Fast Calibration of Generalized Posteriors via Dirichlet Reweighting
Abstract
While generalized Bayesian inference offers a principled approach to uncertainty quantification under general loss functions, its practical performance depends critically on the choice of the learning rate. Generalized posterior calibration, which selects the learning rate so that the resulting credible sets attain nominal frequentist coverage, is computationally intensive because it relies on repeated posterior sampling across bootstrap datasets and candidate learning rates. To overcome this practical difficulty, we propose a fast calibration method that combines Dirichlet-weighted bootstrap perturbations with importance sampling to estimate coverage while reducing the need for repeated posterior sampling. We further extend the framework to hierarchical models, where group-specific learning rates are calibrated and stabilized via shrinkage. Illustrative examples using quantile regression, support vector machines, and hierarchical Huber regression demonstrate that the proposed method substantially reduces computational cost while maintaining accurate calibration.
Related
- Efficient Bootstrap Calibration of Generalized Posteriors with Importance Reweighting
- Computationally Efficient Experimental Design with Generalized Posteriors
- On the asymptotic behaviour of α-divergence generalized posteriors
- Generalized Dirichlet Energy and Graph Laplacians for Clustering Directed and Undirected Graphs
- Generalized Stević–Sharma operators on weighted Dirichlet spaces
BibTeX
@misc{sugasawa2026fast,
title = {{Fast Calibration of Generalized Posteriors via Dirichlet Reweighting}},
author = {Shonosuke Sugasawa},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.35187},
eprint = {2610.35187},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.35187},
}
RIS
TY - UNPB
TI - Fast Calibration of Generalized Posteriors via Dirichlet Reweighting
AU - Shonosuke Sugasawa
PY - 2026
DA - 2026/10/26
PB - arXiv
DO - 10.48550/arxiv.2610.35187
UR - https://arxiv.org/abs/2610.35187
AB - While generalized Bayesian inference offers a principled approach to uncertainty quantification under general loss functions, its practical performance depends critically on the choice of the learning rate. Generalized posterior calibration, which selects the learning rate so that the resulting credible sets attain nominal frequentist coverage, is computationally intensive because it relies on repeated posterior sampling across bootstrap datasets and candidate learning rates. To overcome this practical difficulty, we propose a fast calibration method that combines Dirichlet-weighted bootstrap perturbations with importance sampling to estimate coverage while reducing the need for repeated posterior sampling. We further extend the framework to hierarchical models, where group-specific learning rates are calibrated and stabilized via shrinkage. Illustrative examples using quantile regression, support vector machines, and hierarchical Huber regression demonstrate that the proposed method substantially reduces computational cost while maintaining accurate calibration.
ER -