1
$\begingroup$

I'm reading on Hoeffding's covariance identity, the proof of which is neatly covered here, or, in a similar manner, in this MSE post, but I can't seem to fully understand the trick/property used there.

I.e., assume $(X_1, Y_1)$ and $(X_2, Y_2)$ are two independent vectors with identical distribution. The key point in the proof is to note that we can write

$$ \mathbb{E}[(X_1 - X_2) (Y_1 - Y_2)]$$ as

$$ \mathbb{E}\left(\iint_{\mathbb{R}\times\mathbb{R}} [\mathbb{1}_{u\leq X_1} - \mathbb{1}_{u \leq X_2}] \cdot [\mathbb{1}_{v\leq Y_1} - \mathbb{1}_{v \leq Y_2}]\,du\,dv \right)$$

Why does this hold?

$\endgroup$
2
$\begingroup$

What underlies the equality $\mathbb E(X) = \mathbb E(\int \mathbb 1_{u\le X}\,du)$ is, intuitively, the way one thinks of the Lebesgue integral as coming from partitioning the $y$-axis, whereas the Riemann integral comes from partitioning the $x$-axis.

Think of a reasonable function $f(x)$ (say continuous, but that's not necessary, and nonnegative to be concrete). We think of $\int_{-\infty}^\infty f(x)\,dx$ as the area under the curve $y=f(x)$. Now write this as an iterated integral and then change the order of integration: $$\int_{-\infty}^\infty f(x)\,dx = \int_{-\infty}^\infty\int_0^{f(x)} 1\,dy\,dx = \int_0^\infty \mu(\{x: f(x)\ge y\}\,dy.$$ The $x$ cross-section at height $y$ is precisely the set of points $x$ where $f(x)\ge y$. Here $\mu(E)$ is the (Lebesgue) measure of $E\subset\Bbb R$.

Area by cross-section

$\endgroup$
1
$\begingroup$

For any point $\omega$ in the sample space, we have $$ \int_{\mathbb R \times \mathbb R} [1_{u\leq X_1(\omega) } - 1_{u \leq X_2(\omega)}][1_{v \leq Y_1(\omega)} - 1_{v \leq Y_2(\omega)}] \ du dv =(X_1(\omega) - X_2(\omega))(Y_1(\omega)-Y_2(\omega)) $$

[For example, if $X_1(\omega) > X_2(\omega)$ and $Y_1(\omega) > Y_2(\omega)$, then \begin{align} \int_{\mathbb R \times \mathbb R} [1_{u\leq X_1(\omega) } - 1_{u \leq X_2(\omega)}][1_{v \leq Y_1(\omega)} - 1_{v \leq Y_2(\omega)}] \ du dv &=\int_{\mathbb R \times \mathbb R}1_{X_2(\omega) < u \leq X_1(\omega)}.1_{Y_2(\omega) < v \leq Y_1(\omega)} \ dudv \\ &= \int_{X_2(\omega)}^{X_1(\omega)}\int_{Y_2(\omega)}^{Y_1(\omega)} 1 \ dudv \\ &=(X_1(\omega) - X_2(\omega))(Y_1(\omega)-Y_2(\omega))\end{align} It's not to hard to find analogous arguments for the cases where $X_1(\omega) \leq X_2(\omega)$ or $Y_1(\omega) \leq Y_2(\omega)$.]

Taking expectations on both sides gives the desired result.

$\endgroup$

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.