2
$\begingroup$

Let $p$ and $q$ be probability densities on $\mathbb R$, with respect to the Lebesgue measure $dx$. The corresponding Hellinger integral and distance are $H(p,q):=\int_{\mathbb R}\sqrt{pq}\,dx$ and $\rho(p,q):=\sqrt{\frac12\int_{\mathbb R}(\sqrt{p}-\sqrt q)^2\,dx}=\sqrt{1-H(p,q)}$.

For real $t$, define the $t$-shifted version $p_t$ of $p$ by the formula $p_t(x):=p(x-t)$ for real $x$. A general question is this: Are there broad conditions that guarantee that $H(p_0,p_t)$ will be nonincreasing in $t\ge0$?

A more specific question: if $p$ is unimodal (that is, nondecreasing to the left of some point $c$ and nonincreasing to the right of $c$), will it guarantee that $H(p_0,p_t)$ is nonincreasing in $t\ge0$? This is easy to see if $p$ is also assumed to be symmetric. Numerical experiments (with piecewise-constant $p$) suggest that the unimodality should be enough, even without the symmetry.

(Of course, for (say) saw-like $p$'s, we will not have the desired monotonicity.)

To put this into a context: if one has the strict version of the desired monotonicity, this will allow the convenient reparameterization $[0,\infty)\ni t\mapsto\tau:=\rho(p_0,p_t)$ of the statistical parametric shift (location) family $(p_t)$ of densities.

$\endgroup$
7
  • $\begingroup$ Why don't you simply differentiate w.r.t. $t\geq 0$? $\endgroup$ Jan 5, 2016 at 7:19
  • 1
    $\begingroup$ If the probability density is log-concave function then the claim is true. Indeed, $\sqrt{p(x)p(x+t)}$ is log-concave, then by Prekopa--Leindler $f(t)=\int_{\mathbb{R}}\sqrt{p(x)p(x+t)} dx$ is log-concave, and by change of variables $f(t)$ it is even. Therefore $f(t)$ is decreasing for $t\geq 0$. Actually this argument can be used to show that the claim is true if $B(x,y)=p(x)q(y)$ is quasiconcave function. $\endgroup$ Jan 5, 2016 at 8:35
  • $\begingroup$ leo: so, what should I do after the differentiation? $\endgroup$ Jan 5, 2016 at 14:16
  • $\begingroup$ Paata: Your point about the log-concave case is very nice. However, in the problem from which my question came, the log-concave case can be easily tackled without reparameterization. Can you elaborate on the "quasiconcave" part of your comment? In particular, I don't know what you mean by $q$. In my question, I only have $p$ and its shifts $p_t$. $\endgroup$ Jan 5, 2016 at 14:43
  • $\begingroup$ Iosif, I am sorry I meant $p(x)p(y)$ is quasiconcave. It was a typo in my previous comment. This case does not solve your conjecture for unimodal functions, but at least it covers some cases of it. Yes I will write details later because I am not with my computer. $\endgroup$ Jan 5, 2016 at 16:48

1 Answer 1

2
$\begingroup$

I'm pretty sure that you figured it out by now, but I'll post it for the sake of completeness. Any unimodal $\sqrt{p}$ can be approximated by a sum $\sum c_k\chi_{I_k}$ with $I_1\subset I_2\subset\dots$. Now just notice that the function $\int(\chi_{I_k})(\chi_{I_j})_t$ is non-increasing in $t\ge 0$ for any $k,j$.

$\endgroup$

Your Answer

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

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