6

I have the integral

$$I = \int_0^{2\pi} \mathrm{d}\theta\, \sqrt{k^2\cos^2\theta - \cos\theta + 1}$$

and I would need the asymptotic expansion of the integral for small values of $k$. For $k=0$ we get quite easily $I=4\sqrt{2}$, while for $k\gg 1$, we have $I\sim 4|k|$.

Now, trying to perform the integration using special functions (at first glance, one would think that it could be done in terms of elliptic functions) yields no result. Gradštejn and Ryžik are of no help, nor Mathematica or other softwares.

We can Taylor expand the integrand and integrate each term of the expansion, but the $2n$-th term, for $n>1$, reads

$$ \frac{k^{2n}}{(2n)!}\frac{ (-1)^{n-1} (2n-3)!! }{2^n} \frac{\cos^{2n}\theta}{(1-\cos \theta)^{n-1/2}}$$

having set $k=0$ in the derivative. Of course, integrating each term would mean performing

$$\int_0^{2\pi}\mathrm{d} \theta\,\frac{\cos^{2n}\theta}{(1-\cos \theta)^{n-1/2}}$$

which is divergent for $n>1$.

Any ideas about how to proceed? Are there some kind of 'regularization' techniques that could help?

  • 1
    There is an antiderivative (a big monster which I am unable to evaluate for $\theta=\pi$). Yourapproximation $I\sim 4|k|$ is really good – Claude Leibovici Jun 09 '21 at 10:15
  • Yes, but i need the approximation on the other side (for small $k$...) – fcapoani Jun 09 '21 at 10:36
  • 1
    @Maxim Perhaps you could post it as an answer. Yours is the only proper solution so far. – Gary Jun 10 '21 at 16:33
  • 2
    $I$ satisfies the ODE $$k^2 (1 - 4 k^2) (2 + k^2) I^{(3)}(k) + \ 4 k (1 - 7 k^2 - 6 k^4) I''(k) - \ (4 + 11 k^2 + 12 k^4) I'(k) + \ 3 k (1 + 4 k^2) I(k) = 0$$ (which can be reduced to a 2nd order ODE). The Frobenius method gives a convergent series $$I = \frac {c_0} k + c_1 - \frac {c_1} 4 k^2 \ln k + c_2 k^2 + \sum_{i \geq 1} (a_i k^{2 i + 2} + b_i k^{2 i + 2} \ln k)$$ with $a_i$ and $b_i$ given in terms of $c_1$ and $c_2$. We have $c_0 = 0$, $c_1 = 4 \sqrt 2$, $c_2 = (15 \ln 2 - 5)/(3 \sqrt 2)$. $c_2$ can be found by using this approach. – Maxim Jun 12 '21 at 21:11
  • Thanks @Maxim, I would say this solves the question. Could you post it as an answer so that I could accept it? – fcapoani Jun 14 '21 at 07:45

2 Answers2

3

Surprisingly, this integral can be written in closed form using a linear combination of complete elliptic functions. This doesn't exactly answer the question but at least the integral disappears.

Step 1: Let $u=\cos\theta$ so that $$I=2\int_{-1}^1\sqrt{\frac{1-u+k^2u^2}{1-u^2}}\,du.$$

Step 2: Notice that an appropriate substitution of the form $u=(Av+B)/(Cv+D)$ allows the linear term to be eliminated in the numerator. Choosing $A=D=1$ and $B=C=K$ where $$K=1+k^2-k\sqrt{2+k^2}$$ yields (after extensive simplification) $$I=\frac{MN}{K^2}\int_{-1}^1\frac1{(v+1/K)^2}\sqrt{\frac{1-v^2/N^2}{1-v^2}}\,dv$$ where $$M=2\sqrt{(1-K^2)(K-K^2-k^2)}\quad\text{and}\quad N=\sqrt{\frac{k^2K^2-K+1}{K-K^2-k^2}}.$$

