I am trying to estimate the integral $\int \mathbb{e} ^{-d(x_0,x)^2} \mathbb{d}x$ on a Riemann manifold $(M,g)$, for some arbitrary fixed $x_0 \in M$ and $d$ the usual distance. The only thing that I can think of is to use some coarea theorem, leading to $\int _0 ^\infty \mathbb{e} ^{-r^2} A(x_0, r) \mathbb{d}r$, where $A(x_0, r)$ is the area of the metric sphere of center $x_0$ and radius $r$. The issue now is to have some estimate of $A$. Surprisingly, even though there are plenty of comparison theorems for the volume of metric balls, I haven't been able to find anything usable on the area of spheres. Therefore, the questions:

  • how would you approach this integral (if not by the coarea formula)?
  • is it finite? (I assume so)
  • is $A$ a polynomial in $r$ (of degree $n-1$)? If so, I would expect it not to have a free term, but what about its leading coefficient?
  • can $A$ be computed explicitly in space forms (other than Euclidian, of course)?
  • are there comparison theorems for $A$?

Thank you.

| cite | improve this question | | | | |
  • $\begingroup$ A possible approach: If $\exp_{x_0}$ is a bijection $T_{x_0}M\to M$, then you can turn your integral into some integral over the tangent space (or get an estimate if it is not bijective). You have $d(x_0,\exp_{x_0}v)^2\leq|v|^2$ (with equality if $\exp$ is injective) for $v\in T_{x_0}M$, so you have a Gaussian integral in a Euclidean space with weight coming from pulling the metric to the tangent space. $\endgroup$ – Joonas Ilmavirta Mar 10 '15 at 20:00
  • $\begingroup$ @JoonasIlmavirta: My manifold is not necessarily complete, so $\mathbb{exp _{x_0} }$ cannot even be defined on the whole $T _{x_0} M$. $\endgroup$ – Alex M. Mar 10 '15 at 20:15
  • $\begingroup$ @JoonasIlmavirta: The inequality that you suggest would make my integral "greater or equal" than some Gaussian integral, while I'm looking to (hopefully) get "less or equal". $\endgroup$ – Alex M. Mar 10 '15 at 20:26
  • $\begingroup$ I guess my method would work if $\exp_{x_0}$ is bijective on its maximal domain of definition (which need not be the entire tangent space). If you don't want to assume this much, this approach might be useless. $\endgroup$ – Joonas Ilmavirta Mar 10 '15 at 20:36
  • 1
    $\begingroup$ It seems like things could go quite badly if your manifold is not complete, even in constant curvature. Puncture a round 2-sphere at antipodal points and take its universal cover. Now as $x_0$ heads towards one of the missing points, $A$ grows without bound. $\endgroup$ – John Harvey Mar 10 '15 at 20:38

The integral might be infinite. Indeed, the function $A(r)$ can grow as fast as you want, but in this case the Ricci curvature in some of the radial direction have to go to $-\infty$ as $r\to\infty$. (Needless to say $A$ can grow faster than any polynomial.)

If Ricci curvature is bounded below then $A(r)$ has at most exponential growth --- this follows from the Bishop--Gromov inequality. In this case you integral has to be finite.

| cite | improve this answer | | | | |
  • 4
    $\begingroup$ P.S. if $n=\dim M$ and $\mathrm{Ric}_M\ge -(n-1)$ then Bishop--Gromov inequality says that the function $$r\mapsto A(r)/(\sinh r)^{n-1}$$ is nonincreasing. $\endgroup$ – Anton Petrunin Mar 11 '15 at 1:36
  • $\begingroup$ Let me come back to your answer, please. You assume lower-boundedness of the Ricci curvature, but not completeness. All the proofs of Bishop-Gromov that I have read require completeness. Do you know of any version of it without this requirement? (I had asked a related question a few days ago, and only now did I remember your answer.) $\endgroup$ – Alex M. Sep 16 '15 at 20:16
  • $\begingroup$ @AlexM. One needs completeness, without one you can do pretty much anything, in particular there is no Bishop--Gromov. $\endgroup$ – Anton Petrunin Sep 16 '15 at 21:25

