Since x 0 x_{0} x 0 β is an interior point of C C C , claim 1 of Interior Points in the Metric Topology are Exactly the Centres of Contained Closed Balls provides Ξ΅ β R \varepsilon\in\mathbb{R} Ξ΅ β R with 0 < Ξ΅ 0<\varepsilon 0 < Ξ΅ and B Λ d E ( x 0 , Ξ΅ ) β C \bar{B}_{d_{E}}(x_{0},\varepsilon)\subseteq C B Λ d E β β ( x 0 β , Ξ΅ ) β C ; this Ξ΅ \varepsilon Ξ΅ is fixed for the remainder of the proof.
For i β [ n ] i\in[n] i β [ n ] let e ( i ) β R n e^{(i)}\in\mathbb{R}^{n} e ( i ) β R n be the point whose i i i th coordinate is 1 1 1 and whose other coordinates are 0 0 0 . All sums with a numerical index range are the finite sums of R \mathbb{R} R , and 2 = 1 + 1 2=1+1 2 = 1 + 1 .
Step 1: a coordinate-sum bound. Put c = β k = 1 n 1 c=\sum_{k=1}^{n}1 c = β k = 1 n β 1 . Each summand is nonnegative by claim 1 of Elementary Arithmetic in an Ordered Field , so claim 6 of Properties of Finite Sums gives 1 β€ c 1\le c 1 β€ c ; with 0 < 1 0<1 0 < 1 from claim 6 of Elementary Order Arithmetic in an Ordered Field and claim 2 of that lemma we get 0 < c 0<c 0 < c , so c c c has an inverse c β 1 c^{-1} c β 1 with 0 < c β 1 0<c^{-1} 0 < c β 1 by claim 7 of that lemma. For every y β R n y\in\mathbb{R}^{n} y β R n , claim 4 of Elementary Properties of the Euclidean Norm on R n \mathbb{R}^n R n gives β£ y i β£ β€ β₯ y β₯ |y_{i}|\le\lVert y\rVert β£ y i β β£ β€ β₯ y β₯ for every i β [ n ] i\in[n] i β [ n ] , so comparing the two sums termwise by Comparison and Absolute Value Bounds for Finite Sums of Real Numbers and using claim 3 of Properties of Finite Sums ,
β i = 1 n β£ y i β£ β€ β i = 1 n β₯ y β₯ = c β β₯ y β₯ . (1) \sum_{i=1}^{n}|y_{i}|\le\sum_{i=1}^{n}\lVert y\rVert=c\,\lVert y\rVert. \tag{1} i = 1 β n β β£ y i β β£ β€ i = 1 β n β β₯ y β₯ = c β₯ y β₯ . ( 1 )
Step 2: the coordinate neighbours lie in C C C . Fix i β [ n ] i\in[n] i β [ n ] . By claim 1 of Elementary Properties of the Euclidean Norm on R n \mathbb{R}^n R n and claim 7 of Properties of Finite Sums , applied to the family j β¦ e j ( i ) e j ( i ) j\mapsto e^{(i)}_{j}e^{(i)}_{j} j β¦ e j ( i ) β e j ( i ) β , which vanishes off i i i ,
β₯ e ( i ) β₯ β β₯ e ( i ) β₯ = e ( i ) β
e ( i ) = β j = 1 n e j ( i ) e j ( i ) = 1 , \lVert e^{(i)}\rVert\,\lVert e^{(i)}\rVert=e^{(i)}\cdot e^{(i)}=\sum_{j=1}^{n}e^{(i)}_{j}e^{(i)}_{j}=1, β₯ e ( i ) β₯ β₯ e ( i ) β₯ = e ( i ) β
e ( i ) = j = 1 β n β e j ( i ) β e j ( i ) β = 1 ,
the dot product being that of Difference, Dot Product, and Orthogonality in R n \mathbb{R}^n R n . Writing t = β₯ e ( i ) β₯ t=\lVert e^{(i)}\rVert t = β₯ e ( i ) β₯ , the elementary field identities of Zero Products and Elementary Identities in a Field turn t β t = 1 t\,t=1 t t = 1 into ( t β 1 ) ( t + 1 ) = 0 (t-1)(t+1)=0 ( t β 1 ) ( t + 1 ) = 0 , so t = 1 t=1 t = 1 or t = β 1 t=-1 t = β 1 by that lemma; and claim 4 of Elementary Properties of the Euclidean Norm on R n \mathbb{R}^n R n gives 1 = β£ e i ( i ) β£ β€ t 1=|e^{(i)}_{i}|\le t 1 = β£ e i ( i ) β β£ β€ t , which excludes t = β 1 t=-1 t = β 1 , since 0 < 1 0<1 0 < 1 by claim 6 of Elementary Order Arithmetic in an Ordered Field , hence β 1 < 0 -1<0 β 1 < 0 by the sign reversal of claim 4 of that lemma, hence β 1 < 1 -1<1 β 1 < 1 by claim 1. Hence β₯ e ( i ) β₯ = 1 \lVert e^{(i)}\rVert=1 β₯ e ( i ) β₯ = 1 .
