ITcs.AIcs.LGeess.SP

TL;DR

Wideband mmWave massive hybrid MIMO faces substantial channel-estimation and feedback overhead because channels are high-dimensional and RF chains are limited. The paper combines structured sparsity with learned MMV-LAMP networks and jointly trained pilots or phase-shift networks for TDD and FDD operation. Simulations report improved estimation and feedback performance, including a case where 25% subcarrier feedback outperforms a 50% baseline.

  • Problem

    High-dimensional mmWave massive hybrid MIMO channels must be estimated and fed back despite limited RF chains and prohibitively high pilot and feedback overhead.

  • Method

    The paper uses MDDL networks that combine a priori channel sparsity with learned parameters, jointly training pilots or phase-shift networks and using MMV-LAMP reconstruction for TDD estimation and FDD feedback.

  • Results

    The proposed scheme outperforms conventional channel-estimation and feedback schemes; with feedback compression ratio ρ = 0.25, it outperforms SOMP with ρ = 0.5.

  • Takeaways & Limitations

    Exploiting angle-delay sparsity and learning channel-environment characteristics supports reduced-overhead estimation and feedback in the studied wideband mmWave hybrid MIMO settings.

Abstract

from arXiv · show

This paper proposes a model-driven deep learning (MDDL)-based channel estimation and feedback scheme for wideband millimeter-wave (mmWave) massive hybrid multiple-input multiple-output (MIMO) systems, where the angle-delay domain channels' sparsity is exploited for reducing the overhead. Firstly, we consider the uplink channel estimation for time-division duplexing systems. To reduce the uplink pilot overhead for estimating the high-dimensional channels from a limited number of radio frequency (RF) chains at the base station (BS), we propose to jointly train the phase shift network and the channel estimator as an auto-encoder. Particularly, by exploiting the channels' structured sparsity from an a priori model and learning the integrated trainable parameters from the data samples, the proposed multiple-measurement-vectors learned approximate message passing (MMV-LAMP) network with the devised redundant dictionary can jointly recover multiple subcarriers' channels with significantly enhanced performance. Moreover, we consider the downlink channel estimation and feedback for frequency-division duplexing systems. Similarly, the pilots at the BS and channel estimator at the users can be jointly trained as an encoder and a decoder, respectively. Besides, to further reduce the channel feedback overhead, only the received pilots on part of the subcarriers are fed back to the BS, which can exploit the MMV-LAMP network to reconstruct the spatial-frequency channel matrix. Numerical results show that the proposed MDDL-based channel estimation and feedback scheme outperforms the state-of-the-art approaches.

I. INTRODUCTION

The paper targets high pilot and feedback overhead in wideband mmWave massive hybrid MIMO, where limited RF chains complicate high-dimensional CSI acquisition. It proposes MDDL networks that combine channel sparsity models with learned parameters for estimation and feedback.

  • I. INTRODUCTION: High-dimensional CSI estimation and feedback create excessive overhead in mmWave massive hybrid MIMO, especially with limited RF chains and FDD feedback.
  • I. INTRODUCTION: Existing low-overhead methods exploit channel sparsity, but several prior schemes target low-frequency fully-digital massive MIMO rather than mmWave hybrid systems.
  • I. INTRODUCTION: MDDL combines physical-model knowledge with deep learning, retaining structured sparsity and fewer trainable parameters while learning from data.
  • C. Our Contributions: The proposed scheme jointly trains pilots or RF phase-shift networks with channel estimators, using MMV-LAMP reconstruction networks and redundant dictionaries for sparse multi-subcarrier recovery.
  • C. Our Contributions: The framework also uses a feedback reconstruction network to exploit delay-domain sparsity and reconstruct channels from compressed feedback.
  • C. Our Contributions: Fixed-scattering experiments evaluate whether jointly optimized pilots and estimators can match channel-environment characteristics.

B. Downlink Channel Estimation and Feedback for FDD Systems

For FDD downlink estimation, the base station transmits RF and baseband pilots across OFDM subcarriers, and users collect noisy pilot observations. These observations are post-processed into spatial-frequency channel measurements for subsequent estimation and feedback.

