I'm reading: https://en.wikipedia.org/wiki/Convergence_of_random_variables#Almost_sure_convergence and here it says that

Given a probability space $(\Omega,\mathcal{F},P)$ and a random variable $X:\Omega \rightarrow \mathbb{R}$ almost sure convergence stands for $$P\left(\omega \in \Omega: \lim_{n \rightarrow \infty} X_n(\omega)=X\right)=1.$$ [...] almost sure convergence can also be defined as follows: $$P\left(\limsup_{n \rightarrow \infty} \left\{\omega \in \Omega: |X_n(\omega) - X(\omega)| > \varepsilon\right\}\right)=0, \quad \forall \; \varepsilon>0.$$

My question is, what is the intuition behind this equivalence? I understand the first definition, but why do we use $\limsup$ in the second one to make the equivalence work? Thanks


I don't really see intuition here, the equivalence just follows from using the definition of convergence. For a sequence of sets $(A_n)$ the set $\lim \sup(A_n)=\{A_n\ \ i.o\}$ is the set of elements which belong to infinitely many of the sets $A_n$. The formal definition of this set is $\cap_{n=1}^\infty \cup_{k=n}^\infty A_k$.

Assume $X_n\to X$ almost surely by the first definition and let any constant $\epsilon>0$. Define the sequence $A_{n,\epsilon}:=\{\omega: |X_n(\omega)-X(\omega)|>\epsilon\}$. Note that if $\omega\in\lim\sup A_{n,\epsilon}$ then it means that $|X_n(\omega)-X(\omega)|>\epsilon$ for infinitely many values of $n$, and hence $X_n(\omega)$ obviously does not converge to $X(\omega)$. So $\lim\sup A_{n,\epsilon}\subseteq \{\omega: X_n(\omega)\nrightarrow X(\omega)\}$, and by monotonicity of probability:

$\mathbb{P}(\lim\sup A_{n,\epsilon})\leq \mathbb{P}(\{\omega: X_n(\omega)\nrightarrow X(\omega)\})=0$

Second direction: Now assume $X_n\to X$ by the second definition. For each $k\in\mathbb{N}$ define $B_k=\lim\sup A_{n,\frac{1}{k}}$ where the sets $A_{n,\epsilon}$ are defined like before. Then by assumption $\mathbb{P}(B_k)=0$ for all $k$, and hence $\mathbb{P}(\cup_{k=1}^\infty B_k)=0$. Now suppose we have $X_n(\omega)\nrightarrow X(\omega)$ for some $\omega$. This implies that there must be some $m\in\mathbb{N}$ such that $|X_n(\omega)-X(\omega)|>\frac{1}{m}$ for infinitely many natural numbers $n$, and thus $\omega\in B_m\subseteq\cup_{k=1}^\infty B_k$.

In other words, we have the inclusion $\{\omega: X_n(\omega)\nrightarrow X(\omega)\}\subseteq\cup_{k=1}^\infty B_k$, and so $\mathbb{P}(\{\omega: X_n(\omega)\nrightarrow X(\omega)\})=0$.

  • $\begingroup$ Thank you so much $\endgroup$
    – Ejrionm
    Aug 3 '20 at 1:54


There is not much intuition to be gleaned here. The second definition comes from "massaging" the definition of the [non-random] limit of real numbers (since for a fixed $\omega$, the limit $\lim_{n \to \infty} X_n(\omega)$ is just a non-random limit).

The utility of the second definition is that it is easier to verify because it involves relatively simple sets $\{|X_n(\omega) - X(\omega)| > \epsilon\}$ (fixed $\epsilon$, fixed $n$). You only need to deal with one $n$ at a time to understand this set, and under certain circumstances, bounding the probability of this set for each $n$ can be enough to bound probability of the $\limsup$. By contrast, the set $\{\lim_{n \to \infty} X_n(\omega) = X(\omega)\}$ is difficult to deal with because of the limit inside the event.


Let $A_{n, \epsilon} = \{|X_n(\omega) - X(\omega)| > \epsilon\}$. Note that $$\limsup_{n \to \infty} A_{n, \epsilon} := \bigcap_n \bigcup_{k \ge n} A_{k,\epsilon}$$ by definition.

(1) $\implies$ (2)

Fix $\epsilon > 0$. If $\omega \in \bigcap_n \bigcup_{k \ge n} A_{k, \epsilon}$, then $|X_n(\omega) - X(\omega)| > \epsilon$ for infinitely many $n$, so $\lim_n X_n(\omega) \ne X(\omega)$. Thus $$P(\limsup_n A_{n, \epsilon}) \le P(\lim_n X_n(\omega) \ne X(\omega))$$ for each $\epsilon$. So if almost sure convergence holds in the sense of the first definition, then it holds in the sense of the second definition.

(2) $\implies$ (1)

Conversely, suppose $\omega$ is such that $\lim_n X_n(\omega) \ne X(\omega)$. If you write out the definition of a limit, this means there exists some $\epsilon$ such that $|X_n(\omega) - X(\omega)| > \epsilon$ for infinitely many $n$. That is, there exists $\epsilon$ such that $\omega \in \bigcap_n \bigcup_{k \ge n} A_{k, \epsilon}$. Then $$P(\limsup_n A_{n, \epsilon}) \ge P(\lim_n X_n(\omega) \ne X(\omega))$$ for this particular $\epsilon$. So if almost sure convergence holds in the sense of the second definition, it also holds in the sense of the first definition.


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.