By claim 5 of Elementary Properties of the Euclidean Norm on R n \mathbb{R}^n R n and the value β£ Ξ΅ β£ = β£ β Ξ΅ β£ = Ξ΅ |\varepsilon|=|-\varepsilon|=\varepsilon β£ Ξ΅ β£ = β£ β Ξ΅ β£ = Ξ΅ from Absolute Value in an Ordered Field , both β₯ Ξ΅ e ( i ) β₯ \lVert\varepsilon e^{(i)}\rVert β₯ Ξ΅ e ( i ) β₯ and β₯ β Ξ΅ e ( i ) β₯ \lVert-\varepsilon e^{(i)}\rVert β₯ β Ξ΅ e ( i ) β₯ equal Ξ΅ \varepsilon Ξ΅ . Since d E ( x 0 , x 0 + v ) = β₯ v β₯ d_{E}(x_{0},x_{0}+v)=\lVert v\rVert d E β ( x 0 β , x 0 β + v ) = β₯ v β₯ for every v β R n v\in\mathbb{R}^{n} v β R n , the points x 0 + Ξ΅ e ( i ) x_{0}+\varepsilon e^{(i)} x 0 β + Ξ΅ e ( i ) and x 0 β Ξ΅ e ( i ) x_{0}-\varepsilon e^{(i)} x 0 β β Ξ΅ e ( i ) lie in B Λ d E ( x 0 , Ξ΅ ) \bar{B}_{d_{E}}(x_{0},\varepsilon) B Λ d E β β ( x 0 β , Ξ΅ ) , hence in C C C ; and d E ( x 0 , x 0 ) = 0 β€ Ξ΅ d_{E}(x_{0},x_{0})=0\le\varepsilon d E β ( x 0 β , x 0 β ) = 0 β€ Ξ΅ puts x 0 x_{0} x 0 β in C C C as well.
Step 3: an upper bound at those points. Let g : [ n ] β R g:[n]\to\mathbb{R} g : [ n ] β R be the family
g i = β£ u ( x 0 + Ξ΅ e ( i ) ) β£ + β£ u ( x 0 β Ξ΅ e ( i ) ) β£ , g_{i}=\bigl|u(x_{0}+\varepsilon e^{(i)})\bigr|+\bigl|u(x_{0}-\varepsilon e^{(i)})\bigr| , g i β = β u ( x 0 β + Ξ΅ e ( i ) ) β + β u ( x 0 β β Ξ΅ e ( i ) ) β ,
and put M = β£ u ( x 0 ) β£ + β i = 1 n g i M=|u(x_{0})|+\sum_{i=1}^{n}g_{i} M = β£ u ( x 0 β ) β£ + β i = 1 n β g i β . Absolute values are nonnegative by claim 1 of Properties of the Absolute Value in an Ordered Field , so each g i g_{i} g i β is nonnegative by claim 2 of Elementary Arithmetic in an Ordered Field ; hence claim 5 of Properties of Finite Sums gives 0 β€ β i = 1 n g i 0\le\sum_{i=1}^{n}g_{i} 0 β€ β i = 1 n β g i β and claim 6 of that lemma gives g i β€ β j = 1 n g j g_{i}\le\sum_{j=1}^{n}g_{j} g i β β€ β j = 1 n β g j β for every i β [ n ] i\in[n] i β [ n ] . Using w β€ β£ w β£ w\le|w| w β€ β£ w β£ from claim 3 of Properties of the Absolute Value in an Ordered Field and the additions of claim 3 of Elementary Order Arithmetic in an Ordered Field ,
