$E(F)=\int xf_{k,m}dx$ where $f_{k,m}(t) = \Gamma(t)=\frac{\Gamma((k+m)/2)}{\Gamma (k/2)\Gamma(m/2)}k^{k/2}m^{m/2}t^{k/2 - 1}(m+kt)^{-(k+m)/2}$.

How do you find $E(F)$? Say you have to convert $x*f(k,m)$ to $C * f(k',m')$ where $f(k',m')$ is a pdf itself, which leaves $E(F)=C$ (which is$\frac{m}{m-2}$).


One elementary way to obtain this result is to integrate the survival function. I will prepare the way with some notation and definitions, and then will integrate the survival function by changing the order of a double integral.

The cumulative distribution function of $F$ is

$$F_{k,m}(x) = I\left(\frac{x}{x + \lambda}; k/2, m/2\right)$$

where $\lambda = m/k$ and $I$ is the regularized incomplete Beta function

$$I(x; \alpha, \beta) = \frac{1}{B(\alpha, \beta)}\int_0^x t^{\alpha-1}(1-t)^{\beta-1}dt.$$

(For notational convenience, I will continue to equate $\alpha$ with $k/2$ and $\beta$ with $m/2$.) Notice that because $I$ is a CDF, its full integral must be unity:

$$1 = I(1;\alpha,\beta) = \frac{1}{B(\alpha, \beta)}\int_0^1 t^{\alpha-1}(1-t)^{\beta-1}dt.$$

If you wish, then, you may take $B(\alpha,\beta)$ to be defined by the integral on the right hand side. What will be important in the sequel is the fact that

$$\frac{B(\alpha+1,\beta-1)}{B(\alpha,\beta)} = \frac{\alpha}{\beta-1}$$

for all $\alpha \gt 0$ and $\beta \gt 1$. You can prove this in many ways (such as integration by parts).

Integration by parts demonstrates that the expectation of any positive distribution $F$ is the integral of its survival function $1-F$:

$$E[F] = \int_0^\infty x\, dF(x) = \lim_{t\to\infty}\left(t(1-F(t)) + \int_0^t (1-F(t)) dt\right) = \int_0^t (1-F(t)) dt$$

(provided the expectation is finite). Applying this to the Fisher-Snedecor distribution and changing the order of integration gives

$$\eqalign{ E[F_{k,m}]& = \int_0^\infty \left(1 - I\left(\frac{x}{x + \lambda}; \alpha,\beta\right)\right) dx \\ &= \int_0^\infty \frac{1}{B(\alpha, \beta)}\int_{x/(x+\lambda)}^1 t^{\alpha-1}(1-t)^{\beta-1}dt\,dx \\ &= \frac{1}{B(\alpha, \beta)}\int_0^\infty t^{\alpha-1}(1-t)^{\beta-1}\left(\int_0^{\lambda t/(1-t)}dx\right)dt \\ &= \frac{\lambda}{B(\alpha, \beta)}\int_0^1 t^{\alpha}(1-t)^{\beta-2}dt \\ &= \frac{\lambda}{B(\alpha, \beta)}\int_0^1 t^{\alpha}(1-t)^{\beta-2}dt \\ &= \frac{\lambda B(\alpha+1, \beta-1)}{B(\alpha,\beta)} \\ &= \lambda \frac{\alpha}{\beta-1} = \frac{m}{k}\frac{k/2}{m/2-1}=\frac{m}{m-2}. }$$

(Each step is simple and elementary or has been previously demonstrated.) The penultimate line required $\beta-1 \gt 0$; that is, $m \gt 2$. It is clear that the last integral (on the antepenultimate line) diverges when $m \le 2$.

| cite | improve this answer | |

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.