An integral pairing turns a function into a linear measurement of another function: . Hölder's inequality explains why this measurement is bounded. The reverse question is more subtle: does every bounded linear measurement on arise this way, and is the norm of the measurement exactly the norm of its density? On finite measure spaces, Radon–Nikodym lets us recover the density from the functional's values on indicator functions.
Throughout has . We work with real-valued spaces and real-linear functionals. Functions equal almost everywhere are identified. We use completeness and simple-function density in for , Hölder's inequality, monotone convergence, and the finite nonnegative Radon–Nikodym theorem. The complex case requires the corresponding complex-linear convention and is not proved here.
From a density to a functional
For , let , so ; for put .
Definition 4.2.1
(The functional of a density)
For define
Hölder gives , so is bounded with ; for this reads . These norm estimates are upper bounds until we show they can be approached from below.
Theorem 4.2.2
(Riesz representation of on a finite measure space)
For , every bounded real-linear on is for a unique , and . For , the same assertion holds with and .
The representing function is unique only up to almost-everywhere equality. The proof occupies the next two sections: indicators give a signed measure and an density, and extremal test functions upgrade the density to . On a general -finite space a localization argument is needed; it is treated in the unit on -finite duality.
Build a signed measure from indicators
Fix , put , and define for measurable .
Lemma 4.2.3
(Indicator values form a finite signed measure)
is countably additive, absolutely continuous with respect to , and for every finite measurable partition ,
Proof
(Linearity, continuity and a choice of signs)
Since , every indicator lies in and is finite. A null set's indicator is the zero element of , so vanishes on null sets. If are pairwise disjoint, finite additivity follows from linearity, and
because makes the tail measures tend to zero. Continuity of gives . For the partition bound, choose signs with . Then .
Common mistake
(Defining on sets of infinite measure)
The construction needs . On with Lebesgue measure, , so is meaningless, and the tail estimate in the countable additivity argument also uses . This is why the finite case is proved first.
To apply the positive Radon–Nikodym theorem, split into positive and negative parts directly.
Lemma 4.2.4
(Positive and negative parts and an density)
Put , over measurable , and . Then are finite positive measures absolutely continuous with respect to , and there is with for every simple function .
Proof
(Approximate the supremum on each piece, then apply Radon–Nikodym)
The supremum is finite by the partition bound. Both are nonnegative: gives , and gives . Let be disjoint. Every splits into the sets , so . Conversely choose with ; finite unions show for every . Let and then . Thus is a finite positive measure, and is one by subtraction. Each vanishes on -null sets, since every measurable subset of a null set has zero -value.
Finite Radon–Nikodym gives nonnegative with . Put . Then , and linearity extends this to simple functions.
We have a density; the remaining task is to prove its stronger integrability and extend the identity to all of .
Why the density lies in the conjugate space
The simple-function identity extends to every bounded measurable function: approximate it uniformly by simple functions; their errors tend to zero because , and their integrals against converge.
Lemma 4.2.5
(The density has norm at most )
and , where when .
Proof
(Test against truncated extremal functions)
Assume . For each integer , define the bounded measurable test function
Since , , and . Hence
If , divide by to get ; the inequality is automatic when . Monotone convergence yields and .
For , suppose on a set of positive measure. The bounded test satisfies , while , contradicting . Hence .
Proof X-ray
Where the exponent comes from
Indicator functions produce only an density, and the Radon–Nikodym step uses nothing about . The exponent enters through the test : multiplying by produces , while its th power produces the same expression because . That coincidence is the key move, since it turns the operator bound into , an inequality between powers of a single number. Truncation at level is what keeps each test in before is known; without it the test could fail to be admissible. Finiteness of is used here only to put bounded functions in . The argument closes by monotone convergence in . At the extremal test degenerates to on a set where is large.
The theorem now follows from density and Hölder.
Proof
(Finishing the representation theorem)
Simple functions are dense in , and Hölder makes continuous on , so the identity extends from simple functions to all . Hölder also gives ; combined with the lemma, . If both represent , then for every measurable . Taking and shows almost everywhere.
Examples and the endpoint that changes
Worked example 4.2.6
(Recovering a density from indicators)
On with Lebesgue measure and , let . Then , which is positive on subsets of and negative on subsets of . The supremum defining is attained at , so and . Thus , and , as it should be. For , and .
Worked example 4.2.7
(The extremal test attains the norm)
Continue with and . The test gives and . The ratio , so the upper bound is attained exactly. No truncation is needed here because is bounded.
Worked example 4.2.8
(An unbounded density)
On with Lebesgue measure, take and for . Then
The corresponding functional on has exactly this norm. The condition is essential for this density to lie in ; an unbounded density can nevertheless define a bounded functional. Here the truncation in the lemma is genuinely used, since is unbounded.
The analogous representation for the endpoint fails in general.
Counterexample mode
A functional on L-infinity with no density
Claim: on a finite measure space every bounded functional on is for some . Let with , so . Then is because no singleton is null. On the closed subspace of convergent sequences, the limit functional has norm one. The real Hahn–Banach theorem extends it to a bounded functional on all of . If represented it, testing each coordinate vector would give , so everywhere, but . The theorem is correct only for ; one must not claim in general.
Practice with solutions
Quick check
(Let and on . Find the norm of .)
Use .
Solution · Answer
The theorem gives , attained by .
Quick check
(For and , find the functional norm and a unit-norm test attaining it.)
Concentrate the test where is largest.
Solution · Answer
. The test has and .
Quick check
(Why is every indicator in in the proof, and where is finite total measure used in countable additivity?)
Compute .
Solution · Answer
. Finite measure also makes the measures of the decreasing tails of a disjoint union tend to zero, which is the convergence needed for countable additivity.
Quick check
(In the counterexample, why does testing coordinate vectors rule out every density?)
Evaluate the extension on .
Solution · Answer
Each is convergent with limit zero, so the extension has value zero there. A density would give for every , forcing , contrary to the functional's value one on the constant sequence.