The simplest argument is thus:
When the system is linear with respect to switching frequency and harmonics, those harmonics don't affect the cycle-averaged / DC response.
That is, it obeys superposition.
Such a system can then be decomposed into the baseband / DC / cycle-averaged component, plus Fsw and harmonics.
The high-frequency content can further be ignored entirely, to focus on the baseband response.
Total linearity is not required, merely linearity with respect to the switching. For a typical square-pulse converter, this usually presumes CCM, so that the switch node always has a low impedance, always in the high or low state.
For more general use, some modifications are necessary, such as for DCM, accounting for the effective output resistance of the switches; but as long as the switching node behaves as expected for the approximation used (e.g. a square pulse approximation, with node returning to idle when inductor current falls to zero), this works.
The LC network between/after switches is generally a linear network, so can be cycle-averaged over. When it's also controlling power (resonant types), this gets more complicated, but can still be done.