If I look at an exponential function, $p(t) = e^{-\mu t}$ where the parameter $\mu$ varies over a gamma distribution given by the density function

$f(\mu) = \frac{1}{\Gamma(a)b^a}\mu^{a-1}e^{-\mu/b}$ (parameters $a, b$ and $\Gamma$ the gamma function)

Then I can talk about the "average" of all exponentials with respect to $\mu$ as the integral


I'm reading a paper where they present, without proof, the claim that this has closed form


That is, the "average" of an exponential curve is a power-law curve. (I'm slightly simplifying the actual claim, so I'm hoping I didn't make a mistake)

On one hand, I'm interested to know what results they used in probability calculus to get to this closed form (it is probably pretty basic, but I just never learned this stuff). More importantly, I'm interested in whether the same technique applies to a family of power-law curves. That is if I instead write $p(t) = t^{-\mu}$, can I easily derive a closed form, for this, and if so, is it a power-law?

As a side note, if the distribution is instead uniform, the answer is (sort of) yes: the integral is trivial and the closed form quickly converges to a power-law curve, as it does in the case of an exponential.

  • $\begingroup$ edited my typo of "gamma distribution" to "gaussian distribution" $\endgroup$ – JeremyKun Oct 13 '12 at 4:29
  • $\begingroup$ Your distribution is a gamma distribution, not a Gaussian distribution. You should roll back your edit. $\endgroup$ – Dilip Sarwate Oct 13 '12 at 14:08

Hey man so here's one approach for the first part. Suppose you believe that the Gamma distribution is indeed a probability distribution, i.e. it integrates to 1. This tells you that $$\int_0^\infty x^{a-1}e^{-\frac{x}{b}}dx =\frac{1}{\Gamma(a) b^a}$$ Now when we try to evaluate your integral, up to fiddling with constants we run into the same integral: $$\int_0^\infty \frac{1}{\Gamma(a)b^a} x^{a-1} e^{\frac{-x}{b}} e^{-xt} dx$$$$ = \frac{1}{\Gamma(a)b^a} \int_0^\infty x^{a-1} e^{-x(\frac{1}{b} + t)} dx$$Now just note that $\frac{1}{b} + t = (\frac{b}{tb + 1})^{-1}$, so plugging into the first formula, we our integral equals $$\frac{1}{\Gamma(a)b^a} \cdot (\Gamma(a) (\frac{b}{(tb+1)})^a ) = (tb + 1)^{-a}$$as desired.

You can attack the second part with the same method. First suppose $t > 0$, then we note first that $t^{-x} = (e^{\log t})^{-x} = e^{-x \log t}$, so when we evaluate $$\frac{1}{\Gamma(a)b^a} \int_0^\infty t^{-x} x^{a-1} e^{-\frac{x}{b}} dx = \frac{1}{\Gamma(a)b^a} \int_0^\infty x^{a-1} e^{-x(\frac{1}{b} + \log t)} dx$$So it's the same problem as before except with $\log t$ instead of $t$, and thus we get this equals $$(b \log t + 1)^{-a} \quad \quad t > 0$$(At the corner case $t = 0$, we just get 1)


This follows directly from the definition of the Gamma function as an integral. $$\begin{align} \int_0^{\infty}\frac{1}{\Gamma(a)b^a}\mu^{a-1}\exp(-\mu/b)\exp(-\mu t) \,\mathrm d\mu &= \int_0^{\infty}\frac{1}{\Gamma(a)b^a}\mu^{a-1}\exp(-\mu(b^{-1}+ t)) \,\mathrm d\mu\\ &= \int_0^{\infty}\frac{1}{\Gamma(a)b^a}\frac{x^{a-1}}{(b^{-1} + t)^{a}}\exp(-x) \,\mathrm dx\\ &= \frac{1}{b^a(b^{-1} + t)^a}\int_0^{\infty}\frac{x^{a-1}\exp(-x)}{\Gamma(a)} \,\mathrm dx\\ &= \frac{1}{(1+bt)^a} \end{align}$$

More generally, for any probability density function $f(x)$, the value of the integral $\int e^{-\lambda x}f(x)\,\mathrm dx$ is the moment-generating function of the random variable evaluated at $-\lambda$.


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.