FARTHEST POINTS AND THE FARTHEST DISTANCE MAP

PRADIPTA BANDYOPADHYAY AND S DUTTA

Abstract. In this paper, we consider farthest points and the farthest distance map of a closed bounded set in a Banach space. We show, inter alia, that a strictly convex Banach space has the Mazur-like intersection property for weakly compact sets if and only if every such set is the closed convex hull of its farthest points, and recapture a classical result of Lau in a broader set-up. We obtain an expression for the subdierential of the farthest distance map in the spirit of Preiss Theorem which in turn extends a recent result of Westphal and Schwartz, showing that the subdierential of the farthest distance map is the unique maximal monotone extension of a densely dened monotone operator involving the duality map and the farthest point map.

1. Introduction We work with real scalars. The closed unit ball and the unit sphere of a Banach space X will be denoted by B(X) and S(X) respectively. Our notations are otherwise standard. Any unexplained terminology can be found in [3]. For a closed and bounded set K in a Banach space X, the farthest distance map rK is dened as rK (x) = sup{ z x : z K}, x X. For x X, we dene the farthest point map as QK (x) = {z K : z x = rK (x)}, i.e., the set of points of K farthest from x. Note that this set may be empty. Let D(K) = {x X : QK (x) = }. The set of farthest points of K will be denoted by far(K), i.e., far(K) = {QK (x) : x D(K)}. Call a closed and bounded set K densely remotal if D(K) is norm dense in X. We say that a Banach space X has the Mazur Intersection Property (MIP) if every closed bounded convex set in X is the intersection of closed balls containing it. The MIP is a well studied notion in geometry of Banach space and several authors have studied Mazur-like intersection properties for dierent families of closed bounded convex sets. See [1, 2] and references thereof for a survey and unied treatments. However, no complete characterization is available, in particular, for every weakly compact convex set in X to be intersection of balls.
2000 Mathematics Subject Classication. Primary 46B20; Secondary 47H05. Key words and phrases. Mazur Intersection Property, farthest points, densely remotal sets, farthest distance map, LUR, subdierential.
1

BANDYOPADHYAY AND DUTTA

Lau [7, Theorem 3.3] had shown that a reexive Banach space X has the MIP if and only if every closed bounded convex set in X is the closed convex hull of its farthest points. In Section 2, we show that in a strictly convex Banach space X, every weakly compact convex set is intersection of balls if and only if every such set is the closed convex hull of its farthest points. Similar conclusions hold for compact convex sets, compact convex sets of nite ane dimension. And if X has the Radon-Nikodm Property (RNP), then similar result holds for w*-compact y convex sets in X . For a closed and bounded set K X, x X and > 0, a crescent of K determined by x and is the set C(K, x, ) = {z K : z x > rK (x) }. The simple but crucial observation in proving the main result in our next section is that for each of the above mentioned families of sets, every crescent of such a set contains a farthest point of the set. This also gives Laus result in [7, Theorem 3.3] as an obvious corollary. Recall that the subdierential of a convex function : X R at x X is (x) = {x X : x (y x) (y) (x) for all y X}.
1 The subdierential of the function (x) = 2 x 2 is referred to as the duality map on X and is denoted by D. Since rK is a continuous convex function, rK is a maximal monotone operator dened on X. In [9, Proposition 4.3], the authors showed that if X is a reexive Banach space with X Frchet smooth, then for a closed bounded set K, rK is e the unique maximal monotone extension of D IQK and for each x X, rK

(1)

rK (x) =
>0

co{D

I QK (y) : y x < }. rK

Note that this is actually the Preiss Theorem (see [8]) for rK . For a nonreexive space, such a statement needs qualication as D IQK (y) may rK be empty for some y. Nonetheless, even in nonreexive spaces, for a densely remotal set K (e.g., K weakly compact), D IQK is a well-dened monotone operator with rK dense domain. We show that if X is LUR, then rK remains the unique maximal monotone extension of D IQK and an analogue of (1) is available where we need to rK take the w*-closure and choose y from the set D1 (K) dened below. We believe this provides with the only version of the Preiss Theorem for rK is the non reexive case. Let x X and k QK (x). We say x D1 (K) if k is contained in crescents of K, determined by x of arbitrarily small diameter. It is easy to note that if x D1 (K), then QK (x) is nevessarily singleton.

