Some remarks concerning norm preserving extension and best approximation

Abstract

Let \(X,Y\) be two normed spaces, \(X_{1}\) a subspace of \(X\) and \(A:X \rightarrow Y\) a continuous linear operator. Let us denote \(Z_{1}=Ker\left( \left. A\right \vert _{X_{1}}\right) ,Z=KerA\) and for \(x\in X,E\left( x\right)=\{y\in X:Ax=Ay\) and \(\left \Vert y\right \Vert =\left \Vert Ax\right \Vert/\left \Vert A\right \Vert \}\) and \(E_{1}\left( x\right) =\{y_{1}\in X_{1}:Ax=Ay_{1}\) and \(\left \Vert y_{1}\right \Vert =\left \Vert Ax\right \Vert/\left \Vert A\right \Vert \}\). One gives the relations between the sets \(E\left( x\right)\), \(E_{1}\left(x\right)\) and \(P_{Z}\left( x\right)\), \(P_{Z_{1}}\left( x\right)\) where \(P_{C}\left( x\right) :\{y\in C:\left \Vert x-y\right \Vert =d\left(x,C\right) \}\). An application is considered.

Authors

Costica Mustata
“Tiberiu Popoviciu” Institute of Numerical Analysis, Romanian Academy, Romania

Keywords

Paper coordinates

C. Mustăţa, Some remarks concerning norm preserving extension and best approximation, Rev. Anal. Numer. Theor. Approx., 29 (2000) No. 2, pp. 173-180.

PDF

About this paper

Journal

Revue d’Analyse Numer.Theor. Approx.

Publisher Name

Publishing Romanian Academy

Print ISSN

2457-6794

Online ISSN

2501-059X

google scholar link

[1] Cobzas S., Some Applications of a Distance Formula to the Kernel of a Linear Operator (submitted)
[2] Cobzas, S., Mustata, C., Norm Preserving Extension of Convex Lipschitz Functions, J. Approx. Theory, 24, pÞ.236-244, 1978.
[3] Cobzas, S., Mustata, C., Extension of Bilinear Functionals and Best Approximation in 2-normed, Spaces, Studia Univ. “Babes-Bolyai”, Series Math., XLIII, no. 2, pp. 1-13, 1998.
[4] Czipser, J., Geher, L., Extension of Functions Satisfying a Lipschitz Condition, Acta Math. Acad. Sci. Hungar, 6, pp. 213,220, 1955.
[5] Deutsch, F., Wu Li  and Sizwe Mabizela, Helly Extensions and Best Approximation, Parametric Optimization and Related Topics III (J. Guddat, H. Th. Jongen, B. Kummer and F. No5iaeka Eds., Appromation and Optimization, vol. 3, Verlag Peter Lang, pp. 107-120, Flankfurt 1993.
[6] Engelking, R., General Topology, PWN Warszawa, 1985.
[7] McShane, E. J., Extension of Range of Functions, Bull. Amer. Math., 40, pp. 834-842, L934.
[S] Mustata., C., Best Approximation and Unique Extension of Lipschitz Functions, J. Approx. Theory, 19, pp. 222-230, 1977.
[9] Mustata, C., Norm Preserving Extension of Starshaped, Lipschitz Functions, Mathematica, 19, pp. 183-787, L977.

Paper (preprint) in HTML form

2000-Mustata-Some remarks concerning norm preserving extension-Jnaat

SOME REMARKS CONCERNING NORM PRESERVING. EXTENSIONS AND BEST APPROXIMATION*

COSTICĂ MUSTĂţA

Abstract

Let X , Y X , Y X,Y\mathrm{X}, \mathrm{Y}X,Y be two normed spaces, X 1 X 1 X_(1)X_{1}X1 a subspace of X X XXX and A : X Y A : X Y A:X rarr YA: X \rightarrow YA:XY a continuous linear operator. Let us denote Z 1 = Ker ( A | X 1 ) , Z = Ker A Z 1 = Ker A X 1 , Z = Ker A Z_(1)=Ker(A|_(X_(1))),Z=Ker AZ_{1}=\operatorname{Ker}\left(\left.A\right|_{X_{1}}\right), Z=\operatorname{Ker} AZ1=Ker(A|X1),Z=KerA and for x X , E ( x ) = { y X : A x = A y x X , E ( x ) = { y X : A x = A y x in X,E(x)={y in X:Ax=Ayx \in X, E(x)= \{y \in X: A x=A yxX,E(x)={yX:Ax=Ay and y = A x / A } y = A x / A } ||y||=||Ax||//||A||}\|y\|=\|A x\| /\|A\|\}y=Ax/A} and E 1 ( x ) = { y 1 X 1 : A x = A y 1 E 1 ( x ) = y 1 X 1 : A x = A y 1 E_(1)(x)={y_(1)inX_(1):Ax=Ay_(1):}E_{1}(x)=\left\{y_{1} \in X_{1}: A x=A y_{1}\right.E1(x)={y1X1:Ax=Ay1 and y 1 = A x / A } y 1 = A x / A } ||y_(1)||=||Ax||//||A||}\left\|y_{1}\right\|= \|A x\| /\|A\|\}y1=Ax/A}.

One gives the relations between the sets E ( x ) , E 1 ( x ) E ( x ) , E 1 ( x ) E(x),E_(1)(x)E(x), E_{1}(x)E(x),E1(x) and P Z ( x ) , P Z 1 ( x ) P Z ( x ) , P Z 1 ( x ) P_(Z)(x),P_(Z_(1))(x)P_{Z}(x), P_{Z_{1}}(x)PZ(x),PZ1(x) where P C ( x ) := { y C : x y = d ( x , C ) } P C ( x ) := { y C : x y = d ( x , C ) } P_(C)(x):={y in C:||x-y||=d(x,C)}P_{C}(x):=\{y \in C:\|x-y\|=d(x, C)\}PC(x):={yC:xy=d(x,C)}. An application is considered.

