\( \newcommand{\matr}[1] {\mathbf{#1}} \newcommand{\vertbar} {\rule[-1ex]{0.5pt}{2.5ex}} \newcommand{\horzbar} {\rule[.5ex]{2.5ex}{0.5pt}} \)
deepdream of
          a sidewalk
Show Question
\( \newcommand{\cat}[1] {\mathrm{#1}} \newcommand{\catobj}[1] {\operatorname{Obj}(\mathrm{#1})} \newcommand{\cathom}[1] {\operatorname{Hom}_{\cat{#1}}} \newcommand{\multiBetaReduction}[0] {\twoheadrightarrow_{\beta}} \newcommand{\betaReduction}[0] {\rightarrow_{\beta}} \newcommand{\betaEq}[0] {=_{\beta}} \newcommand{\string}[1] {\texttt{"}\mathtt{#1}\texttt{"}} \newcommand{\symbolq}[1] {\texttt{`}\mathtt{#1}\texttt{'}} \)
Math and science::Analysis::Tao::10: Differentiation of functions

Newton's approximation

Informally, if \( f \) is differentiable at \( x_0 \), then one has the approximation \( f(x) \approx f(x_0) + f'(x_0)(x - x_0) \), and conversely.

The formal version:

Newton's approximation

Let \( X \) be a subset of \( \mathbb{R} \), let \( x_0 \in X \) be a limit point of \( X \), let \( f : X \to \mathbb{R} \) be a function, and let \( L \) be a real number. Then the following statements are logically equivalent:

  1. \( f \) is differentiable at \( x_0 \) with derivative \( f' \).
  2. For any \( \varepsilon > 0 \), there exists a \( \delta > 0 \) such that for all \( x \in \{i \in X : |i - x_0| < \delta \} \), \( f(x) \) is \( \varepsilon |x - x_0| \)-close to \( f(x_0) + f'(x_0)(x-x_0) \).

    In other words, we have:
    \( |f(x) - (f(x_0) + f'(x_0)(x - x_0))| \le \varepsilon|x - x_0| \)

2. can be prased as: fix \( x_0 \). Given a challenge of \( \varepsilon \), it's possible to find a neighbourhood around \( x_0 \) for which \( x_0 \) along with \(f(x_0) \) and \(f'(x_0) \) can be used to approximate \( f \) for all of the neighbourhood, with an approximation error at most \( \varepsilon \).


This seems like such a confusing way to formulate the idea. 

An attempt at the formal version in words as a challenge: for any \( x_0 \) at which \( f \) is differentiable you may choose any \( \varepsilon > 0 \) and I can can give you a \( \delta > 0 \) such that any point that is within \( \delta \) from \( x_0 \), an estimate of \( f(x_0) + f'(x_0)(x - x_0) \) will be at most \( \delta \varepsilon \) off. And in fact, I can do slightly better, the error will be within \( |x - x_0| \varepsilon \). 


Source

p255