2

I'm trying to tackle integrals of the form

$$ \int_{1}^{\infty} \frac{1}{x^{2k}+1} \ dx $$

for large $k \in \mathbb{Z}^{+}$. An easily computable general solution is desired, rather than, say, a massive recursing partial fractions summation.

I've started by considering a slightly tweaked integral with a lower bound of $0$ instead of $1$:

$$ \int_{0}^{\infty}\frac{1}{x^{2k}+1}\ dx $$

Substituting $x^{k}=\tan t$, I can get this into beta function form.

$$ \begin{align*} &\ \int_{\tan^{-1}\left(0\right)}^{\tan^{-1}\left(\infty\right)}\frac{1}{\left(\tan t\right)^{2}+1}\cdot\frac{1}{k}\left(\tan t\right)^{\frac{1}{k}-1}\left(\sec t\right)^{2}\ dt \\ =&\ \frac{1}{k}\int_{0}^{\pi/2}\left(\tan t\right)^{\frac{1}{k}-1}\ dt \\ =&\ \frac{1}{k}\int_{0}^{\pi/2}\left(\sin t\right)^{\frac{1}{k}-1}\left(\cos t\right)^{1-\frac{1}{k}}\ dt \end{align*} $$

So using the fact

$$ B\left(p,\ q\right)=2\int_{0}^{\pi/2}\left(\sin\theta\right)^{2p-1}\left(\cos\theta\right)^{2q-1}\ d\theta $$

and solving for $p$ and $q$, this evaluates to

$$ = \frac{1}{2k}B\left(\frac{1}{2k},\ 1-\frac{1}{2k}\right) $$

This unfortunately isn't quite an exact form, but it's more satisfying than numerical integration. (I should probably mention here that I'm not that familiar with the beta function, this is my first time using it in a solution)

However, when I now change the lower limit to $1$, the substituted limit becomes $\pi/4$.

$$ \frac{1}{k}\int_{\pi/4}^{\pi/2}\left(\sin t\right)^{\frac{1}{k}-1}\left(\cos t\right)^{1-\frac{1}{k}}\ dt $$

Now I'm at a loss for how to proceed. Intuition says split the integral, i.e. do

$$ \int_{0}^{\pi/2} ... dx - \int_{0}^{\pi/4} ... dx $$

but I don't know how to deal with the upper bound of $\pi/4$ either.

I've also considered looking at the graph of $y = (\tan{x})^{\frac{1}{k}-1}$, but this hasn't offered any insight into any symmetries or the like I can leverage.

Perhaps the beta function approach won't work here? I do remember seeing somewhere integrals like these can be solved through contour integration, but I haven't learnt that yet.

If this integral is solvable analytically, please could you offer some hints for how I can discover the solution?

Thanks for your time. Let me know if there are any issues or holes in the question.

Sup2.0
  • 103
  • 3

1 Answers1

3

We're evaluating $$I_k := \int_1^\infty \frac{dx}{1 + x^{2 k}}.$$

First, consider the case that $k$ is an integer (a positive one, which is necessary for convergence).

Mark Viola's answer to a related question gives that the indefinite integral is given by \begin{multline} I_k(x) := \int\frac{dx}{1 + x^{2 k}} \\ = \sum_{n=1}^{2 k}\underbrace{-\frac1{2 k}\left(\frac12 \cos \alpha_n \log(x^2-2\cos \alpha_n \cdot x+1)-\sin \alpha_n \arctan\left(\frac{x-\cos \alpha_n}{\sin \alpha_n}\right)\right)}_{A_n(x)}+C , \end{multline} where for compactness we denote $$\alpha_n := \frac{(2 n - 1) \pi}{2 k} .$$

The Fundamental Theorem of Calculus gives that our definite integral is $$I_k = \lim_{x \to \infty} (I_k(x) - I_k(1)).$$ Now, the individual summands $A_n(x)$ diverge as $x \to \infty$, but owing to the fact that $2 k$ is even, in the "antipodal" sums $A_m(x) + A_{k + m}(x)$ the logarithmic terms cancel as $x \to \infty$, so the limits $\lim_{x \to \infty} (A_m(x) + A_{k + m}(x))$ converge, and our integral is given by $$I_k = \sum_{m = 1}^k \left[A_m(x) + A_{k + m}(x)\right]_{x = 1}^\infty .$$

A straightforward computation evaluation gives $$\lim_{x \to \infty} (A_m(x) + A_{m + k}(x)) = -\frac\pi{2k} \sin \frac{(2 \pi - 1) n}{2 k},$$ and summing over $m$ gives $$\lim_{x \to \infty} I_k(x) = \sum_{m = 1}^k (A_m(x) + A_{m + k}(x)) = \sum_{m = 1}^k -\frac\pi{2k} \sin \frac{(2 \pi - 1) n}{2 k} = \frac{\pi \sin \frac\pi{2 k}}{k \left(1 - \cos \frac\pi k\right)} .$$

