Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

I am analyzing one of the properties of the Poisson Process, that of Competing Processes:

Suppose $N_1(t), t \geq 0$ and $N_2(t), t \geq 0$ are independente Poisson processes with respective rates $\lambda_1$ and $\lambda_2$. Let $S_n^i$ be the time of the $n$th event of process $i, i = 1,2$.

We know that $P[S_n^1 < S_m^2] = \sum_{k=n}^{n+m-1} {n+m-1 \choose k} \left(\frac{\lambda_1}{\lambda_1 + \lambda_2}\right)^k \left(\frac{\lambda_2}{\lambda_1 + \lambda_2}\right)^{n+m-1-k}$. I'm interested in the case where $n=m$ and $ n \rightarrow \infty$. This converges to 1, and my intuition is that the rate of convergence depends on the ratio $\frac{\lambda_1}{\lambda_2}$, but I'm having a hard time proving it out. Any insights into a closed-form rate of convergence would be great.

share|improve this question

1 Answer 1

up vote 2 down vote accepted

First write $$ S_n^1 = X_1^1 + X_2^1 + X_3^1 + \cdots , $$ where the $X_k^1$ are i.i.d. exponential random variables with density function $f_{X_1^1 } (x) = \lambda _1 e^{ - \lambda _1 x} $, $x > 0$, and $$ S_n^2 = X_1^2 + X_2^2 + X_3^2 + \cdots , $$ where the $X_k^2$ are i.i.d. exponential random variables with density function $f_{X_1^2 } (x) = \lambda _2 e^{ - \lambda _2 x} $, $x > 0$, independent also of the $X_k^1$. Now let $Y_k = X_k^1 - X_k^2$. Then, $$ {\rm P}(S_n^1 < S_n^2 ) = {\rm P}(Y_1 + \cdots + Y_n < 0). $$ Next note that the $Y_k$ are i.i.d. random variables with mean $$ \mu = {\rm E}(Y_1) = {\rm E}(X_1^1 ) - {\rm E}(X_1^2 ) = \frac{1}{{\lambda _1 }} - \frac{1}{{\lambda _2 }} $$ and variance $$ \sigma^2 = {\rm Var}(Y_1) = {\rm Var}(X_1^1 ) + {\rm Var}(X_1^2 ) = \frac{1}{{\lambda _1^2 }} + \frac{1}{{\lambda _2^2 }}. $$ It holds $$ \frac{\mu }{\sigma } = \frac{{\lambda _2 - \lambda _1 }}{{\lambda _1 \lambda _2 }}\frac{{\lambda _1 \lambda _2 }}{{\sqrt {\lambda _1^2 + \lambda _2^2 } }} = \frac{{\lambda _2 - \lambda _1 }}{{\sqrt {\lambda _1^2 + \lambda _2^2 } }}. $$ Now let $\tilde S_n = Y_1 + \cdots + Y_n$ and $Z_n = \frac{{\tilde S_n - n\mu }}{{\sigma \sqrt n }}$. Then, $$ {\rm P}(Y_1 + \cdots + Y_n < 0) = {\rm P}\bigg(\frac{{\tilde S_n - n\mu }}{{\sigma \sqrt n }} < \frac{{ - n\mu }}{{\sigma \sqrt n }}\bigg) = {\rm P}\bigg(Z_n < \frac{{\lambda _1 - \lambda _2 }}{{\sqrt {\lambda _1^2 + \lambda _1^2 } }}\sqrt n \bigg). $$ By the central limit theorem, $Z_n$ converges in distribution to the ${\rm N}(0,1)$ distribution. (So if $\lambda_1 > \lambda_2$, the last probability, which is equal to ${\rm P}(S_n^1 < S_n^2 )$, tends to $1$ as $n \to \infty$.) This may help you deduce the asymptotic behavior of ${\rm P}(S_n^1 < S_n^2 )$.

EDIT:

So the above gives rise to the approximation $$ {\rm P}(S_n^1 < S_n^2 ) \approx \Phi \bigg(\frac{{\lambda _1 - \lambda _2 }}{{\sqrt {\lambda _1^2 + \lambda _2^2} }}\sqrt n \bigg), $$ where $\Phi$ is the distribution function of the ${\rm N}(0,1)$ distribution. Of course, the quality of the approximation depends on the parameters involved. We can check it by comparing to results obtained from Monte Carlo simulations. [The probability ${\rm P}(S_n^1 < S_n^2 )$ can be approximated easily and accurately using Monte Carlo simulation, noting that an exponential random variable with density function $f_{X} (x) = \lambda e^{ - \lambda x} $, $x > 0$, can be generated as $-\ln(U)/\lambda$, where $U$ is uniformly distributed on $(0,1)$.] Define the quantity $\zeta = \zeta (n,\lambda_1,\lambda_2)$ by $$ \zeta = \frac{{\lambda _1 - \lambda _2 }}{{\sqrt {\lambda _1^2 + \lambda _2^2 } }}\sqrt n , $$ and let $\hat p = \hat p(n,\lambda_1,\lambda_2)$ denote the Monte Carlo approximation obtained (accurately enough) for the probability ${\rm P}(S_n^1 < S_n^2 )$. The following results were obtained: 1) $n=50$, $\lambda_1=2$, $\lambda_2 = 1$ ($\zeta = \sqrt{10}$): $\hat p = 0.999678$, $\Phi(\zeta) \approx 0.999217$; 2) $n=100$, $\lambda_1=4$, $\lambda_2 = 3$ ($\zeta = 2$): $\hat p = 0.978802$, $\Phi(\zeta) \approx 0.977250$. Finally, letting $\rho = \lambda_1 / \lambda_2$, it holds $$ \frac{{\lambda _1 - \lambda _2 }}{{\sqrt {\lambda _1^2 + \lambda _2^2 } }} = \frac{{\rho - 1}}{{\sqrt {\rho^2 + 1} }}, $$ and thus the original approximation can be written $$ {\rm P}(S_n^1 < S_n^2 ) \approx \Phi \bigg(\frac{{\rho - 1 }}{{\sqrt {\rho^2 + 1} }}\sqrt n \bigg). $$

share|improve this answer
    
@Avneesh: So the last approximation agrees with your intuition... –  Shai Covo Jun 3 '11 at 4:29
    
This is excellent stuff, thanks so much Shai! Sorry I can't vote your answer up, seems like I need more reputation points. –  Avneesh Jun 3 '11 at 22:00
    
@Avneesh: Thanks, glad to help! –  Shai Covo Jun 5 '11 at 4:10
    
@Avneesh: You can formally accept Shai's answer, though, by clicking on the little check mark next to it. –  Mike Spivey Jun 6 '11 at 22:21
    
@Mike: Thanks for pointing this out! –  Shai Covo Jun 6 '11 at 23:01

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.