I personally do not think that you can compute that integral explicitly in general just because the geodesic spheres could become quite ugly. $\newcommand{\bx}{\boldsymbol{x}}$ However, when $M$ has finite volume, you can obtain pretty good asymptotic estimates for the integral

$$ I(\lambda,\bx_0)=\int_M e^{-\lambda d(\bx_0,\bx)^2} dV_g(x), $$

as $\lambda\to \infty$. To obtain these asymptotic estimates you write $dV_g(\bx)$ in normal coordinates $(t^1,\dotsc, t^n)$ at $\bx_0$ in the form

$$ dV_g(x) = \rho(t) dt^1\wedge\cdots \wedge dt^n. $$

The Taylor expansion of $\rho(t)$ at $t=0$ can be described explicitly in terms of geometric invariants at $\bx_0$. The degree 4 Taylor polynomial of $\rho$ at $t=0$ is described in Corollary 9.9 of A. Gray's book Tubes, 2nd Edition, Birhauser, 2006. The Taylor polynomials of degree $>4$ are very complicated.

The integral you are interested could be computed on spaces with rich symmetries, but even in those cases it is not easy.

| cite | improve this answer | | | | |
  • $\begingroup$ Agreed, but normal coordinates cover only a small patch of the whole manifold. What am I going to do with the rest? (And no, I do not hope for an explicit result, just for a "sharp" estimate.) $\endgroup$ – Alex M. Mar 10 '15 at 20:59
  • $\begingroup$ In any case, this makes me think of the formula of steepest descent in Euclidean spaces. Does a similar thing exist for Riemannian manifolds? With Google I can only find articles about optimizations. $\endgroup$ – Alex M. Mar 10 '15 at 21:10
  • $\begingroup$ @AlexM. If the injectivity radius at $x_0$ is small then there is very little you can do. However, if you allow for the large parameter $\lambda$, anything that is outside a small geodesic ball is negligible for large $\lambda$. $\endgroup$ – Liviu Nicolaescu Mar 10 '15 at 21:34

You could integrate the co-area formula by parts, to get: $$\int_0^\infty 2re^{-r^2}V(x_0,r)dr$$ where $V(x_0,r)$ is the volume of the ball at $x_0$ of radius $r$, and then use comparison results for $V$.

| cite | improve this answer | | | | |
  • $\begingroup$ Yes, I realised this after having posted the question. Still, I'm not sure whether this helps: are explicit formulae for the volume of metric balls in space forms known? Or, at least, is $V(x_0, r)$ a polynomial function of $r$? $\endgroup$ – Alex M. Mar 10 '15 at 23:06
  • $\begingroup$ My point is just that you can use comparison results for $V$ in place of comparison results for $A$. As for evaluating your integral explicitly, the integration by parts I mention is probably not very helpful. $\endgroup$ – John Pardon Mar 11 '15 at 0:07
  • 2
    $\begingroup$ @AlexM.: $V(x_0,r)$ is not generally polynomial in $r$ in non-flat space forms. For example, on the $2$-sphere with Gauss curvature $K\equiv 1$, one finds $$V(x_0,r) = 2\pi \,(1{-}\cos r), \qquad 0\le r\le \pi.$$ The general formula involves polynomials in trig (or hyperbolic trig) functions. $\endgroup$ – Robert Bryant Mar 12 '15 at 15:20
  • $\begingroup$ @RobertBryant: I am very interested in the general formula for space forms (at least for small balls). Do you know where I could find it? $\endgroup$ – Alex M. Mar 12 '15 at 21:39
  • 2
    $\begingroup$ I don't know a reference off the top of my head, but straightforward calculation shows that $$V(x_0,r) = V_{n-1}\sum_{\ell=0}^{n-1}\frac{(-1)^\ell}{2^{n-1}(\sqrt{-K})^n}{{n{-}1}\choose{\ell}}{{\bigl({e}^{(n-1-2\ell)\sqrt{-K}\,r}-1\bigr)}\over{(n-1-2\ell)}}$$ when $M$ is an $n$-dimensional space form of sectional curvature $K\not=0$ and $V_{n-1}$ is the volume of the unit $(n{-}1)$-sphere in Euclidean $n$-space. (Note that, even when $K>0$, this is a real-valued function with leading order term $V_{n-1}r^n/n$.) Perhaps this formula simplifies? $\endgroup$ – Robert Bryant Mar 12 '15 at 22:21

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.