4
$\begingroup$

Let's define finite differences of order $n$ in $x_0$ for a function $f$ as: $$\Delta ^1 f(x_0;h_1)= f(x_0+h_1)-f(x_0)$$ \begin{align*}\Delta ^2 f(x_0;h_1,h_2)&= \Delta f^1 (x_0+h_2;h_1)-\Delta f(x_0;h_1)\\ &=f(x_0+h_2+h_1)-f(x_0+h_2)-\left [ f(x_0+h_1)-f(x_0) \right ]\end{align*} $$\vdots$$ $$\Delta ^n f(x_0;h_1,\cdots,h_n)= \Delta f^{n-1} (x_0+h_n;h_1,\cdots,h_{n-1})-\Delta^{n-1} f(x_0;h_1,\cdots,h_{n-1})$$

Let's suppose now that $f\in\mathcal{C}^{n-1}([a,b])$ and $\exists f^{(n)}$ in $(a,b)$. I would like to show that if $x_0$, $x_0+h_1$, $x_0+h_2$, $x_0+h_1+h_2$, $\cdots$, $x_0+h_1+\cdots+h_n$ are always inside $[a,b]$), then there exists a point $\xi$ inside the smallest interval containing all of them such that $$\Delta^nf(x_0;h_1,\cdots,h_n)=f^{(n)}(\xi)h_1\cdots h_n$$ I tried by induction but I can't.

Are there other ways that doesn't use the concept of function of more than one variable? Thanks.

EDIT: By induction I succeeded in proving only that $$\Delta^nf(x_0;h_1,\cdots,h_n)\leq f^{(n)}(\xi)h_1\cdots h_n$$

$\endgroup$
  • $\begingroup$ Did you mean $C^{n}$ instead of $C^{n-1}$? $\endgroup$ – parsiad Mar 1 at 16:21
  • $\begingroup$ No, $n-1$. I just edited the answer to correct a mistake. $\endgroup$ – Nameless Mar 1 at 16:23
  • $\begingroup$ @parsiad how would you show the statement if the hypothesis were what you say? $\endgroup$ – Nameless Mar 1 at 20:14
1
+50
$\begingroup$

$\def\d{\mathrm{d}}\def\peq{\mathrel{\phantom{=}}{}}$Lemma: For any differentiable function $g$ and any $0 \leqslant m \leqslant n$,$$ \frac{\d}{\d x}(∆^m g(x; h_1, \cdots, h_m)) = ∆^m g'(x; h_1, \cdots, h_m). $$

Proof: It will be proved by induction. For $m = 0$, there is $\dfrac{\d}{\d x}(∆^0 g(x)) = \dfrac{\d}{\d x}(g(x)) = g'(x) = ∆^0 g'(x)$. Now assume that it holds for $m$, then by the induction hypothesis,\begin{align*} &\peq \frac{\d}{\d x}(∆^{m + 1}g(x; h_1, \cdots, h_{m + 1})) = \frac{\d}{\d x}(∆^m g(x + h_{m + 1}; h_1, \cdots, h_m) - ∆^m g(x; h_1, \cdots, h_m))\\ &= ∆^m g'(x + h_{m + 1}; h_1, \cdots, h_m) - ∆^m g'(x; h_1, \cdots, h_m) = ∆^{m + 1} g'(x; h_1, \cdots, h_{m + 1}). \end{align*} End of induction.

Now back to the problem. For $1 \leqslant k \leqslant n$, define$$ F_k(x) = ∆^k f^{(n - k)}(x; h_1, \cdots, h_k),\quad G_k(x) = ∆^{k - 1} f^{(n - k)}(x; h_1, \cdots, h_{k - 1}), $$ and also define $F_0(x) = f^{(n)}(x)$, then the lemma implies that $G_k' = F_{k - 1}$.

