15
$\begingroup$

Let $\mathfrak{L}(\mathbb{R})$ be the collection of Lebesgue measurable sets and $\mathfrak{B}(\mathbb{R})$ be the Borel sets.

Question: Is there a nontrivial signed measure on $\mathfrak{L}(\mathbb{R})$ that is trivial on $\mathfrak{B}(\mathbb{R})$?

Obviously, any positive measure that is trivial on $\mathfrak{B}(\mathbb{R})$ is also trivial on $\mathfrak{L}(\mathbb{R})$, since any Lebesgue measurable set is a subset of a Borel set.

For the signed case, I have tried doing Jordan decomposition but it doesn't seem work. It is hard (if ever possible) to show $(\mu|_{\mathfrak{B}(\mathbb{R})})^+ = \mu^+|_{\mathfrak{B}(\mathbb{R})}$ and $(\mu|_{\mathfrak{B}(\mathbb{R})})^- = \mu^-|_{\mathfrak{B}(\mathbb{R})}$.

In fact, If I can deal this problem by decomposition, there must be something special about Borel sets, since the above equalities do not hold in general. Let $\mathfrak{C} = \{\varnothing,\{0\},\{1\},\{0,1\}\}$, $\mathfrak{D} = \{\varnothing, \{0,1\}\}$. The signed measure $\lambda$ on $\mathfrak{C}$ is defined that $\lambda(\{0\})=1$ and $\lambda(\{1\})=-1$. Then $\lambda|_\mathfrak{D}$ is trivial and the equalities fail.

Background: I am trying to prove (or disprove) that if $\mu$ and $\lambda$ are signed measures on $\mathfrak{L}(\mathbb{R})$, then $\mu|_{\mathfrak{B}(\mathbb{R})} = \lambda|_{\mathfrak{B}(\mathbb{R})}$ implies $\mu = \lambda$.

$\endgroup$
6
  • 2
    $\begingroup$ Use the Hahn decomposition instead. $\endgroup$ Oct 20, 2020 at 7:21
  • $\begingroup$ If $(H^+,H^-)$ is a Hahn decompostion of $\mu$. Do you mean I should apply the positive case to $\mu(E\cap H^+)$ and $\mu(E\cap H^-)$? $\endgroup$ Oct 20, 2020 at 12:59
  • 1
    $\begingroup$ So every Borel subset of $H^+$ or $H^-$ is null... but I don't see how that helps. @MichaelGreinecker can you give another hint? $\endgroup$ Oct 20, 2020 at 13:53
  • 2
    $\begingroup$ I guess one note is the following. Suppose a nontrivial measure $\mu$ exists, so $\mu(H^+) > 0$. We can write $H^+ = B^+ \cup N^+$ where $B^+$ is Borel and $N^+$ is null. Since $\mu(B^+) = 0$ by assumption, $\mu(N^+) > 0$. In particular, $N^+$ is not Borel and necessarily uncountable. But every subset of $N^+$ is Lebesgue measurable, so $\mu|_{N^+}$ is a nontrivial countably additive positive measure on $2^{N^+}$. I think this makes $|N^+|$ a real-valued measurable cardinal, or at least implies the existence of one, and this is unprovable in ZFC. $\endgroup$ Oct 20, 2020 at 13:58
  • $\begingroup$ Does @NateEldredge's comment work in reverse? If there is a real-valued measurable cardinal, can we use it to construct an example as in the OP? $\endgroup$ Oct 20, 2020 at 15:45

2 Answers 2

6
$\begingroup$

So, promoting my answer to a comment, this is unprovable in ZFC (assuming ZFC is consistent). I claim that such a signed measure $\nu$ exists only if there exists a nontrivial, atomless, countably additive probability measure $\mu$ on the discrete $\sigma$-algebra of $\mathbb{R}$ (or equivalently $[0,1]$). As I understand it, the latter is equivalent to the existence of a real-valued measurable cardinal of size at most $\mathfrak{c}$, which is independent of ZFC.

Suppose such $\nu$ exists. Consider its Hahn decomposition $\mathbb{R} = H^+ \cup H^-$. Since $H^+ \in \mathfrak{L}(\mathbb{R})$, it can be written $H^+ = B^+ \cup N^+$ where $B^+$ is Borel and $N^+$ is Lebesgue-null. By assumption $\nu(B^+) = 0$ so we must have $\nu(N^+) > 0$, and $\nu$ is positive on $N^+$. Now every subset of $N^+$ is Lebesgue measurable, so $\nu$ is defined for every such subset. Thus define $\mu(A) = \nu(A \cap N^+)$ for any subset $A \subset \mathbb{R}$. This is a nontrivial, countably additive, finite, positive measure on $2^{\mathbb{R}}$, which we may rescale to a probability measure. And since singletons are Borel, and therefore have $\nu$-measure zero, $\mu$ is atomless.

