Chapter 3 · Subchapter 4

Non-Measurable Sets: Vitali Done Honestly

Chapter 0 promised, and twice since we have leaned on it as a shadow. Now every tool is on the table, and the delegate set V can be confronted in full: it has no length, cannot have one, and cannot be named without the axiom of choice.

In this subchapter

  1. What we owe and what we have
  2. The construction, recalled and certified
  3. The theorem: V is not Lebesgue measurable
  4. The stronger statement: no measure sees V
  5. Why no explicit example exists
  6. What the wall is made of

3.4.1What we owe and what we have

The debt ledger. In 0.3 we built the delegate set V: one representative per equivalence class of the relation "x − y is rational" on [0,1]", argued its rational translates V + q are disjoint and blanket [0,1] while staying inside [−1,2], and read off that no assignment of a "length" to V is consistent. That argument was honest but pre-formal: it used the word "length" before any measure existed. Since then we built the real thing — outer measure and the Carathéodory tollgate in 3.1, the measure λ on the σ-algebra ℒ in 3.2 — and the shadow version in 3.1 warned that additivity must fail somewhere inside the translate machine.

This subchapter pays the debt in three steps: recall the construction with every piece now certified by a theorem we have proved; prove V is not Lebesgue measurable; then prove the stronger, more disturbing statement — that no countably additive measure extending length, on any σ-algebra containing the Borel sets, can see V. Along the way we face the question every reader of the construction eventually asks: fine, V exists — but why can nobody ever write one down?

3.4.2The construction, recalled and certified

Lemma — the delegate machine, each piece with its receipt

Define the relation on [0,1]: x ~ y when x − y ∈ ℚ. Then:

  1. ~ is an equivalence relation — the family of 0 is exactly ℚ ∩ [0,1], and √2/2 lives in the family {x : x − √2/2 ∈ ℚ} (the picture of 0.3);
  2. choosing one delegate per family (axiom of choice — flagged) gives V ⊆ [0,1];
  3. the translates V + q, for q ∈ ℚ ∩ [−1,1], are pairwise disjoint, and every point of [0,1] lies in one of them — the blanket: given x ∈ [0,1], its family's delegate v satisfies x − v = q ∈ ℚ ∩ [−1,1];
  4. every translate stays inside [−1, 2]: points of V lie in [0,1] and |q| ≤ 1;
  5. the family is countable — ℚ ∩ [−1,1] is countable — and each translate V + q is measurable if V is (Example 3 of 3.1: shifts preserve measurability).
Example 1 · Where the blanket and the envelope live

It is worth re-seeing the geometry with prices attached, because the whole proof lives in the gap between two covers. The blanket: [0,1] ⊆ ⋃q (V + q), so if the union is measurable, monotonicity prices it at least 1. The envelope: ⋃q (V+q) ⊆ [−1,2], priced at most 3 by subadditivity. The union is squeezed between a floor of 1 and a ceiling of 3 — the same squeeze that broke the naive assignment in Chapter 0, now resting on proved theorems: monotonicity, subadditivity, shift-invariance, countable additivity over the disjoint family.

3.4.3The theorem: V is not Lebesgue measurable

Theorem 1 — Vitali's theorem

The delegate set V is not Lebesgue measurable: V ∉ ℒ. In consequence there are subsets of [0,1] to which λ assigns no value at all.

Proof — suppose it is, and read the meter

Suppose V ∈ ℒ. Then every translate is measurable (Lemma, item 5), and being pairwise disjoint and countable, countable additivity applies: λ(⋃q(V+q)) = ∑q λ(V+q). Shift-invariance of λ (3.1 Example 1: covering prices are shift-invariant, and λ inherits it) makes every summand the same number p = λ(V) ≥ 0. Now the meter reading, from the squeeze:

1 ≤ λ(⋃q (V+q)) = ∑q p ≤ 3

A sum of countably many copies of a constant p ≥ 0 is 0 if p = 0 and ∞ if p > 0 — one of the two, nothing between. Neither 0 nor ∞ sits in [1, 3]. Contradiction — in the blanket, not in the meter: every step was a theorem. So the assumption V ∈ ℒ is what dies. ∎

Warning — the failure is not about size

Do not file this under "V is too big" or "too complicated". The Cantor set is vastly more intricate than V — uncountable, fractal, singular — and it is measurable, of measure zero. The Vitali set fails not by complexity but by arithmetic of overlaps: its translates must tile the interval exactly, and any price per tile is inconsistent with the tiling. Measurability is a knife property (3.1), and V is a bad knife, period. No amount of description-making will fix it: no formula for V will ever appear, as §5 explains.

3.4.4The stronger statement: no measure sees V

Theorem 1 says λ does not see V. The sharper truth — the one that closes the book on Chapter 0's original question — is that the failure has nothing to do with Lebesgue's construction specifically.

