Graph-constrained Analysis for Multivariate Functional Data

Area of Science:

• Statistics
• Machine Learning
• Neuroimaging Analysis

Background:

• Multivariate functional data analysis often requires understanding conditional relationships between variables, typically represented by graphical models.
• Existing functional Gaussian graphical models (GGM) estimate unknown graphs and cannot incorporate prior knowledge of a specified graph structure.
• Prior knowledge of inter-variable relationships is crucial in many applications, such as analyzing functional magnetic resonance imaging (fMRI) data with known brain connectivity.

Purpose of the Study:

• To propose a novel method for multivariate functional data analysis that strictly adheres to a predefined graphical model.
• To establish theoretical connections between functional GGM and graphical Gaussian processes (GP) to leverage existing frameworks.
• To develop algorithms that preserve marginal distributions and computational scalability while respecting graphical constraints.

Main Methods:

• Demonstrated the equivalence between partially separable functional GGM and graphical GP.
• Developed a new algorithm using Dempster's covariance selection for maximum likelihood estimation under graphical constraints.
• Extended the algorithm to address oversmoothing issues associated with low-rank graphical GP approximations, improving marginal distribution preservation.

Main Results:

• Established a theoretical link between functional GGM and graphical GP, enabling the use of GP for constrained covariance function construction.
• Proposed an algorithm that effectively incorporates known graphical structures into the analysis of multivariate functional data.
• Showcased improved preservation of marginal distributions compared to standard low-rank approximations, alongside computational scalability.
• Validated the proposed methods through empirical experiments and a practical neuroimaging application.

Conclusions:

• The proposed method provides a principled way to analyze multivariate functional data when the graphical structure is known.
• The developed algorithms offer a balance between respecting known graphical constraints, preserving marginal distributions, and maintaining computational efficiency.
• This work advances functional data analysis techniques, particularly for applications like neuroimaging where prior structural information is available.