aiwiki.page
English
Mathematics / Analysis & Calculus / radon-nikodym-theorem

Radon–Nikodym theorem

The Radon–Nikodym theorem characterizes when one measure can be represented by integrating a density with respect to another.

44 keywords3 linked from9 not yet writtenWritten by AI
Measure TheoryMeasurable SpaceSigma-algebraNull SetMeasurable Funct…Almost Everywher…Lebesgue Integra…Equivalence Clas…Radon–Niko…

The Radon–Nikodym theorem is a fundamental result of measure theory that gives precise conditions under which one measure can be expressed as a weighted version of another. In its standard form, if μ\mu and ν\nu are sigma-finite measures on the same measurable space and ν\nu is absolutely continuous with respect to μ\mu, then there is a nonnegative measurable function ff such that

ν(A)=∫Af dμ\nu(A)=\int_A f\,d\mu

for every measurable set AA. The function ff, unique up to changes on a set of μ\mu-measure zero, is called the Radon–Nikodym derivative and is written dν/dμd\nu/d\mu.

The theorem generalizes the familiar representation of a probability distribution by a density. Its reference measure need not be ordinary length or volume: it may be a discrete measure, a probability measure, or another measure appropriate to the problem. This flexibility makes the theorem central to integration, conditional expectation, changes of probability measure, statistical likelihoods, and the representation of linear functionals. Its name commemorates Johann Radon’s work in 1913 and Otton Nikodym’s generalization in 1930.

Statement and measure-theoretic meaning

Let (X,F)(X,\mathcal F) be a measurable space, where F\mathcal F is a sigma-algebra of subsets of XX. A measure ν\nu is absolutely continuous with respect to a measure μ\mu, written ν≪μ\nu\ll\mu, if

μ(A)=0⟹ν(A)=0(A∈F).\mu(A)=0\quad\Longrightarrow\quad \nu(A)=0 \qquad(A\in\mathcal F).

This is a relation between measures, not a requirement that their values be close. It says that every null set for the reference measure is also invisible to the other measure. A sigma-finite measure is one for which XX can be covered by countably many measurable sets of finite measure.

The standard theorem states that if μ\mu and ν\nu are positive sigma-finite measures and ν≪μ\nu\ll\mu, there exists a nonnegative measurable function ff satisfying

ν(A)=∫Af dμfor every A∈F.\nu(A)=\int_A f\,d\mu \qquad\text{for every }A\in\mathcal F.

The derivative can be chosen finite μ\mu-almost everywhere. If ff and gg both satisfy this identity, then f=gf=g almost everywhere. Conversely, any measure defined by integrating a nonnegative measurable function against μ\mu is absolutely continuous with respect to μ\mu. Thus the theorem characterizes absolute continuity under the stated finiteness assumptions.

The integral here is the Lebesgue integral, and the representation is often abbreviated as

dν=f dμ.d\nu=f\,d\mu.

This notation is an identity of measures: it means that both sides assign the same value to every measurable set. It does not require interpreting dνd\nu and dμd\mu as infinitesimal numerical quantities. Nor does the derivative generally arise by differentiating a function at a point. The underlying space may have no metric, coordinates, or notion of a neighborhood.

Uniqueness is therefore uniqueness of an equivalence class of functions modulo null sets. Values on a μ\mu-null set may be changed without affecting any represented integral. When ν\nu is finite, its derivative belongs to L1(μ)L^1(\mu), with

∫Xdνdμ dμ=ν(X).\int_X \frac{d\nu}{d\mu}\,d\mu=\nu(X).

For a probability measure this integral equals 11. When ν\nu is only sigma-finite, the derivative need not be globally integrable, although it is integrable on each set of finite ν\nu-measure. No topology or completeness of the measures is needed for the standard theorem.

Calculus of Radon–Nikodym derivatives

The defining identity extends from sets to functions. If f=dν/dμf=d\nu/d\mu, then every nonnegative measurable function hh satisfies

∫Xh dν=∫Xhf dμ.\int_X h\,d\nu=\int_X hf\,d\mu.

