Carathéodory's Measurability Criterion

Basis
Last updated: Tags: Measure Theory

Prerequisites

Outer measure gives every subset of R\mathbb{R} a size, but it is only sub-additive. The problem is that some “bad” sets leak measure across boundaries in a way that makes additivity fail. Constantin Carathéodory’s insight (1914) was to characterise the “good” sets — the measurable ones — by a single geometric condition: a set EE is measurable precisely when it perfectly splits every test set AA into two non-overlapping pieces.

The criterion

Let μ\mu^* be an outer measure on a set XX (see Outer Measure for the abstract definition). A set EXE \subseteq X is Carathéodory-measurable (or simply measurable) with respect to μ\mu^* if

μ(A)=μ(AE)+μ(AEc)for every AX.(1)\mu^*(A) = \mu^*(A \cap E) + \mu^*(A \cap E^c) \qquad \text{for every } A \subseteq X. \tag{1}

The set AA in (1)(1) is called the test set. You are asking: does EE divide AA into two pieces whose outer measures add up exactly to μ(A)\mu^*(A)?

Because A=(AE)(AEc)A = (A \cap E) \cup (A \cap E^c) and the two pieces are disjoint, sub-additivity always gives

μ(A)μ(AE)+μ(AEc).\mu^*(A) \leq \mu^*(A \cap E) + \mu^*(A \cap E^c).

So the content of criterion (1)(1) is really the reverse inequality:

μ(A)μ(AE)+μ(AEc).(1’)\mu^*(A) \geq \mu^*(A \cap E) + \mu^*(A \cap E^c). \tag{1'}

In words: EE does not “create” extra measure when it cuts AA in two.

Remark. The criterion is symmetric in EE and EcE^c: if EE satisfies (1)(1), so does EcE^c (just swap the roles). So measurability is preserved under complementation.

The measurable sets form a σ-algebra

Let M\mathcal{M} denote the collection of all μ\mu^*-measurable sets. You will now verify that M\mathcal{M} satisfies all three σ-algebra axioms.

Axiom 1: XMX \in \mathcal{M}. For any test set AA: AX=AA \cap X = A and AXc=A=A \cap X^c = A \cap \emptyset = \emptyset, so μ(AX)+μ(AXc)=μ(A)+0=μ(A)\mu^*(A \cap X) + \mu^*(A \cap X^c) = \mu^*(A) + 0 = \mu^*(A). ✓

Axiom 2: Closed under complement. Already observed above: criterion (1)(1) is symmetric in EE and EcE^c. ✓

Axiom 3: Closed under countable union. This is the heart of the argument. Here is the key step.

Finite unions first. Suppose E,FME, F \in \mathcal{M}. You want to show EFME \cup F \in \mathcal{M}. For any test set AA, use the measurability of EE with test set AA and then the measurability of FF with test set AEcA \cap E^c:

μ(A)=μ(AE)+μ(AEc)\mu^*(A) = \mu^*(A \cap E) + \mu^*(A \cap E^c) μ(AEc)=μ(AEcF)+μ(AEcFc).\mu^*(A \cap E^c) = \mu^*(A \cap E^c \cap F) + \mu^*(A \cap E^c \cap F^c).

Note EcFc=(EF)cE^c \cap F^c = (E \cup F)^c, so

μ(A)=μ(AE)+μ(AEcF)+μ(A(EF)c).(2)\mu^*(A) = \mu^*(A \cap E) + \mu^*(A \cap E^c \cap F) + \mu^*(A \cap (E \cup F)^c). \tag{2}

Since A(EF)=(AE)(AEcF)A \cap (E \cup F) = (A \cap E) \cup (A \cap E^c \cap F) and these are disjoint, sub-additivity gives μ(A(EF))μ(AE)+μ(AEcF)\mu^*(A \cap (E \cup F)) \leq \mu^*(A \cap E) + \mu^*(A \cap E^c \cap F). Substituting into (2)(2):

μ(A)μ(A(EF))+μ(A(EF)c),\mu^*(A) \geq \mu^*(A \cap (E \cup F)) + \mu^*(A \cap (E \cup F)^c),

which (combined with sub-additivity in the other direction) proves EFME \cup F \in \mathcal{M}. By induction, M\mathcal{M} is closed under finite unions.

