12
$\begingroup$

I am wondering whether there is a closed form for the following integral for $n\in\mathbb{N}$:

$$\gamma(n)=\int_0^1...\int_0^1\frac{1}{\left(1+\sqrt{1+x_1^2+...+x_n^2}\right)^{n+1}}\;dx_1...dx_n\tag{*}$$

Particular values which I am aware of include:

$$\gamma(1)=\frac{4\sqrt{2}-5}{3}$$

$$\gamma(2)=\frac{5}{4}-\frac{9\sqrt{3}}{8}+\frac{\pi}{4}$$

Both of these values were obtained by evaluating different integrals to the above which solved the same problem (see below), and I am not sure how to attack $(*)$ directly; I can see no good way of solving this integral. I am wondering especially about the value of $\gamma(3)$; does it also have a simple closed form? Is there a closed form for all $n$?


Background: I posted a question here about calculating the proportion $p(n)$ of an $n$-cube closer to the centre than to the outside, which seems to me like an interesting problem. The $n=2$ case is simple to solve in terms of the following integral:

$$p(2)=2\int_{0}^{\sqrt{2}-1}\frac{1-x^2}{2}-x\;dx=\frac{4\sqrt{2}-5}{3}$$

I was able to write $p(3)$ as follows:

$$p(3)=6\int_{0}^{\frac{\sqrt{3}-1}{2}}\int_{z}^{\sqrt{2-z^2}-1}\frac{1-x^2-z^2}{2}-x\;dx\;dz$$

and I managed to evaluate this to $\frac{5}{4}-\frac{9\sqrt{3}}{8}+\frac{\pi}{4}$, but I was not able to use my method to solve the problem for higher dimensions. In the comments, however, achille hui made a proposition that we have $p(n)=\gamma(n-1)$ for all $n$ and although I still do not perfectly understand his reasoning, the claim does check out numerically for the two values I know already. Furthermore, the new integral is in a nice simple symmetric form (unlike the methods I had been using which required a case-by-case analysis for every dimension, with ugly bounds on the integrals), which makes me hope for a solution method. However, I really cannot see how to go about it. Thus I ask, is there a method for computing the integral $(*)$?

$\endgroup$
  • $\begingroup$ We can use "spherical coordinates" integrating first wrt the radius and then wrt the surface measure on the sphere $\endgroup$ – Omnomnomnom Feb 2 '17 at 6:59
  • 1
    $\begingroup$ Nice question! @Omnomnomnom Will that work out? The domain is not a ball. $\endgroup$ – mickep Feb 2 '17 at 6:59
  • $\begingroup$ @mickep wow, totally missed that! I'm not so sure anymore. $\endgroup$ – Omnomnomnom Feb 2 '17 at 7:00
  • 1
    $\begingroup$ For whatever it is worth, Mathematica gives the value $$\frac{1}{360} \bigl(170-1827 \sqrt{2} \text{arccot}\,\sqrt{2}+2979 \sqrt{2} \text{arccot}\,2 \sqrt{2}\bigr)\approx 0.032$$ for $n=3$. $\endgroup$ – mickep Feb 2 '17 at 7:41
  • 1
    $\begingroup$ If you are interested in the limit at $n \to \infty$, then for a similar case, @SangchulLee suggested using Strong law of large numbers. $\endgroup$ – Yuriy S Feb 2 '17 at 10:39
6
$\begingroup$

Just for reference, here is the derivation of the formula $p(n) = \gamma(n-1)$ proposed by @achille hui.

  • Consider the cube $\mathcal{C} = [-1, 1]^n$ and define $\mathcal{D} = \{x \in \mathcal{C} : |x| < \operatorname{dist}(x, \partial \mathcal{C}) \}$. Then for each point $x \in \partial\mathcal{C}$, we can find the unique point $q(x) \in \partial\mathcal{D}$ that lies on the line segment joining $0$ and $x$.

    $\hspace{7.5em}$ Point $q(x)$

    Using this, define $r(x)$ as the ratio

    $$ r(x) = \frac{\text{[length of the line segment from $0$ to $q(x)$]}}{\text{[length of line segment from $0$ to $x$]}} = \frac{|q(x)|}{|x|}. $$

  • Let me first give a geometric argument. Let $d\mathcal{S}$ be an infinitesimally small portion of the surface $\partial\mathcal{C}$ near $x$ and consider the cone $d\mathcal{V}$ with the base $d\mathcal{S}$ and the vertex $0$.

