2
$\begingroup$

I am trying to find a tight upper bound the following expression \begin{align} \frac{ \Gamma \left( x; a\right)}{\Gamma(x)} \end{align} where $\Gamma \left( x; a\right)$ is incomplet Gamma function \begin{align} \Gamma \left( x; a\right)= \int_{a}^\infty t^{x-1} e^{-t} dt. \end{align}

Also, assume $x \ge 1/2$. Clearly a trivial upper bound is \begin{align} \frac{ \Gamma \left( x; a\right)}{\Gamma(x)} \le 1 \end{align}

However, I suspect the following bound should be true

\begin{align} \frac{ \Gamma \left( x; a\right)}{\Gamma(x)} \le O( e^{-a} ) \end{align}

$\endgroup$
  • $\begingroup$ You've given a lower bound on $x$. Do you have any bounds on $a$? $\endgroup$ – Eric Towers Apr 12 '16 at 14:20
  • $\begingroup$ @EricTowers I want a general bound that holds for any $a$ and $x$ without any assumption on $x$ and $a$. In fact, I would like to vary $a$ not $x$. $\endgroup$ – Boby Apr 12 '16 at 14:22
  • $\begingroup$ Then this is hopeless. For $x<0$, this ratio is unbounded. $\endgroup$ – Eric Towers Apr 12 '16 at 14:27
  • $\begingroup$ @EricTowers Still with assumption $x>1/2$. $\endgroup$ – Boby Apr 12 '16 at 14:29
2
$\begingroup$

$$\int_{a}^{+\infty}t^{x-1}e^{-t}\,dt = e^{-a}\int_{0}^{+\infty}(t+a)^{x-1}e^{-t}\,dt\leq e^{-a}a^{x-1}\int_{0}^{+\infty}\exp\left(\frac{t(x-1)}{a}-t\right)\,dt$$ hence a simple upper bound is: $$ \Gamma(x;a)\leq \frac{a^x e^{-a}}{a+1-x}.$$ That approximation can be refined, producing a continued fraction representation for the incomplete $\Gamma$ function: $$ \Gamma(x;a) = \frac{a^x e^{-a}}{a+\frac{1-x}{1+\frac{1}{a+\frac{2-x}{\ldots}}}}$$ Have also a look at this paper.

$\endgroup$
  • $\begingroup$ ?? that's not what is asked $\endgroup$ – reuns Apr 7 '16 at 16:37
  • $\begingroup$ If $a+1-x<0$ then the bound is negative, right? This doesn't seem to be correct. $\endgroup$ – Boby Apr 7 '16 at 16:43
  • $\begingroup$ Obviously it is an approximation that holds for large $a$s. $\endgroup$ – Jack D'Aurizio Apr 7 '16 at 16:47
1
$\begingroup$

Your final inequality seems to be false. For instance, $\dfrac{\Gamma(2;a)}{\Gamma(2)} = (a+1)\mathrm{e}^{-a}$ and $\dfrac{\Gamma(3;a)}{\Gamma(3)} = \dfrac{1}{2}(a^2 + 2a + 2)\mathrm{e}^{-a}$. The bound $\dfrac{\Gamma(x;a)}{\Gamma(x)} \leq O(a^{x-1}\mathrm{e}^{-a})$ seems possible, but I haven't taken the time to find a proof.

$\endgroup$
  • $\begingroup$ Thank you. In case you are curios, please do take the time. $\endgroup$ – Boby Apr 12 '16 at 14:28
  • $\begingroup$ For example how would you compute $\lim_{n \to \infty } \frac{\Gamma(n;n)}{\Gamma(n)}$ ? $\endgroup$ – Boby Apr 12 '16 at 17:23
  • 1
    $\begingroup$ @Boby: that limit is $\frac{1}{2}$ by a famous probabilistic argument: you are computing the probability that a limit random variable (fulfilling the hyphotesis of the central limit theorem) takes a value that is less than its expected value. $\endgroup$ – Jack D'Aurizio Apr 13 '16 at 15:56
  • $\begingroup$ @JackD'Aurizio Thank you. What aboutn $\lim_{n \to \infty} \frac{\Gamma(n,\lambda n)}{\Gamma(n)}$. Is it $0$ if $\lambda>1$ and $1$ if $\lambda<1$ ? $\endgroup$ – Boby Apr 13 '16 at 15:59
  • $\begingroup$ @Boby: yes, you're right. In such cases it is enough to use some concentration inequality for the Gamma distribution, whose mean and variance are known. You may also replace the Gamma distribution with a normal distribution and prove the same: the Central Limit Theorem allow us to do so. $\endgroup$ – Jack D'Aurizio Apr 13 '16 at 16:07

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.