The Lebesgue Integral and Convergence Theorems

One line: The Lebesgue integral extends Riemann integration by partitioning the range rather than the domain, yielding a theory in which limit and integral interchange under mild hypotheses (Monotone Convergence, Fatou, Dominated Convergence). Key result in words: If a sequence of measurable functions converges pointwise and is suitably controlled (monotone, or dominated by an integrable function), then the integral of the limit equals the limit of the integrals. Where it applies: Real and harmonic analysis, probability theory (expectation = integral against a probability measure), PDE, functional analysis (

L^{p}

spaces), ergodic theory. Who/when: Henri Lebesgue (thesis 1902), building on Émile Borel's measure theory and Camille Jordan; convergence theorems refined by Pierre Fatou (1906), Beppo Levi (1906), Giuseppe Vitali, and later abstracted by Maurice Fréchet and Johann Radon.

The Lebesgue integral of a measurable function $f : X \to [- \infty, \infty]$ on a measure space $(X, M, μ)$ is defined first for nonnegative measurable functions as the supremum of integrals of simple functions lying below $f$ , then extended to signed and complex functions by splitting into positive and negative parts. Its central virtue over the Riemann integral is completeness of the limit theory: the class of Lebesgue-integrable functions is closed under the limit operations that pervade analysis, and the resulting $L^{p}$ spaces are complete (Banach). A surprising fact is that the Lebesgue integral integrates the Dirichlet function (indicator of the rationals) effortlessly — its integral is $0$ — whereas this function is nowhere Riemann integrable. The convergence theorems are the workhorses: they are what make Lebesgue's theory usable, converting the abstract definition into a calculus of limits.

The Core Insight

Riemann slices the domain into intervals and asks whether the function varies little on each; this fails when $f$ oscillates wildly. Lebesgue slices the range into thin horizontal layers $[y_{k}, y_{k + 1})$ and measures the size $μ (f^{- 1} [y_{k}, y_{k + 1}))$ of the set where $f$ takes those values. Because measure is countably additive and stable under monotone limits, integration inherits exactly the limit-passage properties that Riemann's domain-partition lacks. Integration becomes measuring the region under the graph — but measured with the flexible, $σ$ -additive Lebesgue measure rather than the rigid Jordan content.

Formal Statement

Measure space and measurable functions

A measure space is a triple $(X, M, μ)$ where $M$ is a $σ$ -algebra of subsets of $X$ and $μ : M \to [0, \infty]$ is countably additive:

μ (n = 1 ⨆ \infty E_{n}) = n = 1 \sum \infty μ (E_{n}), E_{n} \in M pairwise disjoint .

⟳This is the defining axiom; finite additivity follows by padding with

\emptyset

, and

μ (\emptyset) = 0

is required to avoid

μ \equiv \infty

A function $f : X \to [- \infty, \infty]$ is $M$ -measurable if $f^{- 1} ((a, \infty]) \in M$ for every $a \in R$ . Equivalently preimages of all Borel sets lie in $M$ .

The Lebesgue integral, in three stages

Stage 1 — Simple functions. A simple function is

φ = \sum_{k = 1}^{n} a_{k} 1_{A_{k}}

with

a_{k} \geq 0

A_{k} \in M

. Its integral is

\int_{X} φ d μ = k = 1 \sum n a_{k} μ (A_{k}) \in [0, \infty],

⟳well-defined independent of representation because two simple representations share a common refinement on which finite additivity of

μ

forces agreement.

Stage 2 — Nonnegative measurable $f \geq 0$ :

\int_{X} f d μ = sup {\int_{X} φ d μ : 0 \leq φ \leq f, φ simple} .

⟳the supremum is taken over all simple minorants; monotonicity of the integral on simple functions makes this a genuine least-upper-bound construction approximating the area under the graph from below.

Stage 3 — General $f$ . Write

f = f^{+} - f^{-}

with

f^{+} = max (f, 0)

f^{-} = max (- f, 0)

. If at least one of

\int f^{+}, \int f^{-}

is finite, define

\int_{X} f d μ = \int_{X} f^{+} d μ - \int_{X} f^{-} d μ .

