A simple derivation for the Lebesgue measure of the euclidean unit ball in $\mathbb{R}^n$ follows from computing $$ \int_{\mathbb{R}^n}e^{-\|x\|^2}\,dx $$ in two different ways. See, for instance, Keith Ball, An Elementary Introduction to Modern Convex Geometry, page $5$. Now I was wondering about the following slightly unusual variation:

Let $X_1,\ldots,X_n$ independent random variables with the Cantor distribution.

What is the probability that $X_1^2+\ldots+X_n^2\leq 1$?

I bet this can be tackled by exploiting the fact that the cumulants of the Cantor distribution are given by:

$$ \kappa_{2n}=\frac{2^{2n-1}(2^{2n}-1)\,B_{2n}}{n (3^{2n}-1)},\tag{CM}$$ but how to prove $(\mathrm{CM})$? - This has been answered, but the main question is still open.



For $\mu(x)$ the Cantor measure supported on the Cantor set $\subset [0,1]$ we have the change of variable formula: $$\int f(x)\, d\mu(x) = \frac{1}{2} \int f(1/3 x)\, d\mu(x) + \frac{1}{2} \int f(1/3 x + 2/3)\, d \mu(x)$$ analogous to $\int_0^1 f(x)\, dx =\frac{1}{2} \int_0^1 f(1/2 x)\, d x + \frac{1}{2} \int_0^1 f(1/2 x + 1/2)\, d x $

$\bf{Added:}$ It's easy to see that the first moment $E(X) = \int x \, d \mu(x)= \frac{1}{2}$, and this can be obtained readily from the above formula for $f(x) = x$.

Consider now the central moment generating function $$F(t)\colon =E[e^{t(X-\frac{1}{2})}] = \int e^ {t(x-\frac{1}{2})} \, d\mu(x) $$ From the above equality for $f_t(x) = e^{t(x-\frac{1}{2})}$ we get $$\int e^ {t(x-\frac{1}{2})} \, d\mu(x) = \frac{1}{2}\left( \int e^ {t(\frac{x}{3}-\frac{1}{2})} \, d\mu(x) + \int e^ {t(\frac{x}{3}+\frac{2}{3}-\frac{1}{2})} \, d\mu(x) \right) $$ Now we notice that \begin{eqnarray} t\,(\frac{x}{3}-\frac{1}{2})= \frac{t}{3}(x - \frac{1}{2}) - \frac{t}{3}\\ t\,(\frac{x}{3}+\frac{2}{3}-\frac{1}{2})= \frac{t}{3}(x - \frac{1}{2}) + \frac{t}{3} \end{eqnarray} Therefore we get the equality $$F(t) = \frac{e^{\frac{t}{3}} + e^{-\frac{t}{3}}}{2} \cdot F(\frac{t}{3}) $$

$\bf{Added:}$ Rewrite the above equality as $$F(3t) = \frac{e^t + e^{-t}}{2} F(t)$$ Let $G(t) = \log F(t)$. From the above we get $$G(3 t) - G(t) = \log ( \frac{e^t + e^{-t}}{2})$$

$\bf{Added:}$ Some (moment) calculations: $$\int x d \mu(x) = \frac{1}{2} \left( \ \int (\frac{1}{3} x + \frac{1}{3} x + \frac{2}{3}) d\mu(x) \right )$$ implies $\int x d \mu(x) = \frac{1}{2}$ as expected.

Let's apply the same formula for $f(x) = (x-\frac{1}{2})^n$. We have \begin{eqnarray} \int (x-\frac{1}{2})^n d \mu(x) = \frac{1}{2}\left( \int ( \frac{x}{3} - \frac{1}{2})^n + ( \frac{x}{3} + \frac{2}{3}- \frac{1}{2})^n d\mu(x) \right ) = \\ =\frac{1}{2\cdot 3^n}\left( \int ( x - \frac{1}{2}-1)^n + ( x-\frac{1}{2} + 1)^n d\mu(x) \right ) \end{eqnarray} that is $$M_n = \frac{1}{3^n}\sum_{k \ge 0} \binom{n}{2k} M_{n-2k}$$ which is basically a formula from above $F(3t) = \frac{e^t + e^{-t}}{2} F(t)$.

We get from here $m_2 = \frac{1}{8}$, $m_4 = \frac{7}{320}$, etc. Note that the formula provided in Wikipedia is for the cumulants, not the central moments, as $\kappa_4 = \frac{1}{40}$.

  • $\begingroup$ And so? $\phantom{}$ $\endgroup$ – Jack D'Aurizio Jun 20 '15 at 23:19
  • 1
    $\begingroup$ @Jack D'Aurizio: Using this, you can calculate the central moments, and perhaps more. $\endgroup$ – Orest Bucicovschi Jun 20 '15 at 23:47
  • 2
    $\begingroup$ Ok, so we have a derivation for the moment generating function, and the Bernoulli numbers just come from expanding $\log\cosh$ as a Taylor series. That is good, but it is not the main point of my question, i.e. to compute $$\mathbb{P}[X_1^2+\ldots+X_n^2]\leq 1.$$ Anyway, I am upvoting your answer since I'm quite curious about it leading to an effective answer. $\endgroup$ – Jack D'Aurizio Jun 23 '15 at 17:09
  • $\begingroup$ When a question contains two or more questions, how do you know what the "main question" is? $\endgroup$ – GEdgar Jun 26 '15 at 15:20
  • 2
    $\begingroup$ @GEdgar: The answer only deals with the cumulants ( central moments). The main question about vol$ ( C^n \cap B_n)$ ( $C^n$ Cantor set with the product measure, $B_n$ the unit ball ) is still unanswered. $\endgroup$ – Orest Bucicovschi Jun 26 '15 at 16:42

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.