1 Introducction
In this article, we take as reference the classic book Gâteaux Differentiability of Convex Functions and Topology: Weak Asplund Spaces 1, in which Marián J. Fabian talks about the Asplund spaces. First, we define the multivalued functions as follows: Let 𝑋 and 𝑌 be sets. A multivalued function from 𝑋 to 𝑌 is a relation which assigns to each 𝑥 ( 𝑋 a subset of 𝑌, denoted 𝑓(𝑥); 𝑓(𝑥) = ø for some 𝑥 ∊ 𝑋 is admitted. The graph of the multivalued functions 𝑓 is the set Graph(𝑓): = {(𝑥, 𝑦) ∊ 𝑋 ( 𝑌: 𝑦 ( 𝑓(𝑥)}. If for each 𝑥 ∊ 𝑋, 𝑓(𝑥) has only one element, we say the function is single-valued. A single-valued function from 𝑋 to 𝑌, is a relation 𝑥 ↦ 𝑦 = 𝑓(𝑥) which assigns to each 𝑥 ∊ 𝑋 a unique 𝑦 ∊ 𝑌. The graph of such functions is Graph(𝑓): = {(𝑥, 𝑦) ∊ 𝑋 ( 𝑌: 𝑦 = 𝑓(𝑥)}.
The objective of this article is the study of the properties of the usco functions. A usco function is a multivalued function 𝑓: 𝑋 → 𝑌 between Banach spaces which is upper semicontinuous and 𝑓(𝑥) is compact for each 𝑥 ∊ 𝑋. A usco function 𝑓: 𝑋 → 𝑌 is minimal if for each usco function 𝑔: 𝑋 → 𝑌 such that
One of the most important examples of the usco functions appears in the study of differentiability of the convex functions defined on an open convex subset Ω of a Banach space 𝑋. It is convenient to note that the subdifferential of a function 𝑓: Ω → ℝ ∪ {+∞} in 𝑥 ∊ Ω is the multivalued function ∂𝑓: 𝑋 → 𝑋´ defined as follows:
It is also important to observe that in the completely regular topological spaces, singletons and closed sets can be separated by open sets. If 𝑋 is a Banach space, the topological dual with the weak topology is a completely regular space. Let us remember that a subset 𝐺 of a topological space 𝑋 is residual if there exists a countable family
The Asplud 𝐺-spaces are Banach spaces in which every continuous convex function, defined on an open convex set Ω, is differentiable in the Gâteaux sense in a residual subset of Ω. We denote by
2 𝛽-Differentiability
Let us start with the following definition:
Definition 2.1 Let 𝑋 be a vector space on a field 𝕂. A vector bornology in 𝑋 is a subset 𝛽 of parts of 𝑋, denoted by ℘(𝑋), satisfying the following axioms:
𝑏𝑜𝑟𝑣 1. The union of all sets of 𝛽 is 𝑋:
𝑏𝑜𝑟𝑣 2. 𝛽 is stable under inclusions. That is, if 𝐵 ( 𝛽 and 𝐷 ⊂ 𝐵, then 𝐷 ( 𝛽.
𝑏𝑜𝑟𝑣 3. 𝛽 is stable under finite union. That is, if
𝑏𝑜𝑟𝑣 4. 𝛽 is stable under the sum operation. That is, if
𝑏𝑜𝑟𝑣 5. 𝛽 is stable under the scalar multiplication operation. That is, if 𝑆 ∊ 𝛽 and λ ( 𝕂, then λ𝑆 ∊ 𝛽.
𝑏𝑜𝑟𝑣 6 is stable under the formation of balanced envelope. That is, if 𝑆 ∊ 𝛽, then 𝑏𝑎𝑙 (𝑆) ∊ 𝛽. Where 𝑏𝑎𝑙 (𝑆) denote the balanced hull of a set 𝑆 defined by
In the context of a topological vector space (𝑋, τ), we have some natural bornologies.
Examples 2.2 Let (𝑋, τ) be an 𝑡.𝑣.𝑠., then the following collections of subsets of 𝑋 are vector bornologies.
1. The 𝐹-bornology
2. The 𝐻-bornology
3. The 𝐺-bornology
From the definitions of these bornologies, it is easy to see that
For a proof of this result see 2.
In the sequel, 𝑋 and 𝑌 denote Banach spaces, Ω is an open subset of 𝑋 and ℒ(𝑋, 𝑌) denotes the space of linear transformation from 𝑋 to 𝑌 that are continuous.
Definition 2.3 Let 𝛽 be a bornology in 𝑋. We say that a function 𝑓: Ω → 𝑌 is 𝛽-differentiable in 𝑎 ∊ Ω if there exists a function 𝑢 ∊ ℒ(𝑋, 𝑌) such that for every 𝑆 ∊ 𝛽
The linear and continuous function 𝑢 is called 𝛽-derivative of 𝑓 at the point 𝑎 ∊ Ω and is denoted by
The following result provides a differentiability criterium with respect to a given bornology:
Theorem 2.4 Let 𝑋 and 𝑌 be 𝑡.𝑣.𝑠., Ω an open subset of 𝑋, and 𝛽 be a bornology in 𝑋. A necessary condition for 𝑓: Ω → 𝑌 to be 𝛽-differentiable in 𝑎 ∊ Ω is that for every 𝑆 ∊ 𝛽
This result follows doing a suitable modification in the proof of proposition 1.23 in 3.
In the case of a convex continuous functions 𝑓:Ω ⊂ 𝑋 → 𝑌, we have the following characterization of the 𝛽-differentiable in 𝑎 ∊ Ω:
Theorem 2.5 Let 𝑋 be an 𝑡.𝑣.𝑠., Ω an open convex subset of 𝑋, and 𝛽 be a bornology in 𝑋. A necessary and sufficient condition for a convex and continuous function 𝑓:Ω → ℝ be 𝛽-differentiable in 𝑎 ∊ Ω is that for every 𝑆 ∊ 𝛽
Similar considerations as theorem 2.4.
2.1 Single-Valued Maps Derivatives
Now, we will study the main notions of differentiation in a 𝑡.𝑣.𝑠. linked to the bornologies mentioned before: the Frechet derivative linked to the bornology of the bounded subsets of a 𝑡.𝑣.𝑠., the derivative of Hadamard linked to the bornology of the compact subsets and the derivative of Gâteaux linked to the bornology of the sets finite. Our interest in this work is to develop this theory in the case of Banach spaces.
Definition 2.1.1 Let 𝑋 and 𝑌 be Banach spaces, Ω ⊂ 𝑋 be an open subset and 𝑓: Ω → 𝑌 be a function.
