next up previous
Next: About this document ...

A Simple Branching Process
The model.
Consider a system of A-particles or individuals subject to the Poisson-distributed random events
\begin{align*}\text{Death} & :\text{ }
\begin{array}[c]{ll}
A\rightarrow0\ \ \...
...gin{array}[c]{ll}
A\rightarrow2A & \text{rate }\beta
\end{array}
\end{align*}
The state of zero individuals is an absorbing state. If the system hits this state it stays there. The corresponding master equation is

 \begin{displaymath}\frac{\partial P_{n}\left( t\right) }{\partial t}=\beta\left(...
...t\right)
-\left( \lambda+\beta\right) nP_{n}\left( t\right)
\end{displaymath} (1)

                         Suppose we start with one individual at $\ t=0$. We'd like to know something about the sizes and durations of the resulting cascades. Interesting quantities.
1. Survival probability.

\begin{displaymath}P_{\text{S}}\left( t\right) =1-P_{0}\left( t\right) ,
\end{displaymath}

or the number of cascades lasting longer than time t, and in particular, its asymptotic value

\begin{displaymath}P_{\infty}=\lim_{t\rightarrow\infty}P_{\text{s}}\left( t\right)
\end{displaymath}

2. Cumulative size distribution
$n\left( s\right) $, the number of cascades larger than size s, or the probability that a cascade involves at least s individuals.


One finds critical behaviour when $\Delta\equiv\beta-\lambda=0$, i.e. all of these quantities behave as power laws
\begin{align*}P_{\text{S}}\left( t\right) & \sim t^{-\delta}\\
P_{\infty} & \sim\Delta^{\theta}\\
n\left( s\right) & \sim s^{-\alpha}
\end{align*}
Survival probability.
The survival probability can be found by using the generating function

\begin{displaymath}G\left( z,t\right) =\sum_{n=0}^{\infty}P_{n}\left( t\right) z^{n}
\end{displaymath} (2)

From the generating function, we can find the moments by taking derivative w.r.t z and then setting z=1.
For example:

\begin{displaymath}\left\langle n\right\rangle =\left. \frac{\partial G(z)}{dz}\right\vert _{z=1}
\end{displaymath}

and

\begin{displaymath}\left\langle n^{2}\right\rangle -\left\langle n\right\rangle =\left.
\frac{\partial^{2}G(z)}{dz^{2}}\right\vert _{z=1}
\end{displaymath}

often the moments

 \begin{displaymath}\phi_{m}=\left\langle n(n-1)(n-2)\cdot\cdot\cdot(n-m+1)\right\rangle =\left.
\frac{d^{m}G(z)}{dz^{m}}\right\vert _{z=1}
\end{displaymath} (3)

are refered to as `factorial' moments. From the factorial moments, one can alway get back the probabilities

\begin{displaymath}P_{n}=\frac{1}{n!}\sum_{m=n}^{\infty}\frac{\left( -1\right) ^{k}}{m!}
\phi_{m}
\end{displaymath}

We can use eqn (1) to obtain an equation for the generating function for this birth-death process. By multiplying by zn and summing over n, one finds

 \begin{displaymath}\frac{\partial G\left( z,t\right) }{\partial t}=\left( z-1\ri...
...mbda\right) \frac{\partial G\left( z,t\right) }{\partial
z}
\end{displaymath} (4)

This can be solved using the method of characteristics.
The method of characteristics.
Let $\left[ z\left( \tau\right) ,t\left( \tau\right) \right] $ be some curve in the $\left( z,t\right) $ plane. The change in G along this curve is

\begin{displaymath}\frac{dG}{d\tau}=\frac{\partial G}{\partial t}\frac{dt}{d\tau}+\frac{\partial
}{\partial z}\frac{dz}{d\tau}
\end{displaymath} (5)

We want to choose a curve such that

\begin{displaymath}\left( \frac{dz}{d\tau}\right) /\left( \frac{dt}{d\tau}\right) =-\left(
z-1\right) \left( \beta z-\lambda\right)
\end{displaymath}

