Skip to content
Sigma-AlgebrasLebesgue IntegralL^p SpacesRadon-NikodymFubini TheoremProbabilityFoundationsdefinesextendsnormsintegratesunderpinsArt. 14: σ-algebra definitionKey: L^p spaces are Banachvlymbooq
Measure Theory Foundations for Competitive Exam Revision Diagram — Shows how σ‑algebras, Lebesgue integration, Lᵖ norms, Radon‑Nikodym, Fubini, and probability intertwine. · Open the full diagram →
🔬 Scientist Pure mathematics — no exam syllabus

Measure Theory — The Mathematics of Size

Measure theory answers the question: how do we measure the "size" of a set? It starts simply — length of an interval, area of a rectangle — but quickly leads to deep questions about infinity, the structure of the real line, and the foundations of integration. Without measure theory, there is no rigorous probability theory, no Lebesgue integral, no modern analysis. It's the invisible scaffolding beneath the mathematics of uncertainty and change.

"Measure theory is the soul of modern analysis. It tells us what we can integrate, what we can average, and what we can know."
— Henri Lebesgue

📏 Sigma-Algebras and Measures — The Foundation

Can Every Set Have a Length? The Banach-Tarski Paradox
In 1902, Lebesgue asked: can we assign a "length" to every subset of the real line that satisfies basic properties — length of an interval is its length, length is translation-invariant, and length is countably additive (the length of a disjoint union equals the sum of lengths)? The answer is no. There exist non-measurable sets — sets so pathological that no consistent notion of length can include them. The Vitali set is the classic example: pick one element from each coset of ℚ in ℝ. It has the bizarre property that countably many disjoint translations of it cover [0,1], so it can't have any consistent length. Even more dramatically, the Banach-Tarski paradox (1924) shows you can decompose a sphere into finitely many pieces, rotate them, and reassemble them into two identical copies of the original sphere — but the pieces are non-measurable. This is why we need σ-algebras: families of "well-behaved" sets that we can measure.
Not every set can be measured. Measure theory tells us which sets are well-behaved enough to have a size — and which are too pathological.

Sigma-Algebras

A σ-algebra (or σ-field) F on a set X is a collection of subsets containing X, closed under complements and countable unions. The pair (X, F) is a measurable space. The Borel σ-algebra on ℝ is the smallest σ-algebra containing all open intervals — it's generated by (a,b), (−∞,a], or any family that generates the open sets. Borel sets include intervals, open sets, closed sets, Gδ sets, Fσ sets, and more — but still not all subsets. The Lebesgue σ-algebra extends the Borel σ-algebra to include all sets that differ from a Borel set by a set of measure zero (the completion).

σ-Algebra Axioms
∅∈F, closed under complements, countable unions
Measure
μ(∅)=0, countable additivity
Borel σ-Algebra
Smallest σ-algebra with all open sets
Probability Measure
μ(Ω) = 1

Constructing the Lebesgue Measure

The Lebesgue measure λ on ℝ is built in stages. Start with the outer measure λ*(A) = inf{ Σ |Iₙ| : A ⊆ ∪ Iₙ } where Iₙ are intervals. A set A is Lebesgue measurable if it splits every set appropriately (Carathéodory's criterion): λ*(E) = λ*(E∩A) + λ*(E\A) for all E ⊆ ℝ. This gives a σ-algebra containing the Borel sets, and λ is countably additive on it. The Lebesgue measure of an interval is its length, it's translation-invariant, and it satisfies all the intuitive properties of "length" — but only on measurable sets.

✨
The Cantor Set — Measure Zero, Yet Uncountable ▶
The Cantor set is constructed by repeatedly removing the middle third of intervals. Start with [0,1]. Remove (1/3, 2/3). Then remove the middle third of each remaining piece, and so on.

The total length removed = 1/3 + 2/9 + 4/27 + ... = 1. So the Cantor set has Lebesgue measure 0. Yet it's uncountable — it has as many points as the entire interval [0,1]!

