21
$\begingroup$

Set $$ I_n :=\int_0^1\left(\frac{1}{\log x} + \frac{1}{1-x}\right)^n \:\mathrm{d}x \qquad n=1,2,3,\cdots. $$

We have $$I_1 =\gamma, \quad I_2 =\log (2 \pi) - \frac 32, \quad I_3 = 6 \log A - \frac{31}{24}, \quad I_4 = 2 \log A + \frac{5 \zeta(3)}{2\pi^{2}}- \frac{49}{72}, \quad ...$$ where $A$ is the Glaisher-Kinkelin constant defined by $$ \begin{equation} \displaystyle A :=\lim_{n\to\infty}\frac{1^22^2\cdots n^n}{e^{-n^2/4}n^{\frac{n^2+n}{2}+\frac{1}{12}}}=1.28242712\cdots. \end{equation} $$

I wonder if there is a "simple" equivalent for $I_n$ as $n$ tends to $+\infty$?

Edit. I had designed the following integral $$\displaystyle \int_0^1\left(\frac{1}{\log x} + \frac{1}{1-x}\right)^2 \mathrm{d}x$$ which I submitted to American Mathematical Monthly (March 2012, problem 11629), the problem was then spread and came in this forum with different interesting solutions (I). An interesting general formula for $I_n$ has been found (II), but I don't think the latter formula is tractable for the above question on asymptotics.

$\endgroup$
  • $\begingroup$ You've done a few others of this form. Can you link them in your question? $\endgroup$ – Semiclassical Jul 24 '14 at 3:34
  • 1
    $\begingroup$ @Mhenni Benghorbal Yes: $$\frac 12 < \frac{1}{\log(1-x)} + \frac{1}{x} < 1, \quad 0<x<1.$$ Thanks. $\endgroup$ – Olivier Oloa Jul 24 '14 at 8:15
  • 1
    $\begingroup$ @Mhenni Benghorbal Please see the edit in the question. $\endgroup$ – Olivier Oloa Jul 24 '14 at 9:26
  • 1
    $\begingroup$ @OlivierOloa: Thanks for these links. $\endgroup$ – Mhenni Benghorbal Jul 24 '14 at 9:31
  • 2
    $\begingroup$ @Venus There is no duplicate, since in the question above, we were looking for an equivalent as $n$ tends to $+\infty$, which is very different from a closed-form, the latter doesn't give a simple equivalent of the integral. So you should remove the inappropriate term "duplicate" here. Thanks. $\endgroup$ – Olivier Oloa Dec 20 '14 at 18:55
21
$\begingroup$

The main idea is that, near the largest point of the integrand (which occurs at $x=0$), we have

$$ \frac{1}{\log x} + \frac{1}{1-x} \approx \frac{1}{\log x} + 1. $$

So we split the integral up as

$$ \int_0^1 \left(\frac{1}{\log x} + \frac{1}{1-x}\right)^n\,dx = \left[\int_0^{1/e} + \int_{1/e}^1 \right] \left(\frac{1}{\log x} + \frac{1}{1-x}\right)^n\,dx. $$

For the second integral we have the bound

$$ 0 < \int_{1/e}^{1} \left(\frac{1}{\log x} + \frac{1}{1-x}\right)^n\,dx < (e-1)^{-n} $$

and for the first integral we write

$$ \begin{align} \int_0^{1/e} \left(\frac{1}{\log x} + \frac{1}{1-x}\right)^n\,dx &= \int_0^{1/e} \exp\left\{n \log\left(\frac{1}{\log x} + \frac{1}{1-x}\right)\right\}dx \\ &= \int_0^{1/e} \exp\left\{n \log\left(\frac{1}{\log x} + 1\right) + n f(x)\right\}dx, \end{align} $$

where

$$ f(x) = \log\left(1+\frac{x}{\left(\frac{1}{\log x}+1\right)(1-x)}\right) = O(x) \quad \text{as $x \to 0$.} $$

We therefore expect that the largest contribution comes from a neighborhood of radius $1/n$ of the point $x=0$. The rest of the integral is bounded by

$$ 0 < \int_{1/n}^{1/e} \left(\frac{1}{\log x} + \frac{1}{1-x}\right)^n\,dx < \left(-\frac{1}{\log n} + \frac{n}{n-1}\right)^n = O(e^{-n/\log n}). $$

For $0 \leq x \leq 1/n$ we have

$$ \exp\{nf(x)\} = 1 + O(nx), $$

so, after combining our estimates, we arrive at

$$ \begin{align} \int_0^1 \left(\frac{1}{\log x} + \frac{1}{1-x}\right)^n\,dx &= \int_0^{1/n} \exp\left\{n \log\left(\frac{1}{\log x} + 1\right)\right\}dx \\ &\qquad + O\left(n \int_0^{1/n} x \exp\left\{n \log\left(\frac{1}{\log x} + 1\right)\right\}dx\right)\\ &\qquad + O(e^{-n/\log n}). \end{align} $$

