Conservative covariance for general-purpose nuclear data evaluation

Abstract

General-purpose evaluated nuclear data libraries requires the production of covariance data that can be safely used for any applications. Propagated uncertainty must be conservative regardless on which specific nuclear data an application is sensitive to. In this paper, the usual analytical and Bayesian marginalization techniques used for the production of covariance data are reviewed, their mathematical justification and practical implementation are discussed. An alternative technique is also presented, it is based on a spectral decomposition of the posterior Bayesian correlation matrix. This new method is shown to be more valid than the usual analytical and Bayesian marginalization techniques and exhibits the required properties for general-purpose applications. Several options in the spectral decomposition method are detailed, allowing a “strict conservatism” to a “best-estimate” justified when experimental data are well reproduced. These options are illustrated on a test-case and compared with the usual methods. The spectral decomposition method is also shown to be able to identify outliers in experimental data.