The Cantor set is a perfect, nowhere-dense set of measure zero. It shows that "size" in the sense of measure is very different from "size" in the sense of cardinality. It also appears in dynamics, fractal geometry, and as the attractor of iterated function systems.

∫ The Lebesgue Integral — Integrating Without Continuity

Why Riemann's Integral Isn't Enough — and What Lebesgue Did About It
The Riemann integral approximates area by slicing vertically — partitioning the x-axis and summing rectangles. This works for continuous functions and many discontinuous ones, but it fails for functions with "too many" discontinuities. The indicator function of the rationals (1 on ℚ, 0 elsewhere) is nowhere continuous, and its Riemann integral doesn't exist. Lebesgue's insight: slice horizontally instead. Instead of partitioning the domain, partition the range and measure the set where the function takes values in each range interval. For the indicator of ℚ, the set where it's 1 is ℚ, which has measure 0 — so the integral is 0. The Lebesgue integral exists for a much wider class of functions — all measurable functions — and satisfies powerful convergence theorems that the Riemann integral lacks.
The Riemann integral asks "what's the value at each x?" The Lebesgue integral asks "how large is the set where the function takes each value?" The second question turns out to be far more powerful.

Measurable Functions and Simple Functions

A function f: X → ℝ is measurable if the preimage of every Borel set is measurable. This is equivalent to {x : f(x) > c} being measurable for all c ∈ ℝ. Almost every function you can reasonably define is measurable — continuous functions, step functions, limits of measurable functions. A simple function is a finite linear combination of indicator functions of measurable sets. The integral of a simple function is Σ aᵢ·μ(Aᵢ). For a non-negative measurable function f, the integral is the supremum of integrals of simple functions ≤ f. For general f, split into positive and negative parts.

The Three Convergence Theorems
Monotone Convergence: fₙ ↗ f ⇒ ∫ fₙ → ∫ f
Dominated Convergence: |fₙ| ≤ g, fₙ → f a.e. ⇒ ∫ fₙ → ∫ f
✨
The Dominated Convergence Theorem in Action ▶
Let fₙ(x) = n·xⁿ for x ∈ [0,1]. Pointwise, fₙ(x) → 0 for x ∈ [0,1), and fₙ(1) = n → ∞. So the pointwise limit is 0 almost everywhere (except at x=1, a set of measure 0).

Can we swap limit and integral? ∫₀¹ n·xⁿ dx = n/(n+1) → 1. The pointwise limit is 0, so ∫ fₙ → 1 ≠ 0 = ∫ 0!

The dominated convergence theorem doesn't apply because there's no integrable dominating function g with |fₙ| ≤ g for all n. Indeed, supₙ fₙ(1) = ∞. This shows why the conditions matter — when they're satisfied, swapping limit and integral is justified; when they're not, all bets are off.

📊 Lᵖ Spaces — Function Spaces with Norms

Lᵖ Spaces — Measuring the Size of a Function
Once we have the Lebesgue integral, we can define norms on function spaces. For p ≥ 1, the Lᵖ norm is ‖f‖ₚ = (∫ |f|ᵖ dμ)^{1/p}. The space Lᵖ consists of all measurable functions with finite Lᵖ norm, identifying functions that agree almost everywhere. L² is a Hilbert space (inner product ⟨f,g⟩ = ∫ f ḡ dμ) and is the natural setting for Fourier series and quantum mechanics. L¹ is the space of integrable functions. L^∞ consists of essentially bounded functions. The Hölder inequality says ∫ |fg| ≤ ‖f‖ₚ·‖g‖₉ for 1/p + 1/q = 1, generalizing Cauchy-Schwarz (p=q=2). Riesz-Fischer theorem: Lᵖ spaces are complete (every Cauchy sequence converges) — they're Banach spaces. This completeness is what makes modern analysis work: we can solve equations by finding approximate solutions and taking limits.
Lᵖ spaces give us a way to measure functions like we measure vectors — with norms, convergence, and completeness.
Lᵖ Norm
‖f‖ₚ = (∫ |f|ᵖ)^{1/p}
Hölder
∫ |fg| ≤ ‖f‖ₚ·‖g‖₉
Minkowski
‖f+g‖ₚ ≤ ‖f‖ₚ + ‖g‖ₚ
Riesz-Fischer
Lᵖ is complete (Banach)