Step 3: The complete elliptic integral of the first, second and third kinds are, respectively: \begin{align}K\left(\frac1N\right)&=\int_0^1\frac1{\sqrt{(1-v^2/N^2)(1-v^2)}}\,dv\\E\left(\frac1N\right)&=\int_0^1\frac{1-v^2/N^2}{\sqrt{(1-v^2/N^2)(1-v^2)}}\,dv\\\Pi\left(n\mid\frac1N\right)&=\int_0^1\frac1{(1-nv^2)\sqrt{(1-v^2/N^2)(1-v^2)}}\,dv.\end{align} Partial fraction decomposition gives \begin{align}\frac{K^2}{MN}I&=\int_{-1}^1\frac{-\frac1{N^2}+\frac2{KN^2}\cdot\frac1{v+1/K}+\left(1-\frac1{K^2N^2}\right)\cdot\frac1{(v+1/K)^2}}{\sqrt{(1-v^2/N^2)(1-v^2)}}\,dv\\&=-\frac2{N^2}K\left(\frac1N\right)+\frac2{KN^2}J_1+\left(1-\frac1{K^2N^2}\right)J_2.\end{align} Notice that \begin{align}aJ_2&=-bJ_1-c\int_{-1}^1\frac{v^2+d}{\sqrt{(1-v^2/N^2)(1-v^2)}}\,dv\\&=-bJ_1+2cN^2E\left(\frac1N\right)-2c(N^2+d)K\left(\frac1N\right)\end{align} where $$a=-\frac{K^{4}N^{2}+1-K^{2}(N^{2}+1)}{K^{4}N^{2}},\quad b=\frac{2-K^{2}(N^{2}+1)}{K^{3}N^{2}},\quad c=\frac1{N^2},\quad d=-\frac1{K^2}.$$ Thus $$\frac{K^2}{MN}I=fK\left(\frac1N\right)+gE\left(\frac1N\right)+h\int_{-1}^1\frac1{(v+1/K)\sqrt{(1-v^2/N^2)(1-v^2)}}\,dv$$ where \begin{align}f&=-\frac2{N^2}-\frac{2c(N^2+d)}a\left(1-\frac1{K^2N^2}\right)\\g&=\frac{2cN^2}a\left(1-\frac1{K^2N^2}\right)\\h&=\frac2{KN^2}-\frac ba\left(1-\frac1{K^2N^2}\right).\end{align}

1

Consider $$f\left( k \right)=2\int\limits_{0}^{\pi }{\sqrt{{{k}^{2}}{{\cos }^{2}}\left( t \right)-\cos \left( t \right)+1}dt}$$ Note that the derivative of ${{\left( {{k}^{2}}{{\cos }^{2}}\left( t \right)-\cos \left( t \right)+1 \right)}^{1/2}}$at $t=0$ doesn’t exist when $k=0$, and this is causing problems. So perhaps “zoom in” to the lower bound by perturbing it by say k, as a start, and then take the limit. For example suppose $$f\left( k \right)\simeq 2\int\limits_{k}^{\pi }{\sqrt{{{k}^{2}}{{\cos }^{2}}\left( t \right)+1-\cos \left( t \right)}dt}=\sqrt{\pi }\sum\limits_{m=0}^{\infty }{\frac{{{k}^{2m}}}{m!\Gamma \left( \tfrac{3}{2}-m \right)}}\int\limits_{k}^{\pi }{\frac{{{\cos }^{2m}}\left( t \right)}{{{\left( 1-\cos \left( t \right) \right)}^{m-\tfrac{1}{2}}}}dt}$$ So for $m=0,1$ say, we get $$f\left( k \right)\simeq 2\int\limits_{k}^{\pi }{\sqrt{1-\cos \left( t \right)}+\frac{{{k}^{2}}{{\cos }^{2}}\left( t \right)}{2\sqrt{1-\cos \left( t \right)}}dt}$$ Mathematica says this is $$f\left( k \right)\simeq -\frac{1}{3\sqrt{1-\cos \left( k \right)}}\left( 6{{k}^{2}}\sin\left( \frac{k}{2} \right)\log\left( \tan \left( \frac{k}{4} \right) \right)+2\sin \left( k \right)\left( {{k}^{2}}+{{k}^{2}}\cos \left( k \right)-6 \right) \right)$$ A series expansion about $k=0$ yields for small $k$ $$f\left( k \right)\simeq 4\sqrt{2}+\frac{12\log \left( 2 \right)-11}{3\sqrt{2}}{{k}^{2}}-\sqrt{2}{{k}^{2}}\log \left( k \right)+O\left( {{k}^{4}} \right)$$

This seems promising but we need to fix this up, so consider $$2\int\limits_{0}^{\pi }{\sqrt{{{k}^{2}}{{\cos }^{2}}\left( t \right)+1-\cos \left( t \right)}dt}\\=2\int\limits_{0}^{\phi }{k\cos \left( t \right)\sqrt{1+\frac{1-\cos \left( t \right)}{{{k}^{2}}{{\cos }^{2}}\left( t \right)}}dt}+2\int\limits_{\phi }^{\pi }{\sqrt{1-\cos \left( t \right)}\sqrt{\frac{{{k}^{2}}{{\cos }^{2}}\left( t \right)}{1-\cos \left( t \right)}+1}dt}$$ Where $\phi =\arccos \left( \frac{-1+\sqrt{1+4{{k}^{2}}}}{2{{k}^{2}}} \right)$ . A change of variables in the first integral and then expansion in $k$ yields $$\begin{aligned} 2\int\limits_{0}^{\phi }{k\cos \left( t \right)\sqrt{1+\frac{1-\cos \left( t \right)}{{{k}^{2}}{{\cos }^{2}}\left( t \right)}}dt}&=\int\limits_{0}^{1}{\cos \left( \phi t \right)\sqrt{1+\frac{1-\cos \left( \phi t \right)}{{{k}^{2}}{{\cos }^{2}}\left( \phi t \right)}}dt}\\& \simeq 2\sqrt{2}{{k}^{2}}\int\limits_{0}^{1}{\sqrt{1+{{t}^{2}}}dt}+O\left( {{k}^{4}} \right) \\ & =\left( 2+\sqrt{2}\arcsin h\left( 1 \right) \right){{k}^{2}}+O\left( {{k}^{4}} \right) \\ \end{aligned}$$ The second integral is tricky since it seems quite difficult to remove k from the terminals. So for example: $$\begin{aligned} 2\int\limits_{\sqrt{2}k}^{\pi }{\sqrt{1-\cos \left( t \right)}\sqrt{\frac{{{k}^{2}}{{\cos }^{2}}\left( t \right)}{1-\cos \left( t \right)}+1}dt}=2\sum\limits_{n=0}^{\infty }{\left( \begin{matrix} \tfrac{1}{2} \\ n \\ \end{matrix} \right){{k}^{2n}}}\int\limits_{\sqrt{2}k}^{\pi }{\frac{{{\cos }^{2n}}\left( t \right)}{{{\left( 1-\cos \left( t \right) \right)}^{n-1/2}}}dt}\end{aligned}$$

And taking $n=0,1$ we obtain for small $k$

$$\begin{aligned} 2\int\limits_{\sqrt{2}k}^{\pi }{\sqrt{1-\cos \left( t \right)}\sqrt{\frac{{{k}^{2}}{{\cos }^{2}}\left( t \right)}{1-\cos \left( t \right)}+1}dt}\\\simeq \int\limits_{\sqrt{2}k}^{\pi }{2\sqrt{1-\cos \left( t \right)}+\frac{{{k}^{2}}{{\cos }^{2}}\left( t \right)}{\sqrt{1-{{\cos }^{2}}\left( t \right)}}dt}\\\simeq 4\sqrt{2}+\frac{\left( 9\log \left( 2 \right)-14-6\log \left( k \right) \right){{k}^{2}}}{3\sqrt{2}}\end{aligned}$$

While higher terms of n contribute to say terms of $O\left( {{k}^{2}} \right)$ this is not the case for the $\log \left( k \right)$ term. So we have then $$2\int\limits_{0}^{\pi }{\sqrt{{{k}^{2}}{{\cos }^{2}}\left( t \right)+1-\cos \left( t \right)}dt}=4\sqrt{2}-\sqrt{2}\log \left( k \right){{k}^{2}}+O\left( {{k}^{2}} \right)$$

  • I'm just not really sure on one thing: why we put the lower integration bound at $k$? Shouldn't it be a generic $\epsilon$? In that case, I'd have in the final integration a term $\sqrt{2}k^2\log\epsilon$ which is the logarithmic divergence (and we can cancel it because as all the higher Taylor terms will generate logarithmic divergences their sum will converge to 0 if the integral converges) – fcapoani Jun 10 '21 at 10:03
  • This answer is incorrect. Why should the lower limit depend on $k$, and even if this is true, why does it have to linear? That is, why can’t the lower limit be something like $k^2$ or $k^3$? – ə̷̶̸͇̘̜́̍͗̂̄︣͟ Jun 10 '21 at 12:03
  • @fcapoani Indeed...why k? The intention of the post was more "what about something along these lines". I'll update it to better reflect this. When i first sketched it out i used the form ${{k}^{b}}$ and quickly determined that $0<b\le 1$ otherwise we don't get the right behavior at the origin. Then i moved to other forms, but its still unclear as to its exact nature as what's occurring is a double limit which seems quite delicate to handle. Note that whatever form it takes should represent a divergence in the derivative, which the above certainly doesnt. – mathstackuser12 Jun 10 '21 at 13:48
  • Let $g(t) = \cos^2(t)/\sqrt{1+\cos t}$. I can safely integrate $g(t)$ from $\epsilon$ to $\pi$. For the integration in $[0,\epsilon]$ I add a 'counterterm' to cancel the divergence: as $g(t)\approx \sqrt{2}/t$ for $t\to 0$, so I integrate from 0 to $\epsilon$ the quantity $(g(t)-\sqrt{2}/t)$, which is a convergent integral. The final sum is $\sqrt{2}(\log 4 -4/3) - \sqrt{2}\log\epsilon$. I would say that I can keep the "convergent" value and discard the $\log\epsilon$ term, which should be canceled summing higher terms of the expansion of $k$. But I'm not sure if this should work or not. – fcapoani Jun 10 '21 at 15:09