1
$\begingroup$

The radial basis function (RBF) kernel is given by

$$K_{\text{RBF}}(\mathbf{x}, \mathbf{y})=\exp[-\gamma\|\mathbf{x}-\mathbf{y}\|^2_2]$$

where $\|\mathbf{x}-\mathbf{y}\|^2_2$ is the squared Euclidean distance. I have three questions. The answer to all three may be generalized RBF kernel, but I would like to make sure.

First, what is the name of the kernel that results if the Euclidean distance is not squared?

$$K(\mathbf{x}, \mathbf{y})=\exp[-\gamma\|\mathbf{x}-\mathbf{y}\|_2]$$

Second, what is the name of the kernel that results if another Minkowski $p$-norm is taken and raised to the $n^{th}$ power?

$$K_{p,n}(\mathbf{x}, \mathbf{y})=\exp[-\gamma\|\mathbf{x}-\mathbf{y}\|_p^n]$$ $$K_{2,2}=K_{\text{RBF}}$$

And finally, what is the name of the kernel that results if that Minkowski $p$-norm is weighted?

$$K_{p,n, \mathbf{w}}(\mathbf{x}, \mathbf{y})=\exp\left[-\gamma\left(\sqrt[p]{\sum_iw_i|x_i-y_i|^p}\right)^n\right]$$

$\endgroup$
1
$\begingroup$

The first one is called (by some) "exponential covariance function". Covariance relates to the interpretation of the RBF interpolation procedure as a regression task. You can read more atthis book (specifically - the chapter that discusses covariance functions).

$\endgroup$

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.