Motivation
Lebesgue measure on the real line has infinite total mass, but every bounded interval has finite measure. An infinite total does not make local information useless. The right question is whether the space can be covered by countably many manageable pieces, and whether the identities obtained on those pieces can be assembled without losing measurability or uniqueness.
For densities, disjoint pieces solve the overlap problem: define the global function by using exactly one local density at each point. For a singular measure, countability is equally important: countably many null supports have a null union. There is also a danger that did not arise for finite measures. One cannot prove uniqueness by subtracting two infinite total masses. The proof must compare restrictions on measurable sets instead.
We use the finite nonnegative Radon–Nikodym theorem and finite Lebesgue decomposition proved in the preceding measure note. Simple-function approximation and monotone convergence are the other inputs. This section gives the -finite extension from those results; it does not assume the extension in its proof.
Definitions
Definition 3.2.1
(-finiteness and restriction)
A nonnegative measure on a measurable space is -finite if there are measurable sets with and for every . The cover need not initially be increasing or disjoint. Replacing by gives an increasing finite-measure cover.
For measurable , the restricted measure is . For measurable ,
This holds first for indicators, then simple functions, then nonnegative measurable functions by monotone convergence.
All measures in the main theorems are nonnegative and defined on the same -algebra. Their total masses may be infinite. Integrals of nonnegative functions take values in ; an integral over a null set is zero even if the function is infinite there.
We retain for absolute continuity: implies . We write if a measurable satisfies and . Neither condition is an informal assertion that the measures have small or disjoint numerical values on every set.
Theorem/Proposition
Theorem 3.2.2
(A common disjoint finite-measure partition)
If and are -finite, there is a measurable partition such that and for every . Empty pieces are allowed.
Theorem 3.2.3
(Radon–Nikodym for two -finite nonnegative measures)
Let be -finite on , and assume . There exists a nonnegative measurable , finite -almost everywhere, such that
It is unique up to -almost-everywhere equality and is denoted . The identity is allowed to be infinite; the theorem does not assert when .
Theorem 3.2.4
(Lebesgue decomposition in the -finite setting)
For -finite nonnegative , without assuming , there are unique nonnegative measures such that
Both components are -finite. Moreover for a nonnegative measurable density finite -almost everywhere and unique in that sense. Uniqueness of the components is equality of measures on every measurable set, not merely equality of their total masses.
Proof sketch or proof idea
Proof
(Make both measures finite on the same pieces)
Choose increasing measurable covers , with and . Set . These sets increase to : a point that enters the two covers at different indices belongs to their intersection at every sufficiently large index. Each is finite for both measures.
Put and . The sets are disjoint, cover and inherit both finite-measure bounds. A cover that is finite for just one of the measures would not justify applying the finite theorem to both restrictions.
Proof
(Construct and identify the global Radon–Nikodym density)
Use the common partition and write , . These are finite measures on the original -algebra. Also : a zero value of forces a zero value of .
The finite theorem gives a nonnegative measurable density with . Define to equal on . It is measurable because the partition is countable. The exceptional set where it is infinite is a countable union of sets null for the corresponding restricted measures, hence null for after intersecting each with .
For any measurable , monotone convergence applied to the partial sums of the disjointly supported nonnegative functions gives
Every equality permits infinity; no subtraction of infinities occurs. Each piece has even when the global integral diverges.
For uniqueness, let be another density with the same measurable-set identities. On each , both and represent the same finite measure relative to the same finite restriction. The finite theorem's uniqueness gives equality -almost everywhere. The union of these countably many exceptional subsets of is -null, proving global uniqueness. In particular the density does not depend, up to null sets, on which common finite partition was used.
Proof X-ray
Why the pieces must be both finite and disjoint
Finiteness lets the Radon–Nikodym theorem for finite measures construct each local density. Disjointness makes the global definition unambiguous without choosing among overlapping representatives. Countability preserves measurability, permits monotone summation and keeps the exceptional union null. Uniqueness is checked on finite restrictions, so it never tries to infer equality by canceling two infinite integrals over the whole space.
The decomposition theorem is proved on the same common partition, followed by a uniqueness argument that only restricts measures to sets.
Proof
(Patch the finite decompositions and compare restrictions)
For decomposition, apply the finite Lebesgue decomposition separately to . Extend its two components by zero outside , and write them as . Choose the singular support inside . This is legitimate because is already concentrated on . Then , not merely for an arbitrary set extending outside the piece.
Define , and . Sums of nonnegative measures are measures: countable additivity follows by interchanging nonnegative series, or by monotone convergence of their finite partial sums. We have , , and . Since each component is bounded above by , a finite- cover is also a finite-measure cover for each component. Thus both are -finite, and the already proved Radon–Nikodym theorem supplies the density of .
To prove uniqueness, suppose two decompositions have singular supports , and set . Both absolutely continuous parts vanish on , and both singular parts vanish outside . Therefore, for either decomposition and every measurable ,
These restriction identities prove uniqueness even when either side has infinite mass. Writing pointwise on sets would instead encounter the undefined expression .
Concept lensStructural
The singular part lives on a largest null set
The decomposition is a restriction of to a set: and for a -null . Such an can be found without densities. On a piece , where is finite, choose -null sets with ; their union is -null and attains . If but , adding that set to would exceed . Hence makes .
