1
$\begingroup$

Show that on the polynomial interpolation of a function $f \in C^3$ in 3 points $x_0 < x_1 < x_2$ with $x_2-x_1 = x_1 - x_0$, the error of interpolation verifies $|E(x)| \le 0.65(x_1-x_0)^3M, \forall x$ between $x_0$ and $x_2$, where $M$ is the absolute value of the maximum of the 3rd derivative of $f$ on the interval $[x_0,x_2]$.

I did

$$\begin{align} |E(x)| &\le 0.65(h)^3M \\ |E(p_2(x))| &\le \max_{x_2,x_0}|(x-x_0)(x-x_1)(x-x_2)|M/3! = \\ \end{align}$$ $$\begin{align} |(x_0+h/2-x_2)(x_0+h/2-x_1)(x_0+h/2-x_2)|M/6&\le 0.65(h)^3M = \\ |(h/2)(-h/2)(3h/2)|M/6 &\le 0.65h^3M \Leftrightarrow \\ 3h^3/8* M/6 &\le 0.65h^3M \end{align}$$

...q.e.d?

How I got to x being equal to $x_0+h/2$ where h = $x_1-x_0$: I drew some 3rd degree polynomials and "kind of" noticed they were symmetrical and that $x_0$, $x_1$ and $x_2$ were always the zeroes. I also noticed that $f(x_1)$ was always equal to zero and that it stood halfway between the other two, and between each, there was a maximum or minimum. So I chose $x = x_0+h/2$.

Is this correct?

$\endgroup$
4
  • $\begingroup$ Although it is true that the max of the $|(x-x_0)(x-x_1)(x-x_2)|$ occurs in some point between $x_0$ and $x_1$, it is not ok to assume that it occurs at the midpoint $x_0+\frac h2$. $\endgroup$ – PierreCarre Feb 23 at 15:23
  • $\begingroup$ @PierreCarre why not? $\endgroup$ – Segmentation fault Feb 23 at 19:12
  • 1
    $\begingroup$ Because you can compute exactly where will the maximum occur and it is not $x_0+\frac h2$. $\endgroup$ – PierreCarre Feb 23 at 19:42
  • 2
    $\begingroup$ If you write $x=x_1+sh$, $s\in[-1,1]$, you get $(x-x_0)(x-x_1)(x-x_2)=(s^3-s)h^3$. The cubic polynomial has derivative $3s^2-1$, which quite easily gives the positions of the extrema. $\endgroup$ – Lutz Lehmann Feb 23 at 20:08
2
$\begingroup$

So, you wrote the error formula $$ |E_2(x)| \leq |(x-x_0)(x-x_2)(x-x_2)| \frac{M}{3!}. $$

Now you can simply maximize the 3rd degree polynomial... The maximum will be attained at a stationary point, that you can compute explicitly. The stationary points are $x_0 + (1 \pm \frac{\sqrt{3}}{3})h$, and the corresponding maximum value will be $\frac{2h^3}{3 \sqrt{3}}$. In general, your choice of $x = x_0 +\frac h2$ is wrong and does not yield the maximum.

Using this information,

$$ |E_2(x)|\leq \frac{M h^3}{9 \sqrt{3}}\approx 0.06415 (x_1-x_0)^3 M. $$

Finally, if this estimate holds, so does the (much worse) proposed estimate.


Regarding the maximization of $|(x-x_0)(x-x_1)(x-x_2)|$ in the interval $[x_0, x_2]$, as noted by Lutz Lehmann, it is equivalent to the maximization of $g(s)=|(s^3-s)h^3|$ for $s \in [-1,1]$. Since $g$ is zero on the boundary and at the branching point for the absolute value, the maximum will occur at a point where $g'(s)=0$, i.e. $s= \pm \frac{\sqrt{3}}{3}$. Substituting back in the function, we conclude that the maximum is $g(\pm \frac{\sqrt{3}}{3}) = \frac{2h^3}{3\sqrt{3}}$.

$\endgroup$
4
  • $\begingroup$ How do you know the maximum is at $x_0 + (1 \pm \frac{\sqrt{3}}{3})h$? $\endgroup$ – Segmentation fault Feb 23 at 19:28
  • $\begingroup$ @Segmentationfault the maximum will occur at a point where the derivative of the polynomial is zero, and you can compute the zeros of the derivative (these are the zeros of a polynomial with degree 2). $\endgroup$ – PierreCarre Feb 23 at 19:43
  • $\begingroup$ I'm trying to calculate the maxima. I got that the derivative is $(3s^2-1)h^2$ where $s \in [-1,1]$. It's zero when $s = \pm 1/\sqrt(3)$. Replacing back the maxima is $(\pm 1/3 \mp 1/\sqrt(3))$. I don't know what to do next. $\endgroup$ – Segmentation fault Feb 23 at 23:29
  • $\begingroup$ @Segmentationfault I'll add to my answer. $\endgroup$ – PierreCarre Feb 24 at 9:44

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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