approximative compactness and asplund property in banach function spaces and in orlicz–bochner...

19
J. Math. Anal. Appl. 421 (2015) 1377–1395 Contents lists available at ScienceDirect Journal of Mathematical Analysis and Applications www.elsevier.com/locate/jmaa Approximative compactness and Asplund property in Banach function spaces and in Orlicz–Bochner spaces in particular with application Shaoqiang Shang a,, Yunan Cui b a Department of Mathematics, Northeast Forestry University, Harbin 150040, PR China b Department of Mathematics, Harbin University of Science and Technology, Harbin 150080, PR China a r t i c l e i n f o a b s t r a c t Article history: Received 2 January 2014 Available online 29 July 2014 Submitted by G. Corach Keywords: Approximative compactness Asplund space Radon–Nikodym property Banach space Orlicz–Bochner function space In this paper it is shown that: (1) If every weak hyperplane of X is approximatively compact, then (a) X is an Asplund space; (b) X has the Radon–Nikodym property. (2) Criteria for approximative compactness of every weakly hyperplane of Orlicz– Bochner function spaces equipped with the Orlicz norm are given. (3) If X has a Fréchet differentiable norm, then (a) Orlicz–Bochner function spaces L 0 M (X ) have the Radon–Nikodym property if and only if M Δ 2 ; (b) Orlicz–Bochner function spaces E N (X) are Asplund spaces if and only if M Δ 2 . (4) We give an important application of approximative compactness to the theory of generalized inverses for operators between Banach spaces and Orlicz–Bochner function spaces. © 2014 Elsevier Inc. All rights reserved. 1. Introduction Let X be a Banach space and X the dual space of X. Denote by B(X) and S(X) the closed unit ball and the unit sphere of X. Let C X be a nonempty subset of X. Then the set-valued mapping P C : X C P C (x)= z C : x z = dist(x, C) = inf yC x y is called the metric projection operator from X onto C . A subset C of X is said to be proximinal if P C (x) = for all x X (see [2]). It is well known that X is reflexive if and only if each closed convex subset of X is proximinal (see [2]). Definition 1. A nonempty subset C of X is said to be approximatively compact if for any {y n } n=1 C and any x X satisfying x y n inf yC x y as n →∞, there exists a subsequence of {y n } n=1 Supported by Heilongjiang Provincial Department of Education Funds 12521070. * Corresponding author. E-mail addresses: [email protected] (S. Shang), [email protected] (Y. Cui). http://dx.doi.org/10.1016/j.jmaa.2014.07.054 0022-247X/© 2014 Elsevier Inc. All rights reserved.

Upload: yunan

Post on 16-Feb-2017

212 views

Category:

Documents


0 download

TRANSCRIPT

J. Math. Anal. Appl. 421 (2015) 1377–1395

Contents lists available at ScienceDirect

Journal of Mathematical Analysis and Applications

www.elsevier.com/locate/jmaa

Approximative compactness and Asplund property in Banach

function spaces and in Orlicz–Bochner spaces in particular with

application ✩

Shaoqiang Shang a,∗, Yunan Cui b

a Department of Mathematics, Northeast Forestry University, Harbin 150040, PR Chinab Department of Mathematics, Harbin University of Science and Technology, Harbin 150080, PR China

a r t i c l e i n f o a b s t r a c t

Article history:Received 2 January 2014Available online 29 July 2014Submitted by G. Corach

Keywords:Approximative compactnessAsplund spaceRadon–Nikodym propertyBanach spaceOrlicz–Bochner function space

In this paper it is shown that: (1) If every weak∗ hyperplane of X∗ is approximatively compact, then (a) X is an Asplund space; (b) X∗ has the Radon–Nikodym property. (2) Criteria for approximative compactness of every weakly∗ hyperplane of Orlicz–Bochner function spaces equipped with the Orlicz norm are given. (3) If X has a Fréchet differentiable norm, then (a) Orlicz–Bochner function spaces L0

M (X∗) have the Radon–Nikodym property if and only if M ∈ Δ2; (b) Orlicz–Bochner function spaces EN (X) are Asplund spaces if and only if M ∈ Δ2. (4) We give an important application of approximative compactness to the theory of generalized inverses for operators between Banach spaces and Orlicz–Bochner function spaces.

© 2014 Elsevier Inc. All rights reserved.

1. Introduction

Let X be a Banach space and X∗ the dual space of X. Denote by B(X) and S(X) the closed unit ball and the unit sphere of X. Let C ⊂ X be a nonempty subset of X. Then the set-valued mapping PC : X → C

PC(x) ={z ∈ C : ‖x− z‖ = dist(x,C) = inf

y∈C‖x− y‖

}

is called the metric projection operator from X onto C.A subset C of X is said to be proximinal if PC(x) �= ∅ for all x ∈ X (see [2]). It is well known that X is

reflexive if and only if each closed convex subset of X is proximinal (see [2]).

Definition 1. A nonempty subset C of X is said to be approximatively compact if for any {yn}∞n=1 ⊂ C

and any x ∈ X satisfying ‖x− yn‖ → infy∈C ‖x− y‖ as n → ∞, there exists a subsequence of {yn}∞n=1

✩ Supported by Heilongjiang Provincial Department of Education Funds 12521070.* Corresponding author.

E-mail addresses: [email protected] (S. Shang), [email protected] (Y. Cui).

http://dx.doi.org/10.1016/j.jmaa.2014.07.0540022-247X/© 2014 Elsevier Inc. All rights reserved.

1378 S. Shang, Y. Cui / J. Math. Anal. Appl. 421 (2015) 1377–1395

converging to an element in C. X is called approximatively compact if every nonempty closed convex subset of X is approximatively compact.

Let us present the history of the approximative compactness and related notions. This notion has been introduced by Jefimow and Stechkin in [13] as a property of Banach spaces, which guarantees the existence of the best approximation element in a nonempty closed convex set C for any x ∈ X. In [16] it was established that the approximative compactness of C in a rotund Banach space X gives also the continuity of the projector operator PC . Fully rotund Banach spaces were defined by Fan and Glicksberg in [6]. In [9] it has been proved that fully rotund Banach spaces are approximatively compact. In [2] it has been proved that approximative compactness of a Banach space X means that X is reflexive and it has the H property.

Some criteria for approximative compactness of every weak∗ hyperplane of Orlicz function spaces equipped with the Luxemburg norm are given (see [22]). Moreover, it is well known that Orlicz func-tion spaces has the Radon–Nikodym property if and only if M ∈ Δ2. However, because of the complicated structure of Orlicz–Bochner function spaces, up to now the criteria for approximative compactness of every weak∗ hyperplane and Radon–Nikodym have not been discussed yet. The paper is organized as follows. In Section 1 some necessary definitions and notations are collected including the Orlicz–Bochner function spaces. In Section 2 we prove that if every weak∗ hyperplane of X∗ is approximatively compact, then (a) X is an Asplund space. (b) X∗ has the Radon–Nikodym property. In Section 3 criteria for approximative compactness of every weakly∗ hyperplane of Orlicz–Bochner function spaces equipped with the Orlicz norm are given. Moreover, we also prove that if X has a Fréchet differentiable norm, then (a) Orlicz–Bochner function spaces L0

M (X∗) have the Radon–Nikodym property if and only if M ∈ Δ2; (b) Orlicz–Bochner function spaces EN (X) are Asplund spaces if and only if M ∈ Δ2. In Section 4 we give an important ap-plication of approximative compactness to the theory of generalized inverses for operators between Banach spaces and Orlicz–Bochner function spaces. The topic of this paper is related to the topic of [6,13,16] and [1–5,7,8,10,11,17–22].

Let us recall some geometrical notions which will be used in the further part of the paper.

Definition 2. Let D be a nonempty open convex subset of X and f a real-valued continuous convex function on D. The function f is said to be Gateaux differentiable at the point x in D if the limit

f ′(x)(y) = limt→0

f(x + ty) − f(x)t

(∗)

exists for all y ∈ X. When this is the case, the limit is a continuous linear function of y, denoted by f ′(x). If the difference quotient in (∗) converges to f ′(x)(y) uniformly with respect to y in the unit ball, then f is said to be Fréchet differentiable at x. X is called a weak Asplund space [Asplund space] or said to have the weak Asplund property if for every f and D as above, f is “generically” Gateaux [Fréchet] differentiable, that is, there exists a dense Gδ subset G of D such that f is Gateaux [Fréchet] differentiable at each point of G.

Definition 3. A Banach space X is said to have a Fréchet differentiable norm if the norm ‖ · ‖ is Fréchet differentiable at every point of S(X).