u ( x 0 ) β€ β£ u ( x 0 ) β£ β€ M , u ( x 0 Β± Ξ΅ e ( i ) ) β€ β£ u ( x 0 Β± Ξ΅ e ( i ) ) β£ β€ g i β€ β j = 1 n g j β€ M , u(x_{0})\le|u(x_{0})|\le M,\qquad u\bigl(x_{0}\pm\varepsilon e^{(i)}\bigr)\le\bigl|u(x_{0}\pm\varepsilon e^{(i)})\bigr|\le g_{i}\le\sum_{j=1}^{n}g_{j}\le M , u ( x 0 β ) β€ β£ u ( x 0 β ) β£ β€ M , u ( x 0 β Β± Ξ΅ e ( i ) ) β€ β u ( x 0 β Β± Ξ΅ e ( i ) ) β β€ g i β β€ j = 1 β n β g j β β€ M ,
where in each chain the omitted terms are nonnegative and claim 1 of Elementary Order Arithmetic in an Ordered Field supplies transitivity.
Step 4: bounds on the coordinate-sum ball. Let
P = { x β R n : β i = 1 n β£ x i β ( x 0 ) i β£ β€ Ξ΅ } . P=\Bigl\{x\in\mathbb{R}^{n}:\sum_{i=1}^{n}\bigl|x_{i}-(x_{0})_{i}\bigr|\le\varepsilon\Bigr\}. P = { x β R n : i = 1 β n β β x i β β ( x 0 β ) i β β β€ Ξ΅ } .
By Steps 2 and 3 the hypotheses of A Convex Function is Bounded Above near a Point by its Values at Coordinate Neighbours hold with r = Ξ΅ r=\varepsilon r = Ξ΅ and with this M M M , so P β C P\subseteq C P β C and u ( w ) β€ M u(w)\le M u ( w ) β€ M for every w β P w\in P w β P .
The point x 0 x_{0} x 0 β lies in P P P , since every term of its defining sum is β£ 0 β£ = 0 |0|=0 β£0β£ = 0 and claim 7 of Properties of Finite Sums makes the sum 0 0 0 . If x β P x\in P x β P and x β² = x 0 + ( x 0 β x ) x'=x_{0}+(x_{0}-x) x β² = x 0 β + ( x 0 β β x ) , then x i β² β ( x 0 ) i = β ( x i β ( x 0 ) i ) x'_{i}-(x_{0})_{i}=-\bigl(x_{i}-(x_{0})_{i}\bigr) x i β² β β ( x 0 β ) i β = β ( x i β β ( x 0 β ) i β ) , so β£ x i β² β ( x 0 ) i β£ = β£ x i β ( x 0 ) i β£ |x'_{i}-(x_{0})_{i}|=|x_{i}-(x_{0})_{i}| β£ x i β² β β ( x 0 β ) i β β£ = β£ x i β β ( x 0 β ) i β β£ by claim 4 of Properties of the Absolute Value in an Ordered Field together with β£ β 1 β£ = 1 |-1|=1 β£ β 1β£ = 1 ; hence x β² β P x'\in P x β² β P . So A Convex Function Bounded Above on a Set Symmetric about a Point is Bounded Below on It applies with D = P D=P D = P and gives
m β€ u ( w ) forΒ everyΒ w β P , whereΒ m = 2 β u ( x 0 ) β M . m\le u(w)\qquad\text{for every }w\in P,\qquad\text{where }m=2\,u(x_{0})-M . m β€ u ( w ) forΒ everyΒ w β P , whereΒ m = 2 u ( x 0 β ) β M .
Since u ( x 0 ) β€ M u(x_{0})\le M u ( x 0 β ) β€ M , multiplying by the nonnegative number 2 2 2 by claim 5 of Elementary Arithmetic in an Ordered Field and adding β M -M β M by claim 3 of Elementary Order Arithmetic in an Ordered Field give m β€ M m\le M m β€ M , so 0 β€ M β m 0\le M-m 0 β€ M β m .