Let X X XXX be a real normed space and M M MMM a nonvoid closed subset of X X XXX. For x X x X x in Xx \in XxX let
d ( x , M ) = inf { x y : y M } d ( x , M ) = inf { x y : y M } d(x,M)=i n f{||x-y||:y in M}d(x, M)=\inf \{\|x-y\|: y \in M\}d(x,M)=inf{xy:yM}
be the distance from x x xxx to M M MMM and let
P M ( x ) := { y M : x y = d ( x , M ) } P M ( x ) := { y M : x y = d ( x , M ) } P_(M)(x):={y in M:||x-y||=d(x,M)}P_{M}(x):=\{y \in M:\|x-y\|=d(x, M)\}PM(x):={yM:xy=d(x,M)}
be the set of nearest points from x x xxx in the set M M MMM.
If P M ( x ) P M ( x ) P_(M)(x)!=O/P_{M}(x) \neq \emptysetPM(x) for every x X x X x in Xx \in XxX then the set M M MMM is called proximinal, if P M ( x ) P M ( x ) P_(M)(x)P_{M}(x)PM(x) is a singleton for every x X x X x in Xx \in XxX then M M MMM is called chebyshevian and if P M ( x ) = P M ( x ) = P_(M)(x)=O/P_{M}(x)=\emptysetPM(x)= for every x X M x X M x in X\\Mx \in X \backslash MxXM then the set M M MMM is called antiproximinal.
For a subspace Y Y YYY of X X XXX let
Y = { x X : x | Y = 0 } Y = x X : x Y = 0 Y^(_|_)={x^(**)inX^(**):x^(**)|_(Y)=0}Y^{\perp}=\left\{x^{*} \in X^{*}:\left.x^{*}\right|_{Y}=0\right\}Y={xX:x|Y=0}
be the annihilator of the subspace Y Y YYY in the conjugate space X X X^(**)X^{*}X of X X XXX.
R.R. Phelps [13] studied the relation between the norm-preserving extension properties of the space Y Y Y^(**)Y^{*}Y with respect to X X X^(**)X^{*}X and the best approximation properties of Y Y Y^(_|_)Y^{\perp}Y. Namely, he proved that every y Y y Y y^(**)inY^(**)y^{*} \in Y^{*}yY has a unique normpreserving extension x X x X x^(**)inX^(**)x^{*} \in X^{*}xX if and only if Y Y Y^(_|_)Y^{\perp}Y is a chebyshevian subspace of X X X^(**)X^{*}X. By the Hahn-Banach extension theorem, every y Y y Y y^(**)inY^(**)y^{*} \in Y^{*}yY has at least
one norm-preserving extension x X x X x^(**)inX^(**)x^{*} \in X^{*}xX. Since then, there have been proved a lot of theorems emphasizing the relations between the extension and best approximation properties for special classes of functions. These results correspond to various extension theorems, such as Tietze extension theorem for continuous function [6], Mc Shane's extension theorem for Lipschitz functions [7], extension theorems for bilinear functionals on 2 -normed spaces [3].
S. Cobzas [1] proved that all the above mentioned results can be derived from a formula for the distance to the kernel of a continuous linear operator.
For normed spaces X , Y X , Y X,YX, YX,Y and A : X Y A : X Y A:X rarr YA: X \rightarrow YA:XY a continuous linear operator, let
(1) Z = Ker A = { x X : A x = 0 } (1) Z = Ker A = { x X : A x = 0 } {:(1)Z=Ker A={x in X:Ax=0}:}\begin{equation*} Z=\operatorname{Ker} A=\{x \in X: A x=0\} \tag{1} \end{equation*}(1)Z=KerA={xX:Ax=0}
be the kernel of the operator A A AAA. Obviously that Z Z ZZZ is a closed subspace of X X XXX.
For x X x X x in Xx \in XxX put
(2) E ( x ) = { y X : A y = A x and y = A x A } (2) E ( x ) = y X : A y = A x  and  y = A x A {:(2)E(x)={y in X:Ay=Ax" and "||y||=(||Ax||)/(||A||)}:}\begin{equation*} E(x)=\left\{y \in X: A y=A x \text { and }\|y\|=\frac{\|A x\|}{\|A\|}\right\} \tag{2} \end{equation*}(2)E(x)={yX:Ay=Ax and y=AxA}
Theorem 1. S. Cobzas, [1]. The following assertions hold: 1 1 1^(@)1^{\circ}1
(3) d ( x , Z ) A x A (3) d ( x , Z ) A x A {:(3)d(x","Z) >= (||Ax||)/(||A||):}\begin{equation*} d(x, Z) \geqslant \frac{\|A x\|}{\|A\|} \tag{3} \end{equation*}(3)d(x,Z)AxA
2 2 2^(@)2^{\circ}2
(4) d ( x , Z ) = A x A (4) d ( x , Z ) = A x A {:(4)d(x","Z)=(||Ax||)/(||A||):}\begin{equation*} d(x, Z)=\frac{\|A x\|}{\|A\|} \tag{4} \end{equation*}(4)d(x,Z)=AxA
if and only if there exists a sequence ( z n z n z_(n)z_{n}zn ) in Z Z ZZZ such that
(5) x z n A x A (5) x z n A x A {:(5)||x-z_(n)||rarr(||Ax||)/(||A||):}\begin{equation*} \left\|x-z_{n}\right\| \rightarrow \frac{\|A x\|}{\|A\|} \tag{5} \end{equation*}(5)xznAxA
(a) If (4) holds then
(6)
P Z ( x ) = x E ( x ) P Z ( x ) = x E ( x ) P_(Z)(x)=x-E(x)P_{Z}(x)=x-E(x)PZ(x)=xE(x)
(b) If there exists z 0 Z z 0 Z z_(0)in Zz_{0} \in Zz0Z such that
