Journal of Approximation Theory AT3102 Journal of Approximation Theory 91, 332348 (1997) Article No. AT963102 On Best Simultaneous Approximation Chong Li* Department of Mathematics, Hangzhou Institute of Commerce, Hangzou, People's Republic of China and G. A. Watson Department of Mathematics and Computer Science, University of Dundee, Dundee DD1 4HN, Scotland Communicated by Andras Kroo Received May 17, 1995; accepted in revised form December 3, 1996 The problem is considered of best approximation of finite number of functions simultaneously. For a very general class of norms, characterization results are derived. The main part of the paper is concerned with proving uniqueness and strong uniqueness theorems. For a particular subclass, which includes the important special case of the Chebyshev norm, a characterization is given of a uniqueness element. 1997 Academic Press 1. INTRODUCTION Let X be a compact Hausdorff space and Y a normed linear space with norm & } & Y . Let C(X, Y ) denote the set of all continuous functions from X to Y, and let & } & A be a norm on C(X, Y ). Let U be defined by U=[a # R l, &a& B 1], where & } & B is a given norm on R l. Define a norm on l-tuples of elements of C(X, Y ) as follows: for any F=(, 1 , ..., , l ) # C(X, Y ) l define &F&=max a#U " l : ai ,i i=1 ", (1.1) A where a=(a 1 , ..., a l ). * Present address: Center for Mathematical Sciences, Zhejiang University, Hangzhou 310027, P. R. China. 332 0021-904597 25.00 Copyright 1997 by Academic Press All rights of reproduction in any form reserved. File: DISTIL 310201 . By:DS . Date:26:11:97 . Time:15:31 LOP8M. V8.0. Page 01:01 Codes: 3467 Signs: 1490 . Length: 50 pic 3 pts, 212 mm ON BEST SIMULTANEOUS APPROXIMATION 333 Now suppose that functions , 1 , ..., , l in C(X, Y ) are given. Then the problem is considered here of approximating these functions simultaneously by functions in S, a subspace of C(X, Y ), in the sense of the minimization of the norm (1.1). In other words, we want to find an l-tuple f =(,, ..., ,), where , # S, to minimize &F& f &. (1.2) If such a function f * exists, it is called a best simultaneous approximation to F=(, 1 , ..., , l ). Problems of simultaneous approximation can be viewed as special cases of vector-valued approximation, and some recent work in this area is due to Pinkus [6], who points out that many questions remain unresolved. He is concerned with the question of when a finite dimensional subspace is a unicity space, for some different norms from those considered here. We are also primarily interested in uniqueness questions. Characterization results for linear problems were recently given in [9] based on the derivation of an expression for the directional derivative, and these generalized earlier work of [8]. We being by showing how these results can be obtained in a simpler and more direct manner, which permits their extension to some nonlinear problems. The rest of the paper is concerned with uniqueness and strong uniqueness of best approximations. In what follows, finite sequences of identical elements identified by (,, ..., ,) will be assumed to be l-tuples. 