1
$\begingroup$

I want to compute the expected norm of a vector-matrix multiplication. I have a vector $x \in \mathbb{R}^n$ with norm one and a matrix $M \in \mathbb{R}^{n \times n}$, whose entries are iid taken from the Gaussian distribution with mean zero and variance $2/n$: $M_{i,j} \sim \mathcal{N}(0,2/n)$.

I need to compute the expected value of the norm of the product, i.e. $ \mathbb{E}[||xM||].$

I have computed $\mathbb{E}[||xM||^2] = 2$, but i have no idea on how to get rid of that square sign.

I could use the chi or Gamma functions, but I'm somehow stuck:

I'd have $$\sqrt{\sum_i\sum_j^n x_i^2 M_{i,j}^2 + \sum_i \sum_j \sum_k x_i x_j M_{i,j} M_{i,k}}$$

I know I can use the gamma function for the first sum, and that the expected value of the second part goes to zero. The problem is that I can't compute the expected value of just the second part since it's under the square root. Any suggestions?

Thank you!

$\endgroup$

1 Answer 1

2
$\begingroup$

The key is the simple observation that, for any orthogonal matrix $Q$, the matrix $QM$ equals $M$ in distribution. Let now $Q$ be any orthogonal matrix such that $xQ=e_1:=[1,0,\dots,0]$. Then $xM$ equals \begin{equation} xQM=e_1M=[M_{1,1},\dots,M_{1,n}] \end{equation} in distribution. So, $\|xM\|$ equals $\sqrt{\frac2n\,X}$ in distribution, where $X\sim\chi^2_n=\text{Gamma}(\frac n2,2)$. It follows that \begin{equation} E\|xM\|=\frac{2\Gamma(\frac{n+1}2)}{\sqrt n\,\Gamma(\frac n2)}, \end{equation} which is close to $\sqrt 2=\sqrt{E\|xM\|^2}$ for large $n$, as should be expected, in view of the measure concentration phenomenon.

$\endgroup$
4
  • $\begingroup$ Sorry, I'm trying to plug numbers inside the equation you provided, but even for large $n$, the final output is not close to $\sqrt2$. How is this possible? Thanks! $\endgroup$
    – Alfred
    May 28, 2019 at 10:55
  • 1
    $\begingroup$ @Alfred : I guess you made a mistake somewhere. I have rechecked the limit manually using Striling's formula, and it is $\sqrt2$ indeed. Look also at the section "Series expansion at $n=\infty$" at wolframalpha.com/input/… $\endgroup$ May 28, 2019 at 17:17
  • 1
    $\begingroup$ @Alfred : You can also look at the numerics at wolframalpha.com/input/… $\endgroup$ May 28, 2019 at 17:23
  • $\begingroup$ I did make a mistake, thank you again! $\endgroup$
    – Alfred
    May 29, 2019 at 9:46

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.