Given a random process $x[n] \in \mathbb{R}$ (say of length $N$) and all correlation functions such as:

\begin{align} \langle x[i]\rangle\\ \langle x[i]x[j]\rangle\\ \langle x[i]x[j]x[k]\rangle\\ \vdots \end{align}

  1. Is it possible to simulate a single realization of the random process and if so how? Thinking abstractly about the problem it seems like any trajectory could be mapped onto a vector in $\mathbb{R}^n$ and there is some probability distribution on $\mathbb{R}^n$. Sampling this distribution and then picking out the corresponding trajectory would then constitute of "simulation" of a single realization that I am looking for. It seems like given information about all of the auto-correlation functions it should be possible to determine the probability distribution. This seems related to the inverse moment problem.
  2. What if we truncate and only have up to $k$-time correlation function with $k<N$? My guess is that it is not possible to uniquely generate the probability distribution but it should be possible to (non-uniquely) simulate a process with matching auto-correlation functions up to $k$.

1 Answer 1


Here is one possible method that you could possibly use.

Let's model your random process $x[n]$ as an auto-regressive (AR) process of order $P$, that is, $$x[n] + \sum_{p=1}^P a_px[n-p] = c + \epsilon[n]$$ where $c$ is some constant that I will set to 0 in what follows for the sake of simplicity, $\{a_p\}_{p=1}^P$ are the parameters of the AR model and $\epsilon[n]$ is some white noise of variance $\sigma^2_\epsilon$.

Multiplying each side of the above equation by $x[n-l]$ for some $l \ge 0$ leads us to $$\left(x[n] + \sum_{p=1}^P a_px[n-p]\right)x[n-l] = \epsilon[n]x[n-l].$$ Now, taking the expectation on each side, we obtain the Yule-Walker equations $$C_x[l] + \sum_{p=1}^P a_pC_x[l-p] = \begin{cases} 0 & \text{if} & l > 0, \\ \sigma^2_\epsilon & \text{if} & l = 0 \end{cases}$$ using the fact that $\epsilon[n]$ and $x[n-l]$ are uncorrelated for $l > 0$ (you can see that from the equation of the AR process above). Using this equation with $l = 0, 1, 2, \dots, P$, you have $P+1$ equations to solve for the $P + 1$ unknowns (i.e., $\{a_p\}_{p=1}^P$ and $\sigma^2_\epsilon$).

You then have a model to generate a time series given the temporal auto-correlation.

  • $\begingroup$ thank you for your answer. It looks like this answers my question in the case that the stochastic process can be modeled as an AR process. I guess this does not work for more complicated processes, for example a process which follows a non-linear stochastic differential (difference) equation. Is this correct? I wonder if there is an approach that would work for a fully general random process. $\endgroup$
    – Jagerber48
    Feb 5, 2019 at 17:50
  • $\begingroup$ I'm not aware of the existence of a "general approach" but I'm not an expert in that field. $\endgroup$
    – anpar
    Feb 6, 2019 at 8:01
  • $\begingroup$ Still, I believe my answer answers your question. You might want to ask a new question for the more general case you describe in the above comment. $\endgroup$
    – anpar
    Feb 11, 2019 at 20:15
  • $\begingroup$ I've asked the question on mathematics stack exchange asking for the more general approach: math.stackexchange.com/questions/3104827/… . There has been no attention there yet. $\endgroup$
    – Jagerber48
    Feb 11, 2019 at 22:03

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.