2. CHARACTERIZATION OF BEST APPROXIMATIONS Let C*(X, Y ) denote the dual space of C(X, Y ), and let W denote the dual unit ball. For F=(, 1 , ..., , l ) # C(X, Y ) l define l g F (a, w)= : a i(w, , i ), for all (a, w) # U_W, i=1 where the inner product notation links elements of C(X, Y ) and its dual. Note that U_W is endowed with the product topology, while W is endowed with the weak * topology. Since for any (a 0, w 0 ) # U_W, } l l | g F (a, w)& g F (a 0, w 0 )| = : a i( w, , i ) & : a 0i ( w 0, , i ) i=1 l i=1 } l : |a i &a 0i | |( w, , i ) | + : |a 0i | |( w&w 0, , i ) |, i=1 i=1 File: DISTIL 310202 . By:DS . Date:26:11:97 . Time:15:31 LOP8M. V8.0. Page 01:01 Codes: 2688 Signs: 1909 . Length: 45 pic 0 pts, 190 mm 334 LI AND WATSON it follows that g F ( } , } ) # C(U_W ) (the space of continuous functions defined on U_W ). Note that for any , # S, f =(,, ..., ,), \ l + g f (a, w)= : a i ( w, ,) # C(U_W ). i=1 Further for any such f, &F& f &=max a#U " l : a i (, i &,) i=1 } " A l =max max : a i( w, , i &,) a#U w#W i=1 } =&g F ( } , } )& g f ( } , } )& C , where & } & C denotes the uniform norm on C(U_W ). Now define Sg =[g f : f =(,, ..., ,), , # S]. It follows that f *=(,*, ..., ,*), ,* # S is a best simultaneous approximation to F=(, 1 , ..., , l ) if and only if g f * # S g is a best approximation to g F in the uniform norm of C(U_W ). Let P S (F ) denote the set of all best simultaneous approximations f =(,, ..., ,), where , # S, to F. In addition, let d(F, S)=inf [&F& f & : f =(,, ..., ,), , # S], d(F, C)=d(F, C(X, Y )). Definition 1. A set S is a sunset for simultaneous approximation if for any F=(, 1 , ..., , l ), and f *=(,*, ..., ,*), ,* # S, f * # P S (F ) implies that f * # P S (F : ) for F : = f *+:(F& f *), and :0. Descriptions of sunsets (or strict suns), and solar properties, are given in [3]. Linear sets are examples of suns, as are convex sets, but also some nonconvex sets, for example rational functions. Theorem 1. Let S/C(X, Y ) be a sunset of simultaneous approximation. Then f * # P S (F ) if and only if for any f =(,, ..., ,), , # S, there exists a # ext(U ), w # ext(W ) such that g F& f *(a, w)=&F& f *&, g f *& f (a, w)0, where ``ext'' denotes the set of extreme points. File: DISTIL 310203 . By:DS . Date:26:11:97 . Time:15:31 LOP8M. V8.0. Page 01:01 Codes: 2542 Signs: 1300 . Length: 45 pic 0 pts, 190 mm 335 ON BEST SIMULTANEOUS APPROXIMATION Proof. If S is a sunset for simultaneous approximation, then S g is a strict sun for uniform approximation in C(U_W ). The result then follows from the generalized Kolmogorov criterion characterizing a best approximation in C(U_W ) with the uniform norm (see, for example [3, Theorem I.2.4]), using the KreinMilman Theorem. K A special case of (1.1) is given by &,& A =max &,(t)& Y , (2.1) t#X for any , # C(X, Y ). This includes the important case of the Chebyshev norm on l-tuples. Let F=(, 1 , ..., , l ) # C(X, Y ) l. Then a Chebyshev norm may be defined by &F&= max max &, i (t)& Y , (2.2) 1il t # X which is the special case of (1.1) when & } & A is given by (2.1) and & } & B is the l 1 norm [see 8, 9]. Define { H(F, f )= t # X: max &a&B =1 " l " i=1 = =&F& f & . : a i (, i (t)&,(t)) Y (2.3) Let Z denote the unit ball in Y*, the dual space of Y, and let ( } , } ) Y denote the inner product linking Y and Y*. Then using the form of points in ext(W ) in this case, we have the following corollary of Theorem 1. Corollary 1. Let & } & A be given by (2.1), and let S/C(X, Y ) be a sunset for simultaneous approximation. Then f * # P S (F ) if and only if for any f =(,, ..., ,), , # S, there