FARTHEST POINTS AND THE FARTHEST DISTANCE MAP

2. Intersection of Balls and Farthest Points Here is the main theorem of this section. As mentioned in the introduction, this, in particular, gives the only known characterization of when every weakly compact convex set in X is intersection of balls. Theorem 2.1. (a) If X is a strictly convex Banach space and C is one of the following families of sets, (i) K = {all compact convex sets in X}. (ii) F = {all compact convex sets in X with nite ane dimension}. (iii) W = {all weakly compact convex set in X}. then every K C is intersection of balls if and only if every K C is the closed convex hull of its farthest points. (b) If X has the RNP, then X has the w*-MIP if and only if every w*compact convex set in X is the w*-closed convex hull of its farthest points. (c) [7] If X is reexive, then X has the MIP if and only if every closed bounded convex set in X is the closed convex hull of its farthest points. Proof. (a) We give the proof for the family W of weakly compact sets. The same proof works in the other cases too. Necessity : Let K W and thus by [7, Theorem 2.3], K is densely remotal. We claim every crescent of K contains a farthest point of K. To see this, let C(K, x, ) be any crescent of K, Choose such that 0 < /2 and then take such that 0 < < 2. Since K is densely remotal, there exists y D(K) such that x y| < . Then C(K, y, ) C(K, x, ) and clearly QK (y) C(K, y, ). Now let L = co(far(K)). Suppose K \ L = . Then L is the intersection of balls containing it as well and hence there exists a crescent C of K disjoint from L. By above obsevation C far(K) = . But, of course, far(K) L. This proves the necessity. Suciency : Suppose there exists K W that is not intersection of balls. Let K = {B : B is a closed ball and K B}. Let x0 K \ K. Choose y0 K and 0 < < 1 such that z0 = x0 + (1 )y0 K. / Let K1 = co(K {z0 }). Then K1 W. We will show that far(K1 ) K, and hence, K1 = co(far(K1 )). Let x X. Then K {u X : u x rK (x)}. Note that rK (x) rK1 (x) rK (x) = rK (x). Clearly, z0 as well as any point of the form e v = z0 + (1 )z, (0, 1], z K,

(2)

BANDYOPADHYAY AND DUTTA

are not extreme points of K, and since X is strictly convex, they are not farthest points as well. Therefore, v x < rK (x). Thus, QK1 (x) K. Since x X was arbitrary, far(K1 ) K. (b) If X has the RNP, by [4, Proposition 3], each w*-compact set K X is densely remotal. Thus, necessity can be proved as in (a). For suciency, note that if there exists a w*-compact set K that is not intersection of balls, since K = >0 [K + B(X )], passing to some K + B(X ) if necessary, we may assume that K has nonempty interior. Now if we choose y0 int(K), then z0 and any point of the form (2) are interior points of K, and hence the result follows as before. The following observation is immediate from the above arguments. Proposition 2.2. Every closed bounded convex set in X is the closed convex hull of its farthest points if and only if (a) X has the MIP; and (b) for every closed bounded convex set K X and every closed ball B with K \ B = , K \ B contains a farthest point of K. Note that the proof of Theorem 2.1 shows that if K is densely remotal, then K satises (b) above. Following example shows that we cannot dispense with (b). Example 2.3. The space c0 has a strictly convex Frchet dierentiable renorming e [3, Theorem 7.1 (ii)] which, thus, has the MIP. However, since the unit ball of the usual norm on c0 lacks extreme points, it must lack farthest points in the new norm. Remark 2.4. This also shows that even if X has the RNP, there may exist a closed bounded convex set in X with far(K) = (Compare this with [4, Proposition 3]). Observe that since the bidual of a space with the MIP has the w*-MIP, Theorem 2.1(b) shows that every w*-compact convex set in , with the bidual of the above norm, is the w*-closed convex hull of its farthest points. Thus, there is a closed bounded convex set K c0 , such that no farthest point of the w*-closure of K in X comes from K. 3. The Farthest Distance Map We begin by collecting some simple properties of the set D1 (K). Recall that a sequence {zn } K is called a maximizing sequence for x if x zn rK (x). Proposition 3.1. Let K be closed bounded set in a Banach space X. (a) x D1 (K) if and only if any maximizing sequence for x converges. (b) If x D1 (K), then QK is single valued and continuous at x and QK (x) is a strongly exposed point of K.

