So what happens is each segment between two knots (in the paper) is made of 2 pieces of parametric quadratic polynomials (i.e. parabolas).
Each knot is at the vertex of the same parabola to each side, which can be broken at the vertex into two parts, each of which belongs to one of the adjacent segments.
The solver in the paper makes sure that at the joint between two parabolas, the absolute magnitude of curvature matches.
This means there aren’t really any smooth inflection points and you always get a discrete jump in curvature at an inflection. This looks bad, so in practice Adobe implemented something different in their software.