2
$\begingroup$

Let $\mathbf{a} \sim \mathcal{GP}(\mathbf{m},\mathbf{C})$ where $\mathbf{a} \in \mathbb{R}^T$ is modeled as Gaussian process with mean $\mathbf{m} \in \mathbb{R}^T$ and prior covariance $\mathbf{C} \in \mathbb{R}^{T \times T}$. Then the marginal probability of $a_t$ being non-zero can be written as \begin{align*} p(a_t > 0) = \phi(m_t/\sqrt{C_{t,t}}) \end{align*} where $\phi$ is normal cummulative distribution function. Now $\mathbf{a}$ is not a one dimension object, but a 2 dimensional object $\mathbf{A} \sim \mathcal{GP}(\mathbf{M},\mathbf{C}_1 \otimes \mathbf{C}_2) $ where $\mathbf{M} \in \mathbb{R}^{T \times K}$ and prior covariances $\mathbf{C}_1 \in \mathbb{R}^{T \times T}$ and $\mathbf{C}_2 \in \mathbb{R}^{K \times K}$. In this case, how do I calculate the marginal probability of $a_{t,k} > 0$. For one dimension, I can marginalize by just dropping the irrelvant variables, but is it true for more than one variable?

$\endgroup$
  • $\begingroup$ In the question as written, $T$ can be any set equipped with a kernel. So $T$ can be replaced by $T \times K$ and the kernel value at $(t,\,t)$ be replaced by the value of the tensor product kernel at $([t, \,k], \, [t, \,k])$. $\endgroup$ – Yves Dec 11 '19 at 9:05
2
$\begingroup$

Yes, by definition a random variable is Gaussian if and only if all linear functionals are normally distributed. We have $a_{t,k}=e_t'Ae_k=tr(Ae_ke_t')=\langle A, e_ke_t'\rangle$, i.e., $a_{t,k}$ is a linear functional of $A$ and is hence normally distributed. Its mean and variance are $\langle M, e_k e_t'\rangle=e_t'Me_k=m_{t,k}$, and variance computed analogously but more tediously as $(C_1\otimes C_2)_{(t,k),(t,k)}=(C_1)_{t,t}(C_2)_{k,k}$. We therefore have \begin{align*} p(a_{t,k} > 0) = \phi\Big(m_{t,k}\Big/\sqrt{(C_1)_{t,t}(C_2)_{k,k}}\Big) \end{align*}

As Yves remarks, the kernels may be defined over much more general spaces than Euclidean ones.

$\endgroup$
0
$\begingroup$

Maybe this is not the answer you were looking for but if we assume that $\mathbf{a}$ is i.i.d and that $p(\mathbf{a}(t) > 0\vert t_{0}) = \phi(m_{t}\,C_{t,t}^{-\frac{1}{2}})$ then for each $\mathbf{a}\in\mathbf{A}$ that should be $\prod_{\mathbf{a}\in\mathbf{A}} \phi_{\mathbf{a}}(m_{t}\,C_{t,t}^{-\frac{1}{2}})$ if I'm not mistaken.

$\endgroup$

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.