0
$\begingroup$

If I am not mistaken, Hellinger distance between P and Q is generally given by:

$$ H^2(P, Q) = \frac12 \int \left( \sqrt{dP} - \sqrt{dQ} \right)^2 .$$

If P and Q, however, are two differently shifted log-normal distributions of the following form $$ {\frac {1}{(x-\gamma)\sigma {\sqrt {2\pi \,}}}}\exp \left(-{\frac {[\ln (x-\gamma)-\mu ]^{2}}{2\sigma ^{2}}}\right) ,$$

how would the Hellinger distance then be formed?

in terms of: $$\gamma1,\gamma2, \mu1, \mu2 .. etc$$

$\endgroup$
  • 2
    $\begingroup$ It's the same formula. Are you asking for an expression in terms of the parameters of the two distributions? $\endgroup$ – Dougal Aug 8 '18 at 13:41
1
$\begingroup$

Note that \begin{align} H^2(P, Q) &= \frac12 \int (\sqrt{dP} - \sqrt{dQ})^2 \\&= \frac12 \int dP + \frac12 \int dQ - \int \sqrt{dP} \sqrt{dQ} \\&= 1 - \int \sqrt{dP} \sqrt{dQ} ,\end{align} and that the density function is 0 if $x \le \gamma$. Thus your question asks to compute \begin{align} 1 - H^2(P, Q) &= \int_{\max(\gamma_1,\gamma_2)}^\infty \sqrt{\frac{1}{(x - \gamma_1) \sigma_1 \sqrt{2 \pi}} \exp\left( - \frac{\left(\ln(x - \gamma_1) - \mu_1\right)^2}{2 \sigma_1^2} \right)} \\&\qquad\qquad \sqrt{\frac{1}{(x - \gamma_2) \sigma_2 \sqrt{2 \pi}} \exp\left( - \frac{\left(\ln(x - \gamma_2) - \mu_2\right)^2}{2 \sigma_2^2} \right)} dx \\&= \sqrt{\frac{1}{2 \pi \sigma_1 \sigma_2}} {\huge\int}_{\max(\gamma_1,\gamma_2)}^\infty \frac{\exp\left( - \frac{\left(\ln(x - \gamma_1) - \mu_1\right)^2}{4 \sigma_1^2} - \frac{\left(\ln(x - \gamma_2) - \mu_2\right)^2}{4 \sigma_2^2} \right)}{\sqrt{(x - \gamma_1)(x - \gamma_2)}} dx .\end{align} Assume (WLOG) that $\gamma_1 \ge \gamma_2$, and do a change of variables to $y = \ln(x - \gamma_1)$, $dy = \frac{1}{x - \gamma_1} dx$. Let $\Delta = \gamma_1 - \gamma_2$, so we have $ x - \gamma_2 = \exp(y) + \Delta .$ Then we get $1 - H^2(P, Q)$ as \begin{align} \sqrt{\frac{1}{2 \pi \sigma_1 \sigma_2}} \int_{-\infty}^\infty \exp\left( - \frac{\left(y - \mu_1\right)^2}{4 \sigma_1^2} - \frac{\left(\ln(\exp(y) + \Delta) - \mu_2\right)^2}{4 \sigma_2^2} \right) \sqrt{\frac{\exp(y)}{\exp(y) + \Delta}} dy .\end{align}

If $\gamma_1 = \gamma_2$ so $\Delta = 0$, this works out to $$ H^2(P, Q) = 1 - \sqrt{\frac{2 \sigma_1 \sigma_2}{\sigma_1^2 + \sigma_2^2}} \exp\left( - \frac{(\mu_1 - \mu_2)^2}{4 (\sigma_1^2 + \sigma_2^2)} \right) .$$ For $\Delta \ne 0$, though, neither I nor Mathematica made any immediate headway.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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