consider something like a polynomial interpolation:

Interpolating polynomial (or most ML models in general) are smooth, and so interpolating between points will result in “overshooting” regional points and “bouncing around”

…as a function of parameters

At a fixed dataset, just increasing the number of parameters will increase