  • B. Downlink Channel Estimation and Feedback for FDD Systems: The BS transmits downlink pilot signals formed from an RF pilot vector and a baseband pilot symbol on each subcarrier.
  • B. Downlink Channel Estimation and Feedback for FDD Systems: The received pilot signal equals the downlink channel applied to the transmitted pilot, plus complex noise.
  • B. Downlink Channel Estimation and Feedback for FDD Systems: Assuming unit-energy baseband pilots, the received observations can be post-processed to remove the pilot symbol and retain channel-dependent measurements.
  • B. Downlink Channel Estimation and Feedback for FDD Systems: Collecting observations over Q time slots forms aggregate measurements for each of K subcarriers and the downlink spatial-frequency channel matrix.

C. Channel Model

The paper models wideband mmWave channels through angle- and delay-related sparsity, then develops an MMV-LAMP estimator that exploits structured sparsity and learned parameters from data.

  • Channel Model: The channel model represents wideband downlink channels across propagation paths, with each path characterized by gain, delay, and angle of departure.
  • Channel Model: The base-station uniform linear array uses a steering vector determined by carrier wavelength and adjacent antenna spacing.
  • Transmit Frame Structure: The proposed frame uses CP-OFDM and divides T time slots into Q pilot slots and T − Q data-transmission slots.
  • MMV-LAMP Network: The network processes multiple measurements whose sparse matrix columns share common sparsity, supporting joint recovery across subcarriers.
  • MMV-LAMP Network: The developed MMV-LAMP network exploits structured sparsity from an a priori model while learning trainable parameters through unfolded AMP iterations.
  • MMV-LAMP Network: A redundant dictionary with finer angular resolution is integrated to mitigate power leakage caused by mismatch between continuous angles and discrete dictionaries.

C. MMV-LAMP Network Based Uplink Channel Estimation

The uplink estimator jointly trains a combining network and an MMV-LAMP channel-recovery network as an end-to-end auto-encoder, using trainable phase parameters and a redundant dictionary.

  • Uplink Channel Estimation: The proposed uplink solution combines a CCN encoder with an MMV-LAMP-based CRN decoder for channel estimation at the base station.
  • Uplink Channel Estimation: The CCN maps the uplink spatial-frequency channel matrix to noiseless received pilot signals using the combining matrix F_UL.
  • Uplink Channel Estimation: The CRN contains T layers with shared trainable parameters and outputs an estimated angle-frequency channel matrix.
  • Uplink Channel Estimation: The overall parameters {Ξ_UL, B_UL, θ_UL} are jointly trained offline in an end-to-end auto-encoder framework.
  • Uplink Channel Estimation: Multiplying the estimated angle-frequency representation by a devised redundant dictionary produces the estimated spatial-frequency channel matrix.
  • Fully-Connected CCN: The combining matrix is implemented through a fully connected CCN whose trainable real-valued parameters represent the phase values of the constant-modulus RF phase-shift network.

2) CRN Based on MMV-LAMP Network:

The CRN uses an MMV-LAMP network and a redundant angle-domain dictionary to recover wideband channel matrices from low-dimensional measurements by exploiting common sparsity across subcarriers.

  • A redundant dictionary with finer angular resolution transforms the spatial-frequency channel matrix into an angle-frequency representation, mitigating power-leakage effects on sparsity.
  • The angle-frequency channel columns are sparse and share common sparsity across subcarriers, yielding an MMV sparse matrix recovery problem.
  • MMV-LAMP replaces conventional sparse-recovery procedures whose greedy variants are reported to provide unsatisfactory channel-estimation accuracy.
  • The proposed T-layer MMV-LAMP network reconstructs the high-dimensional angle-frequency channel matrix from low-dimensional received signals.

3) Learning Strategy:

The learning strategy jointly trains the uplink compression and reconstruction networks layer by layer, using NMSE loss while carrying learned parameters forward between layers.

  • The CCN encoder and CRN decoder are jointly trained through a layer-by-layer learning strategy inspired by auto-encoders.
  • Each layer minimizes an NMSE loss between the target spatial-frequency channel matrix and the MMV-LAMP estimate.
  • The CCN output is used within the uplink channel-estimation pipeline, linking learned compression with iterative MMV-LAMP reconstruction.
  • Parameters learned at one layer initialize the next layer, while the current layer jointly optimizes the CCN and CRN parameters.

