7
$\begingroup$

Background. Let $V = (X,Y)$ be a random vector in 2-dimensions uniformly distributed over two disjoint regions $R_X \cup R_Y$ defined as follows:

$$ \begin{align} R_X &= ([0,1] \times [0,1]) \setminus \left(\bigcup A\right) \\ R_Y &= [1,2] \times [1,2], \end{align} $$

where $$ A = \{[.2,.4] \times [.2,.4], [.2,.4] \times [.6,.8], [.6,.8] \times [.2,.4], [.6,.8] \times [.6,.8]\}. $$ A plot of the regions (green is the density/area associated with each region) is:

enter image description here

I'm interested in how to find $E(XY)$. So, letting $\lambda$ be the Lebesgue measure the associated pdf is

$$ f_{XY}(x,y) = \begin{cases} \frac{1}{\lambda\left(R_X \cup R_Y\right)} = \frac{1}{\lambda([0,1] \times [0,1]) - \lambda\left(\bigcup A\right) + \lambda([1,2] \times [1,2])} \approx \frac{25}{46}, &(x,y) \in R_X\cup R_Y \\ 0, &\text{otherwise} \end{cases} $$

Using the traditional definition

$$E(XY) = \int_{R_X\cup R_Y} xyf_{XY}(x,y)d\lambda = \frac{25}{46}\int_{R_X\cup R_Y} xyd\lambda = \frac{25}{46}\left(\int_{R_X} xydxdy + \int_{R_Y} xydxdy\right).$$

Integrating over $R_Y$ is straightforward. But for the "non-simple" region $R_X$ would we calculate it as

$$\int_{R_X} xydxdy = \int_0^1\int_0^1 xydxdy - \sum_{a \in A} \int_a xydxdy~\text{?} \tag{1}$$

What if there were countably many boxes to remove from $R_X$ instead of 4 finite ones? Does the formula for (1) generalize (I'm assuming here we could use something like the MCT/DCT/etc. to evaluate the sum).


EDIT (after @whuber answer): Quick follow-up points:

  1. It's apparent now that $E(X)$ could be found using this mixture approach too, i.e.

$$\sum_i \omega_i p_i = \frac{21}{25}$$

and

$$E(X) = \frac{25}{21}\left(\frac{1}{2} - \frac{9}{2500} - \frac{21}{2500} - \frac{21}{2500} - \frac{49}{2500}\right) = \frac{23}{42}.$$

  1. For the case where instead of 4 finite regions removed, we have countably many, then we can generalize as

$$E(XY) = \frac{25}{46}\left(\iint_{[0,1]\times [0,1]} xydxdy + \iint_{[1,2]\times [1,2]} xydxdy - \sum_{j}^{\infty}g(x)\mathscr{I}_j(x)\right) = \frac{25}{46}\left(\frac{1}{4} + \frac{9}{4} - \sum_{j}^{\infty}g(x)\mathscr{I}_j(x)\right),$$

for $j > i$ and assuming that $\sum_{j}^{\infty}g(x)\mathscr{I}_{j}(x) < \infty$ (i.e. converges).

$\endgroup$
4
$\begingroup$

Let's solve the more general problem and then apply it to the specific setting as an illustration.

Suppose $g:\mathcal{X}\to\mathbb{R}$ is a measurable function. Let $\{\mathcal{A}_i\}\subset \mathscr{P}(\mathcal{X})$ be a finite or countable collection of subsets of $\mathcal X,$ each with finite positive measure $p_i = \int_{\mathcal{A}_i}\mathrm{d}x.$ Associated with each $\mathcal A_i$ is its indicator function $\mathscr{I}_i.$ Then for any sequence of numbers $(\omega_i)$ -- essentially by construction of the integral -- we have that

$$\int_\mathcal{X} \sum_i g(x)\omega_i\mathscr{I}_i(x)\,\mathrm{d}x = \sum_i \omega_i \int_{\mathcal{A}_i} g(x)\,\mathrm{d}x = \sum_i \omega_i p_i \int_{\mathcal{A}_i} g(x)\,\frac{\mathrm{d}x}{|p_i|}$$

When the $\omega_i$ are positive and the sum of $\omega_i p_i$ is unity, the right hand side is the expectation of $g(X)$ where the distribution of $X$ is a mixture of the uniform distributions on the $\mathcal{A}_i$ with mixture weights $\omega_i p_i.$ I will continue to use this imagery and this language even when some of the $\omega_i$ are negative. Think of this as a "generalized mixture" if you like.

Provided the left hand integrand is never negative and is positive on some set of positive area, we may normalize it to produce a genuine distribution. Evidently its density function is

$$f(x;(\mathcal{A}_i), (\omega_i)) = \frac{1}{\sum_i \omega_i p_i} \sum_i \omega_i g(x) \mathscr{I}_i(x).\tag{*}$$

The first formula therefore gives the expectation $E[g(Z)]$ when $Z$ has this mixture distribution. The right hand side of the formula shows that this expectation is a linear combination of the expectations of the mixture components.


To apply this observation to the example in the question, let $\mathcal X =\mathbb{R}^2$ (with its usual Borel measure) and $g(x,y) = xy.$ Let's begin by getting all the calculations out of the way. They amount to integrating $g$ over various rectangles $[a,b]\times[c,d].$ It is elementary to compute that

  1. The area of a rectangle $[a,b]\times[c,d]$ is $p(a,b,c,d)=(b-a)(d-c).$

  2. $$G(a,b,c,d)=\iint_{[a,b]\times[c,d]}xy\,\mathrm{d}x\mathrm{d}y = \frac{1}{4}(b^2-a^2)(d^2-c^2) = p(a,b,c,d)(a+b)(c+d)/4.$$

The problem can be expressed in terms of six rectangles: the two big ones (which therefore receive weights $\omega_i=1$) from which the four little ones have been removed (by applying weights $\omega_i=-1$). Here is a table of their properties, computed using $(1)$ and $(2)$ above.

$$ \begin{array}[llrrrl] & i & [a,b]& [c,d] & p & G & \omega \\ \hline 1 & [0,1]& [0,1] & 1 & 1/4 & 1 \\ 2 & [1,2]& [1,2] & 1 & 9/4 & 1 \\ 3 & [1/5,2/5] & [1/5,2/5] & 1/25 & 9/100 & -1 \\ 4 & [3/5,4/5] & [1/5,2/5] & 1/25 & 21/100 & -1\\ 5 & [1/5,2/5] & [3/5,4/5] & 1/25 & 21/100 & -1\\ 6 & [3/5,4/5] & [3/5,4/5] & 1/25 & 49/100 & -1 \end{array} $$

The denominator in $(*)$ is

$$\sum_{i=1}^6 \omega_i p_i = 1 + 1 - \frac{1}{25} - \cdots - \frac{1}{25} = \frac{46}{25}.$$

Writing $Z=(X,Y),$ the answer to the question is

$$E[XY] = E[g(Z)] = \frac{25}{46}\left(\frac{1}{4} + \frac{9}{4} - \frac{9}{2500} - \frac{21}{2500} - \frac{21}{2500} - \frac{49}{2500}\right)=\frac{123}{92}.$$

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Thank you. Given your answer, I have 2 quick follow-up questions that I have edited into the OP. Could you verify real quick that my thinking for these is correct? $\endgroup$ – EzioBosso Oct 17 at 21:23
  • 1
    $\begingroup$ Those additional questions appear to be direct applications of this general result. $\endgroup$ – whuber Oct 17 at 22:33

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.