Convex regularization and subdifferential calculus

If (?)\holds, then (?) entails

and (1) follows. Conversely, observe that

and \overline{\limfunc{co}}^{w^{\ast \ast }}(f+g)\leq f+g, so \overline{ \limfunc{co}}^{w^{\ast \ast }}(f+g) is proper. Then, taking the conjugate in (1) and applying twice Moreau’s theorem in (X^{\ast },X^{\ast \ast }), we obtain

(i) If f\in \Gamma _{0}(X), then \partial _{\alpha }f(x)\neq \emptyset for all \alpha >0.

(ii) If f^{\ast } is proper, then f is also proper. Moreover, for every x\in \limfunc{dom}f and x^{\ast }\in \limfunc{dom}f^{\ast }, we have f^{\ast }(x^{\ast })+f(x)-\left\langle x^{\ast },x\right\rangle \geq 0. So, every scalar \alpha such that \alpha >f^{\ast }(x^{\ast })+f(x)-\left\langle x^{\ast },x\right\rangle satisfies x^{\ast }\in \partial _{\alpha }f(x), and we deduce \varepsilon _{f}(x)\leq \alpha .

(iii) Since \overline{\limfunc{co}}^{w^{\ast \ast }}f is proper, its conjugate f^{\ast }=(\overline{\limfunc{co}}^{w^{\ast \ast }}f)^{\ast } is also proper (e.g., [CHL23, Proposition 3.1.4]). Take x^{\ast }\in \limfunc{dom}f^{\ast } and pick any \rho >0. Then there exists some x_{\rho }\in \limfunc{dom}f such that

that is, x^{\ast }\in \partial _{\rho }f(x_{\rho }) and

Hence, \inf_{z\in \limfunc{dom}f}\varepsilon _{f}(z)=0 by the arbitrariness of \rho >0.

The equivalence between (i) and (iii) is established in [CGH12, Proposition 5.5], and we prove next that (i)\Longleftrightarrow (ii), (iv)\implies (i) and (iii)\implies (iv).

(i)\implies (ii). Assume that (i) holds or, equivalently according to Lemma 1,

Fix x\in \limfunc{dom}f\cap \limfunc{dom}g and \varepsilon >\varepsilon _{f+g}(x) (remember that \varepsilon _{f+g}(x)<\infty by Lemma lema2(ii)). Take x^{\ast }\in \partial _{\varepsilon }(f+g)(x) and a \theta-neighborhood U in (X^{\ast },\tau ). Since \varepsilon >\varepsilon _{f+g}(x), there exists some 0<\alpha <\varepsilon such that \partial _{\alpha }(f+g)(x)\neq \emptyset ; that is, \cup _{0<\delta <\varepsilon }\partial _{\varepsilon -\delta }(f+g)(x)\neq \emptyset . So, by applying [CHL23, Proposition 4.1.9] to the function \hat{f}+\hat{g} defined on X^{\ast \ast }, we obtain

Hence, we find \delta \in (0,\varepsilon ) and y^{\ast }\in \partial _{\varepsilon -\delta }(f+g)(x) such that x^{\ast }-y^{\ast }\in U. Moreover, since (remember (Proof))

we find some z^{\ast }\in y^{\ast }+U such that

In other words, there are z_{1}^{\ast }, z_{2}^{\ast }\in X^{\ast } such that z_{1}^{\ast }+z_{2}^{\ast }=z^{\ast } and

that is, rearranging these terms,

Now, we denote

so that \varepsilon _{1}\geq 0 and \varepsilon _{2}\geq g(x)+g^{\ast }(z_{2}^{\ast })-\left\langle z_{2}^{\ast },x\right\rangle \geq 0, and we deduce that \varepsilon _{1}+\varepsilon _{2}=\varepsilon , z_{1}^{\ast }\in \partial _{\varepsilon _{1}}f(x) and z_{2}^{\ast }\in \partial _{\varepsilon _{2}}g(x). Thus,

and (iii) follows by the arbitrariness of U.

(ii)\implies (i). As observed in the proof of Lemma 2, remember that the function \overline{\limfunc{co}}^{w^{\ast \ast }}(f+g) is proper and, so, its conjugate (f+g)^{\ast } is also proper. First, we are going to prove that, for every given x^{\ast }\in \limfunc{dom}(f+g)^{\ast },

