When high-level features are provided by an explanans, they should ideally represent an abstracted, human-understandable form of the input, omitting unnecessary detail. This promotes parsimony by simplifying the input space. Such features can be obtained either through feature selection in FAs or through concept-level representations in CEs.
To quantify this abstraction, [Nguyen and Martínez (2020)] propose measuring the Mutual Information (MI) (see [Cover (1999)]) between the input and the explanans: $MI(x,e_{\theta,x,\hat{y}})$ A lower mutual information score suggests that the explanans captures less granular input detail and thus constitutes a more compact and parsimonious representation.