In a previous homework our task was to prove the regularity of the Lebegue-measure on $\mathbb{R}^d$. More precisely:

Let $(\mathbb{R}^d, \mathcal{M}^*, \lambda)$ be a measure space and $\mathcal{M}^* := \mathcal{M}^*(\mathbb{R}^d)$ be the $\sigma$-algebra of the $\lambda^*$-measurable sets, with $\lambda$ being the by the Lebesgue-outer-measure $\lambda^*$ induced measure.

Show that \begin{align*} \lambda(A) & = \inf\{\lambda(O) \mid O \subset \mathbb{R}^d \text{ is open and } A \subset O \} \\ & = \sup\{\lambda(K) \mid K \subset \mathbb{R}^d \text{ is compact and } K \subset A \}. \end{align*}

Our attempt goes as follows:

For all $A \in \mathcal{M}^*$ we want to show

\begin{equation*} \lambda(A) \ge \inf\{ \lambda(O): O \subset \mathbb{R}^d \text{ open und } A \subset O \} \ge \sup\{ \lambda(K): K \subset \mathbb{R}^d \text{ compact and } K \subset A \} \ge \lambda(A) \end{equation*}

First inequality: If $\lambda(A) = \infty$, because of the monotonicity of the measure, we have $\lambda(O) = \infty$ for all open sets $O$ with $A \subset O$.

If $\lambda(A) < \infty$, we define $A_{\delta} := \bigcup_{a \in A} U_{\delta}(a)$, which is a open and therefore measurable set with $A \subset A_{\delta}$.

Now we have $A_{\delta} \setminus A \xrightarrow{\delta \to 0} \emptyset$ and because of the continuity from below in the emptyset of $\lambda$, and, because $A_{\delta} \setminus A$ is measurable, because it's the difference of measurable sets, \begin{equation*} \forall \varepsilon > 0 \ \exists \delta > 0: \lambda(A_{\delta} \setminus A) < \varepsilon. \end{equation*}

Because $A_{\delta}$ is open and $\sigma$-additivity of $\lambda$ we have \begin{equation*} \inf\{ \lambda(O): O \subset \mathbb{R}^d \text{ open und } A \subset O \} \le \lambda(A_{\delta}) = \lambda(A_{\delta} \setminus A) + \lambda(A) < \varepsilon + \lambda(A). \end{equation*}

Because $\varepsilon > 0$ was arbitrary, the inequality follows.

The second inequality follows from the monotonicity of $\lambda^*$: From $K \subset A \subset O$ we have $\lambda(K) \le \lambda(O)$ for $K$ und $O$ as defined above. Because taking the supremum or infimum doesn't change weak inequalities, the inequality follows.

The third inequality: One can show, that every open set is $\sigma$-compact.

Lemma Let $A_{\delta} \subset \mathbb{R}^d$ be an open subset. Then there exists a countable family of closed bounded axially parallel cubes $(W_k)_{k \in \mathbb{N}}$, so that $A_{\delta} = \bigcup_{n \in \mathbb{N}} W_n$.

Proof Let $a \in A_{\delta} $. Because $A_{\delta} $ is open, there exists an $\varepsilon_{a} > 0$, so that $U_{\varepsilon_a}(a) \subset A_{\delta} $. For every $a \in A_{\delta} $ we choose a bounded closed axially parallel cube $W_a$ with rational midpoint from $\mathbb{Q}^d$ and rational edge length $q \in \mathbb{Q}$, so that \begin{equation*} a \in W_{a} \subset U_{\varepsilon_a}(a) \subset A_{\delta} . \end{equation*} This is always possible, because $\mathbb{Q}$ is dense in $\mathbb{R}$. Therefore, we have $A_{\delta} = \bigcup_{n \in \mathbb{N}} W_n$. Since there are only countable many cubes of this form, the union is countable. $\square$

We let $K_n := \bigcup_{j = 1}^{n} W_j$, which is compact as union of compact sets. Then $K_n \xrightarrow{n \to \infty} A_{\delta}$.