Fix \varepsilon >0 and choose x_{\varepsilon }\in X such that

that is, x^{\ast }\in \partial _{\varepsilon }(f+g)(x_{\varepsilon }) and \varepsilon _{f+g}(x_{\varepsilon })<\varepsilon . So, by (ii) we have

which reads, due to the inclusions \partial _{\varepsilon _{1}}f(x_{\varepsilon })\subset \partial _{\varepsilon _{1}}(\overline{ \limfunc{co}}^{w^{\ast \ast }}f)(x_{\varepsilon }) and \partial _{\varepsilon _{2}}g(x_{\varepsilon })\subset \partial _{\varepsilon _{2}}( \overline{\limfunc{co}}^{w^{\ast \ast }}g)(x_{\varepsilon }),

that is,

At the same time, due to (Proof), the sets \partial _{\varepsilon _{1}}f(x_{\varepsilon }) and \partial _{\varepsilon _{2}}g(x_{\varepsilon }) are nonempty for some \varepsilon _{1},\varepsilon _{2}\geq 0 such that \varepsilon _{1}+\varepsilon _{2}=\varepsilon , thus (e.g., [CHL23, Exercise 62])

and (Proof) and (Proof) entail

Hence, (Proof) follows by the arbitrariness of \varepsilon. Moreover, since the inequality ((\overline{\limfunc{co}}^{w^{\ast \ast }}f)+(\overline{\limfunc{co}}^{w^{\ast \ast }}g))^{\ast }\geq (f+g)^{\ast } always holds, we deduce that ((\overline{\limfunc{co}}^{w^{\ast \ast }}f)+( \overline{\limfunc{co}}^{w^{\ast \ast }}g))^{\ast }=(f+g)^{\ast }. Therefore (i) follows by applying Moreau’s theorem to the last equality in the pair (X^{\ast },X^{\ast \ast }).

(iii)\implies (iv). For any x\in X and \varepsilon >0, (iii) implies that

which is (iv).

(iv)\implies (i). Fix x\in X and \varepsilon >0. Since \partial _{2\varepsilon }f(x)\subset \partial _{2\varepsilon }(\overline{\limfunc{co}} ^{w^{\ast \ast }}f)(x) and \partial _{2\varepsilon }g(x)\subset \partial _{2\varepsilon }(\overline{\limfunc{co}}^{w^{\ast \ast }}g)(x), (iv) entails

Now, since (f+g)^{\ast }=(\overline{\limfunc{co}}^{w^{\ast \ast }}(f+g))^{\ast } is proper, by applying [CGH18, Corollary 3.3 (see, also, Remark 3.2(i))] in X^{\ast \ast } (see, also, [CHP17, CHS16, LoVo14]) we deduce that

for some constant c\in \mathbb{R}. It is clear that c\geq 0 and let us check that c=0. Since \overline{\limfunc{co}}^{w^{\ast \ast }}(f+g) is proper, according to Lemma 2(iii), for every \varepsilon >0 there exists some x_{\varepsilon }\in X such that \varepsilon _{f+g}(x_{\varepsilon })<\varepsilon ; that is, \partial _{\varepsilon }(f+g)(x_{\varepsilon })\neq \emptyset . Thus, by (Proof), the sets \partial _{2\varepsilon }f(x_{\varepsilon }) and \partial _{2\varepsilon }g(x) are nonempty, and so

Thus, (Proof) reads

and taking \varepsilon \downarrow 0 we conclude that c=0.

Pick z\in X^{\ast \ast }, x^{\ast }\in \partial \bar{f}^{w^{\ast \ast }}(z), U\in z+\mathcal{N} , and choose \varepsilon >0 and V\in z+ \mathcal{N} such that V+\sqrt{\varepsilon }B_{X}\subset U. Then

and, so, there exists some x_{\varepsilon }\in V such that

Hence, x^{\ast }\in \partial _{\varepsilon }f(x_{\varepsilon }) and the Brondsted-Rockafellar theorem (see, e.g., [CHL23, Proposition 4.3.7]) gives some z_{\varepsilon }\in x_{\varepsilon }+\sqrt{\varepsilon }B_{X} and z_{\varepsilon }^{\ast }\in x^{\ast }+\sqrt{\varepsilon }B_{X^{\ast }} such z_{\varepsilon }^{\ast }\in \partial f(z_{\varepsilon }). Thus, z_{\varepsilon }\in V+\sqrt{\varepsilon }B_{X}\subset U and

By the arbitrariness of \varepsilon we deduce that x^{\ast }\in \limfunc{ cl}^{\left\Vert \cdot \right\Vert _{\ast }}\left( \cup _{x\in U}\partial f(x)\right) and the arbitrariness of U yields the inclusion \subset in (4).

Conversely, if x^{\ast } is in the right hand side of (4),\then for each U\in \mathcal{N} and \varepsilon >0 there exist some x\in z+U and z^{\ast }\in \partial f(x) such that \left\Vert x^{\ast }-z^{\ast }\right\Vert _{\ast }\leq \varepsilon and f^{\ast }(x^{\ast })\leq f^{\ast }(z^{\ast })+\varepsilon , due to the \left\Vert \cdot \right\Vert _{\ast }-lower semicontinuity of f^{\ast }.\Hence,

and, by taking the supremum over U,

Thus, as \varepsilon \downarrow 0 we obtain that x^{\ast }\in \partial f^{w^{\ast \ast }}(z).

