November 2024 On log-concave approximations of high-dimensional posterior measures and stability properties in non-linear inverse problems
Jan Bohr, Richard Nickl
Author Affiliations +
Ann. Inst. H. Poincaré Probab. Statist. 60(4): 2619-2667 (November 2024). DOI: 10.1214/23-AIHP1397

Abstract

The problem of efficiently generating random samples from high-dimensional and non-log-concave posterior measures arising in nonlinear regression problems is considered. Extending investigations from (J. Eur. Math. Soc. (JEMS) 26 (2024) 1031–1112), local and global stability properties of the model are identified under which such posterior distributions can be approximated in Wasserstein distance by suitable log-concave measures. This allows the use of fast gradient based sampling algorithms, for which convergence guarantees are established that scale polynomially in all relevant quantities (assuming ‘warm’ initialisation). The scope of the general theory is illustrated in a non-linear inverse problem from integral geometry for which new stability results are derived.

On s’intéresse au problème de génération efficace d’échantillons à partir de mesures a posteriori de grande dimension et non log-concave, qui se pose dans des problèmes de régressions non linéaires. En développant les recherches de ((J. Eur. Math. Soc. (JEMS) 26 (2024) 1031–1112), on identifie des propriétés de stabilité locale et globale du modèle, sous lesquelles de telles distributions a posteriori peuvent être approximées dans la distance de Wasserstein par des mesures log-concaves appropriées. Ceci permet l’utilisation d’algorithmes d’échantillonnage rapides basés sur des gradients, pour lesquels on établit des garanties de convergence qui évoluent à une échelle polynomiale en toutes les quantités pertinentes (en supposant un démarrage à chaud). On illustre la portée de cette théorie générale dans un problème inverse non linéaire de géométrie intégrale, pour lequel on obtient de nouveaux résultats de stabilité.

Funding Statement

This article was part of the Doctoral thesis of the first author, who was supported by the EPSRC Centre for Doctoral Training and the Munro-Greaves Bursary of Queens’ College Cambridge.
The second author was supported by ERC grant No.647812 (UQMSI).

Acknowledgements

JB would like to thank François Monard for fruitful discussions leading to the proof of Theorem 4.2. Both authors would like to thank Gabriel Paternain, Sven Wang and two anonymous referees for various helpful remarks.

Citation

Download Citation

Jan Bohr. Richard Nickl. "On log-concave approximations of high-dimensional posterior measures and stability properties in non-linear inverse problems." Ann. Inst. H. Poincaré Probab. Statist. 60 (4) 2619 - 2667, November 2024. https://doi.org/10.1214/23-AIHP1397

Information

Received: 9 August 2021; Revised: 6 August 2022; Accepted: 17 April 2023; Published: November 2024
First available in Project Euclid: 19 November 2024

Digital Object Identifier: 10.1214/23-AIHP1397

Subjects:
Primary: 35R30 , 62G05
Secondary: 65Y20

Keywords: Bayesian non-parametric statistics , Non-Abelian X-ray transform , Non-linear inverse problems , Polynomial time sampling

Rights: Copyright © 2024 Association des Publications de l’Institut Henri Poincaré

Vol.60 • No. 4 • November 2024
Back to Top