Lebesgue Differentiation Theorem: Averages Recover the Function

The precise statement

Let f ∈ L¹_loc(ℝⁿ), meaning f is Lebesgue measurable and ∫_K |f| < ∞ on every compact set K. Write B(x,r) for the open ball of radius r and |B(x,r)| for its Lebesgue measure. The theorem has two levels of strength.

• Weak form (differentiation of the integral): for Lebesgue-a.e. x,  lim_r→0 (1/|B(x,r)|) ∫_B(x,r) f(y) dy = f(x).
• Strong form (Lebesgue points): for a.e. x,  lim_r→0 (1/|B(x,r)|) ∫_B(x,r) |f(y) − f(x)| dy = 0.

A point x where the second equation holds is a Lebesgue point of f. The strong form is genuinely stronger: it says the average of the oscillation around f(x) vanishes, which forbids the average from converging by lucky cancellation of large positive and negative excursions. The balls may be replaced by any family shrinking nicely to x (bounded eccentricity), e.g. cubes.

The picture: zooming in undoes the smear

Integrating f smears its values together; the average A_rf(x) = (1/|B|)∫_B(x,r) f is a blurred snapshot of f near x at resolution r. The theorem says that as you sharpen the resolution (r → 0), the blur disappears and you see the true pixel value f(x) — for almost every pixel.

Why only almost every? Because f is defined only up to sets of measure zero, so no statement can hold at every point (you can wreck any single point by redefining f there). The miracle is that the bad set is not just avoidable — it has measure zero, uniformly, for every integrable f.

• At a continuity point, convergence is obvious: f is nearly constant on a tiny ball.
• The content is that continuity points can be scarce (a generic L¹ function is discontinuous everywhere), yet Lebesgue points are still a.e.

The mechanism: the maximal inequality

The engine is the Hardy–Littlewood maximal function (Hardy & Littlewood 1930), Mf(x) = sup_r>0 (1/|B(x,r)|) ∫_B(x,r) |f|. The key estimate is the weak-(1,1) inequality: for every λ > 0,

|{x : Mf(x) > λ}| ≤ (3ⁿ/λ) ‖f‖_L¹,

proved by a Vitali covering argument — from any cover by balls extract a disjoint subfamily whose 3× dilations cover the whole set, so total measure is controlled by ∑ integrals. Now the standard trick: split f = g + h with g continuous and ‖h‖_L¹ < ε (continuous functions are dense in L¹). For g the averages converge everywhere. The error is controlled by the oscillation operator, which is ≤ 2Mh + |h|; the maximal inequality bounds the measure of {oscillation > λ} by Cε/λ. Since ε is arbitrary, the bad set has measure zero. That density-plus-maximal-bound pattern is one of the most important templates in analysis.

Worked special case: the density theorem

Take f = 1_E, the indicator of a measurable set E ⊂ ℝⁿ. The average over B(x,r) is exactly the local density of E:

D(x,r) = |E ∩ B(x,r)| / |B(x,r)|.

The Lebesgue Differentiation Theorem gives the Lebesgue Density Theorem: lim_r→0 D(x,r) = 1 for a.e. x ∈ E, and = 0 for a.e. x ∉ E. So a measurable set has density either 1 or 0 at almost all points — intermediate densities occur only on a null set.

• Concrete surprise: the fat Cantor set C ⊂ [0,1] has positive measure yet empty interior. Every point of C is a limit of gaps, so it "looks" boundary-like everywhere. Nonetheless a.e. point of C has density 1: the gaps shrink fast enough that they become negligible at small scales.

This corollary alone is the standard route to "a positive-measure set is essentially thick at almost all of its points," used constantly in geometric measure theory.

Why the hypotheses matter

Local integrability is essential. If f ∉ L¹_loc, the averages can diverge on a large set — the very quantity you are averaging is ill-defined. Some integrability is unavoidable to even state the average.

"Nicely shrinking" is essential. The theorem allows shrinking sets of bounded eccentricity (balls, cubes). If you drop this and average over arbitrarily thin, long rectangles centered at x, convergence fails: this is the Zygmund / strong maximal function problem, and the analogue holds only under stronger conditions (e.g. f ∈ L(log L) via Jessen–Marcinkiewicz–Zygmund). Anisotropic averaging genuinely breaks a.e. convergence.

• Relation to the FTC: in n = 1, F(x) = ∫_a^x f satisfies F′ = f a.e. — the measure-theoretic FTC. But note the converse subtlety: the Cantor function is continuous, increasing, has derivative 0 a.e., yet climbs from 0 to 1. So "F′ = f a.e." does not reconstruct F unless F is absolutely continuous.
• It underlies the Radon–Nikodym derivative's pointwise realization and the Besicovitch/differentiation theory for general Radon measures.

Why it matters and what it unlocks

The theorem is a load-bearing wall of modern analysis. It is what makes L^p functions behave like functions with values, not just equivalence classes.

• Harmonic analysis: the Hardy–Littlewood maximal function built for the proof is itself fundamental — it controls approximate identities, gives a.e. convergence of convolutions f ∗ φ_ε → f, and underpins Calderón–Zygmund singular-integral theory.
• Fourier analysis: the same maximal-function machinery is the template for Carleson's theorem on a.e. convergence of Fourier series.
• Radon–Nikodym / probability: gives the pointwise a.e. value of dμ/dλ; the martingale-convergence analogue (Lévy) makes it the deterministic cousin of the martingale differentiation theorem.
• Geometric measure theory: density points, rectifiability, and blow-up arguments all begin here.
• Sobolev / PDE: Lebesgue points are the canonical points at which a Sobolev function has a well-defined precise representative.

In short: it certifies that "the function equals the limit of its averages" almost everywhere — the license to talk about pointwise values in the measure-theoretic world.

The Lebesgue Differentiation Theorem versus neighboring results on recovering a function from its integral

Result	Hypothesis	Conclusion	What it needs
Fundamental Theorem of Calculus (classical)	f continuous on [a,b]	(d/dx)∫ₐˣ f = f(x) everywhere	Continuity — pointwise everywhere
Lebesgue Differentiation Theorem	f ∈ L¹_loc(ℝⁿ)	Averages over B(x,r) → f(x) a.e.	Only integrability; conclusion holds a.e.
Lebesgue Density Theorem	E measurable ⊂ ℝⁿ	Density of E is 1 a.e. on E, 0 a.e. off E	Special case: f = 1_E
FTC for absolutely continuous f	f: [a,b]→ℝ absolutely continuous	f(x)=f(a)+∫ₐˣ f′, f′ exists a.e.	Absolute continuity (rules out Cantor function)
Besicovitch differentiation	μ Radon on ℝⁿ, f ∈ L¹(μ)	μ-averages → f(x) for μ-a.e. x	Besicovitch covering, no doubling needed