Positive Definite Covariance Functions

Updated 17 October 2025

Positive Definite Covariance Functions are symmetric functions that guarantee any covariance matrix formed from them is positive semidefinite, essential for valid Gaussian process models.
They are characterized by fundamental theorems like Bochner’s and Schoenberg’s, enabling translation-invariant and isotropic modeling on various domains including spheres and manifolds.
Operator-based constructions and advanced estimation methods ensure these functions exhibit desired smoothness, sparsity, and numerical stability in high-dimensional applications.

A positive definite covariance function is a symmetric function $C(s, t)$ defined on a domain $D \times D$ such that, for any finite collection of points $s_1, \ldots, s_n \in D$ , the matrix $[C(s_i, s_j)]_{i,j=1}^n$ is positive semidefinite, i.e., all its eigenvalues are nonnegative. This property underpins the feasibility of modeling Gaussian processes and multivariate random fields, as it ensures that the implied covariance matrix is valid for simulation, statistical inference, and prediction. The theory of positive definite covariance functions is central to spatial statistics, Gaussian process modeling, machine learning kernels, and the functional analytic structure of reproducing kernel Hilbert spaces.

1. Definition, Characterizations, and Structural Properties

A function $C : D \times D \rightarrow \mathbb{C}$ is positive definite if for all finite choices $s_1, ..., s_n \in D$ and all $a_1, ..., a_n \in \mathbb{C}$ ,

$\sum_{i=1}^n \sum_{j=1}^n a_i \overline{a_j} C(s_i, s_j) \geq 0.$

This condition is both necessary and sufficient for $C$ to serve as the covariance function of a second-order random field defined over $D$ .

A core result in the construction of positive definite functions is Bochner’s theorem: For $f:\mathbb{R}^d \to \mathbb{C}$ continuous, $f$ is positive definite and translation-invariant (i.e., $f(s, t) = \varphi(s-t)$ ) if and only if

$\varphi(h) = \int_{\mathbb{R}^d} e^{i \langle h, u \rangle} dF(u),$

where $F$ is a finite positive measure on $\mathbb{R}^d$ . Schoenberg’s theorem (Godoy et al., 21 Feb 2025) extends this to isotropic/radial functions: a continuous function $f(|x|)$ is positive definite on $\mathbb{R}^d$ if and only if it has an integral representation involving Bessel functions and nonnegative measures.

Conditional negative definiteness plays a crucial role for extending these concepts to non-Euclidean domains (Godoy et al., 21 Feb 2025): if a metric $d$ on $D$ is such that $d^2$ is conditionally negative definite, then $D$ can be isometrically embedded into a Hilbert space, and popular positive definite families (Matérn, powered exponential, etc.) can be defined over $(D, d)$ by replacing the Euclidean norm with the appropriate metric.

2. Classes, Constructions, and Special Families

Isotropic and Radially Positive Definite Functions

Isotropic covariance functions depend only on the distance between locations: $C(s, t) = \psi(d(s, t))$ . For Euclidean spaces, such functions are characterized by the class $\Phi_d$ (Faouzi et al., 2018, Porcu et al., 2016), and their representation is given by Schoenberg's integral.

On compact two-point-homogeneous spaces such as spheres, classes $\Psi_d$ are used, where isotropic functions are written in terms of positive combinations of Gegenbauer or Legendre polynomials (Jeong et al., 2015, White et al., 2018).

Operator-based Constructions

Functional operators allow transformation and design of covariance functions with prescribed smoothness or support:

The Zastavnyi operator (Faouzi et al., 2018) acts on pairs of scale parameters and allows mixing and tuning of differentiability and potential negative correlation. For instance, applied to Matérn, Generalized Cauchy, and Generalized Wendland families, it enables precise control over smoothness at the origin and introduces more flexible correlation structures.
Difference operators applied to compactly supported radial basis functions (Buhmann class) can increase the smoothness at the origin while preserving both positive definiteness and compact support on the intended ambient dimension (Porcu et al., 2016).

