Signed Measures & the Radon-Nikodym Theorem

Why Radon-Nikodym? Signed Measures, Hahn & Jordan Decompositions Absolute Continuity & Singularity The Radon-Nikodym Theorem Looking Ahead

Why Radon-Nikodym?

Three earlier pages have already used the name "Radon-Nikodym" without proof. In Pushforward Measure, we defined the probability density function as the formal derivative \(f_X = dP_X/d\lambda\) and granted its existence whenever \(P_X \ll \lambda\). In Dual Spaces & Riesz Representation, the duality \((L^p)^* \cong L^q\) was stated as a fact, with the proof of the converse direction — every continuous functional on \(L^p\) arises from some \(L^q\) function — left to a future result. And in the closing paragraphs of Limit Theorems & Product Measures, we previewed the conditional expectation \(\mathbb{E}[X \mid \mathcal{G}]\) as the Radon-Nikodym derivative of a signed measure with respect to \(\mathbb{P}|_{\mathcal{G}}\). All three invocations rest on a single theorem, the proof of which we provide in this chapter.

The theorem answers a sharply posed question: given two measures \(\mu\) and \(\nu\) on the same space, when can \(\nu\) be written as an integral against \(\mu\)? That is, when does there exist a measurable function \(f\) such that \[ \nu(A) \;=\; \int_A f \, d\mu \quad \text{for every measurable set } A \, ? \] The answer turns on a single hypothesis — absolute continuity, written \(\nu \ll \mu\) — and the function \(f\) it produces is the density of \(\nu\) with respect to \(\mu\), denoted \(d\nu/d\mu\). To formulate and prove this, we first need a notion of measure that admits negative values (the assignment \(A \mapsto \int_A f \, d\mu\) for an integrable real-valued \(f\), and the difference \(\nu_1 - \nu_2\) of two measures provided at least one is finite, are central examples). This leads us to signed measures and the Hahn and Jordan decompositions, which split any signed measure into a positive and a negative part. With those in hand, the Radon-Nikodym theorem follows from a celebrated Hilbert-space argument due to von Neumann — recovering, on the Section III side, the dividends of the entire functional-analytic machinery developed in Section II.

The applications in machine learning are pervasive and immediate: Bayesian posterior densities, the score functions driving diffusion-based generative models, importance-sampling reweights, and the KL divergences that regularize policy training in RLHF are all, at the foundational level, Radon-Nikodym derivatives or integrals against them. We will revisit each of these in detail once the theorem is in hand; for now, the message is that the abstract derivative \(d\nu/d\mu\) is the working object underneath nearly every probabilistic construction in modern ML.

Signed Measures, Hahn & Jordan Decompositions

A measure, as introduced in Measure Theory, assigns a non-negative value to each measurable set and is countably additive on disjoint unions. Many natural constructions, however, produce set functions that take both signs. The simplest example: given a measure \(\mu\) and an integrable function \(f \in L^1(\mu)\) that takes both positive and negative values, the assignment \[ \nu_f(A) \;=\; \int_A f \, d\mu \] is countably additive and finite, but \(\nu_f(A)\) can be negative. Differences of two measures, \(\nu = \mu_1 - \mu_2\), behave the same way. To work with such objects on equal footing with measures, we relax the non-negativity axiom while keeping countable additivity intact.

Definition

Definition: Signed Measure

Let \((\Omega, \mathcal{F})\) be a measurable space. A signed measure on \((\Omega, \mathcal{F})\) is a function \(\nu : \mathcal{F} \to [-\infty, \infty]\) satisfying:

  1. \(\nu(\emptyset) = 0\).
  2. \(\nu\) takes at most one of the values \(+\infty\) and \(-\infty\) (so the countable sum in (3) is always well-defined, with no \(\infty - \infty\) ambiguity).
  3. Countable additivity: for any sequence \((A_n)_{n \geq 1}\) of pairwise disjoint sets in \(\mathcal{F}\), \[ \nu\!\left(\bigsqcup_{n=1}^\infty A_n\right) \;=\; \sum_{n=1}^\infty \nu(A_n), \] where the series \(\sum_n \nu(A_n)\) is required to converge absolutely whenever \(\nu(\bigsqcup_n A_n)\) is finite. (When the left-hand side equals \(+\infty\) or \(-\infty\), condition (2) forces all but finitely many terms of the series to share the same sign on the divergent side, so the partial sums diverge unambiguously to the same value.)

The absolute-convergence requirement in (3) deserves a brief comment. When \(\nu(\bigsqcup_n A_n)\) is finite, the left-hand side does not depend on the order in which the disjoint sets \(A_n\) are listed, so the series on the right must converge to the same value under every rearrangement. By Riemann's rearrangement theorem on real series, this is equivalent to absolute convergence. The condition is automatic when \(\nu\) is non-negative — it is only the possibility of cancellation between positive and negative terms that makes it substantive in the signed case.

The example \(\nu_f(A) = \int_A f \, d\mu\) is the prototype. Splitting \(f = f^+ - f^-\) into its positive and negative parts gives \[ \nu_f(A) \;=\; \int_A f^+ \, d\mu \;-\; \int_A f^- \, d\mu \;=\; \nu_{f^+}(A) - \nu_{f^-}(A), \] which writes \(\nu_f\) as the difference of two non-negative measures, supported respectively on \(\{f \geq 0\}\) and \(\{f < 0\}\). The Jordan decomposition asserts that every signed measure admits such a structural splitting canonically, independently of any representing function \(f\).

The Hahn and Jordan Decompositions

The geometric picture is direct: a signed measure \(\nu\) on \(\Omega\) carves the space into a part \(P\) on which \(\nu\) is non-negative and a complementary part \(N\) on which \(\nu\) is non-positive. Once such a decomposition of \(\Omega\) is found, the splitting of \(\nu\) into a non-negative and a non-positive piece comes for free.

A measurable set \(P \in \mathcal{F}\) is called positive for \(\nu\) if \(\nu(A) \geq 0\) for every measurable \(A \subseteq P\); similarly, \(N\) is negative if \(\nu(A) \leq 0\) for every measurable \(A \subseteq N\). It is not enough merely that \(\nu(P) \geq 0\): every measurable subset of \(P\) must inherit non-negativity.

Theorem: Hahn Decomposition

Let \(\nu\) be a signed measure on \((\Omega, \mathcal{F})\). Then there exist disjoint measurable sets \(P, N \in \mathcal{F}\) with \(\Omega = P \sqcup N\), where \(P\) is positive and \(N\) is negative for \(\nu\). The decomposition is unique up to \(\nu\)-null sets: if \(\Omega = P' \sqcup N'\) is another such decomposition, then the symmetric differences \(P \triangle P'\) and \(N \triangle N'\) are \(\nu\)-null in the strong sense (every measurable subset has \(\nu\)-measure zero).

Proof:

We assume without loss of generality that \(\nu\) does not take the value \(+\infty\); the other case is symmetric. The strategy is to extract a positive set of maximal measure and verify that its complement is negative.

Step 1 (subset extraction). We first show: if \(E \in \mathcal{F}\) satisfies \(0 < \nu(E) < \infty\), then \(E\) contains a positive set \(A \subseteq E\) with \(\nu(A) > 0\).

Suppose for contradiction that every measurable \(A \subseteq E\) with \(\nu(A) > 0\) fails to be positive — that is, contains a measurable subset of strictly negative \(\nu\)-value. Set \(E_0 = E\). Let \(n_1\) be the smallest positive integer such that there exists a measurable \(B_1 \subseteq E_0\) with \(\nu(B_1) < -1/n_1\); such an \(n_1\) exists by the contradiction hypothesis. Set \(E_1 = E_0 \setminus B_1\). Inductively, given \(E_{k-1}\), let \(n_k\) be the smallest positive integer such that there exists \(B_k \subseteq E_{k-1}\) with \(\nu(B_k) < -1/n_k\), and set \(E_k = E_{k-1} \setminus B_k\). Continue while such an \(n_k\) exists.

