The list of measurable questions
The previous guide left us with bad news: there is no way to hand every subset of [0,1] a length that respects shifting and countable additivity at once. A non-measurable set exists, and it wrecks any attempt to measure literally everything. The fix is not to give up on probability — it is to give up the fantasy that every subset deserves a probability. We keep a curated list of subsets, the ones we are willing to call genuine events, and we only ever assign probabilities to members of that list.
Think of an event as a yes-or-no question about the random outcome: "did the dart land in the left half?", "is the waiting time under five minutes?". The list of measurable questions is the set of questions we promise to be able to answer with a probability. A sigma-algebra (written with a script F) is exactly such a list — but not just any list. It has to be closed under the natural ways we combine questions, so that asking sensible follow-up questions never throws us off the list.
This is a real shift in attitude from the earlier rungs, where you happily said "let A be any event." Here the sample space comes with a chosen family of allowed events bolted on. The pair (sample space, sigma-algebra) is called a measurable space, and it is the stage on which all of probability will now be performed — before a single probability is even assigned.
Three rules that make a sigma-algebra
A family F of subsets of the sample space Omega is a sigma-algebra if it obeys three rules. First, the whole space Omega is in F (the question "did anything happen?" is always answerable, with answer yes). Second, if a set A is in F, then its complement Omega minus A is in F too (if you can ask "did A happen?", you can ask "did A fail?"). Third — and this is the heart of it — if A_1, A_2, A_3, ... is a sequence of sets in F, then their union A_1 union A_2 union ... is also in F. Crucially that third rule allows a countably infinite list of sets, not just two or three.
A sigma-algebra F over Omega satisfies: (1) Omega is in F (2) A in F => (Omega \ A) in F (complement) (3) A_1, A_2, ... in F => (A_1 u A_2 u ...) in F (countable union) Free consequences: - empty set is in F (complement of Omega) - countable intersections in F (De Morgan: complement of a union)
From these three you get the rest for free. The empty set is in F because it is the complement of Omega. Countable intersections are in F because, by De Morgan's laws, an intersection is the complement of a union of complements — and all three of those operations stay inside F. So a sigma-algebra is automatically closed under unions, intersections, complements, and differences, taken any countable number of times. That "countably infinite" reach is the single feature separating a sigma-algebra from the simpler algebra of events, which only promises closure under finite operations.
Two extremes and the one we actually use
For any sample space there are two trivial sigma-algebras. The smallest possible is { empty set, Omega } — it can only answer "did anything happen?". The largest possible is the power set, the family of all subsets, which tries to answer every conceivable question. On a finite or countable Omega — say a die or a coin sequence — the power set is harmless and we use it without a second thought. The trouble only appears on uncountable spaces like an interval, where, as the last guide showed, the power set contains non-measurable monsters.
So on the real line we deliberately use something between the two extremes: the Borel sigma-algebra. The recipe is elegant. Start with the questions you genuinely care about — every interval, every "is X less than 7?" — and then take the smallest sigma-algebra that contains all of them. "Smallest containing" is a real operation: the intersection of all sigma-algebras that include your starting questions is itself a sigma-algebra, and it is the tightest possible list that still answers everything you asked for. That generated list is the Borel sets.
The payoff is that the Borel sets are vast — every interval, every open and closed set, every countable union and intersection of them, and far stranger sets too — yet they are not everything. The non-measurable set from the previous guide is deliberately left out. We bought consistency by drawing a boundary, and the Borel sigma-algebra is where we draw it: rich enough to contain every set you will ever actually want, lean enough to never contain a contradiction.
Adding the measure: the probability space
Now we put weight on the stage. A probability measure P is a function that takes a set from F and returns a number in [0,1], obeying the Kolmogorov axioms you met in the very first rung — but now seen in their full, measure-theoretic light. First, P(A) is at least 0 for every event A. Second, P(Omega) = 1: something certainly happens. Third, P is countably additive: if A_1, A_2, ... are pairwise disjoint events, then P(A_1 union A_2 union ...) equals P(A_1) + P(A_2) + ... as an infinite sum.
Notice how the sigma-algebra and the measure are made for each other. The third axiom, countable additivity, talks about a countable union of events — and the sigma-algebra guarantees that union is itself an event, so P is entitled to be applied to it. The list of measurable questions and the rule for assigning probabilities are not two separate decisions; the closure rules of F are precisely what makes P's defining property well-posed. They are two halves of one machine.
The complete object — sample space, sigma-algebra, and measure together — is the probability space, written as the triple (Omega, F, P). Every random thing in this course, no matter how elaborate, ultimately lives on some probability space. A coin flip, a Brownian path, a queue running for all time: each is just a different choice of these three ingredients. When a later guide says "let X be a random variable," there is always one of these triples humming quietly underneath.
What the new machinery buys you
This is more than tidy bookkeeping; the axioms work for a living. From countable additivity flows continuity of probability: if events A_1 is contained in A_2 is contained in A_3 ... grow up to a limit event A, then P(A_n) climbs to P(A); and if events shrink down, P shrinks down to match. This is the property that lets you take limits inside probabilities at all — without it you could not say "P(the sequence eventually exceeds 100)" equals the limit of the probabilities, and the laws of large numbers in this very rung would have no ground to stand on.
Here is a small but striking dividend. On the interval [0,1] with Borel sets and ordinary length as P, a single point {x} is a legitimate event, and its probability is 0 — its length is zero. Yet the whole interval, a union of all its points, has probability 1. There is no contradiction, because additivity only adds over countable collections, and the points of [0,1] are uncountable. This is the rigorous foundation for a fact you have used loosely for guides on end: for a continuous variable, any single value has probability zero, and a density is a rate of probability, not a probability itself.
Building one in practice
You almost never specify P on every Borel set by hand — there are uncountably many of them. Instead you pin P down on a small, friendly collection of generating sets and let a theorem extend it to the whole sigma-algebra. On the real line you simply declare P((-infinity, x]) = F(x), where F is the cumulative distribution function you already know from earlier rungs. That assigns probability to intervals only — yet because intervals generate the Borel sets, there is exactly one probability measure agreeing with that choice. The cdf you have used all along is, secretly, the compact instruction set for a full measure.
- Choose the sample space Omega — the set of all possible outcomes (a die's six faces, the interval [0,1], the space of infinite coin sequences).
- Choose the sigma-algebra F — the power set if Omega is countable, the Borel sigma-algebra if Omega is the real line or an interval.
- Specify P only on simple generating sets: probabilities of individual outcomes (discrete), or P((-infinity, x]) = F(x) via a cdf (continuous).
- Invoke the extension theorem: it promises one and only one measure on all of F that matches your specification, so (Omega, F, P) is now fully determined.
With the probability space firmly in place, the next guide can finally say what a random variable really is. It is not a vague "variable that varies" — it is a function on Omega that respects the sigma-algebra, a measurable function. Every question you can ask about the random variable must trace back to a question already on the list F. The careful stage we built here is exactly what makes that definition sing.