6
$\begingroup$

Let $\Omega$ be an open set in a closed manifold, $(M^n, g)$. We can define the perimeter as

$$\text{Per}_g(\Omega) = \sup\bigg\{\int_{\Omega} \text{div}_g(T) dVol_g, \; : \; T \in C^1(M, T M), \quad \sup_{x \in \Omega} |T(x)| \leq 1\bigg\}$$

i.e. the supremum of the integrals of the divergence of all vectors fields on $M$, restricted to $\Omega$. I'd like to show that $\text{Per}(\Omega)$ is continuous with respect to the metric, i.e. if $g, g'$ are two smooth metrics on $M^n$, then

$$|\text{Per}_g(\Omega) - \text{Per}_{g'}(\Omega)| \leq K \|g - g'\|_{C^1}$$

Its okay if $K$ depends on $\Omega$. I guess the main difficulty I foresee is that changing the metric may create an error term depending on derivatives of the vector field. So its unclear how to handle these.

$\endgroup$
14
  • 2
    $\begingroup$ Have you tried writing it all out in local coordinates? $\endgroup$
    – Deane Yang
    Commented Oct 29 at 15:32
  • 1
    $\begingroup$ For some reason, this really didn't occur to me! Thanks $\endgroup$
    – JMK
    Commented Oct 29 at 18:04
  • 3
    $\begingroup$ Unfortunately, that's not surprising. The emphasis these days is on coordinate-free calculations, so it doesn't occur to students to just use coordinates. My comment above is probably the most common I make on math.stackexchange.com. $\endgroup$
    – Deane Yang
    Commented Oct 29 at 18:16
  • 1
    $\begingroup$ I don't see why to expect the perimeter to be Lipschitz in $g$. For example consider a transformation $g \mapsto \lambda^2 g$ and let the dimension be $\geq 4$. Then we obtain $\operatorname{Per}_{\lambda^2 g}(\Omega) = \lambda^{n-1}\operatorname{Per}_{g}(\Omega)$. $\endgroup$
    – user542847
    Commented Oct 31 at 1:37
  • 1
    $\begingroup$ I think that we can get Hölder dependence for small values of $\|g - g'\|_{C^1}$, but large values are a problem if you want a global constant. Would restricting to small values be okay? $\endgroup$
    – Nate River
    Commented Oct 31 at 2:32

2 Answers 2

3
$\begingroup$

As Deane Yang suggested, local coordinates is the way to go. In an arbitrary coordinate system

$$X = X^i \partial_i, \qquad \text{div}_g(X) = X^i_i + X^i \Gamma_{ki}^i(g)$$ $$\implies \text{div}_g(X) - \text{div}_{g'}(X) = X^i (\Gamma_{ki}^i(g) - \Gamma_{ki}^i(g'))$$

where my convention is that $\Gamma_{ki}^j = g^{j \ell} g(\nabla_{\partial k} \partial_i, \partial_{\ell})$. So if $\Omega$ is say compact, then using that $\|X\|_{L^{\infty}} \leq 1$, we have

$$\Big|\int_{\Omega} \text{div}_g(X) - \int_{\Omega} \text{div}_{g'}(X)\Big| \leq \text{Vol}_g(\Omega) \cdot K \|g - g'\|_{C^1}$$

Taking a sequence $X_i$ such that $\text{Per}_g(\Omega) = \lim_i \int_{\Omega} \text{div}_g(X_i)$, then readily shows that

$$\text{Per}_g(\Omega) \leq \text{Per}_{g'}(\Omega) + \text{Vol}_g(\Omega) \cdot K \|g - g'\|_{C^1}$$

EDIT: Actually I was too quick to conclude. To be really careful, we have

$$\int_{\Omega} \text{div}_g(X) dV_g - \int_{\Omega} \text{div}_{g'}(X) dV_{g'} = \int_{\Omega} (\text{div}_g(X) - \text{div}_{g'}(X)) dV_g + \int_{\Omega} \text{div}_{g'}(X) (dV_g - dV_{g'})$$

The first integral is boundable as before. Its tempting to say that

$$\Big| \int_{\Omega} \text{div}_{g'}(X) (dV_g - dV_{g'})\Big| \leq \|g - g'\|_{C^0}\cdot\int_{\Omega} |\text{div}_{g'}(X)|dV_g $$

But a priori, $\int_{\Omega} |\text{div}_{g'}(X)| dV_g$ could be very large!