If at some stage no such \(n_k\) exists, then no measurable subset of \(E_{k-1}\) has strictly negative \(\nu\)-measure, so \(E_{k-1}\) is itself a positive set; moreover \[ \nu(E_{k-1}) \;=\; \nu(E) - \sum_{j < k} \nu(B_j) \;=\; \nu(E) + \sum_{j < k} |\nu(B_j)| \;\geq\; \nu(E) \;>\; 0 \] (each \(\nu(B_j) < 0\), so \(-\nu(B_j) = |\nu(B_j)| \geq 0\)), and the claim holds with \(A = E_{k-1}\). Otherwise, the construction continues for all \(k \geq 1\), and we proceed as follows.

Define \(E_\infty = E \setminus \bigsqcup_{k \geq 1} B_k\), so that \(E = E_\infty \sqcup \bigsqcup_{k \geq 1} B_k\). By countable additivity, \[ \nu(E) \;=\; \nu(E_\infty) + \sum_{k \geq 1} \nu(B_k). \] Since \(\nu(E)\) is finite and \(\nu\) does not take the value \(+\infty\) (by the WLOG assumption), the equation forces both \(\nu(E_\infty)\) and \(\sum_k \nu(B_k)\) to be finite — for if \(\nu(E_\infty) = -\infty\), the right-hand side would be \(-\infty \neq \nu(E)\). With the left-hand side finite, condition (3) of the signed-measure definition gives absolute convergence of \(\sum_k \nu(B_k)\). Each \(\nu(B_k) < -1/n_k < 0\), so \(\sum_k 1/n_k < \infty\), forcing \(n_k \to \infty\). Moreover, \(\nu(E_\infty) = \nu(E) - \sum_k \nu(B_k) \geq \nu(E) > 0\), so in particular \(\nu(E_\infty) > 0\).

We claim \(E_\infty\) is positive. If not, there exists \(C \subseteq E_\infty\) with \(\nu(C) < 0\); choose \(m \in \mathbb{N}\) with \(-1/m > \nu(C)\) (possible since \(\nu(C) < 0\)). For all sufficiently large \(k\), \(n_k > m\), so by minimality of \(n_k\), no measurable subset of \(E_{k-1}\) has \(\nu\)-measure \(< -1/m\). But \(C \subseteq E_\infty \subseteq E_{k-1}\) and \(\nu(C) < -1/m\), a contradiction. Hence \(E_\infty\) is positive with \(\nu(E_\infty) > 0\), proving the claim with \(A := E_\infty\).

Step 2 (maximization). Let \[ s \;=\; \sup\bigl\{\, \nu(P) \,:\, P \in \mathcal{F} \text{ is positive for } \nu \,\bigr\} \;\in\; [0, +\infty]. \] The supremum is over a non-empty family (the empty set is positive with measure \(0\)). The finiteness \(s < \infty\) is not yet established; it will follow at the end of this step from \(\nu(P) = s\) and the WLOG assumption \(\nu < +\infty\). Choose positive sets \(P_n\) with \(\nu(P_n) \to s\), and set \(P = \bigcup_n P_n\).

Each finite union \(P_1 \cup \cdots \cup P_n\) is positive, since a measurable subset of a finite union of positive sets can be partitioned into measurable pieces, each contained in some \(P_i\), and a sum of non-negative numbers is non-negative. Apply continuity from below to the increasing sequence of positive sets \(Q_n := P_1 \cup \cdots \cup P_n \nearrow P\) (this is a special case of countable additivity applied to the disjoint sequence \(P_1, P_2 \setminus P_1, P_3 \setminus (P_1 \cup P_2), \ldots\); the \(\nu\)-values \(\nu(Q_n)\) lie in \([0, s]\), so all quantities are non-negative reals or \(+\infty\) and no \(\infty - \infty\) ambiguity arises). Then every measurable subset \(A \subseteq P\) satisfies \(\nu(A) = \lim_n \nu(A \cap Q_n) \geq 0\) (each term is the \(\nu\)-measure of a measurable subset of the positive set \(Q_n\)), so \(P\) is positive. Moreover, since \(P\) is positive and \(P_n \subseteq P\), the set \(P \setminus P_n\) is a measurable subset of \(P\), hence has non-negative \(\nu\)-measure, giving \(\nu(P) \geq \nu(P_n)\) for each \(n\); letting \(n \to \infty\) yields \(\nu(P) \geq s\), and since \(\nu(P)\) is a candidate in the supremum, \(\nu(P) = s\). Finally, the WLOG assumption \(\nu < +\infty\) gives \(\nu(P) < +\infty\), confirming \(s < \infty\).

Step 3 (complement is negative). Set \(N = \Omega \setminus P\). Suppose for contradiction \(N\) is not negative: there exists \(E \subseteq N\) with \(\nu(E) > 0\). Since \(\nu(E)\) is finite (as \(\nu < +\infty\)), Step 1 produces a positive set \(A \subseteq E\) with \(\nu(A) > 0\). Then \(P \cup A\) is positive, disjointly assembled, with \(\nu(P \cup A) = \nu(P) + \nu(A) = s + \nu(A) > s\), contradicting the definition of \(s\). Hence \(N\) is negative.

Uniqueness. Let \(\Omega = P' \sqcup N'\) be another Hahn decomposition. The set \(P \setminus P' = P \cap N'\) is a subset of the positive set \(P\) and of the negative set \(N'\), so every measurable \(B \subseteq P \setminus P'\) satisfies both \(\nu(B) \geq 0\) and \(\nu(B) \leq 0\), forcing \(\nu(B) = 0\). Thus \(P \setminus P'\) is \(\nu\)-null in the strong sense (every measurable subset has \(\nu\)-measure zero); symmetrically \(P' \setminus P\) is \(\nu\)-null. Hence \(P \triangle P'\) is \(\nu\)-null, and likewise \(N \triangle N'\). \(\blacksquare\)

The Hahn decomposition produces a partition of the underlying space; the Jordan decomposition repackages this as an intrinsic splitting of the measure itself.

Theorem: Jordan Decomposition

Every signed measure \(\nu\) on \((\Omega, \mathcal{F})\) decomposes uniquely as \[ \nu \;=\; \nu^+ - \nu^-, \] where \(\nu^+\) and \(\nu^-\) are non-negative measures and \(\nu^+ \perp \nu^-\) — that is, \(\nu^+\) and \(\nu^-\) are concentrated on disjoint measurable sets.

Proof:

Let \(\Omega = P \sqcup N\) be a Hahn decomposition for \(\nu\). Define \[ \nu^+(A) \;=\; \nu(A \cap P), \qquad \nu^-(A) \;=\; -\,\nu(A \cap N), \qquad A \in \mathcal{F}. \] Since \(P\) is positive and \(N\) is negative, both \(\nu^+\) and \(\nu^-\) are non-negative, and countable additivity of \(\nu\) transfers immediately. For every \(A \in \mathcal{F}\), \[ \nu^+(A) - \nu^-(A) \;=\; \nu(A \cap P) + \nu(A \cap N) \;=\; \nu(A), \] proving the existence of the decomposition. By construction \(\nu^+(N) = 0\) and \(\nu^-(P) = 0\), so \(\nu^+\) and \(\nu^-\) are concentrated on disjoint sets and hence mutually singular.

For uniqueness, suppose \(\nu = \mu_1 - \mu_2\) with \(\mu_1, \mu_2\) non-negative measures concentrated on disjoint measurable sets \(P', N' \in \mathcal{F}\) with \(\Omega = P' \sqcup N'\) (so \(\mu_1(N') = 0\) and \(\mu_2(P') = 0\)).

Step (a): \(\Omega = P' \sqcup N'\) is a Hahn decomposition for \(\nu\). For \(A \subseteq P'\), \(\mu_2(A) \leq \mu_2(P') = 0\) (since \(\mu_2\) is concentrated on \(N'\), disjoint from \(P'\)), so \(\nu(A) = \mu_1(A) \geq 0\); thus \(P'\) is positive. Symmetrically, \(N'\) is negative. By Hahn uniqueness, the symmetric difference \(P \triangle P' = (P \setminus P') \sqcup (P' \setminus P)\) is \(\nu\)-null in the strong sense — every measurable subset has \(\nu\)-measure zero.

