Abstract:We consider random matrix ensembles on the set of Hermitian matrices that are heavy tailed, in particular not all moments exist, and that are invariant under the conjugate action of the unitary group. The latter property entails that the eigenvectors are Haar distributed and, therefore, factorise from the eigenvalue statistics. We prove a classification for stable matrix ensembles of this kind of matrices represented in terms of matrices, their eigenvalues and their diagonal entries with the help of the classification of the multivariate stable distributions and the harmonic analysis on symmetric matrix spaces. Moreover, we identify sufficient and necessary conditions for their domains of attraction. To illustrate our findings we discuss for instance elliptical invariant random matrix ensembles and Pólya ensembles, the latter playing a particular role in matrix convolutions. As a byproduct we generalise the derivative principle on the Hermitian matrices to general tempered distributions. This principle relates the joint probability density of the eigenvalues and the diagonal entries of the random matrix.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the classification of the stable distributions and their domains of attraction for unit - invariant Hermitian random matrix ensembles with heavy - tailed distributions. Specifically, the researchers are concerned with those Hermitian random matrix ensembles that have heavy - tailed properties and are invariant under the conjugation action of the unitary group. The core of these problems lies in:
1. **Classification of stable distributions**: For such matrix ensembles, the researchers hope to find a classification method that can describe the stable distributions of these matrices. Here, "stable" means that when summing multiple independent and identically distributed such matrices, the limit distribution still belongs to the same class of distributions, only requiring appropriate scaling and translation.
2. **Determination of the domain of attraction**: Further, the researchers also hope to be able to determine the domains of attraction of these stable distributions, that is, which matrices with initial distributions will converge to these stable distributions after a large number of independent replications.
### Background and Motivation
- **Application background**: Heavy - tailed random matrices have wide applications in fields such as machine learning, disordered systems, quantum field theory, finance, and statistics. In particular, for Wigner matrices, when their elements have heavy - tailed distributions, their spectral statistical properties show a transition from sine - kernel statistics in the bulk to Poisson statistics in the tails.
- **Existing research**: Existing research has shown that for heavy - tailed Wigner matrices, as the matrix dimension tends to infinity, their eigenvalues show Poisson statistical properties in the tails of the density, and the corresponding eigenvectors also become localized. In addition, there are similar studies for other types of heavy - tailed matrices such as banded matrices, correlation matrices, and group - invariant random matrices.
### Main Contributions
- **Generalization of the derivative principle**: The authors generalize the derivative principle on Hermitian matrices and apply it to general tempered distributions. This principle establishes the relationship between the joint probability density of the eigenvalues of a random matrix and the joint probability density of its diagonal elements.
- **Definition and properties of stable distributions**: The author defines the stable distributions of unit - invariant random matrices and gives the specific form of their characteristic functions. These distributions are described by four parameters: the translation term \(y_0I_N\), the scale parameter \(\gamma\), the stability index \(\alpha\), and the normalized spectral measure \(H\).
- **Classification of the domain of attraction**: By introducing new equivalence conditions, the authors classify the domains of attraction of stable distributions, and in particular, conduct a detailed analysis of the domains of attraction for eigenvalues and diagonal elements.
### Specific Problems
- **How to classify stable distributions**: By defining and analyzing the stable distributions of unit - invariant random matrices, the authors provide a classification method that can describe the stable distributions of these matrices.
- **How to determine the domain of attraction**: The authors conduct a detailed analysis of the domains of attraction of these stable distributions by introducing new equivalence conditions, especially for the domains of attraction for eigenvalues and diagonal elements.
### Conclusion
This paper fills the theoretical gap in this field through in - depth research on the stable distributions and their domains of attraction of unit - invariant Hermitian random matrix ensembles, and provides an important theoretical basis for understanding the spectral statistical properties of heavy - tailed random matrices.