It is well known (see [14]) that a norm ‖ · ‖ in a Banach space X is Fréchet differentiable at every point of S(X) if and only if for any x ∈ S(X), if x∗

n ∈ S(X∗) and x∗n(x) → 1, then there exists an x∗ ∈ S(X∗)

such that ‖x∗n − x∗‖ → 0 as n → ∞.

Definition 4. A Banach space X is said to have the Radon–Nikodym property whenever if (T, Σ, μ) is a nonatomic measure space and v is a vector measure on Σ with values in X which is absolutely continuous with respect to μ and has a bounded variation, then there exists f ∈ L1(X) such that for any A ∈ Σ,

S. Shang, Y. Cui / J. Math. Anal. Appl. 421 (2015) 1377–1395 1379

v(A) =∫A

f(t)dt.

It is well known that if X has a Fréchet differentiable norm, then X is an Asplund space.

Definition 5. A point x ∈ A is said to be an extreme point of A if 2x = y + z and y, z ∈ A imply y = z. The set of all extreme points of A is denoted by ExtA.

Definition 6. A functional fx ∈ S(X∗) is called a supporting functional of x ∈ S(X) if fx(x) = 1. A point x ∈ S(X) is called a smooth point if it has a unique supporting functional fx. If every x ∈ S(X) is a smooth point, then X is called a smooth space.

Definition 7. A function M : R → R+ is called an N -function if it has the following properties:

(1) M is even, continuous convex and M(0) = 0;(2) M(u) > 0 for any u �= 0;(3) limu→0 M(u)u = 0 and limu→∞ M(u)u = ∞.

Let (T, Σ, μ) be a nonatomic finite measurable space. Moreover, given any Banach space (X, ‖ · ‖), we denote by XT the set of all strongly μ-measurable function from T to X, and given any N -function M for each u ∈ XT , we define the modular of u by

ρM (u) =∫T

M(∥∥u(t)

∥∥)dt.Put

LM (X) ={u(t) ∈ XT : ρM (λu) < ∞ for some λ > 0

},

EM (X) ={u(t) ∈ XT : ρM (λu) < ∞ for all λ > 0

}.

It is well known that the Orlicz–Bochner function space LM (X) is a Banach space, when it is equipped with the Luxemburg norm

‖u‖ = inf{λ > 0 : ρM

(u

λ

)≤ 1

}

or with the Orlicz norm

‖u‖0 = infk>0

1k

[1 + ρM (ku)

].

If X = R, then L0M is said to be Orlicz function space. Let p(u) denote the right derivative of M(u) at

u ∈ R+, q(v) be the generalized inverse function of p(u) defined on R+ by

q(v) = supu≥0

{u ≥ 0 : p(u) ≤ v

}

Then we call N(v) =∫ |v|0 q(s)ds the complementary function of M . It is well known that there holds the

Young inequality uv ≤ M(u) + N(v) and uv = M(u) + N(v) ⇔ u = |q(v)| sign v or v = |p(u)| sign u. Moreover, it is well known that M and N are complementary to each other. It is well known that K(u) �= ∅, where

1380 S. Shang, Y. Cui / J. Math. Anal. Appl. 421 (2015) 1377–1395

K(u) ={k > 0 : 1

k

[1 + ρM (ku)

]= ‖u‖0

}.

For every N -function M we define the complementary function N : R → [0, ∞) by the formula

N(v) = supu>0

{u|v| −M(u)

}

for every v ∈ R. We say that an N -function M satisfies condition Δ2 if there exist K > 2 and u0 ≥ 0 such that

M(2u) ≤ KM(u) (u ≥ u0)

In this case, we write M ∈ Δ2 or N ∈ ∇2. Moreover, it is well known that M ∈ Δ2 if and only if for any l > 1, there exist v0 > 0 and δ > 0 such that N(lv) ≥ (l + δ)N(v) whenever v ≥ v0.

Let M be an N -function. An interval [a, b] is called a structural affine interval of M , or simply, SAI of M , provided that M is affine on [a, b] and it is not affine on either in [a − ε, b] or in [a, b + ε] for any ε > 0. Let {[ai, bi]}i be all the SAIs of M . We call SM = R\[

⋃i(ai, bi)] the set of strictly convex points of M .

A continuous function M : R → R is called strictly convex if

M

(u + v

2

)<

12M(u) + 1

2M(v)

for all u �= v.Let us recall some lemmas which will be used in the further part of the paper.

Lemma 1. (See [22].) The following statements are equivalent:

(1) Every weak∗ hyperplane of X∗ is approximatively compact.(2) If x∗

n ∈ S(X∗), x ∈ S(X) and x∗n(x) → 1 as n → ∞, then {x∗

n}∞n=1 is relatively compact.

Lemma 2. (See [22].) X is a smooth space if and only if any x∗ ∈ S(X∗) is an extreme point of B(X∗)provided that x∗ is norm attainable on S(X).

Lemma 3. (See [17].) Suppose that X is a Banach space. Then the following statements are equivalent:

(1) X is an Asplund space.(2) If E is a separable closed subspace of X, then E∗ is a separable space.(3) X∗ has the Radon–Nikodym property.

Lemma 4. (See [7].) Suppose that X∗ has the Radon–Nikodym property. Then (EM (X))∗ = L0N (X∗) and

(E0M (X))∗ = LN (X∗).

Lemma 5. (See [4].) Let A(x) = {x∗ ∈ S(X∗) : x∗(x) = ‖x‖ = 1}, B ⊂ S(X∗) and let B ∩ A(x) be a nonempty set for all x ∈ S(X). Assume that

B ⊂ co(F + αB

(X∗))

for some countable set F ⊂ X∗ and some α < 1. Then X∗ is the closed linear span of F .

S. Shang, Y. Cui / J. Math. Anal. Appl. 421 (2015) 1377–1395 1381

2. Approximative compactness and Asplund property in Banach spaces

Theorem 1. Suppose that every weak∗ hyperplane of X∗ is approximatively compact. Then (1) X is an Asplund space. (2) X∗ has the Radon–Nikodym property.

Proof. Let E be a separable closed subspace of X. We will prove that if x∗n ∈ S(E∗), x ∈ S(E) and

x∗n(x) → 1 as n → ∞, then sequence {x∗

n}∞n=1 is relatively compact. By the Hahn–Banach theorem, there exists y∗n ∈ S(X∗) such that y∗n ∈ S(X∗) and y∗n(z) = x∗

n(z) whenever z ∈ E. Since y∗n ∈ S(X∗), x ∈ S(X)and y∗n(x) → 1 as n → ∞, by Lemma 2, we obtain that the sequence {y∗n}∞n=1 is relatively compact. Hence we obtain that the sequence {x∗

n}∞n=1 is relatively compact. Let A(x) = {x∗ ∈ S(E∗) : ‖x‖ = x∗(x) = 1, x ∈S(E)} and {xn}∞n=1 be a dense subset of S(E). Then for any x ∈ S(E), we have

inf{dist

(A(xn), A(x)

): n ∈ N

}= 0. (1)

In fact, since {xn}∞n=1 is a dense subset of S(E), there exists a subsequence {xnk}∞k=1 of {xn}∞n=1 such that

xnk→ x as k → ∞. Pick x∗

nk∈ A(xnk

). Then

∣∣x∗nk

(x) − 1∣∣ =

∣∣x∗nk

(x) − x∗nk

(xnk)∣∣ ≤ ∥∥x∗

nk

∥∥‖xnk− x‖ → 0 as k → ∞.

From the previous proof, we obtain that {x∗nk}∞k=1 is relatively compact. Hence there exists a subsequence

{x∗nl}∞l=1 of {x∗

nk}∞k=1 such that {x∗

nl}∞l=1 is a Cauchy sequence. Let x∗

nl→ x∗

0 as l → ∞. Noticing that x∗

nk(x) → 1, we have x∗

nl(x) → x∗

0(x) = 1 as l → ∞. This implies that x∗0 ∈ A(x). Hence

inf{dist(A(xn), A(x)) : n ∈ N} = 0.Let {z∗k}∞k=1 ⊂ A(x) for any x ∈ S(E). Then z∗k(x) = 1. From the previous proof, we obtain that {z∗k}∞k=1

has a Cauchy subsequence. Moreover, it is easy to see that A(x) is a closed set. Hence we obtain that A(x)is compact. Then A(x) is separable. Let Dn be a countable dense subset of A(xn) for each n ∈ N . Then D =

⋃∞n=1 Dn is a countable set. Pick α ∈ (0, 1). Therefore, by (1), we obtain that there exist an α(x) ∈ A(x)

and n(x) ∈ N such that dist(A(xn(x)), α(x)) < α for any x ∈ S(E). Hence, for every x ∈ S(E), we obtain that B = {α(x) : x ∈ S(E)} ⊂ S(E∗) and B ∩ A(x) �= ∅ for any x ∈ S(E). Since dist(A(xn(x)), α(x)) < α

and Dn(x) is a countable dense subset of A(xn(x)), we obtain that

dist(α(x), D

)≤ dist

(α(x), Dn(x)

)= dist

(α(x), A(xn(x))

)< α

for any x ∈ S(E). This implies that

α(x) ∈ D + αB(E∗) ⊂ co

(D + αB

(E∗))

for any x ∈ S(E). Noticing that B = {α(x) : x ∈ S(E)}, we have

B ⊂ D + αB(E∗) ⊂ co

(D + αB

(E∗)).

