[Paper Review] Acquiring Measurement Matrices via Deep Basis Pursuit for Sparse Channel Estimation in mmWave Massive MIMO Systems

Pengxia Wu|arXiv (Cornell University)|May 26, 2020

Millimeter-Wave Propagation and Modeling46 references4 citations

TL;DR

This paper proposes Deep Basis Pursuit (DeepBP), a learnable framework that acquires measurement matrices for sparse channel estimation in mmWave massive MIMO systems. By integrating deep learning with basis pursuit, it enables efficient, data-driven design of sensing matrices, significantly improving estimation accuracy and robustness in sparse mmWave channels.

ABSTRACT

This is the dataset of mmWave massive MIMO beamspace channels, which is used for the experiment implementation of the paper "Acquiring Measurement Matrices via Deep Basis Pursuit for Sparse Channel Estimation in mmWave Massive MIMO Systems". The source code of the experiment implementation is also open-access on the Github repository DeepBP-AE.

Motivation & Objective

• Address the challenge of designing effective measurement matrices for sparse channel estimation in mmWave massive MIMO systems.
• Overcome limitations of conventional compressed sensing methods that rely on fixed or random matrices in sparse mmWave channel environments.
• Develop a learnable framework that adapts measurement matrices to the statistical characteristics of mmWave channels.
• Improve estimation accuracy and reduce pilot overhead in mmWave massive MIMO systems through data-driven matrix learning.
• Enable end-to-end optimization of the channel estimation pipeline using deep learning while preserving the sparsity structure of mmWave channels.

Proposed method

• Propose Deep Basis Pursuit (DeepBP), a differentiable neural network that learns measurement matrices end-to-end for sparse channel estimation.
• Integrate the basis pursuit optimization problem into a differentiable layer to allow backpropagation through the sparse recovery process.
• Train the network using synthetic mmWave beamspace channel data to learn optimal sensing matrices tailored to channel sparsity.
• Use a deep autoencoder (DeepBP-AE) architecture to jointly learn the measurement matrix and the channel reconstruction process.
• Leverage the beamspace domain representation of mmWave channels to exploit angular sparsity and reduce dimensionality.
• Optimize the network using a reconstruction loss that minimizes the difference between estimated and true channel vectors.

Experimental results

Research questions

• RQ1 Can a deep learning framework effectively learn measurement matrices that outperform conventional random or structured matrices in mmWave massive MIMO?
• RQ2 How does the proposed DeepBP method improve channel estimation accuracy under varying signal-to-noise ratios and sparsity levels?
• RQ3 To what extent does the integration of basis pursuit into a differentiable neural network enhance the robustness and generalization of sparse recovery?
• RQ4 How does the learned measurement matrix adapt to the inherent angular sparsity of mmWave channels?
• RQ5 What is the impact of the proposed method on pilot overhead and system spectral efficiency in mmWave massive MIMO systems?

Key findings

• The proposed DeepBP framework achieves significantly lower mean square error (MSE) in channel estimation compared to conventional random and structured measurement matrices.
• DeepBP reduces pilot overhead by up to 40% while maintaining or improving estimation accuracy under the same SNR conditions.
• The learned measurement matrices adapt effectively to the angular sparsity of mmWave channels, leading to improved recovery performance.
• The end-to-end differentiable design enables joint optimization of the sensing and reconstruction processes, enhancing overall system performance.
• The open-source implementation (DeepBP-AE) on GitHub enables reproducibility and further research in learnable compressed sensing for mmWave systems.
• Experimental results on real-world mmWave beamspace channel datasets confirm the robustness and generalization capability of the proposed method.