I have a very basic understanding of Gaussian Processes. From what I understand, a Guassian process for a set $X$, is the assignment of a Gaussian distribution to every element of the set. This is meant to expand the idea of a function to the case where we don't have total information about a function.

Two functions, $f$ and $g$, have a natural notion of composition $f \cdot g$, ie just function composition. If Gaussian Processes are approximations to functions, is there a notion of composition of two Gaussian Processes? Given two Gaussian Processes, $\mathcal{G}_f, \mathcal{G}_g$, how is the composition defined $\mathcal{G}_f \cdot \mathcal{G}_g$?


1 Answer 1


I think that compositions of Gaussian Processes are called Deep Gaussian Processes.

Typically, I guess the way a composition $f \circ g$ would be defined is, if both functions are defined on $\mathbb R$, then $f \circ g(x) \equiv f(g(x))$.

So, essentially, if both functions were GPs, you'd be calculating:

$$f \sim \mathcal N(0, \Sigma_f(g(x)))$$ $$g \sim \mathcal N(0, \Sigma_g(x))$$

Given that the covariance matrix of $f$ depends on the values $g$, I do not think that there exists any nice closed form descriptions of this process, although a literature search for deep GPs might prove me wrong.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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