Radon–Nikodym Theorem for Kernels

• Radon–Nikodym Theorem for Kernels is a collection of results that extend the classical density formula to kernels in measurable, RKHS, and operator-valued contexts.
• It employs techniques such as Kolmogorov decomposition, operator factorization, and Hilbert space embeddings to characterize domination and derive density operators.
• The framework supports applications in noncommutative probability, random matrix ensembles, and RKHS-based computational methods in spectral estimation.

The Radon–Nikodym theorem for kernels is a family of representation results that extend the classical measure-theoretic formula $d\mu = f\, d\nu$ to settings where the basic positive object is a kernel rather than a measure. In current usage, the term kernel covers at least three distinct structures: measurable kernels $\kappa:X\rightsquigarrow Y$ in probability and semantics, scalar or operator-valued positive definite kernels $K:X\times X\to \mathcal L(H)$ in RKHS and dilation theory, and completely positive matrix-valued kernels arising from Stinespring- and KSGNS-type constructions. Across these settings, domination is converted into a derivative: either a measurable density $d\kappa/d\lambda$, or a positive contraction acting on a Kolmogorov, reproducing-kernel, or Stinespring space (Vákár et al., 2018, Tian, 24 Sep 2025, Jorgensen et al., 2024).

1. Terminological scope and basic structures

A first technical point is that the phrase kernel is genuinely overloaded. In the measure-theoretic literature, a kernel from $(X,\Sigma_X)$ to $(Y,\Sigma_Y)$ is a map

$\kappa:X\times \Sigma_Y \to [0,\infty]$

such that $A\mapsto \kappa(x,A)$ is a measure for each $x$, and $x\mapsto \kappa(x,A)$ is measurable for each $\kappa:X\rightsquigarrow Y$0 (Vákár et al., 2018). In the RKHS and operator-theoretic literature, a kernel is instead a function on a Cartesian square. A scalar kernel $\kappa:X\rightsquigarrow Y$1 is positive definite if

$\kappa:X\rightsquigarrow Y$2

while an operator-valued kernel $\kappa:X\rightsquigarrow Y$3 is positive definite if

$\kappa:X\rightsquigarrow Y$4

for all finite choices of points and vectors (Tian, 24 Sep 2025).

For positive definite kernels, the basic structural tool is the Kolmogorov–Aronszajn decomposition. In the scalar case one has a Hilbert space $\kappa:X\rightsquigarrow Y$5 and feature map $\kappa:X\rightsquigarrow Y$6 such that

$\kappa:X\rightsquigarrow Y$7

with dense span. In the operator-valued case one has

$\kappa:X\rightsquigarrow Y$8

again with a minimality condition given by density of $\kappa:X\rightsquigarrow Y$9 (Tian, 24 Sep 2025). The same construction appears in a canonical scalarized form in operator-valued kernel theory: if

$K:X\times X\to \mathcal L(H)$0

then $K:X\times X\to \mathcal L(H)$1 is a scalar positive definite kernel with RKHS $K:X\times X\to \mathcal L(H)$2, and the feature map is $K:X\times X\to \mathcal L(H)$3 (Jorgensen et al., 2024).

The relevant order relation also depends on context. For measurable kernels, domination is absolute continuity: $K:X\times X\to \mathcal L(H)$4 for all $K:X\times X\to \mathcal L(H)$5 and measurable $K:X\times X\to \mathcal L(H)$6 (Vákár et al., 2018). For positive definite kernels, domination is the Loewner-type order

$K:X\times X\to \mathcal L(H)$7

equivalently, every finite Gram block for $K:X\times X\to \mathcal L(H)$8 is positive semidefinite (Tian, 24 Sep 2025). A persistent source of confusion is to treat these notions as interchangeable; they are not. They live on different categories of kernels and produce different kinds of Radon–Nikodym derivatives.

