Note: This was originally published on Medium on Mar 31, 2021. I have since taken a numerics lecture and come to know that this is called “piecewise linear interpolation”. I have also learned that solving such a linear system of equations might give awful precision unless you represent your numbers as rationals.
Problem
Given a list of \(n\) Points \(\bold P = (P_1, P_2, …, P_n)\) where \(P_i = (P_{i,x}, P_{i,y})\) ordered along the \(x\)-axis, what function \(f(x)\), when graphed, connects the points in the domain \([P_{1,x}; P_{n,x}]\) with straight lines?
Solution
\[f(x) = \sum_{i = 0}^n a_i \lvert x - P_{i,x} \rvert\]where
\[\begin{bmatrix} a_1 \\ \vdots \\ a_n \\ \end{bmatrix} = \begin{bmatrix} P_{1,y} \\ \vdots \\ P_{n,y} \\ \end{bmatrix} \times \begin{bmatrix} \lvert P_{1,x} - P_{1,x} \rvert & \cdots & \lvert P_{1,x} - P_{n,x} \rvert \\ \vdots & \ddots & \vdots \\ \lvert P_{n,x} - P_{1,x} \rvert & \cdots & \lvert P_{n,x} - P_{n,x} \rvert \\ \end{bmatrix}^{-1} \ .\]Note: I came up with this on my own so there are probably better ways to do this.
Explanation
Consider the function \(g(x) = a \lvert x - b \rvert\): it is like a linear function, \(a\) determines the steepness, but it’s mirrored at \(b\) along the \(y\) axis.
If you sum multiple of these functions with different \(a\) and \(b\) values together, you get a zigzagged line, something that could work as the function we are searching for.
Let’s assume we can use the \(x\)-values of the given points as values for \(b\). Then we get:
\[f(x) = \sum_{i = 0}^n a_i \lvert x - P_{i,x} \rvert\]which is the same as
\[f(x) = a_1 \lvert x - P_{1,x} \rvert + \cdots + a_n \lvert x - P_{n,x} \rvert \ .\]Now we need to find the values for \(a = (a_1, a_2, \ldots, a_n)\).
We know the values of \(f(x)\) at the given points:
\[\forall p \in P \colon f(p_x) = p_y\]which in simpler terms just means
\[\begin{cases} f(P_{1,x}) = P_{1,y} \\ \quad \quad \quad \vdots \\ f(P_{n,x}) = P_{n,y} \\ \end{cases} \ .\]From this we can construct a linear system of equations:
\[\begin{cases} a_1 \lvert P_{1,x} - P_{1,x} \rvert + \cdots + a_n \lvert P_{1,x} - P_{n,x} \rvert = P_{1,y} \\ \quad \quad \quad \vdots \quad \quad \quad \ \ \ \ddots \quad \quad \quad \quad \ \vdots \quad \quad \quad \quad \ \vdots \\ a_1 \lvert P_{n,x} - P_{1,x} \rvert + \cdots + a_n \lvert P_{n,x} - P_{n,x} \rvert = P_{n,y} \\ \end{cases} \ .\]All values for \(P\) are known, so at this point, we could simply use an equation solver to get \(a\). However, it can also be rewritten as a matrix multiplication:
\[\begin{bmatrix} \lvert P_{1,x} - P_{1,x} \rvert & \cdots & \lvert P_{1,x} - P_{n,x} \rvert \\ \vdots & \ddots & \vdots \\ \lvert P_{n,x} - P_{1,x} \rvert & \cdots & \lvert P_{n,x} - P_{n,x} \rvert \\ \end{bmatrix} \times \begin{bmatrix} a_1 \\ \vdots \\ a_n \\ \end{bmatrix} = \begin{bmatrix} P_{1,y} \\ \vdots \\ P_{n,y} \\ \end{bmatrix} \ .\]Solve for \(a\):
\[\begin{bmatrix} a_1 \\ \vdots \\ a_n \\ \end{bmatrix} = \begin{bmatrix} P_{1,y} \\ \vdots \\ P_{n,y} \\ \end{bmatrix} \times \begin{bmatrix} \lvert P_{1,x} - P_{1,x} \rvert & \cdots & \lvert P_{1,x} - P_{n,x} \rvert \\ \vdots & \ddots & \vdots \\ \lvert P_{n,x} - P_{1,x} \rvert & \cdots & \lvert P_{n,x} - P_{n,x} \rvert \\ \end{bmatrix}^{-1} \ .\]And that’s it.
Example
Let’s find \(f(x)\) for the points \((1, 1)\), \((2, 3)\) and \((4, 2)\). First, we need to find the values for \(a\).
Using the above formula, we get:
\[\begin{bmatrix} a_1 \\ a_2 \\ a_3 \\ \end{bmatrix} = \begin{bmatrix} 1 \\ 3 \\ 2 \\ \end{bmatrix} \times \begin{bmatrix} \lvert 1 - 1 \rvert & \lvert 1 - 2 \rvert & \lvert 1 - 4 \rvert \\ \lvert 2 - 1 \rvert & \lvert 2 - 2 \rvert & \lvert 2 - 4 \rvert \\ \lvert 4 - 1 \rvert & \lvert 4 - 2 \rvert & \lvert 4 - 4 \rvert \\ \end{bmatrix}^{-1} \ .\]Simplify:
\[\begin{bmatrix} a_1 \\ a_2 \\ a_3 \\ \end{bmatrix} = \begin{bmatrix} 1 \\ 3 \\ 2 \\ \end{bmatrix} \times \begin{bmatrix} 0 & 1 & 3 \\ 1 & 0 & 2 \\ 3 & 2 & 0 \\ \end{bmatrix}^{-1} \ .\]Solve using SageMath:
V = matrix(QQ, [1, 3, 2])
M = matrix(QQ, [[0, 1, 3], [1, 0, 2], [3, 2, 0]])
a = V * pow(M, -1)
a # [ 3/2 -5/4 3/4]
Now we can use the very first formula to get \(f(x)\):
\[f(x) = \frac 3 2 \lvert x - 1 \rvert - \frac 5 4 \lvert x - 2 \rvert + \frac 3 4 \lvert x - 4 \rvert \ .\]
