I am reading "Gaussian Process Training with Input Noise" by Andrew McHutchon and Carl Edward Rasmussen, where it is assumed that the inputs $x$ are noisy measurements of the actual latent input $\tilde{x}$ , i.e $ x = \tilde{x}+\epsilon_x$ with $\epsilon_x\sim N(0,\Sigma_x)$. The observation is: $y=f(\tilde{x}+\epsilon_x)+\epsilon_y$ under a GP model $f$.

The authors consider taylor expansion of the above expression up to the first order terms as follows: $y = f(x) + \epsilon_x^T\delta_\bar{f}+\epsilon_y$, where $\delta_\bar{f}$ represents the derivative of the mean of the GP function. I`d like to know why this 1st order expansion would be an accurate-enough approximation for the input error propagation.

If anyone could help, I would be very grateful.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.