On a finite measure space, every indicator belongs to for finite . On an infinite space, the indicator of the whole space need not be integrable, so the finite-space representation cannot simply be applied globally. -finiteness provides finite pieces. The essential extra step is to prove that the local representing densities fit together with a single global bound.
Let be -finite, and work with real spaces for . We use the finite-measure duality theorem from the preceding unit, Hölder's inequality and monotone convergence. For put ; for write . Functions equal almost everywhere are identified.
The representation theorem
Theorem 4.8.1
(Lp duality on a -finite space)
Every bounded real-linear functional on has a unique density such that
For the norm on the right is the essential supremum. Conversely, every such density defines a bounded functional by the integral.
Proof
(Local densities, a global test function and a limit)
Local representations. Choose pairwise disjoint measurable sets of finite measure whose union is : take a countable finite-measure cover and subtract preceding sets. Set . Extending functions by zero outside identifies isometrically with a subspace of . The restriction of has norm at most , so finite-measure duality gives with
Choose measurable representatives and define on . This is measurable; null-set changes on countably many pieces change it only on a null set. A separate argument is needed to show when : merely knowing that each piece has norm at most does not make the sum of their th powers finite.
Global estimate for . For put
The function is bounded and supported on a set of finite measure, so . Since , its norm satisfies . Summing the local representations gives . Thus
When , this yields ; the conclusion is trivial if . As increases, the truncated integrands increase to almost everywhere. Monotone convergence gives and .
Global estimate for . Each . For every , outside a null subset of ; their countable union is null. Hence and .
Passing from pieces to all of . For , linearity gives for every . Monotone convergence applied to and finiteness of show in . Continuity of gives , while Hölder gives . Thus both sides converge and the integral formula follows. Hölder also gives , and together with the preceding bound proves norm equality.
For uniqueness, if two densities represent the same functional, their restrictions to each represent the same restricted functional. Finite-measure uniqueness makes them equal almost everywhere on every piece, hence almost everywhere on . Conversely, Hölder immediately proves that any defines a bounded integral functional. Its representing density is itself, so the norm equality just proved gives its norm as .
Proof X-ray
Where $\sigma$-finiteness and the single constant C are used
-finiteness enters only at the start, to cut into countably many finite-measure pieces on which the earlier theorem applies. Those local densities glue to a measurable , but each local bound alone does not control : infinitely many pieces could each contribute up to . The key move is the test function , which assembles finitely many pieces into one admissible input; it is truncated so that it lies in before we know . Evaluating on it compares the total with the single operator norm , uniformly in , and monotone convergence closes the bound. For the same role is played by countable unions of null sets. The last step uses continuity of , since in .
Common mistake
(Gluing local bounds into a global bound)
From for every one cannot conclude . On , has on every unit interval, yet ; it is not the density of any bounded functional on . The bound must come from testing on functions spread over many pieces at once.
Is the norm attained?
Proposition 4.8.2
(When the dual norm is attained)
Let with nonzero. For the norm is attained at a unit vector of . For it need not be attained, but it is the supremum of over unit vectors supported on sets of finite positive measure.
Proof
(An explicit maximizer and an approximate one)
For , the unit vector
satisfies and . For , if and , the set has positive measure. Some intersection with a finite-measure piece has positive finite measure, and
has and . Thus the supremum is still the essential supremum.
Worked example 4.8.3
(An attained norm on L2)
On with Lebesgue measure, has . Its integral functional on therefore has norm one. With the maximizer formula gives , and indeed .
Worked example 4.8.4
(An explicit maximizer when p equals 3)
On let and , so and . Then . The maximizer is , equal to on and on . Check: , and .
Worked example 4.8.5
(An norm that no unit vector attains)
The function defines a functional on of norm . No unit vector attains it: off the countable null set . If , then , so . The approximating vectors of the proposition are with a short interval around , giving values as close to as desired.
Common mistake
(Expecting every dual norm to be attained)
For a maximizer always exists, but for the essential supremum of may be approached without being reached by on any set of positive measure, and then no unit vector attains the norm. Compute the norm as a supremum, not as a maximum.
A non--finite warning
Without -finiteness the theorem can fail already for .
Counterexample mode
A one-point space of infinite measure
Claim: the representation (1) holds on every measure space. Let with and . This measure is not -finite. Its only null set is empty. A function has , finite only for , so , while because . The densities and differ at the one point, hence almost everywhere, but both induce the zero functional on , which has norm . For the uniqueness and norm equality in (1) therefore fail here.
-finiteness is sufficient for the theorem; this example does not show that it is necessary for every space, nor does it establish failure for .
Practice
Quick check
(On , take and . Find the norm of .)
Solution · Answer
, so the functional norm is , attained at .
Quick check
(Why does fail as a universal test function on an infinite-measure space?)
Solution · Answer
when , so it does not belong to the domain.
Quick check
(If and on , give an unit vector attaining the norm.)
Solution · Answer
has and , equal to .
Quick check
(In the one-point infinite-measure example, why are the densities and distinct almost everywhere but equivalent as functionals on ?)
Solution · Answer
No nonempty set is null, so they differ almost everywhere on the single point; nevertheless the domain contains only the zero function, on which both functionals vanish.