5
$\begingroup$

I proved the identity: $$ \lim_{x\rightarrow\infty} x\int_0^1 t^{x t} dt = 1 $$ but my proof is very long and complicated. I'm wondering if there is a simpler way to do it.

Putting a lower bound of 1 is pretty straightforward, because $t^{xt}\ge t^x$ for $t\in[0,1]$: $$ \lim_{x\rightarrow\infty} x\int_0^1 t^{x t} dt \ge \lim_{x\rightarrow\infty} x\int_0^1 t^{x} dt = \lim_{x\rightarrow\infty} \frac{x}{x+1} = 1 $$ I did manage to prove an upper bound of 1 by taking Riemann sums after substituting $y = tx$ but I am wondering if there is a more elegant way to do it.

$\endgroup$
2
  • 1
    $\begingroup$ I guess the first step is to substitute $y = tx$. $\endgroup$
    – PhoemueX
    Jun 15 '18 at 23:00
  • $\begingroup$ Yes that was my first step. $\endgroup$ Jun 16 '18 at 23:16
3
$\begingroup$

With integration by parts one can prove for $x \to \infty$ the asymptotic formula below. The condition for $h(t)$ is that it is increasing on (0,1) and $v$ is assumed to be >0 and bounded. (One may also insert into the integrand a continuous function with certain growth conditions, but I'll leave that generalization aside.)

$$ \int_0^1 \exp{(x \,h(t) )} (1-t)^{v-1} \, dt \sim \frac{\Gamma(v)}{x^v} \frac{\exp{(x h(1) )}}{(h'(1))^v} $$

An earlier answer of mine used this but I was sloppy and used it even though the integrand is not increasing over all of (0,1). Fortunately there is a way to salvage the argument.

enter image description here

The picture illustrates that for large $x$ only the regions around the endpoints of $t=0$ and $t=1$ are important. Split the integrand at $t=1/e,$ the minimum of $t \, \log{t}.$ Furthermore write $$ \exp(x t\,\log{t} ) = \exp{(-x/e)} \exp{\big(x(t\log{t} + 1/e})\big)= \exp{(-x/e)} \exp{\big(x\, h_1(t)\big)}$$ This is done because $h_1(t) \ge 0$ for t in (0,1). The split means $$ I(x)=\int_0^1 \exp{x(\,t\,\log{t})} dt =\exp{(-x/e)}\Big( \underbrace{\int_0^{1/e} \exp{\big(x\, h_1(t)\big)} \,dt}_{I_1} + \underbrace{\int_{1/e}^{1} \exp{\big(x\, h_1(t)\big)} \,dt }_{I_2} \Big)$$ $I_2$ is easy to deal with from the formula on the top of the page. The fact that the lower limit does not extend to 0 can be handled with this argument: We could augment $h_1$ with a function that has a value <0 over (0,1/e) and the criterion of non-decreasing would still be satisfied, and the integral over (0,1/e) would be exponentially smaller than the part we are interested in. Using the formula we find $h(1)=1/e,$ $h'(1) = 1$ and thus $$ \exp{(-x/e)}\, I_2 = \exp{(-x/e)}\, \cdot \frac{ \exp{(x/e)}}{x} \sim \frac{1}{x} $$ Now we want to show $\exp{(-x/e)}I_1$ is exponentially smaller. Rescale the integral to find $$ I_1 = \frac{1}{e}\, \int_0^1 \exp{(x/e \,t\log{(t/e)}) }\, dt= \frac{1}{e}\, \int_0^1 \exp{\big(x/e \,t(\log{t} - 1)\big) }\,dt$$ $$= \frac{1}{e}\, \int_0^1 \exp{\big(x/e \,(1-t)(\log{(1-t)} - 1)\big) }\,dt = $$ $$=\frac{ \exp{(-x/e)}}{e} \int_0^1 \exp{\big(x/e\big(\underbrace{1+ \,(1-t)(\log{(1-t)} - 1)}_{h_2(t)}\big)\big) }\,dt$$ From the first line to the second the substitution $t \to 1-t$ was performed. In the last line of the previous calculation $h_2$ has been defined so that it is $\ge 0 $ on (0,1). If we try to use the formula at the top of page then we find $h_2'(1) = \infty,$ and that is not allowed so the formula is not directly applicable. However, we only want a bound so define $$ h_3(t) = \frac{t^2}{2}(1+t) \ge h_2(t) , \text{ so } h_3'(1) = 5/2, h_3(1)=1.$$ We can thus use the formula and find $$ I_1 < \frac{ \exp{(-x/e)}}{e} \cdot \frac{ \exp{(x/e)}}{5/2 x} = \frac{2}{5 e x} $$ Thus indeed $\exp{(-x/e)} I_1$ is exponentially smaller.