Step (b): \(\mu_1(P \triangle P') = 0\) and \(\mu_2(P \triangle P') = 0\). For \(P' \setminus P \subseteq P'\), the concentration of \(\mu_2\) on \(N'\) gives \(\mu_2(P' \setminus P) = 0\), so \(\nu(P' \setminus P) = \mu_1(P' \setminus P)\); combined with the strong \(\nu\)-nullity from Step (a), \(\mu_1(P' \setminus P) = 0\). For \(P \setminus P' \subseteq N'\) (since \(P \setminus P' \subseteq \Omega \setminus P' = N'\)), the concentration of \(\mu_1\) on \(P'\) gives \(\mu_1(P \setminus P') = 0\) directly; then \(\nu(P \setminus P') = -\mu_2(P \setminus P')\), and the strong \(\nu\)-nullity forces \(\mu_2(P \setminus P') = 0\). Combining, \(\mu_1(P \triangle P') = 0\) and \(\mu_2(P \triangle P') = 0\).

Step (c): \(\mu_1 = \nu^+\) and \(\mu_2 = \nu^-\). Fix \(A \in \mathcal{F}\). Since \(\mu_1\) is concentrated on \(P' = (P \cap P') \sqcup (P' \setminus P)\), \[ \mu_1(A) \;=\; \mu_1(A \cap P') \;=\; \mu_1(A \cap P \cap P') + \mu_1(A \cap (P' \setminus P)) \;=\; \mu_1(A \cap P \cap P'), \] where the last equality uses \(\mu_1(P' \setminus P) = 0\) from Step (b). Similarly, splitting \(P = (P \cap P') \sqcup (P \setminus P')\), \[ \mu_1(A \cap P) \;=\; \mu_1(A \cap P \cap P') + \mu_1(A \cap (P \setminus P')) \;=\; \mu_1(A \cap P \cap P'), \] using \(\mu_1(P \setminus P') = 0\). Hence \(\mu_1(A) = \mu_1(A \cap P)\).

On the other hand, from \(\nu = \mu_1 - \mu_2\), \[ \nu^+(A) \;=\; \nu(A \cap P) \;=\; \mu_1(A \cap P) - \mu_2(A \cap P). \] We claim \(\mu_2(A \cap P) = 0\): split \(A \cap P = (A \cap P \cap P') \sqcup (A \cap (P \setminus P'))\); the first piece satisfies \(\mu_2(A \cap P \cap P') \leq \mu_2(P') = 0\), and the second satisfies \(\mu_2(A \cap (P \setminus P')) \leq \mu_2(P \setminus P') = 0\) by Step (b). Thus \(\nu^+(A) = \mu_1(A \cap P) = \mu_1(A)\). The identity \(\mu_2 = \nu^-\) follows symmetrically. \(\blacksquare\)

The Hahn decomposition is unique only up to \(\nu\)-null sets, but the Jordan decomposition itself is fully unique — the ambiguity in choosing \(P\) versus \(P'\) is invisible from the perspective of \(\nu^+\) and \(\nu^-\), which only see how \(\nu\) acts on sets, not which Hahn-partition was used to construct them.

Definition: Total Variation

The total variation of a signed measure \(\nu\) is the non-negative measure \[ |\nu| \;=\; \nu^+ + \nu^-. \] We say \(\nu\) is a finite signed measure if \(|\nu|(\Omega) < \infty\), and \(\sigma\)-finite if \(\Omega\) is a countable union of sets of finite \(|\nu|\)-measure.

The total variation \(|\nu|\) is the natural "size" of a signed measure: \(|\nu|(A)\) measures the total mass of \(\nu\) on \(A\) without cancellation between the positive and negative parts. For \(\nu_f(A) = \int_A f \, d\mu\) with \(f \in L^1(\mu)\), the Hahn decomposition is given explicitly by \(P = \{f \geq 0\}\) and \(N = \{f < 0\}\), so \(\nu_f^+(A) = \int_A f^+ \, d\mu\) and \(\nu_f^-(A) = \int_A f^- \, d\mu\), giving \(|\nu_f|(A) = \int_A |f| \, d\mu\). In particular \(|\nu_f|(\Omega) = \|f\|_{L^1(\mu)}\). The map \(f \mapsto \nu_f\) is therefore a norm-preserving linear embedding of \(L^1(\mu)\) into the space of finite signed measures absolutely continuous with respect to \(\mu\) (in the sense introduced in the next section). The Radon-Nikodym theorem proved at the end of this chapter will upgrade this embedding to an isometric isomorphism by establishing surjectivity — every such \(\mu\)-AC signed measure is of the form \(\nu_f\) for some \(f \in L^1(\mu)\).

Signed Measures in CS and ML

Signed measures appear wherever a system carries net-flow or signed-mass quantities rather than mass alone.

Optimal transport via Kantorovich-Rubinstein duality. In the Earth Mover's formulation of optimal transport, the difference \(P - Q\) of two probability distributions is a signed measure of total mass zero. The Kantorovich-Rubinstein duality writes the Wasserstein-1 distance as a supremum over Lipschitz functions of integrals against this signed measure: \(W_1(P, Q) = \sup_{f \in \mathrm{Lip}_1} \int f \, d(P - Q)\). The Jordan decomposition \(P - Q = (P - Q)^+ - (P - Q)^-\) identifies the regions of mass surplus (sources) and deficit (sinks) between the two distributions — the regions from which mass must flow and to which it must arrive in any optimal transport plan.

Log-likelihood ratios. In hypothesis testing and classification, when \(\ell \in L^1(P_0)\) (i.e., \(\mathbb{E}_{P_0}[|\log(p_1/p_0)|] < \infty\)), the function \(\ell(x) = \log\!\bigl(p_1(x)/p_0(x)\bigr)\) defines a finite signed measure \(\nu_\ell(A) = \int_A \ell \, dP_0\). Its Jordan decomposition isolates the regions where evidence favors hypothesis \(H_1\) over \(H_0\) and vice versa, and the total variation \(|\nu_\ell|(\Omega) = \mathbb{E}_{P_0}[|\ell|]\) measures the typical magnitude of the log-evidence under \(H_0\) — a quantity related to but distinct from standard divergence-based measures of test difficulty such as the KL divergence \(D_{\mathrm{KL}}(P_0 \| P_1) = -\mathbb{E}_{P_0}[\ell]\) or the total-variation distance \(\mathrm{TV}(P_0, P_1) = \tfrac{1}{2}\|p_0 - p_1\|_{L^1}\) that governs the minimax Bayes risk in binary hypothesis testing (via Le Cam's identity \(\mathcal{R}_{\min} = \tfrac{1}{2}(1 - \mathrm{TV}(P_0, P_1))\)).

Network flows on graphs. For a flow defined on the edges of a finite graph (with sources and sinks), its divergence — the net flow at each vertex (incoming minus outgoing) — is naturally a signed measure on the vertex set, with \(\nu(\{v\})\) the net flow at vertex \(v\); the Hahn decomposition partitions the vertices into net-source and net-sink subsets. The same algebraic structure underlies divergence operators on simplicial complexes — a connection that resurfaces in the simplicial and homological structures developed in Section IV and ahead toward Geometric Deep Learning.

Absolute Continuity & Singularity

With the structural theory of signed measures in hand, we turn to the relation between two measures on the same space. The Radon-Nikodym theorem will assert that one measure can be expressed as an integral against another precisely when the two satisfy the relation defined here: absolute continuity. The opposite extreme — mutual singularity — describes measures that are concentrated on disjoint sets and have nothing to integrate against one another. Together, these two relations partition the qualitative ways that two measures can interact, and the Lebesgue decomposition (mentioned in the Looking Ahead section) shows that they account for every \(\sigma\)-finite case.

