Well, I already bounded my answer, in the part that you snipped. I'll
repeat, in hopes that it sticks this time:
_If_ the system has _just one_ nonlinearity _at the input_ and _is
otherwise substantially linear_ then this method will work well.
Does it make sense now?
If you have a (or some) nonlinearity right on the edge of the plant, i.e.
right at the input or right at the output, and the nonlinearity is
memoryless, and the nonlinearity provides a reasonable one-one mapping
between the command and the internal force exerted (on the input side) or
the plant state(s) and what you read back (on the output side) then
negating the nonlinearity is trivial.
In other words, if your system equation can be expressed as
d
-- x = A x + b(u(t)), y(t) = c(x(t), u(t))
dt
and if b and c are easy functions to "undo", then you can just undo them
and proceed with a linear system design.
On the other hand, if you cannot separate out your system function that
way, then until you know more, all bets are off.