2
$\begingroup$

Does any one know approximations for the lower incomplete gamma function $\gamma(a,bx)$. The problem is this: I want to find the quantile function for the CDF of the gamma distribution. The CDF of the gamma distribution is $$F(x)=\frac{\gamma(a,bx)}{\Gamma(a)}=\frac{b^a}{\Gamma(a)}\sum_{m=0}^{\infty} \frac{(-b)^{m}}{m!} \frac{x^{m+a}}{m+a}.$$

The problem is to find $x$ (the quantile function) such that $F(x)= \epsilon$, $\epsilon$ is a very small value. For a small value of $F(x)$, $x$ is also very small.

$\underline{\mathrm{My~try}}:$

Since $x$ is small, I approximate $F(x)$ using a single term in the above summation, that is,

$$ F(x) \approx \frac{b^a}{\Gamma(a)} \frac{x^{a}}{a}=\epsilon,$$

which gives $x \approx \frac{1}{b} (\epsilon~\Gamma(a)~a)^{\frac{1}{a}}$. Are there any other better approximations than this? Please help.

$\endgroup$
2
$\begingroup$

If I correctly understand, you want an approximation of $x$ such that $F(x)=\epsilon$ and you consider that $x_0=\frac{1}{b} (\epsilon~\Gamma(a)~a)^{\frac{1}{a}}$ is a legitimate approximation when $\epsilon$ is small.

So, to refine the approximation, you could use one step of Newton method writing $$x_1=x_0-\frac{F(x_0)-\epsilon}{F'(x_0)}$$ This should give you $$x_{n+1}=x_n+x_n e^{b x_n}(b x_n)^{-a} (\Gamma (a,b x_n)+(\epsilon -1) \Gamma (a))$$

Another solution could be to use the first two terms and write $$\frac{x^a}{a}=\frac{\epsilon \Gamma(a)}{b^a}+\frac{b x^{a+1}}{a+1}$$ that is to say $$x_{n+1}=\Big(\frac{a\epsilon \Gamma(a)}{b^a}+\frac{ab x_n^{a+1}}{a+1}\Big)^{\frac{1}{a}}$$ If you start with $x=0$ in the rhs, you have your approximation $x_0$. Then, one interation gives $$x_1=\left(a \left(\epsilon b^{-a} \Gamma (a)+\frac{b \left(\left(a \epsilon b^{-a} \Gamma (a)\right)^{\frac{1}{a}}\right)^{a+1}}{a+1}\right)\right)^{\frac{1}{a}}$$

$\endgroup$
  • $\begingroup$ Thank you for this beautiful insight Dr. Claude. May I ask you couple of more questions? 1. I am not clear how do you arrive at the recursive equation $x_{n+1}$, from the first two terms? Pardon me for my ignorance. 2. Your approximation is really helpful to me. For simplicity, can I say $x _{1} \approx \alpha x_{0},$ where $\alpha=1,2 or 3$. Thanks in advance. $\endgroup$ – Oliver Aug 2 '14 at 21:54
  • $\begingroup$ In the second part, I used the first two terms of the expansion. I rewrote putting the first in the lhs a,d the second in the rhs in order to arrive to something looking as a recurrence relation. $\endgroup$ – Claude Leibovici Aug 3 '14 at 2:46
  • $\begingroup$ Thank you for your reply. I am still not convinced with the recurrence relation.In the two term approximation, the $x$ assumes the same value on both sides. How can we take the $x$ (on the L.H.S) as $x_{n+1}$ and the $x$ (on the R.H.S) as $x_{n}$? I am clueless. Thanks. $\endgroup$ – Oliver Aug 3 '14 at 12:44
  • $\begingroup$ I like the second way of approximation, because that does not involve the incomplete gamma functions! $\endgroup$ – Oliver Aug 3 '14 at 12:52

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.