Countable unions. Let E1,E2,ME_1, E_2, \ldots \in \mathcal{M} be pairwise disjoint (the general case reduces to this by writing E~k=Ek(E1Ek1)\tilde{E}_k = E_k \setminus (E_1 \cup \cdots \cup E_{k-1}), which remain in M\mathcal{M} by the closure properties already established). Set Snk=1nEkMS_n \coloneqq \bigsqcup_{k=1}^n E_k \in \mathcal{M}. For any test set AA and each nn:

μ(ASn)=k=1nμ(AEk).(3)\mu^*(A \cap S_n) = \sum_{k=1}^{n} \mu^*(A \cap E_k). \tag{3}

Equation (3)(3) follows by induction on nn using the measurability of EnE_n with test set ASnA \cap S_n. Now let SkEkS \coloneqq \bigcup_k E_k. Since SnSS_n \subseteq S, monotonicity and (3)(3) give:

μ(AS)μ(ASn)=k=1nμ(AEk).\mu^*(A \cap S) \geq \mu^*(A \cap S_n) = \sum_{k=1}^{n} \mu^*(A \cap E_k).

Taking nn \to \infty yields μ(AS)k=1μ(AEk)\mu^*(A \cap S) \geq \sum_{k=1}^{\infty} \mu^*(A \cap E_k), so together with sub-additivity:

μ(AS)=k=1μ(AEk).(4)\mu^*(A \cap S) = \sum_{k=1}^{\infty} \mu^*(A \cap E_k). \tag{4}

Also ScSncS^c \subseteq S_n^c, so μ(ASc)μ(ASnc)\mu^*(A \cap S^c) \leq \mu^*(A \cap S_n^c). Using measurability of SnS_n:

μ(ASn)+μ(ASnc)=μ(A),\mu^*(A \cap S_n) + \mu^*(A \cap S_n^c) = \mu^*(A),

hence μ(ASc)μ(A)k=1nμ(AEk)\mu^*(A \cap S^c) \leq \mu^*(A) - \sum_{k=1}^n \mu^*(A \cap E_k) for every nn. Combining with (4)(4):

μ(A)μ(AS)+μ(ASc),\mu^*(A) \geq \mu^*(A \cap S) + \mu^*(A \cap S^c),

so SMS \in \mathcal{M}. Conclusion: M\mathcal{M} is a σ-algebra. ✓

Countable additivity on M\mathcal{M}

The real payoff of the above argument is equation (4)(4) applied with A=XA = X:

μ ⁣(k=1Ek)=k=1μ(Ek)for pairwise disjoint EkM.(5)\mu^*\!\left(\bigsqcup_{k=1}^{\infty} E_k\right) = \sum_{k=1}^{\infty} \mu^*(E_k) \qquad \text{for pairwise disjoint } E_k \in \mathcal{M}. \tag{5}

This is countable additivity — the defining property of a true measure. The restriction μM\mu^* \restriction_{\mathcal{M}} therefore turns the outer measure into a genuine measure on the σ-algebra M\mathcal{M}.

Completeness

A measure space (X,M,μ)(X, \mathcal{M}, \mu^*) is complete when every subset of a null set is measurable. Carathéodory’s construction automatically delivers completeness.

Proposition. If μ(N)=0\mu^*(N) = 0 and ANA \subseteq N, then AMA \in \mathcal{M}.

Proof. For any test set TT: TANT \cap A \subseteq N, so μ(TA)μ(N)=0\mu^*(T \cap A) \leq \mu^*(N) = 0. Also TTAcT \supseteq T \cap A^c, so μ(TAc)μ(T)\mu^*(T \cap A^c) \leq \mu^*(T) by monotonicity. Thus μ(TA)+μ(TAc)0+μ(T)=μ(T)\mu^*(T \cap A) + \mu^*(T \cap A^c) \leq 0 + \mu^*(T) = \mu^*(T), and the reverse inequality holds by sub-additivity. So AMA \in \mathcal{M}.

This completeness is important in practice: it means you never need to worry about sub-sets of measure-zero sets “falling outside” the σ-algebra.

Summary

  • Carathéodory’s criterion (1)(1) says EE is measurable when it splits every test set AA additively; the non-trivial content is the reverse inequality (1)(1').
  • The collection M\mathcal{M} of all measurable sets is a σ-algebra: it contains XX, is closed under complements (symmetry of the criterion), and is closed under countable unions (the finite-union argument extends by the limit argument above).
  • On pairwise disjoint measurable sets, the outer measure satisfies countable additivity — equation (5)(5). This makes μM\mu^*\restriction_{\mathcal{M}} a genuine measure.
  • The resulting measure space is complete: every subset of a null set is measurable.
  • Next: Lebesgue Measure applies this machinery to μ=λ\mu^* = \lambda^* (the Lebesgue outer measure on R\mathbb{R}) to produce the canonical notion of length.