Usage of kernel function for a special case

[niladridas]

Hello! I need input on the following issue.

I need to calculate the individual matrices: K1 = k(z1,z1) , K2 = k(z1,z2), K3 = k(z2,z2).
Once I have them, my mean and covar is calculated as:
mu = K2^T(K1+epsilonI)^{-1}y
Cov = K3 - K2^T(K1+epsilonI)^{-1}K2 + epsilon*I

I am having difficulty in implementing it.
The length scale parameter is going to be optimized.

Thanks.

[gpleiss]

It sounds like you want to compute a GP posterior? Why not simply use an exact GP, train/condition on z1, and test on z2?