5

Suppose we have the following integral \begin{equation} \int_{-b}^{\infty}\log(t+b)e^{-t}e^{-e^{-t}}dt, \end{equation} where $b$ is a positive constant. It seems very difficult to derive the exact result. So my question is: is there any approximation result? I'm doubting that the approximation result will be just a simple polynomial of the constant $b$. Note that the term $e^{-e^{-t}}$ converge to 1 very fast due to its double exponential structure. Then how to proceed?

UPDATE:I use MATLAB to plot $f_1(b)=\int_{-b}^{\infty}\log(t+b)e^{-t}e^{-e^{-t}}dt$ and $f_2(b)=\log(b)$ in the region $b\in[0, 10]$. The observation is very interesting: they agree very well except a small constant offset...So I'm strongly doubting that this integral is approximately equal to $\log(b)+constant$!

UPDATE 2: please go to Prove $\int_{-b}^{\infty}\log^{\nu}(t+b)e^{-t}e^{-e^{-t}}dt\xrightarrow{b\rightarrow\infty} \log^{\nu}(b)$

3 Answers3

4

This is not an answer, I presume.

Computing the integral for $1\leq b\leq 50$ and playing with the numbers, I had a quite good fit $(R^2=0.999999)$ using $$y=0.979379 \log (b+0.131363)+0.085287$$ $$\begin{array}{clclclclc} \text{} & \text{Estimate} & \text{Standard Error} & \text{Confidence Interval} \\ a & 0.085287 & 0.003956 & \{0.0773,0.0932\} \\ b & 0.979379 & 0.001140 & \{0.9771,0.9817\} \\ c & 0.131363 & 0.007155 & \{0.1170,0.1458\} \\ \end{array}$$

Edit

There are many values of $b$ which makes the numerical integration quite difficult and probably not very accurate. What bothers me more is that, if i cover the range $0\leq b\leq 50$, the resulting fit is quite different $(R^2=0.999742)$ $$1.06113 \log (x+0.939639)-0.222393$$ $$\begin{array}{clclclclc} \text{} & \text{Estimate} & \text{Standard Error} & \text{Confidence Interval} \\ a & -0.222393 & 0.049874 & \{-0.3227,-0.1221\} \\ b & +1.061130 & 0.014504 & \{+1.0320,+1.0903\} \\ c & +0.939639 & 0.074340 & \{+0.7901,+1.0892\} \\ \end{array}$$

Update

As I wrote, I faced quite serious difficulties with the direct numerical integration. Interesting, may be, is to notice that all of them disappear if (as you suggested) we compute $$\int_{-b}^\infty \log (b+t)\,e^{-t-e^{-t}} dt=\int_{-b}^0\log (b+t)\,e^{-t-e^{-t}} dt+\int_0^\infty \log (b+t)\,e^{-t-e^{-t}} dt$$ For the range $0\leq b\leq 200$, each integral seems to follow the same $\alpha+\beta \log(b+\gamma)$ with very high values of the $R^2$ (always larger than $0.999$ and the results are $$\int_{-b}^0\log (b+t)\,e^{-t-e^{-t}} dt\approx 0.393913 \log (x+0.806893)-0.126909\qquad R^2=0.999551$$ $$\int_0^\infty \log (b+t)\,e^{-t-e^{-t}} dt\approx 0.622688 \log (x+0.767432)+0.047395\qquad R^2=0.999995$$ We can notice the similarity of the constant terms in the logarithm and, as a good compromise, a global curve fit leads to $$\int_{-b}^\infty \log (b+t)\,e^{-t-e^{-t}} dt\approx 1.01687 \log (x+0.786348)-0.0808138\qquad R^2=0.999959$$