FARTHEST POINTS AND THE FARTHEST DISTANCE MAP

(c) D1 (K) is a G in X. The following proposition shows that any discussion on D1 (K) naturally require some convexity conditions on the norm. Proposition 3.2. (a) A Banach space X is strictly convex if and only if for every compact set K and k far(K), there exists x D1 (K), such that QK (x) = {k}. (b) A Banach space X is LUR if and only if for every closed bounded set K and k far(K), there exists x D1 (K), such that QK (x) = {k}. Proof. (a) Let K be a compact set in a strictly convex Banach space X. Let k far(K). Then, k QK (x) for some x D(K). Let t > 1. Strict convexity of the norm shows that for y = k + t(x k), QK (y) = {k}. Now compactness shows that y D1 (K). Conversely, if X is not strictly convex, there exists x, y S(X) such that the line segment [x, y] S(X). Clearly, K = [x, y] is compact and K QK (0). But any point of the open segment (x, y) cannot be strongly exposed and therefore, cannot be in the set QK (D1 (K)). (b) Observe that S(X) QB(X) (0). For any x S(X) and any sequence {xn } B(X) that is maximizing for x, we have x + xn 2. So if X is LUR, xn x. Thus x D1 (B(X)) and x QB(X) (x). Now for a closed bounded set K X and k far(K), get x D(K) such that k QK (x), and apply this argument with suitable translation and scaling to the ball B[x, rK (x)]. To prove the converse, let K = B(X). Then, S(X) = far(B(X)). So by the hypothesis, it follows that every point in S(X) is a strongly exposed point of B(X), and therefore, X is strictly convex. Now let x0 S(X). By hypothesis, there exists x D1 (B(X)) such that QK (x) = {x0 }. Then x x0 = rB(X) (x) = 1 + x . By strict convexity, it follows that x = x0 for some R and | 1| = 1 + ||. Therefore, < 0. To show X is LUR, let {xn } B(X) be such that xn + x0 2. For each n consider the function on (0, 1), fn () = 1 xn + (1 )x0 Then for all (0, 1), fn () 0. And by triangle inequality, 2fn (1/2) fn () + fn (1 ) fn () 0 By assumption, fn (1/2) 0. Thus, for any (0, 1), fn () 0. In particular, putting = 1/(1 ), we get xn x0 (1 ), that is {xn } is a maximizing sequence for x = x0 . Hence, xn x0 . The following two lemmas are crucial in proving our main theorem of this section.