Now we have $A_{\delta} \setminus K_n \xrightarrow{n \to \infty} \emptyset$ and with analogous argumentation as above \begin{equation*} \forall \varepsilon > 0 \ \exists N \in \mathbb{N}: \lambda(A_{\delta} \setminus K_n) < \varepsilon \ \forall n > N. \end{equation*} Therefore follows for all $\varepsilon > 0$ \begin{align*} \lambda(A) \le \lambda(A_{\delta}) \le \lambda(A_{\delta}\setminus K_n) + \lambda(K_n) < \varepsilon + \sup\{ \lambda(K): K \subset \mathbb{R}^d \text{ compact und } K \subset A_{\delta} \}, \end{align*}

Because $\varepsilon > 0$ was arbitrary and $\delta$ can be arbitrarily small, the inequality follows.

My problem

I assume the proof of the second inequality is right. But: I not sure if the reasoning at the end of the proof of the last inequality is rigorous enough. Especially, if you take $A := \mathbb{Q}$ in the proof of the first inequality, then for every $\delta > 0$, we have $A_{\delta} = \mathbb{R}$ and therefore, we don't have $A_{\delta} \setminus A \to \emptyset$ or even $\lambda(A_{\delta} \setminus A) \to 0$.

Is there anyway to ''save'' this proof by fixing it and not changing the approach?

Correct Proof

First inequality Case 1: $\lambda(A) = \infty$. As above.

Case 2: $\lambda(A) < \infty$.

Utilising the Caratheodory construction of the Lebesgue measure, we know that $$ \forall \varepsilon > 0 \ \exists (a_n,b_n] := \prod_{i=1}^d \left(a_{n}^{(i)},b_{n}^{(i)}\right]: A \subset \bigcup_{n \in \mathbb{N}} (a_n,b_n] \quad \text{and} \quad \sum_{k=1}^\infty \lambda \left((a_k,b_k]\right) < \lambda(A) + \frac{\varepsilon}{2}, $$ where the $d$-dimensional cubes $(a_n,b_n]$ are pairwise disjoint.

Now let $$ U := \bigcup_{n=1}^\infty (a_n, b_n+ t_n \varepsilon) \qquad \text{with} \qquad t_n := 2^{-n-2d-1} \max\{1,b_{n}^{(i)} -a_{n}^{(i)}\}^{-(d-1)}. $$ Now we have $A \subset \bigcup_{n=1}^\infty (a_n,b_n] \subset U$ and $\lambda(U) < \lambda(A) + \varepsilon$.

Therefore, we have \begin{equation*} \inf\{ \lambda(O): O \subset \mathbb{R}^d \text{ open und } A \subset O \} \le \lambda(U) < \lambda(A) + \varepsilon. \end{equation*} Because $\varepsilon > 0$ was arbitrary, the inequality follows.

Second inequality Same as above

Third inequality

From a previous homework we know that $\mathcal{B}(\mathbb{R}^d) \subset \mathcal{M}^*$ and therefore, that \begin{equation*} \forall M \in \mathcal{M}^* \ \exists B \in \mathcal{B}(\mathbb{R}^d), N \in \mathcal{N}: M = B \cup N, \end{equation*} where $\mathcal{N}$ is the set of borel-null-sets. Now define \begin{equation*} \mathcal{D} := \{ B \in \mathcal{B}(\mathbb{R}^d): \lambda(B) = \sup\{\lambda(K) \mid K \subset \mathbb{R}^d \text{ is compact and } K \subset B \} \} \end{equation*} as the set of all inner regular sets in the borel $\sigma$-algebra, which, by construction is a subset of $\mathcal{B}(\mathbb{R}^d)$.

Now we want to show, that $\mathcal{B}(\mathbb{R}^d) \subset \mathcal{D}$ to conclude $\mathcal{B}(\mathbb{R}^d) = \mathcal{D}$.

  1. From the above lemma we know, that for all open sets $\mathcal{O} \subset \mathbb{R}^d$ we have $\mathcal{O} \in \mathcal{D}$, since by the continuity of the Lebesgue-measure, for all $\varepsilon > 0$ we have $\lambda(A) \le \lambda(\bigcup_{n \in \mathbb{N}} W_n) + \varepsilon$.

