On the phases of a complex matrix

Dan Wang, Wei Chen, Sei Zhen Khong, Li QiuLinear 2020 · DOI · arXiv

Abstract

This paper introduces a rigorous framework for defining the "phases" of complex sectorial matrices, analogous to how singular values are used to describe magnitude. By utilizing a specific sectorial decomposition, the authors establish properties such as congruence invariance, interlacing, and majorization. These theoretical developments find direct applications in feedback stability analysis and matrix completion problems.

Introduction

In basic complex arithmetic, every number is characterized by its magnitude and phase (angle). While matrix theory has a mature understanding of "magnitude" through singular values, a universally accepted, property-rich definition for the "phase" of a matrix has remained elusive. Previous attempts often relied on the unitary part of a standard polar decomposition, but these lacked essential properties like invariance under congruence transformations.

This research addresses this gap by focusing on sectorial matrices—matrices whose numerical range (field of values) $W(C)$ lies within a cone in the complex plane that does not contain the origin. The goal is to create a "phase theory" for matrices that is as robust and useful as the existing "magnitude theory."

The numerical range $W(C)$ of a sectorial matrix $C$ contained within a cone defined by its upper and lower phase bounds.

Methodology

The authors define the phases of a sectorial matrix $C$ as its canonical angles. These are derived from a sectorial decomposition: $C = T^*DT$, where $T$ is invertible and $D$ is a diagonal unitary matrix. The phases correspond to the angles of the diagonal elements of $D$.

To support this definition, the paper introduces and proves the uniqueness of two new factorizations for sectorial matrices:

  1. Symmetric Polar Decomposition (SPD): $C = PUP$, where $P$ is positive definite and $U$ is unitary.
  2. Generalized Cholesky Factorization (GCF): $C = R^*WR$, where $R$ is upper triangular and $W$ is a specific unitary matrix.

Furthermore, the authors utilize the compound numerical range—a sophisticated tool used to track how the eigenvalues of matrix products behave—to establish majorization relations.

Key Insights and Theoretical Properties

The proposed phase theory mirrors the robust "magnitude theory" of singular values through several key properties:

  • Congruence Invariance: Unlike singular values, which are unitarily invariant ($|UAV| = |A|$), matrix phases are "congruence invariant." This means the phases of $C$ are identical to those of $P^*CP$ for any invertible $P$. This is the matrix equivalent of the scalar property $\angle(w^*cw) = \angle c$.
  • The Interlacing Property: Just as the singular values of a submatrix interlace with those of the original matrix, the phases of a "compression" (a specific type of submatrix) and the Schur complement interlace with the phases of the parent matrix.
  • Majorization in Products: A significant finding is a majorization relation for matrix products. The authors show that the phases of the eigenvalues of $AB$ are majorized by the sum of the phases of $A$ and $B$: $\angle \lambda(AB) \prec \phi(A) + \phi(B)$. This mirrors the classic result that the log of singular values of $AB$ is majorized by the sum of the logs of singular values of $A$ and $B$.

Results and Applications

The theoretical framework leads to several practical results and proofs:

  • Uniqueness Proofs: The authors successfully prove that the SPD and GCF are unique for sectorial matrices, providing a stable foundation for the phase definition.
  • Matrix Completion: The paper provides necessary and sufficient conditions for the "sectorial matrix completion problem." It demonstrates exactly when a partial matrix (such as a band of known values) can be filled in to remain a sectorial matrix with specific phase bounds.
  • Control Theory Application: The results on the rank of $I + AB$ provide a "phase-based" robustness margin. This is highly relevant for analyzing the stability of feedback systems in engineering, where phase information can guarantee the invertibility of the system matrix.

Discussion and Future Work

While the current theory is robust for sectorial matrices, it is limited to cases where the numerical range does not touch the origin. Future research may extend these concepts to "semi-sectorial" matrices (where the origin is on the boundary). A major open question involves defining phases for general "unitoid" matrices that aren't sectorial; the authors speculate that Riemann surface theory might be required to handle the "wrap-around" effect of angles exceeding $\pi$.

Summary updated

BibTeX
@inproceedings{wang2020phases,
  author = {Dan Wang and Wei Chen and Sei Zhen Khong and Li Qiu},
  title = {{On the phases of a complex matrix}},
  booktitle = {Linear},
  year = {2020},
  doi = {10.1016/j.laa.2020.01.035},
}
Figure