4
$\begingroup$

Suppose $x$ is sampled from zero-centered Gaussian with $d\times d$ covariance matrix $\Sigma$.

  • is there a name for distribution of $y=\frac{x}{\|x\|}$?

  • is there a closed form expression for covariance of $y$?

$\endgroup$
7
  • 4
    $\begingroup$ Hi Yaroslav. Did you check this CV post on projected normal distribution? Also, you can go through this paper about modelling. $\endgroup$ Aug 27, 2022 at 3:26
  • $\begingroup$ Let me sieve through old CV posts. I guess some more related posts could be found. $\endgroup$ Aug 27, 2022 at 3:27
  • 1
    $\begingroup$ Hello @YaroslavBulatov, kindly check this. $\endgroup$ Aug 27, 2022 at 4:17
  • 4
    $\begingroup$ You asked essentially the same question five years ago. Any motivation for considering this ratio rather than$$\frac{x}{(x^\text T\Sigma^{-1}x) ^{1/2}}$$ for instance? $\endgroup$
    – Xi'an
    Aug 27, 2022 at 7:59
  • 4
    $\begingroup$ This is the Kent distribution. $\endgroup$
    – whuber
    Aug 27, 2022 at 14:24

2 Answers 2

1
$\begingroup$

I think that the distribution of $y=\frac{X}{||X||}$ where $X\sim \mathcal{N}(0, \Psi)$ is the projected normal distribution.

I came across a similar problem and used an approximation that works well in most cases. My approximations also work for the case of non-zero mean.

For reference, here is a previous question of mine, that I responded, for the case of $X\sim \mathcal{N}(\mu, \mathbf{I}\sigma^2)$ (i.e. isotropic noise). See the approach used in that question of approximating each element $i,j$ of the second-moment matrix (i.e. $\mathbb{E}\left(\frac{X_i X_j}{||X||^2}\right)$) as a ratio of quadratic forms that uses a symmetric $A$ in the numerator that is specific to each $i,j$.

For your question, we can use the exact same approach of approximating the elements of the second-moment matrix with that same ratio of quadratic forms, but generalizing the answer above to the case where $X\sim \mathcal{N}(\mu, \Psi)$ where $\Psi$ is any symmetric positive semi definite matrix. For this generalized quadratic form, a second order Taylor approximation can be found in this article, section 3.1.

Using the formulas in the paper above for each matrix $A$ corresponding to each element $i,j$ of the second moment matrix (as done in my other question I linked), and converting the resulting equations to matrix formulas, we get the following approximation:

\begin{equation} \label{nonIso} \mathbb{E}\left( \frac{XX^T}{||X||^2} \right) \approx \frac{\mu_N}{\mu_D} \odot \left( 1 - \frac{\Sigma^{N,D}}{\mu_N\mu_D} + \frac{Var(D)}{\mu_D^2} \right) \end{equation}

where the terms are defined as follows: \begin{equation} \begin{split} & \mu_N = \Psi + \mu \mu^T \\ & \mu_{D} = tr(\Psi) + ||\mu||^2 \\ & Var(D) = 2 tr(\Psi^2) + 4 \mu^T \Psi \mu \\ & \Sigma^{N,D} = 2 \left[\Psi \Psi + \mu \mu^T \Psi + \Psi\mu \mu^T \right] \end{split} \end{equation}

Note that while $\mu_N \in \mathbb{R}^{d\times d}$ and $\Sigma^{N,D} \in \mathbb{R}^{d\times d}$, $\mu_D \in \mathbb{R}$ and $Var(D) \in \mathbb{R}$. Also, importantly, $\odot$ is element-wise multiplication, and the ratios between matrices that appear there are also element-wise.

Maybe for your case of 0 mean you would be able to apply the same approach and find an exact formula for the ratios, instead of this approximation.

Finally, that is the second-moment matrix, not really the Covariance matrix. You can subtract the outer product of the expected value of the projected Gaussian, to obtain the covariance the following way: $$ Cov\left( \frac{X}{||X||} \right) = \mathbb{E}\left( \frac{XX^T}{||X||^2} \right) - \mathbb{E}\left( \frac{X}{||X||} \right) \mathbb{E}\left( \frac{X}{||X||} \right)^T$$

I would suppose that because your variable is centered and its symmetric, the $\mathbb{E}\left( \frac{X}{||X||} \right) = 0$, though I can't say for sure.

