Sabbagh & Stephens prove Bayesian bootstrap validity in semi-parametric models

In a new arXiv paper (arXiv:2608.06670), statisticians Magid Sabbagh and David A. Stephens tackle a fundamental challenge in semi-parametric Bayesian inference: how to make reliable statements about a low-dimensional target parameter when the model also contains highly complex nuisance components. They adopt an estimating function approach and build the posterior using two non-parametric Bayesian tools — the Dirichlet process and the Bayesian bootstrap. Their goal is to ensure that the resulting posterior distribution has strong frequentist properties, meaning it can be trusted in the long run even when the nuisance structure is misspecified.

The key innovation is that they relax the standard assumption of stochastic equicontinuity, a technical condition often required for posterior asymptotic results but hard to verify in practice. Under this weaker framework, they prove that the posterior distribution is asymptotically Normal and concentrates at the true value of the target parameter. This is not just a theoretical exercise — the authors carefully outline exactly which assumptions are needed, and show how relaxing any one of them changes the conclusions. They also back up the analytical findings with simulations, confirming the approach works in finite samples.

For machine learning and statistics practitioners, this work bridges a gap between Bayesian and frequentist methodologies. It provides a rigorous justification for using flexible Bayesian tools (like Dirichlet process mixtures) in semi-parametric settings — common in causal inference, missing data, and high-dimensional nuisance adjustment. The result means practitioners can get calibrated uncertainty estimates even when the nuisance model is only approximately correct, making Bayesian inference more robust in real-world deployments.