I'm going through some notes on how to apply the Hamiltonian formalism to systems with gauge invariance and I found a derivation of Noether's theorem I had never seen before. The idea is roughly that if you have a symmetry transformation that changes the action only by a boundary term:
$$ \delta_{\epsilon} I = \int dt \; \epsilon \frac{dF(t)}{dt}$$
where $\epsilon$ is the infinitesimal parameter of the transformation (constant), then, by promoting $\epsilon$ to a function of the parameter in the worldline $\epsilon (t)$ (satisfying that it vanishes on the boundary of the integral), you should find:
$$ \delta_{\epsilon} I = \int dt \left[ \epsilon(t) \frac{dF(t)}{dt} + \dot{\epsilon}(t) Q(t) \right] $$
This is reasonable since for $\epsilon (t)$ constant you must recover the previous result, and any terms containing higher derivatives of $\epsilon$ can be integrated by parts. Integrating by parts the second term, using that $\epsilon (t)$ is arbitrary and that the variation of the action is zero on shell, you obtain that $F(t)-Q(t)$ is constant along the worldline.
So far so good. My problems start right after the derivation, where a plain sentence asserts that once we have the conserved charge we can recover the symmetry transformation using:
$$ \delta_{\epsilon}f = \epsilon \, \{ f, Q \}_{PB} $$
for any function $f$ on phase space. After a couple of hours struggling with this equation, I can't find why this should be a symmetry transformation. Any ideas? I would be specially interested in arguments not assuming a particular form of the action, since I am dealing with systems with constraints where the action has strange terms included to ensure that these constraints are satisfied. As an example, my action for a relativistic point particle is:
$$ I = \int dt \, \left[ \dot{x}^m p_m - \frac{1}{2}e(p^2 + m^2) \right] $$
with $e(t)$ a Lagrange multiplier.