2. Positive definite kernels and operator-valued densities

For positive definite kernels, the core Radon–Nikodym statement is an operator factorization theorem. If $K:X\times X\to \mathcal L(H)$9 are positive definite and $d\kappa/d\lambda$0, and if

$d\kappa/d\lambda$1

is a minimal Kolmogorov decomposition, then there exists a unique operator

$d\kappa/d\lambda$2

such that

$d\kappa/d\lambda$3

This is the kernel Radon–Nikodym theorem in the form stated in Theorem 3.1 of "Kernel Radon-Nikodym Derivatives for Random Matrix Ensembles" (Tian, 24 Sep 2025). The same functional-analytic pattern appears in the general operator-valued kernel theorem of "Operator-Valued Kernels, Machine Learning, and Dynamical Systems," where the derivative is denoted $d\kappa/d\lambda$4 and acts on the scalarized feature space $d\kappa/d\lambda$5: $d\kappa/d\lambda$6 with the equivalent feature-map relation

$d\kappa/d\lambda$7

(Jorgensen et al., 2024).

The factorization admits an equivalent contraction form. If

$d\kappa/d\lambda$8

is any minimal decomposition of $d\kappa/d\lambda$9, then there is a unique contraction

$(X,\Sigma_X)$0

such that

$(X,\Sigma_X)$1

This shows that the derivative is not a pointwise multiplier on $(X,\Sigma_X)$2, but an operator on the feature space of the dominating kernel (Tian, 24 Sep 2025). In the scalar case, this reduces to the familiar Aronszajn inclusion principle: $(X,\Sigma_X)$3 if and only if $(X,\Sigma_X)$4 embeds contractively into $(X,\Sigma_X)$5, and the derivative is the operator implementing that inclusion (Jorgensen et al., 2024).

A further structural refinement appears when the kernel is equivariant under a representation of a $(X,\Sigma_X)$6-algebra. If a nondegenerate representation $(X,\Sigma_X)$7 satisfies

$(X,\Sigma_X)$8

and similarly for $(X,\Sigma_X)$9, then the derivative lies in the commutant: $(Y,\Sigma_Y)$0 This aligns the kernel theorem with Radon–Nikodym theorems for completely positive maps and Stinespring dilations (Tian, 24 Sep 2025). A common misconception is that the kernel derivative should be a scalar density analogous to $(Y,\Sigma_Y)$1; in the positive-definite setting, the derivative is typically operator-valued because the order relation is encoded in feature-space geometry rather than pointwise integration.

3. Shifted kernels, moment-ratio tests, and random matrix ensembles

A distinctive development of the positive-definite theory is the shifted-kernel formalism on the free semigroup $(Y,\Sigma_Y)$2. If

$(Y,\Sigma_Y)$3

is positive definite, its shift is defined by

$(Y,\Sigma_Y)$4

Given a minimal Kolmogorov decomposition

$(Y,\Sigma_Y)$5

one defines abstract shift operators on the dense span $(Y,\Sigma_Y)$6 by

$(Y,\Sigma_Y)$7

The Radon–Nikodym derivative of the shifted kernel is then

$(Y,\Sigma_Y)$8

and

$(Y,\Sigma_Y)$9

This is the content of Lemma 4.1 and Theorem 4.2 in (Tian, 24 Sep 2025).

In the one-variable bi-unitarily invariant case, the derivative becomes diagonal and the domination criterion reduces to a moment-ratio test. If

$\kappa:X\times \Sigma_Y \to [0,\infty]$0

where

$\kappa:X\times \Sigma_Y \to [0,\infty]$1

then the shift satisfies

$\kappa:X\times \Sigma_Y \to [0,\infty]$2

and hence

$\kappa:X\times \Sigma_Y \to [0,\infty]$3

(Tian, 24 Sep 2025). For Haar unitary matrices, $\kappa:X\times \Sigma_Y \to [0,\infty]$4, so $\kappa:X\times \Sigma_Y \to [0,\infty]$5. For complex Ginibre ensembles with variance $\kappa:X\times \Sigma_Y \to [0,\infty]$6, one has $\kappa:X\times \Sigma_Y \to [0,\infty]$7, and therefore

$\kappa:X\times \Sigma_Y \to [0,\infty]$8

so the shifted-kernel domination criterion is asymptotically equivalent to $\kappa:X\times \Sigma_Y \to [0,\infty]$9 (Tian, 24 Sep 2025).