(7) x z 0 = A x A (7) x z 0 = A x A {:(7)||x-z_(0)||=(||Ax||)/(||A||):}\begin{equation*} \left\|x-z_{0}\right\|=\frac{\|A x\|}{\|A\|} \tag{7} \end{equation*}(7)xz0=AxA
then z 0 P Z ( x ) z 0 P Z ( x ) z_(0)inP_(Z)(x)z_{0} \in P_{Z}(x)z0PZ(x) and (4) and (6) hold.
By specializing the spaces X , Y X , Y X,YX, YX,Y and the operator A A AAA, S. Cobzas obtained in the above quoted paper a lot of duality results of Phelps type as well as other results on best approximation.
There are also some duality results as, e.g., those concerning normpreserving extensions of convex or star-shaped Lipschitz functions (see [2], [ 9 ] [ 9 ] [9][9][9] ) which cannot be derived from the theorem mentioned above. The aim of this paper is to prove a slight extension of Theorem 1 such as to cover these extension results, too.
Let X , Y X , Y X,YX, YX,Y be normed spaces over the same field K ( R K ( R K(R\mathbb{K}(\mathbb{R}K(R or C ) C ) C)\mathbb{C})C), and let A : X Y A : X Y A:X rarr YA: X \rightarrow YA:XY be a continuous linear operator. For a subspace X 1 X 1 X_(1)X_{1}X1 of X X XXX let
Z = Ker A and Z 1 = Ker ( A x 1 ) Z = Ker A  and  Z 1 = Ker A x 1 Z=Ker A" and "Z_(1)=Ker(A∣x_(1))Z=\operatorname{Ker} A \text { and } Z_{1}=\operatorname{Ker}\left(A \mid \mathrm{x}_{1}\right)Z=KerA and Z1=Ker(Ax1)
For x X x X x in Xx \in XxX let
(8) E ( x ) = { y X : A x = A y and y = A x A } (8) E ( x ) = y X : A x = A y  and  y = A x A {:(8)E(x)={y in X:Ax=Ay" and "||y||=(||Ax||)/(||A||)}:}\begin{equation*} E(x)=\left\{y \in X: A x=A y \text { and }\|y\|=\frac{\|A x\|}{\|A\|}\right\} \tag{8} \end{equation*}(8)E(x)={yX:Ax=Ay and y=AxA}
and
(9) E 1 ( x ) = { y X 1 : A x = A y and y = A x A } = X 1 E ( x ) (9) E 1 ( x ) = y X 1 : A x = A y  and  y = A x A = X 1 E ( x ) {:(9)E_(1)(x)={y inX_(1):Ax=Ay" and "||y||=(||Ax||)/(||A||)}=X_(1)nn E(x):}\begin{equation*} E_{1}(x)=\left\{y \in X_{1}: A x=A y \text { and }\|y\|=\frac{\|A x\|}{\|A\|}\right\}=X_{1} \cap E(x) \tag{9} \end{equation*}(9)E1(x)={yX1:Ax=Ay and y=AxA}=X1E(x)
Obviously, Z 1 Z 1 Z_(1)Z_{1}Z1 is a subspace of Z Z ZZZ and E 1 ( x ) E ( x ) E 1 ( x ) E ( x ) E_(1)(x)sube E(x)E_{1}(x) \subseteq E(x)E1(x)E(x), for x X 1 x X 1 x inX_(1)x \in X_{1}xX1.
Theorem 2. 1 1 1^(@)1^{\circ}1. For every x X 1 x X 1 x inX_(1)x \in X_{1}xX1 we have
(10) d ( x , Z 1 ) d ( x , Z ) A x A (10) d x , Z 1 d ( x , Z ) A x A {:(10)d(x,Z_(1)) >= d(x","Z) >= (||Ax||)/(||A||):}\begin{equation*} d\left(x, Z_{1}\right) \geqslant d(x, Z) \geqslant \frac{\|A x\|}{\|A\|} \tag{10} \end{equation*}(10)d(x,Z1)d(x,Z)AxA
2 2 2^(@)2^{\circ}2. For x X 1 x X 1 x inX_(1)x \in X_{1}xX1 we have
(11) d ( x , Z 1 ) = d ( x , Z ) = A x A (11) d x , Z 1 = d ( x , Z ) = A x A {:(11)d(x,Z_(1))=d(x","Z)=(||Ax||)/(||A||):}\begin{equation*} d\left(x, Z_{1}\right)=d(x, Z)=\frac{\|A x\|}{\|A\|} \tag{11} \end{equation*}(11)d(x,Z1)=d(x,Z)=AxA
if and only if there exists a sequence ( z n z n z_(n)z_{n}zn ) in Z 1 Z 1 Z_(1)Z_{1}Z1 such that
(12) x z n A x A . (12) x z n A x A . {:(12)||x-z_(n)||rarr(||Ax||)/(||A||).:}\begin{equation*} \left\|x-z_{n}\right\| \rightarrow \frac{\|A x\|}{\|A\|} . \tag{12} \end{equation*}(12)xznAxA.
3 3 3^(@)3^{\circ}3. (a) If the equalities (11) hold then
(13) P Z 1 ( x ) = x E 1 ( x ) (13) P Z 1 ( x ) = x E 1 ( x ) {:(13)P_(Z_(1))(x)=x-E_(1)(x):}\begin{equation*} P_{Z_{1}}(x)=x-E_{1}(x) \tag{13} \end{equation*}(13)PZ1(x)=xE1(x)
and
(14) P Z ( x ) = x E ( x ) (14) P Z ( x ) = x E ( x ) {:(14)P_(Z)(x)=x-E(x):}\begin{equation*} P_{Z}(x)=x-E(x) \tag{14} \end{equation*}(14)PZ(x)=xE(x)
(b) If there exists z 0 Z 1 z 0 Z 1 z_(0)inZ_(1)z_{0} \in Z_{1}z0Z1 such that
(15) x z 0 = A x A (15) x z 0 = A x A {:(15)||x-z_(0)||=(||Ax||)/(||A||):}\begin{equation*} \left\|x-z_{0}\right\|=\frac{\|A x\|}{\|A\|} \tag{15} \end{equation*}(15)xz0=AxA
then z 0 P Z ( x ) z 0 P Z ( x ) z_(0)inP_(Z)(x)z_{0} \in P_{Z}(x)z0PZ(x) and the equalities (11), (13) and (14) hold.
Proof. 1 1 1^(@)1^{\circ}1. Let x X 1 x X 1 x inX_(1)x \in X_{1}xX1. For every z Z 1 z Z 1 z inZ_(1)z \in Z_{1}zZ1 we have
