A unifying view of contrastive learning, importance sampling, and bridge sampling for energy-based models

A new study presents a unified mathematical framework that connects four distinct estimation methods for energy-based models, a class of probabilistic models where a component of the likelihood cannot be evaluated explicitly [1][2]. The work, submitted to the arXiv preprint repository in April 2026 and revised in June 2026, links noise contrastive estimation, reverse logistic regression, multiple importance sampling, and bridge sampling under a single theoretical structure [1][2]. Energy-based models have become an important class of probabilistic models in recent decades, but parameter estimation in them is challenging for conventional inference methods because a component of the likelihood is intractable [1][2]. The author, Luca Martino, shows that these four methods are equivalent under specific conditions [1][2]. The unified perspective clarifies relationships among existing methods and enables the development of new estimators, with the potential to improve statistical and computational efficiency [1][2]. The study also helps elucidate the success of noise contrastive estimation in terms of its flexibility and robustness, while identifying scenarios in which its performance can be further improved [1][2]. The MATLAB code used in the numerical experiments has been made freely available to support the reproducibility of the results [1][2]. arXiv, where the paper appears, is an open-access repository of electronic preprints and postprints that are approved for posting after moderation but are not peer reviewed [8]. It was begun on August 14, 1991, and by the end of 2021 had surpassed two million articles [8]. The submission rate as of November 2024 was about 24,000 articles per month [8]. The first version of the paper was 144 KB, and the revised second version was 145 KB [1]. The statistical methods addressed in the paper sit within the broader discipline of statistics, which concerns the collection, organization, analysis, interpretation, and presentation of data [3]. Inferential statistics, the branch most relevant to this work, draws conclusions from data that are subject to random variation using the framework of probability theory [3]. Energy-based models themselves are often deployed in artificial intelligence research, a field that develops methods and software enabling machines to perceive their environment and use learning and intelligence to take actions that maximize their chances of achieving defined goals [4]. AI research has employed methods based on statistics, operations research, and economics, among other disciplines [4]. The paper's methodological contribution arrives as scientific fields continue to refine approaches to hypothesis testing and experimental validation [7]. The scientific method involves making conjectures, predicting the logical consequences of a hypothesis, and then carrying out experiments or empirical observations based on those predictions [7]. The unified framework offered in this work provides a new tool for researchers working with energy-based models to test and compare estimation strategies within a common theoretical structure [1][2].