Sigma-Algebras, Measures, and Measurable Functions

One line: A σ-algebra is a collection of subsets closed under countable set operations; a measure assigns consistent "sizes" to those sets; a measurable function pulls back measurable sets to measurable sets. Key result: The triple

(X, A, μ)

is the foundational object of modern integration, probability, and analysis — Lebesgue's framework that tames "size" and "average" beyond Riemann's reach. Where it applies: Probability theory, functional analysis, ergodic theory, PDE, mathematical physics, statistics. Who/when: Émile Borel (1898), Henri Lebesgue (1902), Constantin Carathéodory (1918), Andrey Kolmogorov (1933).

A measure space is a triple $(X, A, μ)$ where $X$ is a set, $A \subseteq 2^{X}$ is a σ-algebra (closed under complement and countable union), and $μ : A \to [0, \infty]$ is a countably additive set function with $μ (\emptyset) = 0$ . This structure replaces the naive notion that "every set has a length/area/probability" with a rigorously consistent one, sidestepping the paradoxes (Vitali, Banach–Tarski) that arise from trying to measure all subsets. A measurable function is the natural morphism between measurable spaces, the right notion of "well-behaved map" for integration. The surprising fact: one cannot define a translation-invariant measure on all subsets of $R$ assigning intervals their length — the axiom of choice forces non-measurable sets to exist.

The Core Insight

"Size" cannot be defined for every subset without contradiction, so we restrict the domain of the size-function to a structured family of sets (the σ-algebra) closed precisely under the operations — complement and countable union — that calculus and probability require. Countability is the magic ingredient: it is strong enough to permit limits, suprema, and series of functions, yet weak enough to avoid the pathologies of uncountable operations. Everything downstream — integration, expectation, convergence theorems — is built on this single compromise.

Formal Statement

σ-Algebra

Let $X$ be a nonempty set. A collection $A \subseteq 2^{X}$ is a σ-algebra on $X$ if:

X \in A

⟳the whole space is always measurable; combined with closure under complement this gives

\emptyset \in A

A \in A ⟹ A^{c} = X ∖ A \in A

⟳closure under complementation — measurability of

A

and "not

A

" must coincide.

A_{1}, A_{2}, \dots \in A ⟹ n = 1 ⋃ \infty A_{n} \in A

⟳countable union; via De Morgan,

⋂_{n} A_{n} = (⋃_{n} A_{n}^{c})^{c} \in A

, so countable intersections are also included.

Elements of $A$ are measurable sets, and $(X, A)$ is a measurable space. A collection satisfying [1]–[2] and closure under finite unions is merely an algebra (field).

Generated σ-Algebra and Borel sets

For any family $E \subseteq 2^{X}$ , the σ-algebra generated by $E$ , written $σ (E)$ , is the smallest σ-algebra containing $E$ :

σ (E) = ⋂ {A : A is a σ -algebra, E \subseteq A}

⟳an arbitrary intersection of σ-algebras is a σ-algebra (verify each axiom termwise), and

2^{X}

contains

E

, so the family is nonempty.

When $X$ is a topological space with topology $τ$ , the Borel σ-algebra is $B (X) = σ (τ)$ , generated by the open sets.

Measure

Given a measurable space $(X, A)$ , a function $μ : A \to [0, \infty]$ is a measure if $μ (\emptyset) = 0$ and it is countably additive (σ-additive): for pairwise disjoint ${A_{n}}_{n \geq 1} \subseteq A$ ,

μ (n = 1 ⨆ \infty A_{n}) = n = 1 \sum \infty μ (A_{n})

⟳this is the defining axiom; the RHS series of nonnegative terms converges in

[0, \infty]

unconditionally, so order is irrelevant.

The triple $(X, A, μ)$ is a measure space. It is finite if $μ (X) < \infty$ , a probability space if $μ (X) = 1$ , and σ-finite if $X = ⋃_{n} X_{n}$ with $μ (X_{n}) < \infty$ .

Elementary Properties of Measures

Let $(X, A, μ)$ be a measure space and $A, B, A_{n} \in A$ . Then:

A \subseteq B ⟹ μ (A) \leq μ (B) (monotonicity)

⟳write

B = A ⊔ (B ∖ A)

and apply finite additivity;

μ (B ∖ A) \geq 0

μ (n = 1 ⋃ \infty A_{n}) \leq n = 1 \sum \infty μ (A_{n}) (countable subadditivity)

⟳disjointify via

B_{1} = A_{1}, B_{n} = A_{n} ∖ ⋃_{k < n} A_{k}

; then

⋃ A_{n} = ⨆ B_{n}

B_{n} \subseteq A_{n}

, apply [5] and [6].

