The ball B ˉ \bar{B} B ˉ is nonempty, since x ^ ∈ B ˉ \hat{x}\in\bar{B} x ^ ∈ B ˉ by claim 1 of Elementary Properties of the Closed Ball in a Metric Space , and compact in R n \mathbb{R}^{n} R n by claim 2 of A Closed Euclidean Ball is Convex and Compact . It is bounded and closed by claims 2 and 3 of Elementary Properties of the Closed Ball in a Metric Space . For p ∈ R n p\in\mathbb{R}^{n} p ∈ R n let ψ p : B ˉ → R \psi_{p}:\bar{B}\to\mathbb{R} ψ p : B ˉ → R be given by ψ p ( x ) = φ ( x ) + p ⋅ x \psi_{p}(x)=\varphi(x)+p\cdot x ψ p ( x ) = φ ( x ) + p ⋅ x ; thus M ( p ) M(p) M ( p ) is the set of points of B ˉ \bar{B} B ˉ at which ψ p \psi_{p} ψ p attains its greatest value on B ˉ \bar{B} B ˉ . Throughout we use the bilinearity and symmetry of the dot product (The Dot Product on R n \mathbb{R}^n R n Is Symmetric, Bilinear and Positive Definite ) and the Cauchy-Schwarz inequality The Cauchy-Schwarz Inequality on R n \mathbb{R}^n R n .
Step 1 (the perturbed functions are continuous). Let p ∈ R n p\in\mathbb{R}^{n} p ∈ R n and S ⊆ B ˉ S\subseteq\bar{B} S ⊆ B ˉ . We check that the restriction of ψ p \psi_{p} ψ p to S S S satisfies the continuity hypothesis of Extreme Value Theorem on a Compact Subset of a Metric Space . Let x ∈ S x\in S x ∈ S and let ε ∈ R \varepsilon\in\mathbb{R} ε ∈ R with 0 < ε 0<\varepsilon 0 < ε . For y ∈ B ˉ y\in\bar{B} y ∈ B ˉ ,
∣ ψ p ( y ) − ψ p ( x ) ∣ ≤ ∣ φ ( y ) − φ ( x ) ∣ + ∣ p ⋅ ( y − x ) ∣ ≤ ∣ φ ( y ) − φ ( x ) ∣ + ∥ p ∥ ∥ y − x ∥ . \bigl|\psi_{p}(y)-\psi_{p}(x)\bigr|\le\bigl|\varphi(y)-\varphi(x)\bigr|+\bigl|p\cdot(y-x)\bigr|\le\bigl|\varphi(y)-\varphi(x)\bigr|+\lVert p\rVert\,\lVert y-x\rVert . ψ p ( y ) − ψ p ( x ) ≤ φ ( y ) − φ ( x ) + p ⋅ ( y − x ) ≤ φ ( y ) − φ ( x ) + ∥ p ∥ ∥ y − x ∥ .
By (H1) there is a positive real δ 1 \delta_{1} δ 1 such that every z ∈ B ˉ z\in\bar{B} z ∈ B ˉ with ∥ z − x ∥ < δ 1 \lVert z-x\rVert<\delta_{1} ∥ z − x ∥ < δ 1 satisfies ∣ φ ( z ) − φ ( x ) ∣ < ε / 2 |\varphi(z)-\varphi(x)|<\varepsilon/2 ∣ φ ( z ) − φ ( x ) ∣ < ε /2 . Let δ \delta δ be the smaller of δ 1 \delta_{1} δ 1 and ε / ( 2 ∥ p ∥ + 2 ) \varepsilon/(2\lVert p\rVert+2) ε / ( 2 ∥ p ∥ + 2 ) , a positive real number. If y ∈ S y\in S y ∈ S satisfies d E ( x , y ) = ∥ y − x ∥ < δ d_{E}(x,y)=\lVert y-x\rVert<\delta d E ( x , y ) = ∥ y − x ∥ < δ , then
∣ ψ p ( y ) − ψ p ( x ) ∣ < ε 2 + ∥ p ∥ ε 2 ∥ p ∥ + 2 ≤ ε 2 + ε 2 = ε , \bigl|\psi_{p}(y)-\psi_{p}(x)\bigr|<\frac{\varepsilon}{2}+\lVert p\rVert\,\frac{\varepsilon}{2\lVert p\rVert+2}\le\frac{\varepsilon}{2}+\frac{\varepsilon}{2}=\varepsilon, ψ p ( y ) − ψ p ( x ) < 2 ε + ∥ p ∥ 2 ∥ p ∥ + 2 ε ≤ 2 ε + 2 ε = ε ,
since 2 ∥ p ∥ ≤ 2 ∥ p ∥ + 2 2\lVert p\rVert\le 2\lVert p\rVert+2 2 ∥ p ∥ ≤ 2 ∥ p ∥ + 2 gives ∥ p ∥ / ( 2 ∥ p ∥ + 2 ) ≤ 1 2 \lVert p\rVert/(2\lVert p\rVert+2)\le\tfrac12 ∥ p ∥ / ( 2 ∥ p ∥ + 2 ) ≤ 2 1 .
