We have a bivariate normal process where $X, Y \sim N(0, \sigma)$, with no covariance.

(For convenience we can assert that $\sigma = 1$, or that we have a good estimate for its value.)

What is the distribution of the random variable $R(n) = \sqrt{\overline{x_i}^2 + \overline{y_i}^2}$ — i.e., the Euclidean distance between the sample center of a n points and the true center (at the origin)?

Note that as defined:

  1. $R(n) \ge 0$
  2. $E[R(n)] \to 0$ monotonically as $n \to \infty$
  3. The Rayleigh distribution gives us $R(1) = \sigma \sqrt{\pi/2} \approx \sigma 1.25$

Furthermore, based on a Monte Carlo simulation for $n \in [2, 25]$ with $\sigma = 1$:

  1. Variance decreases monotonically as n increases
  2. Skewness appears constant across n at 0.63
  3. Kurtosis appears constant across n at about 0.24

(This question is a simplified version of this slightly more complicated one that seems to have gotten derailed in complications.)


The sample means are zero mean normal r.v.'s ,

$$\bar X \sim N(0,\sigma^2/n),\;\; \bar Y \sim N(0,\sigma^2/n)$$

Then we have the r.v.'s

$$Z_x = \left(\frac{\bar X}{\sigma/\sqrt n}\right)^2 = \frac n{\sigma^2}\bar X^2 \sim \chi^2(1),\;\;Z_y = \left(\frac{\bar Y}{\sigma/\sqrt n}\right)^2 =\frac n{\sigma^2}\bar Y^2\sim \chi^2(1),$$

Therefore, the r.v. $$W = Z_x + Z_y =\frac n{\sigma^2}\left(\bar X^2+\bar Y^2\right)\sim \chi^2(2)$$

By the properties of a chi-square random variable, we have

$$W_n=\frac {\sigma^2}nW \sim \text {Gamma}(k=1, \theta = 2\sigma^2/n) = \text{Exp}(2\sigma^2/n)$$ i.e.

$$f_{W_n}(w_n) = \frac {n}{2\sigma^2}\cdot \exp\Big \{-\frac {n}{2\sigma^2} w_n\Big\}$$

Define $R_n = \sqrt {W_n}$. By the change-of-variable formula we have

$$W_n = R_n^2 \Rightarrow \frac {dW_n}{dR_n} = 2R_n$$ and so

$$f_{R_n}(r_n) = 2r_n\frac {n}{2\sigma^2}\cdot \exp\Big \{-\frac {n}{2\sigma^2} r_n^2\Big\} = \frac {r_n}{\alpha^2} \exp\Big \{-\frac {r_n^2}{2\alpha^2} \Big\},\;\;\alpha \equiv \sigma/\sqrt n$$

So $R_n$ also follows a Rayleigh distribution with parameter $\alpha$. We have

$$\begin{align} &E(R_n) = (\sigma/\sqrt n)\sqrt {\pi/2} \Rightarrow \lim_{n\rightarrow \infty} E(R_n) =0\\ &\operatorname {Var}(R_n) = \frac {4-\pi}{2}(\sigma/\sqrt n)^2\Rightarrow \lim_{n\rightarrow \infty} \operatorname {Var}(R_n) =0\\ &\text{skewness} =\frac{2\sqrt{\pi}(\pi - 3)}{(4-\pi)^{3/2}} =0.6311\\ &\text {kurtosis} =-\frac{6\pi^2 - 24\pi +16}{(4-\pi)^2} =0.2450\\ \end{align}$$

as the Monte Carlo simulation has given. Skewness and Kurtosis do not indeed depend on the parameter.

  • $\begingroup$ Interesting: I should have guessed that this was Rayleigh distributed when I saw constant skewness and kurtosis with those values. I wonder how many distributions can be identified via constant moments? $\endgroup$ – feetwet May 2 '14 at 23:38
  • 2
    $\begingroup$ The Logistic and Exponential distributions are two distributions that have higher moments independent of their parameters. $\endgroup$ – Alecos Papadopoulos May 3 '14 at 0:17
  • 2
    $\begingroup$ Generalizations for correlated and biased cases are described in stats.stackexchange.com/a/185204/22077 $\endgroup$ – Felipe G. Nievinski Dec 5 '15 at 18:50
  • $\begingroup$ @AlecosPapadopoulos - since you nailed this, I would be tremendously grateful if you could take a look at this related question. $\endgroup$ – feetwet May 26 '17 at 14:56

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.