The same mechanism extends blockwise. If $A\mapsto \kappa(x,A)$0 and block invariance yields

$A\mapsto \kappa(x,A)$1

then

$A\mapsto \kappa(x,A)$2

and domination holds if and only if

$A\mapsto \kappa(x,A)$3

(Tian, 24 Sep 2025). The paper also states mean-square von Neumann inequalities under shifted domination: $A\mapsto \kappa(x,A)$4 and for $A\mapsto \kappa(x,A)$5 in $A\mapsto \kappa(x,A)$6,

$A\mapsto \kappa(x,A)$7

These formulas connect kernel domination to operator inequalities for noncommutative polynomials (Tian, 24 Sep 2025).

4. Measurable kernels, s-finiteness, and strengthened absolute continuity

For measurable kernels, the Radon–Nikodym theorem takes the classical integral form. If $A\mapsto \kappa(x,A)$8 are $A\mapsto \kappa(x,A)$9-finite kernels and $x$0, then there exists a density

$x$1

such that

$x$2

with uniqueness $x$3-a.e. for each $x$4 (Vákár et al., 2018). Under the mild structural assumptions that $x$5 is countable discrete or $x$6 is standard Borel, the density can be chosen jointly measurable in $x$7 (Vákár et al., 2018).

The s-finite extension is subtler. A kernel is s-finite if it is a countable sum of finite kernels, with no mutual singularity requirement. The paper "On S-Finite Measures and Kernels" shows that for s-finite kernels $x$8, ordinary absolute continuity $x$9 is not sufficient for a Radon–Nikodym density. One needs a strengthened relation, denoted in the paper by $x\mapsto \kappa(x,A)$0, meaning $x\mapsto \kappa(x,A)$1 and preservation of $x\mapsto \kappa(x,A)$2-0–$x\mapsto \kappa(x,A)$3 sets for each $x\mapsto \kappa(x,A)$4 (Vákár et al., 2018). Under this hypothesis there exists

$x\mapsto \kappa(x,A)$5

with

$x\mapsto \kappa(x,A)$6

The strengthened condition is tied to the notion of a top $x\mapsto \kappa(x,A)$7–$x\mapsto \kappa(x,A)$8 set $x\mapsto \kappa(x,A)$9 for an s-finite measure $\kappa:X\rightsquigarrow Y$00. For s-finite kernels, under the same countability or standard-Borel assumptions, there is a measurable set $\kappa:X\rightsquigarrow Y$01 whose vertical sections are the pointwise top $\kappa:X\rightsquigarrow Y$02–$\kappa:X\rightsquigarrow Y$03 sets, and $\kappa:X\rightsquigarrow Y$04 restricted to the complement of $\kappa:X\rightsquigarrow Y$05 is $\kappa:X\rightsquigarrow Y$06-finite (Vákár et al., 2018). Uniqueness of densities is correspondingly refined: derivatives are “a.e. $\kappa:X\rightsquigarrow Y$07-unique,” meaning standard a.e. uniqueness on the $\kappa:X\rightsquigarrow Y$08-finite part and a weaker zero/nonzero agreement condition on the infinite part.

A common misconception is that the classical hypothesis $\kappa:X\rightsquigarrow Y$09 should remain sufficient in the s-finite world. The counterexample in (Vákár et al., 2018) shows otherwise: if

$\kappa:X\rightsquigarrow Y$10

then $\kappa:X\rightsquigarrow Y$11 holds, but no measurable density $\kappa:X\rightsquigarrow Y$12 can satisfy $\kappa:X\rightsquigarrow Y$13. The obstruction is precisely the mismatch of $\kappa:X\rightsquigarrow Y$14–$\kappa:X\rightsquigarrow Y$15 structure.