$\endgroup$
4
  • $\begingroup$ There seems to be another minor problem with this. The sequence $X_i$ you take must satisfy $\sup |X_i(x)|_g \leq 1$ according to your definition. However, this does not imply that $\sup |X_i(x)|_{g'} \leq 1$, and so we can not conclude that $\int_\Omega \operatorname{div}_{g'} (X_i) \mathrm{d}V_g$ is bounded by $\operatorname{Per}_{g'}(\Omega)$. $\endgroup$
    – user542847
    Commented Oct 31 at 0:37
  • $\begingroup$ @OrangeMushroom also a good point. There are many parts to the full proof that are still missing it seems. $\endgroup$
    – JMK
    Commented Oct 31 at 1:13
  • 1
    $\begingroup$ You can write $dV_g' = f dV_g$ where $f\to 1$ in $C^1$. Then $div_g(X)(f-1) dV_g = div_G(X(f-1))dV_g - gX,\nabla (f-1))dV_g$. The first term integrated on $\Omega$ is $o(1)$ since $|X(f-1)| \to 0$ in $C^0$ and then the second is also $o(1)$ since $f\to 1$ in $C^1$ and $|X|\leq 1$. The issue raised by @OrangeMushroom can be fixed by shrinking $X$ a bit to compensate for the fact that $g'$ is measuring things slightly bigger/smaller than $g$ (but this stretch factor is $1+o(1)$ so goes away. $\endgroup$ Commented Oct 31 at 20:15
  • $\begingroup$ Oh clever! I guess your saying $$|\int_{\Omega} \text{div}_g(X(f-1)) dV_g| \leq ||X(f-1)||_{C^0} |\int_{\Omega} \text{div}_g \left(\frac{X(f-1)}{||X(f-1)||_{C^0}} \right) dV_g | \leq ||X(f-1)|| Per_g(\Omega) \leq ||f-1|| Per_g(\Omega)$$ I think this is the step I was missing - thanks! $\endgroup$
    – JMK
    Commented Nov 1 at 16:34
3
+50
$\begingroup$

Since the top-dimensional exterior power is one-dimensional, we may talk about the quotient $\alpha/\beta \in C^1(\mathscr{M})$ of two volume forms. The divergence only depends on the volume form of $g$, and we have the identity $$ \operatorname{div}_{g'}(X) = \frac{\omega_{g\hphantom{}}}{\omega_{g'}} \,\operatorname{div}_{g}\left(\frac{\omega_{g}}{\omega_{g'}} \cdot X\right), $$ where $\omega_g$ and $\omega_{g}'$ are the volume forms of $g$ and $g'$, respectively. For what follows, define $\rho_{g, g'} \in C^1(\mathscr{M})$ by $\rho_{g, g'} = \omega_g/\omega_{g'} - 1$ and define $\lambda_{g,g'} \in C^1(\mathrm{T}\mathscr{M})$ by $\lambda_{g, g'}(v) = |v|_{g}/|v|_{g'}$.

Then for any $C^1$ vector field $X$, we have $\int_{\Omega} \operatorname{div}_{g'} (X)\,\omega_{g'} - \int_{\Omega} \operatorname{div}_{g}(X)\, \omega_g = \int_{\Omega} \operatorname{div}_{g} \left(\rho_{g, g'} X \right)\,\omega_g$; furthermore we have \begin{equation*} \int_{\Omega} \operatorname{div}_{g} (X)\,\omega_{g} = \sup_{\Omega} (\lambda_{g, g'} \circ X) \underbrace{\int_{\Omega} \operatorname{div}_{g} \left(\frac{X}{\sup_{\Omega} (\lambda_{g,g'} \circ X)} \right)\,\omega_{g}}_{\leq \,\operatorname{Per}_{g}(\Omega) \text{ if } \lVert X \rVert_{g'} \leq 1} \end{equation*} and because $\lambda_{g, g'}$ is invariant under scaling, \begin{align*} \int_{\Omega} \operatorname{div}_{g} \left(\rho_{g, g'} X \right)\,\omega_g &= \sup_{\Omega} |\rho_{g, g'}| \int_{\Omega} \operatorname{div}\left(\frac{\rho_{g, g'}X}{\sup_{\Omega} |\rho_{g,g'}| } \right)\,\omega_{g} \\[1ex] &= \sup_{\Omega} (\lambda_{g, g'} \circ X) \sup_{\Omega} |\rho_{g, g'}|\! \underbrace{\int_{\Omega} \operatorname{div}_{g} \left(\frac{\rho_{g, g'} X}{\sup_{\Omega} (\lambda_{g, g'} \circ X) \cdot \sup_{\Omega} |\rho_{g,g'}|} \right)\,\omega_g}_{\leq \, \operatorname{Per}_{g}(\Omega) \text{ if } \lVert X \rVert_{g'} \leq 1}. \end{align*}

Let $\mathrm{U}\mathscr{M}(g') \subseteq \mathrm{T}\mathscr{M}$ be the unit tangent bundle with respect to $g'$, and define $\Lambda_{g, g'} = \sup_{v \in \mathrm{U}\mathscr{M}(g')} |v|_{g}$. Since $\mathrm{U}\mathscr{M}(g')$ is compact, $\Lambda_{g,g'}$ is finite. It follows that $$ \sup_{\Omega} \lambda_{g, g'} X = \sup_{\Omega} \left|\frac{\,X_{\hphantom{g'}}\!}{|X|_{g'}\!}\right|_{\raise{1.5pt}{g}} \leq \Lambda_{g, g'}. $$ It remains to show that $\lim_{g' \to g} \Lambda_{g, g'} = 1$ and $\lim_{g' \to g}\sup_{\Omega} |\rho_{g, g'}| = 0$. To do this, we use the following characterization of $C^0$ convergence:

