State-Parameter Product Kernels

• State-Parameter Product Kernels are reproducing kernels on joint state-parameter spaces, constructed by multiplying individual positive definite kernels.
• They facilitate multivariate scattered-data interpolation using explicit formulas and tensor-product structured numerical linear algebra.
• Their tensorized representation supports spectral analysis and kernel ridge regression, allowing anisotropic modeling across state and parameter dimensions.

A state-parameter product kernel is a reproducing kernel on a joint domain $\mathcal X \times \Theta$ obtained by multiplying a kernel on the state space with a kernel on the parameter space,

$$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$

In the RKHS setting, this construction is governed by Aronszajn’s theory of product kernels: positivity is preserved under products, and the resulting native space is identified with a Hilbert tensor product. In scattered-data interpolation, this yields a multivariate approximation framework with explicit interpolation formulas, strict positive definiteness under factorwise assumptions, and tensor-structured numerical linear algebra; in operator-theoretic treatments of parametric models, the same structure emerges from factorisations of correlation operators and associated Karhunen–Loève or POD decompositions (Albrecht et al., 2023, Matthies et al., 2018).

1. Algebraic form and positivity

Let $k_s$ be a positive (semi-)definite kernel on a state domain $\mathcal X \subset \mathbb R^{d_x}$, and let $k_p$ be a positive (semi-)definite kernel on a parameter domain $\Theta \subset \mathbb R^{d_\theta}$. The joint kernel is

$$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$

Aronszajn’s theorem implies that the product of positive semi-definite kernels is again positive semi-definite. In matrix form, if $A_s=(k_s(x_i,x_j))_{i,j}\succeq 0$ and $A_p=(k_p(\theta_i,\theta_j))_{i,j}\succeq 0$, then the Hadamard–Schur product

$A=A_s\circ A_p$

is again positive semidefinite (Albrecht et al., 2023).

This establishes the basic admissibility of the state-parameter product construction. The key point is that the joint kernel is not introduced ad hoc on $$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$0; it is inherited from two lower-dimensional kernels whose positivity properties are preserved exactly. This factorwise construction is the central reason product kernels are treated as efficient and flexible tools for high-dimensional scattered-data interpolation in the cited work (Albrecht et al., 2023).

2. Native spaces and tensor-product structure

If $$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$1 denotes the RKHS of $$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$2 and $$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$3 denotes the RKHS of $$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$4, then the RKHS $$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$5 of the product kernel satisfies

$$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$6

as a Hilbert space. Concretely, there is a multilinear map

$$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$7

and for simple tensors one has

$$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$8

Hence, if $$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$9, then

$k_s$0

By linearity and density, one recovers the usual infimum-sum norm on the completed tensor product (Albrecht et al., 2023).

An operator-theoretic formulation leads to the same tensor structure. In a linear parametric model, one considers separable Hilbert spaces $k_s$1 and $k_s$2, together with a bounded linear map

$k_s$3

The generalized correlation operator is

$k_s$4

which is self-adjoint and positive; its spectral decomposition yields the singular-value, POD, or Karhunen–Loève factorisation of $k_s$5. On the joint domain $k_s$6, one forms

$k_s$7

whose reproducing kernel is

$k_s$8

This places state-parameter product kernels within a broader functional-analytic framework in which kernel factorisation, correlation operators, and low-rank spectral expansions are directly linked (Matthies et al., 2018).

3. Strict positive definiteness and interpolation

If both component kernels are strictly positive definite, then the product kernel is strictly positive definite as well. Equivalently, for any finite set of distinct state-parameter pairs $k_s$9, the Gram matrix

$\mathcal X \subset \mathbb R^{d_x}$0

is non-singular. The cited proof uses a grid-like embedding trick together with Kronecker-product structure (Albrecht et al., 2023).

Given data $\mathcal X \subset \mathbb R^{d_x}$1, the representer theorem yields the unique minimum-norm interpolant in $\mathcal X \subset \mathbb R^{d_x}$2,

$\mathcal X \subset \mathbb R^{d_x}$3

where the coefficient vector $\mathcal X \subset \mathbb R^{d_x}$4 is obtained from

$\mathcal X \subset \mathbb R^{d_x}$5

This is the standard kernel-interpolation construction transferred directly to the joint state-parameter space (Albrecht et al., 2023).

The interpolation significance of the product form lies in factorwise modelling freedom. The cited numerical discussion emphasizes that one component kernel may be chosen “wide” for rough parameter behavior while the other is chosen “narrow” for highly oscillatory state behavior, yielding an often better MSE-versus-stability trade-off than a single isotropic kernel on $\mathcal X \subset \mathbb R^{d_x}$6 (Albrecht et al., 2023). This suggests that separability is being used not merely for algebraic convenience, but as a mechanism for anisotropic modelling across coordinate blocks.

4. Tensorized Newton bases, conditioning, and complexity

In each component RKHS one may construct a Newton, or “Cholesky,” basis. If $\mathcal X \subset \mathbb R^{d_x}$7 is an orthonormal Newton basis in $\mathcal X \subset \mathbb R^{d_x}$8 for state nodes $\mathcal X \subset \mathbb R^{d_x}$9, and $k_p$0 is an orthonormal Newton basis in $k_p$1 for parameter nodes $k_p$2, then

$k_p$3

is an orthonormal basis of the product space of size $k_p$4 (Albrecht et al., 2023).

This tensor structure has direct computational consequences. The joint Cholesky factor can be assembled as the Kronecker product of the two smaller Cholesky factors, with cost

$k_p$5

Updates under new sample points likewise decouple into two small updates plus one tensoring step. For grid-like or partially separable sampling, matrix assembly, Cholesky factorization, Newton-basis updates, and greedy selection reduce to componentwise operations plus small tensor products (Albrecht et al., 2023).

