5

Let $x_i\in\mathbb{R}^N$ be multivariate normal distributed with mean vector $\mu\in\mathbb{R}^N$ and no correlation: $x_i\sim N(\mu,\sigma^2 I_N)$. Given $T$ iid samples, define the matrix $$X:=\left( \begin{array}{c} x_1' \\ \vdots \\ x_T'\end{array} \right)\in\mathbb{R}^{T \times N}$$ I am interested how the resulting distribution of the elements of the matrix $X'X$ can be characterized. Clearly, the diagonal element $X'X_{i,i}$ contain the sum of independent squared normal variables with mean vector $\mu_i$ and variance $\sigma^2$. Therefore $\sigma^2X'X_{i,i}$ corresponds to a non-central chi squared distribution with $T$ degrees of freedom and non-centrality parameter $\lambda=T\frac{\mu_i^2}{\sigma^2}$. But what can we say about the distribution of the off-diagonal elements $X'X_{i,j}$?

Update: For a better understanding I appreciate every comment that helps to solve the simplified problem with $\mu=0$.

Tim
  • 108,699
  • 20
  • 212
  • 390
muffin1974
  • 1,152
  • 11
  • 24
  • 2
    (+1) A great deal can be said, and already has been (although not to the extent of giving the full multivariate distribution): see http://stats.stackexchange.com/questions/5399, http://stats.stackexchange.com/questions/51699, and http://stats.stackexchange.com/questions/85916. – whuber May 29 '15 at 17:29
  • 1
    One natural way of approaching this problem would be to express $X_i$ in spherical coordinates as the product of a unit vector and a length. This works out nicely when all means $\mu$ are zero, for then the unit vectors are uniformly distributed and the lengths have a $\chi(n)$ distribution. It's going to get trickier with nonzero means... . – whuber Jun 01 '15 at 16:22
  • Okay, I understand that point. The problem with non-zero $\mu$ would be that we do not have a uniform distribution on the unit sphere any more...is there any way to handle this? – muffin1974 Jun 01 '15 at 16:55
  • 1
    Not only is the distribution of unit vectors non-nonuniform, the distribution of the radius becomes a non-central $\chi$ distribution. There are ways to simplify the situation and make it more tractable. For instance, by adopting a suitable coordinate system you may assume $\sigma=1$ and $\mu$ has only one non-zero component. That effectively reduces the heart of the problem to a one-dimensional one. – whuber Jun 01 '15 at 16:58
  • Is there a reference for such a transformation? – muffin1974 Jun 02 '15 at 08:19
  • 2
    What whuber is suggesting it to apply a rotation which sends $\mu$ to $(||\mu||, 0, \dots, 0)'$ without modifying the variance-covariance. You can combine that with a multiplication by $1/\sigma$ to get rid (momentarily) of the variance. – Elvis Jun 02 '15 at 09:13
  • Thank you Elvis for your comment. I get the point and understand that it is possible to apply a passive rotation of the coordinate system such that the vector $\mu$ is transformed onto $(||\mu||,0,\ldots,0)$ whereas I make use of the rotational symmetry of the standard normal distribution and the fact that I can standardize the variance by dividing with $1/\sigma$. But I do not see what to do next: Can I simply apply the results posted by whuber for the $N-1$ dimensional subspace with 0-mean vector? But what about the $N$th term? – muffin1974 Jun 02 '15 at 10:02
  • I don’t think that it’s so easy... I’ll try to have a more serious look at it later. – Elvis Jun 02 '15 at 21:08
  • Would this problem be simplified if we just relay on the first moment of this distribution? – muffin1974 Jun 03 '15 at 12:43

0 Answers0