Let $z_1,z_2,\dots,z_n$ be $n$ complex numbers distributed uniformly and randomly over the unit disc $x^2+y^2 \leq 1$. Let $z$ be the complex number defined by the mean of the of these numbers,that is, $$ z=\frac1n \sum_{i=1}^{n} z_i.$$ Numerical simulations reveal that $E(|z|)$ exists. I want to know if there is some analytical way of finding a closed form expression for the expected radial distance of $z$ in terms of $n$? Any hints/responses will be greatly appreciated . The graph shows the variation of $E(|z|)$ with $n$ enter image description here

  • $\begingroup$ This seems like a difficult version of an "expected distance of random walk from the origin" type of result, but I can't find a clearly related result. Usually the length of each step is fixed (here, it is random), and only the direction is random. $\endgroup$
    – Steve
    Jun 9 at 8:54
  • $\begingroup$ The value for $n=1$ is $\frac{2}{3}$, which can be shown by solving a two-dimensional integral. Solving the 4-dimensional integral for $n=2$ should be withing reach, perhaps. $\endgroup$ Jun 9 at 11:59
  • 1
    $\begingroup$ For $n=1$ it is simply $\int_0^1 \sqrt{x} dx = 2/3$ because the direction does not matter and the distance from zero of each $z_i$ is distributed as the square root of a uniformly distributed random variable on $[0, 1]$. The case $n=2$ seems much harder. $\endgroup$
    – Steve
    Jun 10 at 6:51
  • $\begingroup$ I calculated numerically the value for $n=2$ and I have a conjecture as to what the exact value might be: mathoverflow.net/q/395659 $\endgroup$ Jun 18 at 12:35
  • $\begingroup$ Along the same lines as Kluyver,"A local probability problem" (1906), one can derive that $|\sum_{i=1}^n z_i |$ has density $\rho(x) = \frac{1}{2\pi} \int_0^\infty dt\, t\,J_0(x\, t) ( 2J_1(t) / t)^n$. $\endgroup$ Jun 18 at 15:45

Here is another method.

Since the uniform distribution on the unit disk is rotationally symmetric (invariant under orthogonal transformations), the problem can be reduced to a random walk problem in $\mathbb{R}$.

Let $Z_1,Z_2,\ldots$ be independent, and identically uniformly distributed on the unit disk, with $Z_i=(X_i,Y_i)$. Clearly $X_1,X_2\ldots$ are i.i..d with density $p(x)=\frac{2}{\pi}\sqrt{1-x^2}1_{[-1,1]}(x)$.

For any rotationally symmetric random vector $V=(V_1,V_2)$ with finite expectation of $|V|:=\sqrt{V_1^2+V_2^2}$ it holds that $$\mathbb{E}(|V|)=\frac{\pi}{2}\mathbb{E}(|V_1|)$$ Applying this to $Z^{(n)}:= |Z_1+\ldots+Z_n|$ we get that

\begin{align*} \mathbb{E}\big(\frac{Z^{(n)}}{n}\big)&=\frac{1}{n}\frac{\pi}{2}\mathbb{E}|X_1+\ldots + X_n|\\ &=\frac{1}{n}\big(\frac{2}{\pi}\big)^{n-1}\,\int_{-1}^1\ldots\int_{-1}^1|x_1+\ldots+x_n|\prod_{i=1}^n\sqrt{1-x_i^2}\,dx_1\,\ldots dx_n\\ \end{align*}

The integrals can be solved explicitly for $n=1$ and $n=2$, but that seems to be as far as it goes. But (using the central limit theorem and uniform integrability) it is easy to see that $$\mathbb{E}\big(\frac{Z^{(n)}}{\sqrt{n}}\big)\longrightarrow \sqrt{\frac{\pi}{8}}$$

(1) a related problem appeared here An interesting triple integral
(2) for similar considerations see Feller II (1971), p. 30 ff

  • $\begingroup$ @sajjad veeri: for $n=3$ see my answer to the follow-up post by Moritz Firsching. $\endgroup$
    – esg
    Jun 25 at 18:51

This does not answer the question, but reformulates it as an integration problem and addresses a question asked in the comments. Conditionally on $|z_1| = a_1$, ..., $|z_n| = a_n$, the probability that $x<c$ was computed in Kluyver, "A local probability problem" (1906) to be $$ \mathbb{P}\Big( \big|\sum_{i=1}^n z_i\big| < c \,\Big|\, |z_i|=a_i \Big) = c \int_0^\infty J_1(uc)\prod_{i=1}^n J_0(u a_i)\,\mathrm{d}u. $$ Using that $\int_0^1 2a\,J_0(u a) \mathrm{d}a = 2 J_1(u) /u$ and $\frac{\partial}{\partial c}( c J_1(u c)) = uc J_0(u c)$, this gives the density as $$ \rho(x) = \frac{\partial}{\partial x}\mathbb{P}\Big( \big|\sum_{i=1}^n z_i\big| < x\Big) = \int_0^\infty xu J_0(ux) \,\left(\tfrac{2}{u}J_1(u)\right)^n\mathrm{d}u. $$ (Note: there was a small mistake in the formula given in the comments.) By construction $\int_0^n\rho(x)\,\mathrm{d}x = 1$ as $\rho(x)$ vanishes for $x > n$. Hence \begin{align} \mathbb{E}\Big( \frac{1}{n}\big|\sum_{i=1}^n z_i\big|\Big) &= \frac{1}{n} \int_0^n \mathrm{d}x \int_0^\infty \mathrm{d}u\, x^2u\, J_0(ux) \left(\tfrac{2}{u}J_1(u)\right)^n \\ &= \frac{n^2}{3} \int_0^\infty u\,{_1F_2}(\tfrac{3}{2};1,\tfrac{5}{2}; -\tfrac14 n^2u^2) \left(\tfrac{2}{u}J_1(u)\right)^n, \end{align} where the second integral was kindly provided by mathematica.

One could hope to extract some exact values for $n\geq 3$ from Borwein, Jonathan M. "A short walk can be beautiful." Journal of Humanistic Mathematics 6, no. 1 (2016): 86-109.


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.