2
$\begingroup$

I want to find an asymptotic approximation when $a\to 0^+$ for the integral $$I(a):=\int_0^\infty \int_0^{a/x}e^{-x-y}\ dy\ dx.$$ I found the following approximation: $$C_1\, a\, \mathrm{ln}(1/a) < I(a)< C_2 \sqrt{a},$$ where $C_1$ and $C_2$ are constant terms.

I would like to have a better approximation, something of the type $I(a)=\Theta(\sqrt{a})$ or $I(a)=\Theta(a\, \mathrm{ln}(1/a))$ when $a\to 0^+$.

In the folowing I explain how I found those bounds. First we split our integral: $$I(a):=\int_0^\sqrt{a} \int_0^\sqrt{a} e^{-x-y}\ dy\ dx + \int_0^\sqrt{a} \int_\sqrt{a}^{a/x}e^{-x-y}\ dy\ dx +\int_\sqrt{a}^\infty \int_0^{a/x}e^{-x-y}\ dy\ dx.$$ But the two last integral are equal (clear if you draw the picture of the domain of integration and observe the symmetry), so we have: $I(a)=I_1(a)+2\, I_2(a)$ with $$I_1(a):=\int_0^\sqrt{a} \int_0^\sqrt{a} e^{-x-y}\ dy\ dx\quad \text{ and }\quad I_2(a):=\int_\sqrt{a}^\infty \int_0^{a/x}e^{-x-y}\ dy\ dx.$$ We consider that $a$ is small, so we have $0<a<1$ and in the domain of integration of $I_1(a)$, we have $e^{-2}<e^{-x-y}<1$. Hence $e^{-2} a <I_1(a)< a$. That is $$I_1(a)=\Theta(a).$$ This first integral will not play any role for our bounds, because the second integral will be much bigger (for small values of $a$).

Now we observe that if $x>\sqrt(a)$, then $a/x<a/\sqrt{a}=\sqrt{a}$. Hence $$I_2(a)<\int_\sqrt{a}^\infty \int_0^\sqrt{a}e^{-x-y}\ dy\ dx <\int_\sqrt{a}^\infty e^{-x} \int_0^\sqrt{a}1\ dy\ dx =\sqrt{a}\,e^{-\sqrt{a}}<\sqrt{a}.$$ So we have proved the upper bound, and will show the above bound now. We have for $x>\sqrt{a}>a$ and $a<1$, $$\int_0^{a/x}e^{-y}\ dy> \frac{a}{x} e^{-a/x} >\frac{a}{x} e^{-1}.$$ So $$I_2(a)> a\,e^{-1} \int_\sqrt{a}^\infty \frac{e^{-x}}{x}\ dx >a\,e^{-1} \int_\sqrt{a}^1 \frac{e^{-x}}{x}\ dx >a\,e^{-2} \int_\sqrt{a}^1 \frac{1}{x}\ dx =\frac{e^{-2}}{2}\,a\,\textrm{ln}(1/a).$$ This gives us the lower bound.

$\endgroup$
1
  • 1
    $\begingroup$ It should be noted that your integral is $1-2\sqrt{a}K_1\left(2\sqrt{a}\right)$, where $K_\nu(z)$ is the modified Bessel function of the second kind of order $\nu$. The asymptotics can then be obtained from known series for $K_\nu$, for instance formula (4) in this MathWorld article. $\endgroup$ – Antonio Vargas Nov 18 '13 at 18:32
1
$\begingroup$

Your method will definitely work, we just need to keep track of the errors along the way.

You showed that

$$ \begin{align} I(a) &= 2I_2(a) + \Theta(a) \\ &= 2\int_{\sqrt{a}}^\infty e^{-x} - e^{-a/x-x}\,dx + \Theta(a). \tag{1} \end{align} $$

As you observed, the tail of the integral is insignificant. In fact we have

$$ \begin{align} 0 &\leq \int_{1}^\infty e^{-x} - e^{-a/x-x}\,dx \\ &\leq \int_{1}^\infty e^{-x} - e^{-a/1-x}\,dx \\ &= (1-e^{-a}) \int_1^\infty e^{-x}\,dx, \end{align} $$

so that

$$ \int_{1}^\infty e^{-x} - e^{-a/x-x}\,dx = O(a). \tag{2} $$

In the remaining interval $x \in [\sqrt{a},1]$ we have $e^{-a/x} = 1 - a/x + O(a/x)^2$. Substituting this into the integral yields

$$ \begin{align} \int_\sqrt{a}^1 e^{-x} - e^{-a/x-x}\,dx &= \int_{\sqrt{a}}^1 e^{-x} - e^{-x}\Bigl[1 - a/x + O(a/x)^2\Bigr]\,dx \\ &= a \int_\sqrt{a}^1 \frac{e^{-x}}{x}\,dx + O\left(a^2\int_\sqrt{a}^1 \frac{e^{-x}}{x^2}\,dx\right). \tag{3} \end{align} $$

In the integrals we substitute $e^{-x} = 1 + O(x)$ to get

$$ \int_\sqrt{a}^1 \frac{e^{-x}}{x}\,dx = \int_\sqrt{a}^1 \frac{dx}{x} + O(1) = \frac{\log(1/a)}{2} + O(1) $$

and

$$ \int_\sqrt{a}^1 \frac{e^{-x}}{x^2}\,dx = \frac{1}{\sqrt{a}} + O(\log a). $$

Thus $(3)$ becomes

$$ \int_\sqrt{a}^1 e^{-x} - e^{-a/x-x}\,dx = \frac{a\log(1/a)}{2} + O(a). $$

Combining this with $(1)$ and $(2)$, we conclude that

$$ I(a) = a\log\left(\frac{1}{a}\right) + O(a) $$

as $a \to 0^+$.

$\endgroup$

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.