6

Fix five equally spaced nodes as $x_i = x_0 + ih$ where $h > 0$, $x_0\in\mathbb{R}$, and $i = 0, 1, 2, 3, 4$. Let us also denote $f_i := f(x_i)$.

Exercise. Assume that $f\in \operatorname{C^5}[x_0, x_4].$ Show that there exists some $\xi(x_2)=:\xi\in[x_0, x_4]$ such that $$f'(x_2) = \dfrac{f_0 - 8f_1 + 8f_3 - f_4}{12 h} + \dfrac{h^4}{30}f^\mathrm{V}(\xi).\label{E}\tag{E}$$

Solution. Using a method of undetermined coefficients and approximation by Taylor polynomials with Lagrangian remainders, I believe to have shown that

$$f'(x) = \dfrac{f_0 - 8f_1 + 8f_3 - f_4}{12 h} + \frac{h^4}{30} \frac{16\, f^\mathrm{V}(\xi_2) - 4\, f^\mathrm{V}(\xi_1)}{12} \tag{1}$$

where $\xi_1, \xi_2 \in[x_0, x_4]$, and $x:=x_2.$


Here is the more detailed explanation. (Skip forward to section named Question if you wish). First for $k = 1, 2$ using Taylor polynomials and Lagrangian remainders

$$f(x\pm kh) = f(x) \pm f'(x)\, kh + f''(x)\, \frac{k^2 h^2}{2} \pm f'''(x)\, \frac{k^3 h^3}{6} + f^\mathrm{IV}(x)\,\frac{k^4 h^4}{24} \pm f^\mathrm{V}(\xi_\pm^k)\,\frac{k^5 h^5}{120} \label{A1}\tag{A1}$$

where $\xi_\pm^k$ is between $x$ and $x \pm k h$. Note also that $x_1 = x - h$, $x_3 = x + h$ and so on.

Let us view the expression $Af_0 + Bf_1 + Cf_3 + Df_4$ where $A, B, C, D$ are to be determined. After substituting $f_0, f_1, f_3, f_4$ from the earlier Taylor expansion $\eqref{A1}$ into this expression, one gets after further dividing both sides by $h$ that

\begin{align*}\frac{Af_0 + Bf_1 + Cf_3 + Df_4}{h} = \, (&A + B + C + D)\,\frac{f(x)}{h} + (-2A - B + C + 2D)\, f'(x)\\ +&(4A + B + C + 4 D)\,f''(x)\, \frac{h}{2} \\ +&\, (-8A -B + C + 8D)\,f'''(x)\, \frac{h^2}{6} + (16A + B + C + 16D)\, f^\mathrm{IV}(x)\,\frac{ h^3}{24}\\ +& \left[-32A\, f^\mathrm{V}(\xi_-^2) - B\, f^\mathrm{V}(\xi_-^1) + C\, f^\mathrm{V}(\xi_+^1) + 32D\, f^\mathrm{V}(\xi_+^2)\right]\,\frac{h^4}{120}. \label{A2}\tag{A2}\end{align*}

Next we attempt to determine the coefficients $A, B, C, D$ in such a way that we are left with $f'(x)$ and $h^4$ terms on the RHS of $\eqref{A2}$. This gives us the system

$$ \begin{cases} A + B + C + D = 0,\\ -2A - B + C + 2D = 1, \\ 4A + B + C + 4 D = 0, \\ -8A -B + C + 8D = 0,\\ 16A + B + C + 16D = 0. \end{cases}\label{A3}\tag{A3} $$

The unique solution is $A = - D = \dfrac{1}{12}$, $-B = C = \dfrac{2}{3}.$ If we denote the $h^4$ term by $-\mathcal R(x)$, then substituting the values of the coefficients back into $\eqref{A2}$, we get

$$\dfrac{f_0 - 8f_1 + 8f_3 - f_4}{12 h} + \mathcal R(x) = f'(x).\label{A4}\tag{A4}$$

Comparing this to $\eqref{E}$, what remains to be shown is that the expression

$$\mathcal R(x) = \left[32A\, f^\mathrm{V}(\xi_-^2) + B\, f^\mathrm{V}(\xi_-^1) - C\, f^\mathrm{V}(\xi_+^1) - 32D\, f^\mathrm{V}(\xi_+^2)\right]\,\frac{h^4}{120}\label{A5}\tag{A5}$$

or, after substituting the solution coefficients and simplifying, that the expression

$$\mathcal R(x) = \frac{h^4}{30} \frac{8\, f^\mathrm{V}(\xi_-^2) - 2\, f^\mathrm{V}(\xi_-^1) - 2\, f^\mathrm{V}(\xi_+^1) + 8\, f^\mathrm{V}(\xi_+^2)}{12}\label{A6}\tag{A6}$$

is somehow equal to

$$\dfrac{h^4}{30}f^\mathrm{V}(\xi)\label{A7}\tag{A7}$$

for some $\xi\in[x_0, x_4]$. Because $f^\mathrm{V}$ is continuous, by the intermediate value theorem we get

\begin{align*} f^\mathrm{V}(\xi_-^1) + f^\mathrm{V}(\xi_+^1) = 2 f^\mathrm{V}(\xi_1),\label{A8}\tag{A8}\\ f^\mathrm{V}(\xi_-^2) + f^\mathrm{V}(\xi_+^2) = 2 f^\mathrm{V}(\xi_2),\label{A9}\tag{A9} \end{align*}

where $\xi_1 \in(x - h, x + h)$ and $\xi_2 \in(x - 2h, x + 2h)$. Therefore,

$$\mathcal R(x) = \frac{h^4}{30} \frac{16\, f^\mathrm{V}(\xi_2) - 4\, f^\mathrm{V}(\xi_1)}{12}.\label{A10}\tag{A10}$$


Question.

