4
$\begingroup$

Let $(X_1, X_2, \ldots, X_i, \ldots, X_k)$ be $k$ independent, normally distributed random variables with means $\mu_i$ and variances $1$. Then the random variable $$ \sum_{i=1}^k X_i^2$$ is distributed according to the noncentral chi-squared distribution with parameters: $k$ and $ \lambda=\sum_{i=1}^k \mu_i^2$.

If $(X_1', X_2', \ldots, X_i', \ldots, X_k')$ are also independent normal distributed with variances $1$, and means $\mu'_i$'s s.t. $\sum_{i=1}^k \mu_i'^2 = \lambda$, why are $$ \sum_{i=1}^k X_i'^2$$ and $$ \sum_{i=1}^k X_i^2$$ identically distributed? Thanks.

$\endgroup$
  • $\begingroup$ What kind of explanation are you looking for? For instance, there are many ways to demonstrate that such a sum of squared normal variables has a non-central $\chi^2$ distribution, but since you have accepted this fact, it sounds like you are perhaps looking for something more intuitive. $\endgroup$ – whuber Sep 22 '14 at 16:35
  • $\begingroup$ Yes, I am looking for why the two are identically distributed, so that noncentral chi square distribution can be well defined. $\endgroup$ – Tim Sep 22 '14 at 16:37
5
$\begingroup$

The $X_i$ can be written as

$$X_i = \mu_i + Y_i$$

where the $Y_i$ are independent standard normal variates (of mean zero). By virtue of the Pythagorean Theorem in $k$ dimension, the expression

$$\chi^2_{k;\mu_1,\mu_2,\ldots,\mu_k}= X_1^2 + X_2^2 + \cdots + X_k^2$$

is recognizable as the squared distance between the point $(-\mu_1, -\mu_2, \ldots, -\mu_k)$ and $(Y_1, Y_2, \ldots, Y_k)$. Because distances remain unchanged by rotations (and reflections) and the distribution of $(Y_1, Y_2, \ldots, Y_k)$ also remains unchanged by rotations, the distribution of this sum of squares will remain unchanged upon rotating (or reflecting) $(-\mu_1, -\mu_2, \ldots, -\mu_k)$ to any other point. The only quantity that is left invariant under rotations and reflections is the magnitude, whence the distribution of $\chi^2_{k;\mu_1,\mu_2,\ldots,\mu_k}$ depends only on $\|(-\mu_1,-\mu_2, \ldots, -\mu_k)\|$ = $\sqrt{\mu_1^2 + \mu_2^2 + \cdots + \mu_k^2}$. Conventionally, the square of this magnitude is used to index the non-central $\chi^2$ distribution.

$\endgroup$

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.