Is $f(P_X,P_Y) \equiv 2 \max\{ P( X < Y ), P( Y < X ) \} - 1$ a valid distance measure for one dimensional distributions, & if so does it have a name?

42 Views Asked by At

Suppose $X$ and $Y$ are independent random variables over $\mathbb R$ with distributions $P_X(x)$ and $P_Y(x)$, respectively, and $x \in \mathbb R$. Is $$f(P_X,P_Y) \equiv 2 \max\{ P( X < Y ), P( Y < X ) \} - 1 $$ a valid distance measure for one dimensional distributions? If so, does it have a name?

The idea is that if $P_X$ has little or no overlap with $P_Y$, then either $P( X < Y )$ or $P( Y < X )$, the probability of one r.v. being greater than another, will be close to 1. If they overlap strongly, then $\max\{ P( X < Y ), P( Y < X )\}$ should be close to $\frac{1}{2}$.

The assumption is that $$\max\{ P( X < Y ), P( Y < X )\} \geqslant \frac{1}{2}.$$ This should be true since $P( Y < X ) = 1 - P( X < Y )$, and $\max\{ P( X < Y ), 1 - P( X < Y ) \}$ has to be at least one half for $0 \leqslant P( X < Y ) \leqslant 1$.

Then $f$ as defined above should range from zero to one, depending on the extent of overlap of $P_X(x)$ and $P_Y(x)$. I don't know how to prove this, so I'm posting it as a question here to see if there's a flaw in my argument or not :)

1

There are 1 best solutions below

5
On

there are $X,Y$ so that $f(P_X,P_Y)=0$ while $P_X\neq P_Y.$

A random variable is $Z$ is even if $P_Z(x)=P_Z(-x).$ If $Z$ is even then, $P_Z(x)=P_{-Z}(x).$

If both $X,Y$ are even, continuous and independent, then:

$$P(X<Y)=P(-X<-Y)=P(X>Y)$$

So, $f(P_X,P_Y)=0.$

It is not hard to show there are such $X,Y$ with $P_X\neq P_Y.$

But a metric should not be zero between unequal points in your space.