We say that the function 𝑓: Ω → 𝑌 is 𝐺-differentiable (differentiable in the sense of Gâteaux) at the point 𝑎 ∊ Ω if there exists 𝑢 ∊ ℒ(𝑋, 𝑌) such that
and this limit is uniform on the finite subsets of 𝑋. In this case, we say that 𝑢 is the 𝐺-derivative of 𝑓 in 𝑎 and
We say that the function 𝑓: Ω → 𝑌 is 𝐻-differentiable (differentiable in the sense of Hadamard) at the point 𝑎 ∊ Ω if there exists 𝑢 ∊ ℒ(𝑋, 𝑌) such that
and this limit is uniform on the compact subsets of 𝑋. In this case, we say that 𝑢 is the 𝐻-derivative of 𝑓 in 𝑎 and
We say that the function 𝑓: Ω → 𝑌 is 𝐹-differentiable (differentiable in the sense of Fréchet) at the point 𝑎 ∊ Ω if there exists 𝑢 ∊ ℒ(𝑋, 𝑌) such that
and this limit is uniform on the bounded subsets of 𝑋. In this case, we say that 𝑢 is the 𝐹-derivative of 𝑓 in 𝑎 and
When differentiability holds for any 𝑎 ∊ Ω, we say that 𝑓 is 𝐺-differentiable in Ω (resp. 𝐻-differentiable) (resp. 𝐹-differentiable).
It is straightforward to see that the condition for uniform convergence is expressed as follows:
For all bounded set 𝐵 of 𝑋 and every ε > 0 there is a δ(𝐵, ε) > 0 such that
It is clear that if 𝑓 is 𝐹-differentiable in 𝑎 ∊ Ω with derivative 𝑢, then it is 𝐺-differentiable in 𝑎 with derivative 𝑢.
The following elementary theorem is fundamental to the study of differentiability of convex functions:
Theorem 2.1.2 Let 𝑋 be a normed space. A convex continuous function defined on an open convex set with values in ℝ is necessarily locally Lipschitz.
For a proof of this theorem see the work by R. Phelps in 3.
3 USCO Functions
In this section we will go deeper into the study of the class of usco functions. We start giving some basic definitions.
Definition 3.1 (Multivalued function) Let 𝑋 and 𝑌 be topological spaces. A multivalued function of a set 𝑋 in a set 𝑌 is a correspondence 𝑥 ↦ (𝑥), which assigns to each 𝑥 ∊ 𝑋 a subset 𝑓(𝑥) of the set 𝑌. It is possible that the set (𝑥) is the empty set. The effective domain of this function is the set of the 𝑥 ∊ 𝑋 such that (𝑥) ≠ Ø.
Definition 3.2 (Upper/Lower semicontinuous multivalued function) Let 𝑋 and 𝑌 be topological spaces. A multivalued function b: 𝑋 → 𝑌 is upper semicontinuous in 𝑎 ( 𝑋 if for every open set 𝑊 in 𝑌 such that 𝑊 ( (𝑎) (open neighborhood of 𝑓(𝑎)) there exists an open set 𝑉 of 𝑋 such that 𝑎 ∊ 𝑉 (neighborhood of 𝑎) and 𝑓(𝑥) ( 𝑊 for all 𝑥 ∊ 𝑉. In the case that 𝑓 is upper semicontinuous for any 𝑎 ∊ 𝑋, we say that 𝑓 is upper semicontinuous in 𝑋.
A multivalued function 𝑓: 𝑋 → 𝑌 is lower semicontinuous in 𝑎 ∊ 𝑋 if for every open set 𝑊 in 𝑌 such that 𝑓(𝑎) ∩ 𝑊 ≠ Ø, there exists an open set 𝑉 of 𝑋 such that 𝑎 ∊ 𝑉 (neighborhood of 𝑎) and 𝑓(𝑥) ∩ 𝑊 ≠ Ø for all 𝑎 ∊ 𝑉. In the case that 𝑓 is lower semicontinuous for any 𝑎 ∊ 𝑋, we say that 𝑓 is lower semicontinuous in 𝑋.
The first remark is that 𝑓 is continuous in 𝑎 ∊ 𝑋 if it is both upper semicontinuous and lower semicontinuous at the point 𝑎.
Definition 3.3 (Graph of a multivalued function) Let 𝑋 and 𝑌 be topological spaces. The graph of a multivalued function 𝑓: 𝑋 → 𝑌 is the subset Graph (𝑓) of 𝑋 ( 𝑌 of the pairs (𝑥, 𝑦) ∊ 𝑋 ( 𝑌 such that 𝑦 ∊ (𝑥).
Definition 3.4 The limit values of a net
Where ≽ is the partial order relation in 𝛵.
We note that 𝑦 is a limit value of the indicated net if and only if there exists a subnet that converges to 𝑦.
Theorem 3.5 Let 𝑋 and 𝑌 be topological spaces. A multivalued function 𝑓: 𝑋 → 𝑌 is upper semicontinuous if and only if for any closed set 𝐶 ⊂ 𝑌, the set
is closed on 𝑋.
Proof. Suppose that 𝑓 is upper semicontinuous, we want to show that
Let 𝐶 be a closed subset of 𝑌 and 𝐴 = 𝐶𝐶, then 𝐴 is an open set in 𝑌 and if (𝑥) ∩ 𝐶 = Ø, then 𝑓(𝑥) ( 𝐴. Since 𝑓 is upper semicontinuous, there is a neighborhood open 𝑉 of 𝑥 such that (𝑧) ( 𝐴 for all 𝑧 ∊ 𝑉. But this means that (𝑧) ∩ 𝐶 = Ø for all 𝑧 ( 𝑉 and, therefore,
is open set and so
is closed with 𝐶 is closed, we want to establish that 𝑓 is upper semicontinuous. Let 𝑎 ∊ 𝑋 and 𝑊 is an open subset of 𝑌 such that (𝑎) ( 𝑊. Then, 𝑊𝐶 is a closed set and by hypothesis
is closed in 𝑋 and its complement 𝑉 is an open set. Now, 𝑎 ∊ 𝑉 since (𝑎) ∩ 𝑊𝐶 = Ø and 𝑓(𝑥) ∩ 𝑊𝐶 = Ø for all 𝑥 ∊ 𝑉, so that 𝑓(𝑥) ⊆ 𝑊 for all 𝑋 ∊ 𝑉.
Observe that now we are able to introduce a relation of order on the set of multivalued functions (𝑋, 𝑌) with 𝑋 and 𝑌 being topological spaces. For 𝑓, 𝑔 in (𝑋, 𝑌), se define the order ≼ as
Definition 3.6 (Usco Function) Let 𝑋 and 𝑌 be topological spaces. A multivalued function 𝑓: 𝑋 → 𝑌 is a usco function if 𝑓 is upper semicontinuous such that 𝑓(𝑥) ≠ Ø and is compact for any 𝑥 ∊ 𝑋. We denote the set of usco functions from 𝑋 to 𝑌 by 𝒰 (𝑋,𝑌).
We say that 𝑓 ∊ 𝒰 (𝑋,𝑌) is a minimal usco function if 𝑓 is a minimal element of the ordered set (𝒰 (𝑋,𝑌), ≼). This means that, if 𝑔 ∊ 𝒰 (𝑋, 𝑌) and 𝑔≼ 𝑓, then 𝑓 = 𝑔.
Now, we will establish some results about usco functions.
