I wonder if I make any stupid mistake understanding this, or can there actually a mistake in the bible of Gaussian Processes: Rasmussen & Williams 2006: Gaussian Processes for Machine Learning?
See page 125 below. Since $W$ is defined as negative second partial derivation, shouldn't it change the sign in (5.23) and result in $+{1 \over 2}[(K^{-1}+W)^{-1}]_{ii}{\partial^3 \over \partial f^3_i}\mathrm{log} \ p(y|\hat{f})$ instead?
