3
$\begingroup$

I have a two-factor shot noise process as follows (sorry for the picture of text): enter image description here

I want to simulate from this process. Is it valid to do so in a discretized way? The solution to these SDEs is a compound Poisson process, so I simulated it as follows:

shot_noise <- function(T, n, alpha, lambda, gamma, x0){
  dt = T/n
  dN = rpois(n, lambda*dt)
  x = c(x0)
  j = rep(0, n)
  for (i in 2:(n+1)){
    jumps = sum(rexp(dN[i-1], rate = gamma)*exp(-alpha * runif(dN[i-1], max = dt)))
    j[i-1] = jumps
    x[i] = x[i-1] - alpha*x[i-1] + jumps
  }
  return(x)
}

This would be the code to simulate x_+ and x_-, given the parameters. It seems correct to me, and the output looks like what I think it should. However, the other examples of simulating compound Poisson processes that I see tend not to discretize. I am doing so because this simulation will be used to test whether maximum likelihood estimation for this model is valid, and the observed data is discretized in that case ($\epsilon_t$ at each time-step).

Is the method I've used valid and correct? Is there a better way?

$\endgroup$
5
  • $\begingroup$ Your simulation seems nearly good, but I think that the "skeleton part" of the transition is $x(t_i) = e^{-\alpha \Delta} x(t_{i-1})$ where $\alpha$ (an inverse time) is $a$ in equations, and $\Delta$ stands for the time interval dt in code. Your question could be much simplified by using only one process $x(t)$ (as in the simulation) and possibly the notation $x(t) = \sum_{\{i: \,S_i \leq t\}} \eta_i$ where the $S_i$ are the Poisson arrivals. Note that $[x(t)]_t$ is continuous-time Markov, and $[x(t_i)]_i$ is discrete-time Markov both with gamma margin. $\endgroup$
    – Yves
    Commented Jul 6, 2017 at 11:35
  • $\begingroup$ Sorry I missed the impluse in my formula $x(t) = \sum_{\{i: \,S_i \leq t\}} e^{-\alpha (t- S_i)} \eta_i$. $\endgroup$
    – Yves
    Commented Jul 6, 2017 at 11:44
  • $\begingroup$ @Yves Thank you! I modified that line in the code to: x[i] = x[i-1]*exp(-alpha*dt) + jumps I should have also mentioned, in my case, the jump times are latent, which is where the uniform RV appears (using EM algorithm) $\endgroup$ Commented Jul 6, 2017 at 14:17
  • $\begingroup$ Conditional on their number $n_{i-1} := N(t_i) - N(t_{i-1})$, the successive arrivals $S_k$ falling in the interval $(t_{i-1},\,t_i)$ are such the r.vs $S_k -t_{i-1}$ have the same distribution as order statistics of a sample of the the uniform distribution on $(0,\,\Delta)$. This is a classical property of the PP. Note that your code works when $n_{i-1} =0$ beacuse sum(numeric(0)) == 0. $\endgroup$
    – Yves
    Commented Jul 6, 2017 at 14:55
  • $\begingroup$ I appreciate your exposition. $\endgroup$ Commented Jul 17, 2017 at 22:40

1 Answer 1

1
$\begingroup$

Let us derive the simulation. It will be enough to consider one of the two process $x^{+}_t$ and $x^{-}_t$, that we can denote by $x(t)$. It has the representation $$ x(t) = \sum_{\{k: \, T_k \leq t\}} e^{- \alpha [t - T_k ]} \eta_k $$ where $T_k$ are the arrivals of the Poisson Process. For $\Delta > 0$ we have $$ x(t + \Delta) = \sum_{\{k: \, T_k \leq t\}} e^{- \alpha [t + \Delta - T_k ]} \eta_k + \sum_{\{k: \, t < T_k \leq t + \Delta\}} e^{- \alpha [t + \Delta - T_k ]} \eta_k = e^{-\alpha \Delta} x(t) + \varepsilon . $$

In the second sum denoted as $\varepsilon$, the arrivals $T_k$ and the jumps $\eta_k$ are independent of the history $\{x(s);\, s \leq t \}$, and so is $\varepsilon$. So $x(t)$ is a Markov process. Moreover, conditional on the number of arrivals on $(t,\,t+ \Delta)$, say $J$, we know that the r.vs $T_k - t$ have the distribution of the order statistics of a sample of size $J$ of the uniform distribution on $(0,\,\Delta)$. So

\begin{equation} \varepsilon = \sum_{j=1}^J e^{-\alpha U_j} \eta_j' \tag{1} \end{equation}

where the $\eta_j'$ are i.i.d. from the distribution of $\eta_k$ and the $U_j$ are uniform on $(0,\,\Delta)$. The sum is understood as $0$ when $J =0$.

Therefore the simulation of $x(t + \Delta)$ conditional on $\{x(s);\, s \leq t \}$ can be as follows

  • Draw the number of jumps $J$ on $(t,\,t+ \Delta)$ from the Poisson distribution with mean $\lambda \Delta$.

  • Draw $J$ i.i.d r.vs $U_j$ from the uniform on $(0,\,\Delta)$ and $J$ i.i.d r.vs $\eta_j'$ from the jump distribution.

  • Take $x(t + \Delta) = e^{-\alpha \Delta} x(t) + \varepsilon$ where $\varepsilon$ is obtained by (1).

If $t_i^\star$ is an increasing sequence of observation times which is independent of the process $x(t)$, then we can simulate the sequence $x(t_i^\star)$ by simulating $x(t_i^\star)$ conditional on the history at $t = t_{i-1}^\star$ for $i > 0$, as in the code of the question. The sequence $x(t_i^\star)$ is a Markov chain, and it is time-homogeneous if $t_i^\star-t_{i-1}^\star$ is constant.

Interestingly, it can be shown that when the jump distribution is exponential with scale $1/ \gamma$, the stationary distribution of $x(t)$ is gamma with shape $\lambda / \alpha$ and scale $1 / \gamma$. This distribution can be used to draw the initial value $x(t_0^\star)$, and we have an example of Markov chain with gamma margin as in this question. A classical reference for shot noise is section 5.6 in D.R Cox and V. Isham Point Processes, Chapman & Hall (1980).

$\endgroup$
2
  • $\begingroup$ Thanks @Yves. Seems like that's more or less what I did, with the suggestion you made earlier. Much appreciated. $\endgroup$ Commented Jul 17, 2017 at 22:43
  • $\begingroup$ @ArvindShrivats Yes I just detailed the algorithm that you implemented, up to the autoregression coefficient. $\endgroup$
    – Yves
    Commented Jul 21, 2017 at 10:17

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.