Theorem 3.7 For every 𝑢 ∊ 𝒰 (𝑋,𝑌), there exists a minimal usco function 𝑓 ∊ 𝒰 (𝑋,𝑌) such that 𝑓≼ 𝑢.
Proof. Let 𝑢 ∊ 𝒰 (𝑋,𝑌) and let ℋ be the collection of usco functions ℎ such that ℎ≼𝑢. Let us show that every chain ℒ contained in ℋ is bounded below. If ℱ is a finite subset in ℒ, then
is upper semicontinuous and thus, a usco function that minorizes ℒ. By Zorn’s lemma ℋ has a minimal element.
Lemma 3.8 Let (𝑋, 𝛼) and (𝑌, τ) be topological spaces and 𝑓: 𝑋 → 𝑌 be a usco function. If (𝑥𝑡,𝑦𝑡)𝑡∊𝛵 is a net in 𝐺𝑟𝑎𝑝ℎ(𝑓) and 𝑥𝑡 → 𝑥, then (𝑦𝑡)𝑡∊𝛵 has at least one limit value in 𝑓(𝑥).
Proof. By contradiction, we assuming that the net (𝑦𝑡)𝑡∊𝛵 has not limit values in 𝐺𝑟𝑎𝐺ℎ(𝑓). That is, for all 𝑧 ∊ 𝑓(𝑥), there exists a 𝑡𝑧 such that
As 𝑥𝑡 → 𝑥, there exists a
Theorem 3.9 (Characterization of usco functions) Let (𝑋, 𝛼) and (𝑌, 𝛽) be topological spaces. A multivalued function 𝑓: 𝑋 → 𝑌 is a usco function if and only if its graph 𝐺𝑟𝑎𝑝ℎ(𝑓) is a closed set. Moreover, there exists a usco function 𝑢 ∊ 𝒰 (𝑋, 𝑌) such that 𝑓 ≼ 𝑢.
Proof. Assume that 𝑓 is a usco function. If prove that
In the other direction, if Graph (𝑓) is closed and there is a usco function 𝑢 such that 𝑓 ≼ 𝑢, then 𝑓 is a usco function. By theorem 3.5, it suffices to show that 𝑓(𝑥) is closed for all 𝑥 ∊ 𝑋. Let
The following lemma will be helpful for the main result in this section.
Lemma 3.10 Let (𝑋, τ) be the topological space, (𝑌, σ) Hausdorff space and 𝑓: 𝑋 → 𝑌 a usco function. The following are equivalent:
(i)𝑓 is a minimal usco function.
(ii) If 𝐴 is an open subset of 𝑋, 𝑊 is an open subset of 𝑌 and 𝑓(𝑎) ∩ 𝑊 ≠ Ø for some 𝑎 ∊ 𝐴, then there exists an open nonempty subset 𝑉 ( 𝐴 such that 𝑓(𝑥) ( 𝑊 for all 𝑥 ∊ 𝑉.
(iii)If 𝐴 is an open subset of 𝑋 and 𝐶 is a closed subset of 𝑌 such that 𝑓(𝑎) ∩ 𝐶 ≠ Ø for all 𝑎 ∊ 𝐴, then 𝑓(𝑎) ( 𝐶 for all 𝑎 ∊ 𝐴.
Proof. We will show that (𝑖) ⇒ (𝑖 𝑖). Let 𝑎 ( 𝑋 and 𝑊 ( 𝑌 open subsets as in (ii). We need to establish that exists an 𝑎0 ∊ 𝐴 such that 𝑓(𝑎0) ( 𝑊 because the fact that 𝑓 is upper semicontinuous implies that there is an open neighborhood 𝑉 in 𝐴 of 𝑎0 such that
We argue by contradiction, let us assume that this statement is false and let 𝐶:= 𝑊𝐶 closed in 𝑌. Then, 𝑓(𝑥) ∩ 𝐶≠ Ø for any 𝑥 ∊ 𝐴. We define the function ℎ: 𝑋 → 𝑌 by
It is clear that ℎ(𝑥) ≠ Ø, closed for all 𝑥 ∊ 𝑋 and ℎ ≼ 𝑓. Then, we conclude that ℎ ∊ 𝒰 (𝑋, 𝑌) by theorem 3.9. Since 𝑓 is a minimal usco function by hypothesis, then ℎ = 𝑓 and consequently, ℎ(𝑥)= 𝑓(𝑥) ( 𝐶 for all 𝑥 ∊ 𝐴. This is a contradiction, since by hypothesis 𝑓(𝑎) ∩ 𝑊 ≠ Ø for some 𝑎 ∊ 𝐴.
Now we show that (𝑖 𝑖) ⇒ (𝑖) . By hypothesis 𝑓: 𝑋 → 𝑌 is a usco function. From theorem 3.7, there exists a minimal usco function 𝑔 ≼ 𝑓 and 𝑔 = 𝑓, as we will show below. If 𝑓 and 𝑔 are not equal, there exists 𝑥0 ∊ 𝑋 such that
This construction allows us to see that 𝑥0 ∊ 𝐴 and 𝑓(𝑥0) ∩ 𝛵 ≠ Ø. Then, there is a subset open nonempty 𝑉 ( 𝐴 such that
which is a contradiction, since
Now we see that (𝑖𝑖) ⇒ (𝑖𝑖𝑖). Let 𝐴 be a subset open in 𝑋 and 𝐶 be a subset closed in 𝑌 as in (iii). Suppose there exists 𝑎 ∊ 𝐴 such that 𝑓(𝑎) is not contained in 𝐶. Then, 𝑓(𝑎) ∩ 𝑊≠ Ø for some 𝑎 ∊ 𝐴, where 𝑊:= 𝐶𝐶 is an open subset in 𝑌. For (ii) there exists a nonempty open subset 𝑉 ( 𝐴 such that 𝑓(𝑥) ( 𝑊 for all 𝑥 ∊ 𝑉, i.e., 𝑓 (𝑉) ∩ 𝐶 = Ø , which is a contradiction, since by hypothesis 𝑓(𝑥) ∩ 𝐶≠ Ø for all 𝑥 ( 𝐴.
We will show that (𝑖𝑖𝑖) ⇒ (𝑖𝑖). Let 𝐴 ( 𝑋 and 𝑊 ( 𝑌, with 𝑓(𝑎0) ∩ 𝑊≠ Ø such that 𝑎0 ∊ 𝐴. Let us prove that there exists 𝑎 ∊ 𝐴 such that 𝑓(𝑎) ( 𝑊. Otherwise, we would have 𝑓(𝑎) ∩ 𝑊𝐶 ≠ Ø for all 𝑎 ∊ 𝐴, where 𝐶:= 𝑊𝐶 is a closed subset in 𝑌. Then, for (iii) we have 𝑓(𝑎) ( 𝑊𝐶 for all 𝑎 ∊ 𝐴 and hence 𝑓(𝐴) ∩ 𝑊= Ø, which is a contradiction. Therefore, 𝑊 is an open neighborhood in 𝑌 of 𝑓(𝑎0) and by the upper semicontinuity of 𝑓, there exists an open neighborhood
We also have the following theorem of minimal usco functions:
Theorem 3.11 (Characterization of minimal usco functions) Let (𝑋, τ) and (𝑌,σ) be topological spaces and 𝑓: 𝑋 → 𝑌 be a usco function. Then the following statements are equivalent:
𝑓 is a minimal usco function.