  • Very interesting! But how can we prove that... Just fitting is not very convincing... – kawofengche Nov 12 '16 at 17:05
  • @kawofengche. I totally agree with you. This why I wrote This is not an answer, I presume. – Claude Leibovici Nov 12 '16 at 17:06
  • 1
    @ClaudeLeibovici, the leading order behavior is $e^{-1} \sqrt{2\pi} \log b$ as $b \to \infty$, and since $e^{-1}\sqrt{2\pi} \approx 0.922137$ this seems to more or less confirm the leading constant in your fit. I would expect that if you fit over a range which avoids $b=0$ (like $10 \leq b \leq 200$) the leading constant would agree more closely with $e^{-1}\sqrt{2\pi}$. – Antonio Vargas Nov 14 '16 at 09:03
  • @AntonioVargas Could you please give more detailed information on how to use Laplace method to derive your asymptotic result? – kawofengche Nov 14 '16 at 09:09
  • @AntonioVargas. Interesting remark. I redo the fit for $10 \leq b \leq 200$ and found $\approx 0.999048 \log (b+0.481461)+0.00536057$ and, for $100 \leq b \leq 1000$, $\approx 0.999981 \log (b+0.564595)+0.000140022$ which starts to be quite interesting. – Claude Leibovici Nov 14 '16 at 09:13
  • i think the leading order behaviour for $b\rightarrow \infty$ should be $1\cdot \log b$ – tired Nov 14 '16 at 09:35
  • @tired. This is exactly what I had in mind when I saw the result of the last fit. Cheers :-) – Claude Leibovici Nov 14 '16 at 09:36
  • @Claude, glad to hear that!. But let's see what antonio has to say, he is the absolute expert on asymptotic expansions here – tired Nov 14 '16 at 09:38
  • 1
    @tired, Claude, On closer inspection I agree with tired's asymptotic. – Antonio Vargas Nov 14 '16 at 12:46
  • @AntonioVargas, Claude ... i put everything into an answer – tired Nov 14 '16 at 17:27
4

We want to study the integral

$$I(b)=\int_{-b}^{\infty}\log\left(b+x\right)e^{-x}e^{-e^{-x}}dx$$

for the two cases $b\rightarrow\infty$ and $b\rightarrow 0^+$.

Let's start with the first one. We write

$$ I(b)=\log(b)\int_{-b}^{\infty}e^{-x}e^{-e^{-x}}+\int_{-b}^{\infty}\log\left(1+\frac{x}{b}\right)e^{-x}e^{-e^{-x}}dx $$

the first integral is elementary. Furthermore we sub $x/b=y$ in the second. This yields

$$ I(b)=\log(b)(1-e^{-e^{b}})+b\underbrace{\int_{-1}^{\infty}\log\left(1+y\right)e^{-b y}e^{-e^{-b y}}dy}_{J(b)} $$

Using Taylor expansion we get

$$ J(b)=\sum_{n\geq1}\frac{(-1)^n}{n}\int_{-1}^{\infty} y^n{e^{- b y}} e^{- e^{- b y}}dy $$

by using the fact that $e^{-e^{-b y}}\le 1$ we may show that the constituents of the above sum are bounded by terms of $\mathcal{O}(b^{-n-1})$ so to the first term will yield the dominant correction to the $\log(b)$ term.

$$ J(b)\sim\int_{-1}^{\infty} y{e^{- b y}} e^{- e^{- b y}}dy+\mathcal{O}(b^{-3}) $$

now substitue $e^{-by}=\xi$ and use the defintion of the exponential integral to show that ($\gamma$ is the Euler-Marschoni constant)

$$ J(b)\sim\frac{\gamma+b e^{-e^{b}}-\text{Ei}{(-e^b)}}{b^2}+\mathcal{O}(b^{-3})\sim \frac{\gamma}{b^2}+\mathcal{O}(b^{-3}) $$

and therefore

$$ I(b)\sim\log(b)+\frac{\gamma}{b}+\mathcal{O}(b^{-2})\quad\text{as}\quad b\rightarrow\infty $$


As $b\rightarrow 0$ we rewrite (because i'm in a hurry this part will be a bit more sketchy)

$$ I(b)=e^{b}\int_0^{\infty}\log(y)e^{-y}e^{-e^{b}e^{-y}} $$

For small $b$ we might write $e^{b}=1+b+\mathcal{O(b^2)}$

$$ I(b)=(1+b)\int_0^{\infty}\log(y)e^{-y}e^{-e^{-y}}(1-be^{-y})+\mathcal{O}(b^2) $$

or

$$ I(b)=C+b(C-D)+\mathcal{O}(b^2)\quad\text{as}\quad b\rightarrow 0^+ $$

where $C=\int_0^{\infty}\log(y)e^{-y}e^{-e^{-y}}$ and $D=\int_0^{\infty}\log(y)e^{-2y}e^{-e^{-y}}$ are contstants which have to determined numerically $(C\approx -0.155 ,D\approx0.262)$