Proposition. Let $(s_k)_{k \in \mathbb{N}}$ be a sequence of sections of $\mathrm{T}^*\mathscr{M} \otimes \mathrm{T}^*\mathscr{M}$. Then $(s_k)_{k \in \mathbb{N}}$ uniformly converges to a section $s$ if and only if for every pair of vector fields $X, Y \in \Gamma(\mathrm{T}\mathscr{M})$, we have $s_k(X, Y) \to s(X, Y)$ uniformly.

To use this, first note that there exists an open cover $\{U_\alpha\}_{\alpha \in A}$ of $\mathscr{M}$ such that for each $\alpha \in A$, there exists an orthonormal frame $\xi^\alpha \in \mathrm{F}U_\alpha (g)$.

To show that $\lim_{g' \to g} \Lambda_{g, g'} = 1$, take a convergent sequence $g_k \to g$, and let $\epsilon > 0$. Then there exists some index $K$ such that for all $k \geq K$ and $\alpha \in A$, we have $\lVert \delta_{ij} - g_k(\xi^\alpha_i, \xi^\alpha_j) \rVert_{C^0} < \epsilon$.

Let $x \in \mathscr{M}$; this point must belong to some $U_{\alpha}$. Let $v \in \mathrm{U}_x\mathscr{M}(g_k)$, and write $v^i = g(v, \xi^\alpha_i|_x)$. Then $$ |v|_{g_k}^2 = \left| \sum_{i=1}^{n} v^i \xi^\alpha_i|_x \right|_{g_k} = \sum_{i, j} g_k(\xi^\alpha_i, \xi^\alpha_j) v^iv^j = 1. $$ In particular, if $\Sigma_{k, x}$ are the eigenvalues of $g_k|_x$ in this frame, then $|v|_{g}^2$ is contained in the interval $[(\max \Sigma_{k,x})^{-1}, (\min \Sigma_{k, x})^{-1}]$. However, by the Gershgorin circle theorem, $$\Sigma_{k, x} \subseteq \left[1 - \max_{i} \sum_{j = 1}^{n} |g_k(\xi^\alpha_i, \xi^\alpha_j)|, 1 + \max_{i} \sum_{j = 1}^{n} |g_k(\xi^\alpha_i, \xi^\alpha_j)|\right] \subseteq \left[1 - n\epsilon, 1 + n\epsilon \right].$$ It follows that $|1 - \sup_{v \in \mathrm{U}\mathscr{M}(g_k)} |v|_{g}|$ is close to $0$. Therefore, $\lim_{g' \to g} \Lambda_{g, g'} = 1$.

To show that $\lim_{g' \to g} \sup_\Omega |\rho_{g, g'}| = 0$, take a convergent sequence $g_k \to g$, and let $\epsilon > 0$. As done before, let $x \in \mathscr{M}$; this point must belong to some $U_{\alpha}$. Then $$ \rho_{g, g_k}(x)\hspace{-1pt} = \hspace{-1pt} 1 - \hspace{-1pt} \frac{\omega_{g}(\xi^{\alpha}_1 \wedge \cdots \wedge \xi^{\alpha}_n)}{\omega_{g_k}(\xi^{\alpha}_1 \wedge \cdots \wedge \xi^{\alpha}_n)} = \hspace{-1pt} 1 - \frac{1}{\omega_{g_k}(\xi^{\alpha}_1 \wedge \cdots \wedge \xi^{\alpha}_n)} = \hspace{-1pt} 1 - \frac{1}{\sqrt{\det \big[g_k(\xi^\alpha_i, \xi^{\alpha_j}) \big]_{ij}}}. $$ Applying the Gershgorin circle theorem again, we obtain $\Sigma_{k,x} \subseteq [1-\epsilon, 1 + \epsilon]$. Hence, the product of the eigenvalues is also close to $1$. It follows that $\lim_{g' \to g} \sup_\Omega |\rho_{g, g'}| = 0$.

The conclusion now follows. First take a sequence of vector fields $X_i \in C^1(\mathscr{M}, \mathrm{T}\mathscr{M})$ which realizes the perimeter of $\Omega$ with respect to $g'$. Then $$ \operatorname{Per}_{g'}(\Omega) \leq \Lambda_{g, g'} \operatorname{Per}_{g}(\Omega) + \Lambda_{g, g'} \sup_{\Omega} |\rho_{g, g'}| \operatorname{Per}_{g}(\Omega). $$ Therefore, if $g_k \to g$ uniformly, then $\limsup_{k \to \infty} \operatorname{Per}_{g_k} (\Omega) \leq \operatorname{Per}_{g}(\Omega)$. Thus the perimeter is upper semi-continuous, and hence continuous.

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .