My question is based on question by Robert Israel. There is only one answer by Lucia and also two useful comments (by Lucia too), from which we understand, that if $$\frac{1}{e^x+q}=\sum\limits_{n=0}^{\infty}F_{n}\frac{x^n}{n!}$$ so signs of $F_{n}$ be the same as signs of $$\cos\left((n+1)\arctan\left(\frac{\pi}{\log q}\right)\right)$$ for $q=3$.
Is it obvious from Lucia's answer (not comments)? If not, why is that true? Is it true for any $q$?
To determine the asymptotic behaviour of the MacLaurin coefficients $F_n$ of a function $f(z)$, the answerer implicitly uses an approach due to Darboux (see Olver, Asymptotic and Special Functions (1997) 4.9.2, p. 310). Suppose that $r$ is the distance from the origin of the nearest singularity of $f(z)$, if we can find a ``comparison function'' $g(z)$ with the properties
$g(z)$ is isomorphic in $0<\left|z\right|<r$
$f(z)-g(z)$ is continuous in $0<\left|z\right|\le r$
The coefficients $b_n$ in the Laurent expansion \begin{equation} g(z)=\sum_{-\infty}^\infty b_nz^n \end{equation} have known asymptotic behaviour,
then, for $n\to\infty$, \begin{equation} F_n=b_n+o\left( r^{-n} \right) \end{equation} The function \begin{equation} f(z)=\frac{1}{e^{z}+q} \end{equation} is meromorphic and has poles at \begin{equation} z_n= \ln q+\left( 2n+1 \right)i\pi \end{equation} where $n$ is an integer. The nearest poles are $z_1=\ln q+i\pi $ and $z_{-1}=\bar{z_1}$. Corresponding residues are both $-1/q$. We choose \begin{equation} g(z)=-\frac{1}{q}\left( \frac{1}{z-z_1}+\frac{1}{z-z_{-1}} \right) \end{equation} it satisfies the required conditions with \begin{equation} g(z)=\frac{1}{q}\sum_{n=0}^\infty \left( z_1^{-n-1}+z_{-1}^{-n-1}\right)z^n \end{equation} and thus, by denoting $z_1=\rho e^{i\varphi}$, \begin{equation} b_n=\frac{2}{q}\rho^{-n-1}\cos\left( n+1 \right)\varphi \end{equation} then \begin{equation} F_n\sim\frac{2}{q}\rho^{-n-1}\cos\left( n+1 \right)\varphi \end{equation} Their signs are thus approximatively given by that of $\cos\left(\left( n+1 \right)\arctan\frac{\pi}{\ln q} \right)$
Finally, it is remarked that, when $q=3$, we have $\varphi\simeq \frac{11\pi}{28}$ with a good accuracy. This explains the near 28-periodicity of the coefficients.