This question comes from Bass, Ex 4.15. Given a finite measure space $(X, \mathcal{A}, \mu)$, we can define an outer measure $\mu^*$ as $$ \mu^*(A) := \inf\{ \mu(B) : A \subset B , B \in \mathcal{A} \} \:. $$ (One can check that this is an outer measure and that $\mu^*$ agrees with $\mu$ on $\mathcal{A}$).

I want to show now that if $\{A_n\}_{n \geq 1}$ is a sequence of increasing sets (not necessarily in $\mathcal{A}$) such that $A_n \nearrow A$, then $\mu^*(A_n) \nearrow \mu^*(A)$.

I have this so far. By monotonicity of outer measure, $\mu^*(A_n) \leq \mu^*(A)$ for all $n$, hence $\lim_{n \rightarrow \infty} \mu^*(A_n) \leq \mu^*(A)$. It remains to show that $\mu^*(A) \leq \lim_{n \rightarrow \infty} \mu^*(A_n)$.

My rough idea is as follows. Fix any $\epsilon > 0$. We can pick a sequence of sets $\{G_i\}_{i \geq 1}$ in $\mathcal{A}$ such that $\mu^*(A_i) \geq \mu(G_i) - \epsilon/2^{i}$. If the $\{G_i\}$ were an increasing set then we'd be done by using the continuity of measure from below. However, they are not. But, we can define $H_n := \bigcup_{i=1}^{n} G_i$ which is an increasing sequence by definition. Since $A \subset \bigcup_{n \geq 1} H_n$, we have that $$ \mu^*(A) \leq \mu( \bigcup_{n \geq 1} H_n ) = \lim_{n \rightarrow \infty} \mu(H_n) \:. $$ Now I am not quite sure how to finish off. I want to say something to the effect of $\mu(H_n) \leq \mu^*(A_n) + \epsilon$. However, since the $A_n$'s are not measurable, I am having trouble making such a comparison.

  • $\begingroup$ Your idea is essentially correct. You just need to define $H_n=\bigcap_{i=n}^\infty G_i$, not as $H_n=\bigcup_{i=1}^{n} G_i$. See my answer below for details. $\endgroup$ – Ramiro Dec 16 '15 at 11:38

As you wrote, you have already proved $\lim_{n \rightarrow \infty} \mu^*(A_n) \leq \mu^*(A)$. It remains to show that $\mu^*(A) \leq \lim_{n \rightarrow \infty} \mu^*(A_n)$.

Your idea to complete the proof is essentially correct, all it needs is a small adjustment.

Given any $\epsilon > 0$. We can pick a sequence of sets $\{G_i\}_{i \geq 1}$ in $\mathcal{A}$ such that $A_i\subseteq G_i$ and $\mu^*(A_i) \leq \mu(G_i) \leq \mu^*(A_i) + \epsilon/2^{i}$.

Define $H_n=\bigcap_{i=n}^\infty G_i$. Then $\{H_n\}_{n\geq 1}$ is an increasing sequence of sets in in $\mathcal{A}$.

For any $n \geq 1$, if $i \geq n$ then $A_n\subseteq A_i \subseteq G_i$. So, for any $n \geq 1$, $A_n \subseteq \bigcap_{i=n}^\infty G_i=H_n$. So we have $A_n \subseteq H_n\subseteq G_n$ and we get $$ \mu^*(A_n) \leq \mu(H_n) \leq \mu(G_n) \leq \mu^*(A_n) + \epsilon/2^{n}$$

So we have $\lim_{n \to +\infty} \mu^*(A_n) = \lim_{n \to +\infty} \mu(H_n)$.

Since $A_n \nearrow A$, it is easy to see that that $A\subseteq \bigcup_{n\geq 1} H_n$. But $H_n \nearrow \bigcup_{n\geq 1} H_n$, so using the the continuity of measure from below, we have
$$\mu^*(A)\leq \mu\left(\bigcup_{n\geq 1} H_n\right)=\lim_{n \to +\infty} \mu(H_n)= \lim_{n \to +\infty} \mu^*(A_n)$$

  • $\begingroup$ Une jolie preuve ! $\endgroup$ – Gabriel Romon Nov 14 '16 at 13:52
  • $\begingroup$ I cannot see why do you need $\frac{ε}{2^{i}}$ and not just $\frac{1}{i}$. Is there a reason? $\endgroup$ – richarddedekind Mar 18 at 14:50


Your reasoning is correct. You just need a lemma that estimates the closeness of the approximation of a union by a union of approximants. Details follow:

Let $A\subset X$ and $C \in \mathcal{X}$, $C\supset A$. We have $$\mu^*(A) = \inf \{\mu(B) \ | \ B \in \mathcal{X}, B \supset A\} = \inf \{\mu(B) \ | \ B \in \mathcal{X}, C \supset B \supset A\} $$ Therefore we have $$\mu(C) - \mu^{*}(A) = \sup \{ \mu(C) - \mu(B) \ | \ C \supset B \supset A\} = \sup\{ \mu(D) \ | \ D \in \mathcal{X},D \subset C \backslash A\} $$

Let's define the innner measure of a set $M$ to be $$\mu_*(M)=\sup\{ \mu(D) \ | D \in \mathcal{X},\ D \subset M\} $$ Therefore, for every $A \subset X$ and $C \in \mathcal{X}$, $C \supset A$ $$\mu(C) - \mu^{*}(A) = \mu_*(C \backslash A)$$

Let us prove the following: for $A_1$, $A_2 \subset X$, $C_i \in \mathcal{X}$ and $C_i \supset A_i$ we have $$\mu_*((C_1\cup C_2) \backslash (A_1 \cup A_2)) \le \mu_*(C_1 \backslash A_1) + \mu_*(C_2 \backslash A_2) $$

Indeed, let $D \subset (C_1\cup C_2) \backslash (A_1 \cup A_2)$, $D \in \mathcal{X}$. Then $(D \cap C_i) \subset C_i \backslash A_i$ and therefore $\mu(D) \le \mu(D\cap C_1) + \mu(D\cap C_2) \le \mu_*(C_1 \backslash A_1) + \mu_*(C_2 \backslash A_2)$. Now take the sup over $D$ and get the inequality.

From the above we get

$$\mu(C_1\cup C_2) - \mu^{*} (A_1 \cup A_2) \le (\mu(C_1) - \mu^*(A_1))+ (\mu(C_2) -\mu^*(A_2))$$

Your argument runs like this it seems: let $A_1 \subset A_2 \subset \ldots$. Take $\epsilon > 0$, approximate $A_n$ from the outside with $B_n$ better than $\epsilon/2^{n}$, and take $C_n = B_1 \cup \ldots B_n$ that approximates $A_n$ by better than $\epsilon ( 1/2 + \cdots 1/2^2) < \epsilon$. Now we got an increasing sequence of approximants. Let's show that $\mu(\cup B_n) - \mu^{*}(A) \le \epsilon$. Indeed: the difference is $\mu_* ( \cup B_n \backslash A)$. Let $D \subset \cup B_n \backslash A$, $D \in \mathcal{X}$. Then $D$ is the increasing union of $D \cap B_n$ and each has measure $< \epsilon$, so take the limit. We are about done.


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.