IV. MDDL-BASED FDD DOWNLINK CHANNEL ESTIMATION AND FEEDBACK

For FDD downlink estimation and feedback, the method extends the uplink MMV-LAMP approach to users and adds feedback-specific networks that exploit channel sparsity.

  • The FDD solution contains a CCN at the BS, an FCN at users, and an FCRN at the BS composed of an FRSN and the shared CRN.
  • Because the downlink and uplink measurement formulations are similar, the uplink MMV-LAMP channel estimator can be adapted for downlink estimation at users.
  • The downlink CCN uses trainable phase parameters for the pilot beamforming matrix, while the user-side CRN applies the corresponding downlink inputs and learned parameters.

DLFDL + NT

The feedback pipeline compresses received pilots in the delay domain by selecting part of the subcarriers, then reconstructs the frequency-domain channel before final spatial-frequency recovery.

  • The two-step feedback design first feeds back pilots from only part of the subcarriers, then uses FRSN and CRN reconstruction at the BS.
  • The reconstructed frequency-domain channel is passed to the CRN so the BS can recover the spatial-frequency channel matrix.
  • A DFT transforms the frequency-spatial channel matrix into a delay-spatial matrix, exposing delay-domain sparsity for feedback compression.
  • The FRSN is an MMV-LAMP network that takes compressed feedback pilots as input and reconstructs the frequency-domain channel matrix.

2) Feedback Based Channel Reconstruction Network at BS:

The BS reconstructs the spatial-frequency channel from compressed feedback using a two-stage MDDL architecture: an FRSN first recovers frequency-domain channels, followed by a CRN that completes reconstruction.

  • Feedback Based Channel Reconstruction Network at BS:: The FCRN is composed of an FRSN and a CRN, with the FRSN using a shorter MMV-LAMP network to reconstruct the channel matrix.The FRSN uses T′ layers, whereas the CRN uses T layers.
  • Feedback Based Channel Reconstruction Network at BS:: The BS applies the FRSN to compressed feedback pilots and then passes its frequency-domain output to the CRN for spatial-frequency channel reconstruction.The FRSN and CRN are both based on MMV-LAMP processing.
  • Feedback Based Channel Reconstruction Network at BS:: The feedback reconstruction training optimizes the FRSN parameters layer by layer using batch data generated from the channel model.The trainable parameters are {B′, θ′}, and each layer minimizes its corresponding loss function.
  • Feedback Based Channel Reconstruction Network at BS:: The simulations evaluate the FDD downlink estimation and feedback mechanism using a 256-antenna BS, 4 RF chains, 64 subcarriers, and a 1024-point angle dictionary.The MMV-LAMP CRN has five layers, the FRSN has two layers, and training uses 5000 spatial-frequency channel samples.

B. MDDL-Based FDD Downlink Channel Estimation

The proposed MDDL channel estimator is evaluated against model-based and data-driven schemes across SNR, pilot overhead, channel conditions, and training choices, with results favoring the proposed design.

  • B. MDDL-Based FDD Downlink Channel Estimation: The proposed estimator outperforms competing schemes with smaller pilot overhead and improves NMSE especially in the low-SNR regime.The paper attributes this to a cascaded CCN and MMV-LAMP CRN that combines physical mechanisms, prior model knowledge, and deep learning.
  • B. MDDL-Based FDD Downlink Channel Estimation: At SNR=0dB and SNR=5dB, Q = 40 for the proposed scheme outperforms several methods using Q = 80, reducing pilot overhead by at least 50%.The comparison includes MMV-AMP, LAMP, and data-driven deep learning methods.
  • B. MDDL-Based FDD Downlink Channel Estimation: The estimator trained with L = 8 multipath components remains applicable to channels with L ≠ 8 without retraining the entire network.The reported robustness concerns variation in the number of multipath components.
  • B. MDDL-Based FDD Downlink Channel Estimation: A model trained with K = 64 subcarriers can be applied to K = 128, 256, and 512 by dividing subcarriers into 64-subcarrier groups.The paper identifies this as evidence of robustness to the number of subcarriers.
  • B. MDDL-Based FDD Downlink Channel Estimation: The redundant dictionary improves sparse channel estimation, while training with multi-carrier channel samples performs better than training with single-carrier samples.These effects are evaluated in the two parts of Fig. 13.

