Integration averages a function; differentiation asks whether shrinking averages recover it. For a continuous function the answer is immediate, but an integrable function can be discontinuous on a large set and can be changed on a null set without changing any integral. Thus the appropriate conclusion is recovery almost everywhere, not at every chosen point. We work with Lebesgue measure on the real line and write for the measure of a measurable set.
Averages and a covering estimate
Definition 2.1.1
(Centered maximal function)
For the centered maximal function is
The supremum can be restricted to positive rational , because the displayed average is continuous in ; each fixed-radius average is also continuous in . Hence is measurable. Large values of indicate that some interval centered at contains unusually much of relative to its length.
Lemma 2.1.2
(Weak maximal estimate)
For every and ,
Proof
(Greedy selection of disjoint intervals)
Let . For every choose a centered open interval satisfying . The set is open: the average over a fixed radius varies continuously with its center. If is empty there is nothing to prove. Otherwise take any nonempty compact and extract a finite subcover from these intervals. From that finite collection choose an interval of greatest length, discard every interval meeting it, and repeat among the remaining intervals. The selected intervals are pairwise disjoint. Every discarded interval meets a selected interval at least as long as itself, so it lies in the interval with the same center as the selected interval and three times its length. Consequently
Inner regularity of Lebesgue measure for the open set lets compact subsets exhaust its measure, even when is unbounded. This proves the estimate. The constant is sufficient for the argument; no optimality is claimed.
The differentiation theorem
Fix a measurable representative of that is finite everywhere, assigning arbitrary finite values on any exceptional null set.
Definition 2.1.3
(Lebesgue point)
A point is a Lebesgue point of a locally integrable function if
Every point of continuity is a Lebesgue point, but continuity is not required. Changing the representative of on a null set can change which individual points satisfy this definition.
Common mistake
(Testing only the signed average)
Convergence of the signed average to is weaker than being a Lebesgue point, because cancellation can hide the error. For with , the signed centered average at is for every , yet for every . So is not a Lebesgue point.
Theorem 2.1.4
(One-dimensional Lebesgue differentiation)
If , almost every point is a Lebesgue point of . In particular,
Proof
(Continuous approximation plus the maximal estimate)
First suppose . We use the standard density fact that continuous compactly supported functions are dense in . It follows from Lebesgue regularity and simple-function approximation: truncate an integrable function in size and space, approximate its finitely many level sets in measure by compact subsets and open supersets, and use continuous compactly supported cutoffs between them.
Choose such a continuous and put . Since
and the local oscillation of tends to zero at every point, the limsup of the absolute averages for at is at most . For , the set where lies in , so the weak maximal estimate and Markov's inequality give
The right side can be made arbitrarily small by choosing close to in . Hence the set where the limsup exceeds is null. Take the union over positive rational to conclude that the limsup is zero almost everywhere.
For , apply the result to for each positive integer . At every point strictly inside , sufficiently small averaging intervals see the original function. The countable union of the exceptional null sets is null, and the intervals cover the line.
Proof X-ray
Why the maximal function is the hinge
The statement is easy for the continuous , at every point. The difficulty is the error , which is small only in and can be large at individual points. The proof never controls the limsup of averages of pointwise; it replaces it by the larger quantity , a supremum over all radii, and the weak estimate bounds the measure of the set where that supremum is large by a multiple of . That is the one non-obvious move: a bound uniform in the radius is what survives the limit . The argument closes because the exceptional set for each is contained in sets of arbitrarily small measure, hence is null, and only countably many are needed. The covering selection is where one dimension and centered intervals enter, through the factor .
Worked example 2.1.5
(An indicator function at its jumps)
For , every point except and is a Lebesgue point for the displayed representative: points of and of are points of continuity. At and , the centered average is , while ; the absolute average also equals . The theorem allows this exceptional point; it does not promise to recover values at jumps. Redefining does not help either: the absolute average is then again.
Averages over balls still recover the function almost everywhere, with a dimensional constant in the covering estimate.
A path to generalization
What changes in higher dimensions?
Averages over balls still recover the function almost everywhere, with a dimensional constant in the covering estimate.
In , define with balls . The same greedy selection works, since enlarging a ball three times multiplies its measure by ; it gives , and the proof of the differentiation theorem carries over unchanged. The shape of the averaging sets matters: for some families of shrinking sets that are not comparable to balls, such as thin rectangles in arbitrary directions, the conclusion can fail. Which families work is a subtler question beyond this course.
Indefinite Lebesgue integrals
Let and define
Theorem 2.1.6
(Fundamental theorem, integral-to-derivative direction)
is absolutely continuous on and for almost every .