Now we make the change of variables $-y = \log\left(\frac{1}{\log x} + 1\right)$, transforming our integrals into

$$ \begin{align} &\int_0^{1/n} \exp\left\{n \log\left(\frac{1}{\log x} + 1\right)\right\}dx \\ &\qquad = \int_0^{-\log(1-1/\log n)} e^{-ny} \exp\left\{\frac{1}{e^{-y}-1}\right\} \frac{e^{-y}}{(1-e^{-y})^2}\,dy \end{align} $$

and

$$ \begin{align} &\int_0^{1/n} x\exp\left\{n \log\left(\frac{1}{\log x} + 1\right)\right\}dx \\ &\qquad = \int_0^{-\log(1-1/\log n)} e^{-ny} \exp\left\{\frac{2}{e^{-y}-1}\right\} \frac{e^{-y}}{(1-e^{-y})^2}\,dy. \end{align} $$

We can expand the factor which is independent of $n$ as

$$ \begin{align} \exp\left\{\frac{1}{e^{-y}-1}\right\} \frac{e^{-y}}{(1-e^{-y})^2} &= \exp\left\{-\frac{1}{y}-\frac{1}{2}\right\} \sum_{j=-2}^{\infty} c_j y^j \\ &= \exp\left\{-\frac{1}{y}-\frac{1}{2}\right\} \left( \frac{1}{y^2} -\frac{1}{12 y} - \frac{23}{288} + \frac{427 y}{51840} + \cdots \right), \end{align} $$

so on truncating the expansion at $j=J-1$ we obtain

$$ \begin{align} \int_0^1 \left(\frac{1}{\log x} + \frac{1}{1-x}\right)^n\,dx &= \sum_{j=-2}^{J-1} \frac{c_j}{\sqrt{e}} \int_0^{-\log(1-1/\log n)} e^{-ny} e^{-1/y} y^j \,dy \\ &\qquad + O\left(\int_0^{-\log(1-1/\log n)} e^{-ny} e^{-1/y} y^J \,dy\right) \\ &\qquad + O\left(n \int_0^{-\log(1-1/\log n)} e^{-ny} e^{-2/y} y^{-2} \,dy \right) \\ &\qquad + O(e^{-n/\log n}). \end{align} $$

All that remains is to reattach the tails of the integrals. By doing so we add an error of no more than

$$ \begin{align} &\int_{-\log(1-1/\log n)}^\infty e^{-ny} e^{-1/y} y^j \,dy \\ &\qquad = \int_{-\log(1-1/\log n)}^\infty e^{-ny/2} \left(e^{-ny/2} e^{-1/y} y^j\right) \,dy \\ &\qquad \leq \sqrt{\int_{-\log(1-1/\log n)}^\infty e^{-ny}\,dy} \cdot \sqrt{\int_{-\log(1-1/\log n)}^\infty e^{-ny} e^{-2/y} y^{2j}\,dy} \\ &\qquad = \sqrt{\frac{2}{n}} \left(1-\frac{1}{\log n}\right)^{n/2} \sqrt{\int_{-\log(1-1/\log n)}^\infty e^{-ny} e^{-2/y} y^{2j}\,dy} \\ &\qquad < \sqrt{\frac{2}{n}} \left(1-\frac{1}{\log n}\right)^{n/2} \sqrt{\int_{0}^\infty e^{-ey} e^{-2/y} y^{2j}\,dy} \\ &\qquad = O\left(n^{-1/2} e^{-n/(2\log n)}\right) \end{align} $$

for $n > e$ by the Cauchy-Schwarz inequality, with the same bound when $e^{-1/y}$ is replaced with $e^{-2/y}$. Thus we have

$$ \begin{align} \int_0^1 \left(\frac{1}{\log x} + \frac{1}{1-x}\right)^n\,dx &= \sum_{j=-2}^{J-1} \frac{c_j}{\sqrt{e}} \int_0^\infty e^{-ny} e^{-1/y} y^j \,dy \\ &\qquad + O\left(\int_0^\infty e^{-ny} e^{-1/y} y^J \,dy\right) \\ &\qquad + O\left(n \int_0^\infty e^{-ny} e^{-2/y} y^{-2} \,dy \right) \\ &\qquad + O\left(n^{-1/2} e^{-n/(2 \log n)}\right). \end{align} $$

