5
$\begingroup$

Let $C_{lsc}(\mathbb{R}^n)$ be the space of lower semicontinuous convex functions $\mathbb{R}^n \to \mathbb{R}$. The Legendre-Fenchel (LF) transform of $f \in C_{lsc}(\mathbb{R}^n)$ is: $$ f^*(y) := \sup_{x \in \mathbb{R}^n} (\langle x, y \rangle - f(x)) $$

It is known that the LF transform is continuous and an involution on $C_{lsc}(\mathbb{R}^n)$ (Wijsman 1963). I want to know if the LF transform is Lipschitz. That is, given $f$ and $g$ lsc, is there a way to bound $\|f^* - g^*\|$ by $\|f-g\|$ (under a standard norm)?

I haven't found any sources that do this, so my suspicion is that the LF transform is not Lipschitz. Does anyone know if this is true, or if not a simple counterexample?

It is worth noting that Attouch and Wets ("Isometries of the Legendre-Fenchel transform") constructed norms under which the LF transform is an isometry -- however these norms are not particularly useful for me. I am looking for any results using standard norms (any of the p-norms).

$\endgroup$
2
  • $\begingroup$ What is a standard norm? $\endgroup$ Dec 8, 2022 at 22:57
  • $\begingroup$ Any of the L^p norms $\endgroup$
    – gdavtor
    Dec 8, 2022 at 23:05

1 Answer 1

6
$\begingroup$

This is basically true for sup norm by Fenchel's inequality. Indeed, for all $y$,

$$ f^*(y) = \sup_x\left( \langle x,y\rangle - f(x) \right) \leq \sup_x\left( g(x)+g^*(y) - f(x) \right) \leq \|f-g\|_{\infty} + g^*(y). $$

The same is true when the roles of $f,g$ are reversed, giving essentially what you want. One should take care, though, in writing the final conclusion because it is possible for $f^*(y) = g^*(y) = +\infty$, despite $\|f-g\|_{\infty}=0$ (e.g., take $f=g=0$). That said, if $\|f-g\|_{\infty}<\infty$, then it is clear from above that $$ \{y : f^*(y) = +\infty\} = \{y : g^*(y) = +\infty\}, $$ so it is reasonable to write $$ \|f^*-g^*\|_{\infty} \leq \|f-g\|_{\infty}, $$ provided you adopt the convention $|(+\infty)-(+\infty)| = 0$ for handling the indeterminate form that can appear in defining the sup norm on the LHS.

$\endgroup$
2
  • 1
    $\begingroup$ Lipschitz constant of 1 would then imply that LF is an isometry (on $C_{lsc}(\mathbb{R}^n)$), by the involution property $f^{**} = f$? This makes sense geometrically actually now that I think about it. Thank you! $\endgroup$
    – gdavtor
    Dec 10, 2022 at 1:42
  • 1
    $\begingroup$ Correct (provided one handles the indeterminate case mentioned). $\endgroup$
    – Tom
    Dec 10, 2022 at 3:23

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.