Review pending

Proposition 11.12.12 · Gaussian boundary limit for a finite maximum

Proposition 11.12.12. (Gaussian boundary limit for a finite maximum). Let EE be a finite-dimensional Euclidean space, let v1,…,vsv_1,\ldots,v_s be linearly independent with s≥1s\geq 1, and put δ=∑iR≥0vi\delta=\sum_i\mathbf R_{\geq 0}v_i, T=span⁡(δ)T=\operatorname{span}(\delta) and Q=T⊥Q=T^\perp. Let p:E→Qp:E\to Q be orthogonal projection. Suppose h(x)=max⁡1≤a≤m⟨ma,x⟩≥0,h∣δ=0.h(x)=\max_{1\leq a\leq m}\langle m_a,x\rangle\geq0, \qquad h|_\delta=0. Set Aδ={a:⟨ma,vi⟩=0 for every i}A_\delta=\{a:\langle m_a,v_i\rangle=0\text{ for every }i\}, and hδ(z)=max⁡a∈Aδ⟨ma,z⟩h^\delta(z)=\max_{a\in A_\delta}\langle m_a,z\rangle for z∈Qz\in Q. This set of active indices is nonempty. If ρE\rho_E and ρQ\rho_Q are the standard normalized Gaussian densities and H=h∗ρEH=h*\rho_E, then lim⁡min⁡iti→+∞H(∑itivi+z)=∫Qhδ(z−b)ρQ(b) db.\begin{equation} \lim_{\min_i t_i\to+\infty} H\left(\sum_i t_i v_i+z\right) =\int_Q h^\delta(z-b)\rho_Q(b)\,db. \end{equation} The convergence is uniform for zz in compact subsets of QQ. No monotonicity of the approaching parameters is assumed. Moreover H(x+v)≤H(x)H(x+v)\leq H(x) for v∈δv\in\delta. If h(−v)>0h(-v)>0 for a nonzero v∈δv\in\delta, the inequality is strict.

Argument in the manuscript

Proof. Every ⟨ma,vi⟩\langle m_a,v_i\rangle is nonpositive. Since h(∑ivi)=0h(\sum_i v_i)=0, a maximizing index there has all these coefficients zero, proving that AδA_\delta is nonempty. For y=∑itiviy=\sum_i t_i v_i, every inactive linear form in h(y+z−w)h(y+z-w) tends to −∞-\infty as min⁡iti→∞\min_i t_i\to\infty; active forms annihilate TT. Consequently, for each fixed z,wz,w, h(y+z−w)⟶hδ(z−p(w)).h(y+z-w)\longrightarrow h^\delta(z-p(w)). Write L=max⁡a∥ma∥L=\max_a\|m_a\|. The function hh is LL-Lipschitz and subadditive. For ti≥0t_i\geq 0, its nonnegativity and h(y)=0h(y)=0 give the integrable bound 0≤h(y+z−w)≤h(z−w)≤L(∥z∥+∥w∥).0\leq h(y+z-w)\leq h(z-w)\leq L(\|z\|+\|w\|). Dominated convergence now gives the limit of the integral. Orthogonal decomposition E=T⊕QE=T\oplus Q factors the normalized Gaussian into the product of the two Gaussian densities; integration in TT gives exactly (11.12.3). Both sides, as functions of zz, have Lipschitz constant at most LL. A finite net on a compact subset of QQ upgrades pointwise convergence to uniform convergence. Equivalently, any failure of uniform convergence would give a sequence of parameters and a convergent sequence of points contradicting that bound and the already proved pointwise limit.

Finally, h(x+v)≤h(x)+h(v)=h(x)h(x+v)\leq h(x)+h(v)=h(x), and integration gives monotonicity. If h(−v)>0h(-v)>0, the difference h(x−w)−h(x+v−w)h(x-w)-h(x+v-w) is a nonnegative continuous function of ww and is positive at w=x+vw=x+v: its value there is h(−v)−h(0)>0h(-v)-h(0)>0. It is therefore positive on a nonempty open set of positive Gaussian measure. Its integral is strictly positive. This proves strictness without a separate strict-convexity argument. ◻

Citation

Non-Archimedean Geometry Foundations, Tag NAG-900172.

Proof in chapter