I am taking a class on numerical analysis. While the professor was deriving the theory behind quadratic splines, the professor mentioned that a quadratic spline function has the form: $$ p_{i}(x)=a_{i}x^{2}+b_{i}x+c_{i}\quad\forall i=0,1,n $$ For a dataset with $n+1$ points, this means that we have a total of $3n$ unknowns $a_{i}, b_{i}$ and $c_{i}$. Additionally, we are to enforce constraints on the continuity of the interval nodes between every two consecutive points the same thing also applies for the first derivative so we have $(n-1)+(n-1)=2(n-1)$ constraints. In addition to the interpolation points $p_{i}(x)=y_{i}$ for all $i=0,1,n$ we have $n+1$ conditions to be added to $2(n-1)$ meaning we have a total of $2(n-1)+n+1=3n-1$ constraints.

This means we need to add an additional constraint in order to align the number of unknowns with number of constraints. One approach to do this is by setting $P'(x_{0})=0$ where $P(x)$ is the continuous curve that connects all $p_{i}(x)$ computed. Note that this type of constraint induces a natural quadratic spline.

My professor also mentioned other constraint methods without giving any details:

  1. End Slope Spline
  2. Periodic Spline
  3. Not-a-Knot Spline

My question is can anyone provide me with some insight on each of the three methods. I would be extremely thankful as it will help me expand my knowledge further beyond the scope of this course.


1 Answer 1

  1. End slope for quadratic splines is a generalization of the natural quadratic spline condition: $p'_0(x_0) = v_0$. Alternatively, you could specify instead $p'_n(x_{n+1}) = v_{n+1}$. Since $\mathcal{C}_1$ quadratic splines only have 1 unspecified degree of freedom, you generally can't simultaneously require both of these to be true. For cubic splines, this condition is often called a "clamped" spline as you specify both of these simultaneously. I'm not entirely sure what the generalization beyond cubic splines would be.

  2. Periodic quadratic splines add the constraint $p'_0(x_0) = p'_n(x_{n+1})$. This constraint makes the most sense when $y_0 = y_{n+1}$, since then when you try to connect the spline's end points together you get an overall $\mathcal{C}_1$ spline. For higher order splines where you have more unspecified degrees of freedom you can require the higher derivatives are also equal across the periodic point.

  3. I don't think I've heard of the not-a-knot condition applied to quadratic splines before, however for cubic splines this enforces $p'''_0(x_1) = p'''_1(x_1)$ and $p'''_{n-1}(x_n) = p'''_n(x_n)$ ($C_2$ cubic splines have 2 unspecified degrees of freedom, so you need two extra equations). This results in $p_0(x) = p_1(x)$ and $p_{n-1}(x) = p_{n}(x)$. I suppose the closest you can come to this with a quadratic spline is to enforce $p''_0(x_1) = p''_1(x_1)$ which will result in $p_0(x) = p_1(x)$, or alternatively you can try to enforce $p''_{n-1}(x_n) = p''_{n}(x_n)$, which results in $p_{n-1}(x) = p_{n}(x)$. You generally cannot require both to be true if you want to maintain a $C_1$ quadratic spline. I'm also not entirely sure what the generalization of this for higher order splines would be.


for a natural cubic spline you specify $p''_0(x_0) = 0$ and $p''_n(x_{n+1}) = 0$.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.