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 would like to know if there's a way to compute or approximate the following expectation:


where $X=N(0,\sigma^2)$ and $n,c>0$ (you can also assume that $n$ is a natural number if necessary). After googling a bit I found the following articles: 1, 2 which talk about negative moments of positive random variables but applying the methods yields a divergent series.

share|cite|improve this question
@dp, can you write this similarly to the integral you asked about in your other question? I don't know if you noticed, but I updated my answer in EDIT 2 with a series that works very well for small $b$. – Tom Dickens Mar 19 '13 at 16:00

Since you say you're interested in computing the expectation, here's a way to calculate the moments numerically.

Rearranging a little and defining $Y = X - \log c$, the expectation you're after, which I'll call $I$, is

$$I = c^{-n} \cdot \mathbb{E} \left[ \frac{ 1 }{ \left( 1 + e^{Y} \right)^{n}} \right].$$

Massaging the term inside the integral still further, this can be rewritten

$$ I = c^{-n} \cdot \mathbb{E} \frac{ e^{-nY/2} }{ \left( e^{Y/2} + e^{-Y/2} \right)^{n} } = c^{-n} \cdot \mathbb{E} \left[ e^{-nY/2} \frac{ 1 }{ \left( 2 \cosh (Y/2) \right)^{n} } \right] \ .$$

The point of doing this is that $1/(2\cosh Y/2)^{n}$ has a nice Fourier transform. Specifically, $$\frac{1}{(2 \cosh (Y/2))^{n}} = \int_{-\infty}^{\infty} e^{iYz} \mathscr{F}_{n}(z) \, dz \ , $$ where $$ \mathscr{F}_{n}(z) = \frac{1}{2 \pi} \frac{ \Gamma(n/2+iz)\Gamma(n/2-iz)}{ \Gamma(n) } $$ are rapidly decaying, even functions of $z$. (For integer $n$, they can be expressed more simply: for example, $\mathscr{F}_{1}(z) = \frac{1}{2} \operatorname{sech} \pi z$.)

So $$I = c^{-n} \mathbb{E} \int_{-\infty}^{\infty} e^{-n Y/2} e^{iYz} \mathscr{F}_{n}(z) \, dz \ . $$

Writing out the characteristic function of the Normal random variable $Y$ explicitly, and rearranging a bit, this becomes

$$I = c^{-n/2} \int_{-\infty}^{\infty} \cos \left[ \left( \log c + \frac{1}{2} \sigma^{2} n \right) z \right] e^{ \frac{1}{2} \sigma^{2} \left( \frac{n^{2}}{4} - z^{2} \right) } \mathscr{F}_{n}(z) \, dz \ , $$

which is easy to evaluate numerically. Here's what the moments look like for $\sigma=1$ and $c$ equal to 0.7, 1, and 1.3:

enter image description here

This approach is also useful if $X$ and $Y$ are not Normal, so long as you know their characteristic functions.

share|cite|improve this answer

As $c \to \infty$ for fixed $n$ and $\sigma$, it looks to me like $ {\mathbb E}\left[(c+e^X)^{-n}\right]$ has the asymptotic series $$c^{-n} + \sum_{j=1}^\infty (-1)^j {{n+j-1} \choose j} c^{-n-j} e^{j^2 \sigma^2/2}$$

Yes, like most asymptotic series, this is divergent, but partial sums might still be useful approximations.

share|cite|improve this answer
Thank you for the answer. Unfortunately $c$ is fixed and small so this is not so useful for me (I also found a couple of ways to derive this asymptotic series). – user9121 Mar 6 '13 at 5:06
Another possibility is to use Laplace's method, expanding around the maximum of the integrand $\dfrac{(c+e^x)^{-n}}{\sqrt{2\pi} \sigma} e^{-x^2/(2 \sigma^2)}$, which is at the real root of $(n \sigma^2 + x) e^x + c x$. – Robert Israel Mar 11 '13 at 18:49

Maple 2015 produces a simple answer by

with(Statistics): X := RandomVariable(Normal(0, sigma)):

Mean(c+exp(X))^(-n) assuming c>0, n::posint;

$$ \left( c+{{\rm e}^{1/2\,{\sigma}^{2}}} \right) ^{-n} $$

share|cite|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.