Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

Let $a >0, b >0$, and $r \in \mathbb{R}$. I am trying to find a lower bound for the integral $$\int_a^\infty y^{-r} \exp\left( - b(y-a)^2\right) \,\mathrm dy.$$ After consulting the Wikipedia page for the incomplete gamma function and a couple of other websites I still have not been able to find a reference for a lower bound.

Do there exist a standard or well-known lower bounds for the incomplete gamma function?

share|improve this question
    
Lower or upper? –  Pedro Tamaroff Apr 8 '12 at 1:03
    
Upper notice $a >0$ is the type of integral I am interested in –  user7980 Apr 8 '12 at 1:08
    
Maybe this can help. –  Pedro Tamaroff Apr 8 '12 at 1:17
    
Not sure if it'll help, but in this answer I proved that the quantity $$\Bigl(e^a/(e^a-1)\Bigr)^x \,\Bigl(\Gamma(x)-\Gamma(x,a)\Bigr)$$ is nondecreasing in $a$ when $a > 0$ for fixed $x \geq 1$ and strictly increasing in $a$ when $a > 0$ for fixed $x > 1$. Depending on what you need, you could vary $a$ to get different bounds. –  Antonio Vargas Jul 28 '12 at 23:07

2 Answers 2

up vote 6 down vote accepted

NB: This answer has two parts. The first part addresses the integral in the question and the second provides some simple lower bounds on $\Gamma(s,x)$.

Part I: Bounds on the integral in the question.

Here we provide some tight bounds on the integral $$ \newcommand{\rd}{\,\mathrm d} I(a) := \int_a^\infty y^{-r} e^{-b(y-a)^2}\rd y $$ and, in particular, give an asymptotic equivalence.

Claim: For $a > 0$, $b > 0$, and $r > 0$, $$ \sqrt{\frac{\pi}{4b}}\, \Big(a+\frac{1}{\sqrt{b\pi}} \Big)^{-r} \leq I(a) \leq \sqrt{\frac{\pi}{4b}} \,a^{-r} \>. $$ In particular, we have that $a^r I(a) \uparrow \sqrt{\pi/4b}$ as $a \to \infty$.

Proof. (Lower bound.) Use the substitution $z = \sqrt{2b}(y-a)$ to rewrite the integral in the "canonical form" $$ \sqrt{\frac{\pi}{4 b}} \int_0^\infty (a+z/\sqrt{2b})^{-r} \frac{2 e^{-\frac{1}{2}z^2}}{\sqrt{2\pi}} \rd z = \sqrt{\frac{\pi}{4 b}} \mathbb E (a+Z/\sqrt{2b})^{-r}\>, $$ where $Z$ is a standard half-normal random variable. The function $(a+z/c)^{-r}$ is convex in $z$ and so Jensen's inequality yields the stated lower bound by noting that $\mathbb E Z = \sqrt{2/\pi}$.

(Upper bound). Trivially $(a+z/\sqrt{2b})^{-r} \leq a^{-r}$ and so monotonicity of expectation yields the stated upper bound.

(Asymptotic result). That $I(a) \sim \sqrt{\pi/4b} \,a^{-r}$ as $a \to \infty$ follows directly from the upper and lower bounds. That $a^r I(a) \uparrow \sqrt{\pi/4b}$ follows from the fact that $(1+z/a)^{-r}$ is monotone increasing in $a$ and from monotonicity of expectation.

Part II: Lower bounds on $\Gamma(s,x)$.

Below, we assume real $s$ and $x > 0$. We use similar tricks to those in the previous part. Generally, the bounds we obtain below are asymptotically of the form $g(s) e^{-x} x^{s-1}$ as $x \to \infty$.

First of all, note that $$ \Gamma(s,x) = \int_x^\infty t^{s-1} e^{-t} \rd t = e^{-x} \int_0^\infty (t+x)^{s-1} e^{-t} \rd t = e^{-x} \mathbb E(T+x)^{s-1} \>, $$ where $T \sim \mathrm{Exp}(1)$.

Claim: For $s \leq 1$ and $s \geq 2$, we have $\Gamma(s,x) \geq e^{-x} (1+x)^{s-1}$.

Proof. $(t+x)^{s-1}$ is convex in $t$ for the corresponding $s$. Applying Jensen's inequality to the right-hand side of the above display equation and noting $\mathbb E T = 1$ yields the result.

Claim: For $s \geq 1$, we have $\Gamma(s,x) \geq e^{-x} \Gamma(s) (1+x/s)^{s-1}$. The inequality reverses for $0 < s < 1$.

Proof: The function $f(t) = (1+x/t)^{s-1}$ is convex for $s \geq 1$ and concave for $s < 1$. Apply Jensen's inequality to the right-hand side of $$ \Gamma(s,x) = \int_x^\infty t^{s-1} e^{-t} \rd t = e^{-x} \int_0^\infty (1+x/t)^{s-1} t^{s-1} e^{-t} \rd t = e^{-x} \Gamma(s) \mathbb E(1+x/G)^{s-1} \>, $$ where $G$ is a gamma random variable with shape parameter $s$ and scale parameter 1, and hence mean $\mathbb E G = s$. (For the reverse inequality when $s < 1$, the restriction to $s > 0$ is necessary to maintain the probabilistic interpretation and, hence, the application of Jensen's inequality.)

share|improve this answer

I've been looking at bound and other formulas for $\Gamma(a,x)$ for a few months now. There are two works of reference for this:

  1. "The Incomplete Gamma Functions Since Tricomi" by Gautschi, 1998

  2. "Uniform Bounds for the Complementary Incomplete Gamma Function" by Borwein and Chan, 2009.

The first one is a pretty comprehensive survey of pre-1998 results on the incomplete gamma function, including upper and lower bounds.

The second one is a refinement of existing lower and upper bound to complex numbers. The problem is the bounds only work for a > 1, some of them even need a > 2 or a > 3, which has been useless for me :(

share|improve this answer
    
Gautschi's work is always a good source for gamma-related results. –  cardinal Jul 29 '12 at 0:33
    
@cardinal : Good lower bounds, the part for 0 < s < 1 was very useful for me. Is there anything for -2 < s < 0 ? Also, is any of this published anywhere? –  miladydesummer Jul 29 '12 at 7:40
    
The same bound works for the stated integral whenever $s < 1$ and $x > 0$ with no change in the proof argument. Note that, like in the first part of my answer, we can get a "trivial" (but, useful) upper bound for $s < 1$ via the inequality $(t+x)^{s-1} \leq x^{s-1}$. That is, $\Gamma(s,x) \leq e^{-x} x^{s-1}$. I don't know about published versions of these; but, I have to imagine they exist. –  cardinal Jul 29 '12 at 11:37

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.