I have a question arising from the Moreau-Yosida regularization in Banach spaces. The real Banach space $X$ and its dual $X^*$ are both reflexive strictly convex, $f:X \rightarrow \mathbb{R} \cup \{+\infty\}$ is lower-semicontinuous, proper and convex. $F$ is the duality mapping of $X$, defined by

$$F(x) := \{ x^* \in X^*: \|x\|_X^2=\|x^*\|_{X^*}^2=x^*(x)\}.$$

For $X$ strictly convex, $F$ is always single-valued.

I know that the subdifferential $\partial f(x)$ is maximal monotone and that $$R(\partial f+\lambda F) = X^*,$$ where $R$ is the range of the operator.

Somehow from this should follow that for every $x \in X, ~\lambda>0$ there is an $x_\lambda \in X$ such that $$ 0 \in\partial f(x_\lambda) + \lambda F(x_\lambda - x) $$.

But I don't see how. Well, if $F$ were linear this would be trivial, but it is not.

Does anyone have an idea how this can be shown? Thank you in advance.

  • $\begingroup$ Could you explain better exactly what $F$ is? $\endgroup$ – Sergio Parreiras Dec 8 '13 at 3:22
  • $\begingroup$ Thank you for the suggestion, I've done it now. $\endgroup$ – GenH Dec 8 '13 at 5:06
  • $\begingroup$ By which argument do you know $R(\partial f + \lambda \, F) = X^*$? Can you also show this including a translation $R(\partial f(\cdot+x) + \lambda \, F(\cdot))$? $\endgroup$ – gerw Dec 8 '13 at 8:30

First note that $F$ is the subdifferential of $j(x) = \frac12 \, \|x\|^2$.

Then, your condition is the necessary and sufficient optimality condition of: Minimize (w.r.t. $\tilde x$) $$f(\tilde x) + \frac\lambda2 \, \|\tilde x - x\|^2.$$

Now, it remains to show the existence of a minimizer of this functional. Therefore, you can use the convexity of $f$ and the quadratic growth of $\|\cdot\|^2$ to show that the objective is bounded from below and coercive. Using the lower-semicontinuity of $f$ (and $j$), you can infer the existence of a minimizer using standard arguments.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.