Definitions

Definition: Absolute Continuity

Let \(\mu\) be a non-negative measure and \(\nu\) a signed measure on \((\Omega, \mathcal{F})\). We say \(\nu\) is absolutely continuous with respect to \(\mu\), written \(\nu \ll \mu\), if for every \(A \in \mathcal{F}\), \[ \mu(A) = 0 \;\Longrightarrow\; \nu(A) = 0. \]

The condition \(\nu \ll \mu\) is exactly what is needed for "\(\mu\)-null sets are also \(\nu\)-null sets" — \(\nu\) inherits whatever \(\mu\) declares to be negligible. For signed \(\nu\), the definition is equivalent to \(|\nu| \ll \mu\), and also to the conjunction \(\nu^+ \ll \mu\) and \(\nu^- \ll \mu\). To see (\(\Rightarrow\)): if \(\mu(A) = 0\), then \(\mu(B) = 0\) for every measurable \(B \subseteq A\) (by monotonicity, since \(\mu \geq 0\)); applied to \(B = A \cap P\) and \(B = A \cap N\) from a Hahn decomposition of \(\nu\), this gives \(\nu^+(A) = \nu(A \cap P) = 0\) and \(\nu^-(A) = -\nu(A \cap N) = 0\), hence \(|\nu|(A) = \nu^+(A) + \nu^-(A) = 0\). For (\(\Leftarrow\)): \(\mu(A) = 0 \Rightarrow |\nu|(A) = 0 \Rightarrow |\nu(A)| \leq |\nu|(A) = 0 \Rightarrow \nu(A) = 0\). The Radon-Nikodym theorem will be stated for non-negative \(\nu\); the signed case follows by applying the result to \(\nu^+\) and \(\nu^-\) separately and subtracting.

Definition: Mutual Singularity

Two signed measures \(\mu\) and \(\nu\) on \((\Omega, \mathcal{F})\) are mutually singular, written \(\mu \perp \nu\), if there exists a measurable \(E \in \mathcal{F}\) such that \[ |\mu|(E^c) \;=\; 0 \quad \text{and} \quad |\nu|(E) \;=\; 0. \] Equivalently, \(\mu\) is concentrated on \(E\) and \(\nu\) is concentrated on \(E^c\), so the two measures live on disjoint measurable carriers. (Unlike absolute continuity, mutual singularity is a symmetric relation between two signed measures.)

The two relations are extremes. If \(\nu \ll \mu\), then \(\nu\) is "dominated" by \(\mu\): every \(\mu\)-negligible set is also \(\nu\)-negligible. If \(\nu \perp \mu\), then \(\nu\) and \(\mu\) have no overlap whatsoever. The only signed measure that is simultaneously \(\nu \ll \mu\) and \(\nu \perp \mu\) is the zero measure: from \(\nu \perp \mu\), choose \(E\) with \(|\nu|(E) = 0\) and \(\mu(E^c) = 0\); then \(\nu \ll \mu\) forces \(|\nu|(E^c) = 0\), and so \(|\nu|(\Omega) = 0\), giving \(\nu = 0\).

The ε-δ Characterization

The implication \(\mu(A) = 0 \Rightarrow \nu(A) = 0\) in the definition of absolute continuity is a qualitative, "all-or-nothing" condition: it says that \(\nu\) collapses on \(\mu\)-null sets, but it says nothing about how \(\nu\) behaves on sets of small but positive \(\mu\)-measure. For finite measures, however, the condition tightens: \(\nu(A)\) must in fact be uniformly small whenever \(\mu(A)\) is small. This is the analytic counterpart of the qualitative definition, and it is the form in which absolute continuity most often appears in proofs.

Theorem: ε-δ Characterization of Absolute Continuity

Let \(\mu\) be a non-negative measure and \(\nu\) a finite signed measure on \((\Omega, \mathcal{F})\). Then \(\nu \ll \mu\) if and only if for every \(\varepsilon > 0\) there exists \(\delta > 0\) such that for every \(A \in \mathcal{F}\), \[ \mu(A) < \delta \;\Longrightarrow\; |\nu(A)| < \varepsilon. \]

Proof:

(\(\Leftarrow\)) The ε-δ condition is strictly stronger than \(\nu \ll \mu\): if \(\mu(A) = 0\), then \(\mu(A) < \delta\) for every \(\delta > 0\), so \(|\nu(A)| < \varepsilon\) for every \(\varepsilon > 0\), forcing \(\nu(A) = 0\).

(\(\Rightarrow\)) Suppose \(\nu \ll \mu\) but the ε-δ condition fails. Then there exists \(\varepsilon_0 > 0\) such that for every \(n \geq 1\), some \(A_n \in \mathcal{F}\) satisfies \[ \mu(A_n) < 2^{-n} \quad \text{and} \quad |\nu(A_n)| \geq \varepsilon_0. \] Set \(B_n = \bigcup_{k \geq n} A_k\) and \(B = \bigcap_{n \geq 1} B_n = \limsup_n A_n\). The sets \(B_n\) are decreasing, with \(B_1\) of finite \(|\nu|\)-measure (since \(\nu\) is finite). By \(\sigma\)-subadditivity, \[ \mu(B) \leq \mu(B_n) \leq \sum_{k \geq n} \mu(A_k) < \sum_{k \geq n} 2^{-k} = 2^{-(n-1)}, \] so \(\mu(B) = 0\) on letting \(n \to \infty\). By absolute continuity, \(|\nu|(B) = 0\).

On the other hand, by continuity from above for the finite measure \(|\nu|\), \[ |\nu|(B) \;=\; \lim_{n \to \infty} |\nu|(B_n) \;\geq\; \limsup_{n \to \infty} |\nu|(A_n) \;\geq\; \limsup_{n \to \infty} |\nu(A_n)| \;\geq\; \varepsilon_0, \] where the second inequality uses \(A_n \subseteq B_n\) (so \(|\nu|(A_n) \leq |\nu|(B_n)\) for each \(n\), giving \(\limsup_n |\nu|(A_n) \leq \limsup_n |\nu|(B_n) = \lim_n |\nu|(B_n)\); the last equality holds because \(B_n\) is decreasing) and the third uses \(|\nu(A_n)| \leq |\nu|(A_n)\). This contradicts \(|\nu|(B) = 0\). \(\blacksquare\)

The finiteness hypothesis on \(\nu\) cannot be dropped. Take \(\mu = \lambda\) (Lebesgue measure on \(\mathbb{R}\)) and \(\nu(A) = \int_A x^2 \, d\lambda(x)\), so that \(\nu \ll \lambda\) and \(\nu\) is \(\sigma\)-finite. The sets \(A_n = [n, n + 1/n^2]\) satisfy \(\lambda(A_n) = 1/n^2 \to 0\), but \[ \nu(A_n) \;=\; \int_n^{n + 1/n^2} x^2 \, dx \;\geq\; n^2 \cdot \frac{1}{n^2} \;=\; 1 \] for every \(n\). So \(\nu(A_n)\) does not vanish even as \(\lambda(A_n) \to 0\), and the ε-δ form fails despite \(\nu \ll \lambda\) holding qualitatively.

The Probabilistic Picture

For a probability distribution \(P_X\) on \((\mathbb{R}, \mathcal{B}(\mathbb{R}))\), the relation to Lebesgue measure \(\lambda\) determines the qualitative type of the random variable. If \(P_X \ll \lambda\), the distribution is absolutely continuous in the measure-theoretic sense: it has a Lebesgue density (the Radon-Nikodym derivative \(dP_X/d\lambda\), whose existence we will prove below). If \(P_X \perp \lambda\) — for instance, when \(P_X\) is concentrated on a countable set, or on a fractal of zero Lebesgue measure such as the Cantor set — there is no density relative to length. A mixed distribution, such as \(\tfrac{1}{2}\delta_0 + \tfrac{1}{2}\,\lambda\big|_{[0,1]}\), is neither: it has a singular part (the Dirac mass at \(0\)) and an absolutely continuous part. The Lebesgue decomposition mentioned below asserts that this is the universal pattern: every \(\sigma\)-finite measure splits canonically into an absolutely continuous and a mutually singular piece.

