EUVIMEDEUVIMED entdecken
Crossref

Sequential neural posterior estimation in simulation-based Bayesian inference

R. Peng, C. Genest, D. A. Stephens

Statistics Surveys · 01.01.2026 · 20 (none) · 242-283

DOI: 10.1214/26-ss158 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Statistics Surveys, erschienen am 01.01.2026. Verfasst von R. Peng und 2 weiteren Autor:innen. Die Arbeit nennt 61 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Bayesian inference offers a rigorous framework for systematically updating prior beliefs in light of observed data, but in many scientific models the likelihood function is intractable, making posterior computation difficult. Likelihood-free methods address this challenge by relying on model simulations rather than explicit likelihood evaluations. A well-known family of such methods is Approximate Bayesian Computation (ABC), which produces approximate posterior samples but often suffers from inefficiency and limited accuracy. Recent advances in simulation-based inference have introduced neural approaches, among which Neural Posterior Estimation (NPE) plays a central role. NPE directly estimates the posterior distribution using flexible neural conditional density estimators trained on simulated data, facilitating amortized and computationally efficient Bayesian inference. Sequential NPE, or SNPE for short, builds on NPE by sequentially adapting the simulation distribution to a particular observed dataset. This survey is deliberately centered on SNPE, reviewing its foundations and evolution, including core variants, extensions, and recent methodological advances. Three illustrative examples are also used to discuss the trade-offs among different approaches: a conjugate Bayesian linear regression model with a closed-form posterior for direct benchmarking, a nonlinear regression model with Gaussian errors using MCMC reference samples, and the same model under Student-t errors to probe robustness in a heavy-tailed noise regime. Practical advice is provided for end users.

Zitieren

APA

Peng, R., Genest, C., & Stephens, D. A. (2026). Sequential neural posterior estimation in simulation-based Bayesian inference. Statistics Surveys, 20(none), 242-283. https://doi.org/10.1214/26-ss158

Vancouver

Peng R, Genest C, Stephens DA. Sequential neural posterior estimation in simulation-based Bayesian inference. Statistics Surveys. 2026;20(none):242-283. doi:10.1214/26-ss158

BibTeX
@article{Peng2026,
  author = {Peng, R. and Genest, C. and Stephens, D. A.},
  title = {Sequential neural posterior estimation in simulation-based Bayesian inference},
  journal = {Statistics Surveys},
  year = {2026},
  volume = {20},
  number = {none},
  pages = {242--283},
  doi = {10.1214/26-ss158}
}
Kostenlos starten →