3

Consider the following definitions.

Definition.(UNIFORM INTEGRABILITY). Consider a measurable set $\Omega\subseteq {\bf R}$ such that $\lambda(\Omega)>0$ and a sequence of measurable functions $u_n:\Omega\rightarrow {\bf R}$. We say that a sequence $(u_n)$ is uniformly integrable on $\Omega$ if $\lim_{M\rightarrow+\infty}{\rm sup}_{n\in{\bf N}}\int_{\{|u_n|>M\}}|u_n|=0$ (also known as the condition of "NO ESCAPE TO VERTICAL INFINITY").

Definition.(EQUI-INTEGRABILITY). Consider a measurable set $\Omega\subseteq {\bf R}$ such that $\lambda(\Omega)>0$ and a sequence of measurable functions $u_n:\Omega\rightarrow {\bf R}$. We say that a sequence $(u_n)$ is equi-integrable on $\Omega$ if for every $\varepsilon>0$ there exists $\delta>0$ such that for every measurable set $E\subseteq\Omega$ such that $\lambda(E)\leq\delta$ it follows that for every $n\in {\bf N}$ we have $\int_{E}|u_n|\leq\varepsilon$.

Question: Is it possible to construct a sequence of measurable functions $u_n:{\bf R}\rightarrow {\bf R}$ such that

A (weak version). Let $p>1$ be given.

(A.I) ${\rm sup}_{n\in {\bf N}}||u_n||_{{\rm L}^p({\bf R})}\leq C$ (so $(u_n)$ is bounded in ${\rm L}^p({\bf R})$),

(A.II) $\lim_{M\rightarrow+\infty}{\rm sup}_{n\in {\bf N}}\int_{\{|u_n|>M\}}|u_n|>0$ (so $(u_n)$ is not uniformly integrable, i.e. $(u_n)$ exhibits "escape to vertical infinity").

B (strong version).

(B.I) for every $1<p<+\infty$ we have ${\rm sup}_{n\in {\bf N}}||u_n||_{{\rm L}^p({\bf R})}\leq C_p<+\infty$ (so $(u_n)$ is bounded in ${\rm L}^p({\bf R})$ for every $1<p<+\infty$),

(B.II) $\lim_{M\rightarrow+\infty}{\rm sup}_{n\in {\bf N}}\int_{\{|u_n|>M\}}|u_n|=+\infty$ (so $(u_n)$ is not uniformly integrable, i.e. $(u_n)$ exhibits strong "escape to vertical infinity").

Discussion.

By Theorem B.104 in G. Leoni, A first course in Sobolev Spaces. 2nd ed.}, Graduate Studies in Mathematics 105, American Mathematical Society, Providence RI, 2017. (also cf. Theorem 2.29 in I. Fonseca and G. Leoni, Modern Methods in the Calculus of Variations: ${\rm L}^p$-spaces}, Studies in Advanced Mathematics, Springer, New York 2007.), under the assumption of boundedness in ${\rm L}^1({\bf R})$, equi-integrability is equivalent to uniform integrability.

If $\lambda(\Omega)<+\infty$, then condition A.(I) (by Holder's Inequality) implies that $(u_n)$ is equi-integrable on $\Omega$, and therefore $(u_n)$ is bounded in ${\rm L}^1(\Omega)$, see

A Question About Uniform Integrability

As a consequence, it is not possible to construct an example which satisfies both A.(I) and A.(II) (both B.(I) and B.(II), resp.) on domains $\Omega\subseteq {\bf R}$ of finite measure. See also:

uniform integrability does not imply $L_p$ boundednes

Compare also questions

Yet another definition of uniform integrability. Is it equivalent to the classical definition by G.A. Hunt?

Family of functions that are bounded in $L^1$ but *NOT* Uniformly Integrable

Show that if $p > 1$ and $\sup_{n \geq 1} |X_n| \in L^p$ then $\{X_n, n \geq 1\}$ is uniformly integrable.

Remark1. In the cited questions above, the notions of uniform integrability are defined in different ways, which can be confusing.

Remark2. If $(u_n)$ satisfies the condition $\lim_{M\rightarrow+\infty}{\rm sup}_{n\in {\bf N}}\lambda(|u_n|>M)=0$, then, by Holder's Inequality, we estimate $\int_{\{|u_n|>M\}}|u_n|\leq \Big(\int_{\{|u_n|>M\}}|u_n|^p\Big)^{1\over{p}}\lambda(|u_n|>M)^{1\over{p'}}$, where ${1\over{p}}+{1\over{p'}}=1$, and it follows that condition A.(I) implies $\lim_{M\rightarrow+\infty}{\rm sup}_{n\in {\bf N}}\int_{\{|u_n|>M\}}|u_n|=0$, so A.(II) can not be satisfied.

