I am interested in the behavior of the following sum \begin{align} \sum_{n=1} \frac{1}{n!} e^{an}\log(n), \end{align} as $a$ approaches infinity.

Using $\log(n) \le n$ we can the following upper bound \begin{align} \sum_{n=1} \frac{1}{n!} e^{an}\log(n) \le e^{a+e^a}. \end{align}

However, I think the true behavior is more like $e^{\log(a)+e^a}$. However, I am not sure how to show whether this is true or not.


We expect the largest summand approximately where $n\approx e^a$. Comparing the log to its tangent at that point, $$\ln n\le a-1+\frac{n}{e^a}$$ So $$\begin{align}\sum_{n=1}^\infty\frac1{n!}e^{an}\ln n&\le (a-1)\sum_{n=1}^\infty\frac1{n!}e^{an}+ \sum_{n=1}^\infty\frac1{(n-1)!}e^{a(n-1)}\\&=(a-1)(e^{e^a}-1)+e^{e^a}\\&=e^{\ln a+e^a}-a+1\end{align}$$

  • $\begingroup$ Do you think we can come up with a lower bound of the similar form? $\endgroup$ – Lisa Jun 15 '18 at 15:27

A full asymptotic series for $S$, with $x=\exp(a) \to \infty,$

$$ S:=\sum_{n=1}^\infty \frac{x^n}{n!} \log{n} \sim e^x\big(\, \log(x) - \frac{1}{2x}-\frac{5}{12x^2} + ... \big)$$

can be derived in the following manner.

It is a well-known Laplace transform that can be rearranged as such: $$ \frac{\log{n}}{n} =-\frac{\gamma}{n} - \int_0^\infty e^{-nt} \,\log{t}\, dt. $$ The $\gamma$ is Euler's constant. Insert this into definition for S, interchange order of sum and integration, and sum the series to get $$ S=-\gamma\big(e^x-1)-x\int_0^\infty e^{-t}\exp{(x\, e^{-t})} \log{t} \, dt. $$ This exact representation takes a simpler form on the substitution $e^{-t} \to u,$ $$ S=-\gamma\big(e^x-1)-x\int_0^1 \exp{(x\,u)} \log{(\log{(1/u)})} \, du. $$ Clearly for $x$ large what matters the most is what happens near $u=1.$ Expand the log-log as $$ \log{(\log{(1/u)})} = \log{(1-u)} - \frac{1}{2}(u-1) + \frac{5}{24}(u-1)^2+... $$ Now $$\int_0^1\log{(1-u)}\exp(x\,u)du=-\frac{e^x}{x}\big(\gamma+\Gamma(0,x)+\log(x)\big) \sim -\frac{e^x}{x}\big(\gamma+\log(x)\big)$$ where the incomplete gamma function is neglected because its asymptotic expansion is exponentially small in comparison. Likewise, neglecting terms that are exponentially small, we find that $$ \int_0^1(u-1)\exp(x\,u) du \sim -\frac{e^x}{x^2} \text{ and } \int_0^1(u-1)^2\exp(x\,u) du \sim \frac{2e^x}{x^3} .$$ Algebra completes the solution.

  • $\begingroup$ very nice (+1)! $\endgroup$ – tired Jun 16 '18 at 10:29

Let $N(t)$ be the Poisson process of the unit rate. Then

$$ \sum_{n=0}^{\infty} \frac{e^{an}}{n!}f(n) = e^{e^a}\mathbb{E}[f(N(e^a))]. $$

In our case, one can set $f(n) = \log(n \vee 1)$. In view of the law of large numbers, one may expect that this is approximately

$$e^{e^a}f(\mathbb{E}[N(e^a)]) = e^{e^a}f(e^a) = a e^{e^a}.$$

Indeed, by the Bennett-type concentration inequality, with $\psi(x) = (1+x)\log(1+x)-x $, for any $\epsilon > 0$ we have

$$ \mathbb{P}( \left| N(\lambda) - \lambda \right| \geq \epsilon\lambda) \leq 2e^{-\psi(\epsilon)\lambda}. $$

Then by using the inequality $\log(\lambda \pm a) \leq \log(\lambda+|a|)=\log\lambda + \int_{0}^{|a|/\lambda} \frac{ds}{1+s} $, we have

\begin{align*} \mathbb{E} \left[ \log(N(\lambda)\vee 1) \mathbf{1}_{\{|N(\lambda)-\lambda|\geq\epsilon\lambda\}} \right] &\leq \mathbb{E} \left[ \log \left( \lambda + \left|N(\lambda) - \lambda\right| \right) \mathbf{1}_{\{|N(\lambda)-\lambda|\geq\epsilon\lambda\}} \right] \\ &\leq (\log\lambda)\mathbb{P}( \epsilon\lambda \leq \left|N(\lambda) - \lambda\right|) \\ &\quad + \int_{0}^{\infty} \frac{ds}{1+s} \mathbb{P}( (s\vee\epsilon)\lambda \leq \left|N(\lambda) - \lambda\right|) \\ &\leq 2(\log \lambda)e^{-\psi(\epsilon)\lambda} + \int_{0}^{\infty} \frac{2e^{-\psi(s\vee \epsilon)\lambda}}{1+s}\,ds \end{align*}

This bound vanishes as $\lambda \to \infty$, so with $f(n) = \log(n\vee 1)$, it follows that

$$ \lim_{\lambda\to\infty} \frac{\mathbb{E} \left[ f(N(\lambda)) \right] }{\log\lambda} = 1 $$

This proves the desired asymptotics:

$$ \lim_{a\to\infty} \frac{\sum_{n=1}^{\infty} \frac{1}{n!}e^{an}\log n}{a e^{e^a}} =\lim_{a\to\infty} \frac{\mathbb{E}[f(N(e^a))]}{a} = 1. $$

  • $\begingroup$ I know this is an older question but can you please explain how you got from $ E \left[ \int _0^{ \frac{| N(\lambda)-\lambda|}{\lambda} } \frac{1}{1+s} ds \mathbf{1}_{ \{ |N(\lambda)-\lambda|<\epsilon \lambda \} } \right]$ to $ \int_0^\infty \frac{ds}{1+s} \mathbb{P}[ \max(s,\epsilon) \lambda \le | N(\lambda)-\lambda| ]$. I am assuming you did a change of variable. I just don't see how. Can you give a little more detains on this part. $\endgroup$ – Lisa Jul 15 '18 at 16:46

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.