Gerald's answer, with Michael's comments, seems to be establishing the converse, that the existence of a real-valued measurable cardinal implies the existence of a desired $\nu$. Combining these would show that the original statement is independent of ZFC.

$\endgroup$
3
  • $\begingroup$ Thanks. I see it is impossible to show the existence of such measure in ZFC if ZFC is consistent. However, why is the signed measure $\nu$ constructed in the converse nontrivial on Lebesgue measurable sets? $\endgroup$ Oct 22, 2020 at 13:18
  • $\begingroup$ @Zhang: Responded on the other answer. $\endgroup$ Oct 22, 2020 at 14:10
  • $\begingroup$ Thank you. I have added a new comment below your response. $\endgroup$ Oct 22, 2020 at 14:50
3
$\begingroup$

a converse of Nate Eldridge's comment
not a proof, too long for a comment

Suppose there is a real-valued measurable cardinal. We want to show there is a measure as requested.

There is a probability measure $\mu : \mathfrak P([0,1]) \to [0,1]$. We may assume $\mu([0,t]) = t$ for $0 \le t \le 1$.

Using AC of course, can we show the existence of a set $X \subseteq [0,1]$ with $$ \mu(X \cap [0,t]) = t/2\quad \text{for all }t \in [0,1]\quad? \tag1 $$ We can deduce from this: $$ \mu\big(X \cap B\big) = \frac{1}{2}\lambda\big(B\cap[0,1]\big) \quad\text{for all Borel sets }B. \tag2$$
Then the signed measure we want will be $$ \nu(E) = \mu\big(X \cap E\big) - \mu\big((\,[0,1]\setminus X)\cap E\big) $$ From $(2)$ we can prove that $\nu(B) = 0$ for all Borel sets $B$.

Addendum. If we cannot prove $(1)$ for an arbitrary measure $\mu$ as described, maybe we can construct $\mu$ together with $X$ in order to get $(1)$.

$\endgroup$
9
  • 3
    $\begingroup$ (1) is indeed correct. By a theorem of Gitik-Shelah (found in 5II of Fremlin) the Maharam representation of the measure algebra of $\mu$ is a countable convex combination of uncountable many coin flips each. The Borel $\sigma$-algebra corresponds to countably many factors. In particular, there must be a coin flip independent of all of them. $\endgroup$ Oct 20, 2020 at 22:33
  • $\begingroup$ Your argument works even without the existence of a real-valued measurable cardinal, the non-separable extension of Lebesgue measure constructed by Kakutani and Oxtoby here satisfies (1). $\endgroup$ Oct 20, 2020 at 22:45
  • $\begingroup$ Do you mean $\mu\big(X \cap B\big) = \frac{1}{2}\mu\big(B\cap[0,1]\big)$ in (2)? Also, this might be a silly question, but how can we show that $\nu$ is nontrivial on Lebesgue measurable sets? $\endgroup$ Oct 22, 2020 at 3:34
  • 2
    $\begingroup$ @Zhang: I think to get your desired measure, we should push forward the measure $\nu$ from $[0,1]$ to the Cantor set $C$ under a Borel isomorphism $\phi$. The new measure $\phi_* \nu$ is then defined on all subsets of $C$ and zero on all Borel sets, so it must be nonzero on some other subset of the Cantor set, which in particular is Lebesgue measurable. Now you can extend $\phi_* \nu$ to the rest of $\mathbb{R}$ by making it zero outside $C$. $\endgroup$ Oct 22, 2020 at 14:08
  • 3
    $\begingroup$ If I correctly understand @MichaelGreinecker's comment, it says that even without a real-valued measurable, there exists a $\sigma$-algebra $\mathcal{F} \supsetneq \mathcal{B}$ and a measure $\mu$ on $\mathcal{F}$, such that there is a set $X \in \mathcal{F}$ satisfying (1). But in order to produce Zhang's desired measure, we would actually need to have $\mathcal{F} = 2^{[0,1]}$ or some similar condition, and it is this which can only happen if there is a real-valued measurable cardinal. $\endgroup$ Oct 22, 2020 at 15:16

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.