3
$\begingroup$

This is a follow-up question of the question "When superposition of two renewal processes is another renewal process?".

How can we split a renewal process $P$ into a renewal process $P_1$ and another process $P_2$ (not necessarily renewal), where $P_1$ and $P_2$ can be independent or dependent to $P$?

If the rate (expected value of number of jumps in duration of time $t$) of $P$ is $\lambda$, we want to get the rate $\lambda'$ for $P_1$ for any $\lambda'<\lambda$.

My solution: Assume $x = [x_1 x_2 \dots x_n]$ is generated according to the renewal process $P$ in time interval $[0,t]$, where $x_i$ is the $i^{th}$ jump time that is generated according to the inter-renewal probability distribution, $p(x)$. I think if we randomly take $(\lambda-\lambda') t$ of the jumps and remove them, we obtain another renewal process with rate $\lambda'$ because: 1. since we choose the jumps randomly, the pdf of the jumps will remain identical 2. since each removing of a jump means replacing the two consequent jumps $x_i,x_{i+1}$ by their sum $x_i+x_{i+1}$ that is independent to other elements of $x$.

Is that correct? any idea what is the general answer?

$\endgroup$
  • $\begingroup$ What about using all the even renewals for $P_1$? $\endgroup$ – Michael Nov 24 '15 at 4:59
  • $\begingroup$ @Michael That sounds good. But I am not sure how is the proof. Can you see my solution above? $\endgroup$ – Susan_Math123 Nov 24 '15 at 5:00
  • $\begingroup$ Randomly choosing the renewal events to put to $P_1$ would also work, as long as every choice is made independently and with the same probability. $\endgroup$ – Michael Nov 24 '15 at 5:00
  • 1
    $\begingroup$ Your description of "randomly choosign $m<n$" is a bit ambiguous since a renewal process has an infinite number of renewals, and we do not know what $n$ means. $\endgroup$ – Michael Nov 24 '15 at 5:02
  • $\begingroup$ @Michael Can you show me the math? I want to find a solution which the number of removed jumps over the total jumps $\frac{m}{n}$ can be changed easily. $\endgroup$ – Susan_Math123 Nov 24 '15 at 5:03
3
$\begingroup$

The original renewal process has independent and identically disributed (i.i.d.) inter-renewal times $\{T_1, T_2, T_3, ...\}$. Thus, starting from time 0, renewals occur at times $\{T_1, T_1+T_2, T_1+T_2+T_3, ...\}$. Now fix a probability $p >0$. Independently place each renewal time to $P_1$ with prob $p$. So we get new inter-renewal times $\{Z_1, Z_2, Z_3, ...\}$, where $Z_1 = \sum_{i=1}^GT_i$ and $G$ is an independent geometric random variable with success probability $p$, and $Z_2, Z_3, ...$ are i.i.d. That is because, when you generate them, you generate them independently but using the same probability law (so each is just a random sum of i.i.d. $T_i$ variables).

If the rate of $P$ was $\lim_{t\rightarrow\infty} \frac{N(t)}{t}= \frac{1}{E[T_1]}=\lambda$ (with prob 1), the rate of $P_1$ is $\lim_{t\rightarrow\infty} \frac{N_1(t)}{t} = p\lambda$ (with prob 1).

Here I am of course defining: \begin{align} &N(t) = \mbox{ Total number of renewals from $P$ during $[0,t]$}\\ &N_1(t) = \mbox{ Total number of renewals from $P_1$ during $[0,t]$} \end{align}

A simple example is when $T_i=1/\lambda$ for all $i \in \{1, 2, 3, ...\}$, for a given constant $\lambda>0$. So inter-arrival times are constant (and hence trivially i.i.d.). Then $N(t)$ is a deterministic staircase function and $E[N(t)]=N(t)$ for all $t$, and indeed $$\lim_{t\rightarrow\infty} \frac{N(t)}{t} = \lim_{t\rightarrow\infty} \frac{E[N(t)]}{t} = \lambda $$ Probabilistically splitting this deterministic process $N(t)$ (of rate $\lambda$ arrivals/time) using a probability $p$ gives a random process $N_1(t)$ that has rate $p\lambda$ arrivals/time (and this new process indeed has i.i.d. inter-arrival times).