Step 5: the radius. Put h = Ξ΅ β 2 β 1 h=\varepsilon\,2^{-1} h = Ξ΅ 2 β 1 and Ο = h β c β 1 \rho=h\,c^{-1} Ο = h c β 1 ; here 0 < 2 β 1 0<2^{-1} 0 < 2 β 1 and 0 < h 0<h 0 < h and 0 < Ο 0<\rho 0 < Ο by claims 7 and 5 of Elementary Order Arithmetic in an Ordered Field , and h + h = Ξ΅ ( 2 β 1 + 2 β 1 ) = Ξ΅ h+h=\varepsilon(2^{-1}+2^{-1})=\varepsilon h + h = Ξ΅ ( 2 β 1 + 2 β 1 ) = Ξ΅ . Also h β€ Ξ΅ h\le\varepsilon h β€ Ξ΅ , since Ξ΅ β h = h \varepsilon-h=h Ξ΅ β h = h is nonnegative. If x β B Λ d E ( x 0 , Ο ) x\in\bar{B}_{d_{E}}(x_{0},\rho) x β B Λ d E β β ( x 0 β , Ο ) , then β₯ x β x 0 β₯ β€ Ο \lVert x-x_{0}\rVert\le\rho β₯ x β x 0 β β₯ β€ Ο , so by ( 1 ) (1) ( 1 ) and claim 5 of Elementary Arithmetic in an Ordered Field ,
β i = 1 n β£ x i β ( x 0 ) i β£ β€ c β β₯ x β x 0 β₯ β€ c β Ο = h . (2) \sum_{i=1}^{n}\bigl|x_{i}-(x_{0})_{i}\bigr|\le c\,\lVert x-x_{0}\rVert\le c\,\rho=h .\tag{2} i = 1 β n β β x i β β ( x 0 β ) i β β β€ c β₯ x β x 0 β β₯ β€ c Ο = h . ( 2 )
In particular B Λ d E ( x 0 , Ο ) β P β C \bar{B}_{d_{E}}(x_{0},\rho)\subseteq P\subseteq C B Λ d E β β ( x 0 β , Ο ) β P β C .
Step 6: the Lipschitz estimate. Put L = c β h β 1 ( M β m ) L=c\,h^{-1}(M-m) L = c h β 1 ( M β m ) , a product of nonnegative numbers and hence nonnegative by claim 5 of Elementary Arithmetic in an Ordered Field . Let x , y β B Λ d E ( x 0 , Ο ) x,y\in\bar{B}_{d_{E}}(x_{0},\rho) x , y β B Λ d E β β ( x 0 β , Ο ) .
If x = y x=y x = y , then u ( y ) β u ( x ) = 0 u(y)-u(x)=0 u ( y ) β u ( x ) = 0 and β₯ y β x β₯ = 0 \lVert y-x\rVert=0 β₯ y β x β₯ = 0 by claim 3 of Elementary Properties of the Euclidean Norm on R n \mathbb{R}^n R n , so both sides of the asserted inequality are 0 0 0 .
Suppose x β y x\ne y x ξ = y and put Ξ΄ = β i = 1 n β£ y i β x i β£ \delta=\sum_{i=1}^{n}|y_{i}-x_{i}| Ξ΄ = β i = 1 n β β£ y i β β x i β β£ . The summands are nonnegative, so 0 β€ Ξ΄ 0\le\delta 0 β€ Ξ΄ by claim 5 of Properties of Finite Sums ; and if Ξ΄ \delta Ξ΄ were 0 0 0 the second part of that claim would force β£ y i β x i β£ = 0 |y_{i}-x_{i}|=0 β£ y i β β x i β β£ = 0 , hence y i = x i y_{i}=x_{i} y i β = x i β , for every i β [ n ] i\in[n] i β [ n ] , contradicting x β y x\ne y x ξ = y . So 0 < Ξ΄ 0<\delta 0 < Ξ΄ and Ξ΄ β 1 \delta^{-1} Ξ΄ β 1 exists with 0 < Ξ΄ β 1 0<\delta^{-1} 0 < Ξ΄ β 1 .