To normalize these integrals a little bit we can substitute $y = u/\sqrt{n}$ to get

$$ \begin{align} \int_0^1 \left(\frac{1}{\log x} + \frac{1}{1-x}\right)^n\,dx &= \sum_{j=-2}^{J-1} \frac{c_j}{\sqrt{e}} n^{-(j+1)/2} \int_0^\infty e^{-\sqrt{n}(u+1/u)} u^j \,du \\ &\qquad + O\left(n^{-(J+1)/2} \int_0^\infty e^{-\sqrt{n}(u+1/u)} u^J \,du\right) \\ &\qquad + O\left(n^{3/2} \int_0^\infty e^{-\sqrt{n}(u+2/u)} u^{-2} \,du \right) \\ &\qquad + O\left(n^{-1/2} e^{-n/(2 \log n)}\right). \end{align} $$

With the basic Laplace method estimate

$$ \int_0^\infty e^{-\sqrt{n}(u+\alpha/u)} u^j \,du \sim \alpha^{1/4} \sqrt{\frac{\pi}{n}} e^{-2\sqrt{\alpha n}} $$

we may conclude that the series we've written is actually an asymptotic series, i.e.

$$ \int_0^1 \left(\frac{1}{\log x} + \frac{1}{1-x}\right)^n\,dx \sim \sum_{j=-2}^{\infty} \frac{c_j}{\sqrt{e}} n^{-(j+1)/2} \int_0^\infty e^{-\sqrt{n}(u+1/u)} u^j \,du. $$

Here we recognize that

$$ \int_0^\infty e^{-\sqrt{n}(u+1/u)} u^j \,du = 2 K_{-j-1}\left(2\sqrt{n}\right), $$

where $K_\nu$ is the modified Bessel function of the second kind (see eq. 10.32.10, DLMF).

One asymptotic series for our integral is thus $$ \int_0^1 \left(\frac{1}{\log x} + \frac{1}{1-x}\right)^n\,dx \sim \frac{2}{\sqrt{e}}\sum_{j=-2}^{\infty} c_j n^{-(j+1)/2} K_{-j-1}\left(2\sqrt{n}\right), $$ where the coefficients $c_j$ are defined by $$ \exp\left\{\frac{1}{e^{-y}-1}\right\} \frac{e^{-y}}{(1-e^{-y})^2} = \exp\left\{-\frac{1}{y}-\frac{1}{2}\right\} \sum_{j=-2}^{\infty} c_j y^j. $$

We can transform this asymptotic series into one in terms of elementary functions by using the known asymptotic expansion for $K_\nu(r)$ as $r \to \infty$ (DLMF reference). In our case we have

$$ K_{-j-1}\left(2\sqrt{n}\right) \sim \frac{\sqrt{\pi}}{2} n^{-1/4} e^{-2\sqrt{n}} \sum_{k=0}^{\infty} \frac{a_k(j+1)}{2^k} n^{-k/2}, $$

where $a_0(j+1) = 1$ and

$$ a_k(j+1) = \frac{1}{k!8^k}\prod_{m=1}^{k} \left(4(j+1)^2-(2m-1)^2\right). $$

Substituting this into our asymptotic expansion yields

$$ \begin{align} \frac{2}{\sqrt{e}}\sum_{j=-2}^{\infty} c_j n^{-(j+1)/2} K_{-j-1}\left(2\sqrt{n}\right) &= \sqrt{\frac{\pi}{e}} n^{-1/4} e^{-2\sqrt{n}} \sum_{j=-2}^{\infty} \sum_{k=0}^{\infty} \frac{c_j a_k(j+1)}{2^k} n^{-(j+k+1)/2} \\ &= \sqrt{\frac{\pi}{e}} n^{1/4} e^{-2\sqrt{n}} \sum_{\ell = 0}^{\infty} \left(\sum_{j+k+2=\ell} \frac{c_j a_k(j+1)}{2^k}\right)n^{-\ell/2}, \end{align} $$

So we conclude that

As $n \to \infty$, $$ \begin{align} &\int_0^1 \left(\frac{1}{\log x} + \frac{1}{1-x}\right)^n\,dx \\ &\qquad \sim \sqrt{\frac{\pi}{e}} n^{1/4} e^{-2\sqrt{n}} \sum_{\ell = 0}^{\infty} \left(\sum_{j+k+2=\ell} \frac{c_j a_k(j+1)}{2^k}\right)n^{-\ell/2} \\ &\qquad = \sqrt{\frac{\pi}{e}} n^{1/4} e^{-2\sqrt{n}} \left(1 + \frac{5}{48} n^{-1/2} - \frac{479}{4608} n^{-1} + \frac{15313}{3317760} n^{-3/2} + \cdots \right), \end{align} $$ where the indices of the inner sum range over $j=-2,-1,0,1,2,\ldots$ and $k=0,1,2,3,\ldots$.

$\endgroup$
  • 1
    $\begingroup$ Very nice job! I'm impressed. $\endgroup$ – Olivier Oloa Jul 25 '14 at 10:29

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.