Probability spaces are models, not urn pictures
Objective: For “Probability spaces are models, not urn pictures,” which exact hypothesis justifies the decisive step, and what is the first conclusion lost under this change: Assigning probabilities to an informal list without specifying which unions are events can make later conditioning or limiting events undefined.
Unit 1, chapter 1 begins with one question that can be answered from declared probability data: “Probability spaces are models, not urn pictures.” A probability problem begins by deciding which outcomes are distinct and which questions the model is allowed to answer. We will not start by moving symbols. We first name the population of possible outcomes, the information available, the random object being measured, and the numerical claim that must be recovered.
Read the definition one piece at a time: A probability space is a triple (Ω,F,P): Ω is the outcome set, F is a sigma-algebra of events, and P is a countably additive measure with P(Ω)=1. An expression such as E[X|G], P(X∈A), X_n→X, or sup_t X_t is meaningful only after the probability space, sigma-algebra, version convention, index set, and integrability assumptions needed by that expression have been stated.
The chapter proves this bounded statement: Finite additivity follows from countable additivity, and complements satisfy P(Aᶜ)=1−P(A) for every A∈F. Its role is not to supply a slogan. It turns a list of hypotheses into one output whose mode of equality or convergence is explicit, so almost-sure, L¹, probability, and distributional conclusions cannot be silently exchanged.
The derivation route is: Write Ω and F, split disjoint unions, apply countable additivity with empty tails, then use Ω=A⊔Aᶜ. At each arrow we identify the theorem being used, the quantity held fixed, and the limiting operation. Conditional expectations are compared through their integral identities; convergence claims are compared through the exact topology or norm named in the statement.
A worked calculation makes the abstract statement inspectable: For two binary sensors, take Ω={00,01,10,11} with masses (1,2,3,4)/10; compute the event that at least one sensor is on and its complement. We retain the raw atoms, probabilities, conditioning cells, partial sums, or transition rows before forming derived quantities. The verified endpoint is The event is {01,10,11} with probability 9/10; its complement {00} has probability 1/10, and the two values sum to one.
The first invalid step under changed assumptions is part of the lesson: Assigning probabilities to an informal list without specifying which unions are events can make later conditioning or limiting events undefined. A responsible solution therefore reports what remains true, what additional domination, tightness, uniform-integrability, stopping, or measurability condition would repair the argument, and what the finite calculation does not prove.
Finite additivity follows from countable additivity, and complements satisfy P(Aᶜ)=1−P(A) for every A∈F.
Translate “A probability space is a triple (Ω,F,P): Ω is the outcome set, F is a sigma-algebra of events, and P is a countably additive measure with P(Ω)=1.” into a typed statement on one declared probability space. Mark every event, random variable, sigma-algebra, measure, norm, and limit with its domain and required integrability.
Establish the first implication in the route: Write Ω and F, split disjoint unions, apply countable additivity with empty tails, then use Ω=A⊔Aᶜ. For a conditional object, test the defining integral on generating events; for a limit, isolate the probability, moment, transform, or compactness estimate that is uniform in the index.
Pass from the local estimate to the stated conclusion without strengthening the mode. Null sets may depend on the index until a countable intersection is taken, and weak convergence controls bounded continuous tests rather than arbitrary unbounded functions.
Check the theorem against the finite model: For two binary sensors, take Ω={00,01,10,11} with masses (1,2,3,4)/10; compute the event that at least one sensor is on and its complement. Compute with exact fractions or declared decimal precision, and compare the resulting quantity with the theorem’s prediction rather than with an unrelated simulation trace.
This proves only “Finite additivity follows from countable additivity, and complements satisfy P(Aᶜ)=1−P(A) for every A∈F..” The proof stops before the failure boundary: Assigning probabilities to an informal list without specifying which unions are events can make later conditioning or limiting events undefined. Crossing that boundary requires a new theorem or a weaker claim, not a change of notation.
For two binary sensors, take Ω={00,01,10,11} with masses (1,2,3,4)/10; compute the event that at least one sensor is on and its complement.
- Write the complete finite input for “For two binary sensors, take Ω={00,01,10,11} with masses (1,2,3,4)/10; compute the event that at least one sensor is on and its complement.”: label atoms or states, attach probabilities or transition weights, and verify nonnegativity, total mass, and every stated conditioning event.
- Apply the definition A probability space is a triple (Ω,F,P): Ω is the outcome set, F is a sigma-algebra of events, and P is a countably additive measure with P(Ω)=1. before using the theorem. Preserve exact numerators and denominators so a zero-probability cell, missing tail, or normalization error remains visible.
- Follow the route Write Ω and F, split disjoint unions, apply countable additivity with empty tails, then use Ω=A⊔Aᶜ. and calculate the intermediate conditional mean, transform, bound, variance, or stopping quantity requested by the problem.
- Independently recompute the endpoint and compare both paths. They agree at The event is {01,10,11} with probability 9/10; its complement {00} has probability 1/10, and the two values sum to one. Record the exact hypothesis that made the agreement valid.
Result: The event is {01,10,11} with probability 9/10; its complement {00} has probability 1/10, and the two values sum to one.