On a Brunn-Minkowski principle for first-order information functionals

Abstract

In this work we present a general principle showing that eve\-ry positive homogeneous functional on an abstract convex cone satis\-fying a suitable Brunn-Minkowski inequality gives rise to an information functional, defined as the quotient of the functional by its relative Minkowski content, which satisfies a linear Brunn-Minkowski inequality. We also establish the dual counterpart of this principle for functionals satisfying some convexity, yielding reverse linear Brunn-Minkowski ine\-qualities for the associated information functionals, and their geometric applications in the dual Brunn-Minkowski theory. As applications, we recover a classical result by Gianno\-poulos, Hartzoulaki and Paouris (2002) for the ratio of two consecutive quermassin\-tegrals (which further extends to the setting of mixed discriminants of positive definite symme\-tric matrices) by a different and first-order argument, obtai\-ning moreover a characterization of the equality case. Our approach also yields a new $L_p$ extension ($p \geqslant 1$) of the Brunn-Minkowski inequality by Dembo, Cover and Thomas for the classical information functional, where the relative Minkowski content is now expressed in terms of Lutwak's first $L_p$ quermassintegral. Finally, we show that homogeneity is necessary within this principle by showing a counterexample to the expected result for the information functional associated to the standard Gaussian measure.