Modes of Convergence

In measure theory, there are several distinct notions of convergence for functions. Pointwise a.e. convergence: fₙ(x) → f(x) for almost every x. Convergence in measure: for every ε > 0, μ({|fₙ − f| > ε}) → 0. Convergence in Lᵖ: ‖fₙ − f‖ₚ → 0. These are related: Lᵖ convergence implies convergence in measure; convergence in measure implies a subsequence converges pointwise a.e.; but in general, none of these implications reverse. The Egoroff theorem says that on a finite measure space, pointwise a.e. convergence implies uniform convergence on sets of arbitrarily large measure — almost uniform convergence.

Convergence that Isn't What It Seems
Consider fₙ = n·χ_{[0,1/n]}. Pointwise, fₙ(x) → 0 for all x > 0, but fₙ(0) = n → ∞. Pointwise a.e., fₙ → 0. In measure: for any ε > 0, μ({|fₙ| > ε}) = 1/n → 0. But ∫ fₙ = n·(1/n) = 1, so ‖fₙ‖₁ = 1 — they do NOT converge in L¹ to 0. This function "escapes to infinity" in the L¹ norm even though it converges to 0 pointwise.

🌌 Radon-Nikodym, Fubini, and the Foundations of Probability

The Radon-Nikodym Theorem — When Is One Measure "Denser" Than Another?
Given two measures μ and ν on the same measurable space, when can we write ν(A) = ∫_A f dμ for some measurable function f? The Radon-Nikodym theorem says: if ν is absolutely continuous with respect to μ (ν(A) = 0 whenever μ(A) = 0), then there exists a Radon-Nikodym derivative f = dν/dμ such that ν(A) = ∫_A f dμ. This is the measure-theoretic version of the chain rule. In probability, this gives conditional expectation: E[X|G] is the Radon-Nikodym derivative of a measure with respect to another. The Lebesgue decomposition theorem says any measure can be uniquely decomposed into an absolutely continuous part, a singular continuous part, and a discrete part — the mathematical structure behind the distinction between continuous and discrete probability distributions.
The Radon-Nikodym theorem tells us when one measure can be described as a "density" with respect to another — it's why probability density functions exist.

Product Measures and Fubini's Theorem

Given two measure spaces (X, F, μ) and (Y, G, ν), the product measure μ×ν on X×Y is defined by (μ×ν)(A×B) = μ(A)·ν(B) on rectangles, then extended to the product σ-algebra. Fubini's theorem says that if f is integrable with respect to the product measure, then the iterated integrals exist and are equal: ∫_{X×Y} f d(μ×ν) = ∫_X (∫_Y f(x,y) dν(y)) dμ(x) = ∫_Y (∫_X f(x,y) dμ(x)) dν(y). This is the justification for "changing the order of integration" in multivariable calculus — but it requires integrability, which is not automatic. The Tonelli theorem gives a useful sufficient condition: if f ≥ 0, the iterated integrals are always equal (even if infinite).

Fubini's Theorem
For integrable f, the order of integration doesn't matter: ∬ f = ∫∫ f dx dy = ∫∫ f dy dx
But this is NOT always true for non-integrable functions! Fubini's theorem requires absolute integrability with respect to the product measure.

Probability Theory as Measure Theory

Modern probability theory is entirely built on measure theory. A probability space (Ω, F, P) is just a measure space with P(Ω) = 1. A random variable is a measurable function X: Ω → ℝ. Its expected value is E[X] = ∫_Ω X dP. The law of X is the pushforward measure P_X(A) = P(X ∈ A). Conditional expectation E[X|G] for a sub-σ-algebra G ⊆ F is defined via the Radon-Nikodym theorem. The Kolmogorov extension theorem constructs infinite-dimensional probability spaces (like the space of all Brownian motion paths). The strong law of large numbers and central limit theorem are theorems about convergence of integrals with respect to product measures. Measure theory provides the mathematical rigor that makes probability a branch of mathematics rather than a collection of heuristics.

