Can any of our Monte Carlo experts explain the "unexpected" expectation at the end of this answer?

Ex post facto summary of the other question/answer: if $X_1,\dots,X_n$ are IID random variables and the expectations $\mathrm{E}[X_i/\bar{X}]$ exist, then a simple symmetry argument shows that $\mathrm{E}[X_i/\bar{X}]=1$, but a Monte Carlo experiment with $X_i\sim\mathrm{N}(0,1)$ seems to contradict this proposition.

x <- matrix(rnorm(10^6), nrow = 10^5)

[1] 5.506203

1 Answer 1


The explanation to the Monte Carlo evaluation of the ratio $\mathbb{E}[X_1/(X_1+X_2)]$ taking weird values is that the expectation does not exist. As a transform of a Cauchy $X_1/X_2$ in your Normal example. Indeed, \begin{align*} \mathbb{E}[X_1/(X_1+X_2)] &=\mathbb{E}[1/(1+X_2/X_1)]\\ &=\int_{-\infty}^{+\infty} \frac{1}{1+y}\,\frac{1}{\pi(1+y^2)}\text{d}y \end{align*} which is not integrable at $y=-1$ since equivalent to $(y+1)^{-1}$.

Note that $X_1/\bar{X}$ is not a Cauchy variate but the transform of a Cauchy variate by the function$$f:\ y \to \dfrac{n}{1+\sqrt{n-1}y}$$The reason is that$$(X_2+\ldots+X_n)\sim\text{N}(0,n-1)$$and that $$\frac{X_1}{\bar{X}}=\dfrac{n}{1+(X_2+\ldots+X_n)/X_1}=\dfrac{n}{1+\sqrt{n-1}Z/X_1}$$where $Z\sim\text{N}(0,1)$.

Note that, as $n$ grows to infinity, $X_1/\bar{X}$ converges in distribution to the random variable equal to $\pm \infty$ with probability $1/2$.

  • 2
    $\begingroup$ In the Gamma example, the ratio is bounded by $1$ hence has a finite expectation. $\endgroup$
    – Xi'an
    Oct 14, 2017 at 20:05
  • 4
    $\begingroup$ OK, so the symmetry argument works, but only if the expectation exists in the first place... Of course... $\endgroup$
    – Zen
    Oct 14, 2017 at 20:06
  • 1
    $\begingroup$ @Xi'an: you are right about this not being a Cauchy, and your answer is spot-on. I'll delete my answer, since it's actively misleading. $\endgroup$ Oct 16, 2017 at 7:49

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.