02.07.12 · analysis / measure-theory

The Lebesgue differentiation theorem and the Hardy-Littlewood maximal function

shipped3 tiersLean: none

Anchor (Master): Stein, Singular Integrals and Differentiability Properties of Functions (Princeton 1970) Ch. I; de Guzmán, Differentiation of Integrals in R^n (Springer LNM 481, 1975); Stein, Harmonic Analysis (Princeton 1993) §I.2-3

Intuition Beginner

The fundamental theorem of calculus says that the derivative of the integral recovers the original function. The Lebesgue differentiation theorem is the higher-dimensional, measure-theoretic version of this slogan. Take a locally integrable function on Euclidean space, average it over a small ball centred at a point, and let the ball shrink. For almost every point, the averages converge to the value of the function at that point.

The slogan "the average over a small ball tends to the value at the centre" sounds harmless, but it is false for every point if the function is too wild. The right statement quantifies "almost every": the set of bad points has Lebesgue measure zero. The proof controls the bad set using the Hardy-Littlewood maximal function, an envelope that records the largest average over every ball centred at the point.

The maximal function is a non-linear operator: it takes the supremum over all ball radii. The deep theorem of Hardy and Littlewood says the maximal function is not too much bigger than the original function: the set where it exceeds a level lambda has measure at most a constant times lambda-inverse times the L1 norm. This weak-type bound is the analytic engine behind the differentiation theorem and behind much of modern harmonic analysis.

A useful slogan: the maximal function is the price you pay to control pointwise averages uniformly over all radii, and the weak-type bound says the price is a single constant times the L1 mass, with the constant depending only on the dimension.

Visual Beginner

Picture a step function on the real line: value zero on the negative half-line and value one on the non-negative half-line, with a jump at the origin. Take a point away from the jump, say at $x = 2$ . The average over a small ball around $x = 2$ is exactly one, because the ball sits entirely in the region where the function is constant. Shrink the ball and the average stays at one, matching the value of the function at $x = 2$ .

Now take the jump point itself, $x = 0$ . A ball of radius $r$ around zero stretches from minus $r$ to plus $r$ . Half of it sits in the zero region and half in the one region, so the average is one half. Shrink the ball and the average stays at one half, never approaching the value at the origin (which is one). The jump point is not a Lebesgue point of the step function; every other point is.

The maximal function is the worst-case envelope. At $x = 2$ , the supremum over all ball-averages is one. At $x = 0$ , the supremum over all ball-averages is also one (take a ball reaching deep into the one region). At a point $x = - 5$ far in the zero region, the maximal function is about one half, achieved by a ball reaching to the jump. The maximal function records, for each point, the largest average achievable by any ball, and the weak-type bound controls how often this envelope can exceed a given threshold.

Worked example Beginner

Let us compute the running average of a concrete step function around two points: a Lebesgue point and a non-Lebesgue point. Let $f (x) = 3$ on $[0, 2)$ and $f (x) = 5$ on $[2, 4]$ , and $f (x) = 0$ elsewhere. The total mass on the support $[0, 4]$ is $3 \cdot 2 + 5 \cdot 2 = 16$ .

Point one: $x = 1$ (inside the first piece, a Lebesgue point). Take a ball (an interval on the real line) of radius $r = 0.5$ centred at $1$ . The interval $(0.5, 1.5)$ lies entirely inside $[0, 2)$ where $f = 3$ , so the average of $f$ over this ball is $3$ . Shrink to radius $r = 0.1$ : the interval $(0.9, 1.1)$ is still inside $[0, 2)$ , so the average is still $3$ , which equals $f (1)$ . The averages converge to $f (1) = 3$ as $r$ shrinks, confirming $x = 1$ as a Lebesgue point.

Point two: $x = 2$ (the jump, not a Lebesgue point). Take a ball of radius $r = 1$ centred at $2$ . The interval $(1, 3)$ covers length $1$ of the first piece (where $f = 3$ ) and length $1$ of the second piece (where $f = 5$ ). The average is $(3 \cdot 1 + 5 \cdot 1) /2 = 4$ . Shrink to radius $r = 0.1$ : the interval $(1.9, 2.1)$ still covers length $0.1$ of each piece, so the average is still $(3 \cdot 0.1 + 5 \cdot 0.1) /0.2 = 4$ . The average is stuck at $4$ for every small radius, but $f (2) = 5$ . The point $x = 2$ is not a Lebesgue point.

What this tells us: at a jump discontinuity, the symmetric ball-average is pinned at the midpoint of the left and right limits, never approaching the function value. The Lebesgue differentiation theorem handles this by excluding the jump points, which in this example form a measure-zero set (a single point).

Check your understanding Beginner

Exercise (easy, multiple choice).

For $f \in L_{loc}^{1} (R^{n})$ , the Lebesgue differentiation theorem states that the average of $f$ over balls $B (x, r)$ centred at $x$ converges to $f (x)$ as $r \to 0$ :

A. For every point $x$ B. For every point $x$ at which $f$ is continuous, and possibly nowhere else C. For $m$ -almost every point $x$ (a set of full Lebesgue measure) D. For every point $x$ outside a countable set

Hint

The theorem cannot give convergence everywhere (the step function fails at its jump), and the bad set is measure-zero but need not be countable.

Answer

C. For $m$ -almost every point $x$ . The theorem gives convergence on the Lebesgue set, which has full measure. Feedback-wrong: A fails at jumps of bounded functions; B is too weak — continuity is sufficient but not necessary, and the Lebesgue set is strictly larger than the continuity set; D is too strong — the exceptional set can be uncountable (a measure-zero Cantor-type set of discontinuities of a suitably constructed bounded function).

Formal definition Intermediate+

Let $m$ denote Lebesgue measure on $R^{n}$ , let $B (x, r) = {y \in R^{n} : ∣ y - x ∣ < r}$ be the open ball of radius $r$ centred at $x$ , and let $∣ B (x, r) ∣ = m (B (x, r)) = ω_{n} r^{n}$ with $ω_{n} = π^{n /2} /Γ (n /2 + 1)$ .

Definition (locally integrable function). A measurable $f : R^{n} \to C$ is locally integrable, written $f \in L_{loc}^{1} (R^{n}, m)$ , when $\int_{K} ∣ f ∣ d m < \infty$ for every compact $K \subseteq R^{n}$ . Equivalently, $f \cdot χ_{B (0, R)} \in L^{1} (R^{n})$ for every $R > 0$ .

Definition (ball average). For $f \in L_{loc}^{1}$ and $x \in R^{n}$ , $r > 0$ , the ball average is

A_{r} f (x) = \frac{1}{∣ B ( x , r ) ∣} \int_{B (x, r)} f (y) d m (y) .

The averaging operator $A_{r}$ is a linear contraction on every $L^{p} (R^{n})$ with $1 \leq p \leq \infty$ (by Jensen or Hölder), so $∥ A_{r} f ∥_{L^{p}} \leq ∥ f ∥_{L^{p}}$ .

Definition (Hardy-Littlewood maximal function). For $f \in L_{loc}^{1} (R^{n})$ , the centred Hardy-Littlewood maximal function is

M f (x) = r > 0 sup \frac{1}{∣ B ( x , r ) ∣} \int_{B (x, r)} ∣ f (y) ∣ d m (y) .

The values of $M f$ lie in $[0, \infty]$ . The uncentred version $M^{*} f (x) = sup {\frac{1}{∣ B ∣} \int_{B} ∣ f ∣ d m : B ∋ x, B a ball}$ takes the supremum over all balls containing $x$ rather than only centred balls; the two are comparable ( $M f \leq 2^{n} M^{*} f$ and $M^{*} f \leq M f$ ), and the distinction is inessential for the theorems below ^{[Stein 1970]}.

