Lately I have been investigating numerical methods to approximate fractional composition roots of functions. My intuition tells me that power series expansions will not be very suitable in approximating fractional functions in the sense
$$(f^{\circ 1/k})^{\circ k}(t) = f(t)$$
For example
$g(t) = \sin^{\circ 1/16}(t)$ would be any function satsifying $g^{\circ 16}(t) = \sin(t)$. In other words $\underset{16 \hspace{0.166cm}\text{times}}{\underbrace{g(g(\cdots g(g(t}}))\cdots)) =\sin(t) $
Any well behaved function which is such a solution $t\to g(t)$ will be extremely smooth and rather close to $f(t) = t$ so that a normal power series expansion will be unsuitable. The monomials are simply diverging way too fast.
I have been hypothesizing that families with linear combinations of $f(t) = t$ and in the exponential family with negative exponents could be meaningful. Or maybe logarithmic functions.
But I can't think of a theoretical approach to find reasonable families of functions.
I am curious of where to find sources for how to approach this.
Answers are of course welcome as well.