Bayesian effective dimension from information growth
Statistica Neerlandica · 2026-11 · 80 (4)
DOI: 10.1111/stan.70047 ↗Kurz gefasst
Fachartikel in der Zeitschrift Statistica Neerlandica, online erschienen am 16.11.2026. Verfasst von Sayantan Banerjee. Die Arbeit nennt 33 Literaturangaben.
Abstract (Originalsprache)
The number of parameters effectively learned from the data in a Bayesian analysis can differ substantially from the nominal dimension when some parameters are weakly identified, strongly regularized, or already tightly constrained by the prior. We define an effective dimension using the expected posterior‐to‐prior Kullback–Leibler divergence, normalized so that it converges to the ordinary parameter dimension in regular finite‐dimensional models. The definition is available at each finite information level whenever the mutual information is finite, and depends on both the likelihood and the prior. We derive exact expressions for Gaussian location and regression models and show how the resulting dimension is related to ridge degrees of freedom. For infinite‐dimensional Gaussian models, we characterize their growth in terms of the decay of the signal‐to‐noise eigenvalues, obtaining different rates for polynomial and exponential decay. We also study Gaussian scale mixtures and global–local priors, where the information can be separated into contributions from the latent scales and the conditionally Gaussian parameters. Finally, we describe a practical finite‐sample estimator based on prior‐predictive simulation and posterior‐to‐prior Kullback–Leibler divergence.
Zitieren
APA
Banerjee, S. (2026). Bayesian effective dimension from information growth. Statistica Neerlandica, 80(4). https://doi.org/10.1111/stan.70047
Vancouver
Banerjee S. Bayesian effective dimension from information growth. Statistica Neerlandica. 2026;80(4). doi:10.1111/stan.70047
BibTeX
@article{Banerjee2026,
author = {Banerjee, Sayantan},
title = {Bayesian effective dimension from information growth},
journal = {Statistica Neerlandica},
year = {2026},
volume = {80},
number = {4},
doi = {10.1111/stan.70047}
}
EUVIMED