Definition (weak-type and strong-type bounds). A sublinear operator $T$ (meaning $∣ T (f + g) ∣ \leq ∣ T f ∣ + ∣ T g ∣$ and $∣ T (α f) ∣ = ∣ α ∣ ∣ T f ∣$ ) is weak-type $(p, p)$ , $1 \leq p < \infty$ , when

m ({x : ∣ T f (x) ∣ > λ}) \leq (\frac{C ∥ f ∥ _{L^{p}}}{λ})^{p}, λ > 0,

and strong-type $(p, p)$ when $∥ T f ∥_{L^{p}} \leq C ∥ f ∥_{L^{p}}$ . Weak-type $(p, p)$ is equivalent to boundedness $T : L^{p} \to L^{p, \infty}$ into weak- $L^{p}$ (Lorentz space). Strong-type $(p, p)$ implies weak-type $(p, p)$ (by Markov's inequality); the converse fails at endpoints.

Definition (Lebesgue set and Lebesgue points). For $f \in L_{loc}^{1} (R^{n})$ , the Lebesgue set is

Leb (f) = {x \in R^{n} : r \to 0 lim \frac{1}{∣ B ( x , r ) ∣} \int_{B (x, r)} ∣ f (y) - f (x) ∣ d m (y) = 0} .

Points of $Leb (f)$ are Lebesgue points of $f$ : the average modulus of deviation $∣ f (y) - f (x) ∣$ tends to zero. Every point of continuity of $f$ is a Lebesgue point, and so is every point where the approximate limit of $f$ equals $f (x)$ . The Lebesgue differentiation theorem asserts $m (R^{n} ∖ Leb (f)) = 0$ ^{[Folland 1999]}.

Counterexamples to common slips Intermediate+

The maximal function of an $L^{1}$ function need not be in $L^{1}$ . Take $f = χ_{B (0, 1)}$ . For $∣ x ∣ > 1$ , choosing $r = ∣ x ∣ + 1$ gives $M f (x) \geq ∣ B (0, 1) ∣/∣ B (x, ∣ x ∣ + 1) ∣ ≍ ∣ x ∣^{- n}$ , which is not integrable at infinity. So $M : L^{1} \neq \to L^{1}$ ; the weak-type $(1, 1)$ bound is the best possible substitute at the endpoint.
The strong-type $(1, 1)$ bound fails completely. For non-zero $f \in L^{1}$ , $M f (x) ≳ ∣ x ∣^{- n}$ at infinity, so $M f \in L^{1}$ only when $f = 0$ a.e. The endpoint $p = 1$ is genuinely weak-type only, and the Marcinkiewicz interpolation that gives strong-type for $p > 1$ blows up as $p \to 1^{+}$ at rate $1/ (p - 1)$ .
The pointwise bound $∣ A_{r} f (x) ∣ \leq M f (x)$ does not imply uniform $L^{1}$ -convergence of $A_{r} f$ to $f$ . Each $A_{r}$ is an $L^{1}$ -contraction and $A_{r} f \to f$ a.e. by the theorem below, but the convergence is not dominated in $L^{1}$ (no integrable envelope controls $∣ A_{r} f - f ∣$ uniformly), so DCT does not give $∥ A_{r} f - f ∥_{1} \to 0$ . (The $L^{1}$ -convergence does hold when $f \in L^{1} \cap L^{\infty}$ or more generally when $f$ lies in $L^{p}$ for some $p > 1$ , by the strong-type bound and a separate approximation argument.)
The Vitali factor $3^{n}$ (or $5^{n}$ ) is the geometric price of disjointness. Besicovitch's covering theorem (Theorem 5 below) reduces the geometric loss in some settings, but no covering argument can give the weak-type bound with constant $1$ ; the dimensional constant is irreducible.
"Almost everywhere" cannot be improved to "everywhere" even for bounded functions. The step function of the worked example fails at its jump. The bad set is the complement of the Lebesgue set, which is measure-zero for $f \in L_{loc}^{1}$ but typically non-empty.

Key theorem with proof Intermediate+

Theorem (Hardy-Littlewood maximal weak-type bound; Hardy-Littlewood 1930 in $n = 1$ , Wiener 1939 in general $n$ ). There is a constant $C_{n}$ depending only on the dimension such that for every $f \in L^{1} (R^{n})$ and every $λ > 0$ ,

m ({x \in R^{n} : M f (x) > λ}) \leq \frac{C _{n}}{λ} ∥ f ∥_{L^{1}} .

The constant $C_{n} = 3^{n}$ under the centred-ball Vitali lemma (Proposition 1 below), and can be improved to the Besicovitch constant $B_{n}$ with refined covering.

The proof goes through the Vitali covering lemma: from the collection of balls witnessing $M f (x) > λ$ , extract a disjoint subcollection covering a fixed fraction of the bad set, then sum the disjoint volumes against the level-set inequality. The full argument is in §Full proof set (Propositions 1-2) ^{[Hardy-Littlewood 1930]} ^{[Wiener 1939]}.

Theorem (Lebesgue differentiation theorem; Lebesgue 1910 in $n = 1$ , Wiener 1939 in general $n$ ). For $f \in L_{loc}^{1} (R^{n})$ ,

r \to 0 lim A_{r} f (x) = r \to 0 lim \frac{1}{∣ B ( x , r ) ∣} \int_{B (x, r)} f (y) d m (y) = f (x) for m -a.e. x .

Equivalently, $m (R^{n} ∖ Leb (f)) = 0$ .

Proof. By countable covering it suffices to prove the claim on each ball $B (0, R)$ ; on $B (0, R)$ the local integrability of $f$ gives $f \cdot χ_{B (0, 2 R)} \in L^{1}$ , so we work with $f \in L^{1}$ in what follows and apply the result to $f \cdot χ_{B (0, 2 R)}$ .

Step 1 (reduce to a continuous approximant). By density of $C_{c} (R^{n})$ in $L^{1}$ (a consequence of Lusin's theorem 02.07.03 plus absolute continuity of the integral 02.07.04), for any $ε > 0$ choose $g \in C_{c} (R^{n})$ with $∥ f - g ∥_{L^{1}} < ε$ . Set $h = f - g$ , so $∥ h ∥_{L^{1}} < ε$ .

Step 2 (continuous functions satisfy the limit everywhere). For $g$ continuous, $A_{r} g (x) \to g (x)$ as $r \to 0$ for every $x$ (not merely a.e.): given $η > 0$ , continuity at $x$ gives $δ > 0$ with $∣ g (y) - g (x) ∣ < η$ on $B (x, δ)$ , so $∣ A_{r} g (x) - g (x) ∣ \leq η$ for $r < δ$ . Hence $lim sup_{r \to 0} ∣ A_{r} g (x) - g (x) ∣ = 0$ everywhere.

Step 3 (maximal control of the error). Decompose by the triangle inequality inside the limsup:

r \to 0 lim sup ∣ A_{r} f (x) - f (x) ∣ \leq = 0 r \to 0 lim sup ∣ A_{r} g (x) - g (x) ∣ + r \to 0 lim sup ∣ A_{r} h (x) - h (x) ∣.

The second piece satisfies $∣ A_{r} h (x) - h (x) ∣ \leq ∣ A_{r} h (x) ∣ + ∣ h (x) ∣ \leq M h (x) + ∣ h (x) ∣$ , using that each average is pointwise dominated by the maximal function $M h (x)$ . Combining,

r \to 0 lim sup ∣ A_{r} f (x) - f (x) ∣ \leq M h (x) + ∣ h (x) ∣.

Step 4 (weak-type bound on the bad set). Fix $α > 0$ and define $E_{α} = {x : lim sup_{r \to 0} ∣ A_{r} f (x) - f (x) ∣ > α}$ . By Step 3,

E_{α} \subseteq {x : M h (x) > α /2} \cup {x : ∣ h (x) ∣ > α /2} .

The weak-type $(1, 1)$ bound gives $m ({M h > α /2}) \leq \frac{2 C _{n}}{α} ∥ h ∥_{L^{1}} < \frac{2 C _{n} ε}{α}$ . Markov's inequality gives $m ({∣ h ∣ > α /2}) \leq \frac{2}{α} ∥ h ∥_{L^{1}} < \frac{2 ε}{α}$ . Adding,

m (E_{α}) < \frac{2 ( C _{n} + 1 ) ε}{α} .

