Skip to main content

Lund University Publications

LUND UNIVERSITY LIBRARIES

Matrix variate gamma distributions with unrestricted shape parameter

Kozubowski, Tomasz J. ; Mazur, Stepan LU and Podgórski, Krzysztof LU (2025) In Journal of Multivariate Analysis 209.
Abstract

Matrix gamma distributions are among the most important matrix-variate laws in multivariate statistical analysis, as they encompass the Wishart distributions – the sample covariance distributions under Gaussianity – and provide a natural model for random covariances in Bayesian multivariate methods. A substantial body of literature explores this class of distributions, traditionally characterized by a shape parameter restricted to the (Gindikin) set {i/2,i∈{1,…,k−1}}∪((k−1)/2,∞), where k×k is the dimension of the matrix variate. In this paper, we show that matrix-variate gamma distributions can be naturally extended to allow the entire positive half-line as the domain of the shape parameter. This extension not only unifies the... (More)

Matrix gamma distributions are among the most important matrix-variate laws in multivariate statistical analysis, as they encompass the Wishart distributions – the sample covariance distributions under Gaussianity – and provide a natural model for random covariances in Bayesian multivariate methods. A substantial body of literature explores this class of distributions, traditionally characterized by a shape parameter restricted to the (Gindikin) set {i/2,i∈{1,…,k−1}}∪((k−1)/2,∞), where k×k is the dimension of the matrix variate. In this paper, we show that matrix-variate gamma distributions can be naturally extended to allow the entire positive half-line as the domain of the shape parameter. This extension not only unifies the well-known singular Wishart and non-singular matrix-variate gamma distributions but also introduces new singular matrix-variate distributions with shape parameters outside the Gindikin set. While permutation invariance is no longer preserved in the singular, non-Wishart case, and its scaling properties require special treatment, our unified framework leads to new representations that bypass the restrictions of the Gindikin set. We provide several elegant and convenient stochastic representations for matrix-variate gamma distributions, which are novel even in the non-singular case. Notably, we demonstrate that the lower triangular matrix in the Cholesky factorization of a gamma-distributed matrix – whether singular or not – follows a triangular matrix-variate Rayleigh distribution, introducing a new class of matrix-valued variables that extends the classical univariate Rayleigh distribution to the matrix domain. We also briefly address statistical issues and potential applications to non-elliptical multivariate heavy tailed data.

(Less)
Please use this url to cite or link to this publication:
author
; and
organization
publishing date
type
Contribution to journal
publication status
published
subject
keywords
Matrix gamma distribution, Matrix Laplace distribution, Singular Wishart distribution
in
Journal of Multivariate Analysis
volume
209
article number
105457
publisher
Academic Press
external identifiers
  • scopus:105007624350
ISSN
0047-259X
DOI
10.1016/j.jmva.2025.105457
language
English
LU publication?
yes
id
d0dff3fe-b36d-4370-a3cc-50372bc59ec4
date added to LUP
2025-11-04 11:06:13
date last changed
2025-11-04 11:06:29
@article{d0dff3fe-b36d-4370-a3cc-50372bc59ec4,
  abstract     = {{<p>Matrix gamma distributions are among the most important matrix-variate laws in multivariate statistical analysis, as they encompass the Wishart distributions – the sample covariance distributions under Gaussianity – and provide a natural model for random covariances in Bayesian multivariate methods. A substantial body of literature explores this class of distributions, traditionally characterized by a shape parameter restricted to the (Gindikin) set {i/2,i∈{1,…,k−1}}∪((k−1)/2,∞), where k×k is the dimension of the matrix variate. In this paper, we show that matrix-variate gamma distributions can be naturally extended to allow the entire positive half-line as the domain of the shape parameter. This extension not only unifies the well-known singular Wishart and non-singular matrix-variate gamma distributions but also introduces new singular matrix-variate distributions with shape parameters outside the Gindikin set. While permutation invariance is no longer preserved in the singular, non-Wishart case, and its scaling properties require special treatment, our unified framework leads to new representations that bypass the restrictions of the Gindikin set. We provide several elegant and convenient stochastic representations for matrix-variate gamma distributions, which are novel even in the non-singular case. Notably, we demonstrate that the lower triangular matrix in the Cholesky factorization of a gamma-distributed matrix – whether singular or not – follows a triangular matrix-variate Rayleigh distribution, introducing a new class of matrix-valued variables that extends the classical univariate Rayleigh distribution to the matrix domain. We also briefly address statistical issues and potential applications to non-elliptical multivariate heavy tailed data.</p>}},
  author       = {{Kozubowski, Tomasz J. and Mazur, Stepan and Podgórski, Krzysztof}},
  issn         = {{0047-259X}},
  keywords     = {{Matrix gamma distribution; Matrix Laplace distribution; Singular Wishart distribution}},
  language     = {{eng}},
  publisher    = {{Academic Press}},
  series       = {{Journal of Multivariate Analysis}},
  title        = {{Matrix variate gamma distributions with unrestricted shape parameter}},
  url          = {{http://dx.doi.org/10.1016/j.jmva.2025.105457}},
  doi          = {{10.1016/j.jmva.2025.105457}},
  volume       = {{209}},
  year         = {{2025}},
}