Countable vs. Finite Additivity

Kolmogorov's third axiom asks for countable additivity. Finite additivity — the version the classical calculations actually use — would be strictly weaker, and the strengthening was a deliberate choice, described by Kolmogorov himself as a convention justified by its mathematical fruitfulness rather than by anything intrinsic to the concept.

This page examines the choice from the other side: what a merely finitely additive probability looks like, what becomes possible, and what collapses. The conclusion is not that $\sigma$-additivity is wrong but that it is a substantive axiom with identifiable costs and benefits — the clearest single illustration of §5 of kolmogorov-axioms.md.

References: Dubins–Savage, How to Gamble If You Must; Bingham, "Finite additivity versus countable additivity" (2010); Schervish–Seidenfeld–Kadane on non-conglomerability.

1. The Definition Weakened

Definition. A finitely additive probability (f.a.p.) on an algebra $\mathcal{A}$ of subsets of $\Omega$ is $P:\mathcal{A}\to [0,1]$ with $P(\Omega )=1$ and $P(A\bigsqcup B)=P(A)+P(B)$ for disjoint $A,B$.

Note that the domain need only be an algebra, not a $\sigma$-algebra: closure under finite operations suffices, since no countable operation is ever performed. This is the first structural difference and it is not small — the whole apparatus of generated $\sigma$-algebras, $\pi$–$\lambda$ arguments, and Carathéodory extension (measure-theory.md) is built for countable additivity and has no finitely additive analogue of comparable power.

Equivalence (given finite additivity). Countable additivity $⟺$ continuity along monotone sequences: $A_k\downarrow \varnothing ⟹P(A_k)\to 0$.

So the axiom can be stated as a continuity requirement, which is the form in which its cost is clearest: it forbids probability mass from "leaking out" along a vanishing sequence of events.

2. What Becomes Possible

Theorem (Banach). There exist finitely additive, translation-invariant extensions of Lebesgue measure to all subsets of $\mathbb{R}$ (and of ${\mathbb{R}}^2$).

This is exactly the escape route noted in non-measurable.md §1: the Vitali argument kills countably additive total measures, but not finitely additive ones. In dimensions 1 and 2 the isometry group is amenable and an invariant mean exists; from dimension 3 the Banach–Tarski construction rules even this out.

Uniform distributions on $\mathbb{N}$. There is a f.a.p. on $\mathcal{P}(\mathbb{N})$ assigning every finite set probability $0$ and giving the even numbers probability $1/2$ — e.g. a Banach limit of the natural-density functionals, or any measure induced by a nonprincipal ultrafilter refined appropriately. No countably additive measure can do this (kolmogorov-axioms.md §3), since countable additivity forces $P(\mathbb{N})={\sum }_{k}P(\{k\})=0$.

Similarly there are f.a.p.s that are "uniform on $\mathbb{R}$", and improper priors of the kind used informally in Bayesian statistics acquire a legitimate finitely additive reading. Everything on the "impossible" list generated by $\sigma$-additivity becomes available.

The existence proofs, however, all run through the Hahn–Banach theorem or ultrafilters (normed-banach.md), so these objects are as unexhibitable as the Vitali set — the same non-constructivity, relocated.

3. What Collapses

(a) The convergence theorems. Monotone convergence is countable additivity transposed to functions (lebesgue-integral.md §4). Without it there is no dominated convergence, hence no $L^p$ completeness, no Radon–Nikodym theorem, and therefore no conditional expectation in the sense of this folder.

(b) The limit theorems. The strong law of large numbers is a statement about an event ($\{{\bar{X}}_n\to \mu \}$) that is a countable intersection of countable unions; in a merely finitely additive setting it is typically not in the domain of $P$, and where it is, the Borel–Cantelli argument that proves it (independence.md §5) is unavailable. Part B of this folder does not survive.

(c) Conglomerability. This is the sharpest failure and the most interesting.

Definition. $P$ is conglomerable in a partition $\{H_i\}$ if $P(A\mid H_i)\in [a,b]$ for all $i$ implies $P(A)\in [a,b]$.

Countably additive probabilities are conglomerable in countable partitions — it is the law of total probability, $P(A)={\sum }_{i}P(A\mid H_i)P(H_i)$, i.e. the tower property (conditional-expectation.md §3). Finitely additive ones need not be:

Example (de Finetti). Let $P$ be uniform-in-the-finitely-additive-sense on $\mathbb{N}\times \mathbb{N}$, and let $A=\{(m,n):m\le n\}$. Partitioning by columns gives $P(A\mid \text{column }n)=$ a value tending to $0$; partitioning by rows gives values tending to $1$. Both partitions are countable, both conditional assessments are unambiguous, and they are inconsistent with any single value for $P(A)$.

Non-conglomerability means the total-probability rule fails: one cannot compute a probability by conditioning on a countable partition and averaging. Every iterated argument in probability — every use of the tower property, every martingale, every recursive computation — depends on that rule.

(d) Uniqueness and structure. Where Carathéodory plus $\pi$–$\lambda$ gives a unique extension of a premeasure, finitely additive extensions are wildly non-unique (Hahn–Banach gives many), and there is no canonical choice. Models cease to be pinned down by their specifications.

4. The Ledger

The trade is stark: finite additivity buys a larger domain and the "uniform on an infinite set" idealizations, and pays with essentially the entire analytic content of the subject. That is why mainstream probability takes $\sigma$-additivity, and why the choice is nonetheless a choice.

5. Where the Finitely Additive Theory Survives

It is not a dead end. Dubins and Savage developed a complete theory of finitely additive gambling in which optimal strategies exist without measurability hypotheses; the theory of charges (Bhaskara Rao) is a systematic development; Banach limits and invariant means are standard tools in functional analysis and amenability. And two later frameworks recover some of the wanted idealizations without abandoning countable additivity:

• Non-standard probability (nonstandard-probability.md): a hyperfinite uniform distribution on $\{1,\dots ,N\}$ for infinite $N$ assigns each point the infinitesimal $1/N$, and the associated Loeb measure is genuinely countably additive. "Uniform on an infinite set" is thereby obtained inside a $\sigma$-additive theory.
• Imprecise probability: sets of measures, or lower/upper previsions, weaken additivity in a different direction while keeping the analytic machinery.

6. Where This Is Used

• Kolmogorov's axiomatization §5 — the choice this page audits.
• Non-measurable sets §1 — the trade-off table whose first row is developed here.
• Conditional expectation §3 — the tower property that non-conglomerability destroys.
• Non-standard probability — an alternative route to the same idealizations.