A density turns a measure into an integral with respect to another measure. For example, a nonnegative function on an interval assigns to a measurable set the mass . Such a measure must vanish on every set of Lebesgue measure zero. The converse question is deeper: if one measure vanishes on the null sets of another, must it arise from a density? For finite measures the Radon–Nikodym theorem answers yes.
Our proof uses one substantial external result: the Riesz representation theorem for a Hilbert space. It also uses the standard fact that of a finite measure is a Hilbert space, together with monotone convergence and Cauchy–Schwarz. We state precisely where these inputs enter; the theorem below is proved from them, not from a differentiability assumption on distribution functions.
Null sets and densities
Let be a measurable space, and let be finite nonnegative measures.
Definition 3.1.1
(Absolute continuity of measures)
We write and say that is absolutely continuous with respect to if
The direction matters: it is a statement about on -null sets. It is different from absolute continuity of a real function on an interval, although the fundamental theorem later links these notions.
Common mistake
(Reversing the direction of absolute continuity)
does not imply . On let be Lebesgue measure and . Every -null set is -null, so . But has -measure and -measure , so . A density exists here, while does not.
If is measurable and , then . Indeed, every nonnegative measurable function has integral zero over a -null set. If is finite, must be finite -almost everywhere: otherwise the integral over would be infinite unless that set were null.
Definition 3.1.2
(Mutually singular measures)
Two measures are mutually singular, written , if some measurable satisfies and .
A point mass on is singular with respect to Lebesgue measure : take . It cannot be written as for a measurable density, since its mass on that null set is one.
The theorem and the Hilbert-space construction
Theorem 3.1.3
(Radon–Nikodym theorem, finite case)
If and are finite nonnegative measures on and , then there is a nonnegative measurable , finite -almost everywhere, such that
The function is unique up to -almost-everywhere equality and is denoted . It satisfies .
Proof
(Densities relative to the sum measure)
First put both measures under one finite measure. Set and consider the real Hilbert space . Define . This is well-defined on -equivalence classes because . Cauchy–Schwarz gives
Thus is bounded and linear. By the Hilbert-space Riesz theorem there is such that for all . Indicators belong to since , so for every measurable ,
These identities force -almost everywhere. For instance, a set where with positive measure would have strictly negative integral of , contradicting the first identity; similarly a set where contradicts the second. We choose a measurable representative with everywhere by changing it on a -null set.
Remove the points at which the denominator vanishes. Let . The second identity gives . Since , also , and hence . Define
To integrate this ratio with respect to , note the change-of-density identity
It holds for indicators by the identity for , then for nonnegative simple functions by linearity, and finally for arbitrary nonnegative measurable by monotone approximation and monotone convergence. Taking yields
because . The integral over is finite and equals , so is finite -almost everywhere.
Uniqueness. Suppose nonnegative and give the same finite measure. Both are finite almost everywhere and have finite integrals. If , then for some integer , has positive measure. Consequently , contradicting equality of the measures on , where both integrals are finite. Interchanging proves equality almost everywhere.
Proof X-ray
What each hypothesis does in the construction
Neither measure is initially known to have a density relative to the other, so the proof passes through the sum , which dominates both. Finiteness is used twice: makes bounded, and puts every indicator into , so the Riesz representative can be tested on sets. The key move is that the same gives both densities, for and for , because is a subtraction of finite measures. Positivity of the two measures pins into . Absolute continuity is used exactly once: on the -density vanishes, so , and only forces as well. Without it, carries the singular part of , as the decomposition below shows. The change-of-density identity then converts the ratio into the required -integral.
The proof's ratio is not a pointwise quotient of the original measures. It comes from two densities relative to the auxiliary measure . Finiteness of the reference measure cannot simply be dropped.
Counterexample mode
Absolute continuity without a density
Claim: whenever with finite, has a density with respect to . On with the Borel sets, let be counting measure and Lebesgue measure. The only -null set is , so . Suppose for every Borel . Taking gives for every , hence , a contradiction. Here is infinite, and not even a countable union of finite-measure sets. The repaired statement requires finite, as in the theorem; the -finite extension is a separate result.
Worked example 3.1.4
(The construction on a two-point space)
Let with , , , . Then , , and forces , . Check and . Here , and
so . For the edge case, change the data to , : then , , and . Absolute continuity fails, and is where the singular part of sits.
A decomposition when absolute continuity fails
Theorem 3.1.5
(Lebesgue decomposition for finite measures)
If and are finite nonnegative measures on , then with and , and this decomposition is unique.
Proof
(Split off the set where the -density vanishes)
The same construction works without assuming . With and as above, define
Then , and is singular with respect to . Define on and on . The change-of-density identity gives , so . Here need not be zero: the restriction to is essential.
To prove uniqueness, suppose two decompositions have singular parts supported on -null sets , and put . For either decomposition, its absolutely continuous part gives zero mass to , and its singular part gives zero mass outside . Consequently for every measurable , . The absolutely continuous parts agree, and finite-measure subtraction from makes the singular parts agree as well.
Worked example 3.1.6
(An absolutely continuous part plus an atom)
On let and . The first term is absolutely continuous with density ; the second is concentrated on the -null set , so the decomposition is and , with . Although the integral part assigns no mass to the point , , so the whole measure is not absolutely continuous. Treating as the density of the entire measure would lose its atom.
Changing the reference measure
Worked example 3.1.7
(An unbounded density after changing the reference)
On let and . Here : the weight is positive except at the single Lebesgue-null point 0, so a measurable set with has Lebesgue measure zero. A density is
with any finite value assigned at 0. It is unbounded near 0, yet its -integral is . A finite measure need not have a bounded density. The formula is valid because the factor vanishes only on a null set; dividing by a reference density that vanishes on a set of positive target mass would not be valid.
On the interval, if for , then is absolutely continuous as a function and almost everywhere. This is the fundamental theorem for Lebesgue integrals, a separate result. Radon–Nikodym works on an arbitrary measurable space and does not require an order, intervals or a pointwise derivative.
Practice with solutions
Quick check
(On let , and , . Find and explain its freedom at .)
Solution · Answer
The density at is . Its value at is arbitrary because ; all such choices agree -almost everywhere.
Quick check
(With the same , replace by 1. Does a density with respect to exist?)
Solution · Answer
No: is -null but has positive -mass, so , while any is .
Quick check
(On let and . Compute and state whether .)
Solution · Answer
The integral is . Also : except at the Lebesgue-null point 0, so a set of zero -measure has zero Lebesgue measure.
Quick check
(Suppose and both are finite. If is very small, must also be small uniformly over measurable ? Prove your answer using the density.)
Solution · Answer
Yes. Write with . Given , dominated convergence applied to as lets us choose so that . If , then . This uniform small-set statement is stronger-looking than the null-set definition but follows from it in the finite setting.