A x = A x A z = A ( x z ) A x z A x = A x A z = A ( x z ) A x z ||Ax||=||Ax-Az||=||A(x-z)|| <= ||A||||x-z||\|A x\|=\|A x-A z\|=\|A(x-z)\| \leqslant\|A\|\|x-z\|Ax=AxAz=A(xz)Axz
implying
x z A x A , for all z Z 1 x z A x A ,  for all  z Z 1 ||x-z|| >= (||Ax||)/(||A||)," for all "z inZ_(1)\|x-z\| \geqslant \frac{\|A x\|}{\|A\|}, \text { for all } z \in Z_{1}xzAxA, for all zZ1
so that
d ( x , Z 1 ) A x A d x , Z 1 A x A d(x,Z_(1)) >= (||Ax||)/(||A||)d\left(x, Z_{1}\right) \geqslant \frac{\|A x\|}{\|A\|}d(x,Z1)AxA
Similarly
d ( x , Z ) A x A d ( x , Z ) A x A d(x,Z) >= (||Ax||)/(||A||)d(x, Z) \geqslant \frac{\|A x\|}{\|A\|}d(x,Z)AxA
Since Z 1 Z 1 Z_(1)Z_{1}Z1 is a subspace of Z Z ZZZ it follows that
d ( x , Z 1 ) d ( x , Z ) A x A d x , Z 1 d ( x , Z ) A x A d(x,Z_(1)) >= d(x,Z) >= (||Ax||)/(||A||)d\left(x, Z_{1}\right) \geqslant d(x, Z) \geqslant \frac{\|A x\|}{\|A\|}d(x,Z1)d(x,Z)AxA
for every x X 1 x X 1 x inX_(1)x \in X_{1}xX1.
2 2 2^(@)2^{\circ}2 If d ( x , Z 1 ) = A x A d x , Z 1 = A x A d(x,Z_(1))=(||Ax||)/(||A||)d\left(x, Z_{1}\right)=\frac{\|A x\|}{\|A\|}d(x,Z1)=AxA then, by the definition of d ( x , Z 1 ) d x , Z 1 d(x,Z_(1))d\left(x, Z_{1}\right)d(x,Z1), there exists a sequence ( z n ) z n (z_(n))\left(z_{n}\right)(zn) in Z 1 Z 1 Z_(1)Z_{1}Z1 such that
x z π A x A x z π A x A ||x-z_(pi)||rarr(||Ax||)/(||A||)\left\|x-z_{\pi}\right\| \rightarrow \frac{\|A x\|}{\|A\|}xzπAxA
Conversely, if ( z n ) z n (z_(n))\left(z_{n}\right)(zn) is a sequence in Z 1 Z 1 Z_(1)Z_{1}Z1 such that x z n A x A x z n A x A ||x-z_(n)||rarr(||Ax||)/(||A||)\left\|x-z_{n}\right\| \rightarrow \frac{\|A x\|}{\|A\|}xznAxA, then, since x z n d ( x , Z 1 ) , n N x z n d x , Z 1 , n N ||x-z_(n)|| >= d(x,Z_(1)),n inN\left\|x-z_{n}\right\| \geqslant d\left(x, Z_{1}\right), n \in \mathbb{N}xznd(x,Z1),nN, we get
d ( x , Z 1 ) lim n x z n = A x A d x , Z 1 lim n x z n = A x A d(x,Z_(1)) <= lim_(n rarr oo)||x-z_(n)||=(||Ax||)/(||A||)d\left(x, Z_{1}\right) \leqslant \lim _{n \rightarrow \infty}\left\|x-z_{n}\right\|=\frac{\|A x\|}{\|A\|}d(x,Z1)limnxzn=AxA
which, combined with (10), gives
d ( x , Z 1 ) = A x A d x , Z 1 = A x A d(x,Z_(1))=(||Ax||)/(||A||)d\left(x, Z_{1}\right)=\frac{\|A x\|}{\|A\|}d(x,Z1)=AxA
(a) If x X 1 x X 1 x inX_(1)x \in X_{1}xX1 is such that d ( x , Z 1 ) = A x A d x , Z 1 = A x A d(x,Z_(1))=(||Ax||)/(||A||)d\left(x, Z_{1}\right)=\frac{\|A x\|}{\|A\|}d(x,Z1)=AxA then the following equivalences hold
z P Z 1 ( x ) z Z 1 and x z = d ( x , Z 1 ) = A x A z P Z 1 ( x ) z Z 1  and  x z = d x , Z 1 = A x A z inP_(Z_(1))(x)Longleftrightarrow z inZ_(1)" and "||x-z||=d(x,Z_(1))=(||Ax||)/(||A||)z \in P_{Z_{1}}(x) \Longleftrightarrow z \in Z_{1} \text { and }\|x-z\|=d\left(x, Z_{1}\right)=\frac{\|A x\|}{\|A\|}zPZ1(x)zZ1 and xz=d(x,Z1)=AxA
x z E 1 ( x ) z x E 1 ( x ) . x z E 1 ( x ) z x E 1 ( x ) . Longleftrightarrow x-z inE_(1)(x)Longleftrightarrow z in x-E_(1)(x).\Longleftrightarrow x-z \in E_{1}(x) \Longleftrightarrow z \in x-E_{1}(x) .xzE1(x)zxE1(x).
Taking into account (11) one obtains the equivalences
z P Z ( x ) z Z and x z = d ( x , Z ) = A x A x z E ( x ) z x E ( x ) . z P Z ( x ) z Z  and  x z = d ( x , Z ) = A x A x z E ( x ) z x E ( x ) . {:[z inP_(Z)(x) Longleftrightarrow z in Z" and "||x-z||=d(x","Z)=(||Ax||)/(||A||)],[ Longleftrightarrow x-z in E(x)Longleftrightarrow z in x-E(x).]:}\begin{aligned} z \in P_{Z}(x) & \Longleftrightarrow z \in Z \text { and }\|x-z\|=d(x, Z)=\frac{\|A x\|}{\|A\|} \\ & \Longleftrightarrow x-z \in E(x) \Longleftrightarrow z \in x-E(x) . \end{aligned}zPZ(x)zZ and xz=d(x,Z)=AxAxzE(x)zxE(x).
(b) Let z 0 Z 1 z 0 Z 1 z_(0)inZ_(1)z_{0} \in Z_{1}z0Z1 be such that
x z 0 = A x A = d ( x , Z 1 ) = d ( x , Z ) x z 0 = A x A = d x , Z 1 = d ( x , Z ) ||x-z_(0)||=(||Ax||)/(||A||)=d(x,Z_(1))=d(x,Z)\left\|x-z_{0}\right\|=\frac{\|A x\|}{\|A\|}=d\left(x, Z_{1}\right)=d(x, Z)xz0=AxA=d(x,Z1)=d(x,Z)
It follows that (12) holds for z n = z 0 , n = 1 , 2 , z n = z 0 , n = 1 , 2 , z_(n)=z_(0),n=1,2,dotsz_{n}=z_{0}, n=1,2, \ldotszn=z0,n=1,2,, so that by the point 2 2 2^(@)2^{\circ}2 of the theorem, (11), (13) and (14) hold.

