I have a general question concerning probabilistic models. Imagine that I have a system with a high-dimensional state $x$ with a pdf $p(x)$. The state $x$ is time-dependent, and a propagating stochastic model $f(x_t|x_{t-1})$ describes its evolution through time.

In many applications where an approximation of the pdf $p(x)$ by deterministic samples is employed, such as (for example) the particle filter, the propagating model $f(x_t|x_{t-1})$ is taken as a combination of a determinstic propagating function $g(x)$ operating on the determinstic previous state value(s) plus some error $\epsilon$:

$$f(x_t|x_{t-1}) = g(x_{t-1}) + \epsilon$$

This error $\epsilon$ is often assumed to be, for example, a (multi)gaussian with mean $0$ and (co)variance $\sigma$:


Now my question: How much freedom do I have in choosing the error type? Is it, for example, justified to introduce a more complex error type that may restrict the outcomes to a subspace of the state-space (as opposed to, for example, a gaussian error, which could theoretically move anywhere)?


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.