Matrix-Valued and Cross Covariance Functions

In multivariate or vector-valued settings, covariance functions become matrix-valued and must satisfy blockwise positive definiteness:

For spatial–temporal and cross-variable models, classes extending Gneiting's space–time covariance framework are available (Bourotte et al., 2015), with positive definiteness enforced via scale mixing and explicit parameter ties among cross-covariances.
Sufficient Pólya-type derivative conditions provide tractable validation for bivariate models from powered exponential and generalized Cauchy families (Moreva et al., 2016).

3. Manifold, Non-Euclidean, and Geometric Extensions

Covariance modeling often requires handling domains with non-Euclidean geometry: e.g., spheres, manifolds of SPD matrices.

On spheres, the positive definiteness of isotropic functions using great circle distance is subtle. While projections from Euclidean models via one-to-one mappings are possible, this can lead to physical distortion at large distances. Models defined directly on the sphere (e.g., using canonically normalized Gegenbauer expansions) are positive definite for all spherical dimensions and better suited for global data (Jeong et al., 2015, White et al., 2018).
For the space of Hermitian or symmetric positive definite matrices, the geometry is that of a Riemannian manifold (with affine-invariant metric). Intrinsic data depth functions, such as the intrinsic zonoid and geodesic distance depths, allow robust and efficient statistical inference, rankings, and construction of confidence sets on collections of covariance matrices (Chau et al., 2017).
Multifidelity estimation and regression on the SPD manifold guarantee preservation of positive definiteness and allow for intrinsic Mahalanobis minimization and maximum likelihood estimation under manifold tangent-space error models (Maurais et al., 2023).

4. Estimation, Algorithmic Guarantee, and Sparse Structures

Guaranteeing positive definiteness in practical covariance estimation—especially in high dimensions—is nontrivial.

Penalized Estimation and Constraints: The positive definite $\ell_1$ -penalized estimator (Xue et al., 2012) combines off-diagonal $\ell_1$ regularization (for sparsity) with an eigenvalue constraint $S \succeq \epsilon I$ to ensure strict positive definiteness. The solution is efficiently computed using an alternating direction method with explicit projection onto the PSD cone.
Condition Number Constraints: To produce stably invertible matrices, condition number constraints ( $\lambda_{\max}/\lambda_{\min} \leq \kappa$ ) are employed. These lead to spectral truncation of eigenvalues and efficient solution path algorithms (Oh et al., 2015).
Generalized Nonconvex Penalties: For bias reduction, nonconvex penalties such as SCAD, hard-thresholding, and $L_q$ norm are used in the regularization. Positive definiteness is maintained by spectral projection after each iteration (Wen et al., 2016).
Sparse Linear Models with Fixed Diagonal: Recent approaches formulate penalized linear regression on the vectorized covariance, enforcing exact matching of sample variances (diagonals) and constraining the parameter to the PSD cone. This avoids biasing variance estimation and combines sparsity with validity (Kim et al., 12 Mar 2025).
Nonparametric Regression with Evolutionary Algorithms: Positive definite estimators are constructed from Bochner’s integral representation using kernel-based distribution surrogates, optimized via iterated density estimation evolutionary algorithms; this offers modularity and guarantees by design (Kang, 2023).

5. Harmonic Analysis, Kernel Factorizations, and Boundary Theory

The harmonic-analytic structure of positive definite functions underlies both their representation and applications:

Bochner and Bochner–Schwartz Theorems: Any continuous translation-invariant positive definite function is a Fourier transform of a nonnegative measure (Bochner); this generalizes to tempered distributions via the Bochner–Schwartz theorem (Jorgensen et al., 2017).
RKHS and Kernel Factorizations: Every positive definite kernel defines a unique reproducing kernel Hilbert space (RKHS) (Jorgensen et al., 2019). A new correspondence is established between a kernel $K$ and its factorization via measure spaces $(B, \mu)$ and functions $k_x^{(B)}$ such that

$K(x, y) = \int_B \overline{k_x^{(B)}(b)} k_y^{(B)}(b) d\mu(b).$

This factorization allows embedding the RKHS into $L^2(B, \mu)$ , yielding canonical representations, operator dualities, and interpretations as feature spaces for machine learning models.

