Information Geometry meets Functional ANOVA: An Exact Fisher-Information Decomposition, with an Application to Radio Luminosity Functions

Abstract

Information geometry represents a fitted model as a manifold whose metric is the Fisher information. We show that, for a nonlinear regression model, this metric decomposes exactly across the Hoeffding-Sobol' (functional ANOVA) channels of its covariates. Treating the score contributions of main effects and interactions as vectors in a Hilbert space, their Gram matrix reproduces the (Gauss-Newton) Fisher information metric exactly, and this identity extends to an exact, additive decomposition of…