4

Fix a value $x\in(0,1)$ and recursively define the sequences

$$a_0=0,~b_0=1$$

$$c_{n+1}=\frac{\frac{a_n+b_n}2+x}2=\frac{a_n+2x+b_n}4$$

$$a_{n+1}=\begin{cases}a_n,&c_{n+1}>x\\c_{n+1},&c_{n+1}<x\end{cases}$$

$$b_{n+1}=\begin{cases}c_{n+1},&c_{n+1}>x\\b_n,&c_{n+1}<x\end{cases}$$

Supposing that we never have $c_n=x$, what is the rate of convergence of $a_n$ and $b_n$ as they approach $x$, defined by

$$\mu(x)=\lim_{n\to\infty}(b_n-a_n)^{1/n}$$

as a function of $x$? Specifically I am interested in the infinum, supremum, and geometric mean (see integrals below) of $\mu$ over $(0,1)$.

Attempting to plot this in MATLAB for $n=30$ and $1000$ equally spaced points, I get

enter image description here

with a geometric mean of roughly $0.3613$, which matches up reasonably for $n\in[25,30]$ as well as my integral results (see below). The supremum and infinum for these were found to be approximately

$$0.25\le\mu(x)\le0.3862$$

which suggests the upper bound may be given by $\frac{-1+\sqrt{17}}8$ from the fixed-point below.

Trivially we have the easily proven bounds of $1/4\le\mu\le1/2$, and I can compute this at some points by studying fixed-points. For example, we have

$$\mu\left(\frac{-3+\sqrt{17}}4\right)=\frac{-1+\sqrt{17}}8\simeq0.39039$$

$$\mu\left(\frac{-7+\sqrt{89}}{20}\right)=\sqrt{\frac{-5+\sqrt{89}}{32}}\simeq0.37224$$

which were found by solving $a_n=0$ and $1-x=x/b_n$ for $n=1$ and $n=2$ respectively. By taking the limit of these fixed-points, the general term of which can be found explicitly, we get the infinum by

$$\inf_{x\in(0,1)}\mu(x)=\frac14$$

However, I'm not really sure if this $\mu$ is continuous (as it likely exhibits a fractal-like nature), in which case computing $\mu(x)$ is not very useful to me (see context at the end). Hence, I'm also interested in the "average" rate by instead considering $f_n(x)=b_n-a_n$ as a function of $x$ and taking the geometric average as

$$\lim_{n\to\infty}\exp\left(\frac1n\int_0^1\ln(f_n(x))~\mathrm dx\right)$$

which would be more meaningful to me as a sort of "expected rate of convergence".

For $n\in[1,4]$, we have

\begin{align} \exp\left(\int_0^1\ln(f_1(x))~\mathrm dx\right)&=e^{-1}\simeq0.36788\\ \exp\left(\frac12\int_0^1\ln(f_2(x))~\mathrm dx\right)&=e^{-1/2}2^{9/5}3^{-8/5}\simeq0.36418\\ \exp\left(\frac13\int_0^1\ln(f_3(x))~\mathrm dx\right)&\simeq0.36353\\ \exp\left(\frac14\int_0^1\ln(f_4(x))~\mathrm dx\right)&\simeq0.36306\end{align}

which seems to converge somewhat reasonably to $0.36179$ (by Aitken delta squared acceleration), but this is a fairly tedious computation.

For context, this problem arose while studying a bisection-like algorithm I made by averaging the arithmetic mean with a faster converging approximation of the root, such as the secant's root. That is to say, the $x$ used in $c_{n+1}$ is only an approximation of $x$ that becomes more precise on every iteration.

  • So it looks like you've already partially answered the question; can you please provide a concise summary of what is left to be shown? – Zim Oct 22 '20 at 16:30
  • @Zim I haven't answered the question, there's mostly just heuristics for what the answer appears to be. All of the questions are contained within the block quotes. – Simply Beautiful Art Oct 22 '20 at 17:35
  • It looks like you've already found the infimum? – Zim Oct 22 '20 at 18:14
  • @Zim Sorry haven't looked at this in a while; yes the infimum was found, but aside from this pretty much nothing. – Simply Beautiful Art Oct 22 '20 at 20:00
  • I think $\mu(x)$ could be rewritten as $$\lim_{n \to \infty} \frac{b_{n+1}-a_{n+1}}{b_n-a_n}$$. I got this by taking logs, applying Stolz-Cesaro, and then exponentiating, but I might have made a mistake with Stolz-Cesaro. – Varun Vejalla Oct 22 '20 at 20:57
  • @VarunVejalla I think the troublesome issue with that is that the consecutive terms exhibit fairly chaotic behaviors. The ratio of errors for one iteration to the next does not smoothly predict the limit. – Simply Beautiful Art Oct 22 '20 at 21:28
  • Yeah I just compared the two in Python and my formulation is much worse at converging. I'm not sure that it even does converge, but if it does, it may be easier to work with from an analytic standpoint. – Varun Vejalla Oct 23 '20 at 00:32
  • It might also be more useful for the geometric average. $\exp\left(\int_0^1 \ln\left( \frac{f_{n+1}(x)}{f_n(x)} \right), \mathrm dx\right)$ seems to converge better than $\exp\left(\frac1n\int_0^1\ln(f_n(x)), \mathrm dx\right)$ – Varun Vejalla Oct 23 '20 at 00:54
  • @VarunVejalla Interesting observation, I had not considered that. But I also did not evaluate for very large $n$ to see how well it faired. – Simply Beautiful Art Oct 23 '20 at 01:11

0 Answers0