2
$\begingroup$

For two continuous probability distributions $F,G$ and their densities, $f,g$, the (squared) Hellinger distance/affinity is given by $d^2_H(F,G)=1-\int_{\mathbb{R}} \sqrt{fg}~dx$. Suppose that $f,g$ are two-component mixture densities with mixing probability $\pi$, such that $$ f(x)=\pi f_0(x)+(1-\pi)f_1(x)\\ g(x)=\pi g_0(x)+(1-\pi)g_1(x) $$ The square root term is fairly complicated given this model. Are there any references that tackle the Hellinger distance between these mixture distributions?

$\endgroup$
1
1
$\begingroup$

For $f=\pi f_0+(1-\pi)f_1$, $g=\pi g_0+(1-\pi)g_1$, and $\pi\in(0,1)$, we have \begin{equation} \frac{\partial^2}{\partial\pi^2}\sqrt{fg}= -\frac{\left(f_1 g_0-f_0 g_1\right){}^2}{4 (fg)^{3/2}}\le0. \end{equation} So, $\sqrt{fg}$ is concave in $\pi$ and hence $d^2_H(F,G)$ is convex in $\pi\,$: $$ d^2_H(F,G)\le\pi d^2_H(F_0,G_0)+(1-\pi)d^2_H(F_1,G_1) $$ for $\pi\in[0,1]$, where $F_j,G_j$ are the distributions with densities $f_j,g_j$, respectively, for $j=0,1$.

$\endgroup$

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.