Eqn. (4) implies $G\left( z,t\right) $ is constant along this characteristic. Integrating

\begin{displaymath}\frac{dz}{dt}=-\left( z-1\right) \left( \beta z-\lambda\right)
\end{displaymath}

we find

\begin{displaymath}C=\frac{z-1}{\beta z-\lambda}e^{\left( \beta-\lambda\right) t}
\end{displaymath}

where C is some constant of the curve $z\left( t\right) $. Then $G\left( z,t\right) $ must be some function of C.If we consider initial conditions such that at t=0, the number of individuals was exactly m, or $P_{n}\left( 0\right) =\delta_{n,m}$, we have

\begin{displaymath}G\left( z,0\right) =z^{m}
\end{displaymath}

One finds the solution

\begin{displaymath}G\left( z,t\right) =\left[ \frac{e^{\left( \lambda-\beta\righ...
...\beta z-\lambda\right) +\beta\left( 1-z\right)
}\right] ^{m}
\end{displaymath} (6)

This complicated looking expression tells us everything we need to know about this birth death process. Expanding in a power series, we can find $P_{n}\left( t\right) $. The survival probability is

\begin{displaymath}P_{\text{S}}\left( t\right) =\frac{\lambda-\beta}{e^{\left( \lambda
-\beta\right) t}\lambda-\beta}
\end{displaymath}

For $\beta<\lambda$,

\begin{displaymath}\text{\ }P_{\infty}=0
\end{displaymath}

the population becomes extinct. For $\lambda<\beta$
\begin{align*}P_{\infty} & =1-\frac{\lambda}{\beta}\\
& =\frac{\Delta}{\beta}
\end{align*}
For the critical case, $\lambda=\beta,$ one finds

 \begin{displaymath}P_{\text{S}}\left( t\right) =\frac{1}{1+\beta t}
\end{displaymath} (7)

Exercise. The result for the critical point can also be found by solving the (factorial) moment equations. Using eqn. (3) show that

\begin{displaymath}\frac{d\phi_{m}(t)}{dt}=\frac{\beta k!}{(k-2)!}n_{m-1}
\end{displaymath}

and that for 1-initial particle b.c , these equations can be recursively integrated to find

\begin{displaymath}\phi_{m}\left(t) =m!(\beta t)^{m-1}
\end{displaymath}

Digression:
An easier method
There's actually an easier way to get the duration results. For this method, we need conditional probabilities.
\begin{align*}P_{n/m}\left( t_{2},t_{1}\right) & \equiv\text{prob. of finding }n...
...at
time }t_{2}\text{ }\\
& \text{given }m\text{ at time }t_{1}.
\end{align*}
For Markov processes, one finds Kolmogorov's forward and backward equations.
Forward equation.

\begin{displaymath}\frac{dP_{n/m}(t_2,t_1)}{dt_2}=\sum_{n'}w_{n,n'}
P_{n'/m}(t_2,t_1) -
w_{n',n}P_{n/m}\left(t_2,t_1)
\end{displaymath}

or with t=t2-t1,

\begin{displaymath}\frac{dP_{n/m}\left( t\right) }{dt}=\sum_{n^{\prime}}w_{n,n^{...
...me}/m}\left( t\right) -w_{n^{\prime},n}P_{n/m}\left( t\right)
\end{displaymath}

Backward equation.

\begin{displaymath}\frac{dP_{n/m}\left( t_{2},t_{1}\right) }{dt_{1}}=-\sum_{m^{\...
...1}\right) +w_{m^{\prime}
,m}P_{n/m}\left( t_{2},t_{1}\right)
\end{displaymath}

or with t=t2-t1,

\begin{displaymath}\frac{dP_{n/m}\left( t\right) }{dt}=\sum_{m^{\prime}}w_{m^{\p...
...ime}}\left( t\right) -w_{m^{\prime},m}P_{n/m}\left(
t\right)
\end{displaymath}