tired
  • 12,588
2

Hint. One may write $$ \begin{align} \int_{-b}^{\infty}\log(t+b)e^{-t}e^{-e^{-t}}dt&\stackrel{u=t+b}{=}e^b\int_{0}^{\infty}\log(u)e^{-u}e^{-e^be^{-u}}du \\\\&=e^b\int_{0}^{\infty}\log(u)e^{-u}\sum_{n=0}^\infty \frac{(-1)^n}{n!}e^{nb}e^{-nu}du \\\\&=\sum_{n=0}^\infty \frac{(-1)^n}{n!}e^{(n+1)b}\int_{0}^{\infty}\log(u)e^{-(n+1)u}du \\\\&=\sum_{n=0}^\infty \frac{(-1)^n}{n!}e^{(n+1)b}\left(-\frac{\gamma}{n+1}-\frac{\log(n+1)}{n+1} \right) \\\\&=-\gamma \left(e^{-e^b}-1\right)+\sum_{n=1}^\infty \frac{(-1)^n}{n!}\cdot \log n \cdot e^{nb} \tag1 \end{align} $$ where $\gamma$ is the Euler–Mascheroni constant and where the latter infinite sum, being a standard alternating series, satisfies $$ \left|\sum_{n=1}^\infty\frac{(-1)^n}{n!}\cdot \log n \cdot e^{nb}\,-\,\sum_{n=1}^N\,\frac{(-1)^n}{n!}\cdot \log n \cdot e^{nb}\right|\le \left|\frac{\log N\cdot e^{Nb}}{N!} \right| \tag2 $$ and is then easy to approximate.

From $(1)$, one may deduce that a closed form can be given in terms of the generalized Bell polynomials:

$$ \int_{-b}^{\infty}\log(t+b)e^{-t}e^{-e^{-t}}dt=-\gamma \left(e^{-e^b}-1\right)+e^{-e^b}\left.\frac{\partial}{\partial n}\text{Bell}(n,-e^b)\right|_{n=0}. \tag3 $$

Olivier Oloa
  • 122,789
  • @Oliveir: Great! Thank you very much that you give one of the approximation methods! So I mark your answer as an answer. In addition, do you have any idea on proving our observation that this integral is very near to $\log(b)$? Because I'm really very interesting in that. Thanks again! – kawofengche Nov 12 '16 at 20:28
  • the first term in (1) may be not correct, you missed the summation, could you correct it? And I simulate this summation and found that when $b$ is large, for example, $b=4$, it has a very large error, no matter how much $N$ is... – kawofengche Nov 12 '16 at 22:23
  • So it seems diverge. – kawofengche Nov 12 '16 at 22:32
  • More than impressed by your approach ! I think I have a problem since, repeating all your steps, I arrive to $$\gamma \left(e^{-e^b}-1\right)+\sum_{m=1}^\infty \frac{(-1)^m}{m!}\cdot \log (m) \cdot e^{mb} \tag1$$ – Claude Leibovici Nov 13 '16 at 05:28
  • Following my last comment, the sum does not seem to converge at all. I am probably wrong somewhere. By the way, how would you compute or evaluate $$\left.\frac{\partial}{\partial n}\text{Bell}(n,-e^b)\right|_{n=0}. $$ – Claude Leibovici Nov 13 '16 at 05:51
  • Thank you very much to @kawofengche and to Claude Leibovici: there was a typo, now corrected. – Olivier Oloa Nov 13 '16 at 08:33
  • @ClaudeLeibovici The formal expression $\left.\frac{\partial}{\partial n}\text{Bell}(n,-e^b)\right|{n=0}$ is just computed by $\sum{n=1}^N \frac{(-1)^n}{n!}\cdot \log n \cdot e^{nb}$ with $N$ such that $\left|\frac{\log N\cdot e^{Nb}}{N!} \right| < \varepsilon $. – Olivier Oloa Nov 13 '16 at 08:36
  • OlivierOloa and Claude: I think that the correct form should be $\sum_{n=1}^{\infty}\frac{(-1)^n}{n!}e^{nb}(\gamma+\log n)$. Actually, this series diverge rather than converge due to the term $e^{nb}$, especially when $b$ is large, it diverge very fast, which means this approximation works only when $b$ is very small. – kawofengche Nov 13 '16 at 09:40
  • @kawofengche. I observed exactly the same. – Claude Leibovici Nov 13 '16 at 09:44
  • @ClaudeLeibovici, The reason why we have the term $e^{nb}$ is due to the interval $t\in (-b,0)$. For the interval $t\in (0,\infty)$, I think it will work very well. So the possible solution would be that we divide the integration into two intervals, and calculate them individually. – kawofengche Nov 13 '16 at 09:48
  • @kawofengche. This is possible. Try and add results to your post. Cheers. – Claude Leibovici Nov 13 '16 at 09:55
  • 1
    @kawofengche Sorry but the radius of convergence of $\sum_{n=0}^{\infty}\frac{x^n}{n!}=e^x$ is $R=\infty$, thus the series $\sum_{n=1}^{\infty}\frac{(-1)^n}{n!}e^{nb}(\gamma+\log n)$ is always convergent. – Olivier Oloa Nov 13 '16 at 10:10