  2. Since the set of open sets is a $\cap$-stable generator of $\mathcal{B}(\mathbb{R}^d) \subset \mathcal{P}(\mathbb{R})^d$, we only need to show that $\mathcal{D}$ is a dynkin system. (German Wikipedia article on this line of argument)

    • We have $\mathbb{R}^d \in \mathcal{D}$, because $\lambda(\mathbb{R}^d) = \infty = \sup\{\lambda(K): K \subset \mathbb{R}^d \}$.

    • Let $D \in \mathcal{D}$. Then ??

    • Let $\{ A_n \}_{n \in \mathbb{N}} \subset \mathcal{D}$ be a family of disjoint subsets. Then


1 Answer 1


The first property is called 'outer regularity' and the second 'inner regularity'. First of all, you can not take $A_\delta$ as defined above, as you have already noted. Also the proof of the second equality (i.e. your argument in the section third inequality) is false, because it can happen that $K_n$ is not a subset of $A$! (Take your example: Then $A_\delta = \mathbb{R}$ and $K_n$ will approximate the measure of $\mathbb{R}$.)

To prove the 'outer regularity' use the Caratheodory construction of the Lebesgue measure: If $\lambda(A) < \infty$, then we can find $(a_n,b_n] = \prod_{i=1}^d (a_{n,i},b_{n,i}] $ disjoint ($d$-dimensional) cubes with $$\sum_{k=1}^\infty \lambda ((a_n,b_n]) < \lambda(A) + \varepsilon/2$$ Thus, we can take $U= \bigcup_{n=1}^\infty (a_n,b_n+ t_n \varepsilon)$. (Note $ A \subset \bigcup_{n=1}^\infty (a_n,b_n] \subset U$ and if $t_n$ is taken appropriate, e.g. $t_n = 2^{-n-2d-1} \max\{1,b_{n,i} -a_{n,i}\}^{-(d-1)} $ we get $\lambda(U) < \lambda(A) + \varepsilon$.)

The proof ot the 'inner regularity' is more complicated. First recall that any set $\mathcal{M}^*$ is the completion of the Borel-$\sigma$-algebra. Thus any $A \in \mathcal{M}^*$ can be written as $A= B \cup M$ with a Borel set $B$ and $M \subset N$ with a Borel-nullset $N$. Thus, we only have to prove the 'inner regularity' for Borel-sets. Define the system $$\mathcal{D} := \{ B \in \mathcal{B}(\mathbb{R}^d) : B \text{ is inner regular}\}.$$

  1. Your argument shows that open sets are in $\mathcal{D}$.
  2. Check that $\mathcal{D}$ is a Dynkin-system.
  3. Since the set of all open sets is a $\cap$-stable generator of the Borel-$\sigma$-algebra we can conclude that $\mathcal{D} = \mathcal{B}(\mathbb{R}^d)$.

I additionally added a prove of 2: Note that $\mathbb{R}^d$ is an open set. If $(A_n)_{\in \mathbb{N}} \subset \mathcal{D}$ are disjoint, then we can take compact $K_n \subset A_n$ with $\lambda(A_n) < \lambda(K_n) + \varepsilon 2^{-n}$. Since a finite union of compact sets is compact, but in general not an infinite union, we have to truncate the sums as follows.

First case: Now if $\lambda(\cup_{n=1}^\infty A_n) = \sum_{n=1}^\infty \lambda(A_n) = \infty$, then for any $K>0$ there exists $N$ such that $$\sum_{n=1}^N \lambda(A_n) >K.$$ Thus $\sum_{n=1}^N \lambda(K_n) > K-\varepsilon$. So taking the compact set $K = \cup_{k=1}^N K_n$ shows that the supremum is $>K-1$. Since $K>0$ was arbitrary, we get that the supremum is $\infty$.