exists a # ext(U ), t # H(F, f ), v(t) # ext(Z) such that l : a i( v(t), , i (t)&,*(t)) Y =&F& f *&, i=1 \ l + : a i ( v(t), ,*(t)&,(t)) Y 0. i=1 Returning to the general problem, standard linear theory (for example [7]) gives the following result. Theorem 2. Let S be an n-dimensional subspace of C(X, Y ). Then f * # P S (F) if and only if there exists a j # ext(U ), w j # ext(W ), : j >0, j=1, ..., r with rj=1 : j =1 and 1rn+1 such that File: DISTIL 310204 . By:DS . Date:26:11:97 . Time:15:31 LOP8M. V8.0. Page 01:01 Codes: 2599 Signs: 1527 . Length: 45 pic 0 pts, 190 mm 336 LI AND WATSON g F& f *(a j, w j )=&F& f *&, j=1, ..., r, r : : j g f (a j, w j )=0 for all f # S. j=1 This is just the result given as Theorem 1 in [9]. 3. UNIQUENESS OF BEST APPROXIMATIONS For the general case uniqueness is a consequence of strict convexity of the norm & } & A . This is established next. It is convenient to extract the following result as a preliminary lemma. Lemma 1. Let S/C(X, Y ), and let F=(, 1 , ..., , l ) # C(X, Y ) l. Let f *=(,*, ..., ,*) # P S (F), with a* # U such that " l &F& f *&= : a *(, i i &,*) ". (3.1) A i=1 Then if d(F, C)<d(F, S), li=1 a *{0. i Proof. Assume that d(F, C)<d(F, S) and also that a* satisfying (3.1) is such that li=1 a*i =0. Then there exists , # C(X, Y ) such that max a#U " l l " " " = : a *(, &,) . " " : a i (, i &,) i=1 < : a *(, i i &,*) A A i=1 l i i A i=1 This is a contradiction which proves the result. K Theorem 3. Let & } & A be strictly convex, and let S/C(X, Y ) be a sunset for simultaneous approximation. Then for any F=(, 1 , ..., , l ), d(F, C)= d(F, S) or P S (F ) contains at most one element. Proof. Let F=(, 1 , ..., , l ) be such that d(F, C)<d(F, S). Suppose that f *=(,*, ..., ,*) # P S (F), f =(,, ..., , ) # P S (F ) with ,*{,. Let , 0i =2, i &,, i=1, ..., l, F 0 =(, 01 , ..., , 0l ). Then it follows from the definition of a sunset that f # P S (F 0 ). Also File: DISTIL 310205 . By:DS . Date:26:11:97 . Time:15:31 LOP8M. V8.0. Page 01:01 Codes: 2402 Signs: 1158 . Length: 45 pic 0 pts, 190 mm 337 ON BEST SIMULTANEOUS APPROXIMATION &F 0 & f *&=&2F& f & f *& &F& f &+&F& f *& =2 &F& f & =&F 0 & f &. (3.2) Therefore f * # P S (F 0 ). Further l " " max _" : a (, &,)" +" : a (, &,*)" & &2F& f & f *&=max a#U : a i (2, i &, &,*) A i=1 l l i a#U i i A i=1 0 &F & f &, i A i=1 using (3.2). (3.3) Thus equality holds to (3.3). Let a # U be chosen so that " l : (2, i &, &,*) " =&2F& f & f *&=2d(F, C). A i=1 Thus " l l " : a i (, i &, )+ : a i (, i &,*) i=1 i=1 " l A l = : a i (, i &, ) " +" : a (, &,*)" , i A i=1 i i=1 A which, using the assumption of strict convexity, implies that l l : a i (, i &, )= : a i (, i &,*) i=1 i=1 or \ l + : a i (,*&, )=0. i=1 (3.4) Since by assumption d(F, C)<d(F, S), it follows from Lemma 1 that ( li=1 a i ){0, and so ,*=, and the result is established. K For the special case covered by (2.1), it is possible to give a precise characterization of a uniqueness element, which is defined as follows. File: DISTIL 310206 . By:DS . Date:26:11:97 . Time:15:31 LOP8M. V8.0. Page 01:01 Codes: 2337 Signs: 773 . Length: 45 pic 0 pts, 190 mm 338 LI AND WATSON Definition 2. Let S be a sunset of C(X, Y ), and f *=(,*, ..., ,*), with ,* # S. Then ,* is called a uniqueness element of S if for any F=(, 1 , ..., , l ) # C(X, Y ) l with d(F, C)<d(F, S), f * # P S (F ), then f * is a unique best approximation to F from S. Theorem 4. Let Y be strictly convex, let & } & A be given by (2.1), and let S be a sunset for simultaneous approximation. Let f *=(,*, ..., ,*) # S l. Then ,* is a uniqueness element of S if and only if for any F= (, 1 , ..., , l ) # C(X, Y ) l, with f * # P S (F ) and d(F, C)<d(F, S), ,* is uniquely determined by the set of values in H(F, f *) (that is, if , # S and ,(t)=,*(t) for all t # H(F, f *), then ,=,*). Proof. Let ,* be a uniqueness element of S. Suppose that for some F=(, 1 , ..., , l ) # C(X, Y ) l, with f * # P S (F ) and d(F, C)<d(F, S), there exists f =(,, ..., , ), , # S, , {,*, such that ,(t)=,*(t), for all t # H(F, f *). Define for all t # X _ , 0i (t)=,*(t)+ &,*&,& A & max _ &a&B =1 \ l + : a j &,*(t)&,(t)& Y j=1 , i (t)&,*(t) , &F& f *& & (3.5) and let F 0 =(, 01 , ..., , 0l ). It is easy to verify directly that &F 0 & f *&&,*&,& A . (3.6) Since for any t # H(F, f *), max &a&B =1 " l : a i (, 0i (t)&,*(t)) i=1 " =&,*&,& A , Y then &F 0 & f *&&,*&,& A . (3.7) It follows from (3.6) and (3.7) that &F 0 & f *&=&,*&,& A . File: DISTIL 310207 . By:DS . Date:26:11:97 . Time:15:31 LOP8M. V8.0. Page 01:01 Codes: 2582 Signs: 1157 . Length: 45 pic 0 pts, 190 mm (3.8) 339 ON BEST SIMULTANEOUS APPROXIMATION From Theorem 1, because f * # P S (F ), for any , # S, there exists a # ext(U ), w # ext(W ), such that l \ : a + ( w, ,*&,) 0, w, : a i (, i &,*) =&F& f *&, i=1 (3.9) l i i=1 and so using Corollary 1, for some t # H(F, f *), v(t) # ext(Z), l : a i( v(t), , i (t)&,*(t)) Y =&F& f *&. i=1 From (3.5), l : a i( v(t), , 0i (t)&,*(t)) Y =&,*&,& A i=1 =&F 0 & f *&, using (3.8). Thus t # H(F 0, f *), and we must have l w, : a i (, 0i &,*) =&F 0 & f *&. i=1 (3.10) Equations (3.9) and (3.10) show, using Theorem 1, that f * # P S (F 0 ). Now for any t # X, max &a&B =1 " l " : a i (, 0i (t)&,(t)) i=1 Y l "\ : a + (,*(t)&,(t))" + max "_&,*&,& & max "\ : a + (,*(t)&,(t))" & 1 : a (, (t)&,*(t)) _ " &F& f *& max &a&B =1 i Y i=1 l A &a&B =1 &a&B =1 i i=1 l i i=1 i Y &,*&,& A . File: DISTIL 310208 . By:DS . Date:26:11:97 . Time:15:31 LOP8M. V8.0. Page 01:01 Codes: 2046 Signs: 648 . Length: 45 pic 0 pts, 190 mm Y 340 LI AND WATSON Thus &F 0 & f &&F 0 & f *&, using (3.8), so that f # P S (F 0 ), a contradiction. This proves necessity. Now suppose that for some F=(, 1 , ..., , l ) # C(X, Y ) l with d(F, C)<d(F, S) and f * # P S (F ), f =(,*, ..., ,*), there exists another f # P S (F ), f =(,, ..., , ). Let , 0i =2, i &,, i=1, ..., l, and let F 0 =(, 01 , ..., , 0l ). It follows from the definition of a sunset that f # P S (F 0 ). Now l " " max : a (, &, ) + max : a (, &,*) " " " " &F 0 & f *&= max : a i (2, i &, &,*) &a&B =1 A i=1 l l i &a&B =1 i i &a&B =1 A i=1 i i=1 =2 &F& f & =&F 0 & f &. Thus f * # P S (F 0 ). Now let a # ext(U ), w # ext(W ) such that l w, : a i (2, i &, &,*) =&F 0 & f *&, i=1 which is possible using Theorem 1. Thus for any t # H(F 0, f *), l " " : a (, &, )(t) + : a (, &,*)(t) " " " " &F 0 & f *&= : a i (2, i &, &,*)(t) Y i=1 l l i i i=1 i Y i i=1 &F& f &+&F& f *& =&F 0 & f *&. File: DISTIL 310209 . By:DS . Date:26:11:97 . Time:15:31 LOP8M. V8.0. Page 01:01 Codes: 2260 Signs: 709 . Length: 45 pic 0 pts, 190 mm Y A 341 ON BEST SIMULTANEOUS APPROXIMATION It follows that " l l : a i (, i &, )(t)+ : a i (, i &,*)(t) i=1 i=1 " l " Y l = : a i (, i &, )(t) i=1 " +" : a (, &,*)(t)" . i Y i i=1 Y Therefore using the strict convexity of Y, l l : a i (, i &, )(t)= : a i (, i &,*)(t), i=1 i=1 or equivalently \ l + : a i (,*(t)&,(t))=0, i=1 for all t # H(F 0, f *). Since d(F, C)<d(F, S), by Lemma 1 we must have li=1 a i {0, and so ,*(t)=,(t) for all t # H(F 0, f *). This proves the sufficiency of the stated conditions. K 4. STRONG UNIQUENESS It is possible to establish strong uniqueness for the general problem under a condition which generalizes the Chebyshev set condition for linear best approximation in the uniform norm. The result hinges on the derivation of the analogue of the strong Kolomogorov condition for finite dimensional spaces (see, for example, Wulbert [10] or Nurnberger [5]). Definition 3 [1]. An n-dimensional subspace S of C(X, Y ) is called an interpolating subspace if no nontrivial linear combination of n linearly independent extreme points of W annihilates S. Theorem 5. Let S be an interpolating subspace of C(X, Y ), and let d(F, C)<d(F, S). Then f *=(,*, ..., ,*) # P S (F) is a strongly unique best simultaneous approximation to F, that is, there exists #>0 such that &F& f &&F& f *&+# &,&,*& A for all f =(,, ..., ,), File: DISTIL 310210 . By:DS . Date:26:11:97 . Time:15:31 LOP8M. V8.0. Page 01:01 Codes: 2366 Signs: 1219 . Length: 45 pic 0 pts, 190 mm , # S. 342 LI AND WATSON Proof. We can use Theorem 2. Since f * # P S (F), it follows that there exist a j # ext(U ), w j # ext(W ), : j >0, j=1, ..., r with rj=1 : j =1, and 1rn+1 such that l : a ij(w j, , i &,*) =&F& f *&, j=1, ..., r, (4.1) for all , # S. (4.2) i=1 r : :j j=1 \ l + : a ij ( w j, ,*&,) =0 i=1 With no loss of generality, we can assume that the set [w j, j=1, ..., r] is linearly independent. Because d(F, C)<d(F, S), it follows from Lemma 1 that ; j =: j ( li=1 a ij ){0, j=1, ..., n. Assume that r<n+1. Then we can take an element , 0 # S such that , 0 {,*, and ( w j, ,*&, 0 ) =; j , j=1, ..., r, using the fact that S is an interpolating subspace. This means that r r : ; j( w j, ,*&, 0 ) = : ; 2j >0, j=1 j=1 which is a contradiction. It follows that r=n+1. Now for any , # S, any (a j, w j ) # ext U_ext W satisfying (4.1) and (4.2) assume that we have \ l + : a ij ( w j, ,*&,) 0, i=1 j=1, ..., r. Then it follows from (4.2) that r 0= : ; j(w j, ,*&,) 0, j=1 which in turn implies that ( w j, ,*&,) =0, j=1, ..., r, or ,*=,, since r=n+1. Thus there exists (a, w) # ext U_ext W with l : a i( w, , i &,*) =&F& f *&, i=1 l \ : a + ( w, ,*&,) >0. i i=1 File: DISTIL 310211 . By:DS . Date:26:11:97 . Time:15:31 LOP8M. V8.0. Page 01:01 Codes: 2167 Signs: 995 . Length: 45 pic 0 pts, 190 mm 343 ON BEST SIMULTANEOUS APPROXIMATION Now define l { = M(F, f *)= (a, w) # ext(U )_ext(W ) : : a i( w, , i &,*) =&F& f *& . i=1 Then we have max (a, w) # M(F, f *) \ l + : a i ( w, ,*&,) >0, i=1 for all f # S. Let #= inf max , # S "[,*] (a, w) # M(F, f *) {\ l : ai i=1 + w, ,*&, & f *& f & = . Then #>0, since S is finite dimensional. Further for any , # S, l l \ l + : a i( w, , i &,) = : a i( w, , i &,*) + : a i ( w, ,*&,). i=1 i=1 i=1 For any (a, w) # M(F, f *), it follows that &F& f &&&F& f *& max (a, w) # M(F, f *) \ l + : a i ( w, ,*&,) i=1 # &,*&,& A . This implies that &F& f &&F& f *&+# &,&,*& A and the proof is complete. for all f # S, K Definition 4. For any normed linear space [E, & } &], the