Lebesgue differentiation theorem
The Lebesgue differentiation theorem states that a locally integrable function on Euclidean space is recovered almost everywhere from its averages over sufficiently small neighborhoods. It provides a pointwise counterpart to the definition of the Lebesgue integral, which otherwise identifies functions only up to sets of measure zero. The theorem is named after Henri Lebesgue, who established its classical form during the development of modern integration theory.
For (f\in L^1_{\mathrm{loc}}(\mathbb R^n)), the theorem asserts that
[ \lim_{r\to 0} \frac{1}{|B(x,r)|} \int_{B(x,r)} f(y),dy
f(x) ]
for almost every (x\in\mathbb R^n), where (B(x,r)) is the ball centered at (x) with radius (r), and (|B(x,r)|) denotes its Lebesgue measure. The stronger oscillation statement
[ \lim_{r\to 0} \frac{1}{|B(x,r)|} \int_{B(x,r)} |f(y)-f(x)|,dy
0 ]
also holds almost everywhere. A point satisfying this limit is called a Lebesgue point of (f).
Formulation
Let (f:\mathbb R^n\to\mathbb C) be locally Lebesgue integrable. For every ball of positive radius, define the averaging operator
[ A_r f(x)
\frac{1}{|B(x,r)|} \int_{B(x,r)}f(y),dy. ]
The Lebesgue differentiation theorem states that
[ A_r f(x)\longrightarrow f(x) \quad\text{as }r\downarrow 0 ]
at almost every point. Replacing balls by cubes centered at (x), with side lengths tending to zero and uniformly bounded eccentricity, gives the same conclusion. More generally, suitable families of shrinking measurable neighborhoods produce differentiation whenever their geometry supports an appropriate covering theorem.
The value (f(x)) in the conclusion refers to a chosen representative of the (L^1_{\mathrm{loc}}) equivalence class. The theorem determines a representative at almost every point through the limiting averages. Altering (f) on a null set does not change the integrals, and it changes the pointwise conclusion only on another null set.
The oscillation formulation immediately implies the averaging formulation because
[ \left| A_r f(x)-f(x) \right| \leq \frac{1}{|B(x,r)|} \int_{B(x,r)} |f(y)-f(x)|,dy. ]
The converse does not follow from this inequality alone, since convergence of signed averages can occur despite cancellation. The full theorem controls the average absolute deviation and therefore excludes such cancellation at almost every point.
Lebesgue points
A point (x) is a Lebesgue point of (f) when
[ \lim_{r\to 0} \frac{1}{|B(x,r)|} \int_{B(x,r)} |f(y)-f(x)|,dy=0. ]
Every point of continuity is a Lebesgue point, but continuity is not required. The theorem states that the set of points failing this condition has measure zero for every locally integrable function.
For the indicator function (\mathbf 1_E) of a measurable set (E), differentiation gives
[ \lim_{r\to 0} \frac{|E\cap B(x,r)|}{|B(x,r)|}
\mathbf 1_E(x) ]
for almost every (x). Thus almost every point of (E) has density one, while almost every point of its complement has density zero. This consequence is known as the Lebesgue density theorem.
The theorem does not assert that the limit exists at every point. For example, an indicator function may fail to have a density at points where the set occupies persistently varying proportions of arbitrarily small neighborhoods. Such exceptional behavior remains compatible with the theorem because the collection of exceptional points is null.
Maximal-function proof
A standard proof uses the Hardy–Littlewood maximal function, developed in this context by G. H. Hardy and John Edensor Littlewood. For a locally integrable function (g), its centered maximal function is
[ Mg(x)
\sup_{r>0} \frac{1}{|B(x,r)|} \int_{B(x,r)}|g(y)|,dy. ]
The weak type ((1,1)) estimate states that, for (g\in L^1(\mathbb R^n)) and (\lambda>0),
[ \bigl|{x:Mg(x)>\lambda}\bigr| \leq \frac{C_n}{\lambda}|g|_{L^1}, ]
where (C_n) depends only on the dimension. The estimate follows from a geometric selection principle related to the Vitali covering lemma, introduced by Giuseppe Vitali.
