Why "Function" Isn't Enough: Measurability
A random variable is a function from outcomes to numbers — but not every such function earns the name. The constraint that separates them is one line long, invisible in practice, and the load-bearing wall of the entire theory.
In this subchapter
4.1.1The question every function must pass
Fix a probability space (Ω, F, P) — by now a concrete object: Chapter 3 built ([0,1], ℒ, λ) by hand. A function X: Ω → ℝ wants to be called a random variable. The test is forced by the course's own logic: probabilities are assigned to events, and the questions we ask about X are of the form "is X ≤ t", "is X in (a,b)", "is X positive". Each of these is the inverse image of a Borel set:
X: Ω → ℝ is a random variable if
the inverse image of every half-line (−∞, t] is an event. Unpacking the notation: {X ≤ t} is shorthand for the set of outcomes on which the function's value is at most t. The definition says nothing about the values of X — it is a constraint on which sets of outcomes F can see.
Why half-lines? Three reasons, and each is a theorem from earlier chapters in disguise. First, (−∞, t] generates ℬ(ℝ) — Theorem 1 of 1.3, door (3) — so if the half-line questions are answerable, every Borel question about X is answerable (that is §3). Second, half-lines are exactly what a distribution function reports, so the definition is tailored to the object we will study in 4.2. Third, half-lines form a π-system — the shape on which uniqueness arguments run (1.4).
4.1.2What measurability buys: the transfer theorem
If X is a random variable, then {ω : X(ω) ∈ B} ∈ F for every Borel set B ⊆ ℝ. In words: once the half-line questions are events, every question the Borel grammar can phrase about X is an event.
Let 𝒢 = { B ⊆ ℝ : {X ∈ B} ∈ F }, the collection of target sets whose inverse images are events. Then: ℝ ∈ 𝒢 (inverse image is Ω); B ∈ 𝒢 implies Bc ∈ 𝒢 (inverse image of complement is complement of inverse image); B₁, B₂, … ∈ 𝒢 implies ⋃Bn ∈ 𝒢 (inverse image commutes with unions — draw the commutative square in your head: membership flows through the function either way). So 𝒢 is a σ-algebra, and it contains every half-line by hypothesis. Since the half-lines generate ℬ — the smallest σ-algebra containing them — minimality (1.2) gives ℬ ⊆ 𝒢. ∎
Count the sightings of this argument: the two-quantifier trick in the π–λ theorem (1.4), the good-sets d-system in the uniqueness theorem (1.4), and now inverse images. It is the same skeleton each time: show the well-behaved collection is a σ-algebra (or d-system) containing a generator; conclude by minimality. From here on the course uses the pattern without ceremony, and you should recognize it the way a programmer recognizes a for-loop.
On ([0,1], ℒ, λ): the identity X(ω) = ω is a random variable ({X ≤ t} = [0, t] for t ∈ [0,1], a measurable set — Lemma 4 of 3.2). The floor function ⌊10ω⌋, the first-digit extractor, the indicator of the Cantor set — all random variables, because their level sets are Borel-carved. This is the uniform random real of 0.2 at last promoted from metaphor to object.
4.1.3A genuine non-example
On ([0,1], ℒ, λ) let V be the Vitali delegate set (3.4) and define X = 1V: the function that is 1 on V, 0 elsewhere. Perfectly well-defined as a function — the Vitali set exists. Not a random variable: {X ≤ 1/2} = Vc, and V ∉ ℒ (Theorem 1 of 3.4), so the test fails at t = 1/2. The sentence "is X at most one half?" — a question about coin-flip-shaped data — is one the model cannot hear. The measurability constraint is not a formality: it excludes specific functions, and the excluded ones are exactly the ones grown from non-measurable sets.
Same space, but now equip it with the coarse σ-algebra F = {∅, [0,1/2), [1/2,1], Ω} — the bookmaker of 2.3 Example 1. The identity function is no longer a random variable: {X ≤ 1/4} = [0, 1/4] ∉ F. Nothing about the function changed; the questions the model can hear changed. Measurability is a relationship between a function and a σ-algebra — never say "X is measurable" without mentally appending "with respect to F". This example also previews 4.3: a coarser F means less information, and functions can fail to be measurable simply by knowing too much fine detail for the σ-algebra to record.
