Lemma 21.6.2 · The gradient parametrizes the slope polytope
Lemma 21.6.2. (The gradient parametrizes the slope polytope). For every ϵ>0, ∇fϵ:NR→intP is a diffeomorphism.
Argument in the manuscript
Proof. Let wi be the corresponding positive exponential weights, normalized to sum to one. Differentiation gives ∇fϵ=i∑wiai,D2fϵ=ϵ−1(i∑wiaiait−aat),a=i∑wiai. For a nonzero vector v, its quadratic value is ϵ−1 times the weighted variance of the numbers ⟨ai,v⟩. This variance is positive because the ai affinely span MR. The Hessian is therefore positive definite, and the gradient is locally invertible and globally injective by strict convexity. Its value lies in intP since all weights are positive.
Given p∈intP, some Euclidean ball of radius c>0 about p lies in P. The support function hP(u)=maxi⟨ai,u⟩ satisfies hP(u)−⟨p,u⟩≥c∣u∣. Also fϵ≥f≥hP+minibi. Thus fϵ(u)−⟨p,u⟩ is coercive and has a minimizer. At that point its derivative vanishes, giving ∇fϵ=p. This proves surjectivity, and the inverse function theorem supplies a smooth inverse. ◻