Take the 2-minute tour ×
MathOverflow is a question and answer site for professional mathematicians. It's 100% free, no registration required.

Let $X$ and $Y$ be two positive random variables with $Y < X$; these may be highly correlated. I would like a reasonable condition on $X$ and $Y$ so that the ratio $X/Y$ has a finite moment-generating function. By this I mean that $\mathbb E e^{r X/Y} < \infty$ for all $r \in \mathbb R$.

Here's one answer. Suppose that X and 1/Y both have super-Gaussian tail decay, that is,

$P(X > u) \le C e^{-cu^{2+\delta}},$

for positive constants $c$, $C$ and $\delta$, and similarly for $1/Y$. Then $X/Y$ has super-exponential tail decay:

$P(X/Y > u) = P(X > Yu$ and $Y \le 1/\sqrt{u}) + P(X > Yu$ and $Y > 1/\sqrt{u})$

$\le P(1/Y > \sqrt{u}) + P(X > \sqrt{u})$

$\le 2Ce^{-cu^{(2+\delta)/2}} = 2Ce^{-cu^{1+\delta/2}}.$

This gives a finite moment-generating function by the following simple argument, which uses the fundamental theorem of calculus and Tonelli's theorem. Let $Z = X/Y$.

$\mathbb Ee^{rZ} = \mathbb E \left[ 1 + \int_0^Z re^{ru} ~du \right] = 1 + \int_0^\infty re^{ru} \mathbb P(Z > u) ~du \le 1 + 2Cr \int_0^\infty e^{ru} e^{-cu^{1+\delta/2}} ~du$,

which is finite for all $r \in \mathbb R$. Thus, super-Gaussian tail decay suffices, but this is a very strong condition which I'd like to weaken.

(In fact, I only need that $\mathbb Ee^{rX/Y} < \infty$ for any one positive value of $r = r_0$. In that case, we may choose $r_0 = c/2$, and take $\delta = 0$ in all the above arguments. Then the integral $\int_0^\infty e^{r_0 u} e^{-cu} ~du$ still converges.)

share|improve this question
Leonid, I can tell you exactly the context, if it helps. Let SPD be the set of symmetric, positive-definite d x d matrices. Consider the space Omega = C(R², SPD); this is the space of continuous (better yet, smooth) Riemannian metrics on R². Let P be a "reasonable" probability measure on Omega, and let g be a random variable with distribution P. That is, g is a random Riemannian metric on R². Let X = maximum eigenvalue of g(x) for x in [0,1)², and Y = minimum eigenvalue. Clearly, X and Y should be correlated. I want some understanding of the distribution of X/Y. –  Tom LaGatta Jan 15 '10 at 9:43

1 Answer 1

up vote 2 down vote accepted

Let $D$ denote a set of probability distributions of positive random variables and $r$ a positive real number. Consider the following properties:

  1. For every random variables $X$ and $T$ with probability distributions in $D$, $E(\mathrm{e}^{rXT})$ is finite.
  2. For every random variable $X$ with probability distribution in $D$, $E(\mathrm{e}^{rX^2})$ is finite.

Property 1 implies Property 2 (choose $X=T$). Conversely, if Property 2 holds, note that $$ \mathrm{e}^{ab}\le\mathrm{e}^{a^2}+\mathrm{e}^{b^2}-1$$ for every nonnegative $a$ and $b$ since the inequality is symmetric and, if $a\le b$, then the LHS is $\le\mathrm{e}^{b^2}$. For $a=\sqrt{r}X$ and $b=\sqrt{r}T$, one gets $$ E(\mathrm{e}^{rXT})\le E(\mathrm{e}^{rX^2})+E(\mathrm{e}^{rT^2})-1, $$ which proves that Property 1 holds. Hence Properties 1 and 2 are equivalent.

Coming back to your context, you ask for Property 1 to hold with $T=1/Y$ and with no further constraint on the joint distribution of $(X,T)$ that $XT\ge1$ almost surely, a condition which holds as soon as $X\ge1$ and $T\ge1$ almost surely, for example. Thus, it seems that you are left with the condition that the set of probability distributions $D$ to which the distributions of $X$ and $T=1/Y$ belong, fulfills Property 2.

Basically, this means that one can replace your control $C\mathrm{e}^{-2ru^{2}}$ of the tails $P(X\ge u)$ and $P(1/Y\ge u)$ by $$Cu^{-(2+\delta)}\mathrm{e}^{-ru^2}, $$ as soon as $\delta>0$ and that, without further information on the relationship between $X$ and $Y$, one cannot expect a much better sufficient condition.

share|improve this answer

Your Answer


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.