Google Groups no longer supports new Usenet posts or subscriptions. Historical content remains viewable.
Dismiss

combining Gamma distributions with different k parameters

25 views
Skip to first unread message

FangQ

unread,
Dec 14, 2009, 1:43:52 PM12/14/09
to
hi

My research recently involved the modeling of photon migration in
highly scattering random media. I need some input to understand the
path-lengths of the photons (particles under random walk) exiting from
the domain.

Photon scattering length inside a random medium is modeled by a random
variable (RV) of exponential distribution. As a consequence, I believe
that when the photon exits the domain, the total path-length is a sum
of all exponential RVs. If there are N scattering events before the
exit, the path-length is supposed to be a RV with Gamma (or Erlang)
distribution with k=N. However, the photon does not have to scatter N
times, N can be any number from 0 to inf, so, that means the path-
length of a exit photon should be a "combined distribution" of all
different Gammas with k from 0~inf. I am not entirely sure about the
function form of k(n).

Here are my questions:

1. what kind form do you expect if one want to "combine" these Gamma
distributions with different k parameters? do you think there will be
an analytical form for that? is integrate(Gamma(k(n),theta),n,0,inf) a
reasonable form? or it should be convolution type?

2. I know that adding two RVs with Gamma(k1,theta) and Gamma(k2,theta)
is a Gamma(k1+k2,theta), will this simplify the integrate(Gamma(k
(n),theta),n,inf) to Gamma(integrate(k(n),n,0,inf),theta)?

3. is there a known distribution to describe how many scattering
events happened before a random particle exit a domain at a given
point? (which is k(n) as we mentioned above)

I am great appreciated if someone can point me to the references that
are related to this problem. Any hint or suggestion is also welcome!

thanks

Qianqian

Ray Koopman

unread,
Dec 14, 2009, 5:52:48 PM12/14/09
to

That sounds like N should be treated as a random variable.
Then fX(x) = sum_n fXn(x|n) * fN(n).
Is there theory that suggests a particular distribution for N?

Ray Koopman

unread,
Dec 14, 2009, 5:59:55 PM12/14/09
to

Make that fX(x) = c * sum_n fXn(x|n) * fN(n),
where c is chosen so that fX(x) integrates to 1 over (0,infinity).

Ray Koopman

unread,
Dec 14, 2009, 6:11:11 PM12/14/09
to
On Dec 14, 2:59 pm, Ray Koopman <koop...@sfu.ca> wrote:
> [...]

> Make that fX(x) = c * sum_n fXn(x|n) * fN(n),
> where c is chosen so that fX(x) integrates to 1 over (0,infinity).

If every fXn integrates to 1, and fN sums to 1, then c = 1
and drops out. Looks like it's one of those days :(

FangQ

unread,
Dec 15, 2009, 4:02:51 PM12/15/09
to
On Dec 14, 5:52 pm, Ray Koopman <koop...@sfu.ca> wrote:
> That sounds like N should be treated as a random variable.
> Then fX(x) = sum_n fXn(x|n) * fN(n).
> Is there theory that suggests a particular distribution for N?


thanks for your comments on my questions.

I think (not entirely sure) that the distribution of N is a Poisson
distribution with a mean of x*mus, i.e. fN(n)=(x*mus)^n*exp(-x*mus)/
(n!), where mus is the scattering coefficient of the medium (constant,
it has a unit of 1/length, denoting the probability of hitting a
scattering event per unit length, x is the total path length in the
medium).

Now I don't know if this complicates the situation: fN(n) is now a
function of x; but it does look promising. I will do more derivations
from here.

danh...@yahoo.com

unread,
Dec 15, 2009, 6:07:12 PM12/15/09
to

The answer to part 1 is yes, there is an analytic form that gives a
closed form expression. A random sum of exponentials (which is what
you have) is an exponential when the number of summands
has a geometric distribution. I.e., the probability that there are n
summands has probability
(1-p)p^(n-1), n=1,2,3,... for 0<p<1.

Ray Koopman

unread,
Dec 15, 2009, 6:17:17 PM12/15/09
to

If X|n ~ Gamma[n+1] and N ~ Poisson[m] then Mathematica gets

fX[x,m] = Exp[-(m+x)] * Hypergeometric0F1Regularized[1,m*x],

for which the mean is m+1. (Note that I've taken the scale parameter
-- your 'theta' -- to be 1, so everything will need to be scaled.)

Note also that the number of path segments is n+1, not n. Using just
n would mean that a pathlength of 0 has a nonzero probability.

0 new messages