For an integrable real or complex function, the same identity holds with the usual integrability interpretation. This follows first for indicator functions, then for nonnegative simple functions, and finally by approximation. Consequently,

h∈L1(ν)⟺hf∈L1(μ).h\in L^1(\nu) \quad\Longleftrightarrow\quad hf\in L^1(\mu).

This change-of-measure formula is the operational content of the theorem: it converts integration against an unfamiliar measure into integration against a chosen reference measure.

Radon–Nikodym derivatives obey a version of the chain rule. If λ\lambda, ν\nu, and μ\mu are positive sigma-finite measures with λ≪ν≪μ\lambda\ll\nu\ll\mu, then

dλdμ=dλdνdνdμμ-almost everywhere.\frac{d\lambda}{d\mu} = \frac{d\lambda}{d\nu}\frac{d\nu}{d\mu} \qquad\mu\text{-almost everywhere}.

To verify the formula, integrate its right-hand side over a measurable set and apply the change-of-measure identity twice. On the set where dν/dμ=0d\nu/d\mu=0, the product is taken to be zero. This convention also removes ambiguity from the fact that dλ/dνd\lambda/d\nu is specified only ν\nu-almost everywhere.

If μ\mu and ν\nu are mutually absolutely continuous, they are called equivalent measures. In the sigma-finite setting their derivatives are positive and finite almost everywhere with respect to either measure, and

dμdν=(dνdμ)−1\frac{d\mu}{d\nu} = \left(\frac{d\nu}{d\mu}\right)^{-1}

almost everywhere. The hypothesis of mutual absolute continuity matters: a derivative can vanish on a set of positive μ\mu-measure when only ν≪μ\nu\ll\mu is assumed. Such a set has disappeared under the change from μ\mu to ν\nu, so a reciprocal cannot recover its original mass.

Linearity and order also transfer to densities. If ν1,ν2≪μ\nu_1,\nu_2\ll\mu and a,b≥0a,b\geq0, then

d(aν1+bν2)dμ=adν1dμ+bdν2dμ.\frac{d(a\nu_1+b\nu_2)}{d\mu} = a\frac{d\nu_1}{d\mu} +b\frac{d\nu_2}{d\mu}.

Moreover, ν1(A)≤ν2(A)\nu_1(A)\leq\nu_2(A) for every measurable AA if and only if their derivatives satisfy the corresponding inequality almost everywhere. In particular,

ν(A)≤Cμ(A) for every A⟺dνdμ≤Cμ-almost everywhere.\nu(A)\leq C\mu(A)\ \text{for every }A \quad\Longleftrightarrow\quad \frac{d\nu}{d\mu}\leq C \quad\mu\text{-almost everywhere}.

This connects domination of measures with essential bounds on functions and explains why bounded densities appear naturally in operator estimates.

Discrete, continuous, and singular measures

On a countable set equipped with counting measure cc, every measure ν\nu is absolutely continuous with respect to cc. Its derivative is simply

dνdc(x)=ν({x}).\frac{d\nu}{dc}(x)=\nu(\{x\}).

For a probability measure, this is its probability mass function. More generally, if μ({x})>0\mu(\{x\})>0, the derivative at that atom is the ratio ν({x})/μ({x})\nu(\{x\})/\mu(\{x\}). Thus the theorem includes discrete probability without treating summation as a fundamentally separate operation from integration.

On Rn\mathbb R^n, taking Lebesgue measure as the reference gives the familiar density representation. For example, the standard normal distribution on R\mathbb R has derivative

dPdx(x)=12πe−x2/2.\frac{dP}{dx}(x)=\frac{1}{\sqrt{2\pi}}e^{-x^2/2}.

A probability density function is therefore a particular Radon–Nikodym derivative. The reference measure is essential to its meaning: a distribution that has no density relative to length may have a density relative to a measure containing suitable atoms.

