# Compute the mean of normalized norms of linear transformations of Gaussian random vectors

by DeltaIV   Last Updated October 07, 2018 08:19 AM

if $$M$$ is a $$m\times n$$ constant matrix and $$\eta\sim\mathcal{N}(0,I)$$, then does $$\mathbf{E}_{\eta\sim\mathcal{N}}\left[\frac{\lVert M\eta\rVert}{\lVert\eta\rVert}\right]$$ exist? Also, let $$x\in \mathbb{R}^n_{\ne 0}$$ be an arbitrary non-zero vector. Is it possible to compute the maximum (or at least to find a tight upper-bound) over all $$x$$, of the quantity $$\mathbf{E}_{\eta\sim\mathcal{N}}\left[\frac{\lVert M(x+\lVert x\rVert \eta)-Mx\rVert}{\lVert Mx \rVert}\right]=\lVert x\rVert\mathbf{E}_{\eta\sim\mathcal{N}}\left[\frac{\lVert M \eta\rVert}{\lVert Mx \rVert}\right]$$

Tags :