Application

Let X X XXX be a real normed space and Y Y YYY a nonvoid convex subset of X X XXX containing 0 .
Consider the space
(16) Lip 0 Y = { f : Y R : f Lip 0 Y = { f : Y R : f quadLip_(0)Y={f:Y rarrR:f\quad \operatorname{Lip}_{0} Y=\{f: Y \rightarrow \mathbb{R}: fLip0Y={f:YR:f is a Lipschitz on Y Y YYY and f ( 0 ) = 0 } f ( 0 ) = 0 } f(0)=0}f(0)=0\}f(0)=0}
equipped with the Lipschitz norm
(17)
f Y = sup { | f ( y 1 ) f ( y 2 ) | y 1 y 2 y 1 , y 2 Y , y 1 y 2 } . f Y = sup f y 1 f y 2 y 1 y 2 y 1 , y 2 Y , y 1 y 2 . ||f||_(Y)=s u p{(|f(y_(1))-f(y_(2))|)/(||y_(1)-y_(2)||)y_(1),y_(2)in Y,y_(1)!=y_(2)}.\|f\|_{Y}=\sup \left\{\frac{\left|f\left(y_{1}\right)-f\left(y_{2}\right)\right|}{\left\|y_{1}-y_{2}\right\|} y_{1}, y_{2} \in Y, y_{1} \neq y_{2}\right\} .fY=sup{|f(y1)f(y2)|y1y2y1,y2Y,y1y2}.
The space L i p 0 X L i p 0 X Lip_(0)XL i p_{0} XLip0X and the Lipschitz norm X X ||*||_(X)\|\cdot\|_{X}X are defined similarly.
By the theorem of McShane [7], [4], the space Lip 0 Y Lip 0 Y Lip_(0)Y\operatorname{Lip}_{0} YLip0Y has the extension property with respect to Lip 0 X Lip 0 X Lip_(0)X\operatorname{Lip}_{0} XLip0X, i.e., for every f Lip 0 Y f Lip 0 Y f inLip_(0)Yf \in \operatorname{Lip}_{0} YfLip0Y there exists f f f inf \inf Lip 0 X 0 X _(0)X_{0} X0X such that
F | Y = f and F X = f Y . F Y = f  and  F X = f Y . F|_(Y)=f" and "||F||_(X)=||f||_(Y).\left.F\right|_{Y}=f \text { and }\|F\|_{X}=\|f\|_{Y} .F|Y=f and FX=fY.
Definition 1. A function f f f inf \inf Lip 0 Y 0 Y _(0)Y_{0} Y0Y is called convex if
(18) f ( α y 1 + ( 1 α ) y 2 ) α f ( y 1 ) + ( 1 α ) f ( y 2 ) (18) f α y 1 + ( 1 α ) y 2 α f y 1 + ( 1 α ) f y 2 {:(18)f(alphay_(1)+(1-alpha)y_(2)) <= alpha f(y_(1))+(1-alpha)f(y_(2)):}\begin{equation*} f\left(\alpha y_{1}+(1-\alpha) y_{2}\right) \leq \alpha f\left(y_{1}\right)+(1-\alpha) f\left(y_{2}\right) \tag{18} \end{equation*}(18)f(αy1+(1α)y2)αf(y1)+(1α)f(y2)
for all y 1 , y 2 Y y 1 , y 2 Y y_(1),y_(2)in Yy_{1}, y_{2} \in Yy1,y2Y and all α [ 0 , 1 ] α [ 0 , 1 ] alpha in[0,1]\alpha \in[0,1]α[0,1], and starshaped if
(19) f ( α y ) α f ( y ) (19) f ( α y ) α f ( y ) {:(19)f(alpha y) <= alpha f(y):}\begin{equation*} f(\alpha y) \leq \alpha f(y) \tag{19} \end{equation*}(19)f(αy)αf(y)
for every y Y y Y y in Yy \in YyY and every α [ 0 , 1 ] α [ 0 , 1 ] alpha in[0,1]\alpha \in[0,1]α[0,1].
Obviously that every convex function f f f inf \inf Lip 0 Y 0 Y _(0)Y{ }_{0} Y0Y is starshaped.
Definition 2. A subset C C CCC of a vector space X X XXX is called a convex cone if
(a) x + y C x + y C x+y in Cx+y \in Cx+yC for every x , y C x , y C x,y in Cx, y \in Cx,yC, and
(b) λ x C λ x C lambda x in C\lambda x \in CλxC for every x C x C x in Cx \in CxC and λ 0 λ 0 lambda >= 0\lambda \geq 0λ0.
Denoting by K Y K Y K_(Y)K_{Y}KY (respectively by S Y S Y S_(Y)S_{Y}SY ) the sets of all convex (respectively starshaped) functions in Lip 0 Y Lip 0 Y Lip_(0)Y\operatorname{Lip}_{0} YLip0Y, it follows that K Y K Y K_(Y)K_{Y}KY and S Y S Y S_(Y)S_{Y}SY are convex cones in L i p 0 Y L i p 0 Y Lip_(0)YL i p_{0} YLip0Y.
The sets of convex (starshaped) Lipschitz functions in Lip 0 X Lip 0 X Lip_(0)X\operatorname{Lip}_{0} XLip0X are denoted by K X K X K_(X)K_{X}KX (respectively by S X S X S_(X)S_{X}SX ). Again they are convex cones in L i p 0 X L i p 0 X Lip_(0)XL i p_{0} XLip0X.