For every topological space (𝑍, 𝛼) and any continuous single-valued function 𝑔 : 𝑌→ 𝑍, 𝑔 ∘ 𝑓 is a minimal usco function of 𝑋 in 𝑍.
Proof. We will show 1) ⇒ 2).
(i) 𝑔 ∘ 𝑓 is compact and nonempty for all 𝑥 ∊ 𝑋. Indeed, as 𝑓 is a usco function, then 𝑓(𝑥) is compact for all 𝑥 ∊ 𝑋. Then 𝑔(𝑓(𝑥)) is compact since 𝑔 is continuous.
(ii) 𝑔 ∘ 𝑓 is upper semicontinuous. In fact, let 𝑥0 ∊ 𝑋 and 𝑊 be an open neighborhood in 𝑍 such that 𝑔(𝑓(𝑥0)) ⊆ 𝑊. By continuity, we have 𝐴:= 𝑔-1 (𝑊) is open in 𝑌 such that 𝑓(𝑥0) ⊆ 𝐴. Since 𝑓 is upper semicontinuous, there is a neighborhood open 𝑉 in 𝑋 of 𝑥0 such that 𝑓(𝑥) ⊆ 𝐴 for all 𝑥 ∊ 𝑉 and, therefore, 𝑔(𝑓(𝑥)) ⊆ 𝑊 for any 𝑥 ∊ 𝑉. So, we have shown that (𝑔 ∘𝑓) ∊ 𝒰 (𝑋, 𝑍).
(iii) It remains to show that 𝑔 ∘ 𝑓 is a minimal usco function. Using equivalences established in lemma 3.10, we must verify that 𝑔 ∘ 𝑓 satisfies (ii). Let 𝐴 be an open subset in 𝑋 and 𝑊 subset open in 𝑍 such that ((𝑔 ∘ 𝑓(𝑎)) ∩ 𝑊≠ Ø for some 𝑎 ∊ 𝐴. For continuity of 𝑔, we have 𝑔-1(𝑊) is an open subset in 𝑌 such that 𝑓(𝑎) ( 𝑔-1 (𝑊) (this is because ((𝑔 ∘ 𝑓)(𝐴)) ∩ 𝑊≠ Ø if and only if 𝑓(𝐴) ∩ 𝑔-1 (𝑊) ≠ Ø. Since 𝑓 is upper semicontinuous, there is an open neighborhood 𝑉 ⊆ 𝐴 of 𝑎 such that
𝑓(𝑥) ⊆ 𝑔-1 (𝑊) for any 𝑥 ∊ 𝑉
and, therefore, 𝑔(𝑓(𝑥)) ⊆ 𝑊 for any 𝑥 ∊ 𝑉 .
We will show 2) ⇒ 1). Let 𝑍=𝑌 and 𝑖𝑌: 𝑌→𝑌 be the identity function 𝑖𝑌(𝑦)=𝑦.
We observe that 𝑖𝑌 ∘ 𝑓 = 𝑓, which by hypothesis is a minimal usco function.
Definition 3.12 (Residual space) Let (𝑋, 𝛼) be a topological space. We said that 𝐺 ( 𝑋 is a residual set, if there is a family of open sets (𝒰𝑛)𝑛(ℕ which are dense in 𝑋 such that ⋂𝑛∊ℕ 𝒰𝑛 ⊆ 𝐺.
Definition 3.13 (Baire space) A Baire space is a topological space (𝑋, 𝛼) with the property that any residual is dense in 𝑋.
Regarding Baire space, we have the following well known examples (see 4):
1. Any locally compact topological space is a Baire space.
2. Any open subset of a Baire space is a Baire space.
3. Any complete metric space is a Baire space.
Theorem 3.14 Let (𝑋, 𝛼) be a Baire space and (𝑌, 𝑑) be a complete metric space. Then any minimal usco function 𝑓: 𝑋 → 𝑌 is single-valued in some residual subset 𝑅 of 𝑋.
Proof. Let 𝑓: 𝑋 → 𝑌 be a minimal usco function, we will see that, for every ε > 0, there exists an open set 𝑉 such that diameter of 𝑓(𝑉) is less than ε (𝑑𝑖𝑎𝑚(𝑓(𝑉)) < ε. In fact, let 𝑦 ∊ 𝑓(𝑋) and
Let 𝐴ε = ⋃{ 𝑉 ⊆ 𝑉 : 𝑋 𝑜p𝑒𝑛 and 𝑑𝑖𝑎𝑚(𝑓(𝑉)) < ε}. It is clear that this set is open. We will show that 𝐴ε is dense. Let 𝑥 ∊ 𝑋 and 𝑈 be an open neighborhood of 𝑥. If
and, therefore, 𝑉 ( 𝐴ε. We have shown that 𝑈 ∩ 𝐴ε ≠ Ø. If ε = 𝑛-1, we will write 𝐴𝑛 instead of
and, therefore, 𝑓(𝑥) consists of only one element. That is, 𝑓|𝑅 is a single-valued function.
3.1 Usco Functions and Continuity of Subdifferential of a Convex Function
Definition 3.1.1 (Dual space) Let X be a Banachspace. The topological dual of X is defined as the Banach space
If 𝑋 is a Banach space, the weak-* topology of the dual 𝑋′ denoted by σ (𝑋′, 𝑋) is generated by the family {𝑝𝑥: 𝑥 ∊ 𝑋} of seminorms, where
We note that (𝑋′, σ (𝑋′, 𝑋)) is a completely regular topological space. A fundamental weak neighborhood of 𝑎 ∊ 𝑋 is any set of the form
Where 𝐹′ is a finite subset of 𝑋′ and ε > 0.