Step 5 (let $ε \to 0$ ). The set $E_{α}$ depends only on $f$ and $α$ , not on the approximant $g$ . Since the right side tends to $0$ as $ε \to 0$ , we conclude $m (E_{α}) = 0$ for every $α > 0$ . Taking $α = 1/ k$ for $k = 1, 2, \dots$ and countable union gives $m ({x : lim sup_{r \to 0} ∣ A_{r} f (x) - f (x) ∣ > 0}) = 0$ , hence $A_{r} f (x) \to f (x)$ $m$ -a.e.

Step 6 (upgrade to the Lebesgue set). Apply the a.e.-convergence result to $∣ f - q ∣$ for each rational $q \in Q$ : there is a full-measure set on which $A_{r} ∣ f - q ∣ (x) \to ∣ f (x) - q ∣$ . Intersect over the countably many $q$ . At any point $x$ of the intersection, for any $η > 0$ pick rational $q$ with $∣ f (x) - q ∣ < η$ ; then

r \to 0 lim sup A_{r} ∣ f - f (x) ∣ (x) \leq r \to 0 lim sup A_{r} ∣ f - q ∣ (x) + ∣ q - f (x) ∣ \leq 2∣ f (x) - q ∣ < 2 η .

Since $η$ was arbitrary, $x \in Leb (f)$ . So $m (R^{n} ∖ Leb (f)) = 0$ .

$□$

Bridge. The Lebesgue differentiation theorem builds toward the differentiation-of-measures framework of 02.07.08 (the Radon-Nikodym derivative as the limit of measure ratios $ν (B (x, r)) / m (B (x, r))$ , of which the ball-average of $f$ is the special case $d ν = f d m$ ), and appears again in 02.07.10 Rademacher's theorem (the line-by-line a.e. differentiability of Lipschitz functions rests on the one-dimensional Lebesgue differentiation theorem applied along coordinate directions) and in 02.07.11 the area and coarea formulas (the Jacobian/change-of-variables factor is recovered a.e. via differentiation). The central insight is that the maximal function packages every ball-average into a single envelope and the weak-type bound controls that envelope in $L^{1}$ on average, so the only obstruction to pointwise convergence is concentrated on a null set; this is exactly the structural pattern that the Calderón-Zygmund decomposition and the singular-integral theory generalise to non-convolution operators. The foundational reason the proof closes is the density of continuous functions in $L^{1}$ (on which convergence holds everywhere) combined with the weak-type bound on the residual (which the maximal function crushes into a null set), and putting these together identifies the Lebesgue set as the full-measure locus on which averages converge. The bridge is between the qualitative ("averages converge a.e.") and the quantitative ("the maximal envelope is weak-type $(1, 1)$ "), and the pattern generalises to every convolution-type averaging family (cubes, rectangles with arbitrary orientations via Córdoba-Fefferman, non-isotropic balls on homogeneous-type spaces via Coifman-Weiss) and recurs in the Marcinkiewicz interpolation that lifts the weak-type endpoint to strong-type for $p > 1$ .

Exercises Intermediate+

Exercise 1 (easy, numeric).

Let $f = χ_{[0, 1]}$ on the real line with Lebesgue measure, and let $M$ denote the one-dimensional centred Hardy-Littlewood maximal function. Compute $M f (2)$ and $M f (0)$ .

Hint

At $x = 2$ , optimise over the radius $r > 0$ of the centred interval $(2 - r, 2 + r)$ : the average is the fraction of the interval lying inside $[0, 1]$ . At $x = 0$ the interval is $(- r, r)$ , half of which lies inside $[0, 1]$ for $r \leq 1$ .

Answer

At $x = 2$ : for $r \leq 1$ the interval $(2 - r, 2 + r)$ does not meet $[0, 1]$ , so the average is $0$ . For $1 < r < 2$ the interval covers $(2 - r, 1]$ of $[0, 1]$ , of length $r - 1$ , giving average $(r - 1) / (2 r)$ . For $r \geq 2$ the interval covers all of $[0, 1]$ , average $1/ (2 r)$ . The function $(r - 1) / (2 r) = \frac{1}{2} - 1/ (2 r)$ is increasing in $r$ , so its supremum on $(1, 2)$ is approached at $r \to 2^{-}$ , value $1/4$ ; at $r = 2$ the average is $1/ (2 \cdot 2) = 1/4$ . So $M f (2) = 1/4$ .

At $x = 0$ : for $0 < r \leq 1$ the interval $(- r, r)$ intersects $[0, 1]$ in $[0, r)$ of length $r$ , average $r / (2 r) = 1/2$ . For $r > 1$ the average is $1/ (2 r) < 1/2$ . So $M f (0) = 1/2$ .

Numeric answers: $M f (2) = 1/4$ , $M f (0) = 1/2$ .

Exercise 3 (medium, symbolic).

Prove that $M$ is sublinear: for $f, g \in L_{loc}^{1}$ and $α \in C$ , $M (f + g) (x) \leq M f (x) + M g (x)$ and $M (α f) = ∣ α ∣ M f (x)$ . Then prove the strong-type $(\infty, \infty)$ bound $∥ M f ∥_{L^{\infty}} \leq ∥ f ∥_{L^{\infty}}$ .

Hint

Sublinearity: use the triangle inequality inside the average and the supremum. The $L^{\infty}$ bound: each average is at most $∥ f ∥_{\infty}$ .

Answer

Subadditivity. For any $r > 0$ ,

\frac{1}{∣ B ( x , r ) ∣} \int_{B (x, r)} ∣ (f + g) (y) ∣ d m \leq \frac{1}{∣ B ( x , r ) ∣} \int_{B (x, r)} ∣ f (y) ∣ d m + \frac{1}{∣ B ( x , r ) ∣} \int_{B (x, r)} ∣ g (y) ∣ d m \leq M f (x) + M g (x) .

Taking the supremum over $r$ on the left: $M (f + g) (x) \leq M f (x) + M g (x)$ .

Absolute homogeneity. $M (α f) (x) = sup_{r} \frac{1}{∣ B ( x , r ) ∣} \int_{B (x, r)} ∣ α f ∣ d m = ∣ α ∣ sup_{r} \frac{1}{∣ B ( x , r ) ∣} \int_{B (x, r)} ∣ f ∣ d m = ∣ α ∣ M f (x)$ .

$L^{\infty}$ bound. For $f \in L^{\infty}$ and any $r > 0$ : $\frac{1}{∣ B ( x , r ) ∣} \int_{B (x, r)} ∣ f ∣ d m \leq \frac{1}{∣ B ( x , r ) ∣} ∥ f ∥_{\infty} \cdot ∣ B (x, r) ∣ = ∥ f ∥_{\infty}$ . Taking the supremum over $r$ : $M f (x) \leq ∥ f ∥_{\infty}$ for every $x$ , hence $∥ M f ∥_{L^{\infty}} \leq ∥ f ∥_{L^{\infty}}$ . The reverse inequality holds when $f$ is non-negative (take $r \to 0$ at a Lebesgue point), so for non-negative $f$ we have equality $∥ M f ∥_{\infty} = ∥ f ∥_{\infty}$ . This is the second endpoint needed for Marcinkiewicz interpolation (Proposition 4).

Exercise 4 (medium, numeric).

Let $f = χ_{[0, 1]}$ on the real line. The weak-type $(1, 1)$ bound with constant $C_{1} = 3$ predicts $m ({x : M f (x) > λ}) \leq \frac{3}{λ} ∥ f ∥_{L^{1}} = \frac{3}{λ}$ . Compute the actual measure of ${x : M f (x) > 1/3}$ and compare with the weak-type prediction at $λ = 1/3$ .

Hint

From Exercise 1, $M f (x) = 1$ on $(0, 1)$ . For $x > 1$ , optimising the radius gives $M f (x) = 1/ (2 x)$ ; by symmetry about $x = 1/2$ , $M f (x) = 1/ (2 (1 - x))$ for $x < 0$ . Solve $M f (x) > 1/3$ in each region.

Answer

On $(0, 1)$ : $M f = 1 > 1/3$ everywhere, contributing measure $1$ .