By Lemma 5, we obtain that E∗ = span(D), and so E∗ is separable.This means that if E is a separable closed subspace of X, then E∗ is a separable space. By Lemma 3,

we obtain that X is an Asplund space and X∗ has the Radon–Nikodym property, which completes the proof. �Corollary 1. Suppose that X has a Fréchet differentiable norm. Then (1) X is an Asplund space. (2) X∗

has the Radon–Nikodym property.

1382 S. Shang, Y. Cui / J. Math. Anal. Appl. 421 (2015) 1377–1395

3. Approximative compactness and Asplund property in Orlicz–Bochner function spaces

Theorem 2. The following statements are equivalent:

(1) EN (X) has a Fréchet differentiable norm.(2) Every weak∗ hyperplane of L0

M (X∗) is approximatively compact.(3) (3a) M ∈ Δ2;

(3b) M(u) is strictly convex;(3c) X has a Fréchet differentiable norm.

In order to prove the theorem, we will give some auxiliary lemmas.

Lemma 6. (See [1].) L0M has the Radon–Nikodym property if and only if M ∈ Δ2.

Lemma 7. (See [1].) u ∈ L0M is an extreme point of B(L0

M ) if and only if μ{t ∈ T : ku(t) ∈ R\SM} = 0, where k ∈ K(u).

Lemma 8. The following statements are equivalent:

(a) u ∈ L0M is an extreme point of B(L0

M );(b) If u =

∑∞i=1 tiui where ui ∈ B(L0

M ), ti ∈ (0, 1) and ∑∞

i=1 ti = 1 belongs to S(L0M ), then the sequence

{ui}∞i=1 is relatively compact.

Proof. (b) ⇒ (a). Suppose that u ∈ L0M is not an extreme point of B(L0

M ). Then, by Lemma 7, we have μ{t ∈ T : ku(t) ∈ R\SM} > 0, where k ∈ K(u). Hence there exist an interval (a, b) and ε > 0 such that

μ{t ∈ T : ku(t) ∈ (a + ε, b− ε)

}> 0

and M is affine on [a, b]. Hence there exist p ∈ R and β ∈ R such that M(h) = ph + β for h ∈ [a, b]. Let G = {t ∈ T : ku(t) ∈ (a + ε, b − ε)}. Then it is easy to see that there exist δ > 0 and n0 ∈ N such that

μ

{t ∈ G :

(1 − 1

n0

)ku(t) ∈ (a + δ, b− δ)

}> 0.

Let H = {t ∈ G : (1 − 1/n0)ku(t) ∈ (a + δ, b − δ)}. Decompose H into E11 and E1

2 such that E11 ∪ E1

2 = H, E1

1 ∩E12 = ∅ and

∫E1

1p( k

n0u(t))dt =

∫E1

2p( k

n0u(t))dt. Decompose E1

1 into E21 and E2

2 such that E21 ∩E2

2 = ∅, E2

1 ∪E22 = E1

1 and ∫E2

1p( k

n0u(t))dt =

∫E2

2p( k

n0u(t))dt. Decompose E1

2 into E23 and E2

4 such that E23∩E2

4 = ∅, E2

3 ∪E24 = E1

2 and ∫E2

3p( k

n0u(t))dt =

∫E2

4p( k

n0u(t))dt. Generally, decompose En−1

i into En2i−1 and En

2i such that

En2i−1 ∩En

2i = ∅, En2i−1 ∪ En

2i = En−1i ,

∫E2

2i−1

p

(k

n0u(t)

)dt =

∫E2

2i

p

(k

n0u(t)

)dt,

and

(n = 1, 2, ..., i = 1, 2, ..., 2n−1).

Define

S. Shang, Y. Cui / J. Math. Anal. Appl. 421 (2015) 1377–1395 1383

un(t) =

⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩

u(t), t ∈ T\H(1 − 1

n0)u(t), t ∈ En

1

(1 + 1n0

)u(t), t ∈ En2

... ...

(1 − 1n0

)u(t), t ∈ En2n−1

(1 + 1n0

)u(t), t ∈ En2n ,

u′n(t) =

⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩

u(t), t ∈ T\H(1 + 1

n0)u(t), t ∈ En

1

(1 − 1n0

)u(t), t ∈ En2

... ...

(1 + 1n0

)u(t), t ∈ En2n−1

(1 − 1n0

)u(t), t ∈ En2n ,

and

(yn(t)

)∞n=1 =

(u1(t), u′

1(t), u2(t), u′2(t), ..., un(t), u′

n(t), ...).

Then

‖un‖0 ≤ 1k

[1 + ρM (kun)

]= 1

k

[1 +

∫T\H

M(kun(t)

)dt +

∫En

1

[p(kun(t)

)+ β

]dt +

∫En

2

[p(kun(t)

)+ β

]dt

+ · · · +∫

En2n−1

[p(kun(t)

)+ β

]dt +

∫En

2n

[p(kun(t)

)+ β

]dt

]

= 1k

[1 +

∫T\H

M(ku(t)

)dt +

∫En

1

[p

((1 − 1

n0

)ku(t)

)+ β

]dt +

∫En

2

[p

((1 + 1

n0

)ku(t)

)+ β

]dt

+ · · · +∫

En2n−1

[p

((1 − 1

n0

)ku(t)

)+ β

]dt +

∫En

2n

[p

((1 + 1

n0

)ku(t)

)+ β

]dt

]

= 1k

[1 +

∫T\H

M(ku(t)

)dt +

∫En

1

[p(ku(t)

)+ β

]dt +

∫En

2

[p(ku(t)

)+ β

]dt

+ · · · +∫

En2n−1

[p(ku(t)

)+ β

]dt +

∫En

2n

[p(ku(t)

)+ β

]dt

]

= 1k

[1 +

∫T\H

M(ku(t)

)dt +

∫En

1

M(ku(t)

)dt +

∫En

2

M(ku(t)

)dt

+ · · · +∫

En2n−1

M(ku(t)

)dt +

∫En

2n

M(ku(t)

)dt

]

≤ 1k

[1 + ρM (ku)

]≤ 1.

Similarly, ‖u′n‖0 ≤ 1. Hence ‖yn‖0 ≤ 1. This implies that yn ∈ B(L0

M ). On the other hand, we have

∞∑n=1

(12 · 1

2nun(t) + 12 · 1

2nu′n(t)

)=

∞∑n=1

12n+1

(un(t) + u′

n(t))

=∞∑

n=1

22n+1u(t) = u(t),

and

1384 S. Shang, Y. Cui / J. Math. Anal. Appl. 421 (2015) 1377–1395

∞∑n=1

(12 · 1

2n + 12 · 1

2n

)=

∞∑n=1

(12n

)= 1.

But {yn}∞n=1 is not relatively compact. In fact, since M is a convex function, there exists l > 0 such that M(v) > lv whenever v ≥ a/(kn0). Let km,n ∈ K(um − un). Then [1 + ρM (km,n(um − un))]km,n =‖um − un‖0 ≤ 2. Hence 2km,n ≥ 1. This implies that if ku(t) ∈ G, then

km,n2n0

u(t) ≥ 12

2kkn0

u(t) = 1n0k

ku(t) ≥ a

n0k.

