The setting is the same as in this question. Precisely, we have a strictly convex hypersurface $S\subseteq(\mathbb{R}^{2n},\omega_0)$, where $\omega_0$ is the standard symplectic form, and we define $F$ as follows. We assumed $0\in S$, so any ray issuing from the origin intersects $S$ in precisely one point $\xi$. Thus, for $x\neq0$, there is a unique $\lambda$ such that $\xi=\lambda^{-1}x\in S$. We define $F(x)=\lambda$ for $x\neq0$ and $F(0)=0$. We thus obtain a $\mathcal{C}^2$ function, which makes $H_F$ (the Hessian matrix) symmetric. As seen in the linked question, by Euler's theorem for homogeneous functions we deduce a formula which in turn gives $H_F(x)x=0$ for all $x$. So for no $x$ is this hessian nonsingular. However, according to Hofer-Zehnder (p. 25, preliminaries to the existence theorem for periodic orbits on convex regular compact energy surfaces), we can, by convexity of $S$, conclude that the restriction to the tangent space of $S$ of the bilinear form induced by the Hessian, i.e. $H_F|TS$, is positive. I assume this means $H_F|TS$ is symmetric and positive definite. We have seen it is symmetric by regularity and the Schwarz lemma. But how do I deduce it is positive definite? And above all, how does convexity enter the proof? I really don't know where to start here…

  • $\begingroup$ Is the symplectic form relevant here? $\endgroup$ – Amitai Yuval Sep 9 '15 at 10:57
  • $\begingroup$ @AmitaiYuval I mentioned it for the sake of completeness of context. I guess it might be irrelevant. $\endgroup$ – MickG Sep 9 '15 at 11:00

Sketch: We want to prove that for any $p\in S$ and $v\in T_pS$, the second derivative of $F$ at $p$ in the direction $v$ is positive. Note that for any given such $p,v,$ everything that matters to us happens in the $2$-dimensional space $\mathrm{span}(p,v).$ Hence, it is enough to solve the problem for $n=1$.

So, we have a convex curve $S$ in $\mathbb{R}^2$, and we have some $p\in S$. Let $$l=\{p+t\cdot v|t\in\mathbb{R}\}$$ denote the line tangent to $S$ at $p$. (In other words, $0\neq v\in T_pS$). By definition of $F$, we have $F(p)=1$. By convexity, we have $F(q)>1$ for every $p\neq q\in l$. This means that the function $F|_l$ admits a minimum at $p$, and hence, the second derivative in the direction $v$ is positive. (Well, non-negative, anyway. All you have to do now is use strict convexity in order to show that the second derivative does not vanish at $p$).

Edit: We show that strong convexity of a hypersurface is independent of the defining function. Let $S\subset\mathbb{R}^n$ be defined locally both by $\{f=0\}$ and $\{g=0\}$, where $f,g:U\to\mathbb{R}$ are smooth and both derivatives $df, dg$, don't vanish on $S$. Let $p\in S\cap U$. It follows that there is a smooth $h:U\to\mathbb{R}$, such that $g=hf$. Since $dg_p\neq0,$ it follows from the Leibniz rule that necessarily $h(p)\neq0$.

We assume now that $S$ is strongly convex with respect to $f$, that is, $\nabla^2f_p$ is positive definite on the tangent space $T_pS$. Let $v\in T_pS$. Then$$dg(v)=dh(v)f+hdf(v),$$and at $p$ we have$$\nabla^2g_p(v)=\nabla^2h_p(v)f(p)+2dh_p(v)df_p(v)+h(p)\nabla^2f_p(v).$$By the assumptions, the first and second terms vanish. The last term has the sign of $h(p)$. Thus, if $g$ is such that $h$ is positive, then $\nabla^2g_p$ is positive definite on $T_pS$, and if $h$ is negative, $\nabla^2g_p$ is negative definite on the tangent space.

Edit 2: The above argument works whenever $f$ and $g$ are at least $C^3$. We now treat the case where they are only $C^2$ (then $h$ is $C^2$ far away from $S$, but only $C^1$ on $S$). Since $g$ is $C^2$, we have$$\nabla^2g_p(v)=\lim_{x\to p}\nabla^2g_x(v)=\lim_{x\to p}\nabla^2h_x(v)f(x)+2dh_x(v)df_x(v)+h(x)\nabla^2f_x(v),$$ where we let $x$ approach $p$ from outside of $S$, and so the limit is well defined. It suffices then to prove$$\lim_{x\to p}\nabla^2h_x(v)f(x)=0.$$For that, we change the coordinates and assume $S=\{x_n=0\}.$ Note that when we do so, the expression for the Hessian tensor includes a correction involving the Christoffel symbols. However, this correction is bounded, as it involves only the first derivatives of $h$, and hence vanishes when multiplied by $f$, as $x$ gets closer to $p$. Consequently, we may consider the standard Hessian in our new coordinates.

Outside of $S$, we have $h=g/f$. Hence, for $1\leq i\leq n-1$,$$h_i=\frac{g_if-gf_i}{f^2}.$$Differentiating again, \begin{align} h_{ii}&=\frac{(g_{ii}f+g_if_i-g_if_i-gf_{ii})f^2-(g_if-gf_i)2ff_i}{f^4}\\ &=\frac{g_{ii}}{f}-\frac{gf_{ii}}{f^2}-\frac{2g_if_i}{f^2}+\frac{2gf_i^2}{f^3}, \end{align} and hence, \begin{align} h_{ii}f&=g_{ii}-\frac{g}{f}f_{ii}-2f_i\frac{g_if-gf_i}{f^2}\\ &=g_{ii}-hf_{ii}-2f_ih_i. \end{align} Since on $S$ both $f$ and $g$ are constant, all the terms vanish on $S$. It follows that$$\lim_{x\to p}\nabla^2h_x(v)f(x)=0,$$and the proof is complete.

  • $\begingroup$ By convexity if $q\in l$ and $q\neq p$ $q$ is out of the curve, outside the region the curve bounds, and hence $F(q)>1$, OK. $\endgroup$ – MickG Sep 9 '15 at 12:05
  • $\begingroup$ So we have a minimum, thus non-negative derivative. But I can't really see how the fact that any segment with endpoints on $S$ must be contained in the interior of the region $S$ bounds except for the endpoints, i.e. the strict convexity of $S$, should help in proving the second derivative doesn't vanish at $p$. I mean, I can tell that the level sets of this function, being it homogeneous of rank 1 for positive multipliers, are simply dilations of $S$, so all strictly convex. But even this doesn't seem useful for what I want to prove. Strict convexity implies the tangent line to $p$ has to… $\endgroup$ – MickG Sep 9 '15 at 12:10
  • $\begingroup$ …intersect $S$ only at $p$, and never the interior. Which gives us $F(q)>1$ for all $q\neq p$ in $l$, which we already know. What more can I extract out of strict convexity that convexity doesn't give me? $\endgroup$ – MickG Sep 9 '15 at 12:12
  • 1
    $\begingroup$ @MickG This is not needed. We say the vector $v$ is mapped under the coordinate change to $\partial/\partial x_i$. Then the Hessian of $v$ is equal to the corresponding Hessian of $\partial/\partial x_i$. No need to consider any mixed derivatives. $\endgroup$ – Amitai Yuval Oct 18 '15 at 16:41
  • 1
    $\begingroup$ @MickG Yes, right. $\endgroup$ – Amitai Yuval Oct 18 '15 at 16:45

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.