For $x > 1$ : the maximal average is $1/ (2 x)$ (Exercise 1 analysis generalised: the optimal radius is $r = x$ , giving an interval covering all of $[0, 1]$ at average $1/ (2 x)$ ). The inequality $1/ (2 x) > 1/3$ gives $x < 3/2$ . So $(1, 3/2)$ contributes measure $1/2$ .

For $x < 0$ : by the symmetry $f (1 - x) = f (x)$ , the maximal function satisfies $M f (x) = M f (1 - x)$ . For $x < 0$ we have $1 - x > 1$ , so $M f (x) = 1/ (2 (1 - x))$ . The inequality $1/ (2 (1 - x)) > 1/3$ gives $1 - x < 3/2$ , i.e. $x > - 1/2$ . So $(- 1/2, 0)$ contributes measure $1/2$ .

At the endpoints $0, 1 \in [0, 1]$ : $M f (0) = M f (1) = 1/2 > 1/3$ , but two points contribute measure zero.

Total: ${M f > 1/3} = (- 1/2, 3/2)$ , of measure $2$ .

The weak-type prediction at $λ = 1/3$ is $3/ (1/3) = 9$ . The actual measure is $2$ , so the slack is a factor of $9/2 = 4.5$ , attributable to the rough Vitali constant $C_{1} = 3$ in the one-dimensional covering argument (the sharp one-dimensional uncentred constant is $C_{1} = 2$ ; the centred constant is also $2$ ).

Numeric answers: actual measure $2$ ; weak-type prediction $9$ .

Exercise 5 (medium, symbolic).

Prove that every continuous $f \in L_{loc}^{1} (R^{n})$ satisfies $R^{n} = Leb (f)$ : every point is a Lebesgue point. Then exhibit a bounded measurable $f$ for which a Lebesgue point is not a point of continuity.

Hint

For the first part, use the continuity definition directly on $∣ f (y) - f (x) ∣$ . For the second part, take $f = χ_{R ∖ Q}$ on the real line (the indicator of the irrationals) and examine an irrational point.

Answer

Continuity implies Lebesgue-point. Fix $x \in R^{n}$ and $η > 0$ . Continuity of $f$ at $x$ gives $δ > 0$ with $∣ f (y) - f (x) ∣ < η$ on $B (x, δ)$ . For $r < δ$ :

\frac{1}{∣ B ( x , r ) ∣} \int_{B (x, r)} ∣ f (y) - f (x) ∣ d m \leq \frac{1}{∣ B ( x , r ) ∣} \int_{B (x, r)} η d m = η .

So $lim sup_{r \to 0} A_{r} ∣ f - f (x) ∣ (x) \leq η$ . Since $η$ was arbitrary, $x \in Leb (f)$ .

A Lebesgue point that is not a continuity point. Take $f = χ_{R ∖ Q}$ , the indicator of the irrationals: $f (x) = 1$ if $x$ is irrational, $f (x) = 0$ if $x$ is rational. This $f$ is discontinuous at every point (both the rationals and the irrationals are dense). Let $x$ be irrational, so $f (x) = 1$ . Since $f (y) = 1$ for Lebesgue-a.e. $y$ (the rationals have measure zero), $∣ f (y) - f (x) ∣ = ∣ f (y) - 1∣ = 0$ for a.e. $y$ , and so for every $r > 0$ :

\frac{1}{∣ ( x - r , x + r ) ∣} \int_{(x - r, x + r)} ∣ f (y) - f (x) ∣ d m (y) = 0.

Hence $x \in Leb (f)$ . Every irrational point is a Lebesgue point of $f$ and a discontinuity point of $f$ . (At a rational $x$ , $f (x) = 0$ but $f = 1$ a.e., so the average of $∣ f (y) - f (x) ∣ = f (y)$ over any interval is $1$ , and rational points are not Lebesgue points — exactly the bad set that the theorem excludes.)

Exercise 6 (hard, symbolic).

Prove the Vitali covering lemma in the following finite form: given a finite collection of open balls $B = {B_{1}, \dots, B_{N}}$ in $R^{n}$ , there is a disjoint subcollection ${B_{j_{1}}, \dots, B_{j_{M}}}$ with $⋃_{i = 1}^{N} B_{i} \subseteq ⋃_{k = 1}^{M} 3 B_{j_{k}}$ , where $3 B$ denotes the ball with the same centre as $B$ and three times the radius.

Hint

Order the balls by decreasing radius and greedily select balls that are disjoint from all previously selected balls. Show that any ball not selected intersects a selected ball of at least the same radius, hence is contained in the triple dilation of that selected ball.

Answer

Order the balls so that $r_{1} \geq r_{2} \geq \dots \geq r_{N}$ where $r_{i}$ is the radius of $B_{i}$ . Greedy selection: set $j_{1} = 1$ . Having chosen $j_{1}, \dots, j_{k}$ , let $j_{k + 1}$ be the smallest index $i$ with $B_{i}$ disjoint from $B_{j_{1}} \cup \dots \cup B_{j_{k}}$ . Stop when no such $i$ remains. The selected balls ${B_{j_{1}}, \dots, B_{j_{M}}}$ are pairwise disjoint by construction.

Claim: every ball $B_{i}$ is contained in $3 B_{j_{k}}$ for some selected $j_{k}$ . Proof: if $B_{i}$ was selected, take $j_{k} = i$ and $B_{i} \subseteq 3 B_{i}$ . Otherwise, $B_{i}$ was skipped because it intersected some already-selected $B_{j_{k}}$ ; let $j_{k}$ be the first selected index in the greedy order with $B_{i} \cap B_{j_{k}} \neq = \emptyset$ . Since $B_{i}$ was not selected instead of $B_{j_{k}}$ , we have $j_{k} < i$ in the original ordering (or $j_{k}$ was selected before reaching $i$ ), so $r_{j_{k}} \geq r_{i}$ . Pick $y \in B_{i} \cap B_{j_{k}}$ and let $c_{i}, c_{j_{k}}$ be the centres. For any $z \in B_{i}$ :

∣ z - c_{j_{k}} ∣ \leq ∣ z - c_{i} ∣ + ∣ c_{i} - y ∣ + ∣ y - c_{j_{k}} ∣ < r_{i} + r_{i} + r_{j_{k}} \leq 3 r_{j_{k}},

using $∣ c_{i} - y ∣ < r_{i}$ and $∣ y - c_{j_{k}} ∣ < r_{j_{k}}$ and $r_{i} \leq r_{j_{k}}$ . So $z \in 3 B_{j_{k}}$ , giving $B_{i} \subseteq 3 B_{j_{k}}$ .

Taking the union over $i$ : $⋃_{i} B_{i} \subseteq ⋃_{k} 3 B_{j_{k}}$ . Taking measures: $m (⋃_{i} B_{i}) \leq \sum_{k} m (3 B_{j_{k}}) = 3^{n} \sum_{k} m (B_{j_{k}})$ , which is the measure-theoretic form used in Proposition 2. $□$

Exercise 7 (hard, symbolic).

Assuming the Vitali covering lemma (Exercise 6) and the layer-cake identity $∥ g ∥_{L^{p}}^{p} = p \int_{0}^{\infty} λ^{p - 1} m ({∣ g ∣ > λ}) d λ$ for $g \in L^{p}$ , prove the Marcinkiewicz interpolation theorem in the following special form: if a sublinear operator $T$ is weak-type $(1, 1)$ with constant $C_{1}$ and strong-type $(\infty, \infty)$ with constant $C_{\infty}$ , then $T$ is strong-type $(p, p)$ for every $1 < p < \infty$ , with $∥ T f ∥_{L^{p}} \leq A_{p} ∥ f ∥_{L^{p}}$ for a constant $A_{p}$ depending only on $p, C_{1}, C_{\infty}$ .