C. Channel Estimation Under Non-Ideal Hardware Constraints

The paper tests channel estimation under finite phase-shifter and ADC resolution, and evaluates feedback reconstruction when only part of the subcarriers is returned to the BS.

  • C. Channel Estimation Under Non-Ideal Hardware Constraints: With B_ps = 3-bit phase quantization, the estimator works well, while B_ps = 2 bits causes a small performance loss.The network architecture and training strategy remain unchanged during this test.
  • C. Channel Estimation Under Non-Ideal Hardware Constraints: The practical phase-shifter constraint is handled by quantizing continuously trained combining and beamforming coefficients after offline training.The paper notes that non-differential gradients prevent directly using Adam with quantized phase shifters.
  • C. Channel Estimation Under Non-Ideal Hardware Constraints: With B_adc = 2 or 3 quantization bits, the proposed estimator still works in the low-SNR regime despite degradation in MMV-AMP and SOMP baselines.The paper links this setting to non-ideal ADC operation at the receiver.
  • D. MDDL-Based FDD Downlink Channel Feedback: The user feeds back pilots from only K_c of K subcarriers, after which the BS uses FRSN and CRN to recover the spatial-frequency channel matrix.The feedback compression ratio is defined as ρ = K_c/K.
  • D. MDDL-Based FDD Downlink Channel Feedback: With ρ = 0.25 and K_c = 16, the proposed FCRN outperforms SOMP with ρ = 0.5 and K_c = 32.The comparison is reported for channel reconstruction NMSE versus SNR.

E. Channel Estimation Based on Fixed Scattering Environments

The fixed-scattering-environment experiments evaluate the proposed channel estimation and feedback scheme for TDD uplink and FDD downlink settings. The learned design improves performance while reducing pilot or feedback overhead through structured sparsity and learned combining or beamforming.

  • Experimental setup: The fixed-scattering scenario models a base station, users, and randomly distributed scatterers, with channel parameters determined from geometric characteristics.The setup includes line-of-sight and non-line-of-sight propagation components.
  • Experimental setup: The FSE data set contains 10000 channel samples split into 8000 training, 1000 validation, and 1000 test samples.Samples are generated from randomly distributed user locations and user-array normal directions in a fixed scattering environment.
  • Uplink channel estimation: The FSE-trained uplink estimator learns channel-environment characteristics using less pilot overhead.The evaluated uplink setting uses G = 1024 and Q = 40 and tests the schemes on the FSE test set.
  • Downlink channel estimation: More users achieve larger received SNR with the FSE-trained CCN than with random phases under a fixed scattering environment.The comparison concerns downlink channel estimation using two CCN designs.
  • Overall results: The proposed MDDL-based solution achieves significant improvement in channel estimation and feedback performance over conventional schemes.The scheme exploits angle-delay sparsity, jointly learns hardware-side measurements and estimation, and reconstructs spatial-frequency channels with MMV-LAMP.

APPENDIX

The appendix formulates AMP and MMV-LAMP as sparse-recovery procedures for noisy measurements. It introduces trainable linear and shrinkage parameters, including parameters associated with Bernoulli-Gaussian priors and common sparsity across measurement vectors.

  • AMP: AMP reconstructs a sparse vector from a noisy underdetermined measurement model using iterative residual updates and shrinkage.The model uses y as the noisy measurement, A as the measurement matrix, x as the sparse vector, and n as additive white Gaussian noise.
  • Denoising: The denoiser is based on the MSE-optimal estimator under a zero-mean Bernoulli-Gaussian prior.The prior combines a zero-valued component weighted by 1−γ with a Gaussian component weighted by γ.
  • MMV-LAMP: MMV-LAMP extends the sparse-recovery problem to noisy matrix measurements whose sparse columns share common support.The formulation uses Y ∈ C^M×K, A ∈ C^M×N, X ∈ C^N×K, and AWGN N ∈ C^M×K.
  • MMV-LAMP: The developed MMV-LAMP network trains both the measurement-update matrix B and the shrinkage parameters θ.The parameter vector is defined as θ = [θ1, θ2].
Loading 2104.11052v3…