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.

For a discrete random variable $X$, we have known the exact expression of its PGF $G(z)=E[z^X]$.

The question is how can I get $Pr\{X>k\}$ from this PGF.

I want to have an explicit expression of the probability in terms of $G(z)$.

The Z-Transform is too complex to use. And an explicit expression is preferred rather than a upper bound. Is there some other way?

Thanks!

share|improve this question
    
I found in'Univariate Discrete Distributions, 3ed' another possible solution to this problem. $\Pr\{X\geq k\}=\sum_{j=k}^\infty (-1)^{j+k} C_{j-1}^{k-1} \frac{G^{(j)}(z)|_{z=1}}{j!}. –  Severals-user45972 Oct 26 '12 at 4:07

3 Answers 3

up vote 0 down vote accepted

Using complex integration over the direct unit circle $S^1$, one gets $$\mathbb P(X\gt k)=\frac1{2\mathrm i\pi}\oint_{S^1}\frac{1-G(z)}{1-z}\cdot\frac{\mathrm dz}{z^{k+1}}. $$

To prove this, one can rely on two ingredients:

(1) A distribution transform

For every nonnegative integer valued random variable $X$ integrable and not almost surely zero, there exists a random variable $Y$ such that, for every $k\geqslant0$, $$ \mathbb P(Y=k)=\frac{\mathbb P(X\gt k)}{\mathbb E(X)}. $$ This is because each RHS is nonnegative and their sum is $1$ thanks to the well-known formula $$ \sum_{k\geqslant0}\mathbb P(X\gt k)=\mathbb E(X). $$ If the PGF of $X$ is the function $G$ defined by $G(z)=\mathbb E(z^X)$, then $\mathbb E(X)=G'(1)$ and the PGF of $Y$ is the function $F$ defined by $$ F(z)=\mathbb E(z^Y)=\frac{1-G(z)}{\mathbb E(X)\cdot(1-z)}. $$ (2) The extraction of coefficients through Cauchy formula

For every nonnegative integer valued random variable $Y$ with PGF $F$ and every $k\geqslant0$, $$ \mathbb P(Y=k)=\frac1{2\mathrm i\pi}\oint_{S^1}F(z)\cdot\frac{\mathrm dz}{z^{k+1}}. $$ To see this, expand $F$ as $F(z)=\sum\limits_{n\geqslant0}\mathbb P(Y=n)z^n$ and use the fact that, for every $n\geqslant0$, $$ \oint_{S^1}z^n\cdot\frac{\mathrm dz}{z^{k+1}}=\left\{\begin{array}{ccc}2\mathrm i\pi & \text{if} & n=k,\\ 0 & \text{if} & n\ne k.\end{array}\right. $$ Using (1) and (2) together, one gets $$ \frac{\mathbb P(X\gt k)}{\mathbb E(X)}=\frac1{2\mathrm i\pi}\oint_{S^1}\frac{1-G(z)}{\mathbb E(X)\cdot(1-z)}\cdot\frac{\mathrm dz}{z^{k+1}}, $$ which is equivalent to the desired formula.

share|improve this answer
    
Is this from the Z-Transform thoery? Why the image unit $i$ is not involved? can you give me some reference on this? Thanks. –  Severals-user45972 Oct 25 '12 at 9:33
    
See edited version. –  Did Oct 25 '12 at 10:12
    
Excellent! Thanks! –  Severals-user45972 Oct 25 '12 at 10:27
    
However, I think a factor of $\frac{1}{2\pi i}$ is missed in the formula. –  Severals-user45972 Oct 26 '12 at 4:04
    
Indeed. Thanks. –  Did Oct 26 '12 at 4:52

If $$G(z) = \operatorname{E} (z^X) = \sum_{x=0}^{\infty}\Pr(X=x)z^x$$ then to identify $\Pr(X=x)$ for a particular (non-negative integer) $x$, you take the $x$th derivative of $G(z)$ at $0$ and divide by $x!$.

Or use standard generating function techniques, such as those described in generatingfunctionology.

So for $\Pr(X \gt k)=1 -\Pr(X \le k)$ you can use $$1-\sum_{x=0}^{k}\Pr(X=x) =1- \sum_{x=0}^{k}\frac{G^{(x)}(0)}{x!}$$

share|improve this answer
    
Thanks. But this gives only a particular value. But I want to get $\Pr\{X>k\}$, which is a sum of infinite $p(s)$s. What can I do then? –  Severals-user45972 Oct 25 '12 at 9:21
    
@Severals-user45972: bit added –  Henry Oct 25 '12 at 9:26

This solution is for the positive RV, i.e. RV that takes only positive values. In this case one can show that $$ \sum_{k \geq 0}\mathbf{P}(X \geq k)=\mathbf{E}X=G'_{X}(0) $$

share|improve this answer
    
Yes, you are right. –  Severals-user45972 Oct 25 '12 at 9:37
    
What is $G'_0(X)$? –  Did Oct 25 '12 at 10:11
    
Edited:$\left. \frac{dG_{X}(z)}{dz} \right|_{z=0}$ –  Alex Oct 25 '12 at 10:16

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.