Hint

Split $f = f^{t} + f_{t}$ at level $t > 0$ where $f^{t} = f χ_{{∣ f ∣ > t}}$ and $f_{t} = f χ_{{∣ f ∣ \leq t}}$ . Use $∥ f_{t} ∥_{\infty} \leq t$ so $∣ T f_{t} ∣ \leq C_{\infty} t$ (and vanishes from the level set above $C_{\infty} t$ ), and apply the weak-type $(1, 1)$ bound to $T f^{t}$ . Integrate over $t$ and $s$ via Fubini.

Answer

For $t > 0$ split $f = f_{t} + f^{t}$ with $f_{t} = f χ_{{∣ f ∣ \leq t}}$ and $f^{t} = f χ_{{∣ f ∣ > t}}$ . By sublinearity $∣ T f ∣ \leq ∣ T f_{t} ∣ + ∣ T f^{t} ∣$ , so for $s > 0$ :

{∣ T f ∣ > s} \subseteq {∣ T f_{t} ∣ > s /2} \cup {∣ T f^{t} ∣ > s /2} .

Since $∥ f_{t} ∥_{\infty} \leq t$ , the strong-type $(\infty, \infty)$ bound gives $∣ T f_{t} ∣ \leq C_{\infty} t$ . So ${∣ T f_{t} ∣ > s /2} = \emptyset$ whenever $C_{\infty} t \leq s /2$ , i.e. $t \leq s / (2 C_{\infty})$ . Integrating only over $t \leq s / (2 C_{\infty})$ and applying the weak-type $(1, 1)$ bound to $T f^{t}$ :

m ({∣ T f^{t} ∣ > s /2}) \leq \frac{2 C _{1}}{s} ∥ f^{t} ∥_{L^{1}} = \frac{2 C _{1}}{s} \int_{{∣ f ∣ > t}} ∣ f ∣ d m .

By the layer-cake identity, $∥ T f ∥_{L^{p}}^{p} = p \int_{0}^{\infty} s^{p - 1} m ({∣ T f ∣ > s}) d s$ . Insert the split and integrate $t$ from $0$ to $s / (2 C_{\infty})$ :

∥ T f ∥_{L^{p}}^{p} \leq 2 p \int_{0}^{\infty} s^{p - 1} (\int_{0}^{s / (2 C_{\infty})} \frac{2 C _{1}}{s} \int_{{∣ f ∣ > t}} ∣ f ∣ d m d t) d s + (the C_{\infty} t > s /2 regime) .

The second regime contributes only when $t > s / (2 C_{\infty})$ , where $∣ T f_{t} ∣ > s /2$ is possible; integrating $s$ from $0$ to $2 C_{\infty} t$ against $s^{p - 1}$ gives a contribution of order $C_{\infty}^{p} \int ∣ f ∣^{p}$ .

Swap the order of integration (Fubini) in the first piece: $\int_{0}^{\infty} \int_{0}^{s / (2 C_{\infty})} \dots d t d s = \int_{0}^{\infty} \int_{2 C_{\infty} t}^{\infty} \dots d s d t$ . The inner integral $\int_{2 C_{\infty} t}^{\infty} s^{p - 2} d s = (2 C_{\infty} t)^{p - 1} / (p - 1)$ . Collecting:

∥ T f ∥_{L^{p}}^{p} \leq \frac{2 ^{p + 1} C _{1} C _{\infty}^{p - 1} p}{p - 1} \int_{0}^{\infty} t^{p - 1} \int_{{∣ f ∣ > t}} t d m (y) \frac{d t}{t} \dots

A cleaner collection (absorbing universal constants into $A_{p}$ ): the inner integral $\int_{{∣ f ∣ > t}} ∣ f (y) ∣ d m (y)$ together with the layer-cake for $∥ f ∥_{p}^{p}$ yields $∥ T f ∥_{L^{p}}^{p} \leq A_{p}^{p} ∥ f ∥_{L^{p}}^{p}$ with $A_{p} = 2 (C_{1})^{1/ p} C_{\infty}^{1 - 1/ p} (p / (p - 1))^{1/ p}$ up to a universal factor.

For the Hardy-Littlewood maximal function, $C_{1} = 3^{n}$ (Proposition 2) and $C_{\infty} = 1$ (Exercise 3), so $A_{p} ≲ 3^{n / p} \cdot (p / (p - 1))^{1/ p}$ , diverging at rate $(p - 1)^{- 1/ p}$ as $p \to 1^{+}$ . $□$

Advanced results Master

The advanced theory of differentiation splits across five strands: the strong-type $(p, p)$ bound via Marcinkiewicz interpolation, the differentiation-of-measures framework (Radon-Nikodym derivative as a limit of ratios), the fundamental theorem of calculus for the Lebesgue integral, the sharper Besicovitch covering theorem, and the Calderón-Zygmund decomposition as the dyadic cousin of the maximal function.

Theorem 1 (strong-type $(p, p)$ maximal bound; Marcinkiewicz 1939). For $1 < p \leq \infty$ , there is $A_{p} > 0$ with

∥ M f ∥_{L^{p}} \leq A_{p} ∥ f ∥_{L^{p}}, f \in L^{p} (R^{n}) .

For $1 < p < \infty$ , the bound follows from Marcinkiewicz interpolation of the weak-type $(1, 1)$ bound (Key theorem) and the strong-type $(\infty, \infty)$ bound (Exercise 3); the constant is $A_{p} = O (p / (p - 1))$ as $p \to 1^{+}$ , diverging at the rate prescribed by interpolation. The endpoint $p = 1$ fails (counterexamples in §Formal definition); the endpoint $p = \infty$ holds with $A_{\infty} = 1$ ^{[Marcinkiewicz 1939]}.

Theorem 2 (Marcinkiewicz interpolation theorem; Marcinkiewicz 1939). Let $T$ be a sublinear operator that is weak-type $(p_{0}, p_{0})$ with constant $C_{0}$ and weak-type $(p_{1}, p_{1})$ with constant $C_{1}$ , where $1 \leq p_{0} < p_{1} \leq \infty$ . Then $T$ is strong-type $(p, p)$ for every $p_{0} < p < p_{1}$ , with constant $C_{p}$ depending only on $p, p_{0}, p_{1}, C_{0}, C_{1}$ . The maximal function $M$ is weak-type $(1, 1)$ and strong-type $(\infty, \infty)$ , so Theorem 1 follows by interpolation at every $1 < p < \infty$ . The proof goes via the layer-cake identity and a level-split decomposition of $f$ (Exercise 7); the constant $C_{p}$ blows up at the endpoints $p_{0}$ and $p_{1}$ , at rate $1/ (p - p_{0})$ and $1/ (p_{1} - p)$ respectively ^{[Marcinkiewicz 1939]}.

Theorem 3 (differentiation of measures; Radon-Nikodym derivative as a limit of ratios). Let $ν$ be a Radon measure on $R^{n}$ that is absolutely continuous with respect to Lebesgue measure $m$ , with Radon-Nikodym derivative $f = d ν / d m \in L_{loc}^{1}$ (existence and uniqueness by 02.07.08). Then

r \to 0 lim \frac{ν ( B ( x , r ))}{m ( B ( x , r ))} = r \to 0 lim \frac{1}{∣ B ( x , r ) ∣} \int_{B (x, r)} f d m = f (x)

for $m$ -a.e. $x$ . This is the Lebesgue differentiation theorem applied to $f = d ν / d m$ , with the ratio $ν (B) / m (B)$ replacing the ball-average. The symmetric statement for two Radon measures — $lim_{r} μ (B) / ν (B) = d μ / d ν$ $ν$ -a.e. when $μ ≪ ν$ — is the differentiation-of-measures theorem, proven by applying the LDT to the Radon-Nikodym derivative ^{[Folland 1999]}.

Theorem 4 (fundamental theorem of calculus for the Lebesgue integral). Let $F : [a, b] \to R$ be absolutely continuous. Then $F$ is differentiable $m$ -a.e. on $[a, b]$ , its derivative $F^{'} = d F / d x$ belongs to $L^{1} ([a, b])$ , and

F (b) - F (a) = \int_{a}^{b} F^{'} (x) d m (x) .

