1
$\begingroup$

Similarly to this question I am seeking the derivative of a covariance function with respect to its parameters. However, that question is specifically about the squared exponential kernel.

How can I calculate the hyperparameter gradients for the Matérn 5/2 kernel?

$$k_{\mathrm{Matern}}(r) = \sigma^2 \frac{2^{1-\nu}}{\Gamma(\nu)}\bigg(\frac{\sqrt{2\nu}r}{\ell}\bigg)^\nu K_{\nu}\bigg(\frac{\sqrt{2\nu}r}{\ell}\bigg)$$

$\endgroup$
1
$\begingroup$

How can I calculate the hyperparameter gradients for the Matérn 5/2 kernel?

If the general Matern kernel is written $$ \mathcal{K}_M(r) = \sigma^2\frac{2^{1-\nu}}{\Gamma(\nu)} \left( \frac{r\sqrt{2\nu}}{\ell} \right)^\nu K_\nu\left( \frac{r\sqrt{2\nu}}{\ell} \right) $$ then the $5/2$ kernel is given by $$ \mathcal{K}_{M,5/2}(r) = \sigma^2\left[1 + \frac{r\sqrt{5}}{\ell} + \frac{5r^2}{3\ell^2} \right] \exp\left(-\frac{r\sqrt{5}}{\ell}\right) $$ Then the gradient with respect to $\ell$ and $\sigma$ is \begin{align} \frac{\partial \mathcal{K}_{M,5/2} }{\partial \ell} &= \sigma^2\exp\left(-\frac{r\sqrt{5}}{\ell}\right)\frac{\partial }{\partial \ell} \left[1 + \frac{r\sqrt{5}}{\ell} + \frac{5r^2}{3\ell^2} \right] + \sigma^2\left[1 + \frac{r\sqrt{5}}{\ell} + \frac{5r^2}{3\ell^2} \right] \frac{\partial }{\partial \ell} \exp\left(-\frac{r\sqrt{5}}{\ell}\right) \\ &= \sigma^2\exp\left(-\frac{r\sqrt{5}}{\ell}\right) \left[ \frac{-r\sqrt{5}}{\ell^2} - \frac{10r^2}{3\ell^3} \right] + \sigma^2\left[1 + \frac{r\sqrt{5}}{\ell} + \frac{5r^2}{3\ell^2} \right] \left(\frac{r\sqrt{5}}{\ell^2}\right) \exp\left(-\frac{r\sqrt{5}}{\ell}\right) \\ &= \sigma^2\exp\left(-\frac{r\sqrt{5}}{\ell}\right) \left[ \frac{-r\sqrt{5}}{\ell^2} - \frac{10r^2}{3\ell^3} + \frac{r\sqrt{5}}{\ell^2} + \frac{5{r^2}}{\ell^3} + \frac{5\sqrt{5}r^3}{3\ell^4} \right] \\ &= \sigma^2\exp\left(-\frac{r\sqrt{5}}{\ell}\right) \left[ \frac{5{r^2}}{3\ell^3} + \frac{5\sqrt{5}r^3}{3\ell^4} \right] \\ &= \frac{5{r^2}\sigma^2}{3\ell^3} \exp\left(-\frac{r\sqrt{5}}{\ell}\right) \left[ 1 + \frac{r\sqrt{5}}{\ell} \right] \\ \frac{\partial \mathcal{K}_{M,5/2} }{\partial \sigma} &= \frac{2}{\sigma}\mathcal{K}_{M,5/2}(r) \end{align} Source: Rasmussen and Williams, Gaussian Processes for Machine Learning.

$\endgroup$

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.