  • If I could show that for some $\xi\in [x_0, x_4]$ $$16f^\mathrm{V}(\xi_1) - 4f^\mathrm{V}(\xi_2) = 12f^\mathrm{V}(\xi),\label{Q}\tag{Q}$$ the proof would be complete. Is this achievable?

If it isn't always possible to do, there is probably a mistake somewhere...

StubbornAtom
  • 17,932
  • A simple counterexample shows that the number $\xi$ does not always exist. Let $f^{(5)}(x) = x^2$, $\xi_1=0$, and $\xi_2=h$. Then $4f^{(5)}(\xi_1) - f^{(5)}(\xi_2)=-h^2$. Since $3f^{(5)}(\xi) = 3\xi^2$ can never equal a negative number, $\xi$ does not exist. – J. Heller Jun 19 '20 at 17:19

1 Answers1

3

Apply the extended mean value theorem four times \begin{align} \frac{a(h)}{b(h)}&=\frac{f(x-2h)-8f(x-h)-12hf'(x)+8f(x+h)-f(x+2h)}{h^5} \\ =\frac{a'(h_1)}{b'(h_1)}&=\frac{-2f'(x-2h_1)+8f'(x-h_1)-12f'(x)+8f'(x+h_1)-2f'(x+2h_1)}{5h_1^4} \\ =\frac{a''(h_2)}{b''(h_2)}&=\frac{f''(x-2h_2)-2f''(x-h_2)+2f''(x+h_2)-f''(x+2h_2)}{5h_2^3} \\ =\frac{a'''(h_3)}{b'''(h_3)}&=\frac{-2f'''(x-2h_3)+2f'''(x-h_3)+2f'''(x+h_3)-2f'''(x+2h_3)}{15h_3^2} \\ =\frac{a^{(4)}(h_4)}{b^{(4)}(h_4)}&=\frac{2f^{(4)}(x-2h_4)-f^{(4)}(x-h_4)+f^{(4)}(x+h_4)-2f^{(4)}(x+2h_4)}{15h_4} \end{align} with $h>h_1>h_2>h_3>h_4>0$, $b(h)=h^5$ and $a(0)=a'(0)=a''(0)=a'''(0)=a^{(4)}(0)=0$. Now apportion the last term as \begin{align} &=\frac{2(f^{(4)}(x-2h_4)-f^{(4)}(x-h_4))+(f^{(4)}(x-h_4)-f^{(4)}(x+h_4))+2(f^{(4)}(x+h_4)-f^{(4)}(x+2h_4))}{15h_4} \\ &=\frac2{15}(-f^{(5)}(\xi_1)-f^{(5)}(\xi_2)-f^{(5)}(\xi_3)) \end{align} where $x-2h_4<\xi_1<x-h_4$, $x-h_4<\xi_2<x+h_4$, $x+h_4<\xi_3<x+2h_4$ by the simple mean value theorem. By the intermediate value theorem there is some $\xi\in(x-2h_4,x+2h_4)$ with $3f^{(5)}(\xi)=f^{(5)}(\xi_1)+f^{(5)}(\xi_2)+f^{(5)}(\xi_3)$ so that in total \begin{align} &\frac{f(x-2h)-8f(x-h)-12hf'(x)+8f(x+h)-f(x+2h)}{h^5} \\ &=-\frac2{5}f^{(5)}(\xi), \end{align} the claim follows.

Lutz Lehmann
  • 131,652
  • I have worked through your answer, and it appears correct. Clever! Thank you, (+1). If you have the time and patience, it would be very helpful to me if you could find the mistake(s) in my own derivation in the OP; that way I would avoid such missteps in the future. (Feel free to post an additional, second answer for this purpose should you choose to do so). – Linear Christmas Jun 18 '20 at 10:50
  • 1
    There is nothing wrong in your approach, the direct Taylor approach just leads to subtractions in the remainder terms that can not be collected using the intermediate value theorem. Using the integral form of the remainder term might also lead to a solution, shifting the differences to the weights and thus to showing that the weight difference is still positive. – Lutz Lehmann Jun 18 '20 at 11:17
  • @LutzLehmann: Could you explain how you are using the extended mean value theorem? Thanks – J. Heller Jun 19 '20 at 19:18
  • 1
    @J.Heller : I use $\frac{a(h)}{b(h)}=\frac{a(h)-a(0)}{b(h)-b(0)}=\frac{a'(h_1)}{b'(h_1)}$, $0<h_1<h$, repeatedly, where $a(0)=b(0)=0$ in each instance. – Lutz Lehmann Jun 22 '20 at 08:50