I am working on a project where I estimate transition and measurements models for a kalman filter using Gaussian Processes. In order to linearize the models I require the Jacobian of the estimated Guassian Process.

For the single-output case this is no problem, But I am a little confused about how to do this for the multi- output case.

The posterior mean of the gaussian process would be

\begin{equation} \begin{aligned} \bar{f}_* &= \mathbf{k}(\mathbf{x}_* \, \mathbf{X}) K(\mathbf{X}, \mathbf{X}) ^{-1} \mathbf{y}\\ &\stackrel{\triangle}{=} \mathbf{k}(\mathbf{x}_* \, \mathbf{X}) \mathbf{\alpha} \end{aligned} \end{equation}

And the posterior Covariance

\begin{equation} \begin{aligned} C &= \begin{bmatrix} k(X, X) & k(X, x_*) \\ k(x_*, X) & k(x_*, x_*) \end{bmatrix} \end{aligned} \end{equation}

\begin{equation} \begin{aligned} \text{covariance} &= k(\mathbf{x}_*, \mathbf{x}_*) - k(\mathbf{x}_*, \mathbf{X})k(\mathbf{X}, \mathbf{X})^{-1}k(\mathbf{x}_*, \mathbf{X})^T \\ \end{aligned} \end{equation}

For the single-output case this is no problem

\begin{equation} \begin{aligned} \frac{\partial\bar{f}_*}{\partial \mathbf{x}_*} &= \frac{\partial k(\mathbf{x}_* \, \mathbf{X})}{\partial \mathbf{x}_*} \mathbf{\alpha}\\ &= \left[ \Lambda^{-1} \tilde{\mathbf{X}}^T_* (\mathbf{k}(\mathbf{x}_* \, \mathbf{X})^T \odot \mathbf{\alpha}) \right] \in \mathcal{R}_{D \times 1} \end{aligned} \end{equation}

However, my GP in this case maps $ f: R^2 \rightarrow R^2 $


\begin{equation} \begin{aligned} \mathbf{x}_* &= \begin{bmatrix} x_1 \\ x_2 \end{bmatrix} \\ P\left(\begin{bmatrix} y_1\\ y_2 \end{bmatrix} \ \vert \cdots \right) &= N\left( f( \begin{bmatrix} x_1\\ x_2 \end{bmatrix}) \vert \bar{f}_*, \mathbf{K} \right) \end{aligned} \end{equation}

But I am unsure how to formulate the Jacobian matrix of this function since I do not know the individual functions for each output dimension as I would in a deterministic case.

Any help would be greatly appreciated.


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.