Low-Complexity Iterative Methods for Complex-Variable Matrix Optimization Problems in Frobenius Norm

Sai Wang, Yi GongarXiv 2023 · DOI · arXiv

Abstract

This paper presents a framework for solving complex-variable matrix optimization problems (CMOPs) directly in the complex domain using the Frobenius norm. By leveraging Wirtinger calculus, the authors develop Gradient Descent (GD) and Projection Gradient Descent (PGD) algorithms that avoid the computational overhead of real-domain transformations. Theoretical analysis establishes convergence criteria and step-size bounds. Numerical results validate that these methods offer reduced complexity and faster convergence, making them suitable for large-scale engineering applications.

Introduction

In fields such as medical imaging (MRI), signal processing, and control theory, optimization problems frequently involve complex matrices. The conventional approach transforms these into real-valued problems by splitting variables into real and imaginary components, effectively doubling the problem's dimensionality. This transformation leads to increased computational complexity, a loss of physical intuition regarding the original complex structure, and potentially slower convergence due to more cumbersome mathematical expressions. This study addresses these issues by proposing direct iterative methods in the complex domain.

Methodology

The researchers utilize Wirtinger calculus to define gradients for complex-variable functions without requiring decomposition into real and imaginary parts. The study focuses on two primary scenarios:

  1. Unconstrained CMOPs: These are addressed using a standard Gradient Descent (GD) approach. The authors prove the convexity of the objective function and derive a closed-form solution to serve as a performance benchmark.
  2. Constrained CMOPs: Specifically, problems involving constraints on the diagonal elements of matrix products (common in power-constrained engineering systems) are solved using Projection Gradient Descent (PGD).

A critical component of the methodology is the rigorous proof that these complex-domain iterations satisfy the Lipschitz condition. Convergence stability is maintained by selecting a step size ($\alpha$) within specific theoretical bounds: $(0, 2/L)$ for unconstrained problems and $(0, 1/L)$ for constrained problems, where $L$ is the Lipschitz constant.

Experimental Setup

The theoretical findings were validated using synthetically generated data. Complex matrices $H$ and $A$ were randomly generated with values between -10 and 10. Experiments utilized matrix dimensions of $M=10, N=5$, and $K=8$. The iterative results were compared against derived closed-form optimal solutions to confirm accuracy and stability.

Results and Discussion

Numerical simulations demonstrate the efficacy of the proposed iterative methods. As shown in Figure 1, the function value for unconstrained problems decreases rapidly and stabilizes at the optimal value.

Convergence of the Gradient Descent (GD) algorithm for unconstrained CMOPs across different step sizes.

Similarly, Figure 2 illustrates the convergence of the PGD algorithm for constrained problems. In both cases, the choice of step size $\alpha$ significantly impacts the convergence rate. A larger step size (e.g., $\alpha = 0.0006$) leads to faster convergence than smaller values (e.g., $\alpha = 0.0002$), provided it remains within the theoretically proven stable range.

Convergence of the Projection Gradient Descent (PGD) algorithm for constrained CMOPs across different step sizes.

The results confirm that staying in the complex domain is not only mathematically sound but also computationally more efficient than real-domain reformulation. The PGD method, in particular, exhibits favorable scalability compared to closed-form solutions involving $O(N^3)$ matrix inversions.

Conclusion

This research demonstrates that direct complex-domain optimization using GD and PGD is a viable and efficient alternative to traditional real-domain methods. By preserving the inherent structure of complex data, these methods are particularly well-suited for applications like waveform design and large-scale signal processing. Future research could extend this framework to other norms, such as the $L_1$ or nuclear norms, and explore its applicability to non-convex optimization problems.

Summary updated

BibTeX
@misc{wang2023low,
  author = {Sai Wang and Yi Gong},
  title = {{Low-Complexity Iterative Methods for Complex-Variable Matrix Optimization Problems in Frobenius Norm}},
  year = {2023},
  eprint = {2303.07614},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2303.07614},
}
Figure