Inconsistency between Frequentist and Bayesian inference, approximation of improper priors and permissible priors
Abstract
Improper priors are widely used in objective Bayesian inference, although posterior propriety does not by itself guarantee coherence between Bayesian and frequentist assessments. We investigate this issue by approximating improper priors with proper priors and examining the convergence of the resulting posterior distributions. We introduce monotone q-vague convergence, which ensures pointwise Kullback-Leibler (KL) convergence of the posteriors, and study expected KL (EKL) convergence together with its total variation counterpart (ETV). Through continuous and discrete examples, we show that EKL convergence can depend on the sequence of compact sets used for approximation. In particular, convergence along one compact exhaustion need not extend to others. Our main result connects posterior approximation to inconsistency: strong inconsistency prevents both ETV and EKL convergence for any proper-prior approximation. Consequently, KL-permissible and TV permissible priors cannot be strongly inconsistent. We also introduce weak inconsistency and show that it may persist despite EKL convergence. Finally, a unified discrete framework covering the Flatland and Fraser-Monette-Ng paradoxes illustrates the distinction between these notions. In particular, ETV convergence need not imply EKL convergence, while the absence of weak inconsistency does not guarantee an EKL convergent compact approximation.
Related
- Frequentist and Bayesian population pharmacometrics: similarities, differences, and practical implications
- RECAP: Representation-Based Calibration with Amortized Pooling for Frequentist Simulation-Based Inference
- From Frequentist to Bayesian Contextual Optimization
- Coverage is not enough: Frequentist tests of simulation-based inference for primordial non-Gaussianity
- Recurrent Neural Networks and (Time-Uniform) Universal Approximation of Bayesian Filters
BibTeX
@misc{druilhet2026inconsistency,
title = {{Inconsistency between Frequentist and Bayesian inference, approximation of improper priors and permissible priors}},
author = {Pierre Druilhet and Erwan Saint Loubert Bié},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.26422},
eprint = {2610.26422},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.26422},
}
RIS
TY - UNPB
TI - Inconsistency between Frequentist and Bayesian inference, approximation of improper priors and permissible priors
AU - Pierre Druilhet
AU - Erwan Saint Loubert Bié
PY - 2026
DA - 2026/10/19
PB - arXiv
DO - 10.48550/arxiv.2610.26422
UR - https://arxiv.org/abs/2610.26422
AB - Improper priors are widely used in objective Bayesian inference, although posterior propriety does not by itself guarantee coherence between Bayesian and frequentist assessments. We investigate this issue by approximating improper priors with proper priors and examining the convergence of the resulting posterior distributions. We introduce monotone q-vague convergence, which ensures pointwise Kullback-Leibler (KL) convergence of the posteriors, and study expected KL (EKL) convergence together with its total variation counterpart (ETV). Through continuous and discrete examples, we show that EKL convergence can depend on the sequence of compact sets used for approximation. In particular, convergence along one compact exhaustion need not extend to others. Our main result connects posterior approximation to inconsistency: strong inconsistency prevents both ETV and EKL convergence for any proper-prior approximation. Consequently, KL-permissible and TV permissible priors cannot be strongly inconsistent. We also introduce weak inconsistency and show that it may persist despite EKL convergence. Finally, a unified discrete framework covering the Flatland and Fraser-Monette-Ng paradoxes illustrates the distinction between these notions. In particular, ETV convergence need not imply EKL convergence, while the absence of weak inconsistency does not guarantee an EKL convergent compact approximation.
ER -