Let Tm,n = {t ∈ H : um(t) �= un(t)}. Then

‖un − um‖0 = 1kn,m

[1 + ρM

(kn,m(un − um)

)]≥

∫Tn,m

M(kn,m 2n0

u(t))kn,m

dt

≥∫

Tm,n

M(km,n2n0

u(t))km,n

2n0

u(t)·(

2n0

u(t))dt ≥

∫Tm,n

l2n0

u(t)dt =∫

Tm,n

2lk

k

n0u(t)dt

= 2lkp

∫Tm,n

p

(k

n0u(t)

)dt = l

kp

∫H

p

(k

n0u(t)

)dt > 0

for any m �= n. Hence we obtain that the sequence {yn}∞n=1 is not relatively compact, a contradiction!The implication (a) ⇒ (b) is obvious, which completes the proof. �

Proof of Theorem 2. Since every weak∗ hyperplane of L0M (X∗) is approximatively compact, by Theorem 1,

we obtain that L0M (X∗) has the Radon–Nikodym property. This implies that the subspace L0

M of L0M (X)

has the Radon–Nikodym property. By Lemma 6, we have M ∈ Δ2.Let u be norm attainable on S(EN ), i.e.,

∫Tu(t)v0(t)dt = 1 for some v0(t) ∈ S(EN ). If u =

∑∞i=1 tiui,

then

1 =∫T

u(t) · v0(t)dt =∫T

( ∞∑i=1

tiui(t))

· v0(t)dt =∞∑i=1

ti

∫T

ui(t) · v0(t)dt,

where ui ∈ B(L0M ), ti ∈ (0, 1) and

∑∞i=1 ti = 1. Hence for any i ∈ N , we have

∫Tui(t) · v0(t)dt = 1. This

implies that {ui}∞i=1 is relatively compact by Lemma 1. By virtue of Lemma 8, we obtain that u is an extreme point. By Lemma 2, we obtain that EN is smooth. Hence q(u) is continuous. This implies that M(u) is strictly convex.

We will prove that if x∗n ∈ S(X∗), x ∈ S(X) and limn→∞ x∗

n(x) = 1, then {x∗n}∞n=1 is relatively

compact. In fact, suppose that {x∗n}∞n=1 is not relatively compact. Then there exists r > 0 such that

r = infm�=n ‖x∗n − x∗

m‖ > 0. By the Hahn–Banach theorem, there exist v ∈ S(EN ) and u ∈ S(L0M ) such

that ∫Tu(t)v(t)dt = 1. Define

−→v(t) = x · v(t), −−→un(t) = x∗n · u(t), n = 1, 2, ...

Then −→v ∈ S(EN (X)), −−→un ∈ S(L0M (X∗)) and

∫T

(−−→un(t),−→v(t))dt → 1. Hence

1 ←∫ (−−→un(t),−→v(t)

)dt = 1

kn

∫ (kn

−−→un(t),−→v(t))dt

T T

S. Shang, Y. Cui / J. Math. Anal. Appl. 421 (2015) 1377–1395 1385

≤ 1kn

(∫T

M(kn

−−→un(t))dt +

∫T

N(−→v(t)

)dt

)

= 1kn

(∫T

M(kn

−−→un(t))dt + 1

)= ‖−−→un‖0 = 1,

where kn ∈ K(−−→un). However,

‖−−→un − −−→um‖0 = infk>0

[1k

(∫T

M(k(−−→un(t) − −−→um(t)

))dt + 1

)]

= infk>0

[1k

(∫T

M(kru(t)

)dt + 1

)]

= r‖u‖0 = r > 0,

a contradiction. This implies that if x∗n ∈ S(X∗), x ∈ S(X) and x∗

n(x) → 1 as n → ∞, then {x∗n}∞n=1 is

relatively compact.We next will prove that X is smooth. Suppose that X is not smooth. Then there exist x ∈ S(X),

y∗ ∈ S(X∗) and z∗ ∈ S(X∗) such that y∗ �= z∗ and y∗(x) = z∗(x) = 1. It is easy to see that there exist v ∈ S(EN ) and u ∈ S(L0

M ) such that u(t) > 0 and ∫Tu(t)v(t)dt = 1. Decompose T into T 1

1 , T 12 such that

T 11 ∩ T 1

2 = ∅, T 11 ∪ T 1

2 = T and ∫T 11M(‖y∗ − z∗‖u(t))dt =

∫T 12M(‖y∗ − z∗‖u(t))dt. Decompose T 1

1 into T 21 ,

T 22 such that T 2

1 ∩T 22 = ∅, T 2

1 ∪T 22 = T 1

1 and ∫T 21M(‖y∗ − z∗‖u(t))dt =

∫T 22M(‖y∗ − z∗‖u(t))dt. Decompose

T 12 into T 2

3 , T 24 such that T 2

3 ∩T 24 = ∅, T 2

3 ∪T 24 = T 1

2 and ∫T 23M(‖y∗ − z∗‖u(t))dt =

∫T 24M(‖y∗ − z∗‖u(t))dt.

Generally, decompose Tn−1i into Tn

2i−1 and Hn2i such that

Tn2i−1 ∩ Tn

2i = ∅, Tn2i−1 ∪ Tn

2i = Tn−1i and∫

Tn2i−1

M(∥∥y∗ − z∗

∥∥u(t))dt =

∫Tn2i

M(∥∥y∗ − z∗

∥∥u(t))dt,

where n = 1, 2, ...., i = 1, 2, ...., 2n−1. Set

−−→un(t) =

⎧⎪⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎪⎩

y∗ · u(t), t ∈ Tn1

z∗ · u(t), t ∈ Tn2

... ...

y∗ · u(t), t ∈ Tn2i−1

z∗ · u(t), t ∈ Tn2i,

−→v(t) = x · v(t), n = 1, 2, ...

It is easy to see that −−→un ∈ S(LM (X∗)), −→v ∈ S(EN (X)) and ∫T(−−→un(t), −→v(t))dt = 1. Since every weak∗

hyperplane of L0M (X∗) is approximatively compact, by Lemma 1, we obtain that {−−→un}∞n=1 is relatively

compact. Let Tm,n = {t ∈ T : −−→um(t) �= −−→un(t)}. Then

ρM (um − un) =∫T

M(∥∥−−→un(t) − −−→um(t)

∥∥)dt=

∫Tm,n

M(∥∥y∗ − z∗

∥∥u(t))dt

= 12

∫M

(∥∥y∗ − z∗∥∥u(t)

)dt > 0,

T

1386 S. Shang, Y. Cui / J. Math. Anal. Appl. 421 (2015) 1377–1395

for any m �= n. Therefore, by M ∈ Δ2, we obtain that there exists l > 0 such that ‖−−→un − −−→um‖0 ≥‖−−→un − −−→um‖ ≥ l, a contradiction. This implies that X is smooth. We have proved that if x∗

n ∈ S(X∗), x ∈ S(X) and limn→∞ x∗

n(x) = 1, then {x∗n}∞n=1 is relatively compact. This implies that x∗

n ∈ S(X∗), x ∈ S(X) and x∗

n(x) → 1 as n → ∞, then {x∗n}∞n=1 is a Cauchy sequence. Hence we obtain that X has a

Fréchet differentiable norm.(3) ⇒ (1). Let un ∈ S(L0

M (X∗)), v ∈ S(EN (X)) and limn→∞∫T

(un, v)dt = 1. By the Hahn–Banach theorem and Lemma 4, there exists u ∈ S(L0

M (X∗)) such that ∫T

(u(t), v(t))dt = 1. Hence we have ∫T

(un(t) + u(t), v(t))dt → 2 as n → ∞. The proof requires the consideration of few cases separately.Case 1. Let sup{kn} < ∞, where kn = K(un). We may assume without loss of generality that kn → l as

n → ∞. We claim that ‖un(t)‖ μ−→ ‖u(t)‖ in measure. Otherwise, we may assume without loss of generality that for each n ∈ N , there exists En ⊆ T , ε0 > 0 and σ0 > 0 such that μEn ≥ ε0, where

En ={t ∈ T :

∣∣∥∥un(t)∥∥−

∥∥u(t)∥∥∣∣ ≥ σ0

}.

We define the sets

An ={t ∈ T : M

(∥∥knun(t)∥∥) > 8d

ε0

}and B =

{t ∈ T : M

(∥∥ku(t)∥∥) > 8d

ε0

},

where k ∈ K(u) and d = max{k, sup{kn}}. Then

∥∥knun(t)∥∥ ≥ M−1

(8dε0

), t ∈ An and

∥∥ku(t)∥∥ ≥ M−1

(8dε0

), t ∈ B.

Noticing that ρM (knun) = kn − 1, we have

d ≥∫T

M(∥∥knun(t)

∥∥)dt ≥ ∫An

M(∥∥knun(t)

∥∥)dt ≥ 8dε0

μAn.