By McShane's theorem ([7], [4]), for every f L i p 0 Y f L i p 0 Y f in Lip_(0)Yf \in L i p_{0} YfLip0Y the function is a norm preserving extensions of f f fff, i.e.,
(21)
F | Y = f and F X = f Y F Y = f  and  F X = f Y F|_(Y)=f quad" and "quad||F||_(X)=||f||_(Y)\left.F\right|_{Y}=f \quad \text { and } \quad\|F\|_{X}=\|f\|_{Y}F|Y=f and FX=fY
a) if f K Y f K Y f inK_(Y)f \in K_{Y}fKY then the function F F FFF given by (20) belongs to K X K X K_(X)K_{X}KX, i.e., K Y K Y K_(Y)K_{Y}KY has the extension property with respect to K X K X K_(X)K_{X}KX;
b) if f S Y f S Y f inS_(Y)f \in S_{Y}fSY then F S X F S X F inS_(X)F \in S_{X}FSX, so that S Y S Y S_(Y)S_{Y}SY has the extension property with respect to S X S X S_(X)S_{X}SX, too.
Consider the subspaces
(22) X c = K X K X (22) X c = K X K X {:(22)X_(c)=K_(X)-K_(X):}\begin{equation*} X_{c}=K_{X}-K_{X} \tag{22} \end{equation*}(22)Xc=KXKX
and
(23) X s = S X S X (23) X s = S X S X {:(23)X_(s)=S_(X)-S_(X):}\begin{equation*} X_{s}=S_{X}-S_{X} \tag{23} \end{equation*}(23)Xs=SXSX
generated by the cones K X K X K_(X)K_{X}KX and S X S X S_(X)\mathrm{S}_{X}SX, respectively. We have X c X s Lip p 0 X X c X s Lip p 0 X X_(c)subX_(s)sub Lipp_(0)XX_{c} \subset X_{s} \subset \operatorname{Lip} p_{0} XXcXsLipp0X.
Take in Theorem 2, A A AAA to be the restriction operator r : Lip 0 X Lip 0 Y r : Lip 0 X Lip 0 Y r:Lip_(0)X rarrLip_(0)Yr: \operatorname{Lip}_{0} X \rightarrow \operatorname{Lip}_{0} Yr:Lip0XLip0Y defined by
(24) r ( F ) = F | Y , F L i p 0 X (24) r ( F ) = F Y , F L i p 0 X {:(24)r(F)=F|_(Y)","F in Lip_(0)X:}\begin{equation*} r(F)=\left.F\right|_{Y}, F \in L i p_{0} X \tag{24} \end{equation*}(24)r(F)=F|Y,FLip0X
The operator r r rrr is linear, continuous, and, since
r ( F ) Y = F | Y Y F X , r ( F ) Y = F Y Y F X , ||r(F)||_(Y)=||F|_(Y)||_(Y) <= ||F||_(X),\|r(F)\|_{Y}=\left\|\left.F\right|_{Y}\right\|_{Y} \leq\|F\|_{X},r(F)Y=F|YYFX,
it follows r 1 r 1 ||r|| <= 1\|r\| \leq 1r1.
For f Lip 0 Y f Lip 0 Y f inLip_(0)Yf \in \operatorname{Lip}_{0} YfLip0Y put
(25) E ( f ) = { F L i p 0 X : F | Y = f and F X = f Y } . (25) E ( f ) = F L i p 0 X : F Y = f  and  F X = f Y . {:(25)E(f)={F in Lip_(0)X:F|_(Y)=f" and "||F||_(X)=||f||_(Y)}.:}\begin{equation*} E(f)=\left\{F \in L i p_{0} X:\left.F\right|_{Y}=f \text { and }\|F\|_{X}=\|f\|_{Y}\right\} . \tag{25} \end{equation*}(25)E(f)={FLip0X:F|Y=f and FX=fY}.
By McShane's theorem [7] the set E ( f ) E ( f ) E(f)E(f)E(f) is nonempty for every f Lip 0 Y f Lip 0 Y f inLip_(0)Yf \in \operatorname{Lip}_{0} YfLip0Y.
If f K Y f K Y f inK_(Y)f \in K_{Y}fKY then
(26) E c ( f ) = { F E ( f ) : F K X } (26) E c ( f ) = F E ( f ) : F K X {:(26)E_(c)(f)={F in E(f):F inK_(X)}!=O/:}\begin{equation*} E_{c}(f)=\left\{F \in E(f): F \in K_{X}\right\} \neq \varnothing \tag{26} \end{equation*}(26)Ec(f)={FE(f):FKX}
(27) E s ( f ) = { F E ( f ) : F S X } (27) E s ( f ) = F E ( f ) : F S X {:(27)E_(s)(f)={F in E(f):F inS_(X)}!=O/:}\begin{equation*} E_{s}(f)=\left\{F \in E(f): F \in S_{X}\right\} \neq \varnothing \tag{27} \end{equation*}(27)Es(f)={FE(f):FSX}
It follows that for f S Y f S Y f inS_(Y)f \in S_{Y}fSY
E s ( f ) E ( f ) E s ( f ) E ( f ) E_(s)(f)sube E(f)E_{s}(f) \subseteq E(f)Es(f)E(f)