(vi)\implies (v): Given x\in X and x^{\ast }\in \partial (f+g)(x), we let the sequences (x_{n}), (y_{n})\subset X, x_{n}^{\ast }\in \partial f(x_{n}) and y_{n}^{\ast }\in \partial g(y_{n}) be as in (vi). Then, given any \varepsilon >0, for all n large enough we have

Hence, for any y\in \limfunc{dom}f,

and we deduce that x_{n}^{\ast }\in \partial _{\varepsilon }f(x). Similarly, y_{n}^{\ast }\in \partial _{\varepsilon }g(x)\and, so,

entailing

Thus, for each \delta >0,

and we obtain

implying

Now, applying Lemma 4 successively to f+g and \overline{f} ^{w^{\ast \ast }}+\overline{g}^{w^{\ast \ast }}, we conclude that, for all z\in X^{\ast \ast },

Consequently, by applying Rockafellar’s theorem (see Proposition 9 below) in the Banach space X^{\ast \ast }, the proper w^{\ast \ast }-lsc (hence, \left\Vert \cdot \right\Vert _{\ast \ast }-lsc) convex functions \overline{f+g}^{w^{\ast \ast }} and \overline{f}^{w^{\ast \ast }}+ \overline{g}^{w^{\ast \ast }} coincide up to some additive constant.\ Thus, (?) follows as these two functions coincide with f+g on X.

(v)\implies (vi): Fix x\in X, x^{\ast }\in \partial (f+g)(x), and choose any sequence \varepsilon _{n}\downarrow 0. Then, taking into account Theorem 3(iv) (applied with \tau =\left\Vert \cdot \right\Vert _{\ast }), for each fixed n assertion (v) implies that

that is, there are u_{n}^{\ast }\in \partial _{\frac{(\varepsilon _{n})^{2} }{2}}f(x) and v_{n}^{\ast }\in \partial _{\frac{(\varepsilon _{n})^{2}}{2} }g(x) such that

We introduce the function \varphi \in \Gamma _{0}(X\times X) defined as \varphi (y,z):=f(y)+g(z). Since it can be easily check that (u_{n}^{\ast },v_{n}^{\ast })\in \partial _{(\varepsilon _{n})^{2}}\varphi (x,x), by applying the Brondsted-Rockafellar theorem (see, e.g., [CHL23, Proposition 4.3.7]) in the Banach product space X\times X (endowed with max norm), we find x_{n}, y_{n}\in B_{\varepsilon _{n}}(x), \alpha _{n}\in \lbrack -1,1], y_{n}^{\ast }, z_{n}^{\ast }\in B_{X^{\ast }}, x_{n}^{\ast }\in \partial f(x_{n}) and y_{n}^{\ast }\in \partial g(y_{n}) such that

Thus, by (Proof) and the triangle inequality,

and we deduce that x_{n}^{\ast }+y_{n}^{\ast }\overset{\left\Vert \cdot \right\Vert _{\ast }}{\rightarrow }x^{\ast }. Moreover, thanks to (Proof ), the lower semicontinuity of f and g implies that f(x_{n})\rightarrow f(x) and g(y_{n})\rightarrow g(x). At the same time, since

and

the inequality in (Proof) yields \left\langle x_{n}^{\ast },x_{n}-x\right\rangle \rightarrow 0, as well as \left\langle y_{n}^{\ast },y_{n}-x\right\rangle \rightarrow 0. So, the sequences (x_{n}), (y_{n}) , (x_{n}^{\ast }) and (y_{n}^{\ast }) satisfy the properties required in assertion (vi).

Under the current conditions, for every x\in \limfunc{dom}f\cap \limfunc{ dom}g and \varepsilon \geq 0 we have the following (exact) calculus rule (see, e.g., [CHL23, Theorem 2.8.7] and [Za02, Theorem 2.8.7]):

So, all the conditions of Theorem 3 are satisfied.

According to Corollary 6, we have that

The first equality is an immediate consequence of Corollary 6. To prove the second one we observe first that \overline{f}^{w^{\ast \ast }}, \overline{g}^{w^{\ast \ast }}\in \Gamma _{0}(X^{\ast \ast }), due to existence of continuous affine minorants for f and g in X. Then, thanks to Theorem 3, the functions \overline{f}^{w^{\ast \ast }} and \overline{g}^{w^{\ast \ast }} satisfy (?); i.e., for all x\in X^{\ast \ast } and \varepsilon \geq 0,

