0
$\begingroup$

I came across the following equivalence. Let $g_j$ for $j=1,\ldots,k$ be closed convex functions. For $\lambda_j > 0$ for $j=1,\ldots,k$. Then, the term $$ \big( \sum_{j=1}^k \lambda_j g_j \big)^*(z) $$ is equivalent to

$$ \inf_{\sum z_j = z} \left\{ \sum_{j=1}^k \lambda_j g_j^* \left( \frac{z_j}{\lambda_j} \right) \right\}. $$

It is noted that this is due to the infimal convolution definition applied on the conjugate of the sum of convex functions. I don't see how though. The classical definition as in this Math.SE question is not applicable to this concept, or at least I was not able to show.

How can I show this equivalence, and is there any source online (specifically for this)?

$\endgroup$
0
$\begingroup$

Ok, I solved. here they show that $(\sum\lambda_j g_j)^* =(\lambda_1g_1)^* \Delta (\lambda_2g_2)^*\Delta \ldots $ where $\Delta$ is the infimal convolution operator.

Next, apply the definition of inf conv to have $\inf \{ \sum (\lambda_j g_j)^* (z_j) \}$ or equivalently $\inf \{ \sum \lambda_j (g_j)^* (z_j/\lambda_j) \}$ where $\sum z_j= z$.

$\endgroup$
  • 1
    $\begingroup$ the book on Convex Analysis by Rockafellar is a good reference for these identities $\endgroup$ – LinAlg Oct 15 '19 at 1:21

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.