So, the smoother the function, the easier it is to approximate using SGD? Does that mean there's a relationship between training time and parameters and beyond

So, the smoother the function, the easier it is to approximate using SGD? Does that mean there's a relationship between training time and parameters and beyond a certain parameter count we can discover an equivalent approximation by trading one for the other?