Sample Space, Events, and the Kolmogorov Axioms

What a sample space actually is

Every probability calculation starts with a question that sounds almost too simple to matter: what could happen? The full list of every distinct outcome of a random experiment is called the sample space, usually written Ω (a capital omega). Roll a standard six-sided die and the sample space is Ω = {1, 2, 3, 4, 5, 6}. Flip a coin twice and it's Ω = {HH, HT, TH, TT}. The outcomes in a sample space must be mutually exclusive (only one happens per trial) and collectively exhaustive (between them they cover everything that could possibly occur).

Sample spaces can be finite, as in the die example, countably infinite (the number of coin flips until you first see heads: Ω = {1, 2, 3, ...}), or uncountably infinite (the exact time in seconds until a component fails: Ω = [0, ∞)). The type of sample space determines the mathematical machinery you need — simple counting for finite cases, sums of series for countable ones, and integrals for continuous ones — but the underlying logic stays the same across all three.

Events as subsets of the sample space

An event is any subset of the sample space — a collection of outcomes you've decided to group together because you care about whether one of them happens. "Rolling an even number" is the event A = {2, 4, 6}. "Rolling a number greater than 4" is B = {5, 6}. Because events are just sets, all the usual set operations carry meaning: A ∪ B ("A or B") is the set of outcomes where at least one event happens; A ∩ B ("A and B") is the set where both happen; and the complement A^c is everything in Ω that is not in A.

Two events are called mutually exclusive (or disjoint) if they share no outcomes, meaning A ∩ B = ∅. Rolling an even number and rolling exactly a 1 are mutually exclusive; rolling an even number and rolling a number greater than 4 are not, because they both contain the outcome 6. This distinction turns out to be the single most common source of student error when applying the addition rule, covered below.

The three Kolmogorov axioms

In 1933 the Russian mathematician Andrey Kolmogorov gave probability theory the axiomatic foundation it had lacked for centuries. Instead of trying to define what probability "really is" in a philosophical sense, he specified three rules that any valid probability assignment must obey. Everything else in probability theory — every formula you will ever use — is a logical consequence of these three statements.

Axiom 1 (non-negativity): for every event A, P(A) ≥ 0. Probabilities cannot be negative.

Axiom 2 (normalisation): P(Ω) = 1. Something in the sample space must happen; the total probability of all possible outcomes is exactly 1.

Axiom 3 (countable additivity): for any sequence of mutually exclusive events A₁, A₂, A₃, ... (pairwise disjoint, meaning no two share an outcome), P(A₁ ∪ A₂ ∪ A₃ ∪ ...) = P(A₁) + P(A₂) + P(A₃) + ... . In plain terms: if events cannot happen simultaneously, the probability that at least one of them happens is just the sum of their individual probabilities.

Deriving the rules you already use

Every familiar probability formula falls out of these three axioms with a few lines of algebra. Because A and A^c are disjoint and together make up Ω, Axiom 3 gives P(A) + P(A^c) = P(Ω) = 1, so P(A^c) = 1 − P(A). Because ∅ and Ω are disjoint and their union is Ω, the same reasoning shows P(∅) = 0. And because Ω itself is an event, Axiom 1 combined with Axiom 2 forces every probability to sit in the closed interval [0, 1] — you cannot assign a probability of 1.5 or −0.2 to anything.

The rule most people reach for constantly is the general addition rule: P(A ∪ B) = P(A) + P(B) − P(A ∩ B). This is not a fourth axiom; it is a theorem proved from Axiom 3 by noting that A ∪ B can be split into three disjoint pieces (A only, B only, and A ∩ B), summing them, and then subtracting the overlap that got counted twice. When A and B happen to be mutually exclusive, P(A ∩ B) = 0 and the formula collapses back to simple addition, which is exactly Axiom 3 in its two-event form. This same inclusion–exclusion pattern extends to three or more overlapping events and is the backbone of combinatorial probability problems.

Visualizing sample spaces with Venn diagrams and trees

Two visual tools do most of the heavy lifting when reasoning about events. A Venn diagram draws the sample space as a rectangle and events as overlapping circles inside it; the area of each region is proportional to its probability, which makes the addition rule's "subtract the double-counted overlap" logic visually obvious. A probability tree is better suited to sequential experiments — each branch splits into the possible outcomes of the next stage, branch probabilities multiply along a path, and the probabilities of all branches leaving any single node always sum to 1, a direct visual echo of Axiom 2 applied locally at each decision point.

For a discrete sample space like dice or cards, it also helps to think in terms of counting: if every outcome in a finite Ω is equally likely, then P(A) simply equals |A| / |Ω|, the number of outcomes favourable to A divided by the total number of outcomes. This "classical" definition of probability is not a separate theory — it is one particular way of assigning numbers to outcomes that automatically satisfies all three Kolmogorov axioms, which is precisely why it works.

Why the axioms matter beyond the classroom

The payoff of an axiomatic foundation is consistency. Any time someone proposes a "probability" for something — a weather forecaster's 70% chance of rain, a poker player's odds of hitting a flush, a Bayesian's degree of belief in a hypothesis — the Kolmogorov axioms give a checklist for whether that number is coherent. If someone claims a 60% chance of rain and a 55% chance of no rain, the axioms immediately flag the contradiction, because those two complementary events must sum to exactly 100%. This is also why models in finance, medicine, and machine learning that output "probabilities" are held to the same three rules: a spam filter that assigns overlapping categories probabilities summing to more than 1 has a bug, not a philosophical quirk.