A_{n} ↑ A ⟹ μ (A_{n}) \to μ (A) (continuity from below)

⟳set

C_{1} = A_{1}, C_{n} = A_{n} ∖ A_{n - 1}

; then

A = ⨆ C_{n}

and

μ (A_{n}) = \sum_{k \leq n} μ (C_{k}) \to \sum_{k} μ (C_{k}) = μ (A)

A_{n} ↓ A, μ (A_{1}) < \infty ⟹ μ (A_{n}) \to μ (A) (continuity from above)

⟳apply [8] to

A_{1} ∖ A_{n} ↑ A_{1} ∖ A

and subtract; the finiteness of

μ (A_{1})

is essential to justify subtraction.

The finiteness hypothesis $μ (A_{1}) < \infty$ in [9] cannot be dropped. With Lebesgue measure on $R$ , take $A_{n} = [n, \infty)$ . Then $A_{n} ↓ \emptyset$ but $μ (A_{n}) = \infty$ for all $n$ , so $μ (A_{n}) \to \infty \neq = 0 = μ (\emptyset)$ . Continuity from above genuinely fails for infinite measures.

Measurable Function

Let $(X, A)$ and $(Y, B)$ be measurable spaces. A map $f : X \to Y$ is $(A, B)$ -measurable if

f^{- 1} (B) \in A for every B \in B

⟳preimages commute with all set operations (

f^{- 1} (B^{c}) = (f^{- 1} B)^{c}

f^{- 1} (⋃ B_{n}) = ⋃ f^{- 1} (B_{n})

), so it suffices to check [10] on a generating family

E

B

For $Y = \overline{R} = [- \infty, \infty]$ with the Borel σ-algebra, $f$ is measurable iff for every $a \in R$ ,

{x : f (x) > a} = f^{- 1} ((a, \infty]) \in A

⟳the rays

(a, \infty]

generate

B (\overline{R})

; combined with the preimage-commutation principle this single family suffices.

Why It Holds / Derivation

:::proof[Measurability is checked on generators — proof of the reduction in [10]–[11]] Claim. If $B = σ (E)$ and $f^{- 1} (E) \in A$ for all $E \in E$ , then $f$ is $(A, B)$ -measurable.

Step 1. Define the good sets

G = {B \subseteq Y : f^{- 1} (B) \in A}

. Step 2. Show

G

is a σ-algebra on

Y

. Indeed

f^{- 1} (Y) = X \in A

gives

Y \in G

. If

B \in G

then

f^{- 1} (B^{c}) = (f^{- 1} B)^{c} \in A

since

A

is closed under complement, so

B^{c} \in G

. If

B_{n} \in G

then

f^{- 1} (⋃_{n} B_{n}) = ⋃_{n} f^{- 1} (B_{n}) \in A

by closure under countable unions, so

⋃_{n} B_{n} \in G

. Step 3. By hypothesis

E \subseteq G

. Since

G

is a σ-algebra containing

E

, minimality of

σ (E)

gives

σ (E) = B \subseteq G

. Step 4. Thus

f^{- 1} (B) \in A

for every

B \in B

, which is exactly [10].

■

This "good-sets principle" (a special case of the $π$ - $λ$ / monotone-class machinery) is the workhorse of the entire theory: to prove a measurable-set or measurable-function statement, one verifies it on a convenient generating family and shows the class of validity is a σ-algebra (or λ-system).

Stability of Measurable Functions

Let $f, g : X \to \overline{R}$ be measurable and $(f_{n})$ a sequence of measurable functions. Then $f + g$ , $f g$ , $c f$ , $∣ f ∣$ , $max (f, g)$ , $min (f, g)$ are measurable, and so are

n sup f_{n}, n in f f_{n}, n lim sup f_{n}, n lim inf f_{n} .