This implies that μAn ≤ ε0/8. Similarly, we have μB ≤ ε0/8. We define a bound closed set

C ={

(u, v) ∈ R2 : M(u) ≤ 8dε0

,M(v) ≤ 8dε0

, |u− v| ≥ 14σ0

}

in the two dimensional space R2. Since C is compact, there exists δ > 0 such that

M

(k

k + lu + l

k + lu

)≤ (1 − δ)

[k

k + lM(u) + l

k + lM(v)

]

for any (u, v) ∈ C. Then

M

(k

k + l

∥∥knun(t)∥∥ + l

k + l

∥∥ku(t)∥∥) ≤ (1 − δ)

[k

k + lM

(∥∥knun(t)∥∥) + l

k + lM

(∥∥ku(t)∥∥)]

for t ∈ En\(An ∪B). Let

Wn(t) =

⎧⎨⎩

M( kk+kn

‖knun(t)‖+ knk+kn

‖ku(t)‖)k

k+knM(‖knun(t)‖)+ kn

k+knM(‖ku(t)‖) , t ∈ En\(An ∪B)

0, t ∈ T\(En\(An ∪B))

and

S. Shang, Y. Cui / J. Math. Anal. Appl. 421 (2015) 1377–1395 1387

Qn(t) ={

M( kk+l‖knun(t)‖+ l

k+l‖ku(t)‖)k

k+lM(‖knun(t)‖)+ lk+lM(‖ku(t)‖) , t ∈ En\(An ∪B)

0, t ∈ T\(En\(An ∪B)).

It is easy to see that Wn(t) −Qn(t) → 0 for any t ∈ T . Therefore, by the Egorov theorem, there exist E ⊂ T

and a natural number N such that μE < ε0/16 and |Wn(t) −Qn(t)| < δ/4 whenever n > N and t ∈ T \E. Hence, if t ∈ (En\(An ∪B)) \ E, then

12δ = δ − 1

≤ 1 −M(t, k

k+l‖knun(t)‖ + lk+l‖ku(t)‖)

kk+lM(t, ‖knun(t)‖) + l

k+lM(t, ‖ku(t)‖)− 1

≤ 1 −M(t, k

k+kn‖knun(t)‖ + kn

k+kn‖ku(t)‖)

kk+kn

M(t, ‖knun(t)‖) + kn

k+knM(t, ‖ku(t)‖)

,

whenever n is large enough. This implies that

M

(k

k + kn

∥∥knun(t)∥∥ + kn

k + kn

∥∥ku(t)∥∥) ≤

(1 − 1

2δ)[

k

k + knM

(∥∥knun(t)∥∥) + kn

k + knM

(∥∥ku(t)∥∥)].

Let Hn = (En\(An ∪B)) \ E. Then 4μHn ≥ ε0 and

‖un‖0 + ‖u‖0 − ‖un + u‖0 ≥ 1kn

(1 + ρM (knun)

)+ 1

k

(1 + ρM (ku)

)− kn + k

knk

(1 + ρM

(knk

kn + k(un + u)

))

≥ kn + k

knk

∫Hn

[k

kn + kM

(∥∥knun(t)∥∥) + kn

kn + kM

(∥∥ku(t)∥∥)

−M

(k

kn + k

∥∥knun(t)∥∥ + kn

kn + k

∥∥ku(t)∥∥)]dt

≥ kn + k

knk

∫Hn

12δ

[k

kn + kM

(t,∥∥knun(t)

∥∥) + knkn + k

M(∥∥ku(t)

∥∥)]dt

≥ kn + k

knk

∫Hn

12δ

[M

(k

kn + k

∥∥knun(t)∥∥ + kn

kn + k

∥∥ku(t)∥∥)]dt

≥ kn + k

knk

∫Hn

12δ

[M

(kkn

kn + k

∣∣∥∥un(t)∥∥−

∥∥u(t)∥∥∣∣)]dt

≥ kn + k

knk

∫Hn

12δ

[M

(1

d + kσ0

)]dt

≥ 2dk

12δM

(1

d + kσ0

)μHn

≥ δε0

4dkM(

1d + k

σ0

)

whenever n is large enough. Since ∫T

(un(t) + u(t), v(t))dt → 2, we have ‖un + u‖0 → 2 as n → ∞. This implies that ‖un‖0 + ‖u‖0 − ‖un + u‖0 → 0 as n → ∞, a contradiction! Hence ‖un(t)‖ μ−→ ‖u(t)‖ in measure. By the Riesz theorem, there exists a subsequence {n} of {n} such that ‖un(t)‖ → ‖u(t)‖ μ-a.e. in T . Noticing that

1388 S. Shang, Y. Cui / J. Math. Anal. Appl. 421 (2015) 1377–1395

∣∣(un(t), v(t))∣∣ ≤ ∥∥un(t)

∥∥ ·∥∥v(t)∥∥, lim

n→∞

∫T

(un(t), v(t)

)dt = 1

and ∫T‖un(t)‖ · ‖v(t)‖dt ≤ ‖un‖0 · ‖v‖ ≤ 1, we obtain

limn→∞

∫T

∥∥un(t)∥∥ ·

∥∥v(t)∥∥dt = 1, limn→∞

∫T

[∥∥un(t)∥∥ ·

∥∥v(t)∥∥−(un(t), v(t)

)]dt = 0.

Moreover, it is easy to see that∫T

∣∣∥∥un(t)∥∥ ·

∥∥v(t)∥∥−(un(t), v(t)

)∣∣dt → 0 as n → ∞.

This implies that ‖un(t)‖ · ‖v(t)‖ − (un(t), v(t)) μ−→ 0 in measure. By the Riesz theorem, there exists a subsequence {n} of {n} such that ‖un(t)‖ · ‖v(t)‖ − (un(t), v(t)) → 0 μ-a.e. in T . By ‖un(t)‖ → ‖u(t)‖ μ-a.e. in T , it follows that (un(t), v(t)) → ‖u(t)‖ · ‖v(t)‖ μ-a.e. in T . We may assume without loss of generality that

limn→∞

(un(t)‖u(t)‖ ,

v(t)‖v(t)‖

)= 1

on {t ∈ T : ‖u(t)‖ · ‖v(t)‖ �= 0}. Moreover, μT1 = 0, where T1 = {t ∈ T : ‖v(t)‖ = 0} ∩ {t ∈ T : ‖u(t)‖ �= 0}. In fact, suppose that μT1 > 0. Then

‖u‖0 = 1k

[1 + ρM (ku)

]>

1k

[1 + ρM (kuχT\T1)

]≥ ‖uχT\T1‖0,

where k ∈ K(u). This implies that

1 =∫T

(u, v)dt =∫T

(uχT\T1 , v)dt ≤ ‖uχT\T1‖0 · ‖v‖ < 1,

a contradiction! Hence we may assume without loss of generality that

limn→∞

(un(t)‖u(t)‖ ,

v(t)‖v(t)‖

)= 1, t ∈

{t ∈ T :

∥∥u(t)∥∥ �= 0

}. (2)

Since X has a Fréchet differentiable norm, by (2), we obtain that sequence { un(t)‖u(t)‖}∞n=1 is convergent.

Hence there exists x(t) ∈ S(X) such that un(t)‖u(t)‖ → x(t) on {t ∈ T : ‖u(t)‖ �= 0}. Let

u0(t) ={ ‖u(t)‖x(t), t ∈ {t ∈ T : ‖u(t)‖ �= 0}

0, t ∈ {t ∈ T : ‖u(t)‖ = 0}.

It is easy to see that ‖u0‖0 = 1 and un(t) → u0(t) μ-a.e. in T . Next we will prove that l = h, where h ∈ K(u0) and l = limn→∞ kn. In fact, by the Fatou Lemma, it follows that

1h

[1 + ρM (hu0)

]= ‖u0‖0 = lim

n→∞‖un‖0 = lim

n→∞1kn

[1 + ρM (knun)

]≥ 1

l

[1 + ρM (lu0)

].

Hence ‖u0‖0 = [1 + ρM (hu0)]/h = [1 + ρM (lu0)]/l. Since M(u) is strictly convex, we have that L0M (R) is

strictly convex. Noticing that ‖u0‖0 = [1 + ρM (hu0)]/h = [1 + ρM (lu0)]/l, we have l = h. By the convexity of M , we have

S. Shang, Y. Cui / J. Math. Anal. Appl. 421 (2015) 1377–1395 1389

M(‖knun(t)‖) + M(‖hu0(t)‖)2 −M

(‖knun(t) − hu0(t)‖

2

)≥ 0

for μ-a.e. t ∈ T . Moreover, we have

ρM (knun) = kn‖un‖0 − 1 → h‖u0‖0 − 1 = ρM (hu0) as n → ∞.

