Kolmogorov's Axiomatization

Before 1933, "probability" named a family of computational practices — counting equally likely cases, taking ratios of areas, appealing to limiting frequencies — with no common object and no agreement on what a probability is. Kolmogorov's Grundbegriffe der Wahrscheinlichkeitsrechnung replaced all of it with three lines of measure theory, and the subject has been a branch of analysis ever since.

This page states the axioms, is explicit about what they settle and what they leave free, and recovers the pre-axiomatic models as special cases. Everything here is a translation of measure theory into probabilistic vocabulary; the translation is the content.

References: Kolmogorov, Grundbegriffe (1933); Billingsley, Probability and Measure, ch. 1–2; Williams, Probability with Martingales, ch. 1.

1. The Axioms

Definition (probability space). A probability space is a triple $(\Omega ,\mathcal{F},P)$ where

1. $\Omega$ is a set — the sample space, whose points $\omega$ are outcomes;
2. $\mathcal{F}$ is a $\sigma$-algebra on $\Omega$ — the events;
3. $P:\mathcal{F}\to [0,1]$ is a measure with $P(\Omega )=1$ — the probability measure.

Unpacked, (3) says: $P(A)\ge 0$; $P(\Omega )=1$; and for pairwise disjoint $A_1,A_2,\dots \in \mathcal{F}$,

$$P\left(\underset{k=1}{\overset{\infty }{⨆}}{A}_k\right)=\sum _{k=1}^{\infty }P(A_k).$$

That is the entire foundation. Every elementary rule follows from measure-theory.md §3 with no probabilistic input:

• $P(A^c)=1-P(A)$, $P(\varnothing )=0$, and $A\subseteq B⟹P(A)\le P(B)$;
• inclusion–exclusion $P(A\cup B)=P(A)+P(B)-P(A\cap B)$;
• union bound $P({\bigcup }_k{A}_k)\le {\sum }_{k}P(A_k)$ — countable subadditivity, and the most-used inequality in the subject;
• continuity $A_k\uparrow A⟹P(A_k)\uparrow P(A)$, and likewise downward (no finiteness caveat is needed here: $P(\Omega )=1<\infty$).

A definition, not a discovery. The axioms do not analyse a pre-existing concept; they stipulate a mathematical structure and appropriate the word. Their justification is downstream: the structure supports the limit theorems (Part B), and the classical calculations come out right (§3). Whether this structure is the right model of any particular real-world uncertainty is a question the axioms do not address and cannot settle.

2. What Is Fixed and What Is Free

The axioms fix the calculus of probability completely: given $(\Omega ,\mathcal{F},P)$, every probabilistic question has a determinate answer. They fix nothing else. Three things remain entirely free, and confusion about probability is usually confusion about one of them.

(a) The sample space $\Omega$. Nothing selects it. Two rolls of a die can be modelled on $\{1,\dots ,6{\}}^2$, or on $\{2,\dots ,12\}$ (sums only), or on $[0,1]$ with a suitable $P$. All are legitimate; they differ in which questions can be asked, not in the answers to shared questions. Since random variables show that only the induced law ever matters, $\Omega$ is best thought of as scaffolding.

(b) The $\sigma$-algebra $\mathcal{F}$. For finite or countable $\Omega$ one takes $\mathcal{F}=\mathcal{P}(\Omega )$ and the choice is invisible. For uncountable $\Omega$ it cannot be (non-measurable.md), so $\mathcal{F}$ is a genuine part of the model. It acquires a second reading in conditional-expectation.md: a sub-$\sigma$-algebra $\mathcal{G}\subseteq \mathcal{F}$ represents partial information — the events one can decide. The technical necessity and the informational interpretation are the same object.

(c) The measure $P$. The axioms constrain $P$ only to be a normalized measure. Which measure models a given situation is an input to the theory, never an output of it. §4 shows how badly this bites.

3. The Classical Models Recovered

Finite equally likely cases. $\Omega$ finite, $\mathcal{F}=\mathcal{P}(\Omega )$, $P(A)=|A|/|\Omega |$. Countable additivity is trivial, and one recovers the classical definition — "favourable over possible" — as a special case rather than a definition. The circularity that plagued the classical account (equally likely = equally probable) disappears: uniformity is now a property of a chosen $P$, not a prior notion that defines it.