⟳the splitting is canonical since

f^{+} f^{-} = 0

; we forbid the case

\infty - \infty

, which is when

f

is non-integrable.

f

integrable iff $\int ∣ f ∣ d μ < \infty$ , i.e. both parts finite.

The Three Convergence Theorems

Let $(X, M, μ)$ be a measure space and $f_{n} : X \to [- \infty, \infty]$ measurable.

(MCT) Monotone Convergence (Beppo Levi). If

0 \leq f_{1} \leq f_{2} \leq \dots

and

f_{n} ↑ f

pointwise, then

n \to \infty lim \int_{X} f_{n} d μ = \int_{X} f d μ = \int_{X} n \to \infty lim f_{n} d μ .

⟳monotonicity gives "

\leq

"; the reverse uses that for any simple

φ \leq f

and

0 < c < 1

, the sets

E_{n} = {f_{n} \geq c φ}

increase to

X

, then apply continuity of

μ

from below.

(Fatou) Fatou's Lemma. For

f_{n} \geq 0

\int_{X} n \to \infty lim inf f_{n} d μ \leq n \to \infty lim inf \int_{X} f_{n} d μ .

⟳apply MCT to

g_{n} = in f_{k \geq n} f_{k} ↑ lim inf f_{n}

, and use

g_{n} \leq f_{n}

\int g_{n} \leq \int f_{n}

(DCT) Dominated Convergence (Lebesgue). If

f_{n} \to f

pointwise (a.e.) and there exists integrable

g

with

∣ f_{n} ∣ \leq g

for all

n

, then

f

is integrable and

n \to \infty lim \int_{X} f_{n} d μ = \int_{X} f d μ, moreover n \to \infty lim \int_{X} ∣ f_{n} - f ∣ d μ = 0.

⟳apply Fatou to the nonnegative sequences

g + f_{n}

and

g - f_{n}

; the two resulting inequalities sandwich

\int f

between

lim inf

and

lim sup

\int f_{n}

Why It Holds / Derivation

Monotone Convergence Theorem

Step 1 (the easy inequality). Since

f_{n} \leq f_{n + 1} \leq f

, monotonicity of the integral on nonnegative functions gives

\int f_{n} \leq \int f_{n + 1} \leq \int f

. Hence

L := lim_{n} \int f_{n}

exists in

[0, \infty]

and

L \leq \int f

. Step 2 (setup for the reverse). Fix a simple function

φ = \sum_{k} a_{k} 1_{A_{k}}

with

0 \leq φ \leq f

, and fix a constant

c \in (0, 1)

. Define

E_{n} = {x \in X : f_{n} (x) \geq c φ (x)} \in M .

Step 3 ( $E_{n}$ exhausts $X$ ). Because

f_{n} ↑ f \geq φ > c φ

wherever

φ > 0

(strict since

c < 1

), and trivially where

φ = 0

, every

x

eventually lies in

E_{n}

. Thus

E_{n} ↑ X

. Step 4 (estimate). On

E_{n}

f_{n} \geq c φ

, so

\int_{X} f_{n} d μ \geq \int_{E_{n}} f_{n} d μ \geq c \int_{E_{n}} φ d μ = c k \sum a_{k} μ (A_{k} \cap E_{n}) .

Step 5 (continuity from below). Since

A_{k} \cap E_{n} ↑ A_{k}

σ

-additivity gives

μ (A_{k} \cap E_{n}) ↑ μ (A_{k})

(Equation [1]). Letting

n \to \infty

L \geq c k \sum a_{k} μ (A_{k}) = c \int_{X} φ d μ .

Step 6 (remove $c$ and $φ$ ). Let

c ↑ 1

L \geq \int φ

. Take the supremum over all simple

φ \leq f

: by definition (Equation [3]),

L \geq \int f

. Combined with Step 1,

L = \int f

■

Dominated Convergence from Fatou

Step 1. Since

∣ f_{n} ∣ \leq g

and

f_{n} \to f

a.e.,

∣ f ∣ \leq g

a.e., so

f

is integrable. Consider

h_{n}^{\pm} = g \pm f_{n} \geq 0

. Step 2. Apply Fatou (Equation [6]) to

h_{n}^{+} = g + f_{n}

\int (g + f) \leq n lim inf \int (g + f_{n}) = \int g + n lim inf \int f_{n} .

