I'm interested in the ground for making conditional independence assumption (e.g. that different target dimensions do not covary for given input $x$) when we are modelling some multidimensional signal with Gaussian Process? Why it is natural to assume this? Or it is just our way to manage with the curse of dimensionality?