Definition 3.1.2 (The subdifferential) Let Ω be a convex open subset of a Banach space 𝑋 and let 𝑓: Ω → ℝ ∪ {+∞} be a convex and continuous function, and let 𝑥 ∊ 𝑑𝑜𝑚𝑓. The subdifferential of 𝑓 in the point 𝑎 ∊ Ω, is the set of 𝑢 ∊ 𝑋′ such that
<mml:math style="font-family:'Times New Roman'"><mml:mfenced separators="|"><mml:mi>x</mml:mi></mml:mfenced><mml:mo>-</mml:mo><mml:mi>f</mml:mi><mml:mfenced separators="|"><mml:mi>a</mml:mi></mml:mfenced><mml:mi mathvariant="normal"> </mml:mi><mml:mi mathvariant="normal"> </mml:mi><mml:mo>≥</mml:mo><mml:mi mathvariant="normal"> </mml:mi><mml:mi mathvariant="normal"> </mml:mi><mml:mi>u</mml:mi><mml:mfenced separators="|"><mml:mrow><mml:mi>x</mml:mi><mml:mo>-</mml:mo><mml:mi>a</mml:mi></mml:mrow></mml:mfenced><mml:mo>=</mml:mo><mml:mi>u</mml:mi><mml:mfenced separators="|"><mml:mi>x</mml:mi></mml:mfenced><mml:mo>-</mml:mo><mml:mi>u</mml:mi><mml:mfenced separators="|"><mml:mi>a</mml:mi></mml:mfenced><mml:mi mathvariant="normal"> </mml:mi><mml:mi mathvariant="normal"> </mml:mi><mml:mi>f</mml:mi><mml:mi>o</mml:mi><mml:mi>r</mml:mi><mml:mi> </mml:mi><mml:mi>a</mml:mi><mml:mi>n</mml:mi><mml:mi>y</mml:mi><mml:mi mathvariant="normal"> </mml:mi><mml:mi mathvariant="normal"> </mml:mi><mml:mi>x</mml:mi><mml:mo>∈</mml:mo><mml:mi mathvariant="normal">Ω</mml:mi></mml:math>
We denote by ∂ 𝑓(𝑎) the subdifferential of 𝑓 at point 𝑎. Any function
Definition 3.1.3 Let 𝑋 be a Banach space and Ω be a convex open subset of 𝑋. A function 𝑓: Ω → ℝ called is radially differentiable at 𝑎 ∊ Ω in the direction ℎ if there exists the limit
The value of this limit is called the radial derivative of 𝑓 at point 𝑎 in the direction ℎ. We denote
the radial derivative of 𝑓 at point 𝑎 in the direction ℎ. The function 𝑑+𝑓(𝑎, ℎ) is well defined, it is sublineal, convex and continuous. If 𝑑+𝑓(𝑎, ℎ) exists for all ℎ ∊ 𝑋, then we say that 𝑓 is radially differentiable at 𝑎 ∊ Ω.
We know that every convex real valued function defined in an open is radially differentiable if
then
The radial derivative 𝑑±𝑓(𝑎, ℎ) it allows us to locate the subgradient at the point 𝑎 ∊ Ω in the following sense:
Lemma 3.1.4 Let 𝑋 be a Banach space, Ω be a convex open subset of 𝑋, 𝑓: Ω → ℝ be a convex and continuous function. Then 𝑢 ∊ ∂ 𝑓(𝑎) if and only if
Proof. If 𝑢 ∊ ∂ 𝑓(𝑎), then 𝑢: 𝑋 → ℝ is linear and continuous. Also if 𝑎+ 𝑡ℎ ∊ Ω for 𝑡 small enough and
<mml:math style="font-family:'Times New Roman'"><mml:mi>u</mml:mi><mml:mfenced separators="|"><mml:mrow><mml:mi>t</mml:mi><mml:mi>h</mml:mi></mml:mrow></mml:mfenced><mml:mo>=</mml:mo><mml:mi>t</mml:mi><mml:mi>u</mml:mi><mml:mfenced separators="|"><mml:mi>h</mml:mi></mml:mfenced><mml:mo>=</mml:mo><mml:mi>u</mml:mi><mml:mfenced separators="|"><mml:mrow><mml:mi>a</mml:mi><mml:mo>+</mml:mo><mml:mi>t</mml:mi><mml:mi>h</mml:mi></mml:mrow></mml:mfenced><mml:mo>-</mml:mo><mml:mi>u</mml:mi><mml:mfenced separators="|"><mml:mi>a</mml:mi></mml:mfenced><mml:mi mathvariant="normal"> </mml:mi><mml:mi mathvariant="normal"> </mml:mi><mml:mo>≤</mml:mo><mml:mi mathvariant="normal"> </mml:mi><mml:mi mathvariant="normal"> </mml:mi><mml:mi>f</mml:mi><mml:mfenced separators="|"><mml:mrow><mml:mi>a</mml:mi><mml:mo>+</mml:mo><mml:mi>t</mml:mi><mml:mi>h</mml:mi></mml:mrow></mml:mfenced><mml:mo>-</mml:mo><mml:mi>f</mml:mi><mml:mfenced separators="|"><mml:mi>a</mml:mi></mml:mfenced><mml:mo>.</mml:mo></mml:math>
Dividing by 𝑡 > 0 and taking 𝑡 → 0+, we have 𝑢(ℎ) ≤ 𝑑+𝑓(𝑎, ℎ). On the other hand,
and, therefore, when 𝑡 tends to zero by the left, we obtain that
Consequently,
and, therefore,
Remark 3.1.5 For a Banach space 𝑋, we consider in 𝑋′ the weak-* topology σ (𝑋′, 𝑋). If 𝑓: 𝑋→ ℝ is a convex and continuous function, we will show that ∂𝑓: 𝑋 → 𝑋′ is a usco function, considering in 𝑋 the norm topology and in 𝑋′ the weak-* topology.
Lemma 3.1.6 Let 𝑋 be a normed space, let Ω be an open and convex subset and 𝑓: Ω → ℝ be a convex and continuous function. Then,
1.The subdifferential function is a locally bounded function multivalued. In other words, for all 𝑎 ∊ Ω, exists 𝑟 > 0 and 𝑚 > 0 such that
2. If
For a proof of this lemma see the work by R Phelps 3.
Theorem 3.1.7 Let 𝑋 be a normed space, let Ω be an open and convex subset. If 𝑓: Ω → ℝ is a convex and continuous function, then 𝑥 ↦∂𝑓(𝑥) of 𝑋 in 𝑋′ is a usco function, considering in Ω the norm topology and in 𝑋′ the weak-* topology σ (𝑋′, 𝑋).
Proof. We must show that:
i. ∂𝑓(𝑥) is σ (𝑋′, 𝑋)-compact and ∂𝑓(𝑥) ≠ Ø for all 𝑥 ∊ Ω.
ii. ∂𝑓 is upper semicontinuous.
Let us show (i). Since 𝑓 is continuous at 𝑎 ∊ Ω and convex, then 𝑒𝑝𝑖(𝑓) ( 𝑋 ( ℝ is a convex set such that 𝑖𝑛𝑡(𝑒𝑝𝑖(𝑓)) ≠ Ø. In particular if (𝑎, 𝑓(𝑎)) ∉ 𝑖𝑛𝑡(𝑒𝑝𝑖(𝑓)), then under the Hahn-Banach theorem, it exists 𝑢 ∊ 𝑋′ which separates (but not strictly) the point (𝑎, 𝑓(𝑎)) of the convex set. We can observe that 𝑢 ∊ 𝑋′ is the subgradient of 𝑓 at 𝑎 and, therefore, ∂ 𝑓(𝑥) ≠ Ø for all 𝑥 ∊ Ω. Now, since ∂ 𝑓(𝑥) ( 𝑋′ is bounded, by the Banach-Alaoglu theorem, this set is σ (𝑋′, 𝑋)-relatively compact. Now, to conclude the proof we need to prove that σ (𝑋′, 𝑋)-closed. In addition, ∂ 𝑓(𝑥) is σ (𝑋′, 𝑋)-compact. We recall that, 𝑢 ∊ ∂ 𝑓(𝑥) if and only if 𝑢(𝑦) ≤ 𝑑+𝑓(𝑥, 𝑦) for 𝑦 ( 𝑋 (by lemma 1.3) and
Is σ (𝑋′, 𝑋)-closed since for each 𝑦 ∊ 𝑋, the function 𝑢 ↦ 𝑢 (𝑦) is σ (𝑋′, 𝑋)-continuous.
