I have the following integral $$ I(\varepsilon) = \iint_D \frac{\sqrt{1+|\nabla h(u,v)|^2}}{[(h(u,v)+\varepsilon)^2+u^2+v^2]^2} du dv, $$ where $h$ is a smooth function with $h(0,0)=0 = h_u(0,0) = h_v(0,0)$, $D$ is a disk centered at the origin.

It seems like that the asymptotic of $I(\varepsilon)$ depends only on the first order behavior of $h$ at the origin, so the dominant term of $I(\varepsilon)$ is the same as that with $h\equiv 0$ and $$I(\varepsilon) \sim \frac{\pi}{\varepsilon^2}, \quad \varepsilon \rightarrow 0.$$

Any suggestion on how to prove/disprove it?

  • 1
    $\begingroup$ Does "smooth" mean $C^1$? $\endgroup$ – fedja Feb 26 at 23:57
  • $\begingroup$ I was thinking $C^\infty$, or even just $h(u,v) = \kappa_1 u^2 + \kappa_2 v^2$ as a start. $\endgroup$ – TYp Feb 26 at 23:59
  • 1
    $\begingroup$ Then split the domain of integration into $u^2+v^2\le\varepsilon^{4/3}$ and the complement. $\endgroup$ – fedja Feb 27 at 0:00
  • $\begingroup$ Do you mean that the dominant asymptotic of $I(\varepsilon)$ comes from restricting the integral to the $\varepsilon^{4/3}$ disc? Why is it so? And would it help proving the $\pi/\varepsilon^2$ asymptotic? Thanks. $\endgroup$ – TYp Feb 27 at 0:23
  • 1
    $\begingroup$ $\varepsilon^{2/3}$- disk, that is. Iosif used $3/4$ instead of my $2/3$, but the logic is the same. $\endgroup$ – fedja Feb 27 at 1:55

$\newcommand{\ep}{\varepsilon}$ This is indeed a matter of splitting the integral. In polar coordinates, for some real $R>0$, \begin{equation} I(\ep) = \int_0^{2\pi}(J_t(\ep)+K_t(\ep))\,dt, \end{equation} where \begin{equation} J_t(\ep):=\int_0^{r_*}\frac{\sqrt{1+|\nabla h|^2}}{\big((h+\ep)^2+r^2\big)^2}\,r\,dr, \end{equation} \begin{equation} K_t(\ep):=\int_{r_*}^R\frac{\sqrt{1+|\nabla h|^2}}{\big((h+\ep)^2+r^2\big)^2}\,r\,dr, \end{equation} $h:=h(r\cos t,r\sin t)$, $\nabla h:=\nabla h(r\cos t,r\sin t)$,
$r_*=r_*(\ep)>0$ varies with $\ep\downarrow0$ so that $$\ep<<r_*<<\ep^{1/2}$$ (e.g., one may take $r_*=\ep^{3/4}$), $a\ll b$ means $|a|=O(b)$, and $a<<b$ means $|a|=o(b)$.

We have $h\ll r^2$ and $|\nabla h|\ll r$, so that (as $\ep\downarrow0$) uniformly in $r\in[0,r_*]$ we have $h\ll r_*^2<<\ep$, $h+\ep\sim\ep$, $(h+\ep)^2+r^2\sim\ep^2+r^2$, $|\nabla h|<<1$, and $\sqrt{1+|\nabla h|^2}\sim1$. So, \begin{equation} J_t(\ep)\sim\int_0^{r_*}\frac{r\,dr}{\big(\ep^2+r^2\big)^2} =\frac1{2\ep^2}\,\int_0^{r_*^2/\ep^2}\frac{ds}{(1+s)^2} \sim\frac1{2\ep^2}\,\int_0^\infty\frac{ds}{(1+s)^2} =\frac1{2\ep^2}. \tag{1} \end{equation} On the other hand, \begin{equation} K_t(\ep)\ll\int_{r_*}^\infty\frac{r\,dr}{(r^2)^2}\le\frac1{r_*^2}<<\frac1{\ep^2}. \end{equation} Thus, indeed \begin{equation} I(\ep)\sim\frac\pi{\ep^2}. \end{equation}

  • $\begingroup$ It seems like your claim $J_t(\varepsilon) \sim \int_0^{r_\ast} \frac{r dr}{(\varepsilon^2 + r^2)^2}$ follows solely from the assumption that $r_\ast(\varepsilon) << \varepsilon^{1/2}$ (and that $h$ has a vanishing linearization at the origin.) I could not follow the argument, but now also think that it cannot be true: if all I have is an upper bound on $r_\ast$ , then I can make it zero and then $J_t$ is also zero. (Put differently, the $1/(2\varepsilon^2)$ asymptotic cannot follow just from an upper bound for $r_\ast$.) Am I missing anything? $\endgroup$ – TYp Feb 29 at 17:40
  • $\begingroup$ p.s. The proof for $K_t$ growing strictly slower than $1/\varepsilon^2$ is very clear, and uses only the lower bound for $r_\ast$. $\endgroup$ – TYp Feb 29 at 17:53
  • $\begingroup$ @TYp : I have added details on the first asymptotic equivalence, $\sim$, in (1). In that equivalence, you cannot set $r_*=0$, because then the equivalence would become the nonsensical "equivalence" $0\sim0$. More importantly, the lower bound on $r_*$ given by $\varepsilon<<r_*$ is needed for the second asymptotic equivalence in (1). $\endgroup$ – Iosif Pinelis Mar 1 at 1:55
  • $\begingroup$ Now I see the lower bound is used and the doubt is gone. Thanks. The technical issue I have is how that "interchange of integral and asymptotic" in the first equivalence can be justified. For simplicity, if $h(u,v)$ is the radially symmetric $1/2(u^2+v^2)$. Then $J_t(\varepsilon)$ is exactly $\int_0^{\varepsilon^{3/4}} \frac{\sqrt{1+r^2} r dr}{[(r^2+\varepsilon)^2+r^2]^2}$. And the claim is that for small $\varepsilon$ this is equivalent to $\int_0^{\varepsilon^{3/4}} \frac{r dr}{[\varepsilon^2+r^2]^2}$. I am trying to figure out the (elementary?) missing lemma. $\endgroup$ – TYp Mar 2 at 23:09
  • $\begingroup$ @TYp : When people write (as I did) "for $r\le r_*$ as $\epsilon\to0$", they of course mean "uniformly in $r\in[0,r_*]$ as $\epsilon\to0$" (which is how I have now rewritten that). So, the first equivalence in (1) follows by the general fact that, if $f_\epsilon(r)\sim g_\epsilon(r)>0$ uniformly in $r\in[0,r_*]$ (as $\epsilon\to0$), then $\int_0^{r_*}f_\epsilon(r)\,dr\sim \int_0^{r_*}g_\epsilon(r)\,dr$. This fact should be an easy exercise for you. $\endgroup$ – Iosif Pinelis Mar 3 at 3:25

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.