On the other hand, substituting gives \begin{align} I_k(1) &= -\frac1{2k} \sum_{n = 1}^{2 k} \left(\frac12 \cos \alpha_n \log (2 - 2 \cos \alpha_n) - \sin \alpha_n \arctan \left(\frac{1 - \cos \alpha_n}{\sin \alpha_n}\right)\right) \\ &= -\frac1{2k} \sum_{n = 1}^{2 k} \left(\frac12 \cos \alpha_n \log (2 - 2 \cos \alpha_n) - \frac{(2 n - 1) \pi}{4 k} \sin \alpha_n \right) ; \end{align} here the second equality follows from the tangent half-angle identity: $\frac{1 - \cos \alpha_n}{\sin \alpha_n} = \tan \frac{\alpha_n}2$.

So, our original integral is $$\small \boxed{I_k = \frac{\pi \sin \frac\pi{2 k}}{k \left(1 - \cos \frac\pi k\right)} + \frac1{2 k} \sum_{n = 1}^{2 k} \left(\frac12 \cos \frac{(2 n - 1) \pi}{2 k} \log \left(2 - 2 \cos \frac{(2 n - 1) \pi}{2 k}\right) - \frac{(2 n - 1) \pi}{4 k} \sin \frac{(2 n - 1) \pi}{2 k}\right)} . $$ Perhaps this expression can be simplified further still in terms of elementary functions.

The first several values are $$I_1 = \frac\pi4, \qquad I_2 = \frac1{2 \sqrt 2} (\pi - \operatorname{arcosh} 3), \qquad I_3 = \frac16 (\pi - \sqrt 3 \operatorname{arcosh} 2) .$$


We can extend our results to arbitrary, i.e., not necessarily integral, exponent $k > \frac12$ (that restriction is forced by convergence). Notice that if $k = \frac pq$ is rational ($p, q > 0$), then the substitution $x = u^q$ rationalizes the integrand, giving $q \int_1^\infty \frac{u^{q - 1} \,du}{1 + u^{2 p}}$ so in that case $I_k$ still admits a closed-form expression in terms of elementary functions. For typical irrational values, however, that will not be the case.

Digamma function Rewrite the integrand as a series in $\frac1 x$: $$\frac1{1 + x^{2k}} = \frac1{x^{2k}} \frac1{1 + x^{-2k}} = x^{-2k}\sum_{i=0}^\infty(-1)^i x^{-2 k i} = \sum_{i=0}^\infty(-1)^{i} x^{-2 k (1 + i)}.$$ Integrating gives \begin{multline*} \int_1^\infty \frac{dx}{1 + x^{2k}} = \int_1^\infty \sum_{n=0}^\infty(-1)^n x^{-2 k (1 + n)} \,dx = \sum_{i=0}^\infty(-1)^n \int_1^\infty x^{-2 k (1 + n)} \,dx \\ = \sum_{n=0}^\infty \frac{(-1)^n}{(2 k + 1) n - 1} = \boxed{\frac1{4k} \left[\psi \left(1 - \frac1{4k}\right) - \psi\left(\frac12 - \frac1{4k}\right)\right]} , \end{multline*} where $\psi := \frac{\Gamma'}{\Gamma}$ is the digamma function. The last equality follows from the identity $$\psi(z) = -\gamma + \sum_{n=0}^\infty \left(\frac{1}{1 + n} - \frac1{z + n}\right) ;$$ here $\gamma$ is the Euler–Mascheroni constant. This approach has the advantage that the usual asymptotic series for the digamma function immediately yields an asymptotic formula for $I_k$, namely, $$I_k \sim \frac{\log 2}{2 k} + \frac{\pi^2}{48 k^2} + \frac{3 \zeta(3)}{32 k^3} + \cdots,$$ where $\zeta(3)$ is Apéry's constant and $\cdots$ denotes a remainder in $O(k^{-4})$.

Incomplete beta function Instead substituting $x = \cot^{\frac1k} \theta$ transforms the integral to $$\frac1k \int_0^\frac\pi4 \sin^{1-\frac1k} \cos^{\frac1k-1} \theta \,d\theta = \boxed{\operatorname{B}\left(\frac12; 1 - \frac1{2k}, \frac1{2k}\right)} ,$$ where $\operatorname B$ now denotes the incomplete beta function.

Ordinary hypergeometric function The incomplete beta function is a special case of the ordinary hypergeometric function, ${}_2 F_1$, and so we can write the value of the integral, for example, as $$\frac1{2 k - 1}\cdot{}_2 F_1 \left(1, 1 - \frac1{2k}; 2 -\frac1{2k}; -1\right) .$$

Travis Willse
  • 108,056