Source of my doubt is the section 2.7 of GPML book by Rasmussen, an screenshot of the book is attached below. Much of my confusion is clarified by this discussion.
If mean of GP is not estimated and it is taken as zero or any constant, then outside the range of the observations the estimate would converge to 0/constant. But suppose if I am interested in prediction within the domain of training data.
What would happen if I vary the mean of GP? Would the parameters of covariance function tune themselves accordingly?
