The norm for finite averages the size of a function. The endpoint instead records the smallest bound that can fail only on a null set. This distinction makes the norm insensitive to a single spike and lets it depend on an almost-everywhere equivalence class rather than on arbitrary point values. We prove that this endpoint is complete on an arbitrary measure space.
Let be any measure space, not necessarily finite, -finite or complete. Functions are real- or complex-valued and measurable. Two functions are equivalent if they agree outside a measurable null set.
Definition 4.6.1
(Essential supremum and )
The essential supremum of a measurable is
with . A number in this set is an essential bound of , and is essentially bounded if (1) is finite. The space consists of almost-everywhere equivalence classes of essentially bounded measurable functions.
The norm is an almost-everywhere bound
Lemma 4.6.2
(The infimum is itself an essential bound)
If , then almost everywhere.
Proof
(A countable union of null sets)
For each , the definition of infimum gives an essential bound smaller than . Hence . The set is the union of these countably many null sets, and is null. An infimum of an arbitrary set need not belong to it; here it does because countably many slightly larger bounds suffice.
Worked example 4.6.3
(An essential supremum that is not a maximum)
On with Lebesgue measure, let . For , has positive measure , so is not an essential bound; is, since . Thus , although everywhere, so the value is not attained. If we redefine , its pointwise supremum becomes , while its essential supremum remains : the change is on a null set.
Worked example 4.6.4
(A nowhere-continuous function with essential supremum zero)
On let . It is discontinuous at every point and its pointwise supremum is . But is countable, hence null, so is an essential bound and . In , is the zero class.
The lemma is what makes (1) behave like a norm on classes.
Proposition 4.6.5
(The essential supremum is a norm)
Formula (1) defines a norm on : it is independent of the measurable representative, is homogeneous, obeys the triangle inequality, and vanishes exactly on the zero equivalence class.
Proof
(Combine two almost-everywhere bounds)
If outside a measurable null set, an essential bound for either is also one for the other, so their infima agree. The lemma shows and outside the union of two null sets. Therefore almost everywhere, proving the triangle inequality and closure under addition. Homogeneity follows by scaling essential bounds, with scalar zero handled directly. If , the lemma gives almost everywhere; the converse is immediate.
Completeness
Theorem 4.6.6
(Completeness of )
Every Cauchy sequence in converges in the norm to an element of .
Proof
(A fast subsequence converges uniformly off one null set)
Let be Cauchy and choose measurable representatives. Select increasing indices so that
By the lemma, each inequality holds pointwise outside a measurable null set. Take the union of these sets and a null set outside which is bounded, obtaining one measurable null set . On , the telescoping series
converges uniformly by comparison with . Define as this limit outside and as zero on . It is measurable: all partial sums multiplied by are measurable and converge pointwise to . The uniform bound on makes essentially bounded. For every the tail estimate gives
Finally, given , let be a Cauchy index for . Choose so large that and the right side of (3) is below . Then for every , the Cauchy property gives . The triangle inequality yields in .
Proof X-ray
One null set, chosen once
The proof uses the lemma at its first step: each norm inequality (2) becomes a pointwise inequality off a null set. The key move is to pass to a subsequence with summable gaps, so that only countably many exceptional sets arise; their union is null on any measure space. Off the problem is the classical one of uniform convergence of a series with summable bounds. Measurability of the limit needs care on a noncomplete space: the construction redefines only on the measurable set , so it never relies on subsets of null sets being measurable. The argument closes in two moves: the tail bound (3) gives convergence of the subsequence in norm, and a Cauchy sequence with a convergent subsequence converges.
On a noncomplete space, the indicator of a nonmeasurable subset of a measurable null set is zero almost everywhere but is not measurable; that is why the construction changes values on the measurable null set itself.
Comparing convergence and bounds
Worked example 4.6.7
(Convergence in every finite but not in )
On , let . Then pointwise and, for each finite , . Yet for every , because the set where has positive measure . Thus finite- convergence does not imply convergence, even on a probability space.
Common mistake
(Taking a supremum over uncountably many representatives)
On with Lebesgue measure, each is zero almost everywhere, but the pointwise supremum over all is the constant one. The union of uncountably many null exceptional sets need not be null. One cannot form a pointwise supremum of arbitrary representatives and expect it to depend only on their classes; countability, as in the completeness proof, is what keeps the exceptional set null.
The essential bound also controls integrals.
Proposition 4.6.8
(Endpoint Hölder and finite-measure inclusion)
For and , on any measure space, . If and , then
Proof
(Integrate the essential bound)
By the lemma, and almost everywhere. Integrating gives and ; take th roots for (4). The case is included.
The first inequality is the basic pairing used in – duality. The second one depends on its hypothesis.
Counterexample mode
Without finite measure, bounded does not imply integrable
Claim: for . On with Lebesgue measure, the constant has , but for every finite . Inequality (4) gives no information here, since its right side would be . The correct statement requires , as in (4).
Practice with solutions
Quick check
(If in and , prove in for finite .)
Apply (4) to a difference.
Solution · Answer
By (4) applied to , , and the right side tends to zero.
Quick check
(Why is zero in for each fixed ?)
Measure the set where it is positive.
Solution · Answer
A singleton is Lebesgue measurable and has measure zero, so is an essential bound and ; the function belongs to the zero class.
Quick check
(Can an Cauchy sequence fail to converge pointwise for a badly chosen sequence of representatives?)
Change the representatives on a null set.
Solution · Answer
Yes. On , take . Every term represents the zero class, so the sequence converges in norm to zero, but its chosen values diverge at . The proof of completeness is allowed to discard one common measurable null set.
Quick check
(On with Lebesgue measure, find for on , on , and redefined to be .)
Only matters, and single points are null.
Solution · Answer
on and on ; the point is null. Every is exceeded on , which has measure , while is null. Hence : neither the spike value nor is the answer.