BANDYOPADHYAY AND DUTTA

Lemma 3.3. Suppose X is LUR and K X is densely remotal. Then D1 (K) is a dense G in X. Proof. By Proposition 3.1(d), it suces to show that D1 (K) is dense in X. Let x D(K). Get k QK (x). 0 < < 1. Let y = k + (1 + )(x k). Then, x y = rK (x). It is easy to see that rK (y) = (1 + )rK (x) and by strict convexity, k is a unique farthest point from y. We now claim y D1 (K). Let {zn } K be a maximizing sequence for y. That is, zn y (1 + )rK (x). Then, (zn x) + (k x) rK (x) (1 + ) Then yn = (zn x)/rK (x) B(X), y0 = (k x)/rK (x) S(X), and for = 1/(1 + ), we have yn + (1 )y0 1. Notice that since < 1, 1/2 < < 1. As in the proof of Proposition 3.2(b), let fn () = 1 yn +(1)y0 . By convexity of the norm, fn () (2 2)fn (1/2) 0 Since fn () 0, we have that fn (1/2) 0, that is, yn + y0 2. Since X is LUR, yn y0 and hence, zn k. Remark 3.4. It follows that for any weakly compact set K in a LUR Banach space, D1 (K) is a dense G in X. So our result is more general than [5, Corollary 2.8], where it is proved that if the norm on X is Frchet dierentiable, then for e any closed and bounded subset K X, D1 (K) is residual. Lemma 3.5. Let x D1 (K). Then, rK (x) = D IQK (x). rK Moreover, rK is Gteaux (resp. Frchet) dierentiable at x if and only if the a e norm is Gteaux (resp. Frchet) dierentiable at x QK (x). a e Proof. Let QK (x) = {k} and x D( rxk ). Then x (x k) = rK (x). For z X, K (x) x (z x) = x (z) x (k) rK (x) rK (z) rK (x). Thus x rK (x). Conversely, let x rK (x). Since D IQK (x) is a w*-closed convex subset of rK S(X ), it is enough to show that for any z S(X), there is an x D IQK (x) 0 rK such that x (z) x (z). 0 Let {kn } K be such that x + z/n kn > rK (x + z/n) 1/n2 . Then {kn } is a maximizing sequence for x, and hence, kn k. Now z z z z 1 x ( ) = x (x + ) x (x) rK (x + ) rK (x) < x + kn rK (x) + 2 . n n n n n Choose x D(x + z/n kn ). Then n z z z x ( ) = x (x + kn ) x (x kn ) x + kn rK (x). n n n n n n

FARTHEST POINTS AND THE FARTHEST DISTANCE MAP

Combining the two, we have x (z) x (z) + 1/n. Let x be a w*-cluster point of n 0 {x }. Since x + z/n kn converges to x k in norm, we have x D IQK (x) and n 0 rK x (z) x (z), as desired. 0 Thus, the norm is Gteaux dierentiable is at x k D( rxk ) is singleton a K (x) so is rK (x) rK is Gteaux dierentiable at x. a Now, let {x } = rK (x) = D( rxk ). For any R and z B(X), x (z) K (x) x + z k x k rK (x + z) rK (x). Therefore, x + z k x k rK (x + z) rK (x) x (z) x (z) . Thus, Frchet dierentiability of rK at x implies that of the norm at x k. e Conversely, let the norm be Frchet dierentiable at x k. Let xn x, e x rK (xn ) and x rK (x), then {x } B(X ) and since rK is Gteaux a n n IQK dierentiable at x, xn x in the w*-topology. Since x D rK (x), x is a w*-norm point of continuity of B(X ), and therefore, x x in norm. It follows n that rK is Frchet dierentiable at x. e Remark 3.6. [5, Theorem 3.2(a)] proves only the necessity part of this result. Our proof is also simpler. Combining Lemma 3.5 with Lemma 3.3, it follows that in a Banach space with smooth LUR norm, the farthest distance map rK of a densely remotal set K is Gteaux dierentiable on a dense G . a We now state the main theorem of this section. This gives an expression for rK in the spirit of Preiss Theorem [8]. Note that our result does not need smoothness of the norm and with (Frchet) smoothness, by Theorem 3.5, we get back Preiss e Theorem for rK . Theorem 3.7. Let K be such that D1 (K) is dense in X and x X. Then rK (x) =
>0

co {rk (y) : y D1 (K) and y x < }.