Countable spaces. $\Omega =\{{\omega }_1,{\omega }_2,\dots \}$ with weights $p_k\ge 0$, ${\sum }_k{p}_k=1$; then $P(A)={\sum }_{{\omega }_k\in A}{p}_k$. Every probability measure on a countable space is of this form, and countable additivity is exactly what makes the sum well defined. Note that no uniform distribution on $\mathbb{N}$ exists: constant weights sum to $0$ or $\infty$. This is the first sharp consequence of $\sigma$-additivity, and the starting point of finite-additivity.md.

Geometric probability. $\Omega \subseteq {\mathbb{R}}^n$ with $0<{\lambda }^n(\Omega )<\infty$, $\mathcal{F}={\mathcal{L}}^n{|}_{\Omega }$, and $P(A)={\lambda }^n(A)/{\lambda }^n(\Omega )$: normalized Lebesgue measure. This is the "pick a point at random" model, and its legitimacy is exactly the construction of Lebesgue measure (measure-theory.md §6).

Coin flips. $\Omega =\{0,1{\}}^{\mathbb{N}}$ with the product measure of §5 of product-measures.md. This space exists — a fact requiring proof — and via binary expansion it is measure-isomorphic to $([0,1],{\lambda }^1)$. So the infinite fair-coin sequence and the uniform random real are the same probability space in different clothing.

4. Bertrand's Paradox

The problem (Bertrand, 1889). A chord is drawn "at random" in the unit circle. What is the probability that it is longer than the side $\sqrt{3}$ of the inscribed equilateral triangle?

Three natural constructions give three answers:

Each is a perfectly good probability space; they are different spaces. The "paradox" is that the phrase at random names none of them.

What the paradox shows. Not that probability is ill-founded, but that §2(c) is unavoidable: a probability model is not determined by a verbal description of a random experiment. A distribution must be specified, and no principle internal to the axioms supplies one. Attempts to legislate a canonical answer — the principle of indifference, maximum entropy, invariance under the problem's symmetry group (Jaynes' proposal, which selects $1/2$) — are additional modelling assumptions, defensible or not on a case-by-case basis, and none of them is a theorem.

The same underdetermination recurs whenever "uniform" is applied to a set with no canonical measure, and it is the technical core of the Borel–Kolmogorov paradox about conditioning on null events.

5. Why Countable Additivity?

Finite additivity is what the classical calculations require; the third axiom asks for more. It is the one substantive choice in the definition, and it is worth being explicit about what it buys:

• Continuity along sequences. $A_k\uparrow A⟹P(A_k)\to P(A)$ is equivalent to countable additivity given finite additivity. Without it, "the probability that the sequence eventually does $X$" cannot be approached by finite approximations.
• The convergence theorems. Monotone and dominated convergence (lebesgue-integral.md §4) are countable additivity applied to functions. Every limit theorem in Part B is downstream of them.
• The Borel–Cantelli lemmas (independence.md), hence almost-sure statements about infinite sequences, hence the strong law of large numbers.

What it costs: no uniform distribution on $\mathbb{N}$ or on $\mathbb{R}$; non-measurable events must be excluded; and conglomerability constraints that some have argued are too strong (finite-additivity.md). The axiom is a choice with consequences, not an analytic truth about the word "probability".

6. Events as an Algebra

$\mathcal{F}$ is a Boolean $\sigma$-algebra, and $P$ a normalized, countably additive, strictly positive-on-nonnull functional on it. Two structural remarks that matter downstream:

• Quotienting by null sets yields the measure algebra $\mathcal{F}/\mathcal{N}$, a complete Boolean algebra on which $P$ is strictly positive. Probability is really a theory of measure algebras: the point-set $\Omega$ is, again, scaffolding. (Every nonatomic separable measure algebra is isomorphic to that of $([0,1],{\lambda }^1)$ — all such probability spaces are the same space.)
• Distributivity is a real assumption. $\mathcal{F}$ being Boolean encodes that all events have simultaneous truth values, so joint probabilities always exist. Dropping distributivity — replacing $\mathcal{F}$ by the projection lattice of a Hilbert space — gives quantum probability, where joint distributions for incompatible observables do not exist and Bell/Kochen–Specker say this is not repairable.

7. Where This Is Used

• Random variables and distributions — the next step: what we actually compute with.
• Countable vs. finite additivity — §5 revisited from the other side.
• Quantum probability — §6's distributivity assumption dropped.
• Non-measurable sets — why §2(b) cannot be avoided.