Proof of claim 1. Let p ∈ R n p\in\mathbb{R}^{n} p ∈ R n . By step 1 with S = B ˉ S=\bar{B} S = B ˉ , the function ψ p \psi_{p} ψ p satisfies the hypotheses of Extreme Value Theorem on a Compact Subset of a Metric Space on the nonempty compact set B ˉ \bar{B} B ˉ , so there is x max ∈ B ˉ x_{\max}\in\bar{B} x m a x ∈ B ˉ with ψ p ( y ) ≤ ψ p ( x max ) \psi_{p}(y)\le\psi_{p}(x_{\max}) ψ p ( y ) ≤ ψ p ( x m a x ) for every y ∈ B ˉ y\in\bar{B} y ∈ B ˉ ; that is, x max ∈ M ( p ) x_{\max}\in M(p) x m a x ∈ M ( p ) and M ( p ) ≠ ∅ M(p)\ne\varnothing M ( p ) = ∅ . Now let δ \delta δ be a positive real number. The origin 0 R n 0_{\mathbb{R}^{n}} 0 R n satisfies ∥ 0 R n ∥ = 0 ≤ δ \lVert 0_{\mathbb{R}^{n}}\rVert=0\le\delta ∥ 0 R n ∥ = 0 ≤ δ by claim 3 of Properties of the Euclidean Norm on R n \mathbb{R}^n R n : Square, Distance, Vanishing, Coordinate Bounds, Homogeneity, Triangle Inequalities and Expansion of a Square , so any element of M ( 0 R n ) M(0_{\mathbb{R}^{n}}) M ( 0 R n ) belongs to K δ K_{\delta} K δ , and K δ ≠ ∅ K_{\delta}\ne\varnothing K δ = ∅ .
Proof of claim 2. Let ρ ∈ R \rho\in\mathbb{R} ρ ∈ R with 0 < ρ 0<\rho 0 < ρ .
Suppose first that r < ρ r<\rho r < ρ . Every x ∈ B ˉ x\in\bar{B} x ∈ B ˉ satisfies d E ( x ^ , x ) ≤ r < ρ d_{E}(\hat{x},x)\le r<\rho d E ( x ^ , x ) ≤ r < ρ , so B ˉ ⊆ B d E ( x ^ , ρ ) \bar{B}\subseteq B_{d_{E}}(\hat{x},\rho) B ˉ ⊆ B d E ( x ^ , ρ ) ; since K δ ⊆ B ˉ K_{\delta}\subseteq\bar{B} K δ ⊆ B ˉ for every positive real δ \delta δ , the choice δ ρ = 1 \delta_{\rho}=1 δ ρ = 1 works.
Suppose now that ρ ≤ r \rho\le r ρ ≤ r , and put
A = { x ∈ B ˉ : ρ ≤ ∥ x − x ^ ∥ } . A=\bigl\{x\in\bar{B}:\rho\le\lVert x-\hat{x}\rVert\bigr\}. A = { x ∈ B ˉ : ρ ≤ ∥ x − x ^ ∥ } .