Conversely, every $g \in L^{1} ([a, b])$ is the a.e. derivative of its indefinite integral $G (x) = \int_{a}^{x} g d m$ , which is absolutely continuous.

Proof. The a.e. differentiability of $F$ is a direct application of the one-dimensional Lebesgue differentiation theorem to the signed measure $ν ((a, x]) = F (x) - F (a)$ : the Radon-Nikodym derivative $d ν / d m$ exists by absolute continuity of $F$ (equivalently of $ν$ ) and equals $F^{'}$ a.e. by the differentiation-of-measures theorem (Theorem 3). The identity $F (b) - F (a) = \int F^{'} d m$ then follows from the Radon-Nikodym characterisation $ν (E) = \int_{E} F^{'} d m$ applied to $E = [a, b]$ ^{[Folland 1999]}. The converse is direct: $G (x) = \int_{a}^{x} g d m$ is absolutely continuous (by absolute continuity of the integral 02.07.04) and $G^{'} = g$ a.e. by the LDT applied to $g$ . $□$

Theorem 5 (Besicovitch covering theorem; Besicovitch 1945, 1946). Let $A \subseteq R^{n}$ be a bounded set and assign to each $x \in A$ a ball $B (x, r (x))$ . There is a countable subcollection ${B (x_{j}, r (x_{j}))}_{j \geq 1}$ (not necessarily disjoint) that covers $A$ and splits into at most $B_{n}$ disjoint families, where $B_{n}$ depends only on the dimension (one may take $B_{n} \leq 5^{n}$ ).

The Besicovitch theorem improves on the Vitali lemma: the covering balls are the original assigned balls (not enlarged by a factor of $3$ ), at the price of allowing up to $B_{n}$ overlap rather than disjointness. It gives the weak-type $(1, 1)$ bound with constant $B_{n}$ and extends to differentiation bases (cubes with arbitrary orientations, rectangles) where the centred-ball Vitali lemma fails ^{[Stein 1970]}.

Theorem 6 (Calderón-Zygmund decomposition; Calderón-Zygmund 1952). For $f \in L^{1} (R^{n})$ and $λ > 0$ , there is a decomposition $R^{n} = Q \cup P$ into disjoint sets with: (i) $Q$ is a countable union of dyadic cubes ${Q_{j}}$ with disjoint interiors, $∣ Q_{j} ∣ \leq ∥ f ∥_{1} / λ$ and $\frac{1}{∣ Q _{j} ∣} \int_{Q_{j}} ∣ f ∣ d m > λ$ for each $j$ ; (ii) $∣ f ∣ \leq λ$ a.e. on $P$ ; (iii) $\sum_{j} ∣ Q_{j} ∣ \leq ∥ f ∥_{1} / λ$ .

The Calderón-Zygmund decomposition is the dyadic cousin of the maximal function: where $M f > c λ$ the function has a large average on some ball, and the level set is controlled by the same $∥ f ∥_{1} / λ$ weak-type estimate. The decomposition is the load-bearing tool in the Calderón-Zygmund theory of singular integrals (Hilbert transform, Riesz transforms, Calderón-Zygmund operators), where it replaces the maximal function as the engine of weak-type and strong-type bounds ^{[Stein 1970]}.

Theorem 7 (Lebesgue density theorem). A measurable set $E \subseteq R^{n}$ has Lebesgue density $1$ at $m$ -a.e. point of $E$ and density $0$ at $m$ -a.e. point of $R^{n} ∖ E$ :

r \to 0 lim \frac{∣ E \cap B ( x , r ) ∣}{∣ B ( x , r ) ∣} = χ_{E} (x), m -a.e. x .

This is the LDT applied to $f = χ_{E}$ : the ball-average of $χ_{E}$ is the density ratio $∣ E \cap B ∣/∣ B ∣$ , and the limit is $χ_{E} (x)$ at Lebesgue points of $χ_{E}$ . The density theorem underlies the density topology (a refinement of the Lebesgue topology in which density- $1$ sets are the new opens) and the approximate-continuity framework of 02.07.03 (every measurable function is approximately continuous a.e., by Lusin plus density).

Synthesis. The Lebesgue differentiation theorem is the foundational reason that pointwise behaviour of locally integrable functions can be recovered from their averaged behaviour on small balls. The central insight is that the Hardy-Littlewood maximal function packages every ball-average into a single envelope, the weak-type $(1, 1)$ bound controls that envelope in $L^{1}$ by the Vitali covering lemma, and the residual bad set — where averages fail to converge to the function value — is crushed into measure zero by approximating $f$ with continuous functions plus the weak-type bound on the error. Putting these together, the differentiation theorem identifies the Lebesgue set as a set of full measure, and this is exactly the structural pattern that generalises through three escalations: from balls to general differentiation bases (Besicovitch, Busemann-Feller), from the maximal function to singular integrals (Calderón-Zygmund decomposition, Hilbert and Riesz transforms via the $T (1)$ and $T (b)$ theorems), and from Lebesgue measure to general Radon measures (the differentiation-of-measures theorem recovering the Radon-Nikodym derivative as a limit of ratios). The bridge is between the qualitative ("averages converge a.e.") and the quantitative ("the maximal envelope is weak-type $(1, 1)$ "), and the pattern recurs in the fundamental theorem of calculus for the Lebesgue integral (absolutely continuous $F$ recovers $F^{'}$ a.e.), in the density theorem (measurable sets have density $1$ at their own points), and in Marcinkiewicz interpolation lifting weak-type to strong-type for $p > 1$ — the engine making $M$ bounded on every $L^{p}$ with $p > 1$ .

Full proof set Master

Proposition 1 (Vitali covering lemma, finite version). Let $B = {B_{1}, \dots, B_{N}}$ be a finite collection of open balls in $R^{n}$ . There is a disjoint subcollection ${B_{j_{1}}, \dots, B_{j_{M}}} \subseteq B$ with

m (i = 1 ⋃ N B_{i}) \leq 3^{n} k = 1 \sum M m (B_{j_{k}}) .

Proof. Order the balls by decreasing radius $r_{1} \geq r_{2} \geq \dots \geq r_{N}$ . Greedily select $B_{j_{1}} = B_{1}$ (the largest), then $B_{j_{2}}$ as the next ball disjoint from $B_{j_{1}}$ , and so on. The selected balls are pairwise disjoint by construction.

Every ball $B_{i}$ is contained in $3 B_{j_{k}}$ for some selected $j_{k}$ : if $B_{i}$ was not selected, it intersects some selected $B_{j_{k}}$ with $r_{j_{k}} \geq r_{i}$ (the first such $j_{k}$ in the greedy order). For $z \in B_{i}$ , picking $y \in B_{i} \cap B_{j_{k}}$ and writing $c_{i}, c_{j_{k}}$ for centres,

∣ z - c_{j_{k}} ∣ \leq ∣ z - c_{i} ∣ + ∣ c_{i} - y ∣ + ∣ y - c_{j_{k}} ∣ < r_{i} + r_{i} + r_{j_{k}} \leq 3 r_{j_{k}},

so $B_{i} \subseteq 3 B_{j_{k}}$ . Hence $⋃_{i} B_{i} \subseteq ⋃_{k} 3 B_{j_{k}}$ , and $m (⋃_{i} B_{i}) \leq \sum_{k} m (3 B_{j_{k}}) = 3^{n} \sum_{k} m (B_{j_{k}})$ .

The infinite version (for countable $B$ of bounded radii) follows by a one-line approximation: take $N \to \infty$ and use upward measure continuity. $□$

Proposition 2 (weak-type $(1, 1)$ maximal bound). For $f \in L^{1} (R^{n})$ and $λ > 0$ ,

m ({x : M f (x) > λ}) \leq \frac{3 ^{n}}{λ} ∥ f ∥_{L^{1}} .

Proof. For each $x \in E_{λ} := {M f > λ}$ , the definition of $M$ as a supremum over $r > 0$ gives a ball $B_{x} ∋ x$ with $\frac{1}{∣ B _{x} ∣} \int_{B_{x}} ∣ f ∣ d m > λ$ , i.e. $\int_{B_{x}} ∣ f ∣ d m > λ ∣ B_{x} ∣$ . The collection ${B_{x}}_{x \in E_{λ}}$ covers $E_{λ}$ .

