Bayesian estimation of sparse precision matrices in the presence of Gaussian measurement error (Q2233583): Difference between revisions
From MaRDI portal
Removed claim: author (P16): Item:Q247539 |
Changed an Item |
||
Property / author | |||
Property / author: Subhashis Ghosal / rank | |||
Normal rank |
Revision as of 00:18, 12 February 2024
scientific article
Language | Label | Description | Also known as |
---|---|---|---|
English | Bayesian estimation of sparse precision matrices in the presence of Gaussian measurement error |
scientific article |
Statements
Bayesian estimation of sparse precision matrices in the presence of Gaussian measurement error (English)
0 references
11 October 2021
0 references
Let \(X_1, \ldots, X_n\) denote stochastically independent and identically distributed (i.i.d.) random vectors, each taking its values in \(\mathbb{R}^p\). Assume that \(X_1\) follows the centered normal distribution with positive definite covariance matrix \(\Sigma \in \mathbb{R}^{p \times p}\), and let \(\Omega = \Sigma^{-1}\). The authors' goal is to make inference on the unknown matrix \(\Omega\), especially in the high-dimensional situation when \(p\) is large. A Gaussian measurement error model is considered, meaning that the authors assume that observables \[ Y_{ij} = X_i + Z_{ij} \] for \(1 \leq i \leq n\) and \(1 \leq j \leq m\) are at hand. In this, it is assumed that all \(Z_{ij}\)'s are i.i.d. with \(Z_{11}\) following the centered \(p\)-dimensional normal distribution with covariance matrix \(\nu I_p\), where \(I_p \in \mathbb{R}^{p \times p}\) denotes the identity matrix in dimension \(p\). The ``measurement error variance'' \(\nu\) may be known or unknown, where in the latter case \(m \geq 2\) has to be assumed. Under this sampling model, the authors consider a Bayesian inference approach for \(\Omega\). To this end, they write \(\Omega = \Theta + \kappa I_p\) for a positive semi-definite matrix \(\Theta \in \mathbb{R}^{p \times p}\) and a real constant \(\kappa > 0\), and they assign independent prior distributions to \(\Theta\) and \(\kappa\). Samples from the resulting posterior distribution can be obtained by using Markov chain Monte Carlo (MCMC) methods. Under certain (sparsity) conditions, the authors analyze posterior contraction rates (for \(n \to \infty\)) with respect to the Frobenius distance, and they present a variety of relevant examples which fulfill their conditions. Besides these theoretical contributions, the authors also present some numerical results.
0 references
covariance matrix
0 references
Gaussian graphical model
0 references
high-dimensional inference
0 references
posterior contraction rate
0 references
sparsity
0 references