The set A A A is nonempty: let e 1 ∈ R n e_{1}\in\mathbb{R}^{n} e 1 ∈ R n have first component 1 1 1 and all other components 0 0 0 , so that ∥ e 1 ∥ 2 = 1 \lVert e_{1}\rVert^{2}=1 ∥ e 1 ∥ 2 = 1 and hence ∥ e 1 ∥ = 1 \lVert e_{1}\rVert=1 ∥ e 1 ∥ = 1 by claim 1 of Properties of the Euclidean Norm on R n \mathbb{R}^n R n : Square, Distance, Vanishing, Coordinate Bounds, Homogeneity, Triangle Inequalities and Expansion of a Square ; then z = x ^ + ρ e 1 z=\hat{x}+\rho\,e_{1} z = x ^ + ρ e 1 satisfies ∥ z − x ^ ∥ = ∥ ρ e 1 ∥ = ρ \lVert z-\hat{x}\rVert=\lVert\rho\,e_{1}\rVert=\rho ∥ z − x ^ ∥ = ∥ ρ e 1 ∥ = ρ by claim 5 there, so d E ( x ^ , z ) = ρ ≤ r d_{E}(\hat{x},z)=\rho\le r d E ( x ^ , z ) = ρ ≤ r and z ∈ A z\in A z ∈ A .
The set A A A is compact. Indeed A = B ˉ ∩ C A=\bar{B}\cap C A = B ˉ ∩ C where C = R n ∖ B d E ( x ^ , ρ ) C=\mathbb{R}^{n}\setminus B_{d_{E}}(\hat{x},\rho) C = R n ∖ B d E ( x ^ , ρ ) , because d E ( x ^ , x ) = ∥ x − x ^ ∥ d_{E}(\hat{x},x)=\lVert x-\hat{x}\rVert d E ( x ^ , x ) = ∥ x − x ^ ∥ and B d E ( x ^ , ρ ) B_{d_{E}}(\hat{x},\rho) B d E ( x ^ , ρ ) consists of the points at distance less than ρ \rho ρ from x ^ \hat{x} x ^ . The open ball is open by Open Ball in a Metric Space is Open , so C C C is closed ; B ˉ \bar{B} B ˉ is closed; and the intersection of two closed sets is closed by Complements, Unions and Intersections of Closed Sets in a Topological Space . Moreover A ⊆ B ˉ A\subseteq\bar{B} A ⊆ B ˉ is bounded, directly from Bounded Subset of a Metric Space . By The Heine-Borel Theorem in R n \mathbb{R}^n R n , and Compactness of Closed Balls and the Unit Sphere , A A A is compact.
By step 1 with p = 0 R n p=0_{\mathbb{R}^{n}} p = 0 R n and S = A S=A S = A , and by Extreme Value Theorem on a Compact Subset of a Metric Space , there is x A ∈ A x_{A}\in A x A ∈ A with φ ( x ) ≤ φ ( x A ) \varphi(x)\le\varphi(x_{A}) φ ( x ) ≤ φ ( x A ) for every x ∈ A x\in A x ∈ A . Since 0 < ρ ≤ ∥ x A − x ^ ∥ 0<\rho\le\lVert x_{A}-\hat{x}\rVert 0 < ρ ≤ ∥ x A − x ^ ∥ we have x A ≠ x ^ x_{A}\ne\hat{x} x A = x ^ , so (H2) gives φ ( x A ) < φ ( x ^ ) \varphi(x_{A})<\varphi(\hat{x}) φ ( x A ) < φ ( x ^ ) . Put
η = φ ( x ^ ) − φ ( x A ) > 0 , δ ρ = η 2 r > 0. \eta=\varphi(\hat{x})-\varphi(x_{A})>0,\qquad \delta_{\rho}=\frac{\eta}{2r}>0 . η = φ ( x ^ ) − φ ( x A ) > 0 , δ ρ = 2 r η > 0.