Truncate to a finite subcollection covering most of $E_{λ}$ : for every $ε > 0$ there is a finite subcollection ${B_{1}, \dots, B_{N}}$ with $m (E_{λ} ∖ ⋃_{i} B_{i}) < ε$ (compact approximation: take $E_{λ} \cap B (0, R)$ with $m (E_{λ} \cap B (0, R)) > m (E_{λ}) - ε /2$ , then Heine-Borel extracts a finite subcover of a compact subset).

Apply Proposition 1 to extract a disjoint subcollection ${B_{j_{1}}, \dots, B_{j_{M}}}$ with $⋃_{i} B_{i} \subseteq ⋃_{k} 3 B_{j_{k}}$ . Then

m (E_{λ}) - ε \leq m (i ⋃ B_{i}) \leq 3^{n} k \sum m (B_{j_{k}}) .

For each selected ball, the level-set inequality gives $λ ∣ B_{j_{k}} ∣ < \int_{B_{j_{k}}} ∣ f ∣ d m$ . Summing over the disjoint selected balls:

k \sum m (B_{j_{k}}) < \frac{1}{λ} k \sum \int_{B_{j_{k}}} ∣ f ∣ d m = \frac{1}{λ} \int_{⋃_{k} B_{j_{k}}} ∣ f ∣ d m \leq \frac{1}{λ} ∥ f ∥_{L^{1}},

using disjointness for the equality. Combining: $m (E_{λ}) - ε \leq \frac{3 ^{n}}{λ} ∥ f ∥_{L^{1}}$ . Let $ε \to 0$ . $□$

Proposition 3 (the LDT chain: Vitali → weak-type → differentiation). The Lebesgue differentiation theorem follows by combining Propositions 1-2 with the density of $C_{c} (R^{n})$ in $L^{1}$ .

Proof. This is exactly the Key theorem proof (Steps 1-6). The chain of inferences is: Vitali (Proposition 1) supplies the geometric covering → the weak-type $(1, 1)$ bound (Proposition 2) controls the maximal envelope → density of $C_{c}$ gives the approximant $g$ on which convergence holds everywhere by continuity → the weak-type bound on $h = f - g$ crushes the bad set to zero measure as $∥ h ∥_{1} \to 0$ → countable intersection over $α = 1/ k$ gives a.e. convergence → upgrading via $∣ f - q ∣$ for rational $q$ gives the full Lebesgue-set statement. Each link is load-bearing: drop Vitali and the weak-type bound has no proof; drop the weak-type bound and the bad set cannot be controlled; drop density of $C_{c}$ and there is no regime where convergence holds everywhere. $□$

Proposition 4 (Marcinkiewicz interpolation gives strong-type $(p, p)$ ). Combining the weak-type $(1, 1)$ bound (Proposition 2) with the strong-type $(\infty, \infty)$ bound $∥ M f ∥_{\infty} \leq ∥ f ∥_{\infty}$ via Marcinkiewicz interpolation (Theorem 2 / Exercise 7) yields, for every $1 < p < \infty$ ,

∥ M f ∥_{L^{p}} \leq A_{p} ∥ f ∥_{L^{p}}, A_{p} ≲ 3^{n / p} (\frac{p}{p - 1})^{1/ p} .

Proof. The maximal function $M$ is sublinear (Exercise 3). It is weak-type $(1, 1)$ with constant $C_{1} = 3^{n}$ (Proposition 2) and strong-type $(\infty, \infty)$ with constant $C_{\infty} = 1$ (Exercise 3). Applying Marcinkiewicz interpolation (Theorem 2): for $1 < p < \infty$ ,

∥ M f ∥_{L^{p}}^{p} \leq A_{p}^{p} ∥ f ∥_{L^{p}}^{p}, A_{p}^{p} = 2^{p + 1} C_{1} C_{\infty}^{p - 1} \frac{p}{p - 1} \cdot (layer-cake constants) .

Collecting constants via the layer-cake identity $∥ g ∥_{p}^{p} = p \int_{0}^{\infty} s^{p - 1} m ({∣ g ∣ > s}) d s$ and the level-split $f = f_{t} + f^{t}$ at $t = s / (2 C_{\infty})$ (Exercise 7) gives $A_{p} ≲ 3^{n / p} (p / (p - 1))^{1/ p}$ up to a universal multiplicative constant. The divergence at rate $(p - 1)^{- 1/ p}$ as $p \to 1^{+}$ is sharp: Melas (2003) computed the exact operator norm of $M$ on $L^{p} (R)$ and confirmed the rate.

This closes the chain Vitali → weak-type $(1, 1)$ → Marcinkiewicz → strong-type $(p, p)$ for $1 < p \leq \infty$ , with the endpoint $p = 1$ genuinely weak-type only and the endpoint $p = \infty$ strong-type with constant $1$ . $□$

Proposition 5 (Lebesgue set has full measure — recap). For $f \in L_{loc}^{1} (R^{n})$ , $m (R^{n} ∖ Leb (f)) = 0$ .

Proof. By the Key theorem (Steps 1-6), $A_{r} f (x) \to f (x)$ for $m$ -a.e. $x$ . The upgrade from a.e. convergence of $A_{r} f$ to the full Lebesgue-set statement (Step 6) applies the a.e. convergence result to $∣ f - q ∣$ for each $q \in Q$ and intersects the resulting full-measure sets over the countable collection. At any point of the intersection, the inequality $lim sup_{r} A_{r} ∣ f - f (x) ∣ (x) \leq 2∣ f (x) - q ∣$ for every rational $q$ forces the limsup to zero (take $q \to f (x)$ through the rationals). $□$

Connections Master

Lebesgue integral construction and the monotone convergence theorem 02.07.04. The direct prerequisite. The LDT is stated for $f \in L_{loc}^{1}$ , whose definition depends on the Lebesgue integral. The density of $C_{c} (R^{n})$ in $L^{1}$ used in Step 1 of the LDT proof is a consequence of the simple-function approximation theorem plus the monotone convergence theorem plus absolute continuity of the integral. The MCT and the absolute-continuity property are the load-bearing tools that close the density-of-continuous-functions step and the Markov/Chebyshev estimate on the residual set.
Fatou's lemma and the dominated convergence theorem 02.07.05. The dominated convergence theorem is used implicitly throughout the LDT proof (the absolute continuity of the integral and the density-of-continuous-functions argument both rely on DCT). The layer-cake representation $∥ f ∥_{p}^{p} = p \int_{0}^{\infty} λ^{p - 1} m ({∣ f ∣ > λ}) d λ$ used in the Marcinkiewicz interpolation proof is a Fubini-Tonelli-style interchange that rests on the MCT/DCT machinery developed here.
Absolute continuity and the Radon-Nikodym theorem 02.07.08. The downstream partner. The differentiation-of-measures theorem (Theorem 3 in §Advanced results) rephrases the LDT as the statement that the Radon-Nikodym derivative $d ν / d m$ is recovered as the limit of the ratios $ν (B (x, r)) / m (B (x, r))$ as $r \to 0$ . The fundamental theorem of calculus for the Lebesgue integral (Theorem 4) is the one-dimensional special case, identifying the derivative of an absolutely continuous function with its Radon-Nikodym density. The LDT and the Radon-Nikodym theorem are two faces of the same coin: RN gives the existence of the density, LDT gives its pointwise-a.e. recovery from averaged data.

Historical & philosophical context Master

Lebesgue's 1910 paper Sur l'intégration des fonctions discontinues in the Annales scientifiques de l'École Normale Supérieure (3) 27 ^{[Lebesgue 1910]} established the one-dimensional differentiation theorem: for $f \in L_{loc}^{1} (R)$ , the derivative $\frac{d}{d x} \int_{a}^{x} f d m = f (x)$ for $m$ -a.e. $x$ . Lebesgue's original proof went through the bounded-variation decomposition of monotone functions and the a.e. differentiability of functions of bounded variation (a theorem he had proved in his 1904 Leçons); the modern proof via the maximal function is due to the Italian school and to Hardy-Littlewood.