Let us show (ii). We will argue by contradiction, assuming that 𝑥 ↦ ∂𝑓(𝑥) is not upper semicontinuous in a 𝑎 ∊ Ω. This means that there exists a σ (𝑋′, 𝑋)-neighborhood open 𝑊 of ∂𝑓(𝑎), a sequence
Theorem 3.1.8 Let 𝑋 be a Banach space, Ω an open and convex subset and 𝑓: Ω → ℝ is a convex and continuous function. If 𝑓 is 𝐹-differentiable in 𝑎 ∊ Ω, then the subdifferential function 𝑥↦∂𝑓(𝑥) of Ω in 𝑋′ is upper semicontinuous at 𝑥 = 𝑎 with respect to the norm topologies.
For a proof of this theorem see 3.
Lemma 3.1.9 Let w be a normed space, Ω an open and convex subset and 𝑓:Ω→ℝ a continuous convex function. A sufficient condition for that 𝑓 to be 𝐹-differentiable in 𝑎∊Ω is that there is a continuous selection in 𝑎 with respect to the norm topologies.
For a proof of this lemma see 3.
4 The Asplund Spaces
We recall that a subset 𝐺 of a topological space (𝑋, 𝛼) is a set 𝐺δ if it can be expressed as a countable intersection of open and dense subsets. In complete metric spaces and so, in Banach spaces, any residual subset is dense by the Baire’s theorem.
Definition 4.1 (Asplund 𝐹-space/𝒢-space/𝒢𝒟-space) An Asplund 𝐹-space is a Banach space with the following property: If 𝑓 is a continuous convex function defined in open and convex subset 𝑈 ⊂ 𝑋, then 𝐹-differentiable in a dense subset 𝐺δ of 𝑈.
A 𝒢-Asplund is a Banach space with the following property: If 𝑓 is a continuous convex function defined in open and convex subset 𝑈 ⊂ 𝑋, then 𝐺-differentiable in a dense subset 𝐺δ of 𝑈.
An Asplund 𝒢𝒟-space is a Banach space with the following property: If 𝑓 is a continuous convex function defined in open subset 𝑈 ⊂ 𝑋, then 𝐺-differentiable in a dense subset 𝐺δ of 𝑈.
We note that a 𝒢-Asplund space is also an Asplund 𝒢𝒟-space. We want to point out that in the usual literature, the Asplund 𝐹-spaces are called the Asplund spaces, the 𝒢-Asplund are called weak Asplund spaces and Asplund 𝒢𝒟-spaces are called weak differenciable spaces. We could talk about the 𝐻-Asplund spaces, but it does not make much sense because it has been shown that continuous convex function wich are 𝐺-differentiables are 𝐻-differentiables. In the original article by E. Asplund 5, a 𝒢-Asplund is a weak differentiability spaces (𝑊𝐷𝑆) and a Asplund 𝐹-space is called strong differentiability spaces (𝑆𝐷𝑆). The definition we use comes from R. Phelps 6.
We will denote by 𝔘𝐺 the class of the 𝒢-Asplund and 𝔘𝐹 the class of Asplund 𝐹-spaces.
4.1 Asplund 𝐹-Spaces and Dentables Subset of a Banach Space
In this section we present the proof of the Namioka-Phelps theorem, which characterizes Asplund 𝐹-spaces by means of the geometric condition of dentability. First, we discuss the notion of the dentable set. To do that, if 𝑀 is a subset of a Banach space 𝑋, we define the support function of this set 𝑀 as follows:
of 𝑋′ in ℝ ∪{∞}. If we replace 𝑀 by
Suppose now that 𝑋 is a Banach space and 𝑀′ is a subset of 𝑋′. Proceeding as before, we define the support function of the set 𝑀′ as follows:
of 𝑋′ in ℝ ∪{∞}. If
Definition 4.1.2 (𝑋-slice) Let 𝑋 be a Banach space, 𝛼 > 0 and 𝑥 ∊ 𝑋. For a nonempty bounded subse t 𝑀 of 𝑋, we define the (𝛼, 𝑥′)-slice of 𝑀 as the set
Is clear that every slice is a relatively open set of 𝑀 with the topology σ (𝑋′, 𝑋), i.e., σ (𝑋, 𝑋′)|𝑀 (restriction of σ (𝑋, 𝑋′) to 𝑀).
Definition 4.1.3 (𝑋′-slice) Let 𝑋 be a Banach space, 𝛼 > 0 and 𝑥 ∊ 𝑋. For a nonempty bounded subset 𝑀′ of 𝑋′, we define the weak*- (𝛼, 𝑥)-slice of 𝑀′ as the set
We note that ℛ′(𝛼, 𝑥; 𝑀′) is open in the topology σ (𝑋′, 𝑋)|𝑀′. On the other hand, if 𝛼 < 𝛽, then
Moreover, if
Definition 4.1.3 (Dentable set) Let 𝑋 be a Banach space. A subset 𝑀 ⊂ 𝑋 is called dentable if it admits slices of arbitrarily small diameter. We say that a subset 𝑀′ ⊂ 𝑋′ is weak *-dentable if it admits weak *-(𝛼, 𝑥)-slice of arbitrarily small diameter. A Banach space is dentable if any bounded subset is dentable.
Theorem 4.1.4 (Namioka - Phelps). A Banach space 𝑋 is an Asplund 𝐹-space if and only if its dual 𝑋′ is weak *- dentable.
Proof. Suppose 𝑋 is an Asplund 𝐹-space and show that all bounded subset of 𝑋′ admits slices of arbitrarily small diameter, otherwise, out so there would be a bounded subset 𝑀′ of 𝑋′ in which all (𝛼, 𝑥)-slice has a diameter 𝑟 > 0.
Let
Of 𝑋 in ℝ is not 𝐹-differentiable in any point, which contradicts that 𝑋 is an Asplund 𝐹-space. With this purpose in mind, let
There exists a
We can now choose 𝛼 and 𝛽 appropriately. Let
Since inequality (5) holds for all 𝑡 > 0, in particular if
In virtue of theorem 2.4 with the bornology 𝛽𝐹, we conclude that 𝑝 is not 𝐹-differentiable in 𝑥, which is a contradiction.
Now, we will show that 𝑋 is an Asplund 𝐹-space. Let Ω be an open subset of 𝑋 and 𝑓: Ω → ℝ a convex function. Theorem 2.5 makes possible to locate the points of differentiability of 𝑓. The bornology that we are going to consider is 𝛽𝐹, of all bounded subsets 𝑋 since here the 𝐹-derivative is defined. Let 𝐷(𝑓, ε) be the set of 𝑥 ∊ Ω for which there is a δ (𝑥, ε) > 0 such that
Then:
(i) 𝐷(𝑓, ε) is open for all ε > 0.
