I need to calculate the likelihood $L(\theta)$ of the Gaussian process:

$ y(t) \sim GP(m(t), K((t,t')|\theta)) $

where $m(t)$ is the mean function (this will be zero for my purposes) and $K((t,t')|\theta)$ the covariance function. If the interval $t$ is $[0,1]$ and for points $t_i,\ i = 1..n$ we have a time series of observations $y_i$, the likelihood can be obtained as the PDF of a multivariate normal distribution. However, the calculation becomes very expensive for large $n$ as it requires the calculation of $inv(K)$ and $det(K)$. This is done as part of a Monte Carlo simulation, making it necessary to speed up this calculation as much as possible. I know there are methods to make this less costly, but if possible I would like to avoid having to discretize the time series altogether.

Is there a way to calculate the likelihood for continuous data, i.e. for $y(t)$ being a continuous function over the interval [0,1] for a fully specified GPR? Drawing an analogy to a simpler case, to illustrate what I want to do:

If I have a realization $x$ of a random variable $X$ from a fully specified univariate distribution, the likelihood is the evaluation of the probability density function for that $x$. Is it possible to do something similar for the case of a GP with continuous measurements?


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.