Cancel the finite

\int g

\int f \leq lim inf_{n} \int f_{n}

. Step 3. Apply Fatou to

h_{n}^{-} = g - f_{n}

\int (g - f) \leq n lim inf \int (g - f_{n}) = \int g - n lim sup \int f_{n} .

Cancel

\int g

lim sup_{n} \int f_{n} \leq \int f

. Step 4. Combining:

lim sup_{n} \int f_{n} \leq \int f \leq lim inf_{n} \int f_{n}

, forcing

lim_{n} \int f_{n} = \int f

. Step 5 ( $L^{1}$ convergence). Apply the same argument to

∣ f_{n} - f ∣ \leq 2 g

with pointwise limit

0

\int ∣ f_{n} - f ∣ \to 0

■

Henri Léon Lebesgue | 1875–1941

A French mathematician whose 1902 doctoral thesis Intégrale, longueur, aire at the Sorbonne overturned the foundations of integration. Lebesgue's signature insight — partitioning the range, "summing the money in your pocket by denomination rather than by the order you pull coins out" (a metaphor he himself used) — gave analysis its modern measure-theoretic backbone. He was initially met with skepticism; his contemporaries doubted the need to integrate such "pathological" functions. Lebesgue spent much of his career in pedagogy and held chairs at Rennes, Poitiers, and finally the Collège de France (1921). He was, somewhat ironically, suspicious of the abstract axiomatic generalizations (general measure spaces) his own work spawned, preferring the concrete geometry of the real line.

Building Intuition

Picture the graph of $f \geq 0$ over $X$ . Riemann drops vertical fences at points $x_{0} < x_{1} < \dots$ and sums rectangle areas $f (ξ_{i}) (x_{i + 1} - x_{i})$ ; if $f$ jumps madly within a single $[x_{i}, x_{i + 1}]$ , the choice of sample $ξ_{i}$ matters and the sum need not converge. Lebesgue instead draws horizontal lines at heights $y_{0} < y_{1} < \dots$ , asks "how much of $X$ has value in $[y_{k}, y_{k + 1})$ ?", and sums $y_{k} \cdot μ ({y_{k} \leq f < y_{k + 1}})$ . The set ${y_{k} \leq f < y_{k + 1}}$ can be a hopelessly complicated set — a fractal, the irrationals, anything measurable — and Lebesgue measure handles it. This is why the convergence theorems work: monotone increasing $f_{n}$ correspond to the horizontal layers filling in from the bottom, and countable additivity (continuity from below) lets you take the limit layer by layer without losing any area.

Range-partition (Lebesgue, gold) vs domain-partition (Riemann, gray): Lebesgue measures the preimage set of each horizontal band.

Worked Examples

The Dirichlet function — Riemann fails, Lebesgue trivial

Given:

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

, the indicator of the rationals in

[0, 1]

, with Lebesgue measure

λ

. Step 1.

f

is a simple function:

f = 1 \cdot 1_{A}

with

A = Q \cap [0, 1]

. Step 2. Compute

λ (A)

. The rationals are countable:

A = {q_{1}, q_{2}, \dots}

. By countable additivity (Equation [1]) and

λ ({q_{k}}) = 0

λ (A) = k = 1 \sum \infty λ ({q_{k}}) = k = 1 \sum \infty 0 = 0.

Step 3. By Equation [2],

\int_{[0, 1]} f d λ = 1 \cdot λ (A) = 1 \cdot 0

. Boxed answer:

\int_{[0, 1]} 1_{Q} d λ = 0

Note: the Riemann integral does not exist — upper sums are

1

, lower sums are

0

, for every partition.

Interchanging limit and integral via DCT

Given: Evaluate

n \to \infty lim \int_{0}^{1} \frac{n x}{1 + n ^{2} x ^{2}} d x

. (Lebesgue measure on

[0, 1]

.) Step 1 — pointwise limit. For fixed

x \in (0, 1]

f_{n} (x) = \frac{n x}{1 + n ^{2} x ^{2}} = \frac{x / n}{1/ n ^{2} + x ^{2}} \to 0

n \to \infty

. At

x = 0

f_{n} (0) = 0

. So

f_{n} \to 0

pointwise. Step 2 — find a dominating function. By AM–GM,

1 + n^{2} x^{2} \geq 2 n x

, hence

∣ f_{n} (x) ∣ = \frac{n x}{1 + n ^{2} x ^{2}} \leq \frac{n x}{2 n x} = \frac{1}{2} =: g (x) .

