Using cosmological simulations, I explored in what way the properties of individual galaxies vary when their initial conditions are perturbed ever so slightly, at a machine precision level. I found that a perturbation of even a single particle at redshift z=5 in a cosmological simulation of a (25Mpc/h)^3 volume results in galaxies that have quite different properties compared to an unperturbed simulation. The magnitude of the differences depends on time, simulation resolution, physical model, and the considered galaxy property. Generally speaking, the differences at the final time at redshift of z=0 are of order 2-20%.
This happens to be a scale that is relevant for the scatter in galaxy scaling relations. Namely, the differences between shadow galaxies is not too far from the overall scatter in certain important scaling relations, such as the Tully-Fisher relation between rotation velocity and stellar mass.
If - and this is a big if - this effect exists in the real Universe, then it may imply a fundamental limit to our ability to develop a comprehensive predictive theory for galaxy formation. But even if this effect is purely numerical, extra care must be taken when comparing simulations to each other or to observations.