Consider a random variable $x$ with pdf $f(x)$, and have $x \ge 0$. The moment generating function is defined as $M(t)=\int^{\infty}_{-\infty}e^{-tx}f(x)dx$ (noted that we change the sign of $t$ compared with common case ). For convenient we consider only when $t>0$. The problem is: When would a function be the moment generating function of a distribution? There are some trivial conditions like $M(0)=1$ and $M(t)$ is strictly decreasing. But obviously these are not enough.

An observation is that the $M(t)$ should lie in the convex hull of $e^{-tx}$. From inequality techniques we have $\log M(t)$ up convex. Is this condition sufficient? An related but a bit different problem is given $t_1,t_2,\dots,t_n$ and $M_1,M_2,\dots,M_n$ when there exist $M(t)$ such that $M(t_i)=M_i$?

Add Jul 12:

The reason that the up convex condition is not enough are as follows: Consider a function $\bar M(t)$ that it's logarithm is linear in an interval, using Cauchy inequality we can demonstrate that $\bar M(t)$ is composed of one $e^{-tx}$. But we can find other continuation.

  • 1
    $\begingroup$ Don't restrict yourself to real $t$. $M(it)$ is, up to some constants, the Fourier transform of $f$. $\endgroup$ – Thomas Andrews Jul 12 '13 at 13:02

Take a look at this http://en.wikipedia.org/wiki/Bernstein's_theorem_on_monotone_functions

completely monotone functions for example have a "moment generating function representation"


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.