    $\hspace{7.5em}$ Infinitesimal cone $d\mathcal{V}$

    Then roughly speaking, the portion of $d\mathcal{V}$ that intersects $\mathcal{D}$ is similar to the cone $d\mathcal{V}$ with ratio $r(x)$. Thus their volume roughly satisfies

    $$ |\mathcal{D} \cap d\mathcal{V}| \approx r(x)^n |d\mathcal{V}| = r(x)^n \cdot \frac{1}{n}|d\mathcal{S}|. $$

    From this, we have

    $$ p(n) = \frac{|\mathcal{D}|}{|\mathcal{C}|} = \frac{\int |\mathcal{D} \cap d\mathcal{V}|}{\int |d\mathcal{V}|} = \frac{\int_{\partial \mathcal{C}} r(x)^n \frac{1}{n} \, dA}{\int_{\partial \mathcal{C}} \frac{1}{n} \, dA}. \tag{*} $$

  • Here is a rigorous justification of the argument above. Write

    \begin{align*} |\mathcal{D}| = \int_{\mathcal{C}} \mathbf{1}_{\mathcal{D}}(x) \, dx &= \int_{0}^{1} \int_{\partial[-s,s]^n} \mathbf{1}_{\mathcal{D}}(y) \, dyds \\ &= \int_{0}^{1} \int_{\partial\mathcal{C}} \mathbf{1}_{\mathcal{D}}(s\omega) \, s^{n-1}d\omega ds \\ &= \int_{\partial\mathcal{C}} \int_{0}^{1} s^{n-1} \mathbf{1}_{\mathcal{D}}(s\omega) \, ds d\omega. \end{align*}

    The note that $s\omega \in \mathcal{D}$ if and only if $s|\omega| < |q(\omega)|$, or equivalently, $s < r(\omega)$. From this, we have

    $$ |\mathcal{D}| = \int_{\partial\mathcal{C}} \int_{0}^{r(\omega)} s^{n-1} \, ds d\omega = \int_{\partial\mathcal{C}} \frac{1}{n} r(\omega)^n \, d\omega. $$

    Replacing $\mathbf{1}_{\mathcal{D}}$ by $\mathbf{1}$ from the above computation, we also get

    $$ |\mathcal{C}| = \int_{\partial\mathcal{C}} \frac{1}{n} \, d\omega. $$

    This proves $\text{(*)}$.

  • Now we give an explicit formula for $\text{(*)}$. Consider the top face of $\partial\mathcal{C}$. This face can be written as $[-1,1]^{n-1}\times\{1\}$. Then for each point $x = (x',1) \in [-1, 1]^{n-1}\times\{1\}$, it is not hard to verify that

    $$ r(x) = \frac{1}{1 + \sqrt{|x'|^2+1}}. $$

    (Using the rotational symmetry, it boils down to proving this when the dimension is $n = 2$.) Plugging this back and exploiting the symmetry, we finally have

    $$ p(n) = \int_{[0,1]^{n-1}} r(x',1)^n \, dx' = \int_{[0,1]^{n-1}} \frac{1}{\left( 1 + \sqrt{|x'|^2+1} \right)^n} \, dx' = \gamma(n-1). $$

$\endgroup$
  • $\begingroup$ +1. Thank you for this nice argument; that nicely explains why the formula is true. What do you use to make those pictures? $\endgroup$ – Anon Feb 2 '17 at 22:19
  • 1
    $\begingroup$ @Anon, I'm glad it was helpful. Anyway, I used Mathematica to draw them. $\endgroup$ – Sangchul Lee Feb 2 '17 at 23:23
5
$\begingroup$

Not a full solution, and too long for a comment, but perhaps a plausible strategy forward. The final formula is written in terms of a finite sum of double integrals, which may (or may not) be cracked.

First, rationalize by multiplying up and down by $\left(1-\sqrt{1+x_1^2+...+x_n^2}\right)^{n+1}$ $$ \gamma(n)=(-1)^{n+1}\int_0^1...\int_0^1\frac{\left(1-\sqrt{1+x_1^2+...+x_n^2}\right)^{n+1}}{\left(x_1^2+...+x_n^2\right)^{n+1}}\;dx_1...dx_n\ . $$ Next, use the identity $$ \frac{1}{X^{n+1}}=\int_0^\infty d\xi\frac{\xi ^n \exp (-\xi X)}{\Gamma (n+1)} $$ to rewrite your integral as $$ \gamma(n)=\frac{(-1)^{n+1}}{\Gamma(n+1)}\int_0^\infty d\xi\ \xi^n \int_0^1...\int_0^1\left(1-\sqrt{1+x_1^2+...+x_n^2}\right)^{n+1} e^{-\xi \left(x_1^2+...+x_n^2\right)}\;dx_1...dx_n\ $$ $$ =\frac{(-1)^{n+1}}{\Gamma(n+1)}\int_{1}^{n+1} dr\ (1-\sqrt{r})^{n+1}\int_0^\infty d\xi\ \xi^n \int_0^1...\int_0^1\delta(r-(1+x_1^2+...+x_n^2)) e^{-\xi \left(x_1^2+...+x_n^2\right)}\;dx_1...dx_n $$ $$ =\frac{1}{\Gamma(n+1)}\int_{1}^{n+1} dr\ (\sqrt{r}-1)^{n+1}\int_0^\infty d\xi\ \xi^n F_n(r,\xi)\ , $$ where $$ F_n(r,\xi)=\int_0^1...\int_0^1\delta(r-(1+x_1^2+...+x_n^2)) e^{-\xi \left(x_1^2+...+x_n^2\right)}\;dx_1...dx_n\ . $$ Therefore $$ F_n(r,\xi)=e^{-\xi (r-1)}\int_0^1...\int_0^1\delta(r-(1+x_1^2+...+x_n^2)) \;dx_1...dx_n $$ $$ =e^{-\xi (r-1)}\int_{-\infty}^\infty\frac{dk}{2\pi}e^{\mathrm{i}k(r-1)}\left[\int_0^1 dx\ e^{-\mathrm{i}kx^2}\right]^n\ , $$ using the integral representation of Dirac delta.