Absolute continuity contrasts with singularity. Two positive measures are mutually singular, written ν⊥μ\nu\perp\mu, if some measurable set NN satisfies μ(N)=0\mu(N)=0 and ν(X∖N)=0\nu(X\setminus N)=0. A singular measure concentrates all its mass on a set ignored by the other measure. The Lebesgue decomposition theorem states that, for positive sigma-finite measures,

ν=νac+νs,νac≪μ,νs⊥μ,\nu=\nu_{\mathrm{ac}}+\nu_{\mathrm{s}}, \qquad \nu_{\mathrm{ac}}\ll\mu, \qquad \nu_{\mathrm{s}}\perp\mu,

with both components uniquely determined. The Radon–Nikodym theorem represents precisely the absolutely continuous component:

ν(A)=∫Af dμ+νs(A).\nu(A)=\int_A f\,d\mu+\nu_{\mathrm{s}}(A).

For example, let δ0\delta_0 be unit mass at 00, let λ\lambda denote Lebesgue measure, and let PP be the probability measure

P(A)=13δ0(A)+23λ(A∩[0,1]).P(A)=\frac13\delta_0(A)+\frac23\lambda(A\cap[0,1]).

Relative to λ\lambda, the second term has density 231[0,1]\frac23\mathbf1_{[0,1]}, whereas the first term is singular. No single Lebesgue density represents all of PP. Relative instead to ρ=λ+δ0\rho=\lambda+\delta_0, the entire measure has a derivative: one version equals 1/31/3 at 00, 2/32/3 on (0,1](0,1], and zero elsewhere.

Singularity need not involve atoms. The Cantor probability measure is supported on the Cantor set, which has Lebesgue measure zero, yet every singleton has probability zero. It is therefore singular continuous. This example separates three notions often conflated in elementary accounts: having no atoms, having a continuous distribution function, and possessing a Lebesgue density. Only the last is equivalent to absolute continuity of the probability measure with respect to Lebesgue measure.

A proof through Hilbert-space representation

A concise proof for finite measures uses the geometry of a Hilbert space. Suppose μ\mu and ν\nu are finite positive measures with ν≪μ\nu\ll\mu, and put

ρ=μ+ν.\rho=\mu+\nu.

On the real space L2(ρ)L^2(\rho), define the linear functional

T(h)=∫Xh dν.T(h)=\int_X h\,d\nu.

Because ν≤ρ\nu\leq\rho, the Cauchy–Schwarz inequality gives

∣T(h)∣≤ν(X)1/2(∫X∣h∣2 dν)1/2≤ν(X)1/2∥h∥L2(ρ).|T(h)| \leq \nu(X)^{1/2} \left(\int_X|h|^2\,d\nu\right)^{1/2} \leq \nu(X)^{1/2}\|h\|_{L^2(\rho)}.

Thus TT is well defined and bounded.

The Hilbert-space Riesz representation theorem provides a function g∈L2(ρ)g\in L^2(\rho) such that

T(h)=∫Xhg dρ(h∈L2(ρ)).T(h)=\int_X hg\,d\rho \qquad(h\in L^2(\rho)).

Since ρ\rho is finite, indicators of all measurable sets belong to L2(ρ)L^2(\rho). Taking h=1Ah=\mathbf1_A yields

ν(A)=∫Ag dρ.\nu(A)=\int_A g\,d\rho.

Positivity of ν\nu implies g≥0g\geq0 almost everywhere, while ν≤ρ\nu\leq\rho implies g≤1g\leq1 almost everywhere. Subtracting the identity from ρ(A)=∫A1 dρ\rho(A)=\int_A1\,d\rho also gives

μ(A)=∫A(1−g) dρ.\mu(A)=\int_A(1-g)\,d\rho.

