10

I am trying to evaluate the following integral as a function of $\alpha$: $$ \operatorname{I}\left(\alpha\right) = \int_{0}^{\infty}\frac{{\rm d}x}{x^{2} + 1} {\rm e}^{-\alpha\sqrt{\,x^{2} + 1\,}} $$ $\operatorname{I}\left(0\right) = \pi/2$ and for larger $\alpha$ it should decrease.

  • Since there is seemingly no way to exactly calculate this, I am wondering
    • if it would be possible to find a power series near $\alpha$ equal to zero ( where the result should be slightly smaller than $\pi/2$ )
    • and find the asymptotic form for $\alpha$ much larger than $1$ ( for which, the result should approximately exponentially decay $\mbox{in}\ \alpha$ ).
  • How to calculate the corrections to the integral for small $\alpha$ and how to obtain the asymptotic form for large $\alpha$ ?.
  • Can we interpolate the behavior at large and small $\alpha$ with a single, nice function ?.

Also, what would be the asymptotic form of the following (related) integral? $$ \operatorname J(\alpha) = \int_{0}^{\infty}{\rm d}x {\rm e}^{-\alpha\sqrt{\,x^{2} + 1\,}}$$

  • The function $I$ isn't differentiable at $\alpha = 0$, so there is no power series based there, but the integral might have another asymptotic form as $\alpha \searrow 0$. – Travis Willse Dec 07 '24 at 07:08
  • @TravisWillse There must be a power series for $\alpha > 0$. Plotting the numerically integrated $I(\alpha)$, I can see a finite slope for $\alpha \rightarrow 0^+$. – Archisman Panigrahi Dec 07 '24 at 07:10
  • 1
    There seems to be a logarithmic singularity at $0$: $\int_0^\infty \frac{1}{x^2 + 1} e^{-\alpha \sqrt{x^2 + 1}} ,dx \sim \frac\pi2 + \alpha \log \alpha - (1 + \log 2 - \gamma) \alpha + \cdots$, where $\gamma$ is the Euler–Mascheroni constant. – Travis Willse Dec 07 '24 at 07:44
  • The new integral added to the end of the question should probably be raised in a separate question, but notice that the substitution $x = \sinh t$ transforms it to $\int_0^\infty e^{-\alpha \cosh t} \cosh t ,dt$, which is just the modified Bessel function $K_1(\alpha)$. https://en.wikipedia.org/wiki/Bessel_function#Modified_Bessel_functions:_I%CE%B1,_K%CE%B1 The small-$\alpha$ asymptotics are $K_1(\alpha) \sim \alpha^{-1} + \frac12 \alpha \log \alpha + \cdots$, and the large-$\alpha$ asymptotics are $K_1(\alpha) \sim \sqrt{\frac{\pi}{2 \alpha}} e^{-\alpha} + \cdots$. – Travis Willse Dec 09 '24 at 22:39
  • @TravisWillse It can be easily obtained by twice differentiating the answer you obtained. I was in the process of editing your answer to add two lines, when you wrote this comment. – Archisman Panigrahi Dec 09 '24 at 22:40

2 Answers2

14

Per Gary's comment below, substituting $x = \sinh t$ yields $$I(\alpha) = \int_0^\infty \frac{e^{-\alpha \cosh t}}{\cosh t} \,dt .$$ This integral is a representation of the Bickley (or Bickley–Naylor) function $\operatorname{Ki}_1(\alpha)$ of order $1$; see DLMF $\S$10.43.12 and references therein. Alternatively, $I(\alpha) = \operatorname{Ki}_1(\alpha)$ can be expressed in terms of the Modified Bessel Function of the Second Kind, $K_m(\alpha)$, and the Modified Struve Functions, $L_m(\alpha)$.

