9
$\begingroup$

Though the question here (Partial sums of exponential series - Stack Exchange) is similar, it is more specialized and I rather need a general approximation for an arbitrary partial sum.

Essentially, I am trying to approximate the probability mass function of a particular random variable and I ended up with a Poisson random variable's CDF in the mix. Hence, for my purpose, I need to figure out a reasonable approximation of the sum:

$\displaystyle\sum_{k = 0}^{r} \frac{\lambda^k}{k!}$ OR the tail, i.e. $\displaystyle\sum_{k = r}^{\infty} \frac{\lambda^k}{k!}$

Does someone know some approximations for this? Also, if there exist conditions for those approximations to be valid, I'd like to know them as well.

Thanks in advance!

Addendum: There appears to be a closed form expression for such a partial sum:

$\displaystyle\sum_{k = 0}^{r} \frac{\lambda^k}{k!} = e^\lambda \frac{\Gamma(r + 1, \lambda)}{\Gamma(r + 1)}$,

where $\Gamma(a, x)$ is defined as: $\displaystyle \Gamma(a, x) = \int_x^\infty t^{a - 1} e^{-t} \,dt$ and $\displaystyle \Gamma(a) = \Gamma(a, 0)$.

Is there a simple closed form approximation for the Gamma function? At the end of the day, somehow or the other, I either end up with a summation sign or an integral. I just want to be able to pin down this partial sum as a numeric quantity, that is reasonably approximate.

$\endgroup$
4
  • 2
    $\begingroup$ What language/environment are you using for computations? Chances are someone has already implemented the Gamma function more efficiently and better than you could in a few hours. Unless the approximation is central to the research you're attempting to perform, use a library. For instance, in C++: boost.org/doc/libs/1_42_0/libs/math/doc/sf_and_dist/html/… $\endgroup$
    – dls
    Jul 15, 2012 at 21:28
  • $\begingroup$ The idea is to not use exact values but come up with approximations that are suitable for analytical interpretation. I am attempting to mathematically model a system that has been implemented in real life. In order for the model to be of value, I need to make it as simple as possible. Therefore, to reduce complexity, I am using the mean values rather than the whole PMF, justified by ideas such as mean dynamics to model the dynamics of the system. To be clear, I am asking for an analytical approximation of aforementioned partial sum. Thanks for the response however (May help me in the future) $\endgroup$ Jul 15, 2012 at 21:39
  • $\begingroup$ Gotcha, I misinterpreted your phrase 'numeric quantity'. $\endgroup$
    – dls
    Jul 15, 2012 at 23:07
  • 1
    $\begingroup$ How large are your $r$ and $\lambda$ ? $\endgroup$
    – user65203
    Dec 14, 2015 at 14:17

2 Answers 2

14
$\begingroup$

Use Taylor's series with remainder. We know that $$ e^\lambda = \sum_{k=0}^r \frac{\lambda^k}{k!} + \frac{e^{c\lambda}\lambda^{r+1}}{(r+1)!}, $$ for some $c \in [0,1]$. Therefore $$ \frac{\lambda^{r+1}}{(r+1)!} \leq e^\lambda - \sum_{k=0}^r \frac{\lambda^k}{k!} \leq e^\lambda \frac{\lambda^{r+1}}{(r+1)!}. $$ You can also get these estimates using more elementary means: $$ \sum_{k=r+1}^\infty \frac{\lambda^k}{k!} = \frac{\lambda^{r+1}}{(r+1)!} \left[ 1 + \frac{\lambda}{r+2} + \frac{\lambda^2}{(r+2)(r+3)} + \cdots \right] < \frac{\lambda^{r+1}}{(r+1)!} \sum_{t=0}^\infty \frac{\lambda^t}{t!} = \frac{\lambda^{r+1}}{(r+1)!} e^\lambda. $$ We can get a different upper bound by comparison to a geometric series, when $\lambda < r+2$: $$ \sum_{k=r+1}^\infty \frac{\lambda^k}{k!} \leq \frac{\lambda^{r+1}}{(r+1)!} \sum_{t=0}^\infty \left(\frac{\lambda}{r+2}\right)^t = \frac{\lambda^{r+1}}{(r+1)!} \frac{r+2}{r+2-\lambda}. $$

$\endgroup$
6
  • $\begingroup$ Thank you for giving me an approximation AND a starting point to think in the right direction. Hopefully this approximation in itself is good enough for my purposes. At least, now I know which direction to look for answers. Thanks again! $\endgroup$ Jul 15, 2012 at 21:42
  • 1
    $\begingroup$ Edit: The upper bound didn't seem to be all that tight! Lower bound was better, but still needed tweaking. I then realized I could use the quantity $c$ in $\displaystyle e^{c\lambda}$ as a parameter that can be used for fitting the analytically computed mean value (obtained from the accurate PMF). It seems to work perfectly for a well chosen, small value of $c$. I am quite glad that now, my analysis seems systematic and not ad-hoc as it was earlier! Thanks for the help by the way. $\endgroup$ Jul 15, 2012 at 22:38
  • $\begingroup$ How do you get the reminder of the exponential Taylor series approximation? You write $$e^{c\lambda} \frac{ \lambda^{r+1} }{(r+1)!}$$ which can be no other than the reminder term. How do you get to that? $\endgroup$
    – Dr Potato
    Apr 22, 2020 at 20:43
  • $\begingroup$ thanks for the answer! Is there any meaningful way to pick c as a function of r? That is, can one justify for example the choice c=1/sqrt(r)? Otherwise the bound becomes very lose. Thanks :) $\endgroup$
    – Jonasson
    Nov 19, 2020 at 14:12
  • $\begingroup$ Taylor's theorem only guarantees the existence of some $c$. If you want more accuracy, take more terms in the series. $\endgroup$ Nov 19, 2020 at 14:13
1
$\begingroup$

matlab has a built-in function gammainc($\lambda, a$, 'upper') = ${\Gamma(a, \lambda)\over \Gamma(\lambda)}$.

Also see "A computational procedure for incomplete Gamma functions" by Walter Gautschi.

$\endgroup$

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.