TheoremBase

Adds the quadratic to obtain a convex function, applies the local Lipschitz theorem for convex functions at an interior point, and subtracts the quadratic, which is Lipschitz on a bounded set; the continuity estimate then follows by choosing the radius small.

Proof

Throughout, G:U→RG:U\to\mathbb{R} denotes the function given by

G(z)=φ(z)+μ2 ∥z∥2,G(z)=\varphi(z)+\frac{\mu}{2}\,\lVert z\rVert^{2},

which is convex on UU precisely because φ\varphi is semiconvex on UU with constant μ\mu.

Proof of claim 1. Let x0∈Ux_{0}\in U. The set UU belongs to the topology of The Open Subsets of a Metric Space Form a Topology and satisfies x0∈Ux_{0}\in U and U⊆UU\subseteq U, so x0x_{0} is an interior point of UU in Rn\mathbb{R}^{n} by Interior of a Subset of a Topological Space. Since UU is convex and GG is convex on UU, A Convex Function is Lipschitz on a Ball around an Interior Point supplies ρ,L0∈R\rho,L_{0}\in\mathbb{R} with 0<ρ0<\rho and 0≤L00\le L_{0} such that BˉdE(x0,ρ)⊆U\bar{B}_{d_{E}}(x_{0},\rho)\subseteq U and

∣G(y)−G(x)∣≤L0 ∥y−x∥for all x,y∈BˉdE(x0,ρ).\bigl|G(y)-G(x)\bigr|\le L_{0}\,\lVert y-x\rVert\qquad\text{for all }x,y\in\bar{B}_{d_{E}}(x_{0},\rho).

Put R=∥x0∥+ρR=\lVert x_{0}\rVert+\rho, a nonnegative real number. If x∈BˉdE(x0,ρ)x\in\bar{B}_{d_{E}}(x_{0},\rho) then ∥x−x0∥=dE(x0,x)≤ρ\lVert x-x_{0}\rVert=d_{E}(x_{0},x)\le\rho by Closed Ball in a Metric Space and claim 2 of Properties of the Euclidean Norm on Rn\mathbb{R}^n: Square, Distance, Vanishing, Coordinate Bounds, Homogeneity, Triangle Inequalities and Expansion of a Square, whence by the triangle inequality (claim 6 there)

∥x∥=∥x0+(x−x0)∥≤∥x0∥+∥x−x0∥≤R.\lVert x\rVert=\lVert x_{0}+(x-x_{0})\rVert\le\lVert x_{0}\rVert+\lVert x-x_{0}\rVert\le R .

Next, for arbitrary x,y∈Rnx,y\in\mathbb{R}^{n} the triangle inequality gives ∥y∥≤∥x∥+∥y−x∥\lVert y\rVert\le\lVert x\rVert+\lVert y-x\rVert and ∥x∥≤∥y∥+∥x−y∥\lVert x\rVert\le\lVert y\rVert+\lVert x-y\rVert, while ∥x−y∥=∥(−1)(y−x)∥=∥y−x∥\lVert x-y\rVert=\lVert(-1)(y-x)\rVert=\lVert y-x\rVert by claim 5 there; hence

∣ ∥y∥−∥x∥ ∣≤∥y−x∥.\bigl|\,\lVert y\rVert-\lVert x\rVert\,\bigr|\le\lVert y-x\rVert .

Consequently, for x,y∈BˉdE(x0,ρ)x,y\in\bar{B}_{d_{E}}(x_{0},\rho), using 0≤∥x∥≤R0\le\lVert x\rVert\le R and 0≤∥y∥≤R0\le\lVert y\rVert\le R,

∣ ∥y∥2−∥x∥2 ∣=∣ ∥y∥−∥x∥ ∣⋅(∥y∥+∥x∥)≤2R ∥y−x∥.\bigl|\,\lVert y\rVert^{2}-\lVert x\rVert^{2}\,\bigr|=\bigl|\,\lVert y\rVert-\lVert x\rVert\,\bigr|\cdot\bigl(\lVert y\rVert+\lVert x\rVert\bigr)\le 2R\,\lVert y-x\rVert .

Since φ(z)=G(z)−μ2∥z∥2\varphi(z)=G(z)-\tfrac{\mu}{2}\lVert z\rVert^{2} for z∈Uz\in U, the triangle inequality for the absolute value gives, for all x,y∈BˉdE(x0,ρ)x,y\in\bar{B}_{d_{E}}(x_{0},\rho),

∣φ(y)−φ(x)∣≤∣G(y)−G(x)∣+μ2∣ ∥y∥2−∥x∥2 ∣≤(L0+μR)∥y−x∥.\bigl|\varphi(y)-\varphi(x)\bigr|\le\bigl|G(y)-G(x)\bigr|+\frac{\mu}{2}\bigl|\,\lVert y\rVert^{2}-\lVert x\rVert^{2}\,\bigr|\le\bigl(L_{0}+\mu R\bigr)\lVert y-x\rVert .

So claim 1 holds with this ρ\rho and with L=L0+μRL=L_{0}+\mu R, which is nonnegative.

Proof of claim 2. Let S⊆US\subseteq U, let x∈Sx\in S and let ε∈R\varepsilon\in\mathbb{R} with 0<ε0<\varepsilon. Apply claim 1 at the point x0=xx_{0}=x of UU to obtain ρ,L∈R\rho,L\in\mathbb{R} with 0<ρ0<\rho, 0≤L0\le L, BˉdE(x,ρ)⊆U\bar{B}_{d_{E}}(x,\rho)\subseteq U and

∣φ(w)−φ(z)∣≤L ∥w−z∥for all z,w∈BˉdE(x,ρ).\bigl|\varphi(w)-\varphi(z)\bigr|\le L\,\lVert w-z\rVert\qquad\text{for all }z,w\in\bar{B}_{d_{E}}(x,\rho).

Since 0≤L0\le L we have 0<L+10<L+1, so ε/(L+1)\varepsilon/(L+1) is a positive real number; let δ\delta be the smaller of ρ\rho and ε/(L+1)\varepsilon/(L+1), a positive real number.

Let y∈Sy\in S satisfy ∥y−x∥<δ\lVert y-x\rVert<\delta. Then dE(x,y)=∥y−x∥<δ≤ρd_{E}(x,y)=\lVert y-x\rVert<\delta\le\rho, so y∈BˉdE(x,ρ)y\in\bar{B}_{d_{E}}(x,\rho); also x∈BˉdE(x,ρ)x\in\bar{B}_{d_{E}}(x,\rho) by claim 1 of Elementary Properties of the Closed Ball in a Metric Space. Therefore, using 0≤L0\le L and ∥y−x∥<δ\lVert y-x\rVert<\delta,

∣φ(y)−φ(x)∣≤L ∥y−x∥≤L δ≤LL+1 ε<ε,\bigl|\varphi(y)-\varphi(x)\bigr|\le L\,\lVert y-x\rVert\le L\,\delta\le\frac{L}{L+1}\,\varepsilon<\varepsilon,

the last inequality because 0≤L<L+10\le L<L+1 gives L/(L+1)<1L/(L+1)<1 and 0<ε0<\varepsilon. This is the assertion of claim 2.

Citations

Loading…

Dependencies

Uses0

Loading…

Comments

Log in to comment.

Loading…