The Problem

Given a gaussian random variable $\mathbf{x} = (x_1, ..., x_N)^T \sim \mathcal{N}(0, \Sigma)$ what is the probability that $i = \underset{i\in\{1, \dots, N\}}{argmax}\{|x_1|, \dots, |x_N|\}$?

An Example

Lets say that we are dealing with bivariate normal random variables such that $\mathbf{x} = (x_1, x_2) \sim N(0, \Sigma)$. The problem can be stated geometrically as, what is the probability that the norm of the vector $\mathbf{x}$ falls in region 1 vs region 2 of the unit circle.

Hand drawn picture to help illustrate problem

Note that all internal angles are at 45 degree increments which may be useful for a solution to this problem.

While I illustrate this with a bivariate normal random variable I want an answer that holds for arbitrary finite dimensions.

My Attempts So Far

My initial thoughts are to simplify the problem somewhat by using a reparameterization of $\mathbf{x}$ in terms of the eigendecomposition of $\Sigma = U\Lambda U^T = U\Lambda^{1/2}(U\Lambda^{1/2})^T$. $$\mathbf{x} \sim U\Lambda^{1/2}\mathcal{N}(0, I).$$ My thought is that this problem is easy if $\Sigma = I$ as the answer would then be (by symmetry) that $$P\left(i = \underset{i\in\{1, \dots, N\}}{argmax}\{|x_1|, \dots, |x_N|\}\right) = 1/N$$.

If the problem were framed as an expectation, then this reparameterization would allow us to calculate the difficult integral and the factor $U\Lambda^{1/2}$ would come out of the integral as a constant factor leading to the solution.

Am I onto something here?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.