Worked examples
Worked example 3.2.5
(A finite density on a space of infinite mass)
Let on and . Both measures are -finite, for instance on the increasing cover . The density is , finite everywhere, but . It is therefore not an function on the whole line.
The common partition may be made from the intervals by taking successive differences. On every such piece the integral of is finite, and adding those integrals recovers all of . A density is a local change-of-measure description; -finiteness does not turn an infinite target measure into a finite one.
Worked example 3.2.6
(Infinitely many atoms and an infinite continuous part)
On the real line let and
Every bounded interval has finite -mass, so both measures are -finite. The Lebesgue decomposition is and . The singular part is concentrated on , a countable Lebesgue-null set. Both components have infinite total mass, but their sum is a well-defined measure.
The absolutely continuous density is one. It is not the density of the entire : each singleton integer has zero Lebesgue measure and positive target mass. Nor should the singular component be computed as on the whole line; restricting to its null support identifies it directly.
Worked example 3.2.7
(Changing the reference on a countable space)
Let , let be counting measure, and set
Both are -finite on finite initial segments. Since only the empty set is -null, , and its density is on even integers and zero on odd integers. This density is finite but not summable with respect to counting measure.
Reverse the reference: decompose relative to . Counting on the even integers is absolutely continuous with density there, while counting on the odd integers is singular, supported on the -null odd set. Values of on odd integers are irrelevant to its -integral. Absolute continuity is directional, and a singular part can be -finite without being finite.
What can fail without -finiteness?
Counterexample mode
A finite target can have no density for a non-sigma-finite reference
On with the Lebesgue -algebra, let be counting measure and . Every finite- set is finite, so countably many such sets cannot cover the uncountable interval. Thus is not -finite, whereas is finite. The only -null set is empty, so .
If a density existed, every singleton would give . Hence would vanish everywhere and could not give . This excludes even an extended-valued nonnegative density, not merely an integrable one.
The counterexample above removes -finiteness of the reference measure; the one below keeps the reference finite and removes -finiteness of the target.
Counterexample mode
Without a sigma-finite target the density need not be finite almost everywhere
Keep on and define when , and when . This is a measure: for a disjoint countable union, either one piece has positive Lebesgue measure, or all pieces and their union are null. It is absolutely continuous with respect to . But every set of finite -mass is -null, so no countable finite- cover exists.
An extended density does exist: everywhere. No density finite -almost everywhere can work. If were such a density, the sets would cover almost all the interval, so at least one would have positive Lebesgue measure. On that set , contradicting its infinite -mass. This example defeats the finite-almost-everywhere conclusion, not the existence of every possible extended density.
The hypotheses are a sufficient package, not a claim that every individual non--finite pair must fail. For example any measure has the density one relative to itself. What the counterexamples show is that deleting hypotheses from the general conclusion requires a new theorem, not the same proof with the word “finite” erased.
Common mistakes
Common mistake
(Finite almost everywhere does not mean integrable on finite pieces)
Finite almost everywhere does not imply bounded, globally integrable, or integrable on every set of finite reference measure. The theorem constructs common pieces that are finite for the target as well. On , the density for , set to zero at zero, defines a -finite target but has infinite integral on the entire reference-finite interval.
Common mistake
(A density is identified on every set, not by its total)
Comparing only does not identify a density. The identity is required on every measurable set. Countability is doing mathematical work throughout: an uncountable collection of null singletons can cover a positive-measure space, while the countable exceptional unions in the proof remain null.
Summary
Two -finite measures admit a common disjoint partition on which both restrictions are finite. The finite theorem constructs local densities; monotone convergence assembles their measurable-set identities, and local uniqueness assembles their almost-everywhere agreement. Singular supports combine through a countable null union. Restriction formulas prove decomposition uniqueness without subtracting infinite masses, and the counterexamples distinguish absence of a density from the need for an infinite-valued one.
Exercises
Quick check
(On the real line let and . Verify -finiteness and mutual absolute continuity, find both densities, and compute .)
Check the density by integration against the reference weight, not by dividing the masses of one chosen set.
Solution
Bounded intervals have finite mass for both measures, and each weight is everywhere positive. A zero integral of either weight over a measurable set forces that set to be Lebesgue-null, proving mutual absolute continuity. The densities are and its reciprocal . Multiplication by the relevant reference weight verifies both integral identities. Finally . Neither measure has finite mass on the whole line.
Quick check
(Let be -finite, let be measurable, and define . Prove that is -finite if and only if is finite -almost everywhere.)
Combine finite- sets with level sets of . Treat the set explicitly.
Solution
Suppose first that is finite outside a -null measurable set . Cover by -finite sets . Each has -mass at most , and these countably many sets cover . Also , so adding gives a -finite cover.
Conversely, if cover with , put . Finiteness of forces for each . Their countable union is , hence . This direction itself does not need -finiteness of ; that hypothesis was used to construct the forward cover.
Quick check
(On let for Lebesgue-null and otherwise, and let . Show that the constants and are both densities of relative to but are not equal -almost everywhere.)
The reference measure here is not Lebesgue measure. Explain which uniqueness hypothesis is absent.
Solution
For every measurable , and , since the only possible values are zero and infinity. Thus both constants represent . Yet their inequality set is all of , whose -measure is infinite, so they do not agree -almost everywhere. Every finite- set is Lebesgue-null, making a countable finite-measure cover impossible. The uniqueness theorem for -finite reference and target measures has not been contradicted.