Remark3. If $(u_n)$ satisfies the condition $\lim_{M\rightarrow+\infty}{\rm sup}_{n\in {\bf N}}\lambda(|u_n|>M)<+\infty$, then, by Holder's Inequality, we estimate $\int_{\{|u_n|>M\}}|u_n|\leq \Big(\int_{\{|u_n|>M\}}|u_n|^p\Big)^{1\over{p}}\lambda(|u_n|>M)^{1\over{p'}}$, where ${1\over{p}}+{1\over{p'}}=1$, and it follows that condition B.(I) implies $\lim_{M\rightarrow+\infty}{\rm sup}_{n\in {\bf N}}\int_{\{|u_n|>M\}}|u_n|<+\infty$, so B.(II) can not be satisfied.

Conclusion: B.(II) can not be satisfied, even under assumption (A.I), which is weaker then the assumption B.(I). To see that this is the case, for $M\geq 1$ and for a given $p>1$ we estimate $$ {\rm sup}_{n\in {\bf N}}\lambda(|u_n|>M)={\rm sup}_{n\in {\bf N}}\int_{|u_n|>M}1\leq {\rm sup}_{n\in {\bf N}}\int_{|u_n|>M}|u_n|\leq {\rm sup}_{n\in {\bf N}}\int_{|u_n|>M}|u_n|^p\;. $$ Hence, under assumption (A.I), it holds that $\lim_{M\rightarrow+\infty}{\rm sup}_{n\in {\bf N}}\lambda(|u_n|>M)<+\infty$, and therefore (B.II). can not be satisfied. So the strong version can not be constructed. On the other hand, to satisfy the conditions of the weak version, it is necessary that $(u_n)$ satisfies the condition $$\lim_{M\rightarrow+\infty}{\rm sup}_{n\in {\bf N}}\lambda(|u_n|>M)>0\;.$$

Remark4. If $p>0$, the sequence $u_n(s):=\sum_{k=1}^{n}2^k\chi_{(2^k,2^k+2^{-np})}(s)$, where $s\in {\bf R}$, satisfies: $$ ||u_n||_{{\rm L}^p({\bf R})}^p\leq \Big(2^{p}-1\Big)+\sum_{j=0}^{+\infty}(2^{-p})^j\;, $$

$$ ||u_n||_{{\rm L}^1({\bf R})}=2^{-np}\Big(2^{n+1}-1\Big)\;, $$

$$ \int_{|u_n|>M}|u_n|\geq\int_{\cup_{i=M_0}^{n}\{\sigma\in (2^{i}, 2^{i}+2^{-np}):|u_n(\sigma)|>M\}}|u_n|=\sum_{i=M_0}^{n}\int_{2^i}^{2^i+2^{-np}}2^i=2^{n(1-p)+1}-2^{M_0-np}\;, $$

where $M>2$ and $M_0>{\rm log}_{2}M>1$. Therefore, $(u_n)$ provides the example of a sequence which is bounded in ${\rm L}^p({\bf R})$ for $0<p<1$, but which is not uniformly integrable on ${\bf R}$.

Andrija
  • 587

1 Answers1

3

Regarding question (A).

Suppose $M=\sup_n\|f_n\|_p<\infty$, for some $1<p<\infty$. An application of Hölder's inequality followed my Markov-Chebyshev's inequality yields $$\int_{\{|f_n|>a\}} |f_n|\,d\mu\leq\|f_n\|_p(\mu(|f_n|>a))^{1-1/p}\leq\|f_n\|_p\Big(\frac1{a^p}\|f_n\|^p_p\Big)^{1-1/p}=\|f_n\|^p_p\frac{1}{a^{p-1}}\leq \frac{M^p}{a^{p-1}}$$ Hence $$\lim_{a\rightarrow\infty}\sup_n\int_{\{|f_n|>a\}}|f_n|\,d\mu\xrightarrow{a\rightarrow\infty}0$$ There is no escape to vertical infinity. In general, however, this does not imply uniform integrability in the usual sense (Hunt, G. A. (1966). Martingales et Processus de Markov. Paris: Dunod. p. 254).

