6
$\begingroup$

Consider $\mathcal{F}$ and $\mathcal{G}$ are two continuous distributions and that $\mathcal{G}$ is more dispersed than $\mathcal{F}$ in the sense of satisfying:

$$(1)\quad F^{-1}(\beta)-F^{-1}(\alpha)\leq G^{-1}(\beta)-G^{-1}(\alpha),\quad \forall 0< \alpha < \beta <1.$$

Then, we know (Kochar (2012), eq. 2.16) that (1) is equivalent to

$$(2)\quad\frac{F^\prime(x)}{G^\prime G^{-1}F(x)}\geq 1.$$ (I call the left hand side of (2) the Doksum ratio after (Doksum 1969))

Now consider the ratio:

$$(3)\quad\gamma(\mathcal{F},\mathcal{G})=\underset{x:F^\prime(x)>0}{\min}\;\frac{F^\prime(x)}{G^\prime G^{-1}F(x)}$$

For many distributions I find that, for fixed $\mathcal{F}$, $\gamma(\mathcal{F},\mathcal{G})$ can be made arbitrary large by shifting/rescaling $\mathcal{G}$.

For example, for the normal, student t, Weibull, gamma I find that $\gamma(\mathcal{F},\mathcal{G})$ is an increasing function of $\text{Var}(\mathcal{G})$.

My question is this: Is it always so that for fixed $\mathcal{F}$, $\gamma(\mathcal{F},\mathcal{G})$ can be made arbitrarily large by re-scaling/shifting $\mathcal{G}$?

$\endgroup$
5
$\begingroup$

It is geometrically obvious that shifting $G$ changes nothing and that rescaling will rescale the denominator by the same factor--it's merely a matter of keeping track of units of measurement. The following is a rigorous algebraic demonstration. Your conclusion follows immediately.


Let $a\gt 0$ and $b$ be real numbers and define $G_{(a,b)}$ to be the $a$-scaled, $b$-shifted version of $G$:

$$G_{(a,b)}(x) = G(ax + b).$$

Let $0 \le \alpha \le 1$. Compute that

$$(G_{(a,b)})^{-1}(\alpha) = \frac{G^{-1}(\alpha) - b}{a}$$

and (via the Chain Rule)

$$\frac{d}{dx}G_{(a,b)}(x) = a \left(\frac{d}{dx} G\right)(ax + b).$$

Thus

$$\left(\frac{d}{dx}G_{(a,b)}\right)\left(\left(G_{(a,b)}\right)^{-1}(\alpha)\right) = a \left(\frac{d}{dx}G\right)\left(G^{-1}(\alpha)\right).$$

This shows that shifting does not change the denominator of the Doksum ratio and scaling multiplies the denominator by $a$. As $a\to 0^{+}$, the ratio will grow without bound provided its numerator $F^\prime(x)$ exists and is positive.

$\endgroup$

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.