Lemma 18.3.7 (Almost Subgradient).label Let $\dpn{E, F}{\lambda}$ be a duality over $\real$, $f: E \to (-\infty, \infty]$ be convex and $\sigma(E, F)$-lower semicontinuous, $x \in \bracs{f < \infty}$, and $\alpha < f(x)$, then there exists $(\phi, \gamma) \in E \times \real$ such that:

  1. (1)

    $(\phi, \gamma) \le f$.

  2. (2)

    $\dpn{x, \phi}{\lambda}- \gamma = \alpha$.

In particular, $f^{*}\ne \infty$.

Proof. (1): Since $f$ is convex and $\sigma(E, F)$-lower semicontinuous, $\text{epi}(f)$ is $\sigma(E \times \real, F \times \real)$-closed and convex. By the Hahn-Banach Theorem, there exists $\phi \in F$ and $\mu \in \real$ such that

\[\sup_{(y, \beta) \in \text{epi}(f)}\dpn{y, \phi}{E}- \mu \beta < \dpn{x, \phi}{E}- \mu \alpha\]

For any $(y, \beta) \in \text{epi}(f)$, $\beta$ may be arbitrarily large, $\mu \ge 0$. In particular, since $(x, f(x)) \in \text{epi}(f)$, the strict inequality implies that $\mu > 0$. Thus for each $y \in \bracs{f < \infty}$,

\begin{align*}\dpn{x, \phi}{E}- \mu\alpha&> \dpn{y, \phi}{E}- \mu f(y) \\ -\dpn{y, \phi}{E}+ \dpn{x, \phi}{E}- \mu\alpha&> -\mu f(y)\\ \dpn{y, \mu^{-1}\phi}{E}- \dpn{x, \mu^{-1}\phi}{E}+ \alpha&< f(y)\end{align*}

so $(\mu^{-1}\phi, \dpn{x, \mu^{-1}\phi}{E}- \alpha) \le f$ and

\[\dpn{x, \mu^{-1}\phi}{E}- \dpn{x, \mu^{-1}\phi}{E}+ \alpha = \alpha\]

$\square$

Post a Comment

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