On visualizing the renewal times: I imagine renewal times as if they are things that arrive to a system, like job arrivals in a queueing system. So the original renewal process $P$ can be drawn over a timeline with spikes arising at the renewal times (the times $\{T_1, T_1+T_2, T_1+T_2+T_3, ...\}$). Then $N(t)$ counts the spikes up to time $t$. We can "probabilistically thin" this spikey process by independently including spikes with prob $p$, and throwing the others away. The thinned process $N_1(t)$ counts the number of included spikes, and so $N_1(t)\leq N(t)$ for all $t$.


As an interesting side note: Consider $\{T_1, T_2, T_3, ...\}$ as any random sequence of inter-arrival times (not necessarily i.i.d.) and let $N(t)$ count the number of arrivals up to time $t$. Now probabilistically thin this process to a new process $N_1(t)$ by independently including each arrival with probability $p$. Then $E[N_1(t)] = pE[N(t)]$ for all $t\geq 0$ since: $$ E[N_1(t)] = E[E[N_1(t)|N(t)]] = E[pN(t)] = pE[N(t)] $$ For example, if $E[N(t)]=\lambda t$ for all $t\geq 0$, then $E[N_1(t)]=p\lambda t$ for all $t\geq 0$. An example of such a process $N(t)$ that is not a Poisson process is this: Choose $T_1$ uniformly over $[0,1]$, then define $T_i=1$ for all $i\in\{2,3,4,...\}$.

$\endgroup$
  • $\begingroup$ One really interesting question, likewise here, if we add spikes to $P$ randomly and independently, for example according to Poisson with rate $\lambda'$, will we obtain another renewal process with rate $ \lambda + \lambda'$? $\endgroup$ – Susan_Math123 Nov 24 '15 at 15:34
  • $\begingroup$ To answer my question, the only way that it can happen is when $P$ is also Poisson. How about for a general $P$? How can we add another process with rate $\lambda'$ to obtain a process with rate $\lambda + \lambda'$? $\endgroup$ – Susan_Math123 Nov 24 '15 at 15:44
  • $\begingroup$ If you have two spikey processes, and the counting process for each is $A_1(t)$ and $A_2(t)$, and if $\lim_{t\rightarrow\infty} A_1(t)/t =\lambda_1$ and $\lim_{t\rightarrow\infty} A_2(t)/t = \lambda_2$, then super-imposing them gives $\lim_{t\rightarrow\infty} (A_1(t)+A_2(t))/t = \lambda_1 + \lambda_2$. Now, the sum process $A_1(t) + A_2(t)$ may not be a renewal process anymore, even if both $A_1(t)$ and $A_2(t)$ are independent renewal processes and even if one of them is Poisson. Independent Poisson processes are special since adding them gives a new Poisson process with the sum rate. $\endgroup$ – Michael Nov 24 '15 at 19:32
  • $\begingroup$ Thanks for your comment. I see your point. My question is: if $A_1(t)$ is renewal, then how can we build $A_2(t)$ with parameter $\lambda_2$ such that $A_1(t)+A_2(t)$ is renewal? $\endgroup$ – Susan_Math123 Nov 24 '15 at 19:35
  • $\begingroup$ Another way to scale the rate of a renewal process is this: Suppose $N(t)$ is a renewal process with rate $\lambda$. Fix $r>0$ and define $A(t) = N(rt)$. Then $A(t)$ is also a renewal process, and it has rate $r\lambda$. $\endgroup$ – Michael Nov 24 '15 at 19:38

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.