Non-Measurable Sets

Measure theory opened by asserting that no translation- invariant, countably additive, normalized measure exists on all subsets of ${\mathbb{R}}^n$, and deferred the proof. This page supplies it, and then asks the question the proof provokes: is the obstruction a fact about the real line, or an artefact of the axiom of choice?

The answer — it is an artefact of choice, but one we cannot cheaply discard — is the reason measure theory is organized around $\sigma$-algebras rather than around all subsets, and it connects this spine to the set-theoretic material in ZFC and choice-and-excluded-middle.md.

References: Folland, Real Analysis, §1.1; Wagon, The Banach–Tarski Paradox; Solovay, "A model of set theory in which every set of reals is Lebesgue measurable" (1970).

1. The Vitali Set

Theorem (Vitali, 1905). There is no function $\mu :\mathcal{P}(\mathbb{R})>\to [0,\infty ]$ that is countably additive, translation invariant, and satisfies $\mu ([0,1])=1$.

Proof. Define $x\sim y⟺x-y\in \mathbb{Q}$; this is an equivalence relation on $[0,1]$. By the axiom of choice, pick one representative from each class, forming $V\subseteq [0,1]$. Enumerate $\mathbb{Q}\cap [-1,1]=\{q_1,q_2,\dots \}$ and put $V_k=V+q_k$.

The $V_k$ are pairwise disjoint (two representatives differing by a rational would be in the same class, hence equal), and

$$[0,1]\subseteq \underset{k=1}{\overset{\infty }{⨆}}{V}_k\subseteq [-1,2].$$

By translation invariance every $\mu (V_k)$ equals the same value $c=\mu (V)$. By countable additivity and monotonicity,

$$1\le \sum _{k=1}^{\infty }c\le 3.$$

If $c=0$ the middle sum is $0$, contradicting the left inequality; if $c>0$ it is $+\infty$, contradicting the right. $\square$

Every hypothesis is load-bearing, and the proof is a good map of what can be traded away:

Measure theory takes row three. The others are not idle: row one is the subject of finite-additivity.md, and row four is the content of the rest of this page.

2. It Is Choice That Does the Work

The Vitali construction uses AC in an essential and irreducible way: it selects one point from each of continuum-many classes with no rule for the choice. There is no formula defining $V$, and this is not a failure of ingenuity.

Theorem (Solovay, 1970). If ZFC + "there exists an inaccessible cardinal" is consistent, then so is

$$\mathsf{ZF}+\mathsf{DC}+\text{"every subset of }\mathbb{R}\text{ is Lebesgue measurable"},$$

where DC is the axiom of dependent choice.

DC is enough for essentially all of classical analysis — it justifies recursive constructions, sequential compactness arguments, and the countable choices hidden in $\epsilon$–$\delta$ proofs. So there is a coherent mathematical universe in which every set is measurable and analysis proceeds much as usual. Non-measurable sets are not forced on us by the real line; they are forced on us by full choice over an uncountable index set.

The cost is not zero. In Solovay's model, and in ZF + DC generally, one loses the Hahn–Banach theorem in full strength, the existence of a basis for every vector space, ultrafilters on $\mathbb{N}$ (hence the ultrapower construction of non-standard analysis), and Tychonoff for arbitrary products. Shelah later showed the inaccessible cardinal is necessary for Solovay's result — "all sets measurable" carries genuine consistency strength. The trade is real, and mainstream mathematics takes AC.

The same pattern — a classically indispensable choice principle with constructively unacceptable consequences — appears in choice-and-excluded-middle.md, where AC in a constructive setting outright implies excluded middle.

3. Banach–Tarski

Theorem (Banach–Tarski, 1924). For $n\ge 3$, the unit ball in ${\mathbb{R}}^n$ can be partitioned into finitely many pieces which, moved by rigid motions only, reassemble into two unit balls.

The pieces are necessarily non-measurable — otherwise volume would be conserved — and are produced by a choice-based selection much like Vitali's, applied to orbits of a free subgroup of $SO(3)$ generated by two rotations.

Why $n\ge 3$? The construction needs a free non-abelian subgroup of the isometry group, which $SO(3)$ has and the isometry group of the plane (being solvable) does not. In dimensions 1 and 2 the group is amenable, and a finitely additive, isometry-invariant extension of Lebesgue measure to all subsets exists (Banach) — so no paradoxical decomposition is possible there. The dimensional divide is a fact about groups, not about space; the relevant group-theoretic notions are in group theory.

Banach–Tarski is often presented as a paradox about volume. It is better read as a theorem about which group actions admit invariant means, and as a proof that "finitely additive volume for all sets" is untenable in dimension $\ge 3$ even before countable additivity is requested.

A different route to the same moral: if one asks for a measure on all subsets while dropping only translation invariance, one arrives at real-valued measurable cardinals — the existence of a countably additive extension of ${\lambda }^1$ to $\mathcal{P}(\mathbb{R})$ is equiconsistent with a measurable cardinal, far beyond ZFC. Total measurability is expensive however it is bought.

4. What the $\sigma$-Algebra Is Doing

Reading the design decision back into measure-theory.md:

• Restricting the domain is the price of countable additivity. The $\sigma$-algebra is not a technicality to be tolerated but the exact concession that makes limit theorems — monotone and dominated convergence §4 — available.
• The restriction costs nothing in practice. Every set that can be specified by countably many analytic operations is Borel; every Borel set is Lebesgue measurable; and every analytic (Suslin) set is too. Non-measurable sets cannot be exhibited, only proved to exist.
• Completion is a separate, cheap fix. ${\mathcal{L}}^n⊋\mathcal{B}({\mathbb{R}}^n)$ adds all subsets of null sets. Cardinality shows $|{\mathcal{L}}^1|=2^{\mathfrak{c}}>\mathfrak{c}=|\mathcal{B}(\mathbb{R})|$, so most Lebesgue sets are not Borel — yet the extension is harmless because the additions are null.
• In probability, the $\sigma$-algebra earns a second job. There it also encodes which events are observable, and sub-$\sigma$-algebras represent partial information — the reading that makes conditional expectation and filtrations work. The technical constraint of this page and the informational interpretation of that one are the same object seen twice.

5. Where This Is Used

• Measure theory §1 — the deferred impossibility theorem.
• Countable vs. finite additivity — what survives if row one of §1's table is taken instead.
• The axiom of choice and excluded middle and ZFC — the set-theoretic background.
• The ultrapower construction — another consumer of non-constructive choice, with the same flavour of unexhibitable object.