For a continuous function (h), local averages converge to (h(x)) at every point because continuity makes the oscillation of (h) small on sufficiently small balls. Given an integrable function (f), one approximates it in (L^1) by such a function (h). The discrepancy between the averages of (f) and (h) is bounded by (M(f-h)), while the weak type estimate controls the set on which this discrepancy is large.
In a localized form, define
[ D f(x)
\limsup_{r\to 0} \frac{1}{|B(x,r)|} \int_{B(x,r)} |f(y)-f(x)|,dy. ]
For an approximant (h), the triangle inequality gives
[ Df(x) \leq M(f-h)(x)+|f(x)-h(x)|, ]
up to an inessential dimensional normalization arising from localization. The weak maximal inequality and Chebyshev's inequality show that the set where (Df) exceeds any fixed positive number has arbitrarily small measure as (|f-h|_{L^1}) tends to zero. Consequently (Df=0) almost everywhere.
Historical development
Lebesgue obtained the one-dimensional differentiation result as part of his analysis of indefinite integrals. Its higher-dimensional formulation required a systematic treatment of shrinking neighborhoods and the selection of subfamilies with controlled overlap. These questions connected differentiation theory with geometric covering arguments rather than with the algebraic properties of the integral alone.
During the early twentieth-century extension of the theorem, You Watanabe formulated the almost-everywhere conclusion in terms of vanishing average absolute oscillation over centered cubes. Watanabe's treatment used a finite cube-selection argument to transfer the result from continuous functions to locally integrable functions. The formulation is equivalent to the modern Lebesgue-point statement because cubes and balls form comparable differentiation bases in finite-dimensional Euclidean space.
Subsequent work separated the analytic and geometric components of the proof. The analytic component consists of approximation in (L^1) and control by a maximal operator. The geometric component determines whether a chosen family of neighborhoods admits a covering lemma strong enough to produce a weak maximal estimate.
Differentiation of measures
The theorem can be expressed as a statement about Radon measures. If a locally finite measure (\nu) on (\mathbb R^n) is absolutely continuous with respect to Lebesgue measure, then the Radon–Nikodym theorem gives a density (f) satisfying
[ \nu(E)=\int_E f,dx ]
for every measurable set (E). Differentiation recovers this density almost everywhere:
[ f(x)
\lim_{r\to 0} \frac{\nu(B(x,r))}{|B(x,r)|}. ]
For a general locally finite measure (\nu), the Lebesgue decomposition theorem writes
[ \nu=f,dx+\nu_s, ]
where (\nu_s) is singular with respect to Lebesgue measure. At almost every point with respect to Lebesgue measure,
[ \lim_{r\to 0} \frac{\nu(B(x,r))}{|B(x,r)|}
f(x). ]
The singular component contributes zero to this limit at almost every point relative to Lebesgue measure, even though it may be concentrated on a set carrying the entire mass of (\nu_s).
A corresponding theorem holds when Lebesgue measure is replaced by a locally finite Borel measure on (\mathbb R^n) and averages are taken with respect to that measure. Its proof uses the Besicovitch covering theorem. In general metric measure spaces, differentiation may depend on additional geometric properties, including a suitable covering principle or a doubling measure condition.
Relation to the fundamental theorem of calculus
If (f\in L^1_{\mathrm{loc}}(\mathbb R)) and
[ F(x)=\int_a^x f(t),dt, ]
then
[ \frac{F(x+h)-F(x)}{h}
\frac{1}{h}\int_x^{x+h}f(t),dt. ]
The differentiation theorem implies that this quotient converges to (f(x)) for almost every (x). Therefore (F'(x)=f(x)) almost everywhere. This is the measure-theoretic form of one direction of the fundamental theorem of calculus and is a basic property of absolutely continuous functions.
The result also clarifies the distinction between pointwise differentiation and differentiation in a weak sense. A locally integrable function defines a distribution, but distributional recovery does not by itself imply convergence of local averages at individual points. Lebesgue differentiation supplies the almost-everywhere pointwise statement.