(ii)
The proof of (ii) is obvious. We will proof (i), i.e., 𝐷(𝑓, ε) is open. Since 𝑓 is locally Lipschitzian by theorem 1.1, there exists a 𝑟 > 0 and a constant 𝑘 > 0 such that 𝐵 (𝑥, 𝑟) ( Ω and
By the definition of 𝐷(𝑓, ε), there exists δ > 0 such that
Let
Now let us observe that if 𝑡 = δ, then
Let
Once the points of 𝐹-differentiability of 𝑓 are located, it remains to show that 𝐴ε: = 𝐷(𝑓, ε) is dense for every ε > 0. On the one hand, we have that Ω is a Baire space. If ε = 𝑛−1, we will write
(set of differentiability of 𝑓) is a set 𝐺δ.
We will show that 𝐷(𝑓, ε) is dense for every ε > 0. That is, if 𝑥0 ( Ω (fixed but arbitrary) any open neighborhood 𝑊 of 𝑥0 intersects 𝐷(𝑓, ε). As the subdifferential function is locally bounded, we can suppose that ∂ 𝑓(𝑊) = 𝛵′ is a bounded subset of 𝑋′. As 𝑋′ is weak *-dentable, there exists a weak- *(𝛼, 𝑧)-slice of 𝛵′ of arbitrarily small diameter, let say < ε. If 𝑎′∊ ℛ(𝛼, 𝑧; 𝛵′), then 𝑎′∊ ∂ 𝑓(𝑎) for some 𝑎 in 𝑊, since
Since 𝑊 is open, there exists 𝑟 > 0 (sufficiently small) such that 𝑏:= 𝑎 + 𝑟𝑧 ∊ 𝑊. If 𝑏′ ∊ ∂ 𝑓(𝑏), therefore,
consequently,
As 𝑎 − 𝑏 = −𝑟𝑧 and 𝑏 − 𝑎 = 𝑟𝑧 by replacing in (6)
and how 𝑟 > 0
We have, thus, shown that
Suppose that
If
Now, as 0 < 𝑡 < δ and by the linearity of 𝑢, 𝑣, 𝑤 we have
By (7) we know that 𝑣, 𝑤 ∊ ℛ′(𝛼, 𝑧; 𝛵′) and that the diameter of this set is < ε, then we obtain
This last inequality is true for every ℎ ∊ 𝑋 such that ‖ℎ‖𝑋 ≤ 1 and every 0 < 𝑡 < δ. Therefore, 𝑏 ∊ 𝐷(𝑓, ε), i.e., 𝑊∩ 𝐷(𝑓, ε) ≠ Ø. This means that 𝐷(𝑓, ε) is dense and in consequence
(set of differentiability of 𝑓) is a 𝐺δ-dense set. We have, thus, proved that 𝑋 is an Asplund 𝐹-space.
5 The Class of Stegall 𝕾 on Topological Spaces
This section is the central part of the article, since it introduces an intermediate class between the Asplund 𝐹-spaces and 𝒢-Asplund, called the class of Stegall
Let us recall that a topological space (𝑋, τ) is completely regular, if it is Hausdorff and for each closed set 𝐶 and every point 𝑝 that does not belong to 𝐶, there is a continuous function 𝑓: 𝑋 → [0,1] such that 𝑓|𝐶 = 0 and 𝑓(𝑝) = 1. The Urysohn’s lemma, every metric space is completely regular.
Theorem 5.1 If 𝑋 is a separable Banach space and 𝐾 ⊂ 𝑋′ is bounded, then the topology σ (𝑋′, 𝑋)|𝐾 is metrizable (the topology of the dual is metrizable on the bounded sets).
Theorem 5.2 Let 𝑋 be a Banach space. Then (𝑋′, σ (𝑋′, 𝑋)) is a space completely regular.
For the proof of previuos theorems see 1.
Definition 5.3 (𝑆-space) An 𝑆-space is a completely regular topological space 𝑋 that satisfies the following condition:
If 𝑍 is a Baire space and 𝑓: 𝑍 → 𝑋 is a minimal usco function, then 𝑓 is single-valued in a residual subset of 𝑍. We will denote by 𝔖, the set of all the 𝑆-spaces.
Even though Marian J. Fabian in 1 proves the following theorems, some important details were left out. We present here complete proofs.
Theorem 5.4 Every metric space (𝑋, 𝑑) is an 𝑆-space.
Proof. Let (𝑋, 𝑑) be a metric space, 𝑍 be a Baire space and 𝑓: 𝑍 → 𝑋 is a minimal usco function. For each 𝑛 ∊ ℕ, we define the open set
We will show that
If
Then, Ω ⊂ 𝒰𝑛 and, therefore, Ø ≠ 𝑉𝑥 ∩ 𝒰𝑛, which proof the density of the 𝒰𝑛. So
Theorem 5.5 Let 𝑋 be a completely regular space. If
Proof. Let 𝑍 be a Baire space and 𝑓: 𝑍 → 𝑋 a minimal usco function. For each 𝑛 ∊ ℕ, we define the set
Since 𝑓 is a upper semicontinuous function and by hypothesis each 𝑋𝑛 is closed in 𝑋, then each 𝑍𝑛 is a closed set in 𝑍 by theorem 3.5. If
𝐴𝑛 ≠ for some 𝑛 ∊ ℕ
Let 𝑛 ∊ ℕ be such that 𝐴𝑛 ≠ Ø, and let
1) In general, if 𝑓: 𝑍 → 𝑋 is a usco function and 𝑔: 𝑋 → 𝑌 (with 𝑌 a regular space) is a continuous single-valued function, by theorem 3.11.2, we have that 𝑔 ∘ 𝑓 is a usco function. From this it follows that if 𝑋 is a subspace of 𝑌, then 𝑓|𝑋 is a usco function since
2) We will show that 𝑓𝑛 is minimal. For this we will use lemma 3.10 several times. Let
3)
Then
and as 𝑓 is minimal, by lemma 3.10.(iii) then 𝑓(𝑧) ( 𝐶 for every 𝑧 ∊ 𝑉. So
and, hence, 𝑓𝑛 is minimal (using lemma 3.10 again).
1) 𝑓𝑛(𝐴𝑛) ( 𝑋𝑛. By the definition 𝑍𝑛, we have
Let
𝑅: = {𝑧 ∊ 𝑍: 𝑓(𝑧) is a singleton 𝑓 or all 𝑧}
And 𝑅𝑛: = 𝑅 ∩ 𝐴𝑛.
We note that 𝑅𝑛 is the set of 𝑧 ∊ 𝐴𝑛 such that
Then
Since 𝑅 ∩ 𝐴𝑛 is residual in 𝐴𝑛, there exist a countable collection
The set
This is the final part of the proof. Let
We have, thus, shown that
As
Since
Using (9), we obtain that
In other words, we have shown that 𝑅 is residual, since each 𝛵𝑘 is open and dense, accordingly 𝑋 = ⋃𝑛∊ℕ 𝑋𝑛 ∊ 𝔖.