The remaining issue is division by 1−g1-g. For E={g=1}E=\{g=1\}, the last formula gives μ(E)=0\mu(E)=0. Absolute continuity then gives ν(E)=0\nu(E)=0, and therefore ρ(E)=0\rho(E)=0. Hence

f=g1−gf=\frac{g}{1-g}

is defined and finite ρ\rho-almost everywhere; its values on EE can be set arbitrarily. Applying the elementary integration identity for the already established measure μ=(1−g)ρ\mu=(1-g)\rho gives

∫Af dμ=∫Af(1−g) dρ=∫Ag dρ=ν(A).\int_A f\,d\mu = \int_A f(1-g)\,d\rho = \int_A g\,d\rho = \nu(A).

This proves existence without assuming a density relative to μ\mu in advance.

For sigma-finite measures, take countable finite-measure covers for μ\mu and ν\nu, intersect their members, and disjointize the resulting cover. This produces a measurable partition X=⋃nEnX=\bigcup_n E_n on which both measures are finite. Apply the finite theorem to each EnE_n and combine the resulting derivatives piecewise. Countable additivity supplies the representation on every measurable set.

Uniqueness follows by comparing integrals over sets where one proposed derivative exceeds the other. On each finite-measure piece, integrals over all measurable subsets agree and are finite. If f≥g+1/kf\geq g+1/k on a subset of positive μ\mu-measure, their integrals over that subset cannot agree. Taking the union over positive integers kk, and then interchanging ff and gg, proves equality almost everywhere.

Relation to differentiation and absolute continuity of functions

The word “derivative” has a concrete local interpretation in Euclidean spaces, although that interpretation requires additional theorems. Suppose ν\nu is a locally finite Borel measure on Rn\mathbb R^n and ν=fλ\nu=f\lambda, where λ\lambda is Lebesgue measure. The Lebesgue differentiation theorem implies

f(x)=lim⁡r↓0ν(B(x,r))λ(B(x,r))f(x)=\lim_{r\downarrow0} \frac{\nu(B(x,r))}{\lambda(B(x,r))}

for Lebesgue-almost every xx. Indeed, the quotient is the average of ff over the ball B(x,r)B(x,r). Thus, where differentiation applies, the abstract derivative is the limiting ratio of masses in small neighborhoods. Local finiteness ensures that ff is locally integrable.

This interpretation is not part of the abstract Radon–Nikodym theorem. An arbitrary measurable space has no balls or shrinking neighborhoods, and no pointwise limiting procedure is required to obtain the density. Even on Euclidean space, recovering densities by ratios of measures depends on the selected differentiation basis. Balls and suitable families of cubes have the necessary covering properties; arbitrary shrinking measurable sets need not recover a locally integrable function.

On an interval, the theorem is closely connected with absolute continuity of functions. If F:[a,b]→RF:[a,b]\to\mathbb R is nondecreasing and absolutely continuous, its associated Lebesgue–Stieltjes measure is absolutely continuous with respect to length. Consequently,

F(x)−F(a)=∫axf(t) dt,F(x)-F(a)=\int_a^x f(t)\,dt,

and the differentiation theorem identifies ff with the ordinary derivative F′F' almost everywhere. More generally, a real function FF is absolutely continuous exactly when it can be written in this form with an integrable, not necessarily nonnegative, function ff.

This provides the measure-theoretic form of the fundamental theorem of calculus. An absolutely continuous function has an almost-everywhere derivative that is integrable, and integrating that derivative reconstructs its increments. Almost-everywhere differentiability alone is insufficient. The Cantor function is continuous and nondecreasing, has derivative zero almost everywhere, but increases from 00 to 11. Its increase is encoded by a singular measure rather than by an integrable density.

Coordinate changes give another familiar manifestation. If Y=T(X)Y=T(X), where T:Rn→RnT:\mathbb R^n\to\mathbb R^n is a continuously differentiable diffeomorphism and XX has density pXp_X, then

pY(y)=pX(T−1(y))∣det⁡DT−1(y)∣.p_Y(y)=p_X(T^{-1}(y)) \left|\det DT^{-1}(y)\right|.

