0

From Wikipedia: enter image description here

My question is: How I can get the formula?

From the definition, the only thing I obtained was that $q \approx \frac{\log |\frac{x_{k+1}- L}{x_k-L}|}{\log |\frac{x_{k}-L}{x_{k-1}-L}|}$. How I can continue from here?

Edit:

A more concise explanation of what I've done:

So. Let $\{x_n\}_{n\geq1} \subset \mathbb{R}$ an approximation sequence for $x^*\in\mathbb{R}, x_n \xrightarrow{n \to \infty} x^*$. If $\exists c > 0 $ such that $\lim\limits_{n \to \infty} \frac{|x^* - x_n|}{|x^* - x_{n-1}|^r}= c \,\,(1) \implies$ order of convergence is $r$ and rate of convergence is $c$.

Let $e_n := x_n - x^*$. Because $x_n \xrightarrow{n\to\infty} x^* \implies e_n \xrightarrow{n\to\infty} 0 \implies (1)$ is equivalent to $\exists c > 0$ such that $\lim\limits_{n \to \infty} \frac{|e_n|}{|e_{n-1}|^r}= c$.

When $n \to \infty$ we have $|e_n| \approx c |e_{n-1}|^r$ and $|e_{n-1}| \approx c |e_{n-2}|^r$, so when $n \to \infty$ we have $|\frac{e_n}{e_{n-1}}| \approx |\frac{e_{n-1}}{e_{n-2}}|^r \implies r \approx \frac{\log{|\frac{e_n}{e_{n-1}}}|}{\log{|\frac{e_{n-1}}{e_{n-2}}}|} \implies r \approx \frac{\log{|\frac{x^* - x_n}{x^*-x_{n-1}}}|}{\log{|\frac{x*-x_{n-1}}{x^*-x_{n-2}}}|}$

But, when $n \to \infty$ we have that $|x^* - x_n| \approx |x_{n+1}-x_{n}|$ (because from $x_n \xrightarrow{n \to \infty} x^*$ we have that $\forall \epsilon > 0 \exists n_\epsilon \in \mathbb{N}$ such that $\forall n \in \mathbb{N}, n \geq n_\epsilon : |x_n - x^*| < \epsilon \implies \forall \epsilon > 0 $ we have $||x^* - x_n| - |x_{n+1} - x_n|| \leq |x^* - x_n - x_{n+1} + x_n| = |x^* - x_{n+1}| < \epsilon$).

This implies $r \approx \frac{\log{|\frac{x_{n+1} - x_n}{x_n-x_{n-1}}}|}{\log{|\frac{x_n-x_{n-1}}{x_{n-1}-x_{n-2}}}|}$ and $c \approx \frac{|x_{n+1} - x_n|}{|x_n - x{n-1}|^r}$.

Is something good here or it's messed up totally?

  • 2
    Now consider the fact that usually $L$ is unknown and that the convergence is, presumably, so fast that $x_{k+1}-L$ is small against $x_k-x_{k+1}$. So if you know $L$, then use the formula with $L$, it will give sharper results. This happens when testing the method on example problems. Otherwise if controlling the order is part of an empirical convergence test or so, use the formula without $L$. – Lutz Lehmann Oct 09 '23 at 07:57
  • @LutzLehmann but if the convergence is not so fast? I don't know what to say about this argument. There is no general formula? I've seen this article but it uses something with fixed points. There exists any general case ? https://www.math-cs.gordon.edu/courses/ma342/handouts/rate.pdf – MathLearner Oct 09 '23 at 14:38
  • And.. there exists a formula like this for computing the rate of convergence $q$ ? – MathLearner Oct 09 '23 at 15:27
  • 1
    You mean the rate $r$ as in $x_k\approx L + c\cdot r^{q^k}$? In principle, these are 4 constants, so using 4 iterates should give a system that can be solved for all 4 of them. – Lutz Lehmann Oct 09 '23 at 16:35
  • @LutzLehmann sorry! My bad! I wanted to say rate of convergence $\mu$ from above photo. Thanks! – MathLearner Oct 09 '23 at 17:36
  • 1
    That would be $c$ in my formula. Fitting for the parameters would be more precise, with more iterates. But one has to consider the limits of the number type. With the usual double or float64 the Newton method usually has 3-5 iterates in the quadratic mode before no further advances are made. Use multi-precision – Lutz Lehmann Oct 09 '23 at 18:04
  • @LutzLehmann thanks for all help. Can you check please if what I've done in edit is good or not? – MathLearner Oct 09 '23 at 19:03
  • 1
    Yes, that looks good. If you try it out. I'd expect that the results of the "empirical" formula have a substantial error when rounding to the next integer, but I can't produce ready examples. The experiments for the less dramatic convergence dynamic of quadrature and ODE integration methods shows that the order is rather stable and close to its correct value, while the rate coefficient shows a greater variation. – Lutz Lehmann Oct 09 '23 at 19:18

0 Answers0