Solving a limit sum involving Cantor sets

120 Views Asked by At

Let $P:C\to \mathbb{R}$, where $C$ is the Cantor set and $P$ is continuous. If $C_n$ is the sequence of numerators from the values of endpoints from the defined intervals of iteration $k$

$0, 1, 2, 3, 6, 7, 8, 9, 18, 19, 20, 21, 24, 25, 26, 27, 54, 55, 56, 57, 60, 61, 62, 63, 72, 73, 74, 75, 78, 79, 80, 81, 162, 163, 164, 165, 168, 169, 170, 171, 180, 181, 182, 183, 186, 187, 188, 189, 216, 217, 218, 219, 222, 223, 224, 225, 234, 235, 236, 237, 240, 241, 242, 243,...$

Where $C_1=0$, find the exact value of

$$\lim_{k\to\infty}\frac{\sum\limits_{n=1}^{2^{k+1}} P\left({C_n}/{3^k}\right)}{2^{k+1}}$$

  1. For $P=x^2$ (According to Mathematica, the closed-form is $3/8$).

  2. For $P={1}/{\left[2(x+2)\right]}$. $\left(\text{According to mathematica, it seems the closed form is} \left(\frac{1}{e}\right)^{-1-2e}\cot(e\pi)\cos(e\pi)\ \right)$

  3. For $P=2^x$ (Doesn't appear to have a closed-form in fact most functions don't give a closed form.)

  4. For $P$ in general (In terms of Integrals, Measure, etc.). This should solve my newly defined average for functions defined on the Cantor set.

1

There are 1 best solutions below

0
On

Tl;dr: there is a recursive formula for monomials based on their degree given in a 1992 paper.

To start, I gathered some numerical evidence. For $0\le m,n \le 6$ I asked Mathematica to evaluate $\int_{C_m} x^n H^s(dx)$. For each fixed $n$, the partial sums had the formulas $$ \left\{1,\frac{1}{2},\frac{3^{-2 m}}{8}+\frac{3}{8},\frac{1}{16} 3^{1-2 m}+\frac{5}{16},\frac{9^{1-m}}{32}-\frac{17\ 9^{-2 m}}{320}+\frac{87}{320},-\frac{17\ 9^{-2 m}}{128}+\frac{25\ 9^{-m}}{64}+\frac{31}{128},-\frac{17}{512} 9^{1-2 m}+\frac{29\ 9^{1-m}}{512}+\frac{3253\ 9^{-3 m}}{46592}+\frac{10215}{46592}\right\}$$

Taking $\lim_{m\to \infty}$ you just keep the constant terms:

$$\left\{1,\frac{1}{2},\frac{3}{8},\frac{5}{16},\frac{87}{320},\frac{31}{128},\frac{10215}{46592} \right\}$$

I didn't recognize this sequence and it was somewhat intractable (RSolve, FindSequenceFunction, oeis were stumped). So I googled the sequence; amazingly I got a hit from a 1992 paper. Unfortunately, the paper was in Mandarin, of which I am wholly illiterate. So I asked a good friend to translate and she graciously agreed. Several months later, she emailed me a working draft. I'll give the main points from that paper and will link my edited version and the original (she asked that her version not be distributed).

  • One establishes that integrating with the Hausdorff measure is well-defined and agrees with the Riemann sum definition for continuous $f$.

  • Using a recursion formula for the right endpoints of each subinterval, one obtains the functional relation $$ \int _F f(x) H^s(dx ) = \int _F f(x) H^s(dx ) = \frac{1}{2}\left(\int _F f\left(\frac{x}{3}\right) H^s(dx )+\int _F f\left(\frac{x}{3}+\frac{2}{3}\right) H^s(dx) \right) $$

  • In particular, if we let $\alpha_n =\int_F x^n H^s(dx)$, with $\alpha_0=1$, then

$$\alpha_n = \frac{1}{2(3^n-1)}\sum_{k=1}^{n}\binom{n}{k}2^k\alpha_{n-k};$$

this is immediate from linearity and the binomial theorem. One checks that these values agree with the first few that we computed.

  • For continuous functions $g$ on $[0,1]$, one argues analogously: fix $\epsilon>0$ and use the Stone-Weierstrass theorem to approximate $g$ uniformly by a polynomial to within $\epsilon$. Then use the recursion to integrate the polynomial.

  • The translation was tricky at this point, but I think one can unite the Hausdorff measure with a 'singular' measure. The idea is to take a piecewise constant function on $[0,1]\setminus C_m$ and take the supremum to yield a function $h$ that is continuous with $h'(x)=0$ a.e. This induces a measure $\mu_h$, after which $$ \int _F f(x) H^s(dx) = \int _0^1 f(x) \mu_h(dx) $$

  • The rest of the paper talks about self-similar sets (of which the [ternary] Cantor set is one) and generalizes the similarity transformations to different parameters. In particular, there is a section on the von Koch curve that I have omitted.


I tried and failed to replicate some of the calculations you mentioned in your post. In particular, I'd love to see how you got the value of $\left(\frac{1}{e}\right)^{-1-2e}\cot(e\pi)\cos(e\pi)$ for that one function.