Therefore, by the Fatou Lemma, we obtain the following

ρM (hu0) =∫T

limn→∞

[M(‖knun(t)‖) + M(‖hu0(t)‖)

2 −M

(‖knun(t) − hu0(t)‖

2

)]dt

≤ lim infn→∞

∫T

[M(‖knun(t)‖) + M(‖hu0(t)‖)

2 −M

(‖knun(t) − hu0(t)‖

2

)]dt

= ρM (hu0) − lim supn→∞

ρM

[12(knun − hu0)

].

This implies that ρM (12 (knun − hu0)) → 0 as n → ∞. Since M ∈ Δ2, we have ‖knun − hu0‖0 ≤

2‖knun − hu0‖ → 0 as n → ∞. Using limn→∞ kn = l = h, we obtain ‖un − u0‖0 → 0 as n → ∞.Case 2. Let sup{kn} = ∞, where kn = K(un). We consider the sequence u′

n = 12 (un + u) in place of

{un}∞n=1, because ‖un − u‖0 → 0 as n → ∞ if and only if ‖u′n − u‖0 → 0 as n → ∞. Moreover,

∥∥∥∥12(un + u)

∥∥∥∥0

≤ 12(‖un‖0 + ‖u‖0)

for every n ∈ N . Hence lim supn→∞ ‖12 (un + u)‖0 ≤ 1. Since

∫T

(12(un(t) + u(t)

), v(t)

)dt = 1

2

∫T

(un(t), v(t)

)dt + 1

2

∫T

(u(t), v(t)

)dt → 1,

we conclude that lim infn→∞ ‖12 (un + u)‖0 ≥ 1. Consequently, limn→∞ ‖1

2(un + u)‖0 = 1. Define wn =2knk(kn + k), where k ∈ K(u). Then the sequence {wn}∞n=1 is bounded. Moreover,

∥∥∥∥12(un + u)

∥∥∥∥0

≤ 1wn

[1 + ρM

(wn · un + u

2

)]

= kn + k

2knk

[1 + ρM

(knk

kn + k(un + u)

)]

≤ kn + k

2knk

[1 + ρM

(k

kn + k(knun)

)+ ρM

(kn

kn + k(ku)

)]

≤ 12

[1kn

(1 + ρM (knun)

)+ 1

k

(1 + ρM (ku)

)]

= 12[‖un‖0 + ‖u‖0] → 1 as n → ∞,

whence it follows that

kn + k[1 + ρM

(2knk · 1(un + u)

)]→ 1 as n → ∞.

2knk kn + k 2

1390 S. Shang, Y. Cui / J. Math. Anal. Appl. 421 (2015) 1377–1395

Since∫T

(u(t), v(t))dt = 1 and ∫T

(un(t), v(t))dt → 1, we have ∫T

(u′n(t), v(t))dt → 1. Therefore, we can

prove in the same way as in Case 1 that ‖u′n − u‖0 → 0. So {un}∞n=1 is relatively compact. Hence EN (X)

has a Fréchet differentiable norm.By Lemma 1, the implication (1) ⇒ (2) is obvious, which completes the proof. �

Lemma 9. (See [1].) For any N -function M and ε > 0, there exists a strictly convex N -function M1 such that

M(u) ≤ M1(u) ≤ (1 + ε)M(u), u ∈ R.

Lemma 10. (See [1].) For any N -function M and ε > 0, there exists an N -function M1 such that

M(u) ≤ M1(u) ≤ (1 + ε)M(u), u ∈ R,

and its right derivative p1 is continuous.

Lemma 11. (See [1].) (1) L0M has the Radon–Nikodym property if and only if M ∈ Δ2; (2) EN is an Asplund

space if and only if M ∈ Δ2.

Theorem 3. Let X be a Fréchet differentiable space. Then

(1) L0M (X∗) has the Radon–Nikodym property if and only if M ∈ Δ2;

(2) EN (X) is an Asplund space if and only if M ∈ Δ2.

Proof. Let M ∈ Δ2. By Lemma 9, there exists a strictly convex N -function M1 such that

M(u) ≤ M1(u) ≤ 2M(u), u ∈ R.

Since M ∈ Δ2, there exist K > 2 and u0 ≥ 0 such that M(2u) ≤ KM(u) whenever u ≥ u0. This implies that

M1(2u) ≤ 2M(2u) ≤ 2KM(u) ≤ 2KM1(u)

whenever u ≥ u0. Therefore M1 ∈ Δ2. By Theorem 2, we obtain that every weak∗ hyperplane of (L0M1

(X∗),‖ · ‖0

1) is approximatively compact. So (L0M1

(X∗), ‖ · ‖01) has the Radon–Nikodym property by Theorem 1.

Moreover, we have

‖u‖0 = infk>0

1k

[1 +

∫T

M(∥∥ku(t)

∥∥)dt] ≤ infk>0

1k

[1 +

∫T

M1(∥∥ku(t)

∥∥)dt] = ‖u‖01

and

‖u‖01 = inf

k>0

1k

[1 +

∫T

M1(∥∥ku(t)

∥∥)dt] ≤ infk>0

1k

[1 +

∫T

2M(∥∥ku(t)

∥∥)dt] ≤ 2‖u‖0

for any u ∈ L0M (X∗). This means that L0

M (X∗) has the Radon–Nikodym property.

S. Shang, Y. Cui / J. Math. Anal. Appl. 421 (2015) 1377–1395 1391

Since M ∈ Δ2, then for any l > 1, there exist v0 > 0 and δ > 0 such that N(lv) ≥ (l + δ)N(v) whenever v ≥ v0. By Lemma 10, there exist δ1 = min{δ/4, δ/(4l)} > 0 and an N -function N2 such that

N(v) ≤ N2(v) ≤ (1 + δ1)N(v), v ∈ R,

and its right derivative p2 is continuous. Then M2 is a strictly convex function. Since δ1 = min{δ/4, δ/4l}, we have

δ2 = δ − lδ11 + δ1

≥ δ − l(δ/4l)1 + δ1

> 0

and

N2(lv) ≥ N(lv) ≥ (l + δ)N(v) ≥ (l + δ) 11 + δ1

N2(v) =(l + δ − lδ1

1 + δ1

)N2(v)

= (l + δ2)N2(v)

whenever v ≥ v0. This implies that M2 ∈ Δ2. Therefore, by Theorem 2, we obtain that every weak∗

hyperplane of (L0M2

(X∗), ‖ · ‖02) is approximatively compact. So (EN2(X), ‖ · ‖2) is an Asplund space by

Theorem 1. Moreover, we have ‖v‖0 ≤ ‖v‖02 and

‖v‖02 = inf

k>0

1k

[1 +

∫T

N2(∥∥kv(t)∥∥)dt] ≤ inf

k>0

1k

[1 +

∫T

(1 + δ1)N(∥∥kv(t)∥∥)dt]

≤ (1 + δ1) infk>0

1k

[1 +

∫T

N(∥∥kv(t)∥∥)dt] ≤ (1 + δ1)‖v‖0

for any v ∈ EN (X). Then

‖v‖ ≤ ‖v‖0 ≤ ‖v‖02 ≤ 2‖v‖2 ≤ 2‖v‖0

2 ≤ 2(1 + δ1)‖v‖0 ≤ 4(1 + δ1)‖v‖.

Since ‖v‖ ≤ 2‖v‖2 ≤ 4(1 + δ1)‖v‖ and (EN2(X), ‖ · ‖2) is an Asplund space, we obtain that EN (X) is an Asplund space.

If L0M (X∗) has the Radon–Nikodym property, then L0

M has the Radon–Nikodym property. Therefore, by Lemma 11, we have M ∈ Δ2. If EN (X) is an Asplund space, then EN is an Asplund space. Therefore, by Lemma 11, we have M ∈ Δ2, which completes the proof. �Corollary 2. (1) L0

M has the Radon–Nikodym property if and only if M ∈ Δ2; (2) EN is an Asplund space if and only if M ∈ Δ2.

4. Applications to the theory of generalized inverses

Let T be a linear bounded operator from X into Y . Let D(T ), R(T ) and N(T ) denote the domain, range and null space of T , respectively. If N(T ) �= {0} or R(T ) �= Y , the operator equation Tx = y is generally ill-posed. In applications, one usually looks for the best approximate solution (b.a.s.) to the equation (see [15]).

A point x0 ∈ D(T ) is called the best approximate solution to the operator equation Tx = y, if

‖Tx0 − y‖ = inf{‖Tx− y‖ : x ∈ D(T )

}

1392 S. Shang, Y. Cui / J. Math. Anal. Appl. 421 (2015) 1377–1395

and

‖x0‖ = min{‖v‖ : v ∈ D(T ), ‖Tv − y‖ = inf

x∈D(T )‖Tx− y‖

},

where y ∈ Y (see [15]).Nashed and Votruba [15] introduced the concept of the (set-valued) metric generalized inverse T .

