$\newcommand{\N}{\mathbb{N}}\newcommand{\P}{\mathrm{P}}\newcommand{\esssup}{\mathrm{esssup}}$Let $(Z_j)_{j\in\N}$ be a nonnegative stochastic process ($Z_j\geq 0$ $\P$-a.s) for which the following holds: $$ \esssup\left(\sum_{j=0}^{\infty}Z_j\right)\leq M\tag{1}\label{1}, $$ for some $M\geq 0$.

Using the definition of essential supremum we have that the sequence is almost surely summable: $$ \begin{align} &\P\left[ \{\omega\in\Omega: \sum_{j=0}^{\infty}Z_j > M\}\right] = 0 \\ \Leftrightarrow\ &\P\left[ \{\omega\in\Omega: \sum_{j=0}^{\infty}Z_j \leq M\}\right] = 1 \end{align}$$ and since $\{\omega\in\Omega: \sum_{j=0}^{\infty}Z_j(\omega) \leq M\}\subseteq \{\omega\in\Omega: Z_j(\omega)\to 0\}$, we have that \eqref{1} implies $\P[\{\omega: Z_j(\omega)\to 0\}]=1$, that is, $Z_j\to 0$ a.s., and, a fortiori, $Z_j\to 0$ in probability.

I wonder whether we can prove that $\esssup(Z_j)\to 0$.

If not, there will be a $\delta>0$ and an increasing sequence of integers $(j_k)_k$ so that $\esssup(Z_{j_k})>\delta$ for all $k\in\N$. Since $Z_j$ are nonnegative, I guess (but I am not sure) that we can show that $\esssup(\sum_{k\in\N}Z_{j_k})>\delta$ which is a contradiction. Is this correct? If not, is there a counterexample where \eqref{1} holds but $\esssup(Z_j)$ does not converge to $0$?


No, this need not be the case. For instance, it is possible that with probability $1$, there exists $j$ such that $Z_j=M$ and $Z_k=0$ for all $k\neq j$. For an explicit example, let $X$ be a Poisson random variable and define $Z_j:=\mathbf1_{\{X=j\}}$ for each $j$. Then $\sum_{j=0}^\infty Z_j=1$ almost surely, but $\operatorname{esssup}(Z_j)=1$ for all $j$, so in particular it does not converge to zero.

Note that the conclusion you are trying to draw - that $Z_j\to0$ in $L^\infty$ - is extremely strong. This is rarely something we would try to prove in probability theory.

EDIT: As we have discussed in the comments, it is true that $Z_j\to0$ in $L^p$ for any $1\le p<\infty$. To see this, it is sufficient to show that there is a constant $C$ such that $\sum_jZ_j^p\le C$ almost surely, since this would imply $\sum_j\mathbb E[Z_j^p]=\mathbb E\left[\sum_jZ_j^p\right]\le C$ by the monotone convergence theorem, which of course shows that $\mathbb E[Z_j^p]\to0$.

Let $A=\{j\in\mathbb N:Z_j\le1\}$ and $B=\mathbb N\setminus A$. Notice that $Z_j^p\le Z_j$ for all $j\in A$, $Z_j\le M$ for all $j\in B$ almost surely, and $|B|\le\sum_{j\in B}Z_j\le M$. This implies

$$\sum_jZ_j^p=\sum_{j\in A}Z_j^p+\sum_{j\in B}Z_j^p\le\sum_{j\in A}Z_j+M\cdot M^p\le M+M^{p+1}$$

almost surely, completing the proof.

  • $\begingroup$ Thanks a lot for the answer. Is it true, however, that $Z_j\to 0$ a.s.? Given (1), can we deduce anything additional about $(Z_j)_j$? (perhaps under additional assumptions?) $\endgroup$ – Pantelis Sopasakis Aug 8 '17 at 21:23
  • $\begingroup$ Yes, certainly - $\sum_jZ_j\le M$ almost surely, so clearly $Z_j\to0$ almost surely. As far as other things $(1)$ implies, you can for instance show that $Z_j\to0$ in $L^p$ for all $1\le p<\infty$. Try to think about how you might show this for $p=1$ (this is the easiest case). $\endgroup$ – Jason Aug 8 '17 at 21:31
  • $\begingroup$ For $p=1$ I would start by $\mathbb{E}(\sum_j Z_j) \leq \mathrm{esssup}(\sum_j Z_j) \leq M$ and use the linearity of $\mathbb{E}$. $\endgroup$ – Pantelis Sopasakis Aug 8 '17 at 21:40
  • $\begingroup$ Right. You technically also need the monotone convergence theorem since it's an infinite series, but you get the idea. For $p>1$, it's the same principle - we just need to find an (almost sure) upper bound for $\sum_jZ_j^p$, which i'll leave you to think a little bit about. $\endgroup$ – Jason Aug 8 '17 at 21:45
  • $\begingroup$ For $p=1$ we also have the Beppo-Levi Theorem which applies directly. Let me think about the $L^p$ convergence case... $\endgroup$ – Pantelis Sopasakis Aug 8 '17 at 21:52

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.