Bayesian Inference over the Stiefel Manifold via the Givens Representation

被引:6
作者
Pourzanjani, Arya A. [1 ]
Jiang, Richard M. [1 ]
Mitchell, Brian [1 ]
Atzberger, Paul J. [2 ]
Petzold, Linda R. [1 ]
机构
[1] Univ Calif Santa Barbara, Comp Sci Dept, Santa Barbara, CA 93106 USA
[2] Univ Calif Santa Barbara, Math Dept, Santa Barbara, CA 93106 USA
来源
BAYESIAN ANALYSIS | 2021年 / 16卷 / 02期
基金
美国国家科学基金会;
关键词
orthogonal matrix; transformation; dimensionality reduction; principal component analysis; SIMULATION;
D O I
10.1214/20-BA1202
中图分类号
O1 [数学];
学科分类号
0701 ; 070101 ;
摘要
We introduce an approach based on the Givens representation for posterior inference in statistical models with orthogonal matrix parameters, such as factor models and probabilistic principal component analysis (PPCA). We show how the Givens representation can be used to develop practical methods for transforming densities over the Stiefel manifold into densities over subsets of Euclidean space. We show how to deal with issues arising from the topology of the Stiefel manifold and how to inexpensively compute the change-of-measure terms. We introduce an auxiliary parameter approach that limits the impact of topological issues. We provide both analysis of our methods and numerical examples demonstrating the effectiveness of the approach. We also discuss how our Givens representation can be used to define general classes of distributions over the space of orthogonal matrices. We then give demonstrations on several examples showing how the Givens approach performs in practice in comparison with other methods.
引用
收藏
页码:639 / 666
页数:28
相关论文
empty
未找到相关数据