Some examples:

Example 1 (Yuval Peres): Choose $f\in L^+_p(\mathbb{R},m)\setminus\{0\}$, $1<p<\infty$, where $m$ is Lebesgue's measure and define $f_n=f(\cdot-n)$. Since $\|f_n\|_p=\|f\|_p$ for all $n$, $(f_n:n\in\mathbb{N})$ has no escape to vertical infinity. However, it is not uniform integrable in the usual sense. In fact, the sequence may not even be in $L_1$.

Example 2: (continuation of Example 1): Now suppose that $f\in L^+_1\cap L_p\setminus\{0\}$ so that the sequence of translations $(f_n:n\in\mathbb{N})\subset L_1$ and $\|f\|_1>0$. Consider the density function $\phi$ of the standard normal distribution. For any $m\in\mathbb{N}$, there is a closed symmetric interval $I_m=[-a_n,a_m]$ such that $\int_{I^c_m}\phi=\frac1m$. Notice that for such $I_m$, \begin{align} \int_{I^c_m}f_n&=\int^\infty_{a_m}f(x-n)\,dx+\int^{-a_m}_{-\infty}f(x-n)\,dx\\ &=\int^\infty_{a_m-n} |f(x)|\,dx+\int^{-a_m-n}_{-\infty}f\,dx\xrightarrow{n\rightarrow\infty}\int_{\mathbb{R}} f>0 \end{align} This shows that $(f_n:n\in\mathbb{N})$ is not uniformly integrable (see Theorem A below). In this example, the issue comes from the fact that the underlying measure space has infinite measure and that the sequence is not tight enough, that is $\inf_{K\in\mathcal{K}}\sup_n\int_{\mathbb{R}\setminus K}|f_n|\,dx>0$ where $\mathcal{K}$ is the collection of compact sets in $\mathbb{R}$.

Example 3: When $p>1$ and the underlying measure space is finite there are no issues since in this setting, uniform integrability of a family $\mathcal{F}\subset L_1$ is equivalent to boundedness in $L_1$ and no escape to vertical infinity. If $p=1$ however, there may be issues. Consider the standard Leebsgue space $([0,1],\mathscr{B}([0,1]),m)$, and define the sequence $X_n=2^{n}\mathbb{1}_{[0,2^{-n}]}$. Notice that $\|X_n\|_1=1$, for all $n\in\mathbb{N}$. Now, for $a>0$, \begin{equation} \int_{\{|X_n|>a\}}|X_n|\,dx=\left\{\begin{array}{lcr} 0 &\text{if} & n\leq \log_2(a)\\ 1 &\text{if} & n>\log_2(a) \end{array} \right. \end{equation} Hence, for $a>4$, $$\sup_n\mathbb{E}\big[\mathbb{1}_{\{|X_n|>a\}}|X_n|\big]=1\not\rightarrow0\quad \text{as}\quad a\rightarrow\infty$$ In this case, $(X_n:n\in\mathbb{N})$ fails to be uniformly integrable.


Some basic results:

Theorem A: Suppose $(X,\mathscr{B},\mu)$ is a $\sigma$-finite measure space and let $h\in L_1$ is strictly positive. A family $\mathcal{F}\subset L_1$ is uniformly integrable iff $\mathcal{F}$ is bounded in $L_1$ and for any $\varepsilon>0$, there is $\delta>0$ such that $$\sup_{f\in\mathcal{F}}\int_A|f|\,d\mu<\varepsilon\qquad\text{whenever}\qquad \int_Ah\,d\mu<\delta$$

There are several equivalences of uniform integrability in the sense of Hunt. One that is most commonly used in analysis is the following

Proposition T: A family $\mathcal{F}\subset\mathcal{L}_1$ is uniformly integrable iff

  1. $\sup_{f\in\mathcal{F}}\|f\|_1<\infty$
  2. $\displaystyle \inf_{a>0}\sup_{f\in \mathcal{F}}\int_{\{|f|>a\}}|f|\,d\mu=0$
  3. $\displaystyle \inf_{A\in\mathscr{B}:\mu(A)<\infty}\int_{X\setminus A}|f|\,d\mu=0$

The original definition by Hunt is

Definition H: A family $\mathcal{F}\subset L_1$ is uniformly integrable iff $$ \inf_{g\in L^+_1}\sup_{f\in\mathcal{F}}\int_{\{|f|>g\}}|f|\,d\mu=0$$

For a proof Proposition T here are some postings: (A) last proposition, (B) definition 1 and theorem 3 and (C).

Mittens
  • 46,352