Mahalanobis distances measure the distance of a sample unit from the mean
of a distribution, taking into account the correlation between the units
in the covariance matrix. If $x$ is an observation from a multivariate
distribution with mean $\mu$ and covariance $\Sigma$, the Mahalanobis
squared distance (MSD) of $x$ from $\mu$ is $D^{2}=(x − \mu)^{t}
\Sigma^{-1} (x − \mu)$. When $x$ is from a $v$-dimensional multivariate
normal with known mean and covariance, the population MSD is distributed
as a chi-squared $\chi_{v}^{2}$ random variable with $\nu$ degrees of
freedom (Mardia et al, 1979). Then, to test the deviation of an
observation from the multivariate normal assumption we can compare its
MSD with an appropriate quantile of the chi-squared distribution: the
observation will be considered an outlier if the associated $D^{2}$ value
is larger than the critical value of the chi-squared distribution. There
are known limitations to the application of this cut-off, for example
when the sample is high dimensional and its size is not sufficiently
high. In this case the distribution of the sample MSD is a scaled Beta
distribution (Gnanadesikan and Kettenring, 1972). For continuous
Student-t samples, which account for heavy-tailed distributions, the
appropriate cutoff value depends from a standard Beta distribution with
shape parameters $v/2$ and $\nu/2$, as shown by Barabesi et al (2023).