In a measure space with finite total measure, a family $A$ of r.v's is called U.I if

$$ \lim_{N\to\infty}\sup_{X\in A} \int_{|X|>N} |X|\,d\mu=0 $$

I have some doubts on equivalance of this definition in $L^1$ space. The following is equivalent to U.I in $L^1$:

  1. A is bounded subset of $L^1$.
  2. For every $\epsilon>0$ there exists $\delta>0$ such that for any measurable $E$ if $\mu(E)\le \delta$ then $\int_E |X| d\mu\le \epsilon$ for any $X\in A$.

The following proof is given:

Proof: Suppose $A$ is U.I. Then for any measurable $E$,

$$ \int_E |X| = \int_{E\cap \{|X|>N\}} |X| + \int_{E\cap \{|X|\le N\}} |X|:=B+C $$

Given $\epsilon>0$ we can choose $N$ such that $B<\epsilon /2$ and obviously $C\le N\mu (E)$. Thus we may choose $\delta = \frac{\epsilon}{2N(\mu(\Omega)+1)}$.

And I'm not sure why we can't simply choose $\delta=\epsilon/2N$? And why do we need the condition that $\Omega$ is a finite measure space? I don't particularly see any necessity for it, unless I'm missing something about last line. Probably only place where it might matter is stating $C\le N\mu(E)$, since it might be that $\mu(E)=\infty$. But we are controlling size of $E$ anyway, aren't we?

Edit: It seems for condition $1$ to hold, $E$ must be of finite measure so total measure must be finite, as we only control size of $E$ for proof of $2$.

  • 1
    $\begingroup$ Aren't you confusing uniform continuity with uniform integrability? $\endgroup$ Jun 4 '16 at 2:02
  • $\begingroup$ I don't quite understand what you mean, for which part do you think I'm talking about uniform continuity? It might be that I don't know what that is for a family of r.v, I only encountered that concept for a single function. $\endgroup$
    – user340297
    Jun 4 '16 at 2:06
  • 1
    $\begingroup$ You are right; there is nothing wrong about choosing $\delta = \epsilon/(2N)$. So, uniform integrability implies condition 2, but in order to show condition 1 (i.e. boundedness in $L^1$) you need the finiteness of the measure. $\endgroup$
    – saz
    Jun 4 '16 at 6:08

It seems that the choice $\delta=\varepsilon/(2N)$ works well.

Finiteness of the measure space is used to show boundedness in $\mathbb L^1$ of a uniformly integrable family. Indeed, if the measure space is infinite, then a constant function is not integrable. Therefore, we cannot show that the family $\left(\int \left|X\right|\mathbf 1\{\left|X\right|\lt R\}\right)_{X\in A}$ is bounded. For example, assume that we work with the real line with Lebesgue measure and $X_n=\mathbf 1([n,2n))$. This family would be uniformly integrable if we consider the standard definition, but this is not bounded in $\mathbb L^1$.

What can be misleading in the statement in the opening post is that we talk about random variable, which supposes that we work with a probability space. In particular, there is no need to specify that the measure is finite.

Note that the concept of uniform integrability can be extended to infinite measure spaces. A family of functions $\left(f_i\right)_{i\in I}$ is uniformly integrable if for each positive $\varepsilon$, we can find an integrable function $g$ such that $$\sup_{i\in I}\int\left|f_i\right|\mathbf 1\left\{\left|f_i\right|\gt g\right\}\lt \varepsilon.$$


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.