Let $X_1, \dots, X_n$ be iid standard normal random variables. Consider the vector $X = (X_1, \dots, X_n)$ and the vector

$Y = \frac{1}{\|X\|}(X_1, \dots, X_k)$

for $k < n$. What is $\mathbb{E}[\|Y\|^2] = \mathbb{E}\left[\frac{\sum_{i=1}^k X_i^2}{\sum_{i=1}^n X_i^2} \right]$?

  • 2
    $\begingroup$ By symmetry, the result is exactly $$\color{red}{\bf\frac{k}n}.$$ $\endgroup$ – Did Feb 10 '16 at 11:06
  • $\begingroup$ @Did Thanks, that's exactly the result. Could you please elaborate on your symmetry argument? My reasoning was: The numerator is a Chi-square random variable, with expectation $k$, and the numerator is also a Chi-square random variable with expectation $n$. However, it is not clear to me how to get $k/n$ from this (cc @ecstasyofgold) $\endgroup$ – rodms Feb 10 '16 at 14:27
  • $\begingroup$ @Cm7F7Bb Thanks for your suggestion, I will edit the question. $\endgroup$ – rodms Feb 10 '16 at 14:27
  • 2
    $\begingroup$ For every $i\ne j$, the distributions of $(X_i,\|X\|)$ and $(X_j,\|X\|)$ coincide hence $E(X_i^2\|X\|^{-2})=E(X_j^2\|X\|^{-2})$. Summing these from $i=1$ to $n$ yields the expectation of $(X_1^2+\cdots+X_n^2)\|X\|^{-2}=1$ hence $E(X_i^2\|X\|^{-2})=n^{-1}$ for every $i$ and you are done. $\endgroup$ – Did Feb 10 '16 at 14:30
  • $\begingroup$ @Did Thanks so much, amazing answer! :) $\endgroup$ – rodms Feb 10 '16 at 14:34

I don't know the closed form of this but concentration of measure phenomenon ensures that the expectation will converge to $\frac{k}{n}$ exponentially fast in $k$. So using $k/n$ might be a reasonable approximation.


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.