Since linear systems are virtually always stable

I disagree. See the backwards heat equation for a simple example:

du/dt = -laplacian(u)

The higher the frequencies you allow in the solution, the faster tiny bits of noise blow up on you.