The same paper also proves an s-finite Lebesgue decomposition for kernels: $\kappa:X\rightsquigarrow Y$16 where $\kappa:X\rightsquigarrow Y$17 is absolutely continuous in the strengthened sense, $\kappa:X\rightsquigarrow Y$18 is $\kappa:X\rightsquigarrow Y$19-singular but still dominated, and $\kappa:X\rightsquigarrow Y$20 is singular (Vákár et al., 2018). This decomposition has no analogue in the classical $\kappa:X\rightsquigarrow Y$21-finite setting, where the $\kappa:X\rightsquigarrow Y$22-singular term vanishes.

5. RKHS reformulations and Lebesgue decomposition

A different kernel-theoretic route to Radon–Nikodym theory is to encode measure relations inside reproducing-kernel Hilbert spaces. For finite positive regular Borel measures $\kappa:X\rightsquigarrow Y$23 on the unit circle $\kappa:X\rightsquigarrow Y$24, "A reproducing kernel approach to Lebesgue decomposition" constructs the RKHS $\kappa:X\rightsquigarrow Y$25 of $\kappa:X\rightsquigarrow Y$26-Cauchy transforms on the unit disk $\kappa:X\rightsquigarrow Y$27, with reproducing kernel

$\kappa:X\rightsquigarrow Y$28

(Bal et al., 2023). Here $\kappa:X\rightsquigarrow Y$29 is the Herglotz–Riesz transform of $\kappa:X\rightsquigarrow Y$30.

In this framework, domination is equivalent to RKHS containment. Theorem 3 of (Bal et al., 2023) states that

$\kappa:X\rightsquigarrow Y$31

equivalently $\kappa:X\rightsquigarrow Y$32 embeds in $\kappa:X\rightsquigarrow Y$33 with norm at most $\kappa:X\rightsquigarrow Y$34. Absolute continuity is likewise characterized by intersection density: if

$\kappa:X\rightsquigarrow Y$35

then

$\kappa:X\rightsquigarrow Y$36

(Theorem 6) (Bal et al., 2023).

The Radon–Nikodym derivative appears here as a Toeplitz symbol. When $\kappa:X\rightsquigarrow Y$37 with $\kappa:X\rightsquigarrow Y$38, the kernel identity is

$\kappa:X\rightsquigarrow Y$39

The associated operator on $\kappa:X\rightsquigarrow Y$40 is

$\kappa:X\rightsquigarrow Y$41

so the derivative is realized as the multiplication symbol of a Toeplitz operator implementing the inclusion (Bal et al., 2023). This is structurally parallel to the operator-density formulas $\kappa:X\rightsquigarrow Y$42 of the abstract positive-definite theory.

The same paper gives a kernel-theoretic construction of Lebesgue decomposition via positive quadratic forms. A significant limitation is explicitly identified: the Simon-Lebesgue decomposition of forms need not coincide with the measure-theoretic decomposition unless the intersection space is invariant, equivalently $\kappa:X\rightsquigarrow Y$43-reducing. Example 3 in (Bal et al., 2023) shows that one may have $\kappa:X\rightsquigarrow Y$44 relative to $\kappa:X\rightsquigarrow Y$45 while the form-theoretic absolutely continuous part is nonzero. This is an important caution against assuming that every RKHS decomposition automatically matches measure decomposition without an invariance hypothesis.

6. Noncommutative extensions, stochastic realizations, and computation

The operator-density form of the kernel Radon–Nikodym theorem has several extensions. Every operator-valued positive definite kernel $\kappa:X\rightsquigarrow Y$46 admits a Gaussian realization: $\kappa:X\rightsquigarrow Y$47 and, if $\kappa:X\rightsquigarrow Y$48 is an orthonormal basis of the feature space, one may take

$\kappa:X\rightsquigarrow Y$49

where $\kappa:X\rightsquigarrow Y$50 are i.i.d. standard Gaussians (Jorgensen et al., 2024). If $\kappa:X\rightsquigarrow Y$51 with Radon–Nikodym derivative $\kappa:X\rightsquigarrow Y$52, then the process with covariance $\kappa:X\rightsquigarrow Y$53 is obtained from the process for $\kappa:X\rightsquigarrow Y$54 by inserting $\kappa:X\rightsquigarrow Y$55 in feature space: $\kappa:X\rightsquigarrow Y$56 This realizes domination of kernels as covariance compression (Jorgensen et al., 2024).

