7
$\begingroup$

What is the relationship between a Nonhomogeneous Poisson process and a process that has heavy tail distribution for its inter arrival times?

Any pointer to a resource that can shed some light on this question would be hugely appreciated

$\endgroup$
4
$\begingroup$

Well, if you have a point process that you try modeling as a Poisson process, and find it has heavy tails, there are several possibilities. What are the key assumptions for a Poisson Process:

-There is a constant rate function -Events are memoryless, that is P(E in (t,t+d)) is independent of t and when other events are. -The waiting time until the next event is exponentially distributed (kinda what the previous two are saying)

So, how can you violate these assumptions to get heavy tails?

-Non-constant rate function. If the rate function switches between, say, two values, you'll have too many short wait-times, and too many long wait-times, given the overall rate function. This can show itself as having heavy tails. -The waiting time is not exponentially distributed. In which case, you don't have a Poisson process. You have some other sort of point process.

Note that in the extreme case, any point process can be modeled by a NHPP - put a delta function at each event, and set the rate to 0 elsewhere. I think we can all agree that this is a poor model, having little predictive power. So if you are interested in a NHPP, you'll want to think a bit about whether that is the right model, or whether you are overly-adjusting a model to fit your data.

$\endgroup$
1
$\begingroup$

In general there is none. A Poisson process has inter-arrival times that are exponentially distributed, which does not have heavy tails.

$\endgroup$
  • $\begingroup$ That is only true for homogeneous Poisson Processes. If you have a Non-homogeneous Poisson process (that is not equivalent to some homogeneous Poisson process), then the inter-arrival times will not be exponentially distributed. $\endgroup$ – mpacer Jan 13 '15 at 1:00
1
$\begingroup$

The assumption that

Heavy tail distribution for its inter arrival times

along with the assumption that each inter-arrival time is independent, you have a renewal process. There are different ways that a renewal process can have a inhomogeneous marginal rate. Most frequently this is done by stretch time (time-rescaling theorem), or scale the hazard function (multiplication of the rate and the hazard function at each time point). Poisson process is a special case where the inter-arrival time is exponential as user549 answered. However, making inhomogeneous Possion with either method results in the same inter-spike interval distribution.

In summary, inhomogeneous Poisson process cannot have heavy tails.

$\endgroup$

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.