✨
The Vitali Set — A Non-Measurable Set ▶
Define an equivalence relation on [0,1]: x ~ y if x − y ∈ ℚ. This partitions [0,1] into uncountably many equivalence classes, each a coset of ℚ. Using the axiom of choice, pick one element from each equivalence class to form the Vitali set V.

Key properties: for any rational q ∈ [0,1], the translates V_q = V + q (mod 1) are disjoint and their union over all q ∈ ℚ∩[0,1] is exactly [0,1].

If V were Lebesgue measurable with measure m, then each V_q would also have measure m (translation invariance). Countable additivity would give: 1 = λ([0,1]) = Σ_m λ(V_q) = Σ_m m. If m = 0, the sum is 0. If m > 0, the sum is infinite. Both contradiction. So V is not measurable.

The existence of non-measurable sets requires the axiom of choice. Without it, it's consistent that every set of reals is Lebesgue measurable (Solovay, 1970).
"Probability theory is measure theory with a soul — the soul being the interpretation of measure as degree of belief, and measurable functions as random variables."
— Mark Kac

🎯 Practice — Check Your Measure Theory Intuition

These puzzles build measure theory intuition.

← Back to Math Map

Completion of Measure Spaces and the Lebesgue σ‑algebra

The Carathéodory extension theorem constructs a measure on a σ‑algebra from a pre‑measure defined on an algebra. Starting with the algebra of finite unions of half‑open intervals in ℝ, the pre‑measure λ₀((a,b]) = b−a extends uniquely to a measure on the σ‑algebra generated by that algebra — the Borel σ‑algebra ℬ(ℝ). However, the resulting measure space (ℝ, ℬ, λ) is not complete: there exist subsets of Borel null sets that are not Borel measurable. The completion adds every subset of a λ‑null Borel set to the σ‑algebra and assigns it measure zero. The completed σ‑algebra is the Lebesgue σ‑algebra ℒ, and the completed measure is the Lebesgue measure λ̄. Every Lebesgue measurable set E can be written as E = B ∪ N where B ∈ ℬ and N ⊂ Z for some Borel null set Z. This decomposition is unique up to null sets and shows that ℒ is strictly larger than ℬ (|ℒ| = 2^𝔠 while |ℬ| = 𝔠). Completeness is essential for the dominated convergence theorem and for Fubini’s theorem, because the product of complete measure spaces is again complete after taking the completion of the product σ‑algebra.

Key properties

  • If μ is a measure on (X, Σ), the completion Σ̄ consists of all sets A ⊂ X for which there exist B, C ∈ Σ with B ⊂ A ⊂ C and μ(C B) = 0.
  • The extended measure μ̄(A) = μ(B) = μ(C) is well defined and makes (X, Σ̄, μ̄) a complete measure space.
  • For Lebesgue measure, every set of outer measure zero is Lebesgue measurable; this characterises ℒ as the σ‑algebra of Carathéodory measurable sets for the outer measure λ*.

Signed Measures, Hahn and Jordan Decompositions

A signed measure ν on a measurable space (X, Σ) is a countably additive set function taking values in [−∞, ∞] that never assumes both +∞ and −∞. The Hahn decomposition theorem states that there exist disjoint measurable sets P, N with P ∪ N = X such that ν(E) ≥ 0 for all E ⊂ P and ν(E) ≤ 0 for all E ⊂ N. The pair (P, N) is unique up to ν‑null sets. From a Hahn decomposition one obtains the Jordan decomposition ν = ν⁺ − ν⁻ where ν⁺(E) = ν(E ∩ P) and ν⁻(E) = −ν(E ∩ N) are mutually singular positive measures. The total variation |ν| = ν⁺ + ν⁻ is a positive measure satisfying |ν(E)| ≤ |ν|(E) for all E. The space of finite signed measures equipped with the total variation norm is a Banach space. The Radon‑Nikodym theorem extends to signed measures: if ν ≪ μ then there exists an integrable function f = dν/dμ with ν(E) = ∫_E f dμ, and |ν|(E) = ∫_E |f| dμ.