Writing $\exp(-\mathrm{i}kx^2)=\cos(kx^2)-\mathrm{i}\sin(kx^2)$ and using $$ \int_0^1 dx\cos(k x^2)=\frac{\sqrt{\frac{\pi }{2}} C\left(\sqrt{|k|} \sqrt{\frac{2}{\pi }}\right)}{\sqrt{|k|}} $$ and $$ \int_0^1 dx\sin(k x^2)=\mathrm{sign}(k)\frac{\sqrt{\frac{\pi }{2}} S\left(\sqrt{|k|} \sqrt{\frac{2}{\pi }}\right)}{\sqrt{|k|}}\ , $$ in terms of Fresnel integrals [Mathematica notation], and using the binomial theorem we get $$ F_n(r,\xi)=e^{-\xi (r-1)}\sum_{\ell=0}^n{n\choose \ell}(-\mathrm{i})^{n-\ell} \int_{-\infty}^\infty\frac{dk}{2\pi}e^{\mathrm{i}k(r-1)}\left(\frac{\pi}{2|k|}\right)^{n/2}C^\ell\left(\sqrt{|k|} \sqrt{\frac{2}{\pi }}\right)S^{n-\ell}\left(\sqrt{|k|} \sqrt{\frac{2}{\pi }}\right)(\mathrm{sign}(k))^{n-\ell}\ . $$ Putting everything together $$ \gamma(n)=\frac{1}{\Gamma(n+1)}\sum_{\ell=0}^n{n\choose \ell}(-\mathrm{i})^{n-\ell} \int_{-\infty}^\infty\frac{dk}{2\pi}\left(\frac{\pi}{2|k|}\right)^{n/2}C^\ell\left(\sqrt{|k|} \sqrt{\frac{2}{\pi }}\right)S^{n-\ell}\left(\sqrt{|k|} \sqrt{\frac{2}{\pi }}\right)(\mathrm{sign}(k))^{n-\ell}\times\int_1^{n+1}dr (\sqrt{r}-1)^{n+1} e^{\mathrm{i}k(r-1)}\int_0^\infty d\xi\ \xi^n e^{-\xi (r-1)}\ . $$ Performing the $\xi$-integral first $$ \gamma(n)=\frac{1}{\Gamma(n+1)}\sum_{\ell=0}^n{n\choose \ell}(-\mathrm{i})^{n-\ell} \int_{-\infty}^\infty\frac{dk}{2\pi}\left(\frac{\pi}{2|k|}\right)^{n/2}C^\ell\left(\sqrt{|k|} \sqrt{\frac{2}{\pi }}\right)S^{n-\ell}\left(\sqrt{|k|} \sqrt{\frac{2}{\pi }}\right)(\mathrm{sign}(k))^{n-\ell}\times\int_1^{n+1}dr (\sqrt{r}-1)^{n+1} e^{\mathrm{i}k(r-1)}\frac{1}{(r-1)^{n+1}}\ . $$

$\endgroup$
  • $\begingroup$ Now that's an interesting way of proceeding. +1. It looks like a beast of a Laplace transform though. $\endgroup$ – Anon Feb 2 '17 at 22:22
  • $\begingroup$ The inverse transform of $\frac{\operatorname{erf}{\sqrt{s}}}{\sqrt{s}}$ is $\frac{1}{\sqrt{\pi t}}u(1-t)$ (I think) so the inverse transform of the expression in the brackets will be $\phi(t)=\frac{e^{-\xi t}}{2\sqrt{t}}u(1-t)$ or something. Using convolution should (?) then give us $l(t)=\phi(t)*\phi(t)*...*\phi(t)$ for the inverse transform of the product, so then it seems we get the whole inverse transform as $u(t-1)l(t-1)$. I'm not sure if that's entirely correct or whether it will be useful? $\endgroup$ – Anon Feb 2 '17 at 22:37
  • $\begingroup$ What I said will just bring us back to an $n$-fold integral again, but it seems like it will be a different form. $\endgroup$ – Anon Feb 2 '17 at 22:46
  • $\begingroup$ Please check if after my edit, the new form looks more promising. $\endgroup$ – Pierpaolo Vivo Feb 3 '17 at 16:05
  • 1
    $\begingroup$ It is a complicated formula... I wonder whether the integral in $r$ can be evaluated (it looks the most promising). Note that $\frac{(\sqrt{r}-1)^{n+1}}{(r-1)^{n+1}}=\frac{1}{(\sqrt{r}+1)^{n+1}}$ so we are trying to integrate $\frac{e^{ik(r-1)}}{(\sqrt{r}+1)^{n+1}}$ (I don't know if that helps). $\endgroup$ – Anon Feb 7 '17 at 0:12

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.