To obtain asymptotics for $\operatorname{Ki}_1(\alpha)$, notice first that the form of the integral suggests differentiating under the integral sign w.r.t. $\alpha$, which yields $$\operatorname{Ki}_1'(\alpha) = -\int_0^\infty e^{-\alpha \cosh t} \,dt = - K_0(\alpha), $$ so we can derive the asymptotics of $\operatorname{Ki}_1$ using those of $K_0$.

About $\alpha = 0$ we have $$K_0(\alpha) \sim -\log \alpha + (\log 2 - \gamma) \alpha + \cdots ,$$ where $\gamma$ is the Euler–Mascheroni constant and $\cdots$ denotes terms of order $\alpha^2 \log \alpha$. Integrating and using that $\operatorname{Ki}_1(0) = \frac\pi2$ yields $$\boxed{\operatorname{Ki}_1(\alpha) \sim \frac\pi2 + \alpha \log \alpha + (\gamma - 1 - \log 2) \alpha + \cdots} ,$$ where $\cdots$ denotes a remainder in $O(\alpha^2 \log \alpha)$ (in fact in $O(\alpha^3 \log \alpha)$).

Similarly, expanding $K_0$ at $\infty$ gives $$K_0(\alpha) \sim \sqrt{\frac\pi{2 \alpha}}e^{-\alpha} \left(1 - \frac1{8\alpha} + \frac9{128 \alpha^2} + \cdots\right) ,$$ where $\cdots$ denotes a remainder in $O(\alpha^{-3})$. Integrating and using that $\lim_{\alpha \to \infty} \operatorname{Ki}_1(\alpha) = 0$ yields $$\boxed{\operatorname{Ki}_1(\alpha) \sim \sqrt{\frac\pi{2 \alpha}}e^{-\alpha} \left(1 - \frac5{8 \alpha} + \frac{129}{128 \alpha^2} + \cdots\right)}, $$ where again $\cdots$ denotes a remainder in $O(\alpha^{-3})$. The leading term can alternatively be extracted by instead substituting $x = \tan \theta$ in the original integral, which yields $\operatorname{Ki}_1(\alpha) = \int_0^{\frac\pi2} e^{-\alpha \sec \theta} \,d\theta$, and applying Laplace's Method with $f(\theta) = -\sec \theta$.

The function $\operatorname{Ki}_1(\alpha)$ and some of its asymptotic approximations $$\left\{\begin{array}{l} \operatorname{Ki}_1(\alpha) \\ \color{#7f0000}{\frac{\pi}{2} + \alpha \log \alpha + (\gamma - 1 - \log 2) \alpha} \\ \color{#7f3f00}{\frac{\pi}{2} + \alpha \log \alpha + (\gamma - 1 - \log 2) \alpha + \frac1{12} \alpha^3 \log \alpha + \frac1{36} (3\gamma - 4 - 3 \log 2) \alpha^3} \\ \color{#7f007f}{\sqrt{\frac\pi{2 \alpha}}e^{-\alpha}} \\ \color{#00007f}{\sqrt{\frac\pi{2 \alpha}}e^{-\alpha} \left(1 - \frac5{8\alpha}\right)} \end{array}\right.$$ plot of Ki_1 and some asymptotic approximations

Relative errors of the asymptotic approximations

plot of relative errors in the above approximations


For $k \geq 0$, $$I_k(\alpha) = \int_0^\infty (1 + x^2)^{\frac{k - 1}2} e^{-\alpha \sqrt{1 + x^2}} \,dx = \int_0^\infty e^{-\alpha \cosh t} \cosh^k t \,dt = (-1)^k \frac{d^k}{d\alpha^k} K_0(\alpha) .$$

In particular, the integral $J(\alpha) := \int_{0}^{\infty} e^{-\alpha\sqrt{1 + x^2}}$ is $I_1(\alpha) = -K_0'(\alpha) = K_1(\alpha)$.

