ScalingStacks

Verified tagged author-source HTML · 1912.02360v1 · cited publication edition alignment unverified.

00R4

Proposition 3.27. In the Fermat case, if uu is a locally convex function on ∂Δλ∨\partial\Delta_{\lambda}^{\vee}, which is invariant under the permutation group. Then uu satisfies the extension property.

00R5

Proof. We need to prove the characterisation in Prop. 3.19. Without loss of generality Lλ​(x)L_{\lambda}(x) is achieved by ⟨m0,x⟩+λ⁡(m0)\langle m^{0},x\rangle+\lambda(m^{0}). We need to find p∈Δp\in\Delta, such that u⁡(y)−u⁡(x)≥⟨p,y−x⟩.u(y)-u(x)\geq\langle p,y-x\rangle. For this we study the gradient of the function um0u_{m^{0}} on the various ww-charts.

First, notice for x′,y′x^{\prime},y^{\prime} on the face {Lλ=⟨m0,⟩+λ(m0)}\{L_{\lambda}=\langle m^{0},\rangle+\lambda(m^{0})\}, namely the convex hull of w1,…​wn+1w_{1},\ldots w_{n+1}, the vector y′−x′y^{\prime}-x^{\prime} is parallel to the face, and by convexity of um0u_{m^{0}} the directional derivative ∇um0⋅(y′−x′)\nabla u_{m^{0}}\cdot(y^{\prime}-x^{\prime}) is monotone along the path from x′x^{\prime} to y′y^{\prime}, so must be maximized at y′y^{\prime}. In particular we consider such line segments on the face parallel to wi−wjw_{i}-w_{j} for i,j≥1i,j\geq 1. By the discrete symmetry, ∇um0⋅(wj−wi)\nabla u_{m^{0}}\cdot(w_{j}-w_{i}) must be zero on the plane of reflection bisecting the face. Thus for i,j≥1i,j\geq 1, i≠ji\neq j, the subset of the face

{∇um0⋅wi≥∇um0⋅wj}∩{Lλ=⟨m0,⟩+λ(m0)}\{\nabla u_{m^{0}}\cdot w_{i}\geq\nabla u_{m^{0}}\cdot w_{j}\}\cap\{L_{\lambda}=\langle m^{0},\rangle+\lambda(m^{0})\}

agrees exactly with the half of the face containing wiw_{i}. Therefore the subset of face

{∇um0⋅wi≥∇um0⋅wj,∀j≥1}\{\nabla u_{m^{0}}\cdot w_{i}\geq\nabla u_{m^{0}}\cdot w_{j},\forall j\geq 1\}

is exactly the intersection of Star​(wi)\text{Star}(w_{i}) with the face. Without loss of generality xx lies in Star​(w1)\text{Star}(w_{1}).

We follow the notation in the proof of Prop. 3.26. In the w1w_{1}-chart, denote the gradient of um0u_{m^{0}} as p→\vec{p}, so that for yy in the w1w_{1}-chart,

um0​(y)−um0​(x)≥p→⋅(y−x)w1.u_{m^{0}}(y)-u_{m^{0}}(x)\geq\vec{p}\cdot(y-x)_{w_{1}}.

A priori p→\vec{p} lives in Mℝ/ℝ​m0M_{\mathbb{R}}/\mathbb{R}m^{0}. We lift p→\vec{p} to MℝM_{\mathbb{R}} by demanding ⟨p→,w1⟩=0\langle\vec{p},w_{1}\rangle=0, so by the above discussion ⟨p→,wi⟩≤0\langle\vec{p},w_{i}\rangle\leq 0 for i≥1.i\geq 1. Define p=p→+m0p=\vec{p}+m_{0}, then ⟨p,wi⟩≤1\langle p,w_{i}\rangle\leq 1 for all i≥1i\geq 1. We regard p∈Mℝp\in M_{\mathbb{R}} as the gradient of uu at xx, and write p=∇up=\nabla u as a function of xx. This construction can be made on other faces as well, and on the intersection of two faces the definitions are compatible.