I tried this approximation on simulated data and it works quite well.

$\endgroup$
5
  • $\begingroup$ Interesting! I was particularly interested in special case of zero-centered Gaussian with diagonal covariance, entries 1,1/2,1/3,1/4,....1/d. What do these look like after normalizing? Simulation show new eigenvalues are also harmonically decaying, will check if your approach matches this $\endgroup$ Mar 14 at 16:42
  • $\begingroup$ If you test out this approach, do let me know in the comments how it fares. I'm using this approach for a problem of mine with different covariance structures (diagonal covariance proportional to the mean vector, spatially-dependent correlations, etc), and would be interested in hearing if it does/doesn't work in other cases that may be relevant. $\endgroup$
    – dherrera
    Mar 14 at 17:27
  • $\begingroup$ Your "Var" variable is a scalar, and you are adding it to a matrix, are you assuming automatic broadcasting here? $\endgroup$ Mar 14 at 20:07
  • $\begingroup$ Yes, it assumes automatic broadcasting $\endgroup$
    – dherrera
    Mar 14 at 21:37
  • $\begingroup$ I posted evaluation as an answer, it simplifies considerably for zero-centered trace-normalized Gaussian and gives a nice fit, thanks! $\endgroup$ Mar 14 at 21:39
0
$\begingroup$

By applying dherera's formula to trace-normalized zero-centered Normal with diagonal $d\times d$ covariance matrix $H$, we get the following estimate for the corresponding covariance matrix $H_p$ of projected Normal.

$$H_p = c H -2H^2\\ c=(1+2\operatorname{Tr}(H^2)) $$

When $H$ eigenvalues follow power-law decay with $p=1.1$ we can estimate eigenvalues of $H_p$ using Monte-Carlo and see the following fit:

enter image description here

For $1<p<2$, this estimate is slightly biased.

For instance, when $p=1.1$ using "adjusted" formula by changing $c$ to be $c=(1+1.7\operatorname{Tr}(H^2))$ gives a slightly nicer fit: enter image description here

Error measures mean relative residual squared, tail error only considers smallest $d/2$ eigenvalues.

Notebook

$\endgroup$
6
  • $\begingroup$ The "adjustment" needed to render dherrrera's formula unbiased seems to be dependent on $p$ in p-powerlaw decay eigenvalues, wondering if there's an automatic way to obtain it. Since bias is present in tail eigenvalues, perhaps it could be derived by taking limit $d\to \infty$ $\endgroup$ Mar 14 at 21:58
  • $\begingroup$ Just to share my own results, I'll say that when testing the formula on my own problem with covariance matrices with different structures, I found the approximation to be close to perfect for some covariances, and to have some bias for other covariances. Interestingly, I had found the approximation to work great for diagonal matrices in my case (random diagonals and diagonals proportional to the mean). However, performance evidently depends on the specifics, as your case shows. I'm still working on this problem, so I'll post updates if I find anything useful. $\endgroup$
    – dherrera
    Mar 14 at 22:47
  • $\begingroup$ Sounds good, please comment when you do. I'm assuming random diagonal is going to be close to uniform. I'm looking at diagonal with values 1^p,2^p,3^p,.... Worst case bias is when p=-sqrt(2) $\endgroup$ Mar 14 at 23:16
  • $\begingroup$ Also I'd be curious if your work gave any insight to fourth-moments. For Gaussian, 4th moment factors according to Wick's theorem, I suspect for projected Gaussian it approximately factors -- stats.stackexchange.com/questions/608943/… $\endgroup$ Mar 15 at 0:02
  • 1
    $\begingroup$ The paper, however, seems like a good place to start looking for other references on moments of quadratic forms, which is what the problem ends up boiling to. Either exact formulas for the 0 mean case, or approximations to the second moment of the quadratic form (fourth moment of the X/||X|| variable). $\endgroup$
    – dherrera
    Mar 15 at 0:22

Your Answer

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

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