Differentiating the earlier series for $K_0$ gives for small $\alpha$ that $$\boxed{K_1(\alpha) \sim \frac1\alpha + \frac12 \alpha \log \alpha + \frac14 (2 \gamma - 1 - 2 \log 2) + \cdots} ,$$ where $\cdots$ denotes a remainder in $O(\alpha^3 \log \alpha)$, and for large $\alpha$ that $$ \boxed{K_1(\alpha) \sim \sqrt{\frac{\pi}{2\alpha}}e^{-\alpha} \left(1 + \frac{3}{8 \alpha} - \frac{15}{128 \alpha^2} + \cdots \right)} ,$$ where $\cdots$ denotes a remainder in $O(\alpha^{-3})$.

Travis Willse
  • 108,056
  • 4
    If you substitute $x=\sinh t$, can you express it via https://dlmf.nist.gov/10.43.iii ? It seems to be a primitive of the $K_0$ Bessel function. – Gary Dec 07 '24 at 07:22
  • That looks right---I'd never heard of that Bickley functions. Feel free to write up your comment as an answer, or, if you prefer, I can add expand my own. – Travis Willse Dec 07 '24 at 07:39
  • Incidentally, series expansions are available, and the integral has a logarithmic singularity at $0$: $\int_0^\infty \frac{1}{x^2 + 1} e^{-\alpha \sqrt{x^2 + 1}} ,dx \sim \frac\pi2 + \alpha \log \alpha - (1 + \log 2 - \gamma) \alpha + \cdots$. – Travis Willse Dec 07 '24 at 07:43
  • 1
    Feel free to expand your answer. The DLMF also has some further references. – Gary Dec 07 '24 at 08:37
  • 1
    Fantastic approximation even for small $\alpha$ – Claude Leibovici Dec 07 '24 at 08:53
  • 1
    @Gary I've overhauled my answer (the previous version persists as a comment about an alternative method, just before the plot). Thanks again for the helpful comment! – Travis Willse Dec 07 '24 at 08:57
  • The asymptotics of, $K_0$ doesn't have a negative sign and it's remainder is of order $\alpha^{-3}$. I think if you write $\ldots$, it looks like an asymptotic expansion and hence you should use $\sim$ in both instances. – Gary Dec 07 '24 at 09:03
  • The large $\alpha$ series don't converge. – Gary Dec 07 '24 at 09:10
  • Those are not Laurent series. They can't be as the functions are not analytic in any neighborhood of infinity. Think about the singularities at the origin. The terms of the series diverge like a factorial divided by a power. Nevertheless, they are asymptotic expansions in the sense of Poincaré. The ones near the origin do converge though. – Gary Dec 07 '24 at 09:15
  • @TravisWillse Can you please add the next order term $-(1+\log 2−\gamma)\alpha$ to the plot of small $\alpha$? I plotted and checked that it makes the approximation way better between $0<\alpha < 1$. Thank you – Archisman Panigrahi Dec 07 '24 at 14:52
  • 1
    @ArchismanPanigrahi I've updated the plot and added a second plot showing relative error of the asymptotic approximations. – Travis Willse Dec 07 '24 at 22:26
2

This is too long for a comment.

Starting from @Travis Willse's nice answer, the expansion for small $\alpha$ could write $$I(\alpha)= \frac \pi 2+ (A-1)\alpha+\frac{3 A-4}{36} \alpha ^3 +\frac{ 10 A-17}{3200}\alpha ^5+\frac{ 42 A-83}{677376}\alpha ^7+$$ $$\frac{36 A-79}{47775744}\alpha^9+\frac{660 A-1567}{107053056000}\alpha ^{11} +O\left(\alpha ^{13}\right)$$ where $$A=\log \left(\frac{\alpha}{2}\right)+\gamma$$

Unfortunately, no sequence has been found in $OEIS$ for the different numbers.

The above expansion is quite good for $\alpha < 3$

For $\alpha=2$, the above gives $$I(2)\sim \frac{\pi }{2}+\frac{3461537 }{1247400}\,\gamma-\frac{1661604262}{540280125}=\color{red}{0.09712}12$$