Let δ \delta δ be real with 0 < δ ≤ δ ρ 0<\delta\le\delta_{\rho} 0 < δ ≤ δ ρ and let x ∈ K δ x\in K_{\delta} x ∈ K δ , say x ∈ M ( p ) x\in M(p) x ∈ M ( p ) with ∥ p ∥ ≤ δ \lVert p\rVert\le\delta ∥ p ∥ ≤ δ . Since x ^ ∈ B ˉ \hat{x}\in\bar{B} x ^ ∈ B ˉ we have φ ( x ^ ) + p ⋅ x ^ ≤ φ ( x ) + p ⋅ x \varphi(\hat{x})+p\cdot\hat{x}\le\varphi(x)+p\cdot x φ ( x ^ ) + p ⋅ x ^ ≤ φ ( x ) + p ⋅ x , hence, using Cauchy-Schwarz and ∥ x − x ^ ∥ = d E ( x ^ , x ) ≤ r \lVert x-\hat{x}\rVert=d_{E}(\hat{x},x)\le r ∥ x − x ^ ∥ = d E ( x ^ , x ) ≤ r ,
φ ( x ) ≥ φ ( x ^ ) − p ⋅ ( x − x ^ ) ≥ φ ( x ^ ) − ∥ p ∥ ∥ x − x ^ ∥ ≥ φ ( x ^ ) − δ r ≥ φ ( x ^ ) − η 2 > φ ( x ^ ) − η = φ ( x A ) . \varphi(x)\ \ge\ \varphi(\hat{x})-p\cdot(x-\hat{x})\ \ge\ \varphi(\hat{x})-\lVert p\rVert\,\lVert x-\hat{x}\rVert\ \ge\ \varphi(\hat{x})-\delta\,r\ \ge\ \varphi(\hat{x})-\frac{\eta}{2}\ >\ \varphi(\hat{x})-\eta=\varphi(x_{A}). φ ( x ) ≥ φ ( x ^ ) − p ⋅ ( x − x ^ ) ≥ φ ( x ^ ) − ∥ p ∥ ∥ x − x ^ ∥ ≥ φ ( x ^ ) − δ r ≥ φ ( x ^ ) − 2 η > φ ( x ^ ) − η = φ ( x A ) .
So φ ( x A ) < φ ( x ) \varphi(x_{A})<\varphi(x) φ ( x A ) < φ ( x ) , which rules out x ∈ A x\in A x ∈ A ; as x ∈ B ˉ x\in\bar{B} x ∈ B ˉ , this forces ∥ x − x ^ ∥ < ρ \lVert x-\hat{x}\rVert<\rho ∥ x − x ^ ∥ < ρ , that is, x ∈ B d E ( x ^ , ρ ) x\in B_{d_{E}}(\hat{x},\rho) x ∈ B d E ( x ^ , ρ ) . Hence K δ ⊆ B d E ( x ^ , ρ ) K_{\delta}\subseteq B_{d_{E}}(\hat{x},\rho) K δ ⊆ B d E ( x ^ , ρ ) , proving claim 2.
Proof of claim 3. Let δ \delta δ be a positive real number. Since K δ ⊆ B ˉ K_{\delta}\subseteq\bar{B} K δ ⊆ B ˉ , the set K δ K_{\delta} K δ is bounded. We show that it is closed, using Sequential Characterization of Closed Subsets of a Metric Space .
Let ( x ( q ) ) q ∈ N (x^{(q)})_{q\in\mathbb{N}} ( x ( q ) ) q ∈ N be a sequence with x ( q ) ∈ K δ x^{(q)}\in K_{\delta} x ( q ) ∈ K δ for every q q q , converging to a point x ∈ R n x\in\mathbb{R}^{n} x ∈ R n , and for each q q q choose p ( q ) ∈ R n p^{(q)}\in\mathbb{R}^{n} p ( q ) ∈ R n with ∥ p ( q ) ∥ ≤ δ \lVert p^{(q)}\rVert\le\delta ∥ p ( q ) ∥ ≤ δ and x ( q ) ∈ M ( p ( q ) ) x^{(q)}\in M(p^{(q)}) x ( q ) ∈ M ( p ( q ) ) . As B ˉ \bar{B} B ˉ is closed, x ∈ B ˉ x\in\bar{B} x ∈ B ˉ .
