The Bayes factor surface for searches for new physics

Andrew Fowlie

DOI: https://doi.org/10.1140/epjc/s10052-024-12792-9

2024-04-22

Abstract:The Bayes factor surface is a new way to present results from experimental searches for new physics. Searches are regularly expressed in terms of phenomenological parameters - such as the mass and cross-section of a weakly interacting massive particle. Bayes factor surfaces indicate the strength of evidence for or against models relative to the background only model in terms of the phenomenological parameters that they predict. They provide a clear and direct measure of evidence, may be easily reinterpreted, but do not depend on choices of prior or parameterization. We demonstrate the Bayes factor surface with examples from dark matter, cosmology, and collider physics.

High Energy Physics - Phenomenology,Cosmology and Nongalactic Astrophysics,High Energy Physics - Experiment,Data Analysis, Statistics and Probability

What problem does this paper attempt to address?

This paper proposes a new method for presenting the search results of new physical phenomena in experiments, called the Bayes factor surface. In current practice, experimental results are often represented by a few numbers and charts based on phenomenological parameters such as the mass and cross-section of dark matter particles. The Bayes factor surface directly measures the evidence strength relative to the background model based on these parameters, making it easy to reinterpret and not relying on prior selection or parameterization.

The paper points out that there are two main goals in statistical analysis: testing (determining which models are excluded) and measurement (determining which parameters are excluded). In the frequentist framework, measurement is usually done by constructing confidence intervals, and testing is done by calculating p-values. In the Bayesian framework, measurement is done by considering the posterior distribution of unknown parameters, and testing is done by comparing the Bayes factors of two models.

In the paper, the authors propose using the Bayes factor surface to represent the evidence level of models relative to the background model, especially when experimental results are expressed in terms of phenomenological parameters. For example, if the prediction of a theory lies outside the contour with a Bayes factor of 1/100, then this theory is at least 100 times less credible than the background model. The paper also discusses the interpretation of the Bayes factor surface, computational challenges, comparisons with other statistical methods, and provides examples from the fields of dark matter, cosmology, and particle physics to illustrate its application.

In summary, this paper aims to address how to effectively present the evidence for new physical phenomena in experimental data and proposes a new statistical tool, the Bayes factor surface, to help scientists better understand and evaluate the consistency between different theoretical models and experimental observations.