Let $||\cdot||$ be a norm on $\mathbb{R}^n$, with dual norm $||x||_* :=\max_\limits{y:||y||\leq 1}y^T x$. I'd like to show $$\max_{x \in \mathbb{R}^n}(x^T d-||x||)=\begin{cases}
0 & \text{ if } ||d||_* \leq 1 \\
\infty & \text{ otherwise }
\end{cases}$$
(the Legendre transform of a norm is the characteristic function of the unit ball (wrt the dual norm)
I'm lost because I don't know how to take the gradient of the objective (and I think $||\cdot||$ is non differentiable anyway). I know that the dual of the dual norm is the original norm: $||\cdot||_{**} \equiv ||\cdot||$
Legendre transform of a norm
4.4k Views Asked by Jason https://math.techqa.club/user/jason/detail AtThere are 2 best solutions below
Using the fact that $\|x\| \equiv \underset{y \in \mathbb{R}^n,\|y\|_* \le 1}{\sup }x^Ty$, you immediately get
\begin{equation} \begin{split} \underset{x \in \mathbb{R}^n}{\text{sup }}x^Td - \|x\| &= \underset{x \in \mathbb{R}^n}{\sup}x^Td - \underset{y \in \mathbb{R}^n,\|y\|_* \le 1}{\sup }x^Ty = \underset{y \in \mathbb{R}^n, \|y\|_* \le 1}{\inf }\underset{x\in \mathbb{R}^n}{\sup}x^T(d - y) \\ &= \underset{y \in \mathbb{R}^n, \|y\|_* \le 1}{\inf }\begin{cases}0,&\mbox { if }y = d,\\+\infty, &\mbox{ otherwise}\end{cases}\\ &= \begin{cases}0,&\mbox { if }\|d\|_* \le 1,\\+\infty, &\mbox{ otherwise,}\end{cases} \end{split} \end{equation} where the second equality follows from Sion's minimax theorem (as an easy exercise, the reader should check that the hypotheses of the Theorem are satisfied).
Case 1: $\|d\|_*\le 1$.
You know that $\|x\|=\|x\|_{**}=\max_{\|y\|_*\le 1}x^Ty\ge x^Td$, then
Therefore, $\max=0$.
Case 2: otherwise, i.e. $\|d\|_*>1$.
Let us split $x=\|x\|\cdot\frac{x}{\|x\|}=\|x\|\cdot x_0$. So $x^Td-\|x\|=\|x\|\cdot(x_0^Td-1)$. Maximizing w.r.t. the unit direction $x_0$ first with fixed $\|x\|$ gives $$ \max_{\|x_0\|=1}\|x\|\cdot(x_0^Td-1)=\|x\|\cdot\max_{\|x_0\|=1}(x_0^Td-1)=\|x\|\cdot\underbrace{(\|d\|_*-1)}_{>0}. $$ Now maximizing w.r.t. the length $\|x\|$ gives clearly $+\infty$.
P.S. You can apply the splitting in Case 2 to Case 1 too. Then $\|d\|_*-1\le 0$ and the maximum w.r.t. $\|x\|$ is zero.