$X $ follows $ Po(80)$.

I used a normal approximation to get $P(55\leq X\leq 75)$ which I got correct.

I need to find $P(X=80)$. I tried the Poisson distribution directly; however, my calculator shows an error when calculating $80^{80}$.

The correct answer is 0.0446.

Do we need to use normal distribution? Thanks.

  • $\begingroup$ I've added the self-study tag because your phrasing pretty clearly indicates it falls under self-study, but you should feel free to remove it (or to object and I will remove it). $\endgroup$ – Glen_b -Reinstate Monica Apr 28 '14 at 4:20

Here I discuss several approaches. If I wasn't near my computer (on which I'd just do dpois(80,80) in R), I'd probably use the first one below.

1) You can help your calculator out a bit by doing some algebra first, and trying Stirling's approximation for $80!$

You can try taking the series for the Stirling approximation out several terms but it won't be necessary in this case, just use the "ordinary" approximation ($n! \sim \sqrt{2 \pi n}\left(\frac{n}{e}\right)^n$), which hugely simplifies the calculation, and has about 3 significant figure accuracy in this region (which will be sufficient). [Taking one more term in the series - $n! \sim \sqrt{2 \pi n}\left(\frac{n}{e}\right)^n(1+\frac{1}{12n})$ - gives about six figure accuracy.]

So take the formula for the probability:

$$P(X=80) = \frac{\exp(-80) 80^{80}}{80!}$$

Then just replace 80! by the above approximation, and cancel out all those large terms, leaving a very simple calculation. (I bet you can do it from there.)

2) The "right" way to do that is to use a function (or a series, if you are working on a calculator) that returns a log-factorial (or log-gamma-function), and work on the log-scale until the final step ... which gives the "exact" answer of about 0.04455667. While quite accurate, this may involve a bit of work unless you're using a computer.

3) If you use the normal approximation, the usual way would be to use the continuity correction. Let $X$ be the original Poisson, and let $X^*$ be an approximating normal ($N(80,(\sqrt{80})^2$), and let $Z$ be a standard normal.

$$P(X=80)\approx P(79.5<X^*<80.5)$$ $$= P(\frac{79.5-80}{\sqrt{80}}<\frac{X^*-80}{\sqrt{80}}<\frac{80.5-80}{\sqrt{80}})$$ $$= P(\frac{-0.5}{\sqrt{80}}<Z<\frac{0.5}{\sqrt{80}})$$ $$= P(-0.055902<Z<0.055902)$$ $$= 2\,P(0<Z<0.055902)$$ $$= 2\times 0.02229$$ $$= 0.04458$$

... which is also sufficiently accurate.

4) Another way to do the approximation would be to approximate it by $f_{X^*}(80)\times 1\approx 0.0446$, but you have to be careful to justify that properly (refer to the diagrams and discussion at the link on continuity corrections I gave above).

  • 1
    $\begingroup$ +1 Even Excel has a log gamma function (GAMMALN) for such purposes, suggesting how useful (and basic) taking logarithms is. I'm sure GAMMALN is implemented by means of the asymptotic expansion of $\Gamma$ (generalizing Stirling's approximation). Concerning (4), it might be a little clearer to write your approximation as $f_{X^{*}}(80) \times 1$ or even, quite suggestively, $\int_{80-1/2}^{80+1/2}\left(f_{X^{*}}(80) + O(x-80)^2\right) dx.$ $\endgroup$ – whuber Apr 23 '14 at 14:44
  • $\begingroup$ @whuber I even typed $f_{X^*}(80)\times 1$ at one point but deleted it because I was worried it was perhaps confusing. I think you're right that it is clearer, since it suggests the sort of justification that I was talking about. The integral at the end there explains the whole thing rather neatly, though. $\endgroup$ – Glen_b -Reinstate Monica Apr 23 '14 at 14:52

You could get a better calculator or use logarithms for large powers: $$\log_{10}\left(80^{80}\right)=80\times \log_{10}\left(80\right)\approx 152.247199$$ $$80^{80} \approx 10^{0.247199} \times 10^{152} \approx 1.76685\times 10^{152}$$

If you used a normal approximation, you have two choices:

  • the density at $x=80$ of a normal distribution with mean $80$ and standard deviation $\sqrt{80}$, which is $\tfrac{1}{\sqrt{160 \pi}}$ or about $0.044603$
  • the probability of a value between $79.5$ and $80.5$ from a normal distribution with mean $80$ and standard deviation $\sqrt{80}$, which is about $0.044580$

The actual value from a Poisson distribution is closer to $0.044557$ so these approximations are not bad.


$\ln(Poisson_{80}(80))=\exp(\ln(Poisson_{80}(80)))=\exp(80 \ln 80-80-\ln(80!))$

For $\ln(80!)$ use Ramanujan's approximation.


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.