modulus of convexity is defined by $ E (=)=inf [1& 12 &x+ y& : x, y # E, &x& y&==, &x&=&y&=1], for 0<=2. Definition 5 [2, 4, 11]. E is said to be uniformly convex if $ E (=)>0 for any 0<=2. A uniformly convex space E is p-uniformly convex (or has modulus of convexity of power type p) if for some c>0, $ E (=)c= p. File: DISTIL 310212 . By:DS . Date:26:11:97 . Time:15:31 LOP8M. V8.0. Page 01:01 Codes: 2148 Signs: 848 . Length: 45 pic 0 pts, 190 mm 344 LI AND WATSON Examples of uniformly convex spaces are Hilbert spaces and the L p spaces, 1<p<. In fact, L p spaces are 2-uniformly convex if 1<p2, and p-uniformly convex if p>2. Remark. Let d p =inf { 1 2 &x& p + 12 &y& p && 12 (x+ y)& p , x, y # E, &x& y&>0 . (4.3) &x& y& p = Then it follows from [2] or [11] that d p >0 if and only if E is p-uniformly convex. Theorem 6. Let C(X, Y ) be p-uniformly convex, let S be a convex subset of C(X, Y ), and let F=(, 1 , ..., , l ) # C(X, Y ) l, with d(F, C)<d(F, S). Then f *=(,*, ..., ,*) # P C (F) is a strongly unique best simultaneous approximation of order p to F, that is, there exists # p >0 such that &F& f & p & F& f *& p +# p &,&,*& Ap , Proof. for all f =(,, ..., ,), , # S. Let the stated conditions hold, and let # p( f )= &F& f & p &&F& f *& p . &,&,*& Ap Then it is sufficient to prove that inf [# p( f ) : f =(,, ..., ,), , # S, ,{,*]>0. Without loss of generality, suppose that C(X, Y ) is complete. Suppose also that there exists a sequence [ f n ], with f n =( n , ..., n ), n # S, n {,* such that # p( f n ) 0, as n . (4.4) We will show that this leads to a contradiction. Now # p( f n )= &F& f n & p &&F& f *& p & n &,*& Ap | & f n & f *&&&F& f *& | & n &,*& A p &F& f *& & n &,*& A p _ & _ & &F& f *& &F& f *& max : a & & _ _& &,*& & . & &,*& & l a#U & p p i i=1 n A n File: DISTIL 310213 . By:DS . Date:26:11:97 . Time:15:31 LOP8M. V8.0. Page 01:01 Codes: 2437 Signs: 1134 . Length: 45 pic 0 pts, 190 mm A ON BEST SIMULTANEOUS APPROXIMATION 345 Thus if [ n ] is unbounded, it follows that & n &,*& A , as n . This in turn implies that l lim # p( f n )max : a i >0, n a#U i=1 which contradicts (4.4). Thus [ n ] is bounded, and so &F& f n & &F& f *&, n , as by definition of # p( f ). Thus there exists f =(,, ..., , ), , # clo(S), where clo(S) denotes the closure of S, and a subsequence of [ f n ] (which we do not rename) such that n , weakly. Note that C(X, Y ) is reflexive by the assumption of p-uniform convexity [4]. Now let a=(a 1 , ..., a l ) # U, w # W such that l : a i( w, , i &,*) =&F& f *&, i=1 l : a i( w, ,*&,) 0, for all , # S, (4.5) i=1 using Theorem 1. Since d(F, C)<d(F, S), by Lemma 1 li=1 a i {0. Then lim n " l l : a i (, i & n )+ : a i (, i &,*) i=1 i=1 l " A l w, : a (, & )+ : a (, &,*) = lim w, : a (, & ) +&F& f *& = w, : a (, &, ) +&F& f *&, using (4.5), w, : a (, &,*) +&F& f *&, lim i n i n i=1 i i i=1 l i n i n i=1 l i i i i i=1 l i=1 =2 &F& f *&. File: DISTIL 310214 . By:DS . Date:26:11:97 . Time:15:31 LOP8M. V8.0. Page 01:01 Codes: 2126 Signs: 825 . Length: 45 pic 0 pts, 190 mm 346 LI AND WATSON Also lim n " l : a i (, i & n ) " lim &F& f &=&F& f *&, A i=1 n n and l " : a i (, i &,*) i=1 " =&F& f *&, A and so " l l : a i (, i &, n )& : a i (, i &,*) i=1 i=1 " 0, as n , A using the p-uniform convexity property of C(X, Y ). Since li=1 a i {0, it follows that & f n & f *& 0, n . as Now let the sequence [a n =(a n1 , ..., a nl )] # U be such that " l " : a ni(, i & 12 ( n +,*)) i=1 =&F& 12 ( f n + f *)&. (4.6) A Then &F& 12 ( f n + f *)& 12 " l : a ni(, i & n ) i=1 1 2 " + 12 A " l : a ni(, i &,*) i=1 " A 1 2 &F& f n &+ &F& f *&. Also lim &F& 12 ( f n + f *)&= 12 lim (&F& f n &+&F& f *&)=&F& f *&. n n Thus lim n " l " : a ni(, i & n ) i=1 =&F& f *&. A Since li=1 a ni =0 implies that 12 ( f n + f *) # P C (F ), then d(F, S)&F& 12 ( f n + f *)&=d(F, C), File: DISTIL 310215 . By:DS . Date:26:11:97 . Time:15:31 LOP8M. V8.0. Page 01:01 Codes: 1987 Signs: 589 . Length: 45 pic 0 pts, 190 mm (4.7) 347 ON BEST SIMULTANEOUS APPROXIMATION which contradicts the assumption that d(F, C)<d(F, S). Thus li=1 a ni {0 for any n1. Let } l } *=inf : a ni . n i=1 (4.8) Assume *=0. Then, going to a subsequence if necessary, we must have a ni a 0i , i=1, ..., l with li=1 a 0i =0. Since & f n & f *& 0, " l : a 0i (, i &,*) i=1 " " = lim n A l : a ni(, i & n ) i=1 " A =&F& f *&, using (4.7). This implies that d(F, C)=d(F, S), which is a contradiction. Thus *>0. Now for any n1, &F& f *& p &F& 12 ( f n + f *)& p l l p "_ &" , " : a (, & )" + " : a (, &,*)" &d : a ( &,*) " " 1 2 = : a ni(, i & n )+ : a ni(, i &,*) i=1 l l n i 1 2 i n i 1 2 n A i=1 p i i=1 l by (4.6), A i=1 A p n i p n A i=1 p &F& f n & + &F& f *& p &* p d p & n &,*& Ap , 1 2 1 2 using the above remark and (4.8). Hence &F& f *& p &F& f n & 2 &2* p d p & n &,*& Ap , and # p( f n )2* p d p >0. This is a contradiction and the theorem is proved. K It is easy to give examples of spaces which satisfy the conditions of this Theorem. For example, let Y be a p-uniformly convex Banach space, and let X be a measure space. Then C(X, Y ) is p-uniformly convex with the norm &,& A = {| X &,(t)& Yp = 1 p , 1p<. File: DISTIL 310216 . By:DS . Date:26:11:97 . Time:15:31 LOP8M. V8.0. Page 01:01 Codes: 2304 Signs: 968 . Length: 45 pic 0 pts, 190 mm 348 LI AND WATSON ACKNOWLEDGMENT The work of the first author was supported by the National Nature Science Foundation of China. REFERENCES 1. D. A. Ault, F. R. Deutsch, P. D. Morris, and J. E. Olsen, Interpolating subspaces in approximation theory, J. Approx. Theory 3 (1970), 164182. 2. B. Beauzamy, ``Introduction to Banach Spaces and their Geometry,'' North-Holland, Amsterdam, 1985. 3. D. Braess, ``Nonlinear Approximation Theory,'' Springer-Verlag, Berlin, 1986. 4. J. Lindenstrauss and L. Tzafriri, ``Classical Banach Spaces II,'' Springer-Verlag, Berlin, 1977. 5. G. Nurnberger, ``Approximation by Spline Functions,'' Springer-Verlag, Berlin, 1989. 6. A. Pinkus, Uniqueness in vector-valued approximation, J. Approx. Theory 73 (1993), 1792. 7. I. Singer, ``Best Approximations in Normed Linear Spaces by Elements of Linear Subspaces,'' Springer-Verlag, Berlin, 1970. 8. S. Tanimoto, A characterization of best simultaneous approximations, J. Approx. Theory 59 (1989), 359361. 9. G. A. Watson, A characterization of best simultaneous approximations, J. Approx. Theory 75 (1993), 175182. 10. D. E. Wulbert, Uniqueness and differential characterization of approximations from manifolds, Amer. J. Math. 18 (1971), 350366. 11. H. K. Xu, Inequalities in Banach spaces with applications, Nonlinear Anal. Methods Appl. 16 (1991), 11271138. File: DISTIL 310217 . By:DS . Date:26:11:97 . Time:15:32 LOP8M. V8.0. Page 01:01 Codes: 3758 Signs: 1335 . Length: 45 pic 0 pts, 190 mm