The closed ball B ˉ d E ( 0 R n , δ ) \bar{B}_{d_{E}}(0_{\mathbb{R}^{n}},\delta) B ˉ d E ( 0 R n , δ ) is bounded and closed by claims 2 and 3 of Elementary Properties of the Closed Ball in a Metric Space , and ∥ p ( q ) ∥ = d E ( 0 R n , p ( q ) ) ≤ δ \lVert p^{(q)}\rVert=d_{E}(0_{\mathbb{R}^{n}},p^{(q)})\le\delta ∥ p ( q ) ∥ = d E ( 0 R n , p ( q ) ) ≤ δ places every p ( q ) p^{(q)} p ( q ) in it. By The Bolzano-Weierstrass Theorem in R n \mathbb{R}^n R n there are p ∈ R n p\in\mathbb{R}^{n} p ∈ R n and a strictly increasing sequence ( q t ) t ∈ N (q_{t})_{t\in\mathbb{N}} ( q t ) t ∈ N in N \mathbb{N} N such that ( p ( q t ) ) t ∈ N (p^{(q_{t})})_{t\in\mathbb{N}} ( p ( q t ) ) t ∈ N converges to p p p ; by Sequential Characterization of Closed Subsets of a Metric Space , p ∈ B ˉ d E ( 0 R n , δ ) p\in\bar{B}_{d_{E}}(0_{\mathbb{R}^{n}},\delta) p ∈ B ˉ d E ( 0 R n , δ ) , that is, ∥ p ∥ ≤ δ \lVert p\rVert\le\delta ∥ p ∥ ≤ δ . The subsequence ( x ( q t ) ) t ∈ N (x^{(q_{t})})_{t\in\mathbb{N}} ( x ( q t ) ) t ∈ N converges to x x x .
Fix y ∈ B ˉ y\in\bar{B} y ∈ B ˉ and let ε ∈ R \varepsilon\in\mathbb{R} ε ∈ R with 0 < ε 0<\varepsilon 0 < ε . By (H1) there is a positive real δ 2 \delta_{2} δ 2 such that every z ∈ B ˉ z\in\bar{B} z ∈ B ˉ with ∥ z − x ∥ < δ 2 \lVert z-x\rVert<\delta_{2} ∥ z − x ∥ < δ 2 satisfies ∣ φ ( z ) − φ ( x ) ∣ < ε / 4 |\varphi(z)-\varphi(x)|<\varepsilon/4 ∣ φ ( z ) − φ ( x ) ∣ < ε /4 . By convergence, choose t ∈ N t\in\mathbb{N} t ∈ N so large that
∥ x ( q t ) − x ∥ < δ 2 , ∥ x ( q t ) − x ∥ < ε 4 δ + 4 , ∥ p ( q t ) − p ∥ < ε 4 ∥ x ∥ + 4 ∥ y ∥ + 4 . \lVert x^{(q_{t})}-x\rVert<\delta_{2},\qquad \lVert x^{(q_{t})}-x\rVert<\frac{\varepsilon}{4\delta+4},\qquad \lVert p^{(q_{t})}-p\rVert<\frac{\varepsilon}{4\lVert x\rVert+4\lVert y\rVert+4}. ∥ x ( q t ) − x ∥ < δ 2 , ∥ x ( q t ) − x ∥ < 4 δ + 4 ε , ∥ p ( q t ) − p ∥ < 4 ∥ x ∥ + 4 ∥ y ∥ + 4 ε .