In noncommutative probability, the same pattern becomes a Radon–Nikodym theorem for completely positive maps. For a unital CP map $\kappa:X\rightsquigarrow Y$57 one defines

$\kappa:X\rightsquigarrow Y$58

builds the minimal Stinespring representation $\kappa:X\rightsquigarrow Y$59, and then, for another CP map $\kappa:X\rightsquigarrow Y$60, there exists a unique positive contraction

$\kappa:X\rightsquigarrow Y$61

such that

$\kappa:X\rightsquigarrow Y$62

(Jorgensen et al., 2024). The finite-index Hilbert-module analogue replaces a single CP map by a completely positive $\kappa:X\rightsquigarrow Y$63 matrix of maps and proves that domination corresponds bijectively to a positive contraction in the commutant of the matrix KSGNS dilation; this is the content of the order-isomorphism theorem in (Moslehian et al., 2016).

Computation in the positive-definite setting proceeds through finite Gram data. Given sample points $\kappa:X\rightsquigarrow Y$64, one forms block Gram matrices

$\kappa:X\rightsquigarrow Y$65

tests $\kappa:X\rightsquigarrow Y$66 by checking

$\kappa:X\rightsquigarrow Y$67

in the Loewner sense, constructs feature maps by Cholesky or spectral factorization, and then solves for a contraction $\kappa:X\rightsquigarrow Y$68 mapping the $\kappa:X\rightsquigarrow Y$69-features to the $\kappa:X\rightsquigarrow Y$70-features, finally setting

$\kappa:X\rightsquigarrow Y$71

as the finite-dimensional Radon–Nikodym density on the Kolmogorov span (Tian, 24 Sep 2025). For shifted kernels on $\kappa:X\rightsquigarrow Y$72, one may compute the abstract shift matrices $\kappa:X\rightsquigarrow Y$73 directly and set

$\kappa:X\rightsquigarrow Y$74

If empirical violations are small and attributable to noise, the paper recommends regularization by adding $\kappa:X\rightsquigarrow Y$75 and retesting domination for

$\kappa:X\rightsquigarrow Y$76

(Tian, 24 Sep 2025).

A different computational line arises in RKHS estimation of classical Radon–Nikodym derivatives. If $\kappa:X\rightsquigarrow Y$77 with density $\kappa:X\rightsquigarrow Y$78, then in the RKHS framework of (Nguyen et al., 2023) the derivative solves

$\kappa:X\rightsquigarrow Y$79

and is estimated by a spectral filter: $\kappa:X\rightsquigarrow Y$80 For the iterated Lavrentiev filter,

$\kappa:X\rightsquigarrow Y$81

the paper gives an explicit recursion in terms of the Gram matrix $\kappa:X\rightsquigarrow Y$82 and proves global and pointwise error bounds controlled by regularized Christoffel functions (Nguyen et al., 2023). This is not the same theorem as the operator-density result for positive-definite kernels, but it shows how Radon–Nikodym differentiation interacts with kernel methods algorithmically.

Taken together, these developments show that the phrase Radon–Nikodym theorem for kernels does not designate a single theorem. It denotes a cluster of closely related representation principles: measurable densities for dominated measurable kernels, positive-contraction densities on Kolmogorov or RKHS feature spaces for dominated positive definite kernels, and commutant-valued densities in Stinespring or KSGNS dilations for noncommutative kernels (Vákár et al., 2018, Tian, 24 Sep 2025, Jorgensen et al., 2024, Moslehian et al., 2016).