4.1.4Almost sure equality: the first use of null sets
A statement about outcomes holds almost surely if it holds on an event of probability 1 — that is, the set where it fails is contained in a measurable set of probability zero (contained, thanks to completeness: 3.2 §6). Two random variables are equal almost surely, X = Y a.s., if P({ω : X(ω) ≠ Y(ω)}) = 0.
On ([0,1], ℒ, λ): X(ω) = ω and Y(ω) = ω except Y(ω) = 0 at the single point ω = 1/2. The two functions differ on a singleton — null (3.1 Example 2) — so X = Y a.s. For every purpose the course will ever have — distribution, expectation, convergence theorems — they are the same object. This is the first payout of the entire null-set apparatus of 3.3: the license to say "ignore that", with a proof behind it.
Here is the sharp edge, and the reason the theory keeps the definition strict. Let X ≡ 0 (a random variable) and let Y = 1N where N is a subset of the null Cantor set that is not measurable — such N exists, and completeness is what makes the subset-of-null language precise (3.2 §6, Example 2). Then X = Y a.s. (they differ only on N, contained in a null set), but Y is not measurable — its level set is N itself. So: almost sure equality can promote a non-measurable function to "equal a.s. to a nice one", and the course's remedy is a definition convention rather than a theorem — a function equal a.s. to a measurable function is declared measurable only when its level sets are genuinely in F (in complete spaces, subsets of null sets are measurable, which rescues this case: N is ℒ-measurable. In incomplete spaces the rescue fails). The honest takeaway: a.s. statements live and die by completeness, and the theory's tolerance for null-set fudging is exactly as large as Chapter 3's insurance policy.
4.1.5Operations that preserve measurability
If X, Y are random variables and c ∈ ℝ, then so are: X + c, cX, X + Y, XY, max(X,Y), min(X,Y), and |X|. If X₁, X₂, … are random variables, so are inf Xn, sup Xn, liminf Xn, limsup Xn — and the set where the limit exists, with the limit itself a random variable there.
max(X,Y) ≤ t iff both X ≤ t and Y ≤ t: so the level set is an intersection of two events — closure clause. X + Y < t iff for some rational q, X < q and Y < t − q: a countable union over q ∈ ℚ of events — the density-of-ℚ move again, taming the uncountable. sup Xn ≤ t iff Xn ≤ t for all n: a countable intersection. Liminf/limsup are sups of infs. And "{limit exists}" is the countable sandwich of 2.2 §6 — the proposition there now pays out in functions, not just events. Every rule of Chapter 1's grammar has become a rule of function algebra. ∎
The theorem is all countable. The supremum of an uncountable family of random variables can fail to be a random variable: index over t ∈ ℝ the functions Xt = 1{t} on ([0,1], ℬ, λ) — each is measurable, but supt Xt ≡ 1... choose instead the family Xt = 1(−∞,t) composed with the identity and ask for inf: the point is structural — an uncountable inf/sup has level sets built from uncountable unions, and closure clauses do not reach them. The clean, standard instance: on ([0,1], ℬ), the supremum over all t of 1V+t mod 1-type families drags Vitali sets back in. The moral needs no exotic instance: the Borel grammar is countable; operations respect it; every theorem you use should be checked for a hidden uncountable step — this is exactly the brittleness-flagging habit 1.1 promised to train.
A quant's P&L is a function of market outcomes: P&L = h(X₁, …, Xk) where h is the payoff function — continuous piecewise-linear for options, with kinks. Theorem 2 plus one more reduction (a continuous h of measurable functions is measurable: {h(X) ≤ t} is X pulled back through the Borel set h−1((−∞,t]), open/closed preimages under continuous h are Borel — Theorem 1 does the rest) certifies: the P&L is a random variable, "P&L below −VaR" is an event, and everything in Chapters 5–7 will apply to it. The kinks that terrify Riemann-based intuition (5.3) cost measurability nothing.
A random variable is a function that asks only questions the σ-algebra can hear. Half-lines are enough to ask everything — and the Vitali set is the question that cannot be asked.
Random variables exist, form an algebra closed under countable operations, and respect the a.s. discipline. What we have not done: assign them numbers. The next subchapter builds the distribution function — the complete compression of a random variable's law into one right-continuous, monotone function on ℝ — and proves, via the π–λ machinery of 1.4, that this compression loses nothing.