Applications

  • Decomposition of a complex measure into real and imaginary parts, each a signed measure.
  • Representation of bounded linear functionals on C₀(X) via the Riesz‑Markov theorem, yielding regular signed Borel measures.
  • In probability, the difference of two probability measures is a signed measure of total mass zero; its Jordan decomposition gives the positive and negative parts used in total variation distance.
Worked Example: Hahn Decomposition of a Simple Signed Measure
Let ν be the signed measure on (ℝ, ℬ) defined by ν(E) = ∫_E (1[0,1] − 2·1[2,3]) dλ. Find a Hahn decomposition (P, N) and the Jordan decomposition ν⁺, ν⁻.
The density f = 1[0,1] − 2·1[2,3] is positive on [0,1] and negative on [2,3]; it is zero elsewhere. Take P = [0,1] and N = [2,3] ∪ (ℝ ([0,1] ∪ [2,3])). Then for any Borel E ⊂ P, ν(E) = λ(E) ≥ 0; for any E ⊂ N, ν(E) = −2λ(E ∩ [2,3]) ≤ 0. Thus (P, N) is a Hahn decomposition. The Jordan positive part is ν⁺(E) = ν(E ∩ P) = λ(E ∩ [0,1]), and the negative part is ν⁻(E) = −ν(E ∩ N) = 2λ(E ∩ [2,3]). The total variation is |ν|(E) = λ(E ∩ [0,1]) + 2λ(E ∩ [2,3]).
Exam‑Day Strategy
When a question asks for a Hahn or Jordan decomposition, first identify the sign of the density (or the sign of the measure on generating sets). Write the positive and negative sets explicitly; uniqueness up to null sets means any choice that respects the sign works. For total variation, remember |ν| = ν⁺ + ν⁻ and that ‖ν‖ = |ν|(X).

Differentiation of Measures and the Lebesgue Differentiation Theorem

Given a locally finite Borel measure μ on ℝⁿ, the derivative of μ with respect to Lebesgue measure λ at a point x is defined as the limit (if it exists) of μ(B(x,r))/λ(B(x,r)) as r → 0, where B(x,r) is the open ball. The Lebesgue differentiation theorem states that for any f ∈ L¹_loc(ℝⁿ), the limit of the averages (1/λ(B(x,r))) ∫B(x,r) f(y) dy equals f(x) for λ‑almost every x. Equivalently, if μ ≪ λ with density f = dμ/dλ, then the Radon‑Nikodym derivative is recovered almost everywhere by the limit of ratios of measures of shrinking balls. The proof uses the Hardy‑Littlewood maximal function Mf(x) = supr>0 (1/λ(B(x,r))) ∫B(x,r) |f(y)| dy and the weak‑type (1,1) inequality λ({Mf > α}) ≤ Cₙ α⁻¹ ‖f‖₁. This maximal inequality also yields the Lebesgue differentiation theorem for locally integrable functions and the Besicovitch covering theorem for general Radon measures. A corollary is that for any locally integrable f, the set of Lebesgue points — points where the average of |f(y)−f(x)| over B(x,r) tends to zero — has full measure.

Consequences

  • If μ is a finite Borel measure on ℝⁿ, then for λ‑a.e. x the limit limr→0 μ(B(x,r))/λ(B(x,r)) exists and equals the Radon‑Nikodym derivative dμ/dλ(x).
  • The theorem fails for arbitrary sequences of sets shrinking to x; the balls (or cubes) must satisfy a regularity condition such as bounded eccentricity.
  • In probability, the theorem justifies the interpretation of a density as a limit of relative frequencies in shrinking neighbourhoods.

