Spivak, Ch. 22, Problem 14: Understanding some calculations involving Newton's Method

78 Views Asked by At

Problem 14 of Chapter 22, "Infinite Sequences", of Spivak's Calculus is a bit long. It has five items, and my question is about implicit assumptions in item $(d)$.

Let me show item $(d)$ straight away to cut to the chase, and then show $(a)-(c)$.

14.(d) Let $m=f'(c)=\inf(f')$ on $[c,x_1]$ and let $M=\sup(f'')$ on $[c,x_1]$. Show that Newton's method works if $$x_1-c<\frac{m}{M}\tag{1}$$.

Here is the beginning of the solution from the solution manual. My questions are about hidden/implicit assumptions.

$$\delta_1=x_1-c<\frac{m}{M}\tag{2}$$

Then

$$\frac{M}{m}\delta_1=\alpha<1\tag{3}$$

for some $\alpha$.

"For some $\alpha$" sounds weird. $\alpha$ is simply being defined as the expression $\frac{M}{m}\delta_1$, which is $<1$.

How do we know that $\frac{m}{M}>0$?

By assumption, we are on the interval $[c,x_1]$ and $f'(c)$ is the smallest slope on this interval. Hence $f''(c)$ must be $\geq 0$.

If $f''(c)=0$ then the expression $(1)$ doesn't make sense, so for the purposes of the current problem we needn't consider it (note however, that it isn't a problem in general for $f''(c)=0$).

But what if $f'(c)<0$ and $f''(c)=\sup{(f'')}>0$? Then $\frac{m}{M}<0$.

But then it seems we can't get $(3)$.

Is this line of reasoning correct?

The rest of the proof relies critically on $(3)$. Thus, it seems that the result we are trying to prove isn't a general case. It is the case in which $f'(c)>0$.

Now let me take a step back and show what happens in the previous items (usually items build results that are useful to solve later items).

Items $(a)$ and $(b)$ serve to introduce Newton's Method

In (a), we are asked to show that the tangent line to the graph of $f$ at $(x_1,f(x_1))$ intersects the horizontal axis at $(x_2,0)$, where

$$x_2=x_1-\frac{f(x_1)}{f'(x_1)}$$

If we keep repeating this process of obtaining the tangent and figuring out where it intersects the $x$-axis, we get a sequence $\{x_n\}$ defined by

$$x_{n+1}=x_n-\frac{f(x_n)}{f'(x_n)}$$

We hope that this sequence converges to a root $c$ of $f$.

In (b), we are asked to show that if $f',f''>0$, and we choose $x_1$ with $f(x_1)>0$ then $x_1>x_2>x_3>...>c$.

This is a special case of a convex increasing function. If we start the sequence at a value where $f$ is positive, then the sequence is decreasing. Now, it seems to me that $c$ need not exist (e.g. $e^x$). In any case, it is a simple matter to show that

$$x_1>x_2>x_3>...>x_n>...$$

Here is item (c) in full

Let $\delta_k=x_k-c$. Then

$$\delta_k=\frac{f(x_k)}{f'(\xi_k)}$$

for some $\xi_k\in (c,x_k)$. Show that

$$\delta_{k+1}=\frac{f(x_k)}{f'(\xi_k)}-\frac{f(x_k)}{f'(x_k)}$$

Conclude that

$$\delta_{k+1}=\frac{f(x_k)}{f'(\xi_k)f'(x_k)}\cdot f''(\eta_k)(x_k-\xi_k)$$

for some $\eta_k\in (c,x_k)$

and then that $$\delta_{k+1}\leq \frac{f''(\eta_k)}{f'(x_k)}\delta_k^2$$

I went through the calculations to obtain these results, though the latter aren't at this point particularly enlightening.

1

There are 1 best solutions below

0
On

I've been going through the text and I ran into similar questions. I think a cleaner set of assumptions for part b) (that follow through to the later parts) is

Suppose $f''>0$, $f(x_1)>0$, $f'(x_1)>0$ and let $c$ be a root of $f$.

With these assumptions part b and c work out. In part d, if $x_1-c < m/M$, then it must be that $m=f'(c) >0$, since otherwise $x_1 \leq c$. This can't occur because $f$ is increasing on $[x_1,\infty)$.