3
$\begingroup$

I would like to know the asymptotics of the following sequences of integrals: $$ I_n = \int _0 ^{+ \infty} e^{-t} \left ( \dfrac{t}{1 + t} \right )^n \ dt $$

I have tried using Laplace method ou saddle node method, but I have been unable to conclude anything.

I have tried to find out the behaviour using a software computation. Here is what I found: \begin{equation} \begin{array}{|c|c|c|} \hline \\ n & I_n & - \ln(I_n) \\ \hline 1 = 2^0 & 1.9269472 \cdot 10^{-1} & 1.646648079928304 \\ \hline 2 = 2^1 & 8.7215768 \cdot 10^{-2} & 2.4393701372220464 \\ \hline 4 = 2^2 & 2.6524946 \cdot 10^{-2} & 3.629669596451481 \\ \hline 8 = 2^3 & 4.7442047 \cdot 10^{-3} & 5.35083146190712 \\ \hline 16 = 2^4 & 4.1306898 \cdot 10^{-4} & 7.7918959541372 \\ \hline 32 = 2^5 & 1.3310510 \cdot 10^{-5} & 11.226956600782769 \\ \hline 64 = 2^6 & 1.0697730 \cdot 10^{-7} & 16.05064913913718 \\ \hline 128 = 2^7 & 1.2206772 \cdot 10^{-10} & 22.826445101120278 \\ \hline 256 = 2^8 & 8.8802107 \cdot 10^{-15} & 32.35495110837777 \\ \hline 512 = 2^9 & 1.3241607 \cdot 10^{-20} & 45.77092301635973 \\ \hline 1024 = 2^{10} & 8.1182635 \cdot 10^{-29} & 64.6808514171555 \\ \hline 2048 = 2^{11} & 2.1076879 \cdot 10^{-40} & 91.3578121482943 \\ \hline 4096 = 2^{12} & 9.3011756 \cdot 10^{-57} & 129.01720949348262 \\ \hline 8192 = 2^{13} & 7.3886987 \cdot 10^{-80} & 182.2068558012616 \\ \hline 16384 = 2^{14} & 1.7003331 \cdot 10^{-112} & 257.358706225986 \\ \hline 32768 = 2^{15} & 1.2703122 \cdot 10^{-158} & 363.5691819464147 \\ \hline 65536 = 2^{16} & 7.9749999 \cdot 10^{-224} & 513.7027491850932 \\ \hline 131072 = 2^{17} & 5.2817088 \cdot 10^{-316} & 725.9526396990342 \\ \hline 262144 = 2^{18} & 2.4716651 \cdot 10^{-446} & 1026.0480594180656 \\ \hline 524288 = 2^{19} & 1.2878115 \cdot 10^{-630} & 1450.3756643021711 \\ \hline \end{array} \end{equation} From this tabular, it seems that $\ln I_n \sim - 2 \sqrt{n}$.

Is there any method to study such sequences of integrals?

$\endgroup$

2 Answers 2

5
$\begingroup$

The integral $$I_n=\int_0^\infty e^{f(n,t)}\,dt,\;\;f(n,t)=n\ln t-n\ln(1+t)-t,$$ has a saddle point at $t^*$ where $\partial f(n,t)/\partial t=0$, $$t^\ast=-\tfrac{1}{2}+\tfrac{1}{2}\sqrt{1+4n}.$$ For $n\rightarrow\infty$ the integral tends to $$I_n\rightarrow e^{f(n,t^\ast)} = e^{-2\sqrt{n}+{\cal O}(1)},$$ so $\ln I_n\rightarrow -2\sqrt{n}$, as found numerically.

$\endgroup$
1
$\begingroup$

Let us assume the following conjecture:

Let $h: \mathbb{R}^+ \longrightarrow \mathbb{R}$ be a continuous function. Let also $g_n: \mathbb{R}^+ \longrightarrow \mathbb{R}$ be a $\mathcal{C}^2$ function over $\mathbb{R}^+$, depending of an positive integer $n$, with a unique maximum $\mathcal{U}_n$ such that $g_n''(\mathcal{U}_n) < 0$. Therefore, we have the following estimation: $$ \int _0 ^{+ \infty} h(t) e^{- g_n(t)} \ dt \underset{n \longrightarrow + \infty}{\sim} h(\mathcal{U}_n) \cdot \sqrt { \dfrac{2 \pi}{- g_n''(\mathcal{U}_n)} } \cdot e^{g_n(\mathcal{U}_n)} \ . $$

This is a little adaptation to the classical Laplace method (see [Wikipedia page on Laplace method, section "Other formulation"][1]

In our case, as pointed out by Carlo Beenakker, we have $h(t) = 1$, $g_n(t) = t - n \ln t + n \ln(1 + t)$ and $\mathcal{U}_n = \dfrac{\sqrt{4n + 1} - 1}{2}$. So, we finaly find that $$ I_n \underset{n \longrightarrow + \infty}{\sim} \sqrt{e \pi} n^{\frac{1}{4}} e^{-2 \sqrt{n}} \ . $$ [1]: https://en.wikipedia.org/wiki/Laplace%27s_method

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.