Generalising multivariate Gaussians to anything which has a density function of the form \[ f(x)\propto g((x-\mu )'\Sigma ^{-1}(x-\mu ))\] where \(\mu\) is the mean vector, \(\Sigma\) is a positive definite matrix, and \(g:\mathbb{R}^+\to\mathbb{R}^+\). In fact, we do not need the density function to exist; it’s ok if \(\Sigma\) is positive semi-definite or to allow \(g\) to be a generalised function.
Baby steps, though let us have densities for now. If the mean of such an \(X\sim f\) RV exists, it is \(\mu\), and \(\Sigma\) is proportional to the covariance matrix of \(X\), if such a covariance matrix exists.
I assume they did not invent this idea, but Davison and Ortiz (2019) point out that if you have a least-squares-compatible model, usually it can generalise to any elliptical density, which includes many M-estimator-style robust losses.
Recommended reading
OG paper introduction Cambanis, Huang, and Simons (1981) is basically a textbook on the bits that are important to me at least, and it is not a bad textbook at that. K.-T. Fang, Kotz, and Ng (2017) is an actual textbook.
No comments yet. Why not leave one?