Flexible Bayesian Dynamic Modeling of Correlation and Covariance Matrices.

Area of Science:

• Statistics
• Computational Biology
• Time Series Analysis

Background:

• Modeling covariance matrices is crucial but challenging due to positive-definiteness and high dimensionality.
• Existing methods, like the inverse-Wishart prior, have limitations in flexibility for complex dependence structures.

Purpose of the Study:

• To develop a novel Bayesian framework for flexible modeling of correlation and covariance matrices.
• To extend the framework for spatio-temporal processes and dynamic multivariate time series.
• To introduce efficient computational methods for posterior inference.

Main Methods:

• Decomposition of covariance matrices into correlation and variance components.
• Modeling correlations as products of unit vectors with flexible spherical distributions (e.g., squared-Dirichlet).
• Introduction of unit-vector Gaussian process priors for dynamic correlation structures.
• Development of adaptive Δ-Spherical Hamiltonian Monte Carlo for intractable posteriors.

Main Results:

• The proposed framework provides flexible prior distributions for covariance matrices, surpassing the inverse-Wishart prior.
• The method successfully models complex spatio-temporal dependencies and evolving correlations in multivariate time series.
• Adaptive Δ-Spherical Hamiltonian Monte Carlo efficiently handles posterior intractability.

Conclusions:

• The novel Bayesian framework offers a flexible and powerful approach to covariance matrix modeling.
• The method is validated through simulations of periodic processes and real-world analysis of neural activity.
• This approach advances the analysis of complex dependence structures in high-dimensional spatio-temporal data.