Proof. Let x RHS and > 0. Choose < /3. For z X, choose y D1 (K) and y rK (y) such that y x < and x (z x) < y (z x) + . Thus, x (z x) < y (z x) + = y (z y) + y (y x) + rK (z) rK (y) + 2

rK (z) rK (x) + 3 rK (z) rK (x) + . Since is arbitrary, we have x rK (x). Conversely, let x rK (x). As in Lemma 3.5, we will show given any z S(X) there is an x RHS such that x (z) x (z). 0 0 For each n, get yn D1 (K) such that x + z/n yn < 1/n2 . Then z z z 2 x ( ) rK (x + ) rK (x) rK (yn ) rK (yn ) + 2 . n n n n

BANDYOPADHYAY AND DUTTA

Let x rK (yn ). Let kn QK (yn ). Then n z z z x ( ) = x (yn kn ) x (yn kn ) rK (yn ) rK (yn ). n n n n n n Thus x (z) x (z) + 2/n. Let x be a w*-cluster point of x . Then x RHS n n 0 0 and x (z) x (z). 0 Combining the Lemma 3.3, Lemma 3.5 and Theorem 3.7 we obtain the following: Corollary 3.8. Suppose K is a densely remotal set in a LUR Banach space X. Then rK is the unique maximal monotone extension of the densely dened monotone operator D IQK and for each x X, we have, rK rK (x) =
>0

co {D

I QK (y) : y D1 (K) and y x < }. rK

Remark 3.9. In [9, Proposition 4.3] obtained the similar conclusion for reexive Banach spaces with X Frchet smooth. e We end this section with a result on range of rK . Compare this with [9, Theorem 4.2]. Theorem 3.10. Let X be a smooth (resp. Frchet smooth) Banach space. Let e K X be a closed and bounded set such that D1 (K) is dense in X, then the image of D1 (K) under rK is w*-dense (resp. norm dense) in S(X ). Proof. Let N A(X) denote the set of norm attaining functionals in S(X ). By Bishop-Phelps Theorem, N A(X) is norm dense in S(X ). Let x N A(X) and 0 x0 S(X) such that x (x0 ) = 1. By density of D1 (K), choose xn D1 (K) such 0 that xn nx0 < 1/n and let x rK (xn ). Then xn . Therefore, by [9, n Lemma 4.1], lim xn (xn / xn ) = lim x = 1. But since xn / xn x0 in norm, n x (x0 ) 1 as well. Thus, any w*-cluster point of {x } is in D(x0 ). Since the norm n n is smooth, this set is singleton. Hence, x x in w*-topology. n 0 Now, if the norm on X is Frchet smooth, then x chosen above is a w*-norm e 0 point of continuity of B(X ). Thus rK (D1 (K)) is norm dense in S(X ). References
[1] P. Bandyopadhyay, The Mazur Intersection Property for Families of Closed Bounded Convex Sets in Banach Spaces, Colloq. Math. LXIII (1992), 4556. [2] D. Chen and B. L. Lin, Ball separation properties in Banach spaces, Rocky Mountain J. Math., 28 (1998), 835-873. [3] R. Deville, G. Godefroy, and V. Zizler, Smoothness and renormings in Banach spaces, Pitman Monographs and Surveys in Pure and Applied Mathematics, 64. Longman Scientic & Technical, Harlow; 1993. [4] R. Deville and V. Zizler, Farthest Points in W -Compact Sets, Bull. Austral. Math. Soc. 38 (1988), 433439.

FARTHEST POINTS AND THE FARTHEST DISTANCE MAP

[5] S. Fitzpatrick, Metric Projections and the Dierentiability of Distance Functions, Bull. Austral. Math. Soc. 22 (1980), 291312. [6] J. R. Giles, D. A. Gregory and B. Sims, Characterisation of Normed Linear Spaces with the Mazurs Intersection Property, Bull. Austral. Math. Soc. 18 (1978), 105123. [7] K.-S. Lau, Farthest Points in Weakly Compact Sets, Israel J. Math. 22 (1975), 168174. [8] D. Preiss, Frchet Derivatives of Lipschitz functions, J. Funct. Anal. 91 (1990), 312345. e [9] U. Westphal and T. Schwartz, Farthest Points and Monotone Operators, Bull. Austral. Math. Soc. 58 (1998), 7592. (Pradipta Bandyopadhyay) StatMath Division, Indian Statistical Institute, 203, B. T. Road, Kolkata 700 108, India, E-mail : [email protected] (S Dutta) StatMath Division, Indian Statistical Institute, 203, B. T. Road, Kolkata 700 108, India, E-mail : sudipta [email protected]