Probability and Gaussian Processes: Any p.d. kernel is the covariance kernel of a Gaussian process; canonical isometries (Ito–isometry) map kernel sections to the process realizations in $L^2$ (Jorgensen et al., 2019).
Boundary Analysis and Partial Orders: Measure-theoretic “boundaries” for kernels (including fractal boundaries from iterated function systems, e.g., the Cantor set) induce decompositions of kernels. Partial orders among kernel families based on quadratic forms allow for comparative and constructive analysis.
Harmonic Expansion on Gelfand Pairs: Invariant positive definite functions on products of homogeneous spaces are characterized via integral/spectral expansions with respect to the Plancherel measure on the duals, generalizing well-known spatio-temporal covariance results to very general settings (Berg, 2020).

6. Multivariate, Space-Time, and Nonseparable Extensions

Extensions to multivariate and space–time settings require joint positive definiteness conditions:

Parametric Matrix-Valued Covariances: Sufficient conditions for positive definiteness in cross-covariance (multivariate) space–time models are developed via mixture/scale representations and explicit time–space parameterizations (e.g., multivariate Gneiting–Matérn, Cauchy families). Sufficient parameter relations guarantee positive definiteness (Bourotte et al., 2015).
Nonseparable Kernels and Multivariate Interpolation: Abstract constructions using bounded completely monotone functions and multivariate matrix–vector structures yield a flexible class of nonseparable positive definite kernels for cross-domain problems (Menegatto et al., 2021).
Pólya-type Criteria: Sufficient Pólya-type derivative conditions for bivariate/multivariate models provide constructible and flexible powered exponential and Cauchy class cross-covariances (Moreva et al., 2016).

7. Independence, Kernel Methods, and Positive Definite Independent Kernels

Kernel-based criteria for independence, such as HSIC and distance covariance, are traditionally built from the Kronecker product of positive definite or conditionally negative definite kernels. Recent work generalizes this by developing positive definite independent (PDI) kernels (Guella, 2022):

PDI kernels are defined on $X \times Y \times X \times Y$ with a double-centering positivity condition. Radial PDI kernels are completely characterized as continuous Bernstein functions of two variables:

$g(\|x-y\|^2, \|u-v\|^2)$

with an explicit integral representation.

This characterization encompasses classical tests as special cases and allows for the construction of new, nonseparable independence criteria, broadening the arsenal of dependence measures in statistics and machine learning.

References to Major Results

(Godoy et al., 21 Feb 2025): Sufficient conditions and frameworks for extending PD covariance functions to non-Euclidean domains.
(Jorgensen et al., 2019): Factorization, partial order, and boundaries for p.d. kernels; applications to Gaussian processes and feature space identification.
(Xue et al., 2012, Wen et al., 2016, Kim et al., 12 Mar 2025): Sparse and high-dimensional covariance estimation with positive definiteness guarantee.
(Jeong et al., 2015, White et al., 2018): Covariance models on the sphere and space–time extensions; impact of geometry on PD conditions and spatial prediction.
(Bourotte et al., 2015): Matrix-valued space–time cross-covariances with flexible parametric structure and sufficient PD conditions.
(Berg, 2020): Harmonic analysis, Gelfand pairs, and spectral expansion for covariance construction.
(Guella, 2022): PDI kernels, Bernstein characterizations, and generalizations of HSIC/dCov for independence testing.

Positive definite covariance functions thus constitute a rich mathematical structure highly relevant for probabilistic modeling, spatial statistics, machine learning, and functional analysis. The interplay between geometry, harmonic analysis, algebraic and analytic conditions, and computational implementation continues to drive discovery and application across fields.