⟳

{sup_{n} f_{n} > a} = ⋃_{n} {f_{n} > a} \in A

— a countable union, hence measurable; the others follow since

in f f_{n} = - sup (- f_{n})

and

lim sup f_{n} = in f_{m} sup_{n \geq m} f_{n}

. Arithmetic operations use measurability of continuous maps

R^{2} \to R

composed with the measurable

(f, g)

Think of a σ-algebra as the list of "questions you are allowed to ask" about a point $x \in X$ — each measurable set $A$ encodes the yes/no question "is $x \in A$ ?". The axioms say: you may ask the trivial question ( $X$ ), you may negate any question ( $A^{c}$ ), and you may combine countably many questions with "or" ( $⋃ A_{n}$ ). A measure assigns each allowable question a "weight of yes-answers." A measurable function is a quantity whose level sets are askable questions — you can always ask "is $f (x)$ above the threshold $a$ ?". The countable cap is what lets you take limits: a limit involves infinitely many comparisons, but only countably many, so the answer stays askable. Uncountable operations would let you single out individual points and reconstruct every subset, collapsing the whole structure.

Containment hierarchy: from semiring to measure space, and the action of a measurable function

Worked Examples

Lebesgue–Stieltjes measure of an interval

Given. Let

F (x) = arctan x

R

, a bounded increasing right-continuous function. It induces the Lebesgue–Stieltjes measure

μ_{F}

B (R)

via

μ_{F} ((a, b]) = F (b) - F (a)

. Step 1. Compute

μ_{F}

(0, 1]

μ_{F} ((0, 1]) = arctan 1 - arctan 0 = \frac{π}{4} - 0 = \frac{π}{4} .

Step 2. Total mass: by continuity from below ([8]) with

(- n, n] ↑ R

μ_{F} (R) = n \to \infty lim (arctan n - arctan (- n)) = \frac{π}{2} - (- \frac{π}{2}) = π .

Step 3. Normalize to a probability measure

P = μ_{F} / π

. Then the probability of

(0, 1]

(π /4) / π = 1/4

. This is exactly the standard Cauchy distribution on

(0, 1]

μ_{F} ((0, 1]) = \frac{π}{4}, μ_{F} (R) = π, P ((0, 1]) = \frac{1}{4}

A function and its level sets: measurability of the Dirichlet function

Given. Let

f = 1_{Q}

X = ([0, 1], B ([0, 1]), λ)

(Lebesgue measure

λ

), i.e.

f (x) = 1

x \in Q

, else

0

. Step 1. Check [11]. For

a \geq 1

{f > a} = \emptyset \in B

. For

0 \leq a < 1

{f > a} = Q \cap [0, 1]

, which is countable, hence Borel (a countable union of singletons). For

a < 0

{f > a} = [0, 1] \in B

. All level rays are measurable ⇒

f

is Borel measurable. Step 2. Its Lebesgue integral. Since

λ (Q \cap [0, 1]) = 0

(countable set),

f = 0

λ

-a.e., so

\int_{[0, 1]} f d λ = 1 \cdot λ (Q \cap [0, 1]) + 0 \cdot λ ([0, 1] ∖ Q) = 1 \cdot 0 + 0 \cdot 1 = 0.

Step 3. Contrast:

f

is not Riemann integrable (upper sums

= 1

, lower sums

= 0

). The measure-theoretic framework integrates it effortlessly.

f is Borel measurable, \int_{[0, 1]} f d λ = 0

Common Misconceptions

"Every subset of $R$ has a length." False. Assuming the axiom of choice, there exist non-measurable sets (Vitali sets) — no translation-invariant, countably additive extension of length can be defined on all of

2^{R}

. This is the reason we restrict to a σ-algebra rather than

2^{X}

. "Measurable means continuous-ish." No. Measurable functions can be wildly discontinuous (e.g.

1_{Q}

). Conversely Lusin's theorem says a measurable function agrees with a continuous one off a set of small measure, which is the precise sense in which measurability is "almost continuity." " $σ$ -additivity is the same as finite additivity." It is strictly stronger. Finitely additive set functions exist (e.g. via Banach limits / ultrafilters) that are not countably additive; only σ-additivity yields the convergence theorems.

A finitely additive "measure" that is not a measure

On $X = N$ , take $A = 2^{N}$ and let $U$ be a free ultrafilter. Define $ν (A) = 1$ if $A \in U$ , else $0$ . Then $ν$ is finitely additive and $ν (N) = 1$ , but it is not countably additive: writing $N = ⨆_{n} {n}$ gives $\sum_{n} ν ({n}) = \sum_{n} 0 = 0 \neq = 1 = ν (N)$ , because no singleton belongs to a free ultrafilter. Thus finite additivity does not imply [5]; σ-additivity is a genuine extra axiom.