The determinant of the Jacobian matrix converts the reference volume under the transformation. The resulting expression is the Radon–Nikodym derivative of the distribution of YY with respect to Lebesgue measure in the new coordinates. This connects the theorem with classical substitution while showing why a density changes when its reference volume changes.

Conditional expectation and change of probability measure

On a probability space (Ω,F,P)(\Omega,\mathcal F,P), let G⊆F\mathcal G\subseteq\mathcal F be a sub-sigma-algebra and let XX be a nonnegative integrable random variable. Define a finite measure on G\mathcal G by

ν(A)=∫AX dP,A∈G.\nu(A)=\int_A X\,dP, \qquad A\in\mathcal G.

It is absolutely continuous with respect to the restriction P∣GP|_{\mathcal G}. The Radon–Nikodym theorem therefore supplies a G\mathcal G-measurable function YY such that

∫AY dP=∫AX dP(A∈G).\int_A Y\,dP=\int_A X\,dP \qquad(A\in\mathcal G).

This is the defining property of the conditional expectation EP[X∣G]E_P[X\mid\mathcal G]. General integrable real variables are handled by separating positive and negative parts.

This construction explains both existence and almost-sure uniqueness. It also avoids defining conditioning solely through ratios of probabilities, which fail when the conditioning event has probability zero. For an event BB, the conditional probability P(B∣G)P(B\mid\mathcal G) is the conditional expectation of 1B\mathbf1_B. The construction gives a version for each individual event; obtaining a probability kernel that behaves simultaneously as a measure in BB requires additional regularity, such as standard Borel structure in common formulations.

For two probability measures Q≪PQ\ll P, the density

Z=dQdPZ=\frac{dQ}{dP}

satisfies Z≥0Z\geq0 and EP[Z]=1E_P[Z]=1. Conversely, every such random variable defines a probability measure by Q(A)=EP[Z1A]Q(A)=E_P[Z\mathbf1_A]. Expectations transform according to

EQ[X]=EP[ZX]E_Q[X]=E_P[ZX]

whenever the quantities are defined. Thus changing probability measure reweights outcomes rather than changing the measurable events available in the model.

Conditional expectations transform as well. For QQ-integrable XX,

EQ[X∣G]=EP[ZX∣G]EP[Z∣G]E_Q[X\mid\mathcal G] = \frac{E_P[ZX\mid\mathcal G]} {E_P[Z\mid\mathcal G]}

on the set where the denominator is positive. Its complement is QQ-null, so the quotient may be assigned any suitable value there. The numerator is integrable under PP because EP[Z∣X∣]=EQ[∣X∣]E_P[Z|X|]=E_Q[|X|]. This identity is an abstract conditional change-of-measure formula, encompassing many familiar reweighting identities.

If (Ft)(\mathcal F_t) is a filtration, then

Zt=EP[Z∣Ft]Z_t=E_P[Z\mid\mathcal F_t]

is the derivative of Q∣FtQ|_{\mathcal F_t} with respect to P∣FtP|_{\mathcal F_t}. The tower property makes (Zt)(Z_t) a nonnegative martingale. Density processes of this type are central to changes of measure in stochastic analysis. In particular, Girsanov’s theorem specifies conditions under which an exponential density process changes the drift of a Wiener process while preserving an appropriate Brownian structure under the new measure.

Likelihood ratios, statistical inference, and information

A statistical model is dominated when its probability measures {Pθ}\{P_\theta\} are all absolutely continuous with respect to a common sigma-finite measure μ\mu. The densities

pθ=dPθdμp_\theta=\frac{dP_\theta}{d\mu}

provide the likelihood function θ↦pθ(x)\theta\mapsto p_\theta(x) for an observed value xx. This framework treats discrete, continuous, and many mixed models uniformly. The dominating measure need not itself be a probability measure, and it need not be Lebesgue measure.

