Skip to content

Lecture 7

Order of Convergence and Modified Newton

Key results

  • Order \(p\): \(e_{n+1} \approx C\,e_n^{\,p}\); on a log-log plot of \(e_{n+1}\) against \(e_n\) the slope is \(p\).
  • One-point methods: \(g'(r) = \cdots = g^{(p-1)}(r) = 0\), \(g^{(p)}(r) \ne 0\) \(\Rightarrow\) order exactly \(p\).
  • At a root of multiplicity \(m\), Newton has \(g'(r) = 1 - \frac1m\) (linear); modified Newton \(x_{n+1} = x_n - m\frac{f(x_n)}{f'(x_n)}\) restores order \(2\).

Order of convergence

A sequence \(x_n \to r\) converges with order \(p \ge 1\) if

\[ |r - x_{n+1}| \le C\,|r - x_n|^p , \]

with \(C < 1\) required when \(p = 1\) (then \(C\) is the rate). With order \(1\) each step adds a fixed number of correct digits; with order \(p\) the number of correct digits is multiplied by about \(p\).

Taking logarithms, \(\log e_{n+1} \approx p\log e_n + \log C\): plotted against each other, the errors lie on a line of slope \(p\). From three consecutive errors,

\[ p \approx \frac{\log(e_{n+1}/e_n)}{\log(e_n/e_{n-1})} . \]

For Newton on \(x^3 + 2x^2 - x - 1\) from \(x_0 = 1\) this gives \(1.87,\ 1.99,\ 2.00\); for the fixed-point iteration \((2x+5)^{1/3}\) it gives \(1.00\).

Method Order
Bisection, Regula Falsi, fixed point (\(g'(r) \ne 0\)) \(1\)
Illinois \(\approx 1.44\)
Secant \(\frac{1+\sqrt5}{2} \approx 1.618\)
Newton (simple root) \(2\)

Order of one-point methods

If \(x_{n+1} = g(x_n)\), \(g(r) = r\) and \(g'(r) = g''(r) = \cdots = g^{(p-1)}(r) = 0\), then the order is at least \(p\), and

\[ \lim_{n\to\infty}\frac{x_{n+1} - r}{(x_n - r)^p} = \frac{g^{(p)}(r)}{p!} . \]

Example. For \(\sqrt a\), \(g(x) = \frac{x}{2}\left(1 + \frac{a}{x^2}\right)\): \(g'(x) = \frac12 - \frac{a}{2x^2}\) so \(g'(\sqrt a) = 0\), and \(g''(x) = \frac{a}{x^3}\) so \(g''(\sqrt a) = \frac{1}{\sqrt a} \ne 0\). The order is exactly \(2\), with \(\epsilon_{n+1}/\epsilon_n^2 \to \frac{1}{2\sqrt a}\) (about \(0.354\) for \(a = 2\)).

Multiple roots

\(r\) has multiplicity \(m\) if \(f(x) = (x - r)^m q(x)\) with \(q(r) \ne 0\). If \(f(r) = f'(r) = \cdots = f^{(m-1)}(r) = 0\) and \(f^{(m)}(r) \ne 0\), then \(r\) has multiplicity \(m\). For \(f(x) = e^{x^2} - 1\): \(f(0) = 0\), \(f'(0) = 0\), \(f''(0) = 2\), so \(0\) is a double root.

For Newton's \(g(x) = x - f(x)/f'(x)\) one finds \(g'(r) = 1 - \frac1m\): at a multiple root Newton is only linear, with rate \(1 - \frac1m\) (\(\frac12\) for a double root, \(\frac23\) for a triple root). Conversely, an observed ratio \(\rho\) suggests \(m \approx \frac{1}{1 - \rho}\).

Modified Newton

\[ x_{n+1} = x_n - m\,\frac{f(x_n)}{f'(x_n)} \]

has \(g'(r) = 0\), so quadratic convergence is restored. For \(f(x) = e^{x^2} - 1\), \(m = 2\), \(x_0 = 0.5\):

\(n\) Newton \(x_n\) Modified Newton \(x_n\)
\(0\) \(0.50000\) \(5.0000\times10^{-1}\)
\(1\) \(0.27880\) \(5.7602\times10^{-2}\)
\(2\) \(0.14468\) \(9.5454\times10^{-5}\)
\(3\) \(0.07309\) \(4.3486\times10^{-13}\)
\(4\) \(0.03664\)
\(5\) \(0.01833\)
\(6\) \(0.00917\)

Here the convergence is even cubic: \(g(x) = x - \frac{1 - e^{-x^2}}{x} = \frac{x^3}{2} + \cdots\), so \(g'(0) = g''(0) = 0\).

Unknown \(m\). Apply Newton's method to \(u(x) = \dfrac{f(x)}{f'(x)}\), which has a simple root at \(r\) for any \(m\). This converges quadratically, but \(u' = 1 - \dfrac{f f''}{(f')^2}\) needs \(f''\).

Cancellation

For small \(x\), \(e^{x^2}\) is very close to \(1\), and computing \(e^{x^2} - 1\) directly subtracts nearly equal numbers, losing digits. The naive formula gives \(x_3 \approx 1.1\times10^{-12}\) instead of \(4.35\times10^{-13}\). A dedicated routine for \(e^y - 1\) (usually called expm1) avoids this. Lecture 9 explains why.