3
$\begingroup$

It is well-known that CDFs (Cumulative Distribution Functions) of one-dimensional random variables are Borel measurable. But does the same apply to CDFs of multi-dimensional random variables (rvecs)? It suffices, for my purposes, to consider finite dimensional rvecs.


Relevant definitions

Let $n$ be an integer $\geq 2$.

Call a probability measure over the Borel field on $\mathbb R^n$ an $n$-dimensional probability measure.

To every $n$-dimensional probability measure, $m$, define the following function $F:\mathbb R^n\rightarrow\mathbb R$, called $m$'s CDF: $F(x)=m\left((-\infty, x]\right)$ where $(-\infty, x]$ is the set of all $y$ in $\mathbb R^n$ such that $y\leq x$ component-wise.

An $n$-dimensional CDF is a function $F:\mathbb R^n\rightarrow\mathbb R$ that can be obtained as some $n$-dimensional probability measure's CDF.

$n$-dimensional CDFs are known to be characterized by certain properties.

$\endgroup$
3
$\begingroup$

For any $x \in \mathbb R^n$, consider the measurable box-like function $M_x(y) = m((-\infty,x]) \cdot \mathbb1_{[x,\infty)}(y)$. Here $\mathbb1_{[x,\infty)}(y)$ is the indicator function of "$y \ge x$ component-wise".

It should be easy to show that $F(y) = \sup_{x \in \mathbb Q^n} M_x(y)$ for any $y$ (take an increasing sequence of $x$ converging to $y$), and this makes $F$ the countable supremum of measurable functions.

$\endgroup$
  • 1
    $\begingroup$ Thank you. What a great idea, though i suspect your $F(y)$ fails to work, since it's left-continuous, whereas a CDF should be right-continuous. However, a slight modification will fix it: define $M_x(y) = m((-\infty,x])\cdot\mathbb1_{(-\infty,x]}(y) + (1+\epsilon)\cdot\mathbb1_{\mathbb R^n \setminus (-\infty,x]}(y)$ and $F(y) = \inf_{x\in\mathbb Q^n}M_x(y)$. $\endgroup$ – Evan Aad Aug 3 '12 at 15:01
  • 1
    $\begingroup$ @EvanAad Oops, of course you are correct! Thanks for the fix. How silly of me to try to write $(-\infty,x]$ as a countable union rather than a countable intersection. $\endgroup$ – Erick Wong Aug 3 '12 at 22:21
  • $\begingroup$ Actually the $\epsilon$ is redundant. $\endgroup$ – Evan Aad Aug 4 '12 at 5:38
1
$\begingroup$

Actually the Wikipedia properties are not sufficient. You also need n-monotonicity, which is defined here on pp. 265-266: https://books.google.fi/books?id=GjPUBwAAQBAJ&printsec=frontcover&hl=fi#v=onepage&q&f=false A more readable treatment is here: http://home.uchicago.edu/hickmanbr/uploads/PT_REVIEW_CH3.pdf but the monotonicity condition in Theorem 3 is too weak, use the n-monotonicity instead. However, in Theorem 2 the n-monotonicity is correctly stated for R^2: F(b, d) − F(b, c) − F(a, d) + F(a, c) ≥ 0 for all a < b, c < d.

Beware: the Springer link it seems to use left-continuous CDFs.

$\endgroup$

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.