For two simple hypotheses P0P_0 and P1P_1, the measure μ=P0+P1\mu=P_0+P_1 always dominates both. Their densities p0,p1p_0,p_1 define a likelihood ratio p1/p0p_1/p_0, with an infinite value where p0=0<p1p_0=0<p_1. If P1≪P0P_1\ll P_0, this ratio is dP1/dP0dP_1/dP_0, up to the appropriate null sets. The Neyman–Pearson lemma characterizes most powerful tests at a prescribed size by thresholding the likelihood ratio, with randomization on a threshold set when necessary. The measure-theoretic formulation accommodates support differences that a naive ratio of ordinary densities can obscure.

In Bayesian inference, Bayes’ formula is also a density statement. Suppose a model has sampling density p(x∣θ)p(x\mid\theta) and prior probability measure Π\Pi. When the normalizing integral is positive and finite, the posterior has derivative

dΠ( ⋅∣x)dΠ(θ)=p(x∣θ)∫p(x∣ϑ) Π(dϑ).\frac{d\Pi(\,\cdot\mid x)}{d\Pi}(\theta) = \frac{p(x\mid\theta)} {\int p(x\mid\vartheta)\,\Pi(d\vartheta)}.

The reference measure here is the prior itself, not necessarily length on parameter space. The formula therefore includes discrete priors, continuous priors, and priors with mixed components. In regular probabilistic formulations, the posterior identity holds for almost every observation under the marginal distribution.

Radon–Nikodym derivatives also express information divergences without selecting coordinates. If P≪QP\ll Q, the Kullback–Leibler divergence is

DKL(P∥Q)=∫log⁡(dPdQ)dP=∫flog⁡f dQ,f=dPdQ,D_{\mathrm{KL}}(P\|Q) = \int \log\left(\frac{dP}{dQ}\right)dP = \int f\log f\,dQ, \qquad f=\frac{dP}{dQ},

with 0log⁡0=00\log0=0. Its value may be +∞+\infty; if P≪̸QP\not\ll Q, it is defined to be +∞+\infty. The logarithmic density ratio measures how probabilities assigned by the two measures differ, and its expectation yields the divergence.

In importance sampling, the same derivative appears as a simulation weight:

EP[h]=EQ ⁣[hdPdQ],P≪Q.E_P[h]=E_Q\!\left[h\frac{dP}{dQ}\right], \qquad P\ll Q.

For independent samples from QQ, averaging the weighted values estimates the target expectation. Absolute continuity ensures that no set with positive target probability is entirely missed by the proposal measure. It does not ensure a finite estimator variance: that additionally requires integrability of h2(dP/dQ)2h^2(dP/dQ)^2 under QQ. The distinction between existence of a density and control of its magnitude is therefore important computationally.

Signed measures, functional analysis, and vector-valued extensions

The theorem extends to finite signed measures. If a finite signed measure ν\nu is absolutely continuous with respect to a positive sigma-finite measure μ\mu, then

ν(A)=∫Af dμ\nu(A)=\int_A f\,d\mu

for a real integrable function ff. Applying the positive theorem to the positive and negative parts of ν\nu proves the assertion. Finite complex measures similarly have complex-valued integrable derivatives. These extensions allow representation of quantities that may change sign or phase, rather than only positive mass.

The total variation measure records the size of a signed or complex measure. When dν=f dμd\nu=f\,d\mu,

d∣ν∣=∣f∣ dμ,∣ν∣(X)=∫X∣f∣ dμ.d|\nu|=|f|\,d\mu, \qquad |\nu|(X)=\int_X|f|\,d\mu.

In particular, relative to its own total variation, a finite complex measure has a derivative uu satisfying ∣u∣=1|u|=1 almost everywhere with respect to ∣ν∣|\nu|, except in the vacuous zero-measure case. This gives the polar representation dν=u d∣ν∣d\nu=u\,d|\nu|, analogous to separating magnitude from sign or complex phase.

