Bayesian Analysis of QENS data: From parameter determination to model selection

TL;DR

This paper advocates for a Bayesian approach to analyze quasielastic neutron scattering spectra, providing a probabilistic framework that captures parameter correlations, model likelihoods, and improves physical interpretation over traditional chi-square methods.

Contribution

It demonstrates how Bayesian analysis yields comprehensive probability distributions for parameters and models, enhancing data interpretation in QENS analysis.

Findings

Bayesian PDFs naturally account for parameter correlations.

Likelihood calculations enable probabilistic model selection.

The method improves parameter error estimation and model fitting.

Abstract

The extraction of any physical information from quasielastic neutron scattering spectra is generally done by fitting a model to the data by means of chi-square minimization procedure. However, as pointed out by the pioneering work of D.S. Sivia et al., also another probabilistic approach based on Bayes theorem can be employed. In a nutshell, the main difference between the classical chi-square minimization and the Bayesian approach is the way of expressing the final results: In the first case, the result is a set of values of parameters with a symmetric error and a figure of merit such as chi-square, whereas in the second case the results are presented as probability distribution functions (PDF) of both, parameters and merit figure. In this contribution, we demonstrate how final PDFs are obtained by exploring all possible combinations of parameters that are compatible with the experimental error. Three advantages of this method will be emphasized: First, correlations between parameters are automatically taken into account, which implies, for example, that parameter errors are correctly calculated, correlations show up in a natural way and ill defined parameters are immediately recognized from their PDF. Second, it is possible to calculate the likelihood of a determined physical model, and therefore to select the one among many that fits the data best with a minimal number of parameters, in a correctly defined probabilistic way.