Asymptotic theory for maximum likelihood estimates in reduced-rank multivariate generalized linear models.

Area of Science:

• Statistics
• Machine Learning
• Dimensionality Reduction

Background:

• Reduced-rank regression is a key dimensionality reduction technique with established theory for multivariate linear models.
• Theoretical results for reduced-rank multivariate generalized linear models are scarce.
• Existing methods lack robust theoretical underpinnings for complex data structures.

Purpose of the Study:

• To develop novel M-estimation theory for reduced-rank multivariate generalized linear models.
• To establish consistency and asymptotic distribution for maximum likelihood estimators in these models.
• To provide a theoretical framework for analyzing data with joint response and predictor distributions.

Main Methods:

• Developed M-estimation theory for concave criterion functions over non-convex, non-closed parameter spaces.
• Derived asymptotic properties of estimators under joint distribution assumptions.
• Applied the theory to a real-world classification problem with binary covariates.

Main Results:

• Established the consistency and asymptotic distribution of maximum likelihood estimators for reduced-rank multivariate generalized linear models.
• Demonstrated the applicability of the developed theory through a practical data classification example.
• Provided a robust theoretical foundation for a class of models with limited prior research.

Conclusions:

• The developed M-estimation theory provides a significant advancement for reduced-rank multivariate generalized linear models.
• The findings offer reliable methods for parameter estimation and inference in complex statistical modeling.
• This work bridges a critical gap in the theoretical understanding of these models, enhancing their practical utility.