Writing p ( q t ) ⋅ x ( q t ) − p ⋅ x = p ( q t ) ⋅ ( x ( q t ) − x ) + ( p ( q t ) − p ) ⋅ x p^{(q_{t})}\cdot x^{(q_{t})}-p\cdot x=p^{(q_{t})}\cdot(x^{(q_{t})}-x)+(p^{(q_{t})}-p)\cdot x p ( q t ) ⋅ x ( q t ) − p ⋅ x = p ( q t ) ⋅ ( x ( q t ) − x ) + ( p ( q t ) − p ) ⋅ x and using Cauchy-Schwarz,
∣ ψ p ( q t ) ( x ( q t ) ) − ψ p ( x ) ∣ ≤ ∣ φ ( x ( q t ) ) − φ ( x ) ∣ + δ ∥ x ( q t ) − x ∥ + ∥ p ( q t ) − p ∥ ∥ x ∥ < 3 ε 4 , \bigl|\psi_{p^{(q_{t})}}(x^{(q_{t})})-\psi_{p}(x)\bigr|\le\bigl|\varphi(x^{(q_{t})})-\varphi(x)\bigr|+\delta\,\lVert x^{(q_{t})}-x\rVert+\lVert p^{(q_{t})}-p\rVert\,\lVert x\rVert<\frac{3\varepsilon}{4}, ψ p ( q t ) ( x ( q t ) ) − ψ p ( x ) ≤ φ ( x ( q t ) ) − φ ( x ) + δ ∥ x ( q t ) − x ∥ + ∥ p ( q t ) − p ∥ ∥ x ∥ < 4 3 ε ,
while ∣ p ( q t ) ⋅ y − p ⋅ y ∣ ≤ ∥ p ( q t ) − p ∥ ∥ y ∥ < ε / 4 \bigl|p^{(q_{t})}\cdot y-p\cdot y\bigr|\le\lVert p^{(q_{t})}-p\rVert\,\lVert y\rVert<\varepsilon/4 p ( q t ) ⋅ y − p ⋅ y ≤ ∥ p ( q t ) − p ∥ ∥ y ∥ < ε /4 . Since x ( q t ) ∈ M ( p ( q t ) ) x^{(q_{t})}\in M(p^{(q_{t})}) x ( q t ) ∈ M ( p ( q t ) ) and y ∈ B ˉ y\in\bar{B} y ∈ B ˉ ,
φ ( y ) + p ( q t ) ⋅ y ≤ φ ( x ( q t ) ) + p ( q t ) ⋅ x ( q t ) , \varphi(y)+p^{(q_{t})}\cdot y\ \le\ \varphi(x^{(q_{t})})+p^{(q_{t})}\cdot x^{(q_{t})} , φ ( y ) + p ( q t ) ⋅ y ≤ φ ( x ( q t ) ) + p ( q t ) ⋅ x ( q t ) ,
and combining the three displays,
φ ( y ) + p ⋅ y < φ ( y ) + p ( q t ) ⋅ y + ε 4 ≤ φ ( x ( q t ) ) + p ( q t ) ⋅ x ( q t ) + ε 4 < φ ( x ) + p ⋅ x + ε . \varphi(y)+p\cdot y\ <\ \varphi(y)+p^{(q_{t})}\cdot y+\frac{\varepsilon}{4}\ \le\ \varphi(x^{(q_{t})})+p^{(q_{t})}\cdot x^{(q_{t})}+\frac{\varepsilon}{4}\ <\ \varphi(x)+p\cdot x+\varepsilon . φ ( y ) + p ⋅ y < φ ( y ) + p ( q t ) ⋅ y + 4 ε ≤ φ ( x ( q t ) ) + p ( q t ) ⋅ x ( q t ) + 4 ε < φ ( x ) + p ⋅ x + ε .
As ε \varepsilon ε was an arbitrary positive real number, φ ( y ) + p ⋅ y ≤ φ ( x ) + p ⋅ x \varphi(y)+p\cdot y\le\varphi(x)+p\cdot x φ ( y ) + p ⋅ y ≤ φ ( x ) + p ⋅ x : were the reverse strict inequality to hold, taking ε \varepsilon ε to be half of the positive difference would contradict the display. Since y ∈ B ˉ y\in\bar{B} y ∈ B ˉ was arbitrary, x ∈ M ( p ) x\in M(p) x ∈ M ( p ) with ∥ p ∥ ≤ δ \lVert p\rVert\le\delta ∥ p ∥ ≤ δ , so x ∈ K δ x\in K_{\delta} x ∈ K δ .
By Sequential Characterization of Closed Subsets of a Metric Space the set K δ K_{\delta} K δ is closed, and being bounded it is compact by The Heine-Borel Theorem in R n \mathbb{R}^n R n , and Compactness of Closed Balls and the Unit Sphere . This proves claim 3.