Note that we can obtain the master equation by multiplying the both sides of the forward equation by Pm and summing over m, so that one only need worry about state probabilities at time t2 and not conditional probabilities. For this simple branching process, the backward equation for 1-initial individual boundary conditions is

\begin{displaymath}\frac{dP_{n,1}\left( t\right) }{dt}=-\left( \lambda+\beta\right)
P_{n/1}+\beta P_{n/2}+\lambda P_{n/0}
\end{displaymath}

Using the generating functions
\begin{align*}G_{1}\left( z\right) & =\sum_{n=0}^{\infty}P_{n/1}\left( t\right) ...
...\left( z\right) & =\sum_{n=0}^{\infty}P_{n/2}\left( t\right) z^{n}
\end{align*}
the backward equation becomes

\begin{displaymath}\frac{\partial G_{1}( z,t) }{\partial t}=-(\lambda
+\beta)G_{1}+\beta G_{2}(z,t)+\lambda,
\end{displaymath}

but since

\begin{displaymath}G_{m}\left( z,t\right) =\left[ G_{1}\left( z,t\right) \right] ^{m}
\end{displaymath}


\begin{displaymath}\frac{\partial G_{1}\left( z,t\right) }{\partial t}=-\left( \...
...\beta\right) G_{1}+\beta G_{1}^{2}\left( z,t\right) +\lambda,
\end{displaymath}

subject to the initial condition

\begin{displaymath}G_{1}\left( z,0\right) =z
\end{displaymath}

hence

\begin{displaymath}G_{1}\left( z,t\right) =\frac{e^{\left( \lambda-\beta\right) ...
...ht) t}\left( \beta z-\lambda\right) +\beta\left( 1-z\right) }
\end{displaymath}

Size distribution. In order to find the size distribution, we look at a generic tree.
\begin{figure}
\epsfysize=140pt
\epsffile{tree.ps}
\end{figure}

The probability of a tree with n leaves is
\begin{align*}P_{n} & =\left(
\begin{array}[c]{c}
\text{\char93  of trees}\\ 
...
...ight) ^{2}
}\right] ^{n}\left( \frac{\lambda+\beta}{\beta}\right)
\end{align*}
Using Stirling's formula

\begin{displaymath}P_{n}\simeq\frac{1}{2\sqrt{\pi}}n^{-3/2}\left[ \frac{4\lambda...
...^{2}}\right] ^{n}\left( \frac{\lambda+\beta}{2\beta
}\right)
\end{displaymath}

or

\begin{displaymath}n\left( s\right) \simeq\frac{1}{2\sqrt{\pi}}s^{-3/2}e^{-s/a}
\end{displaymath}

with the characteristic size

\begin{displaymath}a=-\frac{1}{\ln\left( \frac{4\lambda\beta}{\left( \lambda+\beta\right)
^{2}}\right) }
\end{displaymath}

So the cumulative size distribution

\begin{displaymath}n\left( s\right) \sim s^{-1/2}
\end{displaymath}

Exercise. Suppose you didn't know Cayley's formula. You can still calculate the size distribution by using a slightly modified model.

\begin{displaymath}\begin{array}{lll}
Death:&A\rightarrow R & rate\;\lambda\\
Birth:&A\rightarrow 2A &rate\; \beta\\
\end{array}\end{displaymath}

where R is a `removed' individual. Using a two species generating function

\begin{displaymath}G(y,z)=\sum_{n,l}P_{n,l}y^{l}z^{n}
\end{displaymath}

One can use either the backward or forward equation and the relevant method above to solve for the $t\rightarrow\infty,\;R$ probabilities.

\begin{displaymath}P_{l}=\sum_{n}P_{n,l}
\end{displaymath}



Click here for Return to title page
Click here for 7. Absorbing boundaries


 
next up previous
Next: About this document ...
Birger Bergersen
1998-10-03