The constant

g = \frac{1}{2}

is integrable on

[0, 1]

since

\int_{0}^{1} \frac{1}{2} d x = \frac{1}{2} < \infty

. Step 3 — apply DCT (Equation [7]). Domination

∣ f_{n} ∣ \leq g

and

f_{n} \to 0

give

n \to \infty lim \int_{0}^{1} f_{n} d x = \int_{0}^{1} n \to \infty lim f_{n} d x = \int_{0}^{1} 0 d x .

Boxed answer:

n \to \infty lim \int_{0}^{1} \frac{n x}{1 + n ^{2} x ^{2}} d x = 0

(Sanity check: the exact integral is

\frac{1}{2 n} ln (1 + n^{2}) \to 0

, confirming DCT.)

Common Misconceptions

Domination is not optional

Without an integrable dominating function $g$ , the conclusion of DCT can fail even when $f_{n} \to 0$ pointwise everywhere and each $\int f_{n}$ exists. Pointwise convergence alone never guarantees interchange. Likewise, MCT requires the sequence to be monotone increasing and nonnegative — a decreasing sequence to $0$ can have $\int f_{n} \neq \to 0$ (mass escapes to infinity). Fatou's inequality can be strict; equality is the exception, not the rule.

The escaping bump — DCT and MCT both can fail without hypotheses

On $(R, λ)$ let $f_{n} = 1_{[n, n + 1]}$ . Then $f_{n} \to 0$ pointwise (for each fixed $x$ , $f_{n} (x) = 0$ once $n > x$ ), but

\int_{R} f_{n} d λ = λ ([n, n + 1]) = 1 \neq \to 0 = \int_{R} 0 d λ .

Fatou is satisfied (and strict):

\int lim inf f_{n} = 0 < 1 = lim inf \int f_{n}

. DCT fails because any

g \geq sup_{n} f_{n} = 1_{[1, \infty)}

has

\int g = \infty

— no integrable dominator exists. A second classic: the "tall thin spike"

g_{n} = n 1_{[0, 1/ n]}

has

\int g_{n} = 1

for all

n

but

g_{n} \to 0

a.e.; mass escapes vertically. Both show: pointwise convergence controls nothing about integrals.

Connections

The Lebesgue integral is the analytic engine behind several neighbouring theories. In probability, expectation $E [X] = \int_{Ω} X d P$ is literally a Lebesgue integral against a probability measure; the convergence theorems become the Monotone/Dominated Convergence for expectations, and Fatou underlies many a.s.-convergence arguments. The completeness of Lp Spaces (Riesz–Fischer theorem) is proved precisely via the convergence theorems applied to rapidly Cauchy subsequences. Fubini–Tonelli extends single integrals to product measures, with Tonelli's half resting on MCT and Fubini's on DCT. The Radon-Nikodym Theorem characterizes when one measure is an integral of a density against another, generalizing the integral itself. And Differentiation Under the Integral Sign (Leibniz rule) is just DCT applied to difference quotients.

Real-World Applications

Quantum mechanics — Wavefunctions live in $L^{2} (R^{3})$ ; normalization $\int ∣ ψ ∣^{2} d x = 1$ and expectation values $⟨ \hat{A} ⟩ = \int \overset{ˉ}{ψ} \hat{A} ψ$ are Lebesgue integrals, and the Hilbert-space completeness of $L^{2}$ (needing Lebesgue) underpins the spectral theorem.
Signal processing & Fourier analysis — The Fourier transform $\hat{f} (ξ) = \int f (x) e^{- 2 π i x ξ} d x$ requires $L^{1}$ / $L^{2}$ theory; Plancherel's theorem and the inversion formula depend on DCT and $L^{2}$ completeness.
Financial mathematics — Option pricing under risk-neutral measures computes $e^{- r T} E^{Q} [payoff]$ , a Lebesgue integral against the martingale measure; the Girsanov theorem changes measure via Radon–Nikodym densities.
Machine learning / statistics — Maximum likelihood and Bayesian posteriors integrate densities; the law of large numbers and consistency proofs invoke DCT and the monotone convergence of partial expectations.
Ergodic theory & dynamical systems — Birkhoff's pointwise ergodic theorem and the equality of time and space averages are statements about Lebesgue integrals invariant under measure-preserving maps.
PDE & numerical analysis — Weak (Sobolev) solutions are defined by integral identities; Galerkin and finite-element convergence proofs use DCT to pass to limits in the weak formulation.
Image processing — Total-variation denoising and the coarea formula treat images as $B V$ functions integrated against Lebesgue measure, decomposing them by level sets exactly as Lebesgue's range-partition suggests.