Definition 8. Let X, Y be Banach spaces and T be a linear operator from X to Y . The set-valued mapping T ∂ : Y → X defined by

T ∂(y) ={x0 ∈ D(T ) : x0 is a best approximation solution to T (x) = y

}for any y ∈ D(T ∂), is called the (set-valued) metric generalized inverse of T , where

D(T ∂

)=

{y ∈ Y : T (x) = y has a best approximation solution in X

}.

Definition 9. A Banach space X is said to have the H property whenever if {xn}∞n=1 ⊂ S(X), x ∈ S(X)and xn

w−−→ x, then xn → x.

During the last three decades, the linear generalized inverses of linear operators in Banach spaces and their applications have been investigated by many authors (see [12,23–25]). In this section, we will investigate continuity of the metric generalized inverse of a linear operator.

Theorem 4. Let X be a smooth spaces, T : X → Y a linear bounded operator, N(T ) a one-dimensional subspace, N(T ∗) a Chebychev set and R(T ) a closed subspace. Then D((T ∗)∂) = X∗ and (T ∗)∂ is single-valued. Moreover, if every weak∗ hyperplane of X∗ is approximatively compact and Y ∗ is approximatively compact, then (T ∗)∂ is continuous.

In order to prove this theorem, we give first some lemmas.

Lemma 12. Let X be a smooth space. Then for any x ∈ S(X), if x∗n ∈ S(X∗) and x∗

n(x) → 1, then there exists a supporting functional x∗ of x such that x∗

nw∗−−→ x∗ as n → ∞.

Proof. Since X is a smooth space, we obtain that B(X∗) is w∗ sequentially compact. Hence, if x∗n(x) → 1,

then there exist x∗ ∈ B(X∗) and a subsequence {x∗nk}∞k=1 of {x∗

n}∞n=1 such that x∗nk

w∗−−→ x∗ and x∗(x) = 1. This implies that x∗ is a supporting functional of x. We claim that x∗

nw∗−−→ x∗ as n → ∞. Otherwise, there

exist ε0 > 0, y ∈ X and a subsequence {nl} of {n} such that |x∗nl

(y) − x∗(y)| > ε0. Since x∗nl

(x) → 1, there exists a subsequence {ni} of {nl} such that x∗

ni

w∗−−→ x∗ as i → ∞, a contradiction, which completes the proof. �Lemma 13. Let X be a smooth space. Then every weak∗ hyperplane of X∗ is a Chebychev set.

Proof. By [22], we know that every weakly∗ closed convex set is a proximinal set. Let H∗ = {x∗ ∈ X∗ :x∗(x) = k} be a weak∗ hyperplane. For any x∗ /∈ H∗, we may assume without loss of generality that x∗ = 0. Suppose that there exists z∗1 �= z∗2 such that ‖z∗1‖ = ‖z∗2‖ = dist(0, H∗). Let {y∗n}∞n=1 = (z∗1 , z∗2 , z∗1 , z∗2 , ...). We claim that {y∗n}∞n=1 is a weakly∗ Cauchy sequence. In fact, since H∗ is a weakly∗ closed set, we have dist(0, H∗) = r > 0. Pick y∗0 ∈ PH∗(0). Then

r = dist(0, PH∗(0)

)=

∥∥y∗0∥∥, Br(0) ∩H∗ = ∅, Br(0) ∩H∗ = PH∗(0).

S. Shang, Y. Cui / J. Math. Anal. Appl. 421 (2015) 1377–1395 1393

We may assume without loss of generality that k ≤ 0. Hence

k = sup{x(y∗)

: y∗ ∈ H∗} ≤ inf{x(y∗)

: y∗ ∈ Br(0)}

= −‖x‖ ·∥∥y∗0∥∥.

In fact, suppose that there exists y∗0 ∈ Br(0) such that x(y∗0) < k. Then there exists λ ∈ (0, 1) such that x(λy∗0) = k. It is easy to see that λy∗0 ∈ Br(0) and λy∗0 ∈ H∗, a contradiction. Since y∗0 ∈ PH∗(0) ⊂ H∗, we have

−‖x‖ ·∥∥y∗0∥∥ ≤ x

(y∗0)≤ sup

{x(y∗)

: y∗ ∈ H∗}≤ inf

{x(y∗)

: y∗ ∈ Br(0)}

= −‖x‖ ·∥∥y∗0∥∥.

This implies that

−‖x‖ ·∥∥y∗0∥∥ = x

(y∗0)

= sup{x(y∗)

: y∗ ∈ H∗}.This means that the inequality x(y∗0) ≥ x(y∗n) holds. Therefore

∥∥0 − y∗0∥∥ = x

(0 − y∗0

)≤ x

(0 − y∗n

)≤

∥∥0 − y∗n∥∥ → dist

(0, H∗) =

∥∥0 − y∗0∥∥ =

∥∥y∗0∥∥.Hence

∥∥y∗n∥∥ →∥∥y∗0∥∥ as n → ∞

and

x(0 − y∗n

)→

∥∥0 − y∗0∥∥ as n → ∞.

Furthermore, we have

x

(− y∗n‖y∗n‖

+ y∗n‖y∗0‖

)=

(1

‖y∗0‖− 1

‖y∗n‖

)· x

(y∗n

)→ 0 as n → ∞,

which shows that

x

(− y∗n‖y∗n‖

)→ 1 as n → ∞.

By Lemma 12, we obtain that {y∗n}∞n=1 is a weakly∗ Cauchy sequence. However, it is easy to see that {y∗n}∞n=1 = (z∗1 , z∗2 , z∗1 , z∗2 , ...) is not a weakly∗ Cauchy sequence, a contradiction. This implies that every weak∗ hyperplane of X∗ is a Chebychev set, which completes the proof. �Proof of Theorem 4. Since R(T ) is closed, we obtain that R(T ∗) = N⊥ = {x∗ ∈ X∗ : x∗(x) = 0, x ∈ N(T )}by the closed range theorem. Since dimN(T ) = 1, there exists x ∈ S(X) ∩N(T ) such that R(T ∗) = {x∗ :x∗(x) = 0}. This implies that R(T ∗) is a weakly∗ hyperplane of X∗. By Lemma 13, we obtain that R(T ∗) is a Chebychev set. Let x∗ ∈ X∗. Then there exists a unique z∗ ∈ R(T ∗) such that PR(T∗)(x∗) = {z∗}. Hence there exists y∗ ∈ Y ∗ such that (T ∗)−1(z∗) = y∗ + N(T ∗). Since N(T ∗) be a Chebychev set, we obtain that D((T ∗)∂) = X∗ and (T ∗)∂ is single-valued.

Let y∗n = (T ∗)∂(x∗n), y∗0 = (T ∗)∂(x∗

0) and x∗n → x∗

0. It is easy to see that T ∗y∗n = PR(T∗)(x∗n) and

T ∗y∗0 = PR(T∗)(x∗0). Since every weak∗ hyperplane of X∗ is approximatively compact and R(T ∗) is a

1394 S. Shang, Y. Cui / J. Math. Anal. Appl. 421 (2015) 1377–1395

Chebychev set, we obtain that PR(T∗) is continuous. Then T ∗y∗n → T ∗y∗0 as n → ∞. Since T is a linearbounded operator, we obtain that N(T ∗) is a closed subspace. Put

T ∗ : Y ∗/N(T ∗) → R

(T ∗), T ∗

[y∗]

= T ∗(y∗).By the inverse operator theorem, we have

[y∗n

]= T ∗−1(PR(T∗)

(x∗n

))→ T ∗−1(PR(T∗)

(x∗

0))

=[y∗0]

as n → ∞. (3)

This implies that ‖[y∗n]‖ → ‖[y∗0 ]‖ as n → ∞. Since

y∗n =(T ∗)∂(x∗

n

),

∥∥[y∗n]∥∥ = infz∗∈N(T∗)

∥∥x∗n + z∗

∥∥, T ∗[x∗n

]= T ∗(y∗n + z∗

)= PR(T∗)

(y∗n

),

y∗0 =(T ∗)∂(x∗

0),

∥∥[x∗0]∥∥ = inf

z∗∈N(T∗)

∥∥x∗n + z∗

∥∥, T ∗[x∗

0]

= T ∗(x∗0 + z∗

)= PR(T∗)

(y∗0),

we have ‖[y∗n]‖ = ‖y∗n‖ and ‖[y∗0 ]‖ = ‖y∗0‖. Noticing that ‖[y∗n]‖ → ‖[y∗0 ]‖, we have ‖y∗n‖ → ‖y∗0‖. We will derive limn→∞ (T ∗)∂(x∗

n) = (T ∗)∂(x∗0) for each of the following two cases.

