6
$\begingroup$

For a seminar I am working on a Moreau-Yosida regularization in Banach spaces.

The regularization is defined by

$$f_\lambda(x) := \inf \left \{ \frac{\|x-y\|^2}{2\lambda} +f(y) : y \in X \right \}, ~ x \in X$$

where $X$ is a reflexive and strictly convex Banach space, $f: X \rightarrow \mathbb{R} \cup \{+\infty\}$ lower-semicontinuous, proper and convex, $\lambda > 0$.

In Convexity and Optimization in Banach Spaces, V. Barbu and T. Precupanu, Springer, 2012, the authors prove that the infimum defining $f_\lambda(x)$ is attained for all $x \in X$.

From that they deduce that $f_\lambda$ is convex and lower-semicontinious, but, unfortunately, they don't elaborate the argument. However, I do not see how it follows from what is given.

As far as I am aware, most literature deals with the case where $X$ is Hilbert, I could not finde the result stated above for the general case in another publication.

Question: Maybe someone can recommed me a paper where this result is proven? Or is there an obvious argument which I am missing at the moment?

Thank you in advance

$\endgroup$
5
$\begingroup$

I deleted a previous wrong and misleading answer.

The Moreau-Yoshida envelope is a special case of th infimal convolution of two convex functions $f$ and $g$ which is defined as $$ f\Box g(x) = \inf_{y\in Y} f(y) + g(x-y). $$ In other words: The infimal convolution of $f$ and $g$ is the largest (extended) real valued functional whose epigraph contains the (Minkowski) sum of the epigraphs of $f$ and $g$. Consequently, it is convex and lsc in $f$ and $g$ are.

A more verbatim argument for the convexity: In general if $F:X\times X\to ]-\infty,\infty]$ is convex, then $f(x) = \inf_y F(x,y)$ is convex: Take $x_1$ and $x_2$ such that $f(x_i)$ is finite and $\xi_1> f(x_1)$, $\xi_2> f(x_2)$. Then there exist $y_1$, $y_2$ such that $F(x_i,y_i)<\xi_i$. By convexity of $F$ it holds for $0<\lambda<1$ that $$\begin{array}{rl} f(\lambda x_1 + (1-\lambda)x_2) & \leq F(\lambda x_1 + (1-\lambda)x_2, \lambda y_1 + (1-\lambda)y_2)\\ & \leq \lambda F(x_1,y_1) + (1-\lambda)F(x_2,y_2)\\ & \leq \lambda \xi_1 + (1-\lambda)\xi_2. \end{array} $$ Letting $\xi_i\to f(x_i)$ shows convexity of $f$. Now apply this to $F(x,y) = f(y) + g(x-y)$.

By the way: One calls $f$ the inf-projection of $F$. The paper "Lipschitz Continuity of inf-Projections" by Roger J.-B. Wets also deals with the continuity of the inf projection.

$\endgroup$
  • $\begingroup$ Thank you very much, this helps a lot. But I have a question: In the convexity proof, do you really need the $\xi_i$'s? I would just replace the last line in the inequality with $\lambda f(x_1) + (1-\lambda)f(x_2)$, which gives us the convexity. $\endgroup$ – GenH Nov 29 '13 at 23:34
  • $\begingroup$ It does not hold that $F(x_i,y_i) \leq f(x_i)$. $\endgroup$ – Dirk Nov 30 '13 at 12:33
1
$\begingroup$

I remember some works by Crandal Pazy on semiflows of accretive operator on Banach spaces and they used the Yosida approximation for their purpose...

$\endgroup$

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.