0
$\begingroup$

I'm trying to obtain the Ridge Regression solution from the mean of predictive distribution of a Gaussian Process with a linear kernel.

The mean of the predictive distribution of a GP is $$\boldsymbol k^T(\boldsymbol K+\sigma^2I)^{-1}y$$

where $\boldsymbol K=k(\boldsymbol X^T, \boldsymbol X)$ is the Kernel Gram Matrix (Symetric Positive definite) and $\boldsymbol k = k(x_{n+1},\boldsymbol X)$. (Note that $x_{n+1}$ is a vector while $\boldsymbol X$ is a matrix

Now I'm assuming a linear kernel, thus $k(\boldsymbol X^T, \boldsymbol X) = \boldsymbol X^T \boldsymbol X$ and $\boldsymbol k = k(x_{n+1},\boldsymbol X) = x_{n+1}^T* \boldsymbol X$. So I can write $$\boldsymbol k^T(\boldsymbol K+\sigma^2*I)^{-1}y = (x_{n+1} \boldsymbol X^T) (\boldsymbol X^T \boldsymbol X+\sigma^2*I)^{-1}y $$

Now the issue: The Ridge Regression should be $$x_{n_1}(X^TX+\sigma ^2 I)^{-1}\boldsymbol X^T y$$ while I have $$(x_{n+1} \boldsymbol X^T) (\boldsymbol X^T \boldsymbol X+\sigma^2*I)^{-1}y $$

So I need to commute $\boldsymbol X^T$ and $(\boldsymbol K+\sigma^2I)^{-1}$ but I know matrix cannot commute, so I'm stucked.

What am I missing? Can you help me?

Note that $(\boldsymbol K+\sigma^2I)^{-1}$ is a symmetric matrix.

$\endgroup$

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Browse other questions tagged or ask your own question.