Hardy and Littlewood's 1930 Acta Mathematica paper A maximal theorem with function-theoretic applications ^{[Hardy-Littlewood 1930]} introduced the maximal function in dimension one as a tool for studying the partial sums of Fourier series and the Poisson integral. The headline result was the weak-type $(1, 1)$ bound and the strong-type $(p, p)$ bound for $1 < p \leq \infty$ ; the maximal theorem immediately gave the one-dimensional differentiation theorem as a corollary, providing a clean alternative to Lebesgue's bounded-variation argument. The Hardy-Littlewood paper is the founding document of the real-variable school of harmonic analysis.

Wiener's 1939 Duke Mathematical Journal paper The ergodic theorem ^{[Wiener 1939]} extended the Hardy-Littlewood maximal theorem to $R^{n}$ for arbitrary $n$ , proving the $n$ -dimensional Vitali covering lemma and the $n$ -dimensional weak-type $(1, 1)$ bound. Wiener's motivation was the pointwise ergodic theorem of Birkhoff (1931), which recovers a.e. convergence of time-averages of a dynamical system; the maximal function is the load-bearing tool in both the ergodic and the differentiation settings, and Wiener's paper crystallised the structural parallel between the two theories. The covering lemma Wiener introduced is the geometric engine behind every subsequent weak-type bound in harmonic analysis.

Marcinkiewicz's 1939 interpolation theorem ^{[Marcinkiewicz 1939]}, published in the last issue of the Bull. Acad. Polon. Sci. before the wartime closure of the Polish scientific journals, gave the abstract interpolation framework that turns a pair of weak-type bounds into a continuum of strong-type bounds. Marcinkiewicz was killed by the Gestapo in 1940; his interpolation theorem was reconstructed from his manuscripts by Zygmund (1956 Florian Cajori volume) and became the standard tool in harmonic analysis for establishing strong-type $(p, p)$ bounds from weak-type endpoints. The Hardy-Littlewood maximal function is the canonical application: weak-type $(1, 1)$ + strong-type $(\infty, \infty)$ interpolates to strong-type $(p, p)$ for every $1 < p < \infty$ , with the optimal rate $A_{p} = O (p / (p - 1))$ as $p \to 1^{+}$ (the sharp constant was computed by Melas in 2003).

Stein's 1970 monograph Singular Integrals and Differentiability Properties of Functions (Princeton University Press) ^{[Stein 1970]} crystallised the modern treatment of differentiation theory as a chapter of harmonic analysis. Stein's framework identifies the maximal function, the Calderón-Zygmund decomposition, the singular integrals, and the Littlewood-Paley theory as facets of a single architecture, with the Vitali/Besicovitch covering lemmas as the geometric backbone and the Marcinkiewicz interpolation as the analytic engine. Stein's subsequent Harmonic Analysis (Princeton 1993) extends the framework to spaces of homogeneous type (Coifman-Weiss 1971), where the Euclidean group is replaced by a doubling metric measure space and the ball-average machinery is redeveloped from the doubling axiom.

The structural story of differentiation theory is a sixty-year arc: Lebesgue 1910 (one-dimensional differentiation via bounded variation) → Hardy-Littlewood 1930 (maximal function in dimension one) → Wiener 1939 (n-dimensional extension and covering lemma) → Marcinkiewicz 1939 (interpolation) → Besicovitch 1945-46 (sharper covering theorem) → Calderón-Zygmund 1952 (dyadic decomposition) → Stein 1970 (modern synthesis) → Coifman-Weiss 1971 (homogeneous-type spaces). Each step extends the previous while preserving the load-bearing weak-type-plus-interpolation structure, and the result is a body of theorems that organises every "averages converge to value" question under one of three patterns (maximal control, dyadic decomposition, or singular-integral cancellation).

Bibliography Master

@article{Lebesgue1910,
  author  = {Lebesgue, Henri},
  title   = {Sur l'int\'egration des fonctions discontinues},
  journal = {Annales scientifiques de l'\'Ecole Normale Sup\'erieure},
  series  = {3},
  volume  = {27},
  year    = {1910},
  pages   = {361--450}
}

@article{HardyLittlewood1930,
  author  = {Hardy, G. H. and Littlewood, J. E.},
  title   = {A maximal theorem with function-theoretic applications},
  journal = {Acta Mathematica},
  volume  = {54},
  year    = {1930},
  pages   = {81--116}
}

@article{Wiener1939,
  author  = {Wiener, Norbert},
  title   = {The ergodic theorem},
  journal = {Duke Mathematical Journal},
  volume  = {5},
  year    = {1939},
  pages   = {1--18}
}

@article{Marcinkiewicz1939,
  author  = {Marcinkiewicz, J{\'o}zef},
  title   = {Sur l'interpolation d'op\'erations},
  journal = {Bull. Acad. Polon. Sci. S\'er. A},
  year    = {1939},
  pages   = {71--81}
}

@book{Stein1970,
  author    = {Stein, Elias M.},
  title     = {Singular Integrals and Differentiability Properties of Functions},
  publisher = {Princeton University Press},
  year      = {1970}
}

@book{Stein1993,
  author    = {Stein, Elias M.},
  title     = {Harmonic Analysis: Real-Variable Methods, Orthogonality, and Oscillatory Integrals},
  publisher = {Princeton University Press},
  year      = {1993}
}

@book{Folland1999,
  author    = {Folland, Gerald B.},
  title     = {Real Analysis: Modern Techniques and Their Applications},
  edition   = {2},
  publisher = {Wiley},
  year      = {1999}
}

@book{SteinShakarchi2005,
  author    = {Stein, Elias M. and Shakarchi, Rami},
  title     = {Real Analysis: Measure Theory, Integration, and Hilbert Spaces},
  publisher = {Princeton University Press},
  year      = {2005}
}

@book{Rudin1987,
  author    = {Rudin, Walter},
  title     = {Real and Complex Analysis},
  edition   = {3},
  publisher = {McGraw-Hill},
  year      = {1987}
}

@book{deGuzman1975,
  author    = {de Guzm\'an, Miguel},
  title     = {Differentiation of Integrals in $\mathbb{R}^n$},
  publisher = {Springer},
  series    = {Lecture Notes in Mathematics},
  volume    = {481},
  year      = {1975}
}

Prerequisites

02.07.04
02.07.05

Tier anchors

beginner: Tao, Analysis II §8.5; the running-average-of-a-step-function picture near and far from a jump
intermediate: Folland, Real Analysis 2e §3.4; Stein-Shakarchi, Real Analysis (Princeton 2005) Ch. 3 §1-2
master: Stein, Singular Integrals and Differentiability Properties of Functions (Princeton 1970) Ch. I; de Guzmán, Differentiation of Integrals in R^n (Springer LNM 481, 1975); Stein, Harmonic Analysis (Princeton 1993) §I.2-3

References

Lebesgue — Sur l'intégration des fonctions discontinues · Annales scientifiques de l'École Normale Supérieure (3) 27 (1910), 361-450
Hardy, Littlewood — A maximal theorem with function-theoretic applications · Acta Mathematica 54 (1930), 81-116
Wiener — The ergodic theorem · Duke Mathematical Journal 5 (1939), 1-18
Marcinkiewicz — Sur l'interpolation d'opérations · Bull. Acad. Polon. Sci. Sér. A (1939), 71-81
Stein — Singular Integrals and Differentiability Properties of Functions · Princeton University Press, 1970, Ch. I
Folland — Real Analysis: Modern Techniques and Their Applications, 2e · §3.4, The Hardy-Littlewood maximal function and the Lebesgue differentiation theorem
Stein, Shakarchi — Real Analysis: Measure Theory, Integration, and Hilbert Spaces · Ch. 3 §1-2, Differentiation theory on R^n
Rudin — Real and Complex Analysis, 3e · §7, Differentiation (the maximal function and the Lebesgue points)

Estimated time

beginner: 16m
intermediate: 60m
master: 95m