11
$\begingroup$

If $Y \sim N(\mu,\sigma^2)$ is normally distributed, then $X=\mathrm{e}^Y$ is lognormally distributed. To get the log-$\mu$ and log-$\sigma$ of this lognormal distribution you calculate $$\sigma^2 = \ln\left( \frac{\mathit{Var}}{E^2} + 1 \right)$$ and $$\mu = \ln(E)-\frac{\sigma^2}{2}$$.

But how is this in the case of a bivariate Normaldistribution, $Y \sim N((\mu_1, \mu_2),\Sigma)$, $\Sigma=((\sigma_1,\rho),(\rho,\sigma_2))$ How do I get $\mu, \sigma, \rho$? And more general in the case of a multivariate distribution?

$\endgroup$
2
  • $\begingroup$ What is $E$ in your notation in the expressions for $\sigma^2$ and $\mu$? $\endgroup$
    – Confounded
    Apr 16, 2020 at 9:29
  • $\begingroup$ I presume that $E$ is meant to be the population mean of $X$, i.e. $E=E[X]$ and similarly $Var$ is meant to be $\text{Var}[X]$ $\endgroup$
    – Glen_b
    Nov 8, 2021 at 2:05

1 Answer 1

15
$\begingroup$

$\text{Cov}(X_1,X_2)=E(X_1X_2)-E(X_1)E(X_2)$

$E(X_1X_2)=E(e^{Y_1+Y_2})$

Now the distribution of $Y_1+Y_2$ is normal (and straightforward), so $E(e^{Y_1+Y_2})$ is just the expectation of a univariate lognormal.

The $E(X_1)E(X_2)$ term you can already do.

As a result, it's straightforward to write $\text{Cov}(X_1,X_2)$ in terms of $\mu,\sigma$ and $\rho$ and thereby to solve for $\rho$.


$Y_1+Y_2\sim N(\mu_1+\mu_2,\sigma_1^2+\sigma_2^2+2\rho\sigma_1\sigma_2)$, so

$e^{Y_1+Y_2}\sim logN(\mu_1+\mu_2,\sigma_1^2+\sigma_2^2+2\rho\sigma_1\sigma_2)$,

which has expectation $\exp[\mu_1+\mu_2+\frac{1}{2}(\sigma_1^2+\sigma_2^2+2\rho\sigma_1\sigma_2)]$.

$E(X_i)=\exp(\mu_i+\frac{1}{2}\sigma_i^2)$

So $\text{Cov}(X_1,X_2)=E(X_1)E(X_2)[\exp(\rho\sigma_1\sigma_2)-1]$

And hence:

$\exp(\rho\sigma_1\sigma_2)-1=\frac{\text{Cov}(X_1,X_2)}{E(X_1)E(X_2)}$

$\rho=\log(\frac{\text{Cov}(X_1,X_2)}{E(X_1)E(X_2)}+1)\cdot\frac{1}{\sigma_1\sigma_2}$

You can extend this approach to calculating $\rho_{ij}$ from $\text{Cov}(X_i,X_j)$ and the other quantities.

However, if you're trying to do this to estimate parameters from a sample, using sample moments of a lognormal to do parameter estimation (i.e. method-of-moments) doesn't always perform all that well. (You might consider MLE if you can.)

$\endgroup$
3
  • $\begingroup$ It seems from your formulas for the covariance of two log-normal RVs that when $\rho=1$, the covariance does not reduce to the product of two standard deviations, i.e. $\sqrt{Var(X_1)Var(X_2)}$, and so the correlation between the log-normals is not 1. Is this correct? What would $\rho$ have to be for the log-normals to be 100% correlated then? $\endgroup$
    – Confounded
    Aug 25, 2020 at 9:25
  • $\begingroup$ In general when $\rho=1$ the relationship between the lognormals is curved and so the Pearson correlation won't generally reach 1. You'd need $\rho=1$ and $\sigma_1=\sigma_2$ for it to be straight. $\endgroup$
    – Glen_b
    Aug 25, 2020 at 10:15
  • $\begingroup$ In short, the Pearson correlation of the $X$'s is different from the Pearson correlation of the $Y$'s (the latter is denoted $\rho$ above) $\endgroup$
    – Glen_b
    Sep 22, 2022 at 1:01

Your Answer

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

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