This paper explores the geometric properties of regular exponential-family distributions, which are fundamental in statistics. The authors establish a general method for constructing equivalent dual Bregman representations, linking log-densities to Bregman divergences in dual coordinate systems. These representations are shown to preserve conjugacy for prior distributions and offer natural reference geometries for analyzing relative smoothness. The research provides new geometric insights into various statistical models, including an application to the Poisson distribution that clarifies existing algorithms and reference geometries. AI
RANK_REASON The item is an academic paper detailing theoretical advancements in statistics and information geometry. [lever_c_demoted from research: ic=1 ai=0.7]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →