2

Let $f: \mathbb{R} \to \mathbb{R}$ be linear spline (continuous, piecewise polynomial of degree $\le 1 $) at knots $x_0 < x_1 < … < x_n$. Prove $f$ can be uniquely represented at form:

$$\displaystyle f(x) = a+bx+ \sum_{j=0}^n c_j|x-x_j|$$ for some $a,b,c_j$.

I've tried to prove it by contradiction assuming there are two representation but I haven't figured out anything so far. I'll be grateful for some hints, since I've got stuck at this problem and don't have any idea

Jessy
  • 293
  • @CarlChristian why have you deleted your answer ? I haven't written anything yet, because I haven't figured out some details so far – Jessy May 14 '16 at 15:25
  • When it appeared that my hints would cause unnecessary complications, I decided to delete them. If I have time to reconsider, and I find a suitable alternative, then I will write it up. I am sorry for having wasted your time. – Carl Christian May 16 '16 at 07:56
  • This result has practical implications as well. Normally, we evaluate the linear spline by using table lookup. Given $x$, you determine by comparisons which interval $x$ belongs to. Then you load the appropriate coefficients and compute the relevant linear function. This is operation which is difficult to vectorize. In contract, the expression for $f$ given in this problem is one which vectorizes trivially. This is important if you application requires a large number of function evaluations, say, a simulation. – Carl Christian May 18 '16 at 09:39

1 Answers1

1

Given a set $S=\{x_0,\ldots,x_n\}\subset{\mathbb R}$ of $n+1$ knots denote by $V_S$ the vector space of linear splines $f\colon\>{\mathbb R}\to{\mathbb R}$ with knots in $S$. I claim that ${\rm dim}(V_S)=n+3$.

Proof. In order to fix an $f\in V_S$ we can freely choose $n+2$ slopes $m_i$ in the $n+2$ intervals $$(-\infty,x_0), \ (x_0,x_1), \ \ldots, \ (x_{n-1},x_n), \ (x_n,\infty)\ ,$$ which then will determine $f'$ wherever defined. In this way $f$ is defined up to an additive constant, which is finally determined by fixing $f(0)$.$\qquad\square$

On the other hand, the $n+3$ functions $$1,\quad x,\qquad\phi_j(x):=|x-x_j|\quad(0\leq i\leq n)\tag{1}$$ belong to $V_S$. I claim that they are linearly independent.

Proof. If $$a+b x+\sum_{j=0}^n c_j \phi_j(x)\equiv0$$ then the $c_j$ have to vanish individually, because each $\phi_j$ has a property that cannot be annihilated by a linear combination of the other involved functions: It has a corner at $x_j$. From $ax+b\equiv0$ it then follows that $a=b=0$.$\qquad\square$

It follows that the functions $(1)$ form a basis of $V_S$. This implies that any $f\in V_S$ has a unique representation as a linear combination of the functions $(1)$.

  • Thank you for this nice solution. Would you perhaps add a few words on the role of the corner property of $\phi_j$? Given your solution I thought to write $- c_0 \phi_0(x) = a + bx + \sum_{j=1}^n c_j \phi_j(x)$. Then the right hand side is differentiable at $x=x_0$, but the left hand side is not differentiable at $x_0$ unless $c_0 = 0$. – Carl Christian May 16 '16 at 18:34
  • @CarlChristian: You said it. – Christian Blatter May 16 '16 at 18:38
  • When I read the word "corner" I immediately knew what you meant, but it was the first time that I had seen it used in this context. We have different word in my native language (Danish): "knækpunkt", i.e. bend or break-point. We do not have a word for "cusp", so we use the English word as far as I know. I see "corner" used only sporadicly online, but I might be looking in the wrong places. So my question: Is "corner" a standard word in this context? – Carl Christian May 18 '16 at 09:31