MathOverflow is a question and answer site for professional mathematicians. It's 100% free, no registration required.

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

There is probably terminology for this, and I apologize that I don't know it, and part of my question is what the standard terminology for the concepts I'm giving is. This is a pretty open-ended question.

Let $C$ be some class of random variables, on the the same or different probability spaces, taking values in some topological space $V$ (the case that interests me most is $R^\omega$).

Let $D_C$ be the set of distributions of the variables in $C$, i.e., equivalence classes under probability isomorphism ($X:\Omega\to V$ and $Y:\Upsilon\to V$ are equivalent iff there is a probability isomorphism $\alpha:\Omega\to\Upsilon$ such that $X = Y\circ \alpha$).

Say that $C$ is "single-point recoverable" provided that there is a function $f:V\to D_C$ such that for every $X\in C$, $f(X)=[X]$ with probability one. I.e., for all $X\in C$, if $X:\Omega\to V$, then $f(X(\omega)) = [X]$ for almost all $\omega\in\Omega$.

Question 1: Is there a standard term for "single-point recoverable"?

Fact: If $V=\mathbb R^\infty$ (with product topology and a Borel $\sigma$-algebra) and $C_n$ is the class of all vector-valued random variables $X$ of the form $X=(X_i)_{i\in \mathbb N}$ where the $X_i$ are i.i.d. and have values in $R^n$, then $C_n$ is single-point recoverable. (Proof sketch: Rectangles are a Vapnis-Chervonenkis class, and we can a.s. recover the distribution of $X_1$, and hence of $X$, via a uniform strong LLN by using the frequencies of visits of $X_i(\omega)$ to rectangles.)

Question 2: Is there a standard name for the Fact? An elementary proof? (I may be missing something completely obvious.)

(The Fact is a sort of partial vindication of frequentism: It shows that if we know that an infinite sequence of observations was generated by an i.i.d. sequence, a.s. we can recover the distribution from the observations.)

Question 3: Are there any interesting generalizations of the Fact? For instance, if $E$ is the class of all sequences of $\mathbb R^n$-valued i.r.v.'s, then $E$ isn't single-point recoverable. But is there any result like this: There is a function $f:\mathbb R^\infty \to D_C$ such that for all $X\in E$, for almost all $\omega$, $f(X(\omega))$ is "asymptotically close" to $[X]$, in some precise sense? (Or at least for some subclass of $E$.) Is there any result that weakens independence?

share|cite|improve this question
up vote 2 down vote accepted

Regarding Question 1, I think the notion you are looking for is that $C$ consists of random variables having mutually singular distribution measures.

Regarding Question 2 (standard name for the Fact), I would say just Strong Law of Large Numbers, applied to the probabilities of events which determine the distribution.

share|cite|improve this answer
Thanks!Yeah, SLLN is all one needs in Q1: just take rectangles with rational corner coordinates. I missed that because the philosophical application (a defense of frequentism against certain objectins) I was thinking of made it inappropriate to single out a particular set of rectangles in the recovery. – Alexander Pruss Apr 26 '14 at 14:24

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.