Next it will be proved by induction on $n \geqslant m \geqslant 0$ that $F_n(x_0) = F_m(ξ_m) \prod\limits_{k = m + 1}^n h_k$ for some $ξ_m$. For $m = n$, it suffices to take $ξ_n = x_0$. Assume that it holds for $m$ ($m \geqslant 1$). The mean value theorem implies that there exists $ξ_{m - 1}$ satisfying$$ F_m(ξ_m) = G_m(ξ_m + h_m) - G_m(ξ_m) = G_m'(ξ_{m - 1}) h_m = F_{m - 1}(ξ_{m - 1}) h_m, $$ thus $F_n(x_0) = F_m(ξ_m) \prod\limits_{k = m + 1}^n h_k = F_{m - 1}(ξ_{m - 1}) \prod\limits_{k = m}^n h_k$. End of induction.

Finally, taking $m = 0$ yields$$ ∆^n f(x_0; h_1, \cdots, h_n) = F_n(x_0) = F_0(ξ_0) \prod_{k = 1}^n h_k = f^{(n)}(ξ_0) \prod_{k = 1}^n h_k. $$

$\endgroup$
  • $\begingroup$ $$ F_m(ξ_m) = G_m(ξ_m + h_m) - G_m(ξ_m) = G_m'(ξ_{m - 1}) h_m = F_{m - 1}(ξ_{m - 1}) h_m, $$ Here, $G_m'$ is w.r.t. $x$ (the first point), but $ξ_m$ depends on $x$, so I have a composite derivative, no? $\endgroup$ – Nameless Mar 19 at 22:36
  • $\begingroup$ @Nameless No, $G_m'$ is with respect to its independent variable, be it $G_m'(x)$, $G_m'(y)$ or $G_m'(ξ)$. That's why it's not written as $\dfrac{\mathrm d}{\mathrm dx}(G_m(ξ_m+h_m))$ here. $\endgroup$ – Saad Mar 20 at 0:29
  • $\begingroup$ Okay, I see. Thanks. $\endgroup$ – Nameless Mar 20 at 21:14
2
$\begingroup$

I had an idea that was longer than comment. But not a complete solution. I felt it may be helpful to you so I’ve provided it below:

This is essentially applying the mean value theorem with induction.

Let’s define the “divided difference” of a function as

$$D_h(f) = \frac{f(x+h)-f(x)}{h} $$

If we consider the interval $[x_0, x_0+h]$ (and the order here might be wrong since h could be negative, so we should really say “minimum interval containing both”) and let $f$ be differentiable over this interval then the mean value theorem applies and we have that there exists a $\eta \in [x_0, x_0+h]$ such that

$$ f’(\eta) = D_h(f)(x_0) = \frac{f(x_0 + h)-f(x_0)}{h} $$ Or in your notation (and you’ll see why we are doing this)

$$ h f’(\eta) = f(x_0 + h)-f(x_0) = \Delta f(x_0 : h)$$

You can go further and say that there is some function $\eta(x)$ such that:

$$ h f’(\eta(x)) = \Delta f(x:h)$$

And we have that $\eta(x)$ is always contained in the smallest interval containing $x$,$x+h$

So now the to continue building our intuition let’s Say we have some g(x) = $ \Delta f(x: h)$. It is the case that over the minimum interval containing $x_0, x_0+h, x_0+h_2$ that there is some $c$ in the interval such that

$$ \Delta(g:h_2)(x_0) = h_2 g’(c)$$

Filling in the definition of g we have that:

$$ \Delta(f: h_1 : h_2)(x_0) = h_2 ( \Delta f(x: h))’(c) $$

$$ \Delta(f: h_1 : h_2)(x_0) = h_2 h_1 (f’(\eta(x)))’(c) $$

Which means:

$$ \Delta(f: h_1 : h_2)(x_0) = h_2 h_1 f’’(\eta(c)) \eta’(c)$$

Now the problem is that $\eta’(c)$ at the very end. There should be some way to absorb it by showing that $f’’(\eta(x)) \eta’(x) = f’’(m) $ for some suitable choice of m, at which point you can then use induction on the strategy we have demonstrated to cover the arbitrary case.

$\endgroup$
  • $\begingroup$ Thank you too @frogeyedpeas (+1). $\endgroup$ – Nameless Mar 20 at 21:16

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.