The Radon-Nikodym Theorem

We arrive at the central result. Given two \(\sigma\)-finite measures with \(\nu \ll \mu\), the Radon-Nikodym theorem asserts that \(\nu\) is the integral of some non-negative measurable function against \(\mu\). The function — unique up to \(\mu\)-null sets — is the abstract derivative \(d\nu/d\mu\), and it is the working object underneath every "density" in probability and statistics.

Statement

Theorem: Radon-Nikodym

Let \((\Omega, \mathcal{F})\) be a measurable space, and let \(\mu, \nu\) be \(\sigma\)-finite non-negative measures on \((\Omega, \mathcal{F})\) with \(\nu \ll \mu\). Then there exists a non-negative measurable function \(f : \Omega \to [0, \infty]\), unique up to \(\mu\)-a.e. equality, such that \[ \nu(A) \;=\; \int_A f \, d\mu \quad \text{for every } A \in \mathcal{F}. \] The function \(f\) is called the Radon-Nikodym derivative of \(\nu\) with respect to \(\mu\), denoted \[ f \;=\; \frac{d\nu}{d\mu}. \]

For signed \(\nu\), the result extends by Jordan decomposition: writing \(\nu = \nu^+ - \nu^-\) with \(\nu^\pm \ll \mu\) (which follows from \(\nu \ll \mu\), as observed in the previous section), we obtain \(d\nu/d\mu = d\nu^+/d\mu - d\nu^-/d\mu\), an extended-real measurable function, finite \(\mu\)-a.e. when \(\nu\) is finite.

Proof via von Neumann's L² Method

The proof we present is due to von Neumann and is among the most striking applications of the Hilbert-space machinery developed in Section II. The key idea is to construct \(f\) as the solution of a single linear-algebraic problem in \(L^2\) and then transport it back to a relation between measures. The integral against \(\nu\) defines a bounded linear functional on a suitable \(L^2\) space; the Riesz Representation Theorem delivers a representing function; algebraic manipulation extracts \(f\) from that representative.

Reduction to finite measures. Since both \(\mu\) and \(\nu\) are \(\sigma\)-finite, we can write \(\Omega = \bigsqcup_{n \geq 1} \Omega_n\) with \(\Omega_n\) pairwise disjoint, measurable, and satisfying \(\mu(\Omega_n), \nu(\Omega_n) < \infty\). (Take a \(\mu\)-exhaustion and a \(\nu\)-exhaustion separately, intersect, and disjointify.) If we can produce a Radon-Nikodym derivative \(f_n\) of \(\nu|_{\Omega_n}\) with respect to \(\mu|_{\Omega_n}\) on each \(\Omega_n\), extend each \(f_n\) by zero outside \(\Omega_n\); the function \(f = \sum_n f_n \mathbf{1}_{\Omega_n}\) is measurable and non-negative, and the disjointness of \(\{\Omega_n\}\) together with MCT gives \(\nu(A) = \sum_n \nu(A \cap \Omega_n) = \sum_n \int_{A \cap \Omega_n} f_n \, d\mu = \int_A f \, d\mu\) for every \(A \in \mathcal{F}\). Thus it suffices to prove the theorem when both \(\mu\) and \(\nu\) are finite.

Step 1: The auxiliary measure \(\rho = \mu + \nu\). Assume henceforth that \(\mu\) and \(\nu\) are finite. Define a new measure \[ \rho \;=\; \mu + \nu, \qquad \rho(A) = \mu(A) + \nu(A). \] Then \(\rho\) is finite, and \(\mu \ll \rho\) and \(\nu \ll \rho\) hold trivially (any \(\rho\)-null set has both \(\mu\)- and \(\nu\)-measure zero, since \(\mu, \nu \geq 0\)).

Step 2: A bounded linear functional on \(L^2(\rho)\). Define \[ T : L^2(\rho) \to \mathbb{R}, \qquad T(g) \;=\; \int_\Omega g \, d\nu. \] The functional \(T\) is linear by linearity of the integral. To verify boundedness, apply the Cauchy-Schwarz inequality on \(L^2(\rho)\) using the constant function \(\mathbf{1} \in L^2(\rho)\) (which lies in \(L^2(\rho)\) because \(\rho(\Omega) < \infty\)): \[ |T(g)| \;\leq\; \int |g| \, d\nu \;\leq\; \int |g| \, d\rho \;=\; \int |g| \cdot 1 \, d\rho \;\leq\; \|g\|_{L^2(\rho)} \cdot \rho(\Omega)^{1/2}. \] Here the second inequality uses \(\nu \leq \rho\) (as measures), and the final inequality is Cauchy-Schwarz. Thus \(T\) is a continuous linear functional on the Hilbert space \(L^2(\rho)\), with operator norm at most \(\rho(\Omega)^{1/2}\).

Step 3: Riesz representation. By the Riesz Representation Theorem, there exists a unique \(h \in L^2(\rho)\) such that \[ T(g) \;=\; \langle g, h \rangle_{L^2(\rho)} \;=\; \int_\Omega g \, h \, d\rho \quad \text{for every } g \in L^2(\rho). \] Unwinding the definition of \(T\), this reads \[ \int_\Omega g \, d\nu \;=\; \int_\Omega g \, h \, d\rho \quad \text{for every } g \in L^2(\rho). \tag{$\star$} \] The function \(h\) is so far an abstract \(L^2(\rho)\)-element; the next two steps pin down its geometry.

Step 4: \(0 \leq h \leq 1\) holds \(\rho\)-a.e. We prove the two bounds separately by testing (\(\star\)) against indicator functions.

For the lower bound, let \(E_- = \{h < 0\}\), and substitute \(g = \mathbf{1}_{E_-} \in L^2(\rho)\) (this is in \(L^2(\rho)\) because \(\rho\) is finite). Then \[ \nu(E_-) \;=\; \int_{E_-} 1 \, d\nu \;=\; \int_{E_-} h \, d\rho \;\leq\; 0, \] while \(\nu(E_-) \geq 0\) since \(\nu\) is non-negative. Hence \(\nu(E_-) = 0\) and \(\int_{E_-} h \, d\rho = 0\). But \(h < 0\) strictly on \(E_-\), so \(\int_{E_-} h \, d\rho < 0\) unless \(\rho(E_-) = 0\). Combined with \(\int_{E_-} h \, d\rho = 0\), this forces \(\rho(E_-) = 0\), i.e., \(h \geq 0\) holds \(\rho\)-a.e.

For the upper bound, let \(E_+ = \{h > 1\}\) and substitute \(g = \mathbf{1}_{E_+}\). Then \[ \nu(E_+) \;=\; \int_{E_+} h \, d\rho \;\geq\; \int_{E_+} 1 \, d\rho \;=\; \rho(E_+) \;=\; \mu(E_+) + \nu(E_+). \] Since \(\nu(E_+) \leq \nu(\Omega) < \infty\) (we are in the finite case), the term \(\nu(E_+)\) can be subtracted from both sides, forcing \(\mu(E_+) \leq 0\), hence \(\mu(E_+) = 0\). By absolute continuity \(\nu \ll \mu\), this gives \(\nu(E_+) = 0\), and therefore \(\rho(E_+) = 0\). Hence \(h \leq 1\) holds \(\rho\)-a.e.

We may modify \(h\) on a \(\rho\)-null set without disturbing equation (\(\star\)), so we redefine \(h\) to take values in \([0, 1]\) everywhere.

Step 5: Rewriting (\(\star\)) and isolating \(d\nu/d\mu\). Substitute the definition \(\rho = \mu + \nu\) into (\(\star\)) and use linearity of the integral with respect to the sum measure (\(\int \phi \, d(\mu + \nu) = \int \phi \, d\mu + \int \phi \, d\nu\) for non-negative or \(\rho\)-integrable \(\phi\)): \[ \int_\Omega g \, d\nu \;=\; \int_\Omega g \, h \, d\mu + \int_\Omega g \, h \, d\nu, \] which rearranges to \[ \int_\Omega g \, (1 - h) \, d\nu \;=\; \int_\Omega g \, h \, d\mu \quad \text{for every } g \in L^2(\rho). \tag{$\star\star$} \] We will use (\(\star\star\)) on indicator functions to identify \(d\nu/d\mu\).

