MathOverflow is a question and answer site for professional mathematicians. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

I'm sure that many MO users would answer "Oh, yes, I'm more intelligent than the average intelligence of the population that has an intelligence greater than the (absolute) average". And someone, less modestly: "Even more than the average of those who are more intelligent than the average of those who are more intelligent than the average". And so on.

Anyway, the silly title was created just to attract curiosity a bit: my question is not about the intelligence of MO users, on which I have no doubts! :)

So, take a quantity $X$ that we suppose normally distributed (b.t.w., I have no deep knowledge of probability theory), i.e. it's described by a gaussian distribution that we suppose standardized and call $f(x)$.

Now, define:

$M_0:= \int_{-\infty}^{\infty}f(x)dx=1$


and, inductively,

$M_{n+1}:= \int_{\mu_n}^{\infty}f(x)dx$


I think this describes the situation in which your $X$ (tallness? Weight?...) has the value $\mu_n$ precisely when you're as $X$ as the average of those who are more $X$ than the average of those who are more $X$ than...... (n times). If not, please explain why.

So my questions:

  1. How does the sequence $\mu_n$ behave asymptotically? Does it converge?
  2. If yes, is there a nice expression for the limit?
  3. Is there even a reasonably explicit expression ("closed form") for $\mu_n$ as a function of $n$?
share|cite|improve this question
I'd only change "...the average of who is more..." into "...the average of those who are..." – Mariano Suárez-Alvarez Sep 20 '10 at 17:58
@Mariano: thanks for the English grammar correction. I've just edited. – Qfwfq Sep 20 '10 at 18:09
Suppose $f(x) := 1_{\mathbb{R}_+}(x) \cdot \exp(-\lambda x)/\lambda$. Now $M_{n+1} = \exp(-\lambda \mu_n)$ and $\mu_{n+1} = \exp(-\lambda[\mu_n-\mu_{n-1}])\cdot (\lambda\mu_n +1)/\lambda$. MATLAB goes nuts and spits out NaNs when I try to get more than a handful of terms for various values of $\lambda$. – Steve Huntsman Sep 20 '10 at 18:54
The title should be changed. – Martin Brandenburg Sep 21 '10 at 0:31
up vote 22 down vote accepted

As in Nate's answer, we are interested in iterating the function $$G(y) := \frac{ \int_{y}^{\infty} x e^{- x^2} dx}{\int_{y}^{\infty} e^{- x^2} }.$$

The numerator is $e^{-y^2}/2$ (elementary). The denominator is $e^{-y^2}/2 \cdot y^{-1} \left( 1-(1/2) y^{-2} + O(y^{-4}) \right)$ (see Wikipedia). So $G(y) = y + (1/2) y^{-1} + O(y^{-3})$.

Set $z_n = \mu_n^2$. Then $$z_{n+1} = (\mu_n+\mu_n^{-1}/2 + O(\mu_n^{-3}))^2 = \mu_n^2 + 1 + O(\mu_{n}^{-2}) = z_n + 1 + O(z_n^{-1}).$$ So $z_n \approx n$ and we see that $\mu_n \to \infty$ like $\sqrt{n}$.

I haven't checked the details, but I think you should be able to get something like $\mu_n = n^{1/2} + O(1)$.

share|cite|improve this answer

We have $\mu_n \uparrow \infty$. Proof: let $$G(y) = \frac{\int_y^\infty x f(x) dx}{\int_y^\infty f(x) dx}$$ so that $\mu_{n+1} = G(\mu_n)$. Clearly $G$ is a continuous function and $G(y) > y$ for all $y$. But if $\mu_n \to \mu$ for some finite $\mu$ we must have $G(\mu) = \mu$, a contradiction.

More generally, this should show that if $X$ is a continuous random variable with essential supremum $M$, and we define $G(y) = E[X | X \ge y]$ for $y < M$, then the iterates $G^n(y) \to M$.

share|cite|improve this answer
Nice. So it seems this question also deserves a "Dynamical systems" tag... – Qfwfq Sep 20 '10 at 19:21

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.