The Vitali set: non-measurability

Define $x \sim y$ iff $x - y \in Q$ on $[0, 1]$ . Using AC, pick one representative from each class to form $V$ . The translates ${V + q mod 1 : q \in Q \cap [0, 1)}$ are countably many, pairwise disjoint, and cover $[0, 1]$ . If $V$ were Lebesgue measurable with measure $m$ , translation invariance forces $λ ([0, 1]) = \sum_{q} m$ , i.e. $1 = \sum_{q} m$ — impossible: the sum is $0$ if $m = 0$ and $\infty$ if $m > 0$ . Hence $V \in / L$ , the Lebesgue σ-algebra.

Connections

The triple $(X, A, μ)$ is the launchpad for the Lebesgue Integral, where measurable functions become integrands via the "horizontal slicing" of their level sets. With $μ (X) = 1$ it is a probability space, so Probability Spaces and Random Variables are nothing but measure spaces and measurable functions in disguise (a random variable is a measurable map). The construction of measures from premeasures is governed by Carathéodory's Extension Theorem and outer measures. The interplay of two measures on the same space is captured by Radon-Nikodym Theorem (densities) and by Product Measures and Fubini's Theorem (multivariable integration). Topological refinements live in Borel and Lebesgue Sigma-Algebras and Radon Measures.

Real-World Applications

Probability & statistics — Kolmogorov's axioms are a measure space with $μ (X) = 1$ ; every theorem about expectations, conditional distributions, and martingales rests on measurable functions and σ-algebras (filtrations).
Mathematical finance — Stochastic calculus and the Black–Scholes model use filtrations (increasing families of σ-algebras) to encode "information available by time $t$ "; equivalent martingale measures (change of measure) come straight from Radon-Nikodym Theorem.
Signal processing & information theory — Shannon entropy $H = - \sum p_{i} lo g p_{i}$ and differential entropy are integrals against measures; ergodic theory (measure-preserving transformations) underlies data compression bounds.
Quantum mechanics — Projection-valued measures (spectral theorem) assign measurable subsets of $R$ to projection operators; the Born rule is a measure on the spectrum of an observable.
Statistical mechanics — The Gibbs measure on phase space and the microcanonical/canonical ensembles are concrete measures; partition functions are total masses.
Machine learning — Generalization bounds, PAC learning, and generative models (GANs, diffusion) are framed via measures, pushforwards, and the Kullback–Leibler divergence (a Radon–Nikodym quantity).
Dynamical systems / climate models — Invariant (SRB) measures describe long-run statistical behavior of chaotic flows where individual trajectories are unpredictable.

Historical Development

1883 — Georg Cantor and others develop content/Jordan measure, exposing limits of "elementary" length and area.
1898 — Émile Borel introduces the σ-algebra of Borel sets and countably additive length on them, the first systematic countable-additivity framework.
1902 — Henri Lebesgue's thesis Intégrale, longueur, aire defines Lebesgue measure and integral, revolutionizing analysis by integrating far broader classes of functions.
1905 — Giuseppe Vitali constructs the first non-measurable set, proving full $2^{R}$ cannot be measured translation-invariantly.
1909 — Johann Radon unifies Lebesgue and Stieltjes integration; Radon measures on locally compact spaces emerge.
1913–1915 — Maurice Fréchet and others abstract measure to general spaces, divorcing it from $R^{n}$ .
1918 — Constantin Carathéodory formalizes outer measures and the extension theorem, giving the modern construction recipe.
1930 — Otton Nikodym and Johann Radon's theorem on densities (Radon–Nikodym) is completed.
1933 — Andrey Kolmogorov's Grundbegriffe der Wahrscheinlichkeitsrechnung axiomatizes probability as measure theory, cementing the framework's centrality.

Subtleties & Limits

Measurable cardinals and the limits of "everything is measurable." Solovay (1970) showed (assuming an inaccessible cardinal) the consistency of a model of ZF + dependent choice in which every subset of

R

is Lebesgue measurable — so non-measurable sets are precisely an artifact of full AC. The existence of a nontrivial σ-additive ${0, 1}$ -valued measure on all subsets of some set is equivalent to the existence of a measurable cardinal, a large-cardinal axiom unprovable in ZFC. Banach's measure problem (does a countably additive, isometry-invariant measure on

2^{R^{n}}

extending Lebesgue measure exist?) connects to these. Practically: the boundary of "what can be measured" is entangled with the foundations of set theory itself, and how much choice and how large the universe is determine the answer.