Conditional Expectation as Orthogonal Projection in L²

Let (Ω, ℱ, P) be a probability space and 𝒢 ⊂ ℱ a sub‑σ‑algebra. For X ∈ L²(Ω, ℱ, P) the conditional expectation E[X | 𝒢] is the unique 𝒢‑measurable random variable Y ∈ L²(Ω, 𝒢, P) satisfying ∫_G Y dP = ∫_G X dP for all G ∈ 𝒢. The subspace L²(Ω, 𝒢, P) is closed in L²(Ω, ℱ, P); therefore the orthogonal projection Π_𝒢 onto this subspace exists and coincides with the conditional expectation operator. This geometric view yields immediate properties: ‖E[X | 𝒢]‖₂ ≤ ‖X‖₂ (contraction), the tower property E[E[X | ℋ] | 𝒢] = E[X | 𝒢] for 𝒢 ⊂ ℋ, and the fact that X − E[X | 𝒢] is orthogonal to every 𝒢‑measurable L² variable. For general X ∈ L¹, conditional expectation is defined by approximation with L² variables or via the Radon‑Nikodym theorem applied to the signed measure ν(G) = ∫_G X dP. The projection interpretation also explains why conditional expectation is the best predictor in mean‑square sense: among all 𝒢‑measurable random variables Y, E[(X−Y)²] is minimised uniquely by Y = E[X | 𝒢].

Key identities

  • Linearity: E[aX + bY | 𝒢] = aE[X | 𝒢] + bE[Y | 𝒢].
  • Taking out what is known: if Z is 𝒢‑measurable and bounded, E[ZX | 𝒢] = Z E[X | 𝒢].
  • Independence: if X is independent of 𝒢, then E[X | 𝒢] = E[X] a.s.

Martingale Convergence Theorems

A filtration (ℱₙ)n≥0 is an increasing sequence of sub‑σ‑algebras of ℱ. A sequence (Mₙ) of integrable random variables adapted to (ℱₙ) is a martingale if E[Mn+1 | ℱₙ] = Mₙ a.s. for all n; it is a sub‑martingale (super‑martingale) if the conditional expectation is ≥ (≤) Mₙ. Doob’s martingale convergence theorem states that if (Mₙ) is an L¹‑bounded martingale (i.e. supₙ E[|Mₙ|] < ∞), then Mₙ converges almost surely and in L¹ to an integrable limit M_∞. For non‑negative super‑martingales the L¹‑boundedness is automatic, so they converge a.s. to a finite limit. The upcrossing inequality provides a quantitative proof: for any a < b, the expected number of upcrossings of the interval [a,b] by (Mₙ) is bounded by (E[(M₀−a)⁻])/(b−a). This yields a.s. convergence for any super‑martingale bounded in L¹. If the martingale is bounded in L^p for some p > 1, then convergence also holds in L^p and the limit satisfies Mₙ = E[M_∞ | ℱₙ] for all n (the martingale is closed). The theorem extends to backward martingales and to martingales indexed by continuous time under right‑continuity of the filtration. Applications include the strong law of large numbers (via the martingale Mₙ = Sₙ/n), the convergence of likelihood ratios in statistics, and the construction of stochastic integrals.

Important corollaries

  • Lévy’s upward theorem: if ℱₙ ↑ ℱ_∞ and X ∈ L¹, then E[X | ℱₙ] → E[X | ℱ_∞] a.s. and in L¹.
  • Lévy’s downward theorem: if ℱₙ ↓ ℱ_∞, then E[X | ℱₙ] → E[X | ℱ_∞] a.s. and in L¹.
  • Kolmogorov’s 0‑1 law follows from the martingale convergence theorem applied to the tail σ‑algebra.
  • Confusing the completion of a measure space with the Carathéodory extension; the former adds subsets of null sets, the latter builds a measure from a pre‑measure.
  • Assuming a signed measure can take both +∞ and −∞ on different sets; by definition it cannot.
  • Forgetting that the Hahn decomposition is unique only up to ν‑null sets, not pointwise.
  • Applying the Lebesgue differentiation theorem to arbitrary shrinking families without a regularity condition (e.g., rectangles with unbounded eccentricity).
  • Treating conditional expectation as a pointwise operation rather than an equivalence class in L¹; two versions differ on a null set.
  • Using the martingale convergence theorem without verifying L¹‑boundedness (or non‑negativity for super‑martingales).
  • Misidentifying the orthogonal projection in L² with the conditional expectation for non‑square‑integrable variables; the projection argument only works directly in L².