2
$\begingroup$

Let $f$ be a semi-concave function on an Alexandrov space $X$. Denote $\gamma_p(t)$ the $f$-gradient curve with $\gamma_p(0)=p$, i.e. $$ \gamma^+_p(t)=\nabla_{\gamma_p(t)}f. $$ If $X$ is a Riemannian manifold, $\nabla_{\gamma_p(t)}f=\nabla f(\gamma_p(t))$.

For any $p\in X$ and sufficiently small $t\geqslant 0$, there is $p'\in X$ such that $\gamma_{p'}(t)=p$. In other words, there exists a maximal $T>0$, $\gamma_p(t)$ can be defined on $(-T,0]$ with $\gamma_p(0)=p$.

Let $q=\lim_{t\to -T}\alpha_p(t)$. Then $q$ may be a critical point of $f$, i.e. $\nabla_q f=0$. Or $|\nabla_q f|>0$, but if we connect gradient curve $\gamma_{q}(t)$ defined on $(-T',0]$ ($\gamma_q(0)=q$) with gradient curve $\gamma_p(t)$ at $q$, this is not a gradient curve since it's possible that $$ \nabla_q f\neq \gamma_p^+(-T). $$

Can one show some examples for this case?

$\endgroup$

1 Answer 1

1
$\begingroup$

The answer depends on your definition of semi-concave functions. If you only require them to be semi-concave on geodesics then an obvious example is given by $X$ equal to the closed unit ball in $\mathbb R^n$ and $f=d(\cdot, \partial X)$. Gradient curves starting on the boundary can not be extended to the left.

If you require the function to be semi-concave on the double of $X$ then such example is impossible and gradient curves can always be extended to the left at points where $\nabla_pf\ne 0$. This follows by a standard homology argument. I'll leave you to fill the details but briefly it goes like this. Suppose we have a point $p\in X^n$ such that $\nabla_pf\ne 0$ but the gradient curve starting at $p$ can not be extended to the left. Suppose $p\notin\partial X$. Then $H_n(X,X\backslash \{p\},\mathbb Z_2)\cong H_{n-1}(\Sigma_pX,\mathbb Z_2)\ne 0$. On the other hand since the gradient curve starting at $p$ does not extend to the left we can use the gradient flow $\phi_t$ of $f$ to push any relative cycle in $C_n(X,X\backslash \{p\},\mathbb Z_2)$ into $X\backslash \{p\}$ which implies that it's equal to zero in homology which gives a contradiction. If $p\in\partial X$ then the same argument applies to the flow of the double of $f$ on the double of $X$ which has no boundary.

$\endgroup$
6
  • $\begingroup$ Thank you! I know that for any $p$ with $\nabla_p f\neq 0$, there exist $q$ such that the gradient flow starting at $q$ go through $p$. I'm sorry that may question was not clear, so I update it right now. Please read it again. $\endgroup$ Jan 5, 2020 at 7:47
  • $\begingroup$ Sorry, I misunderstood the question and I am still not sure what you are asking. Are you asking for an example when the gradient curve starting at $q$ doesn't pass through $p$? That's not possible because forward gradient flow is Lipschitz. $\endgroup$ Jan 5, 2020 at 14:42
  • $\begingroup$ Thank you! I want to find an example, that, for any $t>-T$, $\nabla_{\gamma(t)} f=\gamma^+(t)$, but at the limit point $\gamma(-T)$, the equality fails. $\endgroup$ Jan 7, 2020 at 15:04
  • $\begingroup$ This is impossible because the limit of gradient curves is a gradient curve. So for $t_i\to -T+$ and $q_i=\gamma(t_i)$ the gradient curves starting at $q_i$ will converge to the gradient curve starting at $q=\gamma(-T)$. This means that the curve $\gamma$ on $[-T,0] is actually a gradient curve. $\endgroup$ Jan 7, 2020 at 17:15
  • $\begingroup$ Thank you! But if the gradient curve $\gamma$ can be defined on $[-T,0)$, we know that there is $q'$ such that the $f$-gradient flow starting at $q'$ goes through $\gamma(-T)$. This contradicts to the maximal of $T$. On the other hand, I check the proof of "the limit of gradient curves is a gradient curve" in Petrunin's paper "Semiconcave...". If gradient curves $\alpha_n(t)\to \alpha(t)$, we can only get that $\alpha^+(t)=\nabla_{\alpha(t)}f$ for almost every $t$. So I am puzzled now. $\endgroup$ Jan 8, 2020 at 3:24

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.