5.1 The Class of Stegall
The class
Next, we give a detailed proof of the Stegall’s theorem whose proof can be found in 1 citing numerous articles. We show that the usco multivalued functions, particulary the subdifferentials of convex functions, play an important role in this proof.
Theorem 5.1.1 The Stegall class lies between the Asplund 𝐹-spaces and the 𝒢-Asplund. In other words,
Proof. Let us first show that
A weak-*(𝛼, 𝑥)-slice of 𝑀 is a set σ (𝑋′, 𝑋)|𝑀 open due to the continuity of the function 𝑥′ ↦ ⟨ 𝑥, 𝑥′⟩ (i.e., the continuity of the support function).
Suppose 𝑋 is an Asplund 𝐹-space and let
(since 𝑋′ is fitted with the topology
Our aim now is to show that 𝛵 is single-valued in a residual subset. For this, let us consider the set defined in (8).
We claim that 𝒰𝑛 is an open set in (𝑀, τ), where τ = σ (𝑋′, 𝑋)|𝑀. In fact, If 𝑤 ∊ 𝑉, then 𝑉 is an open neighborhood of 𝑤.
Now, we claim that 𝒰𝑛 is dense in 𝑍. To prove this statement, let 𝒰 be an open and nonempty subset of 𝑍. Recall that 𝑋 is an Asplund space, so from theorem 1.1, 𝑋′ is weak-*-dentable. Now, since 𝛵(𝒰) ( 𝑀, then given ε > 0 there exists a weak-*(𝛼, 𝑥)-slice ℛ(𝛼, 𝑥; 𝛵(𝒰)) of arbitrarily small diameter, let say < ε (i.e., 𝑑𝑖𝑎𝑚[ ℛ(𝛼, 𝑥; 𝛵(𝒰))] < ε). If
then 𝑊 is a subset τ-open of 𝑀 and
By lemma 3.10.(ii), there exists an open nonempty subset 𝑆 ( 𝒰 such that 𝛵(𝑆) ( 𝑊 and, therefore, 𝛵(𝑆) ( 𝛵(𝒰) and 𝛵(𝑆) ( 𝑊 ∩ 𝛵(𝒰) = ℛ(𝛼, 𝑥; 𝛵(𝒰)). So that,
which shows that 𝒰𝑛 is dense in 𝑍. From previous claims, we have that {𝒰𝑛: 𝑛∊ℕ} is a countable collection of open and dense subsets in 𝑍. We remark that
Indeed, if
that is, 𝛵(𝑧) is a singleton for all 𝑧 from the dense residual subset 𝒟 of 𝑍. We have, thus, shown that 𝛵 is single-valued in a residual subset of 𝑍, and so
Now, we will establish that
By theorem 3.1.5, the subdifferential ∂ ψ is a usco function considering in Ω the norm topology and in 𝑋′ the weak-* topology σ (𝑋′, 𝑋). By theorem 3.7, there exists a minimal usco function 𝛵: Ω → 𝑋′ such that 𝛵 ≼ ∂ ψ. Since Ω is a Baire space, (𝑋′, σ (𝑋′, 𝑋)) is completely regular and 𝛵: Ω → 𝑋′ is a minimal usco function, then 𝛵 is single-valued on a residual subset 𝒟 of Ω. It remains to verify that ψ is 𝐺-differentiable in 𝒟. Let 𝑥 ∊ 𝒟 (arbitrary but fixed) and ℎ ∊ 𝑋 . Then 𝛵(𝑥) = {𝑢}. As }.(𝑥) ( ∂ ψ (𝑥), then 𝑢 ∊ ∂ ψ (𝑥) and, therefore, for a 𝑡 > 0 small enough we have to
Let
which implies that
Since 𝑡 > 0, from (10) and (11), we obtain that
and, therefore,
Now, let ε > 0 be such that
is a σ (𝑋′, 𝑋)-open neighborhood 𝑢 ∊ 𝑋′. Using the upper semicontinuity of 𝛵 in 𝑥 ∊ Ω, there exists an open neighborhood 𝑉 ( Ω of 𝑥 such that
There exists a δ > 0 such that
In virtue of (12) and (13), we conclude that
Since 𝑢 ∊ 𝑋′, so
We have, thus, shown that ψ is 𝐺-differentiable in a 𝐺δ-dense subset 𝒟 and hence 𝑋∊ 𝔘𝐺.
Proposition 5.1.2 Every separable Banach space belongs to the Stegall class.
Proof. Suppose 𝑋 is a separable Banach space. Let
By the Banach-Alaoglu theorem, the closed unitary ball 𝐵′ de 𝑋′ is σ (𝑋′, 𝑋)-compact, then by theorem 5.1 we have that (𝐵𝑋′, τ) is metrizable and (𝐵𝑋′, τ) ∊ 𝔖. As
Then (𝑋′, σ (𝑋′, 𝑋)) ∊ 𝔖 by theorem 5.5 and therefore
Remarks 5.1.3
1) In general
2) In general
3) There exist Banach spaces that are not 𝒢-Asplund spaces, for example, the nonseparable Banach space 𝑋 = 𝓁∞( ℝ) is not a 𝒢-Asplund space.
5.2 Example of Usco Function
The Supremum Mapping
As it is well known, the problem of the differentiability of convex functions is well addressed by Fabian in 1. Let 𝑋 be a compact and Hausdorff topological space, and 𝒞(𝑋) be the Banach space of the continuous functions of 𝑋 in ℝ with the norm
φ: 𝒞(𝑋) → ℝ
This function is sublineal and, therefore, convex. In addition, we define
ψ: 𝒞(𝑋) → 𝑋
This is a multivalued function that assigns to each function 𝑓 the set of points 𝑡 ∊ 𝑋 in which the function attains the supremum. We will call ψ the supremum mapping. We remark that the function 𝑓 ↦ ψ(𝑓) is a usco function. Indeed,
I. It is clear that ψ(𝑓) is compact because
is closed in 𝑋. As 𝑋 is compact,
and, therefore, ψ(𝑓) ≠ Ø for all 𝑓 ( 𝒞(𝑋).
II. We will show that ψ is upper semicontinuous. We argue by contradiction assuming that for some point 𝑔 ∊ 𝒞(𝑋) there is an open neighborhood ψ(𝑔) ( 𝑊 such that on every open ball 𝐵(𝑔, ε) there exists a ℎ such that ψ(ℎ) is not contained in 𝑊. That is, there exists a sequence (𝑔𝑛)𝑛∊ℕ such that
As 𝑊𝐶 is compact, the sequence (𝑡)𝑛∊ℕ has limit value 𝑡 ∊ 𝑊𝐶. Since 𝑔 is a continuous function, 𝑔(𝑡) is an adhesion value of the sequence (𝑔(𝑡)) 𝑛∊ℕ in ℝ. Therefore, there is a subsequence
In summary, ψ is a usco function. We emphasize that this example is closely related to the differentiability of the supremum norm in 𝒞(𝑋).