Put ΞΌ = h β Ξ΄ β 1 \mu=h\,\delta^{-1} ΞΌ = h Ξ΄ β 1 and z = y + ΞΌ β ( y β x ) z=y+\mu\,(y-x) z = y + ΞΌ ( y β x ) . By claim 4 of Properties of the Absolute Value in an Ordered Field , by β£ ΞΌ β£ = ΞΌ |\mu|=\mu β£ ΞΌ β£ = ΞΌ , and by claim 3 of Properties of Finite Sums ,
β i = 1 n β£ z i β y i β£ = β i = 1 n ΞΌ β β£ y i β x i β£ = ΞΌ β Ξ΄ = h . \sum_{i=1}^{n}|z_{i}-y_{i}|=\sum_{i=1}^{n}\mu\,|y_{i}-x_{i}|=\mu\,\delta=h . i = 1 β n β β£ z i β β y i β β£ = i = 1 β n β ΞΌ β£ y i β β x i β β£ = ΞΌ Ξ΄ = h .
The triangle inequality of claim 5 of Properties of the Absolute Value in an Ordered Field gives β£ z i β ( x 0 ) i β£ β€ β£ z i β y i β£ + β£ y i β ( x 0 ) i β£ |z_{i}-(x_{0})_{i}|\le|z_{i}-y_{i}|+|y_{i}-(x_{0})_{i}| β£ z i β β ( x 0 β ) i β β£ β€ β£ z i β β y i β β£ + β£ y i β β ( x 0 β ) i β β£ for every i β [ n ] i\in[n] i β [ n ] , so Comparison and Absolute Value Bounds for Finite Sums of Real Numbers , claim 2 of Properties of Finite Sums and the bound ( 2 ) (2) ( 2 ) applied to y y y give
β i = 1 n β£ z i β ( x 0 ) i β£ β€ h + β i = 1 n β£ y i β ( x 0 ) i β£ β€ h + h = Ξ΅ , \sum_{i=1}^{n}\bigl|z_{i}-(x_{0})_{i}\bigr|\le h+\sum_{i=1}^{n}\bigl|y_{i}-(x_{0})_{i}\bigr|\le h+h=\varepsilon , i = 1 β n β β z i β β ( x 0 β ) i β β β€ h + i = 1 β n β β y i β β ( x 0 β ) i β β β€ h + h = Ξ΅ ,
so z β P z\in P z β P .
Put Ξ» = Ξ΄ β ( Ξ΄ + h ) β 1 \lambda=\delta\,(\delta+h)^{-1} Ξ» = Ξ΄ ( Ξ΄ + h ) β 1 , which is defined and nonnegative because 0 < Ξ΄ + h 0<\delta+h 0 < Ξ΄ + h . From Ξ΄ β€ Ξ΄ + h \delta\le\delta+h Ξ΄ β€ Ξ΄ + h and claim 5 of Elementary Arithmetic in an Ordered Field , multiplying by the nonnegative number ( Ξ΄ + h ) β 1 (\delta+h)^{-1} ( Ξ΄ + h ) β 1 gives Ξ» β€ 1 \lambda\le1 Ξ» β€ 1 ; and 1 β Ξ» = h β ( Ξ΄ + h ) β 1 1-\lambda=h\,(\delta+h)^{-1} 1 β Ξ» = h ( Ξ΄ + h ) β 1 , since Ξ΄ ( Ξ΄ + h ) β 1 + h ( Ξ΄ + h ) β 1 = ( Ξ΄ + h ) ( Ξ΄ + h ) β 1 = 1 \delta(\delta+h)^{-1}+h(\delta+h)^{-1}=(\delta+h)(\delta+h)^{-1}=1 Ξ΄ ( Ξ΄ + h ) β 1 + h ( Ξ΄ + h ) β 1 = ( Ξ΄ + h ) ( Ξ΄ + h ) β 1 = 1 . For every i β [ n ] i\in[n] i β [ n ] , using Ξ΄ ΞΌ = h \delta\mu=h Ξ΄ ΞΌ = h ,
Ξ΄ β z i + h β x i = Ξ΄ β y i + h β ( y i β x i ) + h β x i = ( Ξ΄ + h ) β y i , \delta\,z_{i}+h\,x_{i}=\delta\,y_{i}+h\,(y_{i}-x_{i})+h\,x_{i}=(\delta+h)\,y_{i}, Ξ΄ z i β + h x i β = Ξ΄ y i β + h ( y i β β x i β ) + h x i β = ( Ξ΄ + h ) y i β ,
and multiplying by ( Ξ΄ + h ) β 1 (\delta+h)^{-1} ( Ξ΄ + h ) β 1 gives Ξ» z i + ( 1 β Ξ» ) x i = y i \lambda z_{i}+(1-\lambda)x_{i}=y_{i} Ξ» z i β + ( 1 β Ξ» ) x i β = y i β . Hence y = Ξ» z + ( 1 β Ξ» ) x y=\lambda z+(1-\lambda)x y = Ξ» z + ( 1 β Ξ» ) x .