$\endgroup$
1
  • $\begingroup$ Am I right that $$I_1 =\frac{1}{e}\, \int_0^1 \exp{(x/e \,t\log{(t/e)}) }\, dt$$ originates from the full $t\log t$ exponent and not from $I_1$ defined via $h_1$ above? Since then you are multiplying with $e^{-x/e}$ one times too much and your actual estimate $2/(5ex)$ is not exponentially smaller. The integral from $0$ to $1/e$ indeed vanishes, but this is precisely because $(t\log t)'=1+\log t$ goes to $-\infty$ for $t\rightarrow 0$. Any finite slope at $0$ would yield a finite result, which happens to be the case for your function $h_3$ about $t=1$ (after your inversion $t\rightarrow(1-t)$. $\endgroup$
    – Diger
    Oct 24 '18 at 16:48
1
$\begingroup$

I would first substitute $t={\rm e}^{-u}$ to obtain $$ x\int_{0}^{\infty }{{\rm e}^{-u \left( x{{\rm e}^{-u}}+1 \right) }} \,{\rm d}u $$ and split the integral $$ x\int_{0}^{\frac{1}{\sqrt{x}}}{{\rm e}^{-u \left( x{{\rm e}^{-u}}+1 \right) }} \,{\rm d}u + x\int_{\frac{1}{\sqrt{x}}}^\infty{{\rm e}^{-u \left( x{{\rm e}^{-u}}+1 \right) }} \,{\rm d}u \, . $$

For $x\rightarrow \infty$ you can replace ${\rm e}^{-u}$ by $1$ in the first term and carry out the integral which yields $1+{\cal O}\left(\frac{1}{x}\right)$. The second integral can be shown to be ${\cal O}\left(\frac{1}{x}\right)$.

For a more rigorous proof you can sandwich the first term with ${\rm e}^{-u} \geq 1-u$ and ${\rm e}^{-u}\leq 1-\sqrt{x}\left(1-{\rm e}^{-\frac{1}{\sqrt{x}}}\right) u$ and use the series expansion of the arising error function.


Another argument goes as follows: Again split the integral

$$ \underbrace{x\int_0^1 e^{xt\log t} \, {\rm d}t}_{I} = \underbrace{x\int_0^{1/2} e^{xt\log t} \, {\rm d}t}_{I_1} + \underbrace{x\int_{1/2}^1 e^{xt\log t} \, {\rm d}t}_{I_2} \, . $$ Then for any $0<t_1<t_2\leq\frac{1}{2}$ $$ x\int_{t_1}^{t_2} e^{xt \log t} \, {\rm d}t \leq x\int_{t_1}^{t_2} e^{xt \log t_2} \, {\rm d}t = \frac{e^{xt_2 \log t_2}-e^{xt_1 \log t_2}}{\log t_2} $$ which vanishes for $x\rightarrow \infty$. Thus for any $\delta > 0$ we can choose $\epsilon > 0$ such that $$ x\int_0^\epsilon e^{xt \log t} \, {\rm d}t \leq x\int_0^\epsilon e^{xt \log \epsilon} \, {\rm d}t = \frac{e^{x\epsilon \log \epsilon} - 1}{\log \epsilon} \leq -\frac{1}{\log \epsilon} < \delta $$ for $x\epsilon >> 1$ and so $$\lim_{x \rightarrow \infty} I_1 = 0 \, .$$

On the other hand since $t\leq e^{t-1}$ for all $1/2 \leq t \leq 1$ we have $$ I_2 \leq x\int_{1/2}^1 e^{(t-1)xt} \, {\rm d}t = -\frac{i}{2} \, \sqrt{\pi x} \, e^{-x/4} \, {\rm erf}\left(\frac{i}{2}\,\sqrt{x}\right) = 1 + \frac{2}{x} + {\cal O}\left(x^{-2}\right) $$ and therefore $$ \lim_{x \rightarrow \infty} I \leq 1 \, . $$

$\endgroup$

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.