2
$\begingroup$

The author in a book uses this result.

If $X_n\to X$ in probability and $(X_n)$ is bounded in $L^2(\Omega)$ (that is, $E(X_n^2)<K$ for all $n$), then the sequence $(X_n)$ is uniformly integrable.

My question is:

If we only assume that $X_n\to X$ in probability and that $(X_n)$ is bounded in $L^1(\Omega)$, do we still have that $(X_n)$ is uniformly integrable?

When trying to prove this, I end up using the $L^2$ boundedness. I used that when using Hölder, I highlight when I use it, is it possible to solve it without using Hölder and proving the statement for $L^1$-boundedness, or is it false then? Or can I finish my proof without using Hölder and the $L^2$-boundedness?

Proof-attempt:

First I prove that for all n, and all $\epsilon$, there exists a $k$, such that $P(|X_n|>k)<\epsilon$. By using subsequences and a.s. convergence it is easy to see that $X$ must also be bounded by the same constant in as the sequence in $L^2(\Omega)$. Since $X$ is integrable there is a $k'$ such that $E[|X|1_{|X|>k'}]<\epsilon/2$, by choosing $k'$ bigger than 1, we have that $P(|X|>k')<\epsilon/2$. By convergence in probability, there is an $N$, so that if $n \ge N$, then $P(|X_n-X|>k')<\epsilon /2$. For all $i<N$, since $X_i$ is integrable, there are $k_i$ such that $P(|X_i|>k_i)<\epsilon$. Choose k = $\max\{2k',k_1,\ldots,k_{N-1}\}$. Then $P(|X_n|>k)<\epsilon, \forall n$.

Now I prove the statement. Given $\epsilon$ Since $X$ is integrable, there is a $\delta$, so that if $P(A)<\delta$, then $E[|X|1_A]<\epsilon /2$. By the above argument, there is $k_1$ so that for all $n$ $P(|X_n|>k_1)<\delta$, and there is a $k_2$ such that for all $n$, $P(|X_n|>k_2)<(\frac{\epsilon}{4K})^2$, let $k=\max\{k_1,k_2\}.$

Then for all $n$ we have that $$E[|X_n|1_{|X_n|>k}]\le E[|X|1_{|X_n|>k}]+E[|X-X_n|1_{|X_n|>k}].$$ The first term on the RHS is less than $\epsilon/2$, by the $\delta$, $k_1$. The second term on the RHS is also less than $\epsilon/2$ by Hölder: $$E[|X-X_n|1_{|X_n|>k}]\le E[|X_n-X|^2]^{0.5}P(|X_n|>k_2)^{0.5}\le(\|X_n\|_2+\|X\|_2)\frac{\epsilon}{4K}\le\epsilon /2.$$

$\endgroup$
3
  • 2
    $\begingroup$ See this question, noting that the $X_n$ in the answer converge to $0$ in probability: math.stackexchange.com/questions/1709355/… $\endgroup$ Aug 1 '16 at 16:36
  • $\begingroup$ @carmichael561 Thank you very much. $\endgroup$
    – user119615
    Aug 1 '16 at 17:04
  • $\begingroup$ @user119615 can you please tell me the name of the said book ? $\endgroup$ Jan 25 '19 at 14:50
1
$\begingroup$

Your proof under the $\mathbb L^2$-boundedness is correct. The first paragraph can be shortened a little bit more, noticing that by Markov's inequality, $$\tag{*} \mathbb P\left(\left|X_n\right|\gt k\right)\leqslant k^2\mathbb E\left[X_n^2\right]\leqslant K/k^2.$$

The second part can also be shortened, using Cauchy-Scharz inequality in order to bound $\mathbb E\left[\left|X_n\right|\mathbf 1\left\{\left|X_n\right|\gt k\right\} \right]$ by $ \sqrt{\mathbb E\left[X_n^2\right]} \cdot \sqrt{\mathbb P\left(\left|X_n\right|\gt k\right)} \overset{\mbox{by }(*) }{\leqslant}K/k $. This actually only needs boundedness in $\mathbb L^2$. The convergence in probability is not needed.

The answer given here by carmichael561 gives a counter-example where the assumption on boundedness in $\mathbb L^2$ is relaxed to boundedness in $\mathbb L^1$ (with convergence in probability to $0$) :

Let $\Omega=[0,1]$ with Lebesgue measure, and $X_n=n1_{[0,\frac{1}{n}]}$. Then the $X_n$ are bounded in $L^1$ but not uniformly integrable.

$\endgroup$
1
  • $\begingroup$ Thank you, very nice how you shortened the proof. $\endgroup$
    – user119615
    Aug 1 '16 at 20:40

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.