Proof
(Lebesgue points give the derivative; truncation gives absolute continuity)
Extend by zero outside . At every Lebesgue point and for sufficiently small nonzero ,
The last expression tends to zero, proving the derivative claim. The estimate treats both signs of and does not require to be continuous.
For absolute continuity we must find, for every , a such that any finite collection of pairwise disjoint intervals with total length below satisfies . Given , choose with ; this follows from dominated convergence. Put . If is the union of the intervals, then
Thus small total length, not merely short individual intervals, controls the increments.
For a continuous , the same derivative conclusion holds at every interior point, directly from continuity. Almost-everywhere language is essential in general, as the following computation shows.
Worked example 2.1.7
(An indefinite integral with corners)
Let and on . Then on , on , and on . At an interior point , is differentiable with . At the one-sided derivatives are and , and at they are and , so does not exist there, even though has assigned values there. The exceptional set is null, as the theorem permits.
Absolute continuity is stronger than continuity
Proposition 2.1.8
(Lipschitz implies absolutely continuous implies uniformly continuous)
A Lipschitz function on is absolutely continuous, and an absolutely continuous function on is uniformly continuous.
Proof
(Choose from the Lipschitz constant)
If with , the sum of increments over disjoint intervals is at most times their total length, so works; a constant function is immediate. Applying the definition of absolute continuity to a single interval gives uniform continuity.
Neither implication reverses in general.
Worked example 2.1.9
(Absolutely continuous but not Lipschitz)
is absolutely continuous on : it is the indefinite integral of for , with an arbitrary value at , and is integrable since . But is not Lipschitz near , because as .
Constructing the Cantor function
The counterexample to the second reverse implication uses a function that rises from to even though it is locally constant almost everywhere. Here is a construction, so that these properties are consequences rather than assumptions.
Start with on . At stage , the retained set consists of closed intervals of length . On each such interval , is affine and rises by . Construct by leaving unchanged outside the retained intervals and replacing its graph on each by three pieces: rise linearly through half the old increment on the left third, remain constant on the middle third, and rise linearly through the other half on the right third. The two outer thirds are the retained intervals at the next stage. The endpoint values agree with the old graph, so is continuous and nondecreasing; it is also constant on every gap deleted by stage .
Within any stage- retained interval, both and take values between the same two endpoint values, which differ by . Elsewhere they agree. Hence
The sum of these bounds converges. The uniform Cauchy criterion therefore gives a uniform limit ; as a uniform limit of continuous functions it is continuous, and it is nondecreasing because each is. All later stages preserve the values at the endpoints of every retained interval and leave every deleted gap unchanged. Thus , , rises by across each stage- retained interval, and is constant on every deleted open interval.
Let be the intersection of the retained sets. Their total lengths are , so . Every point outside lies in a deleted open interval, where is constant; consequently there, and almost everywhere. For example, the first removed interval is , where ; continuity gives . The first removed interval inside is , where .
Counterexample mode
A derivative zero almost everywhere does not force a constant
Claim: a continuous nondecreasing function on with derivative almost everywhere is constant. The Cantor function refutes it: off the null set , yet . The same function is continuous but not absolutely continuous. The stage- retained intervals have pairwise disjoint interiors of total length , while . Taking in the definition, no works. The repaired claim assumes absolute continuity: an absolutely continuous function with derivative zero almost everywhere is constant.
The converse fundamental theorem says that every absolutely continuous on has an integrable derivative almost everywhere and satisfies . It is a deeper result, proved in the unit on absolute continuity and the fundamental theorem of calculus; the integral-to-derivative theorem above does not alone prove it. It explains why absolute continuity rules out the Cantor function's singular increase.
Common mistake
(Confusing short intervals with small total length)
Continuity makes the increment over each single short interval small. Absolute continuity asks for more: the increments over many disjoint intervals must have a small sum once their total length is small. The Cantor function satisfies the first requirement and fails the second.
Practice with solutions
Quick check
(If and for every , prove that almost everywhere.)
Differentiate the indefinite integral.
Solution · Answer
The indefinite integral is identically zero, so its derivative is zero at every interior point. The integral-to-derivative theorem identifies that derivative with almost everywhere, so almost everywhere.
Quick check
(Is the Cantor function Lipschitz? Explain using its construction-stage intervals.)
Compare increment with length on a retained interval.
Solution · Answer
No. On a stage- retained interval the ratio of its increment to its length is , which is unbounded in . A single Lipschitz constant cannot bound these ratios.
Quick check
(Let for and . Is locally integrable, and is a Lebesgue point of ?)
Compute the absolute average at .
Solution · Answer
is locally integrable, since . At , , so is not a Lebesgue point for any choice of . Every is a point of continuity, hence a Lebesgue point; the exceptional set is null, as the theorem allows.