The points x x x , y y y and z z z all lie in P P P , on which m β€ u β€ M m\le u\le M m β€ u β€ M by Step 4, so Increment Bound for a Convex Function through an Extended Point applies with D = P D=P D = P and gives u ( y ) β u ( x ) β€ Ξ» ( M β m ) u(y)-u(x)\le\lambda(M-m) u ( y ) β u ( x ) β€ Ξ» ( M β m ) .
It remains to bound Ξ» \lambda Ξ» . From h β€ Ξ΄ + h h\le\delta+h h β€ Ξ΄ + h and 0 β€ Ξ» 0\le\lambda 0 β€ Ξ» , claim 5 of Elementary Arithmetic in an Ordered Field gives Ξ» h β€ Ξ» ( Ξ΄ + h ) = Ξ΄ \lambda h\le\lambda(\delta+h)=\delta Ξ»h β€ Ξ» ( Ξ΄ + h ) = Ξ΄ , and multiplying by the nonnegative number h β 1 h^{-1} h β 1 gives Ξ» β€ Ξ΄ β h β 1 \lambda\le\delta\,h^{-1} Ξ» β€ Ξ΄ h β 1 . Multiplying by the nonnegative number M β m M-m M β m , then using ( 1 ) (1) ( 1 ) with y β x y-x y β x in place of y y y and multiplying by the nonnegative number h β 1 ( M β m ) h^{-1}(M-m) h β 1 ( M β m ) , and finally using claim 1 of Elementary Order Arithmetic in an Ordered Field ,
u ( y ) β u ( x ) β€ Ξ» β ( M β m ) β€ Ξ΄ β h β 1 ( M β m ) β€ c β β₯ y β x β₯ β h β 1 ( M β m ) = L β β₯ y β x β₯ . u(y)-u(x)\le\lambda\,(M-m)\le\delta\,h^{-1}(M-m)\le c\,\lVert y-x\rVert\,h^{-1}(M-m)=L\,\lVert y-x\rVert . u ( y ) β u ( x ) β€ Ξ» ( M β m ) β€ Ξ΄ h β 1 ( M β m ) β€ c β₯ y β x β₯ h β 1 ( M β m ) = L β₯ y β x β₯ .
Interchanging the roles of x x x and y y y leaves Ξ΄ \delta Ξ΄ unchanged, because β£ x i β y i β£ = β£ y i β x i β£ |x_{i}-y_{i}|=|y_{i}-x_{i}| β£ x i β β y i β β£ = β£ y i β β x i β β£ , and leaves β₯ y β x β₯ \lVert y-x\rVert β₯ y β x β₯ unchanged, by claim 5 of Elementary Properties of the Euclidean Norm on R n \mathbb{R}^n R n with the factor β 1 -1 β 1 ; the same argument therefore gives u ( x ) β u ( y ) β€ L β₯ y β x β₯ u(x)-u(y)\le L\lVert y-x\rVert u ( x ) β u ( y ) β€ L β₯ y β x β₯ . Claim 6 of Properties of the Absolute Value in an Ordered Field now yields
β£ u ( y ) β u ( x ) β£ β€ L β β₯ y β x β₯ . \bigl|u(y)-u(x)\bigr|\le L\,\lVert y-x\rVert . β u ( y ) β u ( x ) β β€ L β₯ y β x β₯ .