We claim p∈Δp\in\Delta: it suffices to show ⟨p,w0⟩≤1\langle p,w_{0}\rangle\leq 1. Notice w0=−∑1n+1wi=∑i=2n+1(w1−wi)−(n+1)w1w_{0}=-\sum_{1}^{n+1}w_{i}=\sum_{i=2}^{n+1}(w_{1}-w_{i})-(n+1)w_{1}. Consider the line segment in the face joining xx to the boundary of the face in the direction ∑i=2n+1(w1−wi)\sum_{i=2}^{n+1}(w_{1}-w_{i}), which stays inside Star​(w1)\text{Star}(w_{1}), and along which ∇u⋅∑i=2n+1(w1−wi)\nabla u\cdot\sum_{i=2}^{n+1}(w_{1}-w_{i}) increases, or equivalently ⟨∇u,w0⟩\langle\nabla u,w_{0}\rangle increases. But the boundary of the face {Lλ=⟨m0,⟩+λ(m0)}\{L_{\lambda}=\langle m^{0},\rangle+\lambda(m^{0})\} lies also on a different face, and we can use the information from this new face to deduce ⟨∇u,w0⟩≤1\langle\nabla u,w_{0}\rangle\leq 1 there.

By construction for yy in the w1w_{1}-chart,

u⁡(y)−u⁡(x)≥⟨p→​(x),(y−x)w1⟩+⟨m0,y−x⟩=⟨∇u​(x),y−x⟩.u(y)-u(x)\geq\langle\vec{p}(x),(y-x)_{w_{1}}\rangle+\langle m_{0},y-x\rangle=\langle\nabla u(x),y-x\rangle.

We claim that in fact u⁡(y)−u⁡(x)≥⟨∇u​(x),y−x⟩u(y)-u(x)\geq\langle\nabla u(x),y-x\rangle holds for all y∈∂Δλ∨y\in\partial\Delta_{\lambda}^{\vee}. We are left to check for yy on the face {Lλ=⟨m1,⟩+λ(m1)}\{L_{\lambda}=\langle m^{1},\rangle+\lambda(m^{1})\}, namely the complement of the w1w_{1}-chart. Consider the wiw_{i}-chart for i>1i>1. We can write according to the decomposition Nℝ=(m1)⟂⊕ℝ​wiN_{\mathbb{R}}=(m^{1})^{\perp}\oplus\mathbb{R}w_{i}, that

y−x=(y−x)wi,m1+⟨m1,y−x⟩​wi.y-x=(y-x)_{w_{i},m^{1}}+\langle m^{1},y-x\rangle w_{i}.

By local convexity, in the wiw_{i}-chart um1u_{m^{1}} is convex, so there is some p→′\vec{p}^{\prime}, such that for any y′y^{\prime} in the wiw_{i}-chart

um1​(y′)−um1​(x)≥p→′⋅(y′−x)wi,m1.u_{m^{1}}(y^{\prime})-u_{m^{1}}(x)\geq\vec{p}^{\prime}\cdot(y^{\prime}-x)_{w_{i},m^{1}}.

But a gradient vector of um1u_{m^{1}} at xx is ∇u​(x)−m1\nabla u(x)-m^{1}, so we may take p→′=∇u​(x)−m1\vec{p}^{\prime}=\nabla u(x)-m^{1}. Thus

u⁡(y)−u⁡(x)≥p→′⋅(y−x)wi,m1+⟨m1,y−x⟩=⟨∇u​(x),y−x⟩−⟨p→′,wi⟩​⟨m1,y−x⟩.u(y)-u(x)\geq\vec{p}^{\prime}\cdot(y-x)_{w_{i},m^{1}}+\langle m^{1},y-x\rangle=\langle\nabla u(x),y-x\rangle-\langle\vec{p}^{\prime},w_{i}\rangle\langle m^{1},y-x\rangle.

Now ⟨m1,y−x⟩≥0\langle m^{1},y-x\rangle\geq 0 as in the proof of Prop. 3.26, and ⟨p→′,wi⟩≤0\langle\vec{p}^{\prime},w_{i}\rangle\leq 0 by ∇u∈Δ\nabla u\in\Delta. This implies u⁡(y)−u⁡(x)≥⟨∇u​(x),y−x⟩u(y)-u(x)\geq\langle\nabla u(x),y-x\rangle as required.

We have verified the characterisation in Prop. 3.19, hence the extension property. ∎

Original mathematics by the credited authors. Source collection and HTML conversion remain in progress.