Historical Development

1854 — Bernhard Riemann, in his Habilitation, gives the first rigorous definition of the integral via domain partitions and the Riemann criterion.
1881–1884 — Vito Volterra and others construct functions (e.g. Volterra's function) exposing the limits of Riemann integration vis-à-vis the Fundamental Theorem of Calculus.
1892–1898 — Camille Jordan and Giuseppe Peano develop "Jordan content," a finitely-additive notion of measure foreshadowing but inadequate for limits.
1898 — Émile Borel introduces countably-additive measure on the Borel sets of $R$ in his Leçons sur la théorie des fonctions, the decisive conceptual step.
1902 — Henri Lebesgue's thesis Intégrale, longueur, aire defines Lebesgue measure and integral, proving the bounded convergence theorem.
1906 — Pierre Fatou's thesis on trigonometric series states and uses Fatou's Lemma; Beppo Levi proves the Monotone Convergence Theorem the same year.
1907 — Frigyes Riesz and Ernst Fischer prove the completeness of $L^{2}$ (Riesz–Fischer theorem), revealing $L^{p}$ as Banach spaces.
1910–1913 — Johann Radon and Maurice Fréchet abstract integration to general measure spaces; Radon measures and the Radon–Nikodym derivative emerge.
1933 — Andrey Kolmogorov's Grundbegriffe founds probability on measure theory, making the Lebesgue integral the definition of expectation.

Subtleties & Limits

The Lebesgue integral is absolutely convergent — it integrates $f$ only when $\int ∣ f ∣ < \infty$ — and so cannot directly assign $\int_{0}^{\infty} \frac{s i n x}{x} d x = \frac{π}{2}$ (an improper, conditionally convergent Riemann integral). This motivates gauge / Henstock–Kurzweil integration, which recovers all such conditionally convergent integrals and contains Lebesgue's theory as the absolutely-integrable subcase; the precise functional-analytic structure of the Henstock space remains less developed than $L^{p}$ theory. A second frontier is non-measurable phenomena: assuming the Axiom of Choice, Vitali sets and Banach–Tarski paradoxes show Lebesgue measure cannot be extended to all subsets of $R^{n}$ translation-invariantly; the consistency strength of "every set is Lebesgue measurable" (Solovay's model, requiring an inaccessible cardinal) ties measure theory to large-cardinal axioms. Finally, infinite-dimensional analogues fail: there is no translation-invariant $σ$ -finite measure on an infinite-dimensional Banach space, forcing the use of Gaussian (Wiener) measures and abstract Wiener spaces, where the right convergence theorems are still actively studied.

The Lebesgue Integral and Convergence Theorems

The Core Insight

Formal Statement

Measure space and measurable functions

The Lebesgue integral, in three stages

The Three Convergence Theorems

Why It Holds / Derivation

Monotone Convergence Theorem

Dominated Convergence from Fatou

Henri Léon Lebesgue | 1875–1941

Building Intuition

Range-partition (Lebesgue, gold) vs domain-partition (Riemann, gray): Lebesgue measures the preimage set of each horizontal band.

Worked Examples

The Dirichlet function — Riemann fails, Lebesgue trivial

Interchanging limit and integral via DCT

Common Misconceptions

Domination is not optional

The escaping bump — DCT and MCT both can fail without hypotheses

Connections

Real-World Applications

Historical Development

Subtleties & Limits

See Also

Before this

Where this leads