Review pending

Lemma 21.6.2 · The gradient parametrizes the slope polytope

Lemma 21.6.2. (The gradient parametrizes the slope polytope). For every ϵ>0\epsilon>0, ∇fϵ:NR→int⁡P\nabla f_\epsilon:N_{\mathbb R}\to\operatorname{int}P is a diffeomorphism.

Argument in the manuscript

Proof. Let wiw_i be the corresponding positive exponential weights, normalized to sum to one. Differentiation gives ∇fϵ=∑iwiai,D2fϵ=ϵ−1(∑iwiaiait−a‾ a‾t),a‾=∑iwiai.\nabla f_\epsilon=\sum_iw_i a_i, \quad D^2 f_\epsilon=\epsilon^{-1} \left(\sum_iw_i a_i a_i^t-\overline a\,\overline a^t\right), \qquad \overline a=\sum_iw_i a_i. For a nonzero vector vv, its quadratic value is ϵ−1\epsilon^{-1} times the weighted variance of the numbers ⟨ai,v⟩\langle a_i,v\rangle. This variance is positive because the aia_i affinely span MRM_{\mathbb R}. The Hessian is therefore positive definite, and the gradient is locally invertible and globally injective by strict convexity. Its value lies in int⁡P\operatorname{int}P since all weights are positive.

Given p∈int⁡Pp\in\operatorname{int}P, some Euclidean ball of radius c>0c>0 about pp lies in PP. The support function hP(u)=max⁡i⟨ai,u⟩h_P(u)=\max_i\langle a_i,u\rangle satisfies hP(u)−⟨p,u⟩≥c∣u∣h_P(u)-\langle p,u\rangle\geq c|u|. Also fϵ≥f≥hP+min⁡ibif_\epsilon\geq f\geq h_P+\min_i b_i. Thus fϵ(u)−⟨p,u⟩f_\epsilon(u)-\langle p,u\rangle is coercive and has a minimizer. At that point its derivative vanishes, giving ∇fϵ=p\nabla f_\epsilon=p. This proves surjectivity, and the inverse function theorem supplies a smooth inverse. ◻

Citation

Non-Archimedean Geometry Foundations, Tag NAG-900409.

Proof in chapter