What is the mean of a normal random variable transformed via $g(x) = \frac{1}{a+x}$

35 Views Asked by At

Consider normal random variable $\lambda \sim N(\mu, \sigma^2)$ ($\lambda \in \mathbb{R}$) whose PDF is:

$$ f_\lambda(\lambda) = \frac{1}{\sigma \sqrt{\pi}} \exp \left[ {-\frac{1}{2}\left( \frac{\lambda - \mu}{\sigma} \right)^2} \right] $$

Transformation $g(x) = \frac{1}{a+x}$ is applied and a new random variable $\Lambda$ is generated as:

$$ \Lambda = g(\lambda) \implies \Lambda = \frac{1}{a + \lambda} $$

With $a \in \mathbb{R}$.

The problem

I want to know the expected value of $\Lambda$: $E[\Lambda]$.

My attempt so far was by following the theory. The new PDF is computed according to well known formula:

$$ f_\Lambda(\lambda) = f_\lambda[g^{-1}(\lambda)] \left| \frac{d}{d\lambda}g^{-1}(\lambda) \right| $$

Which means, after a few calculations:

$$ f_\Lambda(\lambda) = \frac{1}{\sigma \sqrt{\pi}} \frac{1}{\lambda^2} \exp \left[ -\frac{1}{2} \left( \frac{1}{\sigma \lambda} - \frac{a}{\sigma} - \frac{\mu}{\sigma} \right)^2 \right] = \frac{\alpha}{\lambda^2} \exp \left[ -\frac{1}{2} \left( \frac{1}{\sigma \lambda} - \beta \right)^2 \right] $$

Having $\alpha = \frac{1}{\sigma \sqrt{\pi}}$ and $\beta = \frac{a}{\sigma} + \frac{\mu}{\sigma}$. Now I can proceed calculating the mean value of $\Lambda$ since I know its PDF by means of: $E[\Lambda] = \int_{-\infty}^{+\infty} \lambda f_\Lambda(\lambda) d\lambda$:

$$ E[\Lambda] = \int_{-\infty}^{+\infty} \frac{\alpha}{\lambda} \exp \left[ -\frac{1}{2} \left( \frac{1}{\sigma \lambda} - \beta \right)^2 \right] d\lambda = \alpha \int_{-\infty}^{+\infty} \frac{1}{\lambda} \exp \left[ -\frac{1}{2} \left( \frac{1}{\sigma \lambda} - \beta \right)^2 \right] d\lambda $$

And here is the problem, this integration is kinda tricky:

$$ \int_{-\infty}^{+\infty} \frac{1}{x} \exp \left[ -\frac{1}{2} \left( \frac{1}{c_1 x} - c_2 \right)^2 \right] dx $$

How to solve this?


Integration attempts

Attempt 1 I have tried with substitutions, but it is horrible because I might get rid of the negative power of the variable in the exponential, but then I get higher order negative powers of the variable multiplying the exponential :(

Attempt 2 I know that the exponential integral function is defined as:

$$ \text{Ei}(x) = -\int_{-x}^{+\infty} \frac{e^{-t}}{t} dt $$

And I have a feeling that the primitive of the integrand might be a function of $\text{Ei}$, but I am not sure.

Attempt 3 Using complex calculus and the Residue Theorem?