In the second case, we have for some $N \in \mathbb{N}$, because the series is convergent, that $$\lambda(\cup_{n=1}^\infty A_n) < \sum_{n=1}^N \lambda(A_n) + \varepsilon.$$ Thus $$\lambda(\cup_{n=1}^\infty A_n) < 2 \varepsilon + \lambda(\cup_{n=1}^N K_n) $$ and therefore we can take the compact set $K = \cup_{n=1}^N K_n$. This proves that $\cup_{n=1}^\infty A_n \in \mathcal{D}$.

Let $A,B \in \mathcal{D}$ with $B \subset A$. For the next argument we need that both $B$ and $A$ have finite measure. Of course, this is satisfied if the sets are bounded. Thus take $$A_i = A \cap \prod_{i=1}^d (n_i,n_i+1] \quad \text{and} \quad B_i = B \cap \prod_{i=1}^d (n_i,n_i+1]$$ with $n_i \in \mathbb{Z}$ instead of $A$ and $B$ to get bounded sets. (If $A_i \setminus B_i \in \mathcal{D}$. then also the union of this disjoint sets is in $\mathcal{D}$ by the previous argument. Moreover, show that also $A_i,B_i \in \mathcal{D}$.)

So we can assume that both $B$ and $A$ are bounded. Now take an open set $U$ with $B \subset U$ and $\lambda(U \setminus B) < \varepsilon$ (this is possible, because we already know that $\lambda$ is outer regular) and a compact set with $K \subset A$ and $\lambda(A \setminus K) < \varepsilon$. Define $L = K \setminus U$. Then $L$ is compact and $$\lambda( (A \setminus B) \setminus L) \le \lambda(U \setminus B) + \lambda(A \setminus K) < 2 \varepsilon.$$ Therefore $A \setminus B \in \mathcal{D}$.

  • $\begingroup$ For the first inequality, we had the idea that we can instead choose $A_{\delta}$ in such similar fashion to the way it is possible to cover the rationals of a compact interval in $\mathbb{R}$ with countable many open subsets, whose measure is smaller than an arbitrary $\varepsilon > 0$. Would this work? $\endgroup$
    – Ramanujan
    Jan 13, 2019 at 13:48
  • 1
    $\begingroup$ Yes, I have edited my answer. In the case of the rational numbers you can, in fact, take for example $U = \bigcup_{n=1}^\infty (q_n -2^{-n} \varepsilon, q_n +2^{-n} \varepsilon)$, where $(q_n)_n$ is an enumeration of $\mathbb{Q}$. But this argument only applies for countable sets. If $A=C$ is a (thin) Cantor set, you have uncountable many intervals in your union, but $C$ is a nullset. $\endgroup$
    – p4sch
    Jan 13, 2019 at 13:54
  • $\begingroup$ I have added the construction of the $W_j$ with the proof of the $\sigma$-compactness of the open set $A_{\delta}$. Can you please explain why it is a problem, that $K_n \not\subset A$? $\endgroup$
    – Ramanujan
    Jan 13, 2019 at 15:05
  • 1
    $\begingroup$ The lemma after the 'third inequality' shows that, if $U$ is open, there exists a countable family of closed bounded axially parallel cubes $(W_k)_{k \in \mathbb{N}}$ such that $A=\bigcup_{n \in \mathbb{N}} W_n$. Thus, by the continuity of measures, we have $\lambda(U) < \lambda(\bigcup_{n=1}^N W_n) + \varepsilon$ for some $N \in \mathbb{N}$. Note that $K:=\bigcup_{n=1}^N W_n$ is a compact set. $\endgroup$
    – p4sch
    Jan 13, 2019 at 20:40
  • 1
    $\begingroup$ Yes, the rest of your corrected proof is error-free. Note that the characterization of the completion (in the sense of $A= B \cup M$ and $B$ is Borel measurable and there exists a Borel-nullset $N$ with $M \subset N$) is not just a consequence of $\mathcal{B}(\mathbb{R}^d) \subset \mathcal{M}^*$, but a theorem about the characterization of the completion of a measure space. $\endgroup$
    – p4sch
    Jan 13, 2019 at 21:23

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.