and for f K Y f K Y f inK_(Y)f \in K_{Y}fKY
E c ( f ) E s ( f ) E ( f ) E c ( f ) E s ( f ) E ( f ) E_(c)(f)subeE_(s)(f)sube E(f)E_{c}(f) \subseteq E_{s}(f) \subseteq E(f)Ec(f)Es(f)E(f)
Let
Y = { F L i p 0 X : F | Y = 0 } Y c = Y X c Y = F L i p 0 X : F Y = 0 Y c = Y X c {:[Y^(_|_)={F in Lip_(0)X:F|_(Y)=0}],[Y_(c)^(_|_)=Y^(_|_)nnX_(c)]:}\begin{gathered} Y^{\perp}=\left\{F \in L i p_{0} X:\left.F\right|_{Y}=0\right\} \\ Y_{c}^{\perp}=Y^{\perp} \cap X_{c} \end{gathered}Y={FLip0X:F|Y=0}Yc=YXc
and
Y s = Y X s Y s = Y X s Y_(s)^(_|_)=Y^(_|_)nnX_(s)Y_{s}^{\perp}=Y^{\perp} \cap X_{s}Ys=YXs
We have the following result
Proposition 1. (a) If F S X F S X F inS_(X)F \in S_{X}FSX then F | Y S Y , E s ( F | Y ) F Y S Y , E s F Y F|_(Y)inS_(Y),E_(s)(F|_(Y))!=O/\left.F\right|_{Y} \in S_{Y}, E_{s}\left(\left.F\right|_{Y}\right) \neq \varnothingF|YSY,Es(F|Y) and r | X s = 1 r X s = 1 ||r|_(X_(s))||=1\left\|\left.r\right|_{X_{s}}\right\|=1r|Xs=1. Furthermore
(28)
d ( F , Y s ) = d ( F , Y ) = F | Y d F , Y s = d F , Y = F Y d(F,Y_(s)^(_|_))=d(F,Y^(_|_))=||F|_(Y)||d\left(F, Y_{s}^{\perp}\right)=d\left(F, Y^{\perp}\right)=\left\|\left.F\right|_{Y}\right\|d(F,Ys)=d(F,Y)=F|Y
and
(29) F E s ( F | Y ) = P Y s ( F ) P Y ( F ) = F E ( F | Y ) (29) F E s F Y = P Y s ( F ) P Y ( F ) = F E F Y {:(29)F-E_(s)(F|_(Y))=P_(Y_(s)^(_|_))(F)subeP_(Y^(_|_))(F)=F-E(F|_(Y)):}\begin{equation*} F-E_{s}\left(\left.F\right|_{Y}\right)=P_{Y_{s}^{\perp}}(F) \subseteq P_{Y^{\perp}}(F)=F-E\left(\left.F\right|_{Y}\right) \tag{29} \end{equation*}(29)FEs(F|Y)=PYs(F)PY(F)=FE(F|Y)
(b) If F K X F K X F inK_(X)F \in K_{X}FKX then F | Y K Y , E c ( F | Y ) F Y K Y , E c F Y F|_(Y)inK_(Y),E_(c)(F|_(Y))!=O/\left.F\right|_{Y} \in K_{Y}, E_{c}\left(\left.F\right|_{Y}\right) \neq \varnothingF|YKY,Ec(F|Y) and r | X c = 1 r X c = 1 ||r|_(X_(c))||=1\left\|\left.r\right|_{X_{c}}\right\|=1r|Xc=1. Furthermore
(30) F E c ( F | Y ) = P Y c ( F ) P Y s ( F ) = F E s ( F | Y ) P Y ( F ) = F E ( f ) (30) F E c F Y = P Y c ( F ) P Y s ( F ) = F E s F Y P Y ( F ) = F E ( f ) {:[(30)F-E_(c)(F|_(Y))=P_(Y_(c)^(_|_))(F)subeP_(Y_(s)^(_|_))(F)],[=F-E_(s)(F|_(Y))subeP_(Y^(_|_))(F)=F^(-)-E(f)]:}\begin{gather*} F-E_{c}\left(\left.F\right|_{Y}\right)=P_{Y_{c}^{\perp}}(F) \subseteq P_{Y_{s}^{\perp}}(F) \tag{30}\\ =F-E_{s}\left(\left.F\right|_{Y}\right) \subseteq P_{Y^{\perp}}(F)=F^{-}-E(f) \end{gather*}(30)FEc(F|Y)=PYc(F)PYs(F)=FEs(F|Y)PY(F)=FE(f)
Proof. (a) If F E s ( F | Y ) F E s F Y F inE_(s)(F|_(Y))F \in E_{s}\left(\left.F\right|_{Y}\right)FEs(F|Y) then G = F F ¯ Y s G = F F ¯ Y s G=F- bar(F)inY_(s)^(_|_)G=F-\bar{F} \in Y_{s}^{\perp}G=FF¯Ys and
F G X = F ¯ X = F | Y Y = r ( F G ) Y r | X s F G Y F G X = F ¯ X = F Y Y = r ( F G ) Y r X s F G Y ||F-G||_(X)=|| bar(F)||_(X)=||F|_(Y)||_(Y)=||r(F-G)||_(Y) <= ||r|_(X_(s))||||F-G||_(Y)\|F-G\|_{X}=\|\bar{F}\|_{X}=\left\|\left.F\right|_{Y}\right\|_{Y}=\|r(F-G)\|_{Y} \leq\left\|\left.r\right|_{X_{s}}\right\|\|F-G\|_{Y}FGX=F¯X=F|YY=r(FG)Yr|XsFGY
implying r | X s = 1 r X s = 1 ||r|_(X_(s))||=1\left\|\left.r\right|_{X_{s}}\right\|=1r|Xs=1. By Theorem 2.(b), G P Y s ( F ) G P Y s ( F ) G inP_(Y_(s)^(_|_))(F)G \in P_{Y_{s}^{\perp}}(F)GPYs(F) so that (28) and (29) hold. The proof of (b) is similar.

