Uncertainty components in profile likelihood fits

TL;DR

This paper clarifies the distinction between impact and actual uncertainty components in profile-likelihood fits, proposing methods to accurately determine these components for better uncertainty propagation in high-energy physics analyses.

Contribution

It introduces new methods to identify true uncertainty components in profile-likelihood fits, improving the accuracy of uncertainty decomposition beyond impact estimates.

Findings

Impacts are not equivalent to actual uncertainty components.

Methods are established to determine true uncertainty components.

Enhanced accuracy in uncertainty propagation for physics analyses.

Abstract

When a measurement of a physical quantity is reported, the total uncertainty is usually decomposed into statistical and systematic uncertainties. This decomposition is not only useful to understand the contributions to the total uncertainty, but also required to propagate these contributions in subsequent analyses, such as combinations or interpretation fits including results from other measurements or experiments. In profile-likelihood fits, widely applied in high-energy physics analyses, contributions of systematic uncertainties are routinely quantified using "impacts", which are not adequate for such applications. We discuss the difference between impacts and actual uncertainty components, and establish methods to determine the latter in a wide range of statistical models.