# Smoothness and Hessian of convex conjugate/Legendre transform

Let $p:[0,\infty) \times \mathbb{R}^n \to \mathbb{R}$ be a smooth convex function for all $x \in \mathbb{R}^n$.

We define its Legendre transform (or convex conjugate) as

$$p^*(t,y)=\sup_{x \in \mathbb{R^n}}\left\{ x \cdot y - p(t,x)\right\}$$

I know differentiability and convexity are dual concepts, so convexity of $p$ implies differentiability of $p^*$. However, I should be able to get smoothness of $p^*$ and Hessian $D^2_y p^*$ bounded away from zero and $+\infty$. How do I do this?

-

Correction: strict convexity of $p$ implies differentiability of $p^*$.
Think in terms of the gradient map $\nabla p\colon\mathbb R^n\to\mathbb R^n$. The Hessian of $p$ is the Jacobian determinant of $\nabla p$. Now, $\nabla p^*$ is $(\nabla p)^{-1}$, the inverse map. The Jacobian determinant of the inverse map is, of course, the reciprocal of the Jacobian determinant of $\nabla p$. Thus, the inequality $c\le \det D^2 p^*\le C$ is equivalent to $C^{-1}\le \det D^2 p\le c^{-1}$.
There are some sufficient conditions for $C^{-1}\le \det D^2 p\le c^{-1}$... For example: if $p(x)-\epsilon |x|^2$ is convex for some $\epsilon>0$, you have a lower bound. And if the second derivatives are bounded, you have an upper bound. Of course, these are only sufficient conditions, they are not necessary.
Leonid, any idea where I might find a proof of the fact that strict convexity of $p$ implies differentiability of $p^*$? Thx! – chango Jun 10 '12 at 11:09