Densities with Gaussian Tails

Abstract

Consider densities fi(t), for i = 1, …, d, on the real line which have thin tails in the sense that, for each i, fi(t) ∼ γi(t)e−ψi(t), where γi behaves roughly like a constant and ψi is convex, C2, with ψ′ → ∞ and ψ″ > 0 and l/√ψ″ is self-neglecting. (The latter is an asymptotic variation condition.) Then the convolution is of the same form ft * … *fd(t) ∼ γ(t)e − ψ(t) Formulae for γ, ψ are given in terms of the factor densities and involve the conjugate transform and infimal convolution of convexity theory. The derivations require embedding densities in exponential families and showing that the assumed form of the densities implies asymptotic normality of the exponential families.