0
$\begingroup$

Assume a time series composed by many recurring events coming from many different poisson process each with a different rate. Lets assume for simplicity no overlap between events. Is there any math/statistical tool to decompose the given time series and get the various poisson processes. Ideally like fourier analysis can yield the harmonics composing a periodic signal, the tool should provide a list of poisson processes composing the whole sequence.

$\endgroup$
  • $\begingroup$ Do you only observe their sum? $\endgroup$ – dsaxton Mar 2 '16 at 16:16
  • $\begingroup$ Yes, you can only observe the sum. But events never overlaps $\endgroup$ – Davide C Mar 2 '16 at 16:49
  • 1
    $\begingroup$ My first impression is that you can't. The sum itself will be Poisson$(\lambda)$, but there is an unlimited number of ways, all apparently equally valid, in which you could decompose this total count into a sum of Poisson random variables. $\endgroup$ – dsaxton Mar 2 '16 at 16:57
  • 1
    $\begingroup$ Sum of Possions is Poisson. So, no, you can't decompose the sum... Unless the rates change in time in different ways, and you know something about this $\endgroup$ – Aksakal Mar 2 '16 at 16:57
1
$\begingroup$

Say, you observe $y_t$, where $y_t=x_t+z_t$, where $x_t\sim Poisson(\lambda^x_t)$ and $z_t\sim Poisson(\lambda^z_t)$. This means that $$y_t\sim Poisson(\lambda^y_t),$$ where $\lambda^y_t=\lambda^x_t+\lambda^z_t$

Let's say you can get $\hat\lambda^y_t$, how would you extract then $\hat\lambda^x_t,\hat\lambda^z_t$? There's no way unless you know something about the processes for $\lambda^x_t,\lambda^z_t$.

$\endgroup$

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.