Let \(E_1 = \{h = 1\}\). Substituting \(g = \mathbf{1}_{E_1}\) in (\(\star\star\)) gives \(0 = \int_{E_1} h \, d\mu = \mu(E_1)\), so \(\mu(E_1) = 0\). By absolute continuity \(\nu \ll \mu\), also \(\nu(E_1) = 0\). Hence \(h < 1\) holds \(\mu\)-a.e. and \(\nu\)-a.e., and redefining \(h\) to take a value in \([0, 1)\) on the \(\rho\)-null set \(E_1\) leaves (\(\star\star\)) intact.

Set \[ f \;=\; \frac{h}{1 - h}, \] a non-negative measurable function with values in \([0, \infty)\), defined \(\mu\)-a.e. (we set \(f = 0\) on the \(\mu\)-null set \(E_1\) where \(h = 1\) to make \(f\) defined everywhere; the choice is irrelevant). The identity \(h = (1-h) f\) holds \(\mu\)-a.e., and will be used in Step 6 to extract \(f\) as the desired Radon-Nikodym derivative.

Step 6: From (\(\star\star\)) to the Radon-Nikodym identity. We first extend (\(\star\star\)) from \(L^2(\rho)\) to all non-negative measurable \(g\). Both sides of (\(\star\star\)) are integrals of non-negative functions against non-negative measures (since \(0 \leq h \leq 1\) \(\rho\)-a.e., so \(g(1-h) \geq 0\) and \(gh \geq 0\) for \(g \geq 0\)). For indicator functions \(g = \mathbf{1}_A\), (\(\star\star\)) holds since \(\mathbf{1}_A \in L^2(\rho)\); by linearity it holds for non-negative simple \(g\); and by the standard simple-function approximation \(g_n \uparrow g\) and MCT applied to each side, it extends to all non-negative measurable \(g\) (whether or not \(g \in L^2(\rho)\)). Call this extended identity (\(\star\star'\)).

Now apply (\(\star\star'\)) with \(g = \mathbf{1}_A / (1 - h)\) — formally, via the truncations \(g_N := \mathbf{1}_A \cdot \min(N, 1/(1-h))\), each bounded and measurable (hence in \(L^2(\rho)\) since \(\rho\) is finite, and certainly admissible for the extended (\(\star\star'\))). The function \(1/(1-h)\) is undefined on the set \(E_1 = \{h = 1\}\), but \(E_1\) is both \(\mu\)-null and \(\nu\)-null (Step 5), hence \(\rho\)-null, so the value assigned on \(E_1\) (say 0) is irrelevant for both integrals. Substituting \(g_N\) into (\(\star\star'\)): \[ \int_A \min(N(1-h), 1) \, d\nu \;=\; \int_A \min(N, 1/(1-h)) \cdot h \, d\mu. \] On \(\{h < 1\}\), \(\min(N(1-h), 1) \uparrow 1\) and \(\min(N, 1/(1-h)) \cdot h \uparrow h/(1-h) = f\) as \(N \to \infty\). Applying MCT to each side, \[ \nu(A) \;=\; \int_A 1 \, d\nu \;=\; \int_A f \, d\mu \quad \text{for every } A \in \mathcal{F}, \] completing the proof for finite \(\mu, \nu\). The reduction at the start of the proof extends the result to the \(\sigma\)-finite case. \(\blacksquare\)

What Hilbert Space Bought Us

The proof never constructs \(f\) directly. The function \(f = h / (1 - h)\) appears only in the last step; until then, all the work happens with the auxiliary function \(h \in L^2(\rho)\) produced by Riesz representation. This is the strategic content of von Neumann's argument: a problem about existence of a density is converted into a problem about existence of a representing vector for a continuous linear functional on a Hilbert space — a problem already solved in Dual Spaces & Riesz Representation. The entire functional-analytic apparatus of Section II — completeness of \(L^2\) ( Riesz-Fischer), orthogonal projection, the duality \((L^2)^* \cong L^2\) — collapses, in this single proof, into a result about probability densities. Section III thereby harvests, in this single proof, what Section II's chapters on Hilbert spaces, dual functionals, and \(L^p\) completeness developed.

Uniqueness

Uniqueness of the Radon-Nikodym derivative is a short, standalone argument that uses only the integrability of \(f\) and the zero-integral lemma from Section II.

Suppose \(f_1, f_2\) are two non-negative measurable functions, both satisfying \(\nu(A) = \int_A f_i \, d\mu\) for every \(A \in \mathcal{F}\). We first prove uniqueness when \(\nu\) is finite; the \(\sigma\)-finite case follows by applying the finite argument to each \(\Omega_n\) of a \(\sigma\)-finite exhaustion. Assume henceforth \(\nu(\Omega) < \infty\). Then \(\int_\Omega f_i \, d\mu = \nu(\Omega) < \infty\), so each \(f_i\) is \(\mu\)-integrable and therefore finite \(\mu\)-a.e. Setting \(g = f_1 - f_2\) (well-defined \(\mu\)-a.e. and assigned the value 0 on the \(\mu\)-null exceptional set where either \(f_i\) is infinite; the choice is irrelevant), we have \[ \int_A g \, d\mu \;=\; \nu(A) - \nu(A) \;=\; 0 \quad \text{for every } A \in \mathcal{F}. \] Apply this to \(A = \{g > 0\}\) and to \(A = \{g < 0\}\) separately. On the first set, \(g^+ = g\) and \(g^- = 0\), so \(\int g^+ \, d\mu = \int_{\{g > 0\}} g \, d\mu = 0\). Since \(g^+\) is non-negative, the Zero Integral Lemma forces \(g^+ = 0\) \(\mu\)-a.e., that is, \(\mu(\{g > 0\}) = 0\). The argument on \(\{g < 0\}\) is symmetric and gives \(\mu(\{g < 0\}) = 0\). Hence \(g = 0\) \(\mu\)-a.e., that is, \(f_1 = f_2\) \(\mu\)-a.e.

The Radon-Nikodym derivative is therefore well-defined as an element of the equivalence class of measurable functions modulo \(\mu\)-null modifications. In particular, equations of the form "\(d\nu/d\mu = f\)" are always understood up to \(\mu\)-a.e. equality.

Chain Rule and Change of Variables

The Radon-Nikodym derivative is genuinely a derivative: it satisfies a chain rule under composition of absolute-continuity relations, and it intertwines with integration in exactly the way the Leibniz notation \(d\nu/d\mu\) suggests. These two properties — the chain rule and the change-of-variables formula — are the computational engines that make the derivative usable in practice. Every Bayesian update, every importance-sampling estimator, every KL divergence computation rests on the second of these.

Proposition: Chain Rule and Change of Variables

Let \(\mu, \nu, \lambda\) be \(\sigma\)-finite non-negative measures on \((\Omega, \mathcal{F})\).

  1. Chain rule. If \(\nu \ll \mu \ll \lambda\), then \(\nu \ll \lambda\) and \[ \frac{d\nu}{d\lambda} \;=\; \frac{d\nu}{d\mu} \cdot \frac{d\mu}{d\lambda} \quad \lambda\text{-a.e.} \]
  2. Change of variables. If \(\nu \ll \mu\), then for every measurable \(g : \Omega \to [0, \infty]\), \[ \int_\Omega g \, d\nu \;=\; \int_\Omega g \cdot \frac{d\nu}{d\mu} \, d\mu. \] The same identity holds for measurable \(g : \Omega \to \mathbb{R}\) provided either side is finite when \(g\) is replaced by \(|g|\) (i.e., \(g \in L^1(\nu)\) iff \(g \cdot d\nu/d\mu \in L^1(\mu)\), and the integrals agree).
Proof:

We prove (ii) first; (i) follows as a special case.

Proof of (ii). Set \(f = d\nu/d\mu\), so that \(\nu(A) = \int_A f \, d\mu\) for every \(A \in \mathcal{F}\). We extend this from indicator-set integrals to general non-negative \(g\) by the standard three-step procedure.