Theorem 2 — no length-like measure sees V

Let μ be any countably additive measure on any σ-algebra F with ℬ(ℝ) ⊆ F, such that μ assigns finite positive mass to [0,1] and is shift-consistent: μ(A + t) = μ(A) for all shifts t and all measurable A. Then V ∉ F.

Proof — the same ledger, no Lebesgue anywhere

The proof of Theorem 1 used exactly four properties of λ: countable additivity over the disjoint family; shift-invariance; monotonicity; and the interval prices μ([0,1]) = m ∈ (0, ∞), μ([−1,2]) = 3m (the last from additivity over the three unit intervals). Every one is in the hypothesis list of Theorem 2. Run the same argument: if V ∈ F, the disjoint countable family of translates gives m ≤ ∑q p ≤ 3m, with p = μ(V) constant across translates — and the constant-sum dichotomy kills it again. So V belongs to no σ-algebra that carries any shift-faithful probability of the line. The wall is not Lebesgue's taste; it is arithmetic. ∎

Example 2 · What dies with V, concretely

Concretely, the sentence "the uniform random real lands in V" is not a question the model can hear: {uniform real ∈ V} is not an event of ([0,1], ℒ, λ). Not probability zero — not anything. This is the precise meaning of the 101-era slogan "not all subsets are events": there is a specific, constructible-in-principle set that no probability agreeing with length can weigh. Every restriction of the domain we have imposed since 1.1 — the σ-algebra, the tollgate, the whole architecture — exists to route the theory around exactly this object.

3.4.5Why no explicit example exists

Proposition — measurability is description-proof

Any subset of ℝ that can be built from intervals by countably many operations of complement, countable union, countable intersection, and pointwise limits of sequences of previously built sets — the full grammar of "writing down a set" in this course — is Borel, hence measurable. Every describable set is measurable; non-measurability is undescribable by construction.

Proof sketch — the grammar never leaves ℬ

ℬ is closed under each operation in the list: the first three by definition (1.1, 1.2); pointwise limits via the countable-sandwich construction of {Xn → X} from 2.2 — the convergence event of measurable functions is measurable, so limsup and liminf of sequences of Borel-describable sets are Borel. Induction over the (countable) grammar depth closes it: every describable set is Borel, and Borel ⊆ ℒ. So a non-measurable set can only be reached by leaving the grammar — which is exactly what the axiom of choice does when it selects one delegate per equivalence class, all at once, with no rule. ∎

Solovay's caveat, honestly stated

In the other direction, a deep theorem of Solovay (1970) says: in a set-theoretic universe without the axiom of choice (but with an inaccessible cardinal), every definable set of reals is Lebesgue measurable — no Vitali set can be built at all. The two statements bracket the situation: with choice, non-measurable sets exist but can never be named; without choice, everything nameable is measurable but the theory of 2.3 and beyond loses tools. Probability chose its side: keep choice, keep the wall, and route all questions through the σ-algebra. We flag it once here and never again.

3.4.6What the wall is made of

Example 3 · Positive inventory: what the wall permits

End Chapter 3 with the inventory of what we may now assert without flinching. Length exists, is unique, and lives on ℒ — a σ-algebra so large it contains every set the operations of analysis can produce (1.2's closure, 2.2's limits, this subchapter's grammar), is complete (3.2), and is shift-faithful. Null sets — uncountably large ones like the Cantor set of 3.3 — are handled and safely ignorable. And the wall itself is now surveyed with precision: exactly the sets that would tile the line inconsistently, all of them choice-made, none of them nameable. The modeler's contract is exact: declare your questions in the grammar of 1.2, and every question you can ever actually ask gets a number.

Example 4 · The wall in higher dimensions

The same construction, run with rotations instead of shifts, builds non-measurable sets in the plane — this is the seed of the Banach–Tarski paradox: a ball can be cut into finitely many choice-made pieces and reassembled into two balls of the original size. The pieces are non-measurable Vitali-style sets; the paradox is precisely the price of pretending they have volume. One more reason the course's response — restrict the domain to the measurable, keep the theorems — is not bureaucracy but self-defense.

The non-measurable sets are the receipt for the axiom of choice: they exist because we can choose, cannot be named because naming is countable, and cannot be measured because arithmetic does not allow it.

Where we are, and what is next

Chapter 3 is complete: price, tollgate, measure, null sets, and the wall. The sample space now has honest continuous models. What is missing is the main character: the random variable — a function from outcomes to numbers — and the quiet measurability constraint that decides which functions deserve expectations, distributions, and limit theorems. Chapter 4 begins with that constraint, and it is stricter than 101 ever let on.

Powered by Webloyt | Report abuse