In the context of Smoothing Splines, Im trying to show that the expected value of the cross-validation can approximate the predictive square error. More specifically, I want to show that

  1. $$E[(y_i - \hat{f}^{-i}_\lambda(x_i))^2] = \sigma^2 + E[(\hat{f}^{-i}_\lambda(x_i) - f(x_i))^2] $$ And that
  2. $$E[(Y_i^* - \hat{f}^{-i}_\lambda(x_i))^2] = \sigma^2 + E[(\hat{f}_\lambda(x_i) - f(x_i))^2] $$

So that if $\hat{f}^{-i}_\lambda(x_i))^2 \approx \hat{f}_\lambda(x_i))^2$ then they are approximately the same.

Note that $\hat{f}^{-i}_\lambda$ is the model fitted with smoothing parameter $\lambda$ and without the ith observation, and $Y_i^*$ is an imaginary new observation at $x_i$ obtained through the same DGP.

I can prove 1 by doing the following to the LHS and expanding out $$E[(y_i - f(x_i) + f(x_i) - \hat{f}^{-i}_\lambda(x_i))^2]$$

However I am unable to prove 2, even though I know it can be done by a similar argument. The issue is mostly that the $\hat{f}^{-i}_\lambda(x_i)$ doesnt appear in the RHS of the second equation?



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.