Consider the transition rate of the evolution of the state $|i\rangle$ to a state $|f\rangle$ - the Fermi golden rule: $$ d \omega_{if} = 2\pi |\mathcal{M}_{fi}|^{2} \delta(E_{f} - E_{i})d\nu, \quad \mathcal{M}_{fi} = \mathcal{V}_{fi}+\sum_{n\neq i}\frac{\mathcal{V}_{fn}\mathcal{V}_{ni}}{E_{i}-E_{n}}+\dots $$ Here $V$ is the perturbation operator, with $V_{fi}(t) = e^{i\omega_{fi}t}\mathcal{V}_{fi}$, $|i,f\rangle$ are initial and final states being the eigenstates of the free Hamiltonian $H_{0}$ and $|n\rangle$ is the intermediate state.
The energy is conserved in the transition. However, it is not conserved for the intermediate states $|n\rangle$. The total Hamiltonian $H = H_{0}+V$ itself is a generator of time translations corresponding to the conservation of the energy. What is a formal reason for this? Has this anything to do with the approach of the perturbation theory (within which we divide the total Hamiltonian into two parts)?