REFERENCES

[1] COBZAŞ, S., Some Applications of a Distance Formula to the Kernel of a Linear Operator (submitted)
[2] COBZAŞ, S., MUSTĂTA, C., Norm Preserving Extension of Convex Lipschitz Functions, J. Approx. Theory, 24, pp. 236-244, 1978.
[3] COBZAŞ, S., MUSTÃTA, C., Extension of Bilinear Functionals and Best Approximation in 2-normed Spaces, Studia Univ. "Babeş-Bolyai", Series Math., XLIII, no. 2, pp. 1-13, 1998.
[4] CZIPSER, J., GÉHÉR, L., Extension of Functions Satisfying a Lipschitz Condition, Acta Math. Acad. Sci. Hungar, 6, pp. 213-220, 1955.
[5] DEUTSCH, F., WU LI and SIZWE MABIZELA, Helly Extensions and Best Approximation, Parametric Optimization and Related Topics III (J. Guddat, H. Th. Jongen, B. Kummer and F. Nošiaeka Eds., Appromation and Optimization, vol. 3, Verlag Peter Lang, pp. 107-120, Frankfurt 1993.
[6] ENGELKING, R., General Topology, PWN Warszawa, 1985.
[7] McSHANE, E. J., Extension of Range of Functions, Bull. Amer. Math., 40, pp. 834-842, 1934.
[8] MUSTATA, C., Best Approximation and Unique Extension of Lipschitz Functions, J. Approx. Theory, 19, pp. 222-230, 1977.
[9] MUSTĂTA, C., Norm Preserving Extension of Starshaped Lipschitz Functions, Mathematica, 19, pp. 183-187, 1977.
[10] MUSTĂTA, C., Extension of Bounded Lipschitz Functions and Some Related Problems of Best Approximation, "Babeş-Bolyai" Univ., Seminar on Func. Anal. and Numerical Methods Preprint, no. 4, pp. 93-99, 1981.
[11] MUSTĂTA, C., Extension of Holder Functions and Some Related Problems of Best Approximation, Seminar on Mathematic:al Analysis "Babeş-Bolyai" Univ. Preprint, no. 7, pp. 71-86, 1991.
[12] SUNG-HO, PARK, Quotient Mappings, Helly Extensions, Hahn-Banach Extension, Tietze Extensions, Lipschitz Extension and Best Approximation, J. Koreean Math. Soc., 29, no. 2, pp. 239-250, 1992.
[13] PHELPS, R. R., Uniqueness of Hahn-Banach Extensions and Best Approximations, Trans. Amer. Math. Soc., 95, pp. 238-255, 1960.
Received September 21, 1999
"T. Popoviciu" Institute
of Numerical Analysis
С. P. 68,
str. Gh. Bilaşcu 37
3400 Cluj-Napoca, Romania

  1. 2000 AMS Subject Classification: 41A65.
    *Supported by the Ministry of Research and Technology (GR 4122/1999).
2000

Related Posts