Theorem 18.3.8 (Fenchel-Moreau).label Let $\dpn{E, F}{\lambda}$ be a duality over $\real$, and $f: E \to (-\infty, \infty]$ with $f \ne \infty$ and $f^{*}\ne \infty$, then:

  1. (1)

    For each $x \in E$,

    \[f^{**}(x) = \sup\bracs{\dpn{x, \phi}{\lambda} - \alpha|(\phi, \alpha) \in F \times \real, (\phi, \alpha) \le f}\]

  2. (2)

    $\text{epi}(f^{**})$ is the $\sigma(E \times \real, F \times \real)$-closed convex hull of $\text{epi}(f)$.

  3. (3)

    $f^{**}$ is the greatest convex and $\sigma(E, F)$-lower semicontinuous function bounded above by $f$.

  4. (4)

    $f = f^{**}$ if and only if $f$ is convex and $\sigma(E, F)$-lower semicontinuous.

Proof. (1): Let $\phi \in F$ such that $f^{*}(\phi) < \infty$, then for each $x \in E$,

\[\dpn{x, \phi}{\lambda}- f^{*}(\phi) \le f(x)\]

so $\dpn{\cdot, \phi}{\lambda}- f^{*}(\phi) \le f$, and

\begin{align*}f^{**}(x)&= \sup_{\phi \in F}\dpn{x, \phi}{\lambda}- f^{*}(\phi) = \sup_{\substack{\phi \in F \\ f^*(\phi) < \infty}}\dpn{x, y}{\lambda}- f^{*}(\phi) \\&\le \sup\bracs{\dpn{x, \phi}{\lambda} - \alpha|(\phi, \alpha) \in F \times \real, (\phi, \alpha) \le f}\end{align*}

On the other hand, let $\phi \in F$ and $\alpha \in \real$ such that $(\phi, \alpha) \le f$, then $f^{*}(\phi) \le \alpha$, and

\[f^{**}(x) \ge \dpn{x, \phi}{\lambda}- f^{*}(\phi) \ge \dpn{x, \phi}{\lambda}- \alpha\]

Therefore

\[f^{**}(x) \ge \sup\bracs{\dpn{x, \phi}{\lambda} - \alpha| \phi \in F, \alpha \in \real, \dpn{\cdot, y}{\lambda} - \alpha \le f}\]

(2): By Lemma 18.3.4, $f^{**}$ is lower semicontinuous and convex with $f^{**}\le f$, so $\text{epi}(f^{**}) \supset \text{epi}(f)$ and $\text{epi}(f^{**}) \supset \ol{\text{Conv}}(\text{epi}(f))$. Thus it is sufficient to show that $\text{epi}(f^{**}) \subset \ol{\text{Conv}}(\text{epi}(f))$, or equivalently,

\[E \times \real \setminus \ol{\text{Conv}}(\text{epi}(f)) \subset E \times \real \setminus \text{epi}(f)\]

To this end, let $A = \ol{\text{Conv}}(\text{epi}(f))$ and $(x, \alpha) \in E \times \real \setminus A$. By the Hahn-Banach Theorem, there exists $\phi \in F$, $\mu \in \real$, and $\alpha_{0} \in (\alpha, \infty)$ such that

\[\sup_{(y, \beta) \in A}\dpn{y, \phi}{\lambda}- \mu \beta \le \dpn{x, \phi}{\lambda}- \mu \alpha_{0} \le \dpn{x, \phi}{\lambda}- \mu \alpha\]

For any $(y, \beta) \in A$, $\beta$ may be arbitrarily large by Lemma 18.3.6, so $\mu \ge 0$.

In the case that $\mu > 0$, for each $y \in \bracs{f < \infty}$,

\begin{align*}\dpn{x, \phi}{\lambda}- \mu\alpha_{0}&\ge \dpn{y, \phi}{\lambda}- \mu f(y) \\ -\dpn{y, \phi}{\lambda}+ \dpn{x, \phi}{\lambda}- \mu\alpha_{0}&\le - \mu f(y) \\ \dpn{y, \mu^{-1}\phi}{\lambda}- \dpn{x, \mu^{-1}\phi}{\lambda}+ \alpha_{0}&\le f(y)\end{align*}

so $(\mu^{-1}\phi, \dpn{x, \mu^{-1}\phi}{\lambda}- \alpha_{0}) \le f$ and

\[f^{**}(x) \ge \dpn{x, \mu^{-1}\phi}{\lambda}- \dpn{x, \mu^{-1}\phi}{\lambda}+ \alpha_{0} > \alpha\]

Now suppose that $\mu = 0$. Given that $f^{*} \ne \infty$, there exists at least one pair $(\phi_{0}, \gamma_{0}) \in F \times \real$ such that $(\phi_{0}, \gamma_{0}) \le f$. Let

\[\gamma = \sup_{(y, \beta) \in A}\dpn{y, \phi}{\lambda}< \dpn{x, \phi}{\lambda}\]

For each $t > 0$, let $\Phi_{t} = \phi_{0} + t\phi$ and $\Gamma_{t} = \gamma_{0} + t\gamma$, then for each $y \in \bracs{f < \infty}$,

\[\dpn{y, \Phi_t}{\lambda}- \Gamma_{t} \le f(y) + t\underbrace{(\dpn{y, \phi}{\lambda} - \gamma)}_{\le 0}\le f(y)\]

so $(\Phi_{t}, \Gamma_{t}) \le f$. By (1),

\[f^{**}(x) \ge \dpn{x, \Phi_t}{\lambda}- \Gamma_{t} = \dpn{x, \phi_0}{\lambda}- \gamma_{0} + t\underbrace{(\dpn{x, \phi}{\lambda} - \gamma)}_{> 0}\]

As the above holds for all $t > 0$, $f^{**}(x) = \infty > \alpha$.

Thus $f^{**}(x) > \alpha$ and $(x, \alpha) \not\in \text{epi}(f^{**})$ for all $(x, \alpha) \in E \times \real \setminus A$. Therefore

\[E \times \real \setminus A \subset E \times \real \setminus \text{epi}(f^{**})\]

and $\text{epi}(f^{**}) \subset \ol{\text{Conv}}(\text{epi}(f))$.$\square$

Post a Comment

Name:Email:
Please enter the tag of the current page (12D) to post the comment.
Tag: