Generalizing the Eigenvalue Interlacing Theorem to Pseudo-Similarity Transformations

TL;DR

This paper extends the Eigenvalue Interlacing Theorem to include pseudo-similarity transformations, showing interlacing can occur between Hermitian and non-Hermitian matrices, even with dimensionally inflated matrices.

Contribution

It generalizes the theorem to pseudo-similarity transformations involving Moore-Penrose pseudoinverses, broadening the scope of eigenvalue interlacing results.

Findings

Eigenvalue interlacing applies to pseudo-similarity transformations.

Interlacing can occur between Hermitian and non-Hermitian matrices.

Interlacing can happen with dimensionally inflated matrices.

Abstract

The current general form of the well-known Eigenvalue Interlacing Theorem states that, given an $N \times N$ Hermitian matrix $P$, the eigenvalues of the matrix product $Q^{H} P Q$ will interlace those of $P$ if the columns of the $N \times L$ matrix $Q$ (with $L \le N$) are unitary. This note further generalizes this theorem to include pseudo-similarity transformations, namely products of the form $H^{\dagger} P H$, where $H$ is a general $N \times K$ matrix and "$\dagger$" denotes the Moore-Penrose pseudoinverse. This implies that, while the product $Q^{H} P Q$ is Hermitian and is generally a deflated version of $P$ (both in dimensionality and in the number of non-zero eigenvalues), this is not the case for $H^{\dagger} P H$, which, although generally a deflated version of $P$ in terms of the number of non-zero eigenvalues, will not necessarily be so in dimensionality, nor will it in general be Hermitian. Thus, this note not only generalizes the Eigenvalue Interlacing Theorem but also shows that eigenvalue interlacing may occur between Hermitian and non-Hermitian matrices and even in the presence of dimensionally inflated matrices.