Case 1. y∗0 = 0. By ‖y∗n‖ → ‖y∗0‖, we have ‖y∗n‖ → 0 as n → ∞. This implies that (T ∗)∂(x∗n) → (T ∗)∂(x∗

0)as n → ∞.

Case 2. y∗0 �= 0. By x∗n → x∗

0, we obtain that the sequence {x∗n}∞n=1 is a bounded sequence. Since

‖x∗n − PR(T∗)(x∗

n)‖ → dist(x∗0, R(T ∗)), we obtain that {PR(T∗)(x∗

n)}∞n=1 is a bounded sequence. Since T ∗−1

is a linear bounded operator, we obtain that {[y∗n]}∞n=1 is a bounded sequence. By ‖[y∗n]‖ = ‖y∗n‖, we obtain that {y∗n}∞n=1 is a bounded sequence. Since Y ∗ is approximatively compact, Y ∗ is reflexive. Hence there exists a subsequence {y∗nk

}∞k=1 of {y∗n}∞n=1 such that y∗nk

w−−→ y∗ as k → ∞. This implies that T ∗y∗nk

w−−→ T ∗y∗ as k → ∞. By T ∗y∗n → T ∗y∗0 , we have T ∗y∗0 = T ∗y∗. By the Hahn–Banach theorem, there exists y∗∗ ∈ S(X∗∗)such that y∗∗(y∗) = ‖y∗‖. Then

∥∥y∗∥∥ = y∗∗(y∗)

= limk→∞

y∗∗(y∗nk

)≤ lim

k→∞

∥∥y∗nk

∥∥ = limk→∞

∥∥[y∗nk

]∥∥ =∥∥[y∗0]∥∥ =

∥∥y∗0∥∥.Hence ‖y∗‖ = ‖y∗0‖. Noticing that y∗0 = (T ∗)∂(x∗

0), we have y∗0 = y∗. Then y∗nk

w−−→ y∗0 as k → ∞. Since X∗

is approximatively compact, by [11], we obtain that X∗ has the H property. Therefore, by y∗nk

w−−→ y∗0 and ‖y∗nk

‖ → ‖y∗0‖, we have y∗nk→ y∗0 as k → ∞. Then y∗n → y∗0 as n → ∞. In fact, suppose that there exist

ε > 0 and a subsequence {ni} of {n} such that ‖y∗ni− y∗0‖ > ε. Then there exists a subsequence {nl} of

{ni} such that y∗nl→ y∗0 as l → ∞, a contradiction.

By Case 1 and Case 2, we have (T ∗)∂(x∗n) → (T ∗)∂(x∗

0) as n → ∞, which completes the proof. �Lemma 14. (See [7].) L0

M (X) is approximatively compact if and only if

(1) M ∈ Δ2 and N ∈ Δ2;(2) M(u) is strictly convex;(3) X is approximatively compact and rotund.

Theorem 5. Suppose that every weak∗ hyperplane of L0M1

(X∗) is approximatively compact, L0M2

(Y ∗) is approximatively compact and T is a linear bounded operator from EN1(X) to EN2(Y ). Then if N(T ) is a one-dimensional subspace and R(T ) is a closed subspace, then (T ∗)∂ is single-valued and continuous, and D((T ∗)∂) = L0

M1(X∗).

Proof. By Lemma 14 and Theorem 2, we obtain that if L0M2

(Y ∗) is approximatively compact, then EN2(Y )has a Fréchet differentiable norm and is reflexive. This implies that L0

M (Y ∗) is rotund. Hence N2(T ∗) is a

2

S. Shang, Y. Cui / J. Math. Anal. Appl. 421 (2015) 1377–1395 1395

Chebychev set. Moreover, by Theorem 2, we obtain that EN1(X) has a Fréchet differentiable norm. Then EN1(X) is smooth. Therefore, by Theorem 4, we obtain that D((T ∗)∂) = L0

M1(X∗), (T ∗)∂ is single-valued

and continuous, which completes the proof. �References

[1] S.T. Chen, Geometry of Orlicz spaces, Dissertations Math., Warszawa, 1996.[2] S. Chen, H. Hudzik, W. Kowalewski, Y. Wang, M. Wisla, Approximative compactness and continuity of metric projector

in Banach spaces and applications, Sci. China Ser. A 50 (2) (2007) 75–84.[3] S.T. Chen, X. He, H. Hudzik, A. Kamin’ska, Monotonicity and best approximation in Orlicz–Sobolev spaces with the

Luxemburg norm, J. Math. Anal. Appl. 344 (2) (2008) 687–698.[4] M. Contreras, R. Paya, On upper semicontinuity of duality mappings, Proc. Amer. Math. Soc. 121 (2) (1994) 451–459.[5] M. Denker, H. Hudzik, Uniformly non-l(1)n Musielak–Orlicz sequence spaces, Proc. Indian Acad. Sci. 101 (1991) 71–86.[6] K. Fan, I.G. Glicksberg, Fully convex normed linear spaces, Proc. Natl. Acad. Sci. USA 41 (1955) 947–953.[7] W. Gonga, Z. Shi, Drop properties and approximative compactness in Orlicz–Bochner function spaces, J. Math. Anal.

Appl. 344 (1) (2008) 748–756.[8] H. Hudzik, W. Kurc, Monotonicity properties of Musielak–Orlicz spaces and dominated best approximation in Banach

lattices, J. Approx. Theory 95 (3) (1998) 115–121.[9] H. Hudzik, B.X. Wang, Approximative compactness in Orlicz spaces, J. Approx. Theory 95 (1) (1998) 82–89.

[10] H. Hudzik, W. Kurc, M. Wisla, Strongly extreme points in Orlicz function spaces, J. Math. Anal. Appl. 189 (1995) 651–670.[11] H. Hudzik, W. Kowalewski, G. Lewicki, Approximative compactness and full rotundity in Musielak–Orlicz space and

Lorentz–Orlicz space, Z. Anal. Anwend. 25 (1) (2006) 163–192.[12] H. Hudzik, Y.W. Wang, W. Zheng, Criteria for the metric generalized inverse and its selections in Banach spaces, Set-Valued

Anal. 16 (2008) 51–65.[13] N.W. Jefimow, S.B. Stechkin, Approximative compactness and Chebyshev sets, Soviet Math. 2 (1961) 1226–1228.[14] Pei-Kee Lin, Kochner Function Spaces, Birkhäuser, Boston, 2004.[15] M. Nashed, G. Votruba, A unified approach to generalized inverses of linear operator: II Extremal and proximinal prop-

erties, Bull. Amer. Math. Soc. 80 (1974) 831–835.[16] E.V. Oshman, Characterization of subspaces with continuous metric projection into a normed linear space, Soviet Math.

13 (6) (1972) 1521–1524.[17] R.R. Phelps, Convex Functions, Monotone Operators and Differentiability, Lecture Notes in Math., vol. 1364, Springer-

Verlag, New York, 1989.[18] S. Shang, Y. Cui, Locally uniform convexity in Musielak–Orlicz function spaces of Bochner type endowed with the Lux-

emburg norm, J. Math. Anal. Appl. 378 (2) (2011) 432–441.[19] S. Shang, Y. Cui, Y. Fu, Nonsquareness and locally uniform nonsquareness in Orlicz–Bochner function spaces endowed

with Luxemburg norm, J. Inequal. Appl. 2011 (2011) 1–15.[20] S. Shang, Y. Cui, Y. Fu, Midpoint locally uniform rotundity of Musielak–Orlicz–Bochner function spaces endowed with

the Luxemburg norm, J. Convex Anal. 19 (1) (2012) 213–223.[21] S. Shang, Y. Cui, Y. Fu, P -convexity of Orlicz–Bochner function spaces endowed with the Orlicz norm, Nonlinear Anal.

74 (1) (2012) 371–379.[22] S. Shang, Y. Cui, Y. Fu, Smoothness and approximative compactness in Orlicz function spaces, Banach J. Math. Anal.

8 (1) (2014) 26–38.[23] Y.W. Wang, J. Liu, Metric generalized inverse for linear manifolds and extremal solutions of linear inclusion in Banach

spaces, J. Math. Anal. Appl. 302 (2005) 360–371.[24] Y.W. Wang, S. Pan, An approximation problem of the finite rank operator in Banach spaces, Sci. China 46 (2003) 245–250.[25] H. Wang, Y.W. Wang, Metric generalized inverse of linear operator in Banach space, Chin. Ann. Math. 24 (2003) 509–520.