Let $X$ be a random vector in $\mathbb{R}^n$ whose entries are joint Gaussian with zero mean and covariance matrix $K.$ Is there a closed form expression for $\mathbb{E}||X||_2,$ as there is for the absolute deviation of a standard Gaussian in a 1-dimensional space?
Probability – Expected Norm of a Random Gaussian Vector
probability
Related Solutions
If $X$ is a random variable with density $f$, and $\phi$ is a measurable function, then $E[\phi(X)]=\int_{\Bbb R}\phi(t)f(t)dt$. As the $L^1$ norm of a random variable $X$ is $E[|X|]$, we have, when $X$ is normally distributed, the announced result.
Consider the random vector $X=(X_1,\dots,X_k)$ such that $X\sim N_k(\mu,\Sigma)$.
Set $Q(X)=X^TAX=\sum_{i=1}^kX_i^2$, where $A=\mathrm I_k$. A good reference for the study of $Q(X)$ is "Quadratic forms in random variables" by Mathai and Provost. Specifically, denote by $P$ an orthogonal matrix that diagonalizes $\Sigma$, and write $P^T\Sigma P=\mathrm{Diag}(\lambda_1,\dots,\lambda_k)$. Also define $$ b=P^T\Sigma^{-1/2}\mu. $$ Then, $$ Q(X)=\sum_{i=1}^k\lambda_i\left(U_i+b_i\right)^2,\qquad(1) $$ where $U_i\sim N_k(0,\mathrm I_k)$ (equation $(4.1.1)$ of the reference). This is sometimes called a generalized Chi-squared distribution, and the length of the vector $\sqrt{Q(X)}$ is thus called generalized Chi distribution. The Laplace transform of $Q(X)$ is also obtained in equation $(4.2b.6)$ as $$ \mathrm L(s)=\exp\left(-\frac12\sum_{i=1}^kb_i^2\right)\exp\left(\frac12\sum_{i=1}^kb_i\frac1{1+2s\lambda_i}\right)\prod_{i=1}^k\frac1{\sqrt{1+2s\lambda_i}}, $$ for $\left|2s\lambda_i\right|<1$. Now, in general $(1)$ does not follow a well-known distribution. As you mention in OP, if $\Sigma=\sigma^2\mathrm I_k$ and $\mu\neq0$, then $Q(X)$ follows a non-central chi-squared distribution. Another case that has a closed-form solution is if $\mu=0$, and $\Sigma$ is diagonal with elements $\sigma_1^2,\dots,\sigma_k^2$. Then (c.f. wikipedia as well as $[5]$ therein), if $\sigma_i\neq\sigma_j$, the density of $Q(X)$ can be computed as $$ f(x)=\sum_{i=1}^{k} \frac{e^{-\frac{x}{\sigma_i^2}}}{\sigma_i^2 \prod_{j=1, j\neq i}^{k} (1- \frac{\sigma_j^2}{\sigma_i^2})}1_{x\ge0}. $$ From this you can deduce the distribution of $\sqrt{Q(X)}$. Similar calculations can be done if $\sigma_i=\sigma_j$ for some $i,j\in\{1,\dots,k\}$.
To answer your question is a general setting is more difficult since to my knowledge, these distributions do not follow other known laws. Therefore, it depends on what application you have in mind. If you want to calculate moments, then you might be able to exploit formula $(1)$. If you want to approximate the pdf of $\sqrt{Q(X)}$, then you can use some of the asymptotic expansions of Chapter $4$ of the book that I mentioned in the beginning of this post.
Best Answer
If you can settle with a diagonal covariance matrix, then please check "Multidimensional Gaussian Distributions" by Kenneth S. Miller (1964 edition, chapter 2, section 2, RAYLEIGH DISTRIBUTIONS). Otherwise you need to deal with a lot more complicated equations. This reference could be a good start :
"Properties of Generalized Rayleigh Distributions"
L. E. Blumenson and K. S. Miller
The Annals of Mathematical Statistics
Vol. 34, No. 3 (Sep., 1963), pp. 903-910
You can find a copy of this paper at JSTOR (free sign up!).