Step 1: Indicator functions. For \(g = \mathbf{1}_A\), the identity \[ \int_\Omega \mathbf{1}_A \, d\nu \;=\; \nu(A) \;=\; \int_A f \, d\mu \;=\; \int_\Omega \mathbf{1}_A \cdot f \, d\mu \] is the defining property of \(f = d\nu/d\mu\).

Step 2: Non-negative simple functions. Let \(g = \sum_{i=1}^n c_i \mathbf{1}_{A_i}\) with \(c_i \geq 0\) and \(A_i \in \mathcal{F}\) pairwise disjoint. By linearity of the integral on both sides and Step 1, \[ \int_\Omega g \, d\nu \;=\; \sum_{i=1}^n c_i \nu(A_i) \;=\; \sum_{i=1}^n c_i \int_\Omega \mathbf{1}_{A_i} \cdot f \, d\mu \;=\; \int_\Omega g \cdot f \, d\mu. \]

Step 3: Non-negative measurable functions. Let \(g : \Omega \to [0, \infty]\) be measurable. By the standard simple-function approximation, there exists an increasing sequence \(g_n \uparrow g\) of non-negative simple functions. Since \(f \geq 0\), the sequence \(g_n f \uparrow g f\) is also increasing pointwise. Apply MCT on both sides: \[ \int_\Omega g \, d\nu \;=\; \lim_{n \to \infty} \int_\Omega g_n \, d\nu \;=\; \lim_{n \to \infty} \int_\Omega g_n \, f \, d\mu \;=\; \int_\Omega g \, f \, d\mu, \] where the middle equality is Step 2 applied to each \(g_n\). This proves (ii) for non-negative \(g\).

Step 4: Real-valued case. For measurable \(g : \Omega \to \mathbb{R}\), apply Step 3 separately to \(g^+\) and \(g^-\) and subtract. Both \(\int g^+ \, d\nu = \int g^+ f \, d\mu\) and \(\int g^- \, d\nu = \int g^- f \, d\mu\) hold, and the integrability hypothesis \(g \in L^1(\nu)\) (i.e., \(\int |g| d\nu < \infty\)) makes both finite by the non-negative identity applied to \(|g|\). The subtraction then yields \[ \int_\Omega g \, d\nu \;=\; \int_\Omega g \, f \, d\mu, \] with both sides finite. The integrability equivalence \(g \in L^1(\nu) \iff g f \in L^1(\mu)\) follows by applying the non-negative identity to \(|g|\).

Proof of (i). Set \(f = d\nu/d\mu\) and \(k = d\mu/d\lambda\). For any \(A \in \mathcal{F}\), apply (ii) with \(\nu \ll \mu\) and the test function \(g = \mathbf{1}_A\), then apply (ii) again with \(\mu \ll \lambda\) and the test function \(g' = \mathbf{1}_A f\) (non-negative and measurable): \[ \nu(A) \;\stackrel{\text{(ii)}_{\nu \ll \mu}}{=}\; \int_\Omega \mathbf{1}_A f \, d\mu \;\stackrel{\text{(ii)}_{\mu \ll \lambda}}{=}\; \int_\Omega \mathbf{1}_A f \, k \, d\lambda \;=\; \int_A f k \, d\lambda. \] Hence \(\nu(A) = \int_A (f k) \, d\lambda\) for every \(A \in \mathcal{F}\). The product \(fk\) is measurable and non-negative; although the value of \(f\) on \(\mu\)-null sets is undetermined (and such sets need not be \(\lambda\)-null), the integral against \(\lambda\) is unaffected by this ambiguity: for any two representatives \(f, f'\) of \(d\nu/d\mu\), applying (ii) with \(\mu \ll \lambda\) gives \(\int_A (f - f') k \, d\lambda = \int_A (f - f') \, d\mu = 0\), since \(f = f'\) holds \(\mu\)-a.e. Thus \(\int_A fk \, d\lambda = \nu(A)\) is well-defined regardless of the choice of representatives, exhibiting \(fk\) as a density of \(\nu\) with respect to \(\lambda\). By uniqueness of the Radon-Nikodym derivative with respect to \(\lambda\), \(d\nu/d\lambda = fk = (d\nu/d\mu)(d\mu/d\lambda)\) holds \(\lambda\)-a.e.

The hypothesis \(\nu \ll \lambda\) needed to apply uniqueness here is itself a consequence of \(\nu \ll \mu \ll \lambda\): if \(\lambda(A) = 0\), then \(\mu(A) = 0\), and then \(\nu(A) = 0\). \(\blacksquare\)

Two notational conveniences are worth noting. First, the change-of-variables formula (ii) is often written in the suggestive Leibniz form \[ \int g \, d\nu \;=\; \int g \, \frac{d\nu}{d\mu} \, d\mu, \] in which "\(d\nu = (d\nu/d\mu) \, d\mu\)" reads as a formal cancellation of differentials. This is the source of all calculations of the type \(\mathbb{E}_P[g(X)] = \mathbb{E}_Q[g(X) \cdot dP/dQ(X)]\) used in importance sampling and Monte Carlo methods, where the target distribution \(P\) is integrated against by drawing samples from a different distribution \(Q\) and reweighting by the Radon-Nikodym derivative \(dP/dQ\). Second, the chain rule (i) gives the Radon-Nikodym derivative the algebraic structure of a classical derivative under composition: transitivity of \(\ll\) at the level of measures lifts to multiplicativity of \(d/d\) at the level of densities.

A subtlety distinguishes this change-of-variables formula from the pushforward change-of-variables proved in Pushforward Measure. The pushforward identity \(\int g \, d(X_*\mathbb{P}) = \int (g \circ X) \, d\mathbb{P}\) transports integrals across a measurable map between two different measurable spaces; the function \(g\) and its lift \(g \circ X\) live on different spaces, and the derivative \(dX_*\mathbb{P}/d\mathbb{P}\) is not even defined (the measures live on different \(\sigma\)-algebras). The Radon-Nikodym change-of-variables, by contrast, transports integrals between two measures on the same space, paid for by multiplication by the density. Both are change-of-variables formulas, but they operate in orthogonal directions: pushforward changes the carrier space, Radon-Nikodym changes the weighting.

Connection to Machine Learning

Almost every probabilistic quantity in modern machine learning is, at the foundational level, a Radon-Nikodym derivative — and almost every algorithm involving densities is, at the foundational level, an instance of the change-of-variables formula (ii). We collect four examples that span supervised, generative, and reinforcement learning.

Bayesian posterior density. Given a prior \(\Pi\) on the parameter space \(\Theta\) and a likelihood \(p(x | \theta)\) — viewed as a density of the conditional data distribution with respect to a reference measure on the data space — the Bayesian posterior \(\Pi_{\theta | x}\) is defined as a measure on \(\Theta\) — not, a priori, as a density. When \(\Pi \ll \lambda_\Theta\) (with \(\lambda_\Theta\) a reference measure on \(\Theta\), often Lebesgue measure), Bayes' rule produces the posterior density \(d\Pi_{\theta|x}/d\lambda_\Theta\) by the formula \[ \frac{d\Pi_{\theta|x}}{d\lambda_\Theta}(\theta) \;=\; \frac{p(x | \theta) \cdot (d\Pi/d\lambda_\Theta)(\theta)}{\int_\Theta p(x|\theta') \, (d\Pi/d\lambda_\Theta)(\theta') \, d\lambda_\Theta(\theta')}. \] If the prior is supported on a discrete set, \(\Pi \perp \lambda_\Theta\) and no Lebesgue density exists; the posterior is then a discrete measure whose density relative to counting measure on the support plays the analogous role. The framework is unified by the choice of dominating measure.