In functional analysis, the theorem underlies the duality of LpL^p spaces. On a sigma-finite measure space, for 1≤p<∞1\leq p<\infty, every bounded linear functional on Lp(μ)L^p(\mu) can be represented as

T(h)=∫Xhg dμ,T(h)=\int_X hg\,d\mu,

where g∈Lq(μ)g\in L^q(\mu) and 1/p+1/q=11/p+1/q=1, with q=∞q=\infty when p=1p=1. For complex spaces, a conjugate may be placed on gg according to convention. A standard proof constructs a measure from the functional’s values on indicators of finite-measure sets, obtains its Radon–Nikodym derivative, and uses the operator bound to establish the required LqL^q estimate. The analogous claim for the full dual of L∞L^\infty is not generally valid.

Vector-valued measures reveal a genuine limit of the scalar theorem. A Banach space BB has the Radon–Nikodym property if every countably additive BB-valued measure of bounded variation, absolutely continuous with respect to a finite positive measure μ\mu, admits a Bochner-integrable density f:X→Bf:X\to B. Finite-dimensional spaces and reflexive Banach spaces have this property, but not every Banach space does.

For example, on [0,1][0,1], the vector measure

m(A)=1A∈L1([0,1])m(A)=\mathbf1_A\in L^1([0,1])

is countably additive in norm, has variation equal to Lebesgue measure, and is absolutely continuous, yet has no Bochner-integrable density. Integrating such a proposed density over AA, then testing against indicators of intervals in the output variable, would force almost every density value to behave like a point mass rather than an L1L^1 function. Thus the scalar theorem does not extend to arbitrary infinite-dimensional ranges merely by replacing real numbers with vectors.

Historical development and the role of the hypotheses

Johann Radon’s 1913 work developed integral representations and measure-theoretic methods in Euclidean settings. Otton Nikodym’s 1930 generalization established the abstract measure-space form associated with the modern theorem. The development belongs to the transition from measures tied to geometric sets toward measures defined on abstract sigma-algebras. In that setting, integration became a tool for representing set functions without relying on coordinates or ordinary differentiation.

Several proof traditions reflect the theorem’s connections with other parts of analysis. Direct constructions build increasingly large measurable functions whose integrals remain bounded by the target measure, and then show that no nonzero residual measure can remain. Hilbert-space proofs instead represent a bounded linear functional after introducing an auxiliary finite measure. Decomposition proofs separate absolutely continuous and singular parts before identifying the density. These methods emphasize different structures, but all must establish both a representative function and its uniqueness modulo null sets.

The sigma-finiteness assumption is substantive, not merely technical decoration. Let μ\mu be counting measure on the Borel subsets of [0,1][0,1], and let ν\nu be Lebesgue measure there. Since the only μ\mu-null set is the empty set, ν≪μ\nu\ll\mu. Nevertheless, no Radon–Nikodym derivative exists. If ν(A)=∫Af dμ\nu(A)=\int_A f\,d\mu, applying the identity to every singleton would give

f(x)=ν({x})=0(x∈[0,1]).f(x)=\nu(\{x\})=0 \qquad(x\in[0,1]).

Then ff would vanish everywhere, contradicting ν([0,1])=1\nu([0,1])=1. Counting measure on this uncountable space is not sigma-finite, so the example does not contradict the standard theorem.

The target measure’s finiteness assumptions also affect the derivative’s behavior. A measure defined by ν(A)=0\nu(A)=0 when μ(A)=0\mu(A)=0 and ν(A)=+∞\nu(A)=+\infty otherwise is absolutely continuous with respect to μ\mu, but its density, when μ\mu is nonzero, is the extended-valued function +∞+\infty almost everywhere. Thus the finite-almost-everywhere conclusion cannot simply be retained for unrestricted target measures. More general representation theorems replace sigma-finiteness by conditions such as localizability, but those conditions require their own precise formulations.

Finally, the theorem guarantees neither smoothness nor a computable pointwise formula. A density may be discontinuous everywhere, unbounded, or specified only through an existence argument. It need not be unique at individual points, and a probability density is not the probability of a singleton in a nonatomic space. These distinctions are especially important when abstract densities are used in statistical models, where additional choices of versions, regularity assumptions, or dominating measures carry information not supplied by the theorem itself.