A Peaceman-Rachford Splitting Approach with Deep Equilibrium Network for Channel Estimation

Dingli Yuan, Shitong Wu, Haoran Tang, Lu Yang, Chenghui PengIcct 2024 · DOI

TL;DR

This paper introduces a novel channel estimation method for MIMO systems by reformulating the problem as a fixed-point equation and solving it using a Deep Equilibrium Network (DEQ). This approach leverages the Peaceman-Rachford splitting method to ensure theoretical convergence and offers significantly lower computational complexity and memory overhead compared to existing deep learning-based methods.

What Problem Does This Paper Address?

Accurate channel estimation is crucial for the performance of Multiple-Input Multiple-Output (MIMO) systems, especially with the advent of new technologies like 6G and the need to handle complex scenarios such as hybrid far- and near-field channels. Traditional methods like Least Squares (LS) and Minimum Mean-Squared Error (MMSE) have limitations: LS offers poor performance, while MMSE is computationally prohibitive. While deep learning approaches, particularly those based on deep unfolding networks, have shown promise, they suffer from high memory and computational costs during training and lack theoretical guarantees due to truncated iterations. This paper aims to bridge this gap by developing a robust, efficient, and theoretically sound channel estimation method for MIMO systems.

Methods

The core of this paper lies in its novel approach to channel estimation:

  • Dual Formulation and Fixed-Point Equation: Instead of directly minimizing the channel estimation problem, the authors reformulate it in its dual form. They then apply the Peaceman-Rachford (PR) splitting method to this dual problem to derive a fixed-point equation. This equation is designed to have a non-expansive property, which is crucial for guaranteeing convergence.
  • Deep Equilibrium Network (DEQ): The derived fixed-point equation is then implemented using a Deep Equilibrium (DEQ) network. DEQ models are characterized by a single "fixed layer" that represents the equilibrium state of a fixed-point equation. This design inherently avoids the need for deep unfolding, leading to constant memory complexity and significantly reduced training complexity (O(1) compared to O(T) for unfolded networks).
  • NN Approximation of Proximal Operator: The PR splitting method involves proximal operators. The paper proposes replacing the nonlinear proximal operator with a trainable neural network (NN), denoted as $R_{\theta}$. This NN is designed using residual blocks and aims to approximate the proximal operator with arbitrary precision.
  • PR-DEN Algorithm: The combined approach, named Peaceman-Rachford Splitting method implemented with Deep Equilibrium Network (PR-DEN), unfolds a single iteration of the PR splitting algorithm. The nonlinear proximal term is replaced by the trained NN, effectively creating a fixed layer within the DEQ framework.

Datasets

The paper does not explicitly mention specific public datasets for training or evaluation. Instead, it describes simulations conducted using a hybrid far- and near-field channel model. For training the neural networks, a dataset of 80,000 samples was used, with an additional 5,000 samples for validation. The testing was performed on a comprehensive dataset of 5,000 samples.

Key Insights and Ideas

  • Shifting to the Dual Domain for Robustness: By reformulating the problem in the dual domain and leveraging the convexity of Fenchel conjugates, the authors ensure a more robust optimization landscape for the PR splitting method.
  • DEQ's Efficiency Advantage: The use of DEQ is a key insight, as it elegantly sidesteps the memory and computational burdens associated with deep unfolding networks by treating the entire iterative process as a single equilibrium state.
  • Theoretical Guarantees for Deep Learning: The paper provides rigorous theoretical analysis, proving convergence and optimality. This is a significant contribution, as many deep learning-based signal processing methods lack such strong theoretical backing.
  • NN as a Proximal Operator Learner: The idea of using a neural network to learn the complex proximal operator is a clever way to integrate data-driven learning with model-based optimization principles.

Key Results

  • Superior NMSE Performance: The proposed PR-DEN method significantly outperforms traditional methods like LS and MMSE across various Signal-to-Noise Ratio (SNR) levels.
  • Competitive with State-of-the-Art Deep Learning: PR-DEN shows notable improvement over ISTA-Net+ and achieves approximately a 5% performance gain compared to FPN-OAMP, another advanced deep learning-based method.
  • Rapid Convergence: Simulations show that PR-DEN achieves rapid convergence, often within 4 iterations, outperforming other iterative methods in terms of NMSE early in the process.
  • Computational Efficiency: PR-DEN demonstrates significantly lower computational time compared to other iterative algorithms, especially when compared to network unfolding methods like ISTA-Net+. It is comparable to FPN-OAMP in terms of execution time.

Limitations and Future Work

  • Approximation Precision: The theoretical guarantees of PR-DEN rely on the neural network $R_{\theta}$ sufficiently approximating the proximal operator. The paper acknowledges that the precision of this approximation is crucial.
  • Convexity Assumption: The proof of convergence for the primal solution relies on the assumption that the regularization term $g(p)$ is convex. While this is a common assumption, it might limit applicability in scenarios with non-convex regularization.
  • Future Work: The authors suggest that their approach can be extended to other optimization problems where fixed-point equations with non-expansive properties can be formulated. Further exploration of different network architectures for approximating proximal operators and investigating non-convex regularization scenarios could be future directions.

Summary updated

BibTeX
@inproceedings{yuan2024peaceman,
  author = {Dingli Yuan and Shitong Wu and Haoran Tang and Lu Yang and Chenghui Peng},
  title = {{A Peaceman-Rachford Splitting Approach with Deep Equilibrium Network for Channel Estimation}},
  booktitle = {Icct},
  year = {2024},
  doi = {10.1109/icct62411.2024.10946331},
}
Figure