Score function in diffusion models. Score-based generative models — including denoising diffusion probabilistic models and score-based stochastic differential equations — train a neural network \(s_\phi\) to estimate the gradient of the log-density, called the Stein score (in the diffusion-modeling literature, often simply "score"): \[ s_\phi(x, t) \;\approx\; \nabla_x \log p_t(x) \;=\; \nabla_x \log \frac{dP_t}{d\lambda}(x), \] where \(P_t\) is the distribution of the noised data at time \(t\) and \(\lambda\) is Lebesgue measure on \(\mathbb{R}^d\). The score is well-defined as a measurable function on \(\mathbb{R}^d\) when \(P_t \ll \lambda\) for \(t > 0\) (and, in the diffusion-model setting, smooth thanks to the Gaussian noise injection along the forward diffusion). This noise injection ensures the absolute continuity even when the original data distribution \(P_0\) lies on a low-dimensional manifold and is singular to \(\lambda\). The reverse-time generative SDE is then driven by \(s_\phi\), which integrates exactly the Radon-Nikodym derivative \(dP_t/d\lambda\) into the dynamics.

Importance sampling. The most direct application of the change-of-variables formula is importance sampling: to estimate \(\mathbb{E}_P[g(X)]\) using samples drawn from a different distribution \(Q\), one rewrites \[ \mathbb{E}_P[g(X)] \;=\; \int g \, dP \;=\; \int g \cdot \frac{dP}{dQ} \, dQ \;=\; \mathbb{E}_Q\!\left[g(X) \cdot \frac{dP}{dQ}(X)\right] \] and replaces the expectation under \(P\) by the empirical mean under \(Q\) of the reweighted integrand \(g(X) \cdot dP/dQ(X)\). This is precisely the change-of-variables proposition with \(\nu = P\) and \(\mu = Q\). The estimator is unbiased whenever \(P \ll Q\) (and \(g \in L^1(P)\), so that both expectations are finite); when \(P \not\ll Q\), the ratio \(dP/dQ\) is undefined on a set of positive \(P\)-measure, and the estimator misses contributions from that region — the practical manifestation of an absolute-continuity violation.

KL divergence and RLHF regularization. The Kullback-Leibler divergence between two probability measures \(P\) and \(Q\) is defined by \[ D_{\mathrm{KL}}(P \| Q) \;=\; \begin{cases} \displaystyle \int_\Omega \log \frac{dP}{dQ} \, dP & \text{if } P \ll Q, \\ +\infty & \text{otherwise}. \end{cases} \] The Radon-Nikodym derivative \(dP/dQ\) is what the divergence is integrating; without absolute continuity, \(dP/dQ\) does not exist as a Radon-Nikodym derivative on the portion of the space where \(Q\) places no mass but \(P\) does, and \(D_{\mathrm{KL}}(P \| Q)\) is set to \(+\infty\) by convention. In reinforcement learning from human feedback (RLHF), the regularizer \(D_{\mathrm{KL}}(\pi_\theta \| \pi_{\mathrm{ref}})\) — penalizing the trained policy \(\pi_\theta\) for drifting from a reference policy \(\pi_{\mathrm{ref}}\) — is finite precisely when \(\pi_\theta \ll \pi_{\mathrm{ref}}\), which is enforced by ensuring the trained policy cannot place mass where the reference places none (e.g., via shared support or by parameterizing \(\pi_\theta\) as \(\pi_{\mathrm{ref}}\) times a strictly positive learned ratio).

The recurring pattern: the Radon-Nikodym derivative is the object that algorithms compute with, and absolute continuity is the hypothesis that makes the computation well-posed. The Radon-Nikodym theorem proved above is the existence proof that licenses every one of these constructions.

Looking Ahead

The Radon-Nikodym theorem closes three open accounts in the curriculum and opens several new lines of development. We summarize both.

Retroactive Closures

First, in Pushforward Measure, we defined the probability density function \(f_X\) of a continuous random variable as the formal derivative \(dP_X/d\lambda\), assuming its existence. With the Radon-Nikodym theorem proved here, that assumption is now justified: whenever \(P_X \ll \lambda\) — the measure-theoretic definition of "continuous random variable" — the function \(f_X\) exists, is unique \(\lambda\)-a.e., and is given by \(f_X = dP_X/d\lambda\). The PDF is no longer a primitive object but a derived one, obtained from the structural relation \(P_X \ll \lambda\) between two measures.

Second, in Dual Spaces & Riesz Representation, the duality \((L^p)^* \cong L^q\) was stated as a fact, with the converse direction — every continuous functional on \(L^p\) arises from some \(\psi \in L^q\) — left without proof. The standard proof of that converse takes a continuous functional \(\Lambda\) on \(L^p(\mu)\), constructs the signed measure \(\nu(A) = \Lambda(\mathbf{1}_A)\), shows \(\nu \ll \mu\), and applies Radon-Nikodym to obtain \(\psi = d\nu/d\mu\). The present theorem fills that gap, completing the standard \(L^p\)-duality statement for \(1 \leq p < \infty\).

Third, in Limit Theorems & Product Measures, the closing paragraphs previewed conditional expectation \(\mathbb{E}[X | \mathcal{G}]\) as a Radon-Nikodym derivative of the signed measure \(A \mapsto \int_A X \, d\mathbb{P}\) (defined for \(A \in \mathcal{G}\)) with respect to \(\mathbb{P}|_{\mathcal{G}}\). With Radon-Nikodym established, this construction is now fully licensed; its development is the subject of the next chapter.

The Lebesgue Decomposition

Absolute continuity (\(\ll\)) and mutual singularity (\(\perp\)) are extreme relations between two measures. The Lebesgue decomposition theorem, which we state without proof, asserts that every \(\sigma\)-finite signed measure decomposes uniquely into an absolutely continuous part and a mutually singular part with respect to any reference measure: given a \(\sigma\)-finite non-negative measure \(\mu\) and a \(\sigma\)-finite signed measure \(\nu\), there exist unique signed measures \(\nu_{\mathrm{ac}}, \nu_{\mathrm{s}}\) with \(\nu = \nu_{\mathrm{ac}} + \nu_{\mathrm{s}}\), \(\nu_{\mathrm{ac}} \ll \mu\), and \(\nu_{\mathrm{s}} \perp \mu\). The proof, available in standard references such as Folland's real analysis text or Durrett's probability text, is a refinement of the von Neumann argument used here for Radon-Nikodym. Combined with our theorem, the Lebesgue decomposition gives a complete structural classification: \(\nu\) splits into a "density part" (an integral against \(\mu\)) and a "singular part" (concentrated where \(\mu\) is not) — and nothing else.

The Next Chapter

In the upcoming page on conditional expectation, we will apply the Radon-Nikodym theorem to define the conditional expectation \(\mathbb{E}[X | \mathcal{G}]\) of an integrable random variable \(X \in L^1(\mathbb{P})\) with respect to a sub-\(\sigma\)-algebra \(\mathcal{G} \subseteq \mathcal{F}\). The signed measure \(\nu_X(A) = \int_A X \, d\mathbb{P}\) on \((\Omega, \mathcal{G})\) is absolutely continuous with respect to the restricted probability measure \(\mathbb{P}|_{\mathcal{G}}\); the conditional expectation is then defined as the Radon-Nikodym derivative \[ \mathbb{E}[X | \mathcal{G}] \;=\; \frac{d\nu_X}{d \mathbb{P}|_{\mathcal{G}}}. \] This rephrasing does several things at once: it makes \(\mathbb{E}[X | \mathcal{G}]\) a \(\mathcal{G}\)-measurable function (rather than an event-by-event computation), it explains why conditional expectation is unique only \(\mathbb{P}\)-a.s. (Radon-Nikodym uniqueness), and it provides a single unified definition that subsumes both the discrete case (\(\mathbb{E}[X | A]\) for an event \(A\)) and the continuous case. From this foundation, the martingale theory that drives stochastic calculus, optimal stopping, and modern asymptotic statistics becomes available.

Further Horizons

Beyond conditional expectation, several future directions follow naturally from the Radon-Nikodym framework:

Each of these directions extends the same theorem in a different style: martingales push it into the time domain, Girsanov pushes it into path space, variational inference pushes it into the optimization landscape over distributions, and information geometry pushes it into differential geometry. The Radon-Nikodym theorem is, in this sense, the central hub from which the deeper probabilistic structure of modern machine learning radiates.