where

Take x^{\ast }\in \partial _{\varepsilon }\left( \overline{f}^{w^{\ast \ast }}+\overline{g}^{w^{\ast \ast }}\right) (x)\and consider any net (\alpha _{V})_{V\in \mathcal{N}}\subset ]0,1[ such that \alpha _{V}\downarrow \varepsilon (here, \mathcal{N}:=\mathcal{N}_{X^{\ast }}(\tau ) is partially-ordered by descending inclusions). Hence, for all V\in \mathcal{N} , (Proof) implies that x^{\ast }\in \limfunc{cl}\nolimits^{\tau }\left( A_{\alpha _{V}}\right) and we can find x_{V}^{\ast }\in A_{\alpha _{V}} such that x_{V}^{\ast }\in x^{\ast }+V. Given any W\in \mathcal{N} , by taking V_{0}=W we see that

that is, (x_{V}^{\ast }) is \tau-convergent to x^{\ast }. Moreover, for each V\in \mathcal{N} there are \varepsilon _{V,1}, \varepsilon _{V,2}\geq 0, y_{V}^{\ast }\in \partial _{\varepsilon _{V,1}}\overline{f} ^{w^{\ast \ast }}(x) and z_{V}^{\ast }\in \partial _{\varepsilon _{V,2}} \overline{g}^{w^{\ast \ast }}(x) such that \varepsilon _{V,1}+\varepsilon _{V,2}=\alpha _{V} and x_{V}^{\ast }=y_{V}^{\ast }+z_{V}^{\ast }. By the current continuity hypothesis, we choose x_{0}\in \limfunc{dom}g and \theta-neighborhood U_{0}\subset X such that

Then, for all u\in U_{0} (\subset X),

At the same time, there exists some M\geq 0 such that

otherwise, since \left\langle x_{V}^{\ast },x_{0}-x\right\rangle =\left\langle y_{V}^{\ast }+z_{V}^{\ast },x_{0}-x\right\rangle converges to \left\langle x^{\ast },x_{0}-x\right\rangle and z_{V}^{\ast }\in \partial _{\varepsilon _{V,2}}\overline{g}^{w^{\ast \ast }}(x), we get a contradiction with

Therefore (Proof) reads

and, thanks to the Alaoglu-Bourbaki theorem, we may assume that (y_{V}^{\ast })_{V} is \sigma (X^{\ast },X)-convergent to some y^{\ast }\in X^{\ast }; hence, (z_{V}^{\ast })_{V} is \sigma (X^{\ast },X) -convergent to x^{\ast }-y^{\ast }. We can also assume that \varepsilon _{V,1}\downarrow \varepsilon _{1} and \varepsilon _{V,2}\downarrow \varepsilon _{2} with \varepsilon _{1}+\varepsilon _{2}=\varepsilon . More precisely, since for all y\in X and V\in \mathcal{N} we have

by taking the limit on V we infer that

Thus, knowing that X is \sigma (X^{\ast \ast },X^{\ast })-dense in X^{\ast \ast }, we deduce that

that is, y^{\ast }\in \partial _{\varepsilon _{1}}\overline{f}^{w^{\ast \ast }}(x). Similarly, we show that x^{\ast }-y^{\ast }\in \partial _{\varepsilon _{2}}\overline{g}^{w^{\ast \ast }}(x), and we conclude that

Thus, we are done because the inclusion \supset in (8) is straightforward.

(i) Assume that f is continuous at x_{0}\in \limfunc{dom}f. Given x\in X, we introduce the functions \varphi , \psi \in \Gamma _{0}( \mathbb{R}) defined as

Since 0\in \limfunc{int}(\limfunc{dom}\psi ), [CHL23, Lemma 4.4.2] entails

which in turn implies f(x)=g(x)+c_{0}.

(ii) Assume that X is Banach and fix x_{0}\in \limfunc{dom}f\cap \limfunc{dom}g together with r>\left\Vert x_{0}\right\Vert . Given z\in X^{\ast \ast }, by Theorem 8 (applied twice) and Lemma lembb, we have that

where \mathcal{N}:=\mathcal{N}_{X^{\ast \ast }}(w^{\ast \ast }). Hence, using the current assumption,

and, proceeding as above, we conclude that

Moreover, using Moreau’s theorem, we have that

and similarly (\bar{g}^{w^{\ast \ast }}+\mathrm{I}_{rB_{X^{\ast \ast }}})^{\ast }=g^{\ast }\square (r\left\Vert \cdot \right\Vert _{\ast })=:g_{r}; hence, for all x^{\ast }\in X^{\ast },

Therefore, since f_{r} is (norm-)continuous, by (i) there exists some c_{r}\in \mathbb{R} such that

So, taking the conjugate with respect to the pair (X,X^{\ast }) and using Moreau’s theorem, we obtain

In particular, c_{r}=f(x_{0})-g(x_{0})=:c_{0} and we deduce that f+ \mathrm{I}_{rB_{X}}=g+\mathrm{I}_{rB_{X}}+c_{0}, for all sufficiently large r; that is, f=g+c_{0}.