When we have a [[Multivariate Gaussian|Gaussian Vector]] $X \in \mathbb R^d$, and we have $n$ copies of that, then $\mu$\bar X_n$ denotes the average of each element of $X$ over all $n$ copies.
- $\bar X_n \in \mathbb R^d$ is a random vector, which converges to a multivariate Gaussian.
- $\mu \in \mathbb R^d$ is a deterministic vector as it holds the population means. When we subtract the vector by its mean, we are centering the multivariate Gaussian around zero on each of the $d$ dimensions.
$ \sqrt n \,(\bar X_n-\mu) \xrightarrow[n \to \infty]{(d)}\mathcal N_d(0, \Sigma) $
To arrive at a statement of convergence to a multivariate standard Gaussian, we take following transformations:
$
\begin{align}
\sqrt n \,(\bar X_n-\mu) &\xrightarrow[n \to \infty]{(d)}\mathcal N_d(0, \Sigma) \tag{1}\\[4pt]
\Sigma^{-1/2}\sqrt n \,(\bar X_n-\mu) &\xrightarrow[n \to \infty]{(d)} \Sigma^{-1/2} \mathcal N_d(0, \Sigma) \tag{2} \\[4pt]
\Sigma^{-1/2}\sqrt n \,(\bar X_n-\mu) &\xrightarrow[n \to \infty]{(d)} \mathcal N_d \big(0, \Sigma^{-1/2} \, \Sigma (\Sigma^{-1/2})^T \big) \tag{3} \\[4pt]
\Sigma^{-1/2} \sqrt n \,(\bar X_n-\mu) &\xrightarrow[n \to \infty]{(d)}\mathcal N_d(0, \mathbf I_d) \tag{4}
\end{align}
$
where:
- (2) Multiply both sides with $\Sigma^{-1/2}$.
- (3) Apply [[Covariance Matrix#Affine Transformation|Affine Transformation to Covariance Matrix]].
- (4) Since a covariance matrix is always positive definite, the variance of $\mathcal N_d$ simplifies to the identity $\mathbf I_d$.