The same paper also records a conditioning formula: $k_p$6 Accordingly, choosing well-conditioned component kernels gives finer control of the joint condition number than treating the problem as a single kernel construction in the combined dimension $k_p$7 (Albrecht et al., 2023). In this sense, the tensor product is not only a representational device but also a numerical pre-structuring of the interpolation problem.

5. Spectral analysis and kernel ridge regression on product spaces

On a product input space $k_p$8 with product measure $k_p$9, and with continuous Mercer kernels $\Theta \subset \mathbb R^{d_\theta}$0 and $\Theta \subset \mathbb R^{d_\theta}$1, the product kernel

$\Theta \subset \mathbb R^{d_\theta}$2

is itself a Mercer kernel. If the marginal kernels admit $\Theta \subset \mathbb R^{d_\theta}$3-expansions with eigenpairs $\Theta \subset \mathbb R^{d_\theta}$4 and $\Theta \subset \mathbb R^{d_\theta}$5, then the integral operator factorizes as

$\Theta \subset \mathbb R^{d_\theta}$6

with eigenpairs

$\Theta \subset \mathbb R^{d_\theta}$7

Hence the Mercer expansion on $\Theta \subset \mathbb R^{d_\theta}$8 is doubly indexed by the two marginal spectra (Zhou et al., 14 May 2026).

In kernel ridge regression, this factorization yields an explicit bias-variance decomposition in the product eigenbasis. The exact leading-order formulas are

$\Theta \subset \mathbb R^{d_\theta}$9

where $$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$0 and $$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$1 are series over $$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$2. The summary further states that the “effective dimension” splits as

$$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$3

so that the spectral complexity of the joint estimator is inherited multiplicatively from the marginals (Zhou et al., 14 May 2026).

Under a source condition $$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$4, the reported phenomena include minimax-optimality when $$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$5, saturation plateau for $$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$6 in certain $$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$7-intervals, and periodic plateaux in $$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$8 together with multiple-descent behavior in $$k\bigl((x,\theta),(x',\theta')\bigr)=k_s(x,x')\,k_p(\theta,\theta').$$9 as $A_s=(k_s(x_i,x_j))_{i,j}\succeq 0$0 varies (Zhou et al., 14 May 2026). These results place state-parameter product kernels within the asymptotic theory of large-dimensional KRR rather than limiting them to interpolation and deterministic approximation.

6. Characteristicness, coupled systems, and terminological scope

For a two-component state-parameter pair $A_s=(k_s(x_i,x_j))_{i,j}\succeq 0$1, let

$A_s=(k_s(x_i,x_j))_{i,j}\succeq 0$2

If $A_s=(k_s(x_i,x_j))_{i,j}\succeq 0$3 and $A_s=(k_s(x_i,x_j))_{i,j}\succeq 0$4 are both characteristic, then $A_s=(k_s(x_i,x_j))_{i,j}\succeq 0$5 is $A_s=(k_s(x_i,x_j))_{i,j}\succeq 0$6-characteristic, so the associated HSIC characterizes independence of $A_s=(k_s(x_i,x_j))_{i,j}\succeq 0$7 and $A_s=(k_s(x_i,x_j))_{i,j}\succeq 0$8. On locally compact Polish domains, $A_s=(k_s(x_i,x_j))_{i,j}\succeq 0$9 is $A_p=(k_p(\theta_i,\theta_j))_{i,j}\succeq 0$0-universal if and only if each factor is $A_p=(k_p(\theta_i,\theta_j))_{i,j}\succeq 0$1-universal. For continuous, bounded, translation-invariant kernels on $A_p=(k_p(\theta_i,\theta_j))_{i,j}\succeq 0$2 and $A_p=(k_p(\theta_i,\theta_j))_{i,j}\succeq 0$3, characteristicness of the factors, characteristicness of the product, $A_p=(k_p(\theta_i,\theta_j))_{i,j}\succeq 0$4-characteristicness, $A_p=(k_p(\theta_i,\theta_j))_{i,j}\succeq 0$5-characteristicness, and $A_p=(k_p(\theta_i,\theta_j))_{i,j}\succeq 0$6-characteristicness are equivalent (Szabo et al., 2017).

A limitation is also explicit in the same source: for $A_p=(k_p(\theta_i,\theta_j))_{i,j}\succeq 0$7 factors, mere characteristicness of each factor need not suffice for the product to be $A_p=(k_p(\theta_i,\theta_j))_{i,j}\succeq 0$8-characteristic, whereas universality of each factor does suffice in arbitrary dimension and component count (Szabo et al., 2017). This addresses a common misconception that tensorisation preserves every desirable statistical property automatically once the marginals are individually well behaved.

In coupled-system models, one may have

$A_p=(k_p(\theta_i,\theta_j))_{i,j}\succeq 0$9

If the subsystems are uncoupled, the joint kernel is block-diagonal; if there are genuine cross-coupling terms through a coupling operator, off-diagonal blocks appear, and each block again factorises as a state-kernel and parameter-kernel pair while encoding cross-coupling (Matthies et al., 2018). Recursive factorisations of the parameter space similarly give hierarchical product kernels underpinning tensor-train or hierarchical-Tucker approximations (Matthies et al., 2018).

The term “state-parameter product kernel” also has a distinct usage in coagulation theory, where $A=A_s\circ A_p$0 denotes a collision kernel interpolating between constant-kernel and pure product-kernel regimes (Łepek et al., 2020). That usage is mathematically separate from RKHS product kernels on $A=A_s\circ A_p$1; the shared terminology reflects multiplicative structure, not a common reproducing-kernel framework.