Review pending

Theorem 21.27.1 · A finite measure with prescribed masses

Theorem 21.27.1. (A finite measure with prescribed masses). Let x1,…,xs∈NRx_1,\ldots,x_s\in N_\mathbb R be distinct, and let mi>0m_i>0 with ∑imi=vol⁡(P)\sum_i m_i=\operatorname{vol}(P). There is a finite convex function ff with f−hPf-h_P bounded and MA⁡R(f)=∑i=1smiδxi.\operatorname{MA}_{\mathbb R}(f)=\sum_{i=1}^s m_i\delta_{x_i}. It is obtained by minimizing, modulo a common additive constant in a=(ai)a=(a_i), the function Φ(a)=∑imiai+∫Pmax⁡i(⟨p,xi⟩−ai) dp.\begin{equation} \Phi(a)=\sum_i m_i a_i+ \int_P\max_i\bigl(\langle p,x_i\rangle-a_i\bigr)\,dp. \end{equation} The minimizer is unique modulo that common constant.

Argument in the manuscript

Proof. Write ga(p)=max⁡i(⟨p,xi⟩−ai)g_a(p)=\max_i(\langle p,x_i\rangle-a_i) on PP. The function Φ\Phi is convex and continuous. Because the total masses agree, Φ(a+c1)=Φ(a)\Phi(a+c\mathbf 1)=\Phi(a). Normalize min⁡iai=0\min_i a_i=0, so all ai≥0a_i\geq 0. Choose CC with ∣⟨p,xi⟩∣≤C|\langle p,x_i\rangle|\leq C for all p∈Pp\in P and all ii. Some aj=0a_j=0, and hence ga(p)≥⟨p,xj⟩≥−Cg_a(p)\geq\langle p,x_j\rangle\geq-C. Consequently Φ(a)≥∑imiai−Cvol⁡(P).\Phi(a)\geq\sum_i m_i a_i-C\operatorname{vol}(P). Since every mim_i is positive, the normalized sublevel sets are bounded and closed. A minimum exists. This coercivity argument does not require any general-position condition on the xix_i.

The cells associated with aa are Ci(a)={p∈P:⟨p,xi⟩−ai≥⟨p,xj⟩−aj for every j}.C_i(a)=\{p\in P:\langle p,x_i\rangle-a_i \geq\langle p,x_j\rangle-a_j\text{ for every }j\}. The equality set of two distinct planes lies in a hyperplane and has Lebesgue measure zero. Almost every p∈Pp\in P therefore has one active index. Differentiating the integral in a coordinate direction is justified by dominated convergence, since the difference quotient of a maximum is bounded by one. It gives ∂Φ∂ai=mi−vol⁡(Ci(a)).\frac{\partial\Phi}{\partial a_i}=m_i-\operatorname{vol}(C_i(a)). At a minimizer, the derivative vanishes in every direction in Rs\mathbb R^s. One may take an unrestricted representative, since the functional is constant along 1\mathbf 1. Thus each cell has the prescribed volume.

Extend gag_a by +∞+\infty outside PP, and put f(x)=sup⁡p∈P(⟨p,x⟩−ga(p)).f(x)=\sup_{p\in P}\bigl(\langle p,x\rangle-g_a(p)\bigr). The function gag_a is continuous on PP, convex, and closed after this extension. Hence f∗=gaf^*=g_a. Boundedness of gag_a on PP gives boundedness of f−hPf-h_P. Since ga(p)≥⟨p,xi⟩−aig_a(p)\geq\langle p,x_i\rangle-a_i, one has f(xi)≤aif(x_i)\leq a_i. Equality holds on the nonempty cell Ci(a)C_i(a), and therefore f(xi)=ai,∂f(xi)=Ci(a).f(x_i)=a_i,\qquad \partial f(x_i)=C_i(a). Almost every point in the interior of PP belongs to a unique active cell and is a differentiability point of gag_a, with gradient xix_i. The subgradient duality p∈∂f(x)⟺x∈∂ga(p)p\in\partial f(x)\Longleftrightarrow x\in\partial g_a(p) then implies that subgradients over points x∉{x1,…,xs}x\notin\{x_1,\ldots,x_s\} occupy only a null subset of PP. The boundary of a full-dimensional convex body also has Lebesgue measure zero. Thus the Alexandrov measure of ff is exactly the asserted atomic measure.

For uniqueness of the minimizer, let a,ba,b be two minimizers. Convexity implies equality in the integral convexity inequality at their midpoint. The pointwise nonnegative difference between the two sides consequently vanishes almost everywhere. At a point where the active indices of both maxima are unique, equality forces those indices to be the same. On the resulting full-measure set, ga−gbg_a-g_b belongs to the finite set {bi−ai}\{b_i-a_i\}. By continuity, this holds everywhere on PP. Since PP is connected, ga−gbg_a-g_b is constant. Each Ci(a)C_i(a) has positive volume, so its interior contains a point with the same unique active index for bb; hence every bi−aib_i-a_i equals that constant. ◻

Citation

Non-Archimedean Geometry Foundations, Tag NAG-900443.

Proof in chapter