The E-Posterior
The e-Posterior uses e-variables to provide frequentist risk bounds, outperforming Bayesian posteriors in robustness under model misspecification.
Key Findings
Methodology
This paper introduces an e-Posterior based on e-variables, which serve as a generalization of likelihood ratios. By defining Sθ as an e-variable with expectation ≤ 1, the inverse 1/Sθ forms the e-Posterior. This structure supports arbitrary loss functions and guarantees frequentist risk bounds valid across all parameter values. The approach leverages the supermartingale property of e-variables to establish stochastic upper bounds on risks, ensuring robustness even when models or priors are poorly specified. The framework extends classical Bayesian inference, replacing subjective priors with collections of e-variables, and provides a quasi-conditional paradigm that guarantees risk control without overconfidence.
Key Results
- Simulation studies in nonparametric regression and hypothesis testing show that e-Posterior risk bounds maintain coverage probabilities within 5% error at sample size n=100, outperforming traditional Bayesian credible intervals, especially under model misspecification. In high-dimensional settings, the bounds adapt well to data-dependent weights, maintaining validity where Bayesian methods fail. The experiments demonstrate that carefully designed e-variables produce tighter risk bounds, matching frequentist minimax rates up to small constants, and exhibit robustness against model deviations.
- In hypothesis testing scenarios, e-Posteriors based on likelihood ratio e-variables replicate frequentist error control, with risk bounds remaining valid under optional stopping. For example, in simple-vs.-simple tests with sample size n=100, the risk bounds stay below 0.05, comparable to classical p-values but with guaranteed coverage. The framework also unifies Bayesian and frequentist approaches, interpreting classical conditional error probabilities as special cases of e-Posteriors, providing a new perspective on test calibration.
- The analysis of different e-collections reveals that tailored design of e-variables can optimize the tightness of risk bounds, balancing robustness and informativeness. The approach generalizes to nonparametric and high-dimensional models, offering a versatile tool for modern statistical inference, with potential applications in adaptive testing, model selection, and risk management.
Significance
This work addresses a fundamental challenge in statistical inference: how to obtain risk assessments that are valid regardless of model or prior misspecification. By leveraging e-variables, the authors develop a framework that guarantees frequentist risk bounds while maintaining flexibility across loss functions. This bridges the gap between Bayesian and frequentist paradigms, offering a robust alternative that is particularly valuable in complex, high-dimensional, or nonparametric settings. The approach enhances decision-making safety, reduces overconfidence, and broadens the applicability of statistical inference in scientific research, industry, and policy-making.
Technical Contribution
The paper introduces a novel class of risk bounds based on the reciprocal of e-variables, establishing a supermartingale-based risk assessment that holds uniformly over all parameters. This generalizes classical confidence distributions and Bayesian posteriors, providing a unified framework with theoretical guarantees. The method supports arbitrary loss functions and decision rules, enabling minimax strategies that are both robust and computationally feasible. It also offers insights into the design of e-collections, linking them to prior-like structures, and demonstrates how to achieve tight bounds in nonparametric and high-dimensional models, significantly advancing the theoretical landscape of statistical inference.
Novelty
This research is the first to systematically embed e-variables into a risk assessment framework akin to Bayesian posteriors, but with guaranteed frequentist validity under model misspecification. Unlike traditional confidence distributions, the e-Posterior supports arbitrary loss functions and provides stochastic upper bounds on risks, making it a versatile tool for robust inference. Its innovative use of supermartingale properties of e-variables to establish risk bounds marks a significant departure from existing methods, offering a new paradigm that unifies Bayesian and frequentist principles in a theoretically sound manner.
Limitations
- Constructing suitable e-variables requires careful design; poorly chosen e-collections can lead to overly conservative risk bounds, reducing informativeness.
- Computational complexity increases with model complexity and data dimensionality, posing challenges for large-scale applications.
- The framework assumes the existence of valid e-variables, which may be difficult in highly complex or nonparametric models, limiting immediate applicability in some scenarios.
Future Work
Further research will focus on automating e-variable construction, especially in high-dimensional and nonparametric contexts. Developing scalable algorithms for real-time inference and extending the framework to handle dependent data and complex null hypotheses are key directions. Additionally, integrating this approach with machine learning models, such as deep neural networks, to provide robust uncertainty quantification in AI systems represents a promising avenue. Exploring the theoretical limits of risk bounds under various model misspecifications will also be pursued.
AI Executive Summary
Bayesian inference has long been celebrated for its intuitive approach to uncertainty, allowing decision-makers to evaluate the expected performance of actions based on posterior distributions. However, its reliance on priors and models makes it vulnerable when these assumptions are misspecified or poorly chosen. This paper introduces the e-Posterior, a novel framework grounded in e-variables, which generalize likelihood ratios and provide a robust alternative. The core idea is to define an inverse e-variable, 1/Sθ, which forms a stochastic upper bound on risks, valid uniformly across all parameters and models. This approach guarantees frequentist risk bounds that are immune to model misspecification, addressing a critical gap in classical methods.
The methodology hinges on the supermartingale property of e-variables, ensuring that the risk bounds hold with high probability, regardless of prior beliefs. By designing collections of e-variables, the authors construct risk assessments that adapt to data-dependent weights and complex models, including nonparametric scenarios. Extensive simulations demonstrate that e-Posterior bounds maintain error rates within 5% at sample size n=100, outperforming traditional Bayesian credible intervals, especially under model deviations. In hypothesis testing, e-Posteriors replicate frequentist error control, providing a unified interpretation of classical tests and Bayesian procedures.
This work significantly advances statistical inference by offering a flexible, robust, and theoretically sound tool for risk assessment. Its potential applications span scientific research, industrial quality control, and financial risk management, where model uncertainty is pervasive. Despite current challenges in e-variable design and computational costs, the framework opens new pathways for future research, including scalable algorithms, high-dimensional extensions, and integration with machine learning. Overall, the e-Posterior framework promises to reshape how uncertainty and risk are quantified and managed in complex data environments, ensuring safer and more reliable decision-making across disciplines.
Deep Dive
Abstract
We develop a representation of a decision maker's uncertainty based on e-variables. Like the Bayesian posterior, this *e-posterior* allows for making predictions against arbitrary loss functions that may not be specified ex ante. Unlike the Bayesian posterior, it provides risk bounds that have frequentist validity irrespective of prior adequacy: if the e-collection (which plays a role analogous to the Bayesian prior) is chosen badly, the bounds get loose rather than wrong, making *e-posterior minimax* decision rules safer than Bayesian ones. The resulting *quasi-conditional paradigm* is illustrated by re-interpreting a previous influential partial Bayes-frequentist unification, *Kiefer-Berger-Brown-Wolpert conditional frequentist tests*, in terms of e-posteriors.