Connecting reflective asymmetries in multivariate spatial and spatio-temporal covariances

In the analysis of multivariate spatial and univariate spatio-temporal data, it is commonly recognized that asymmetric dependence may exist, which can be addressed using an asymmetric (matrix or space-time, respectively) covariance function within a Gaussian process framework. This paper introduces a new paradigm for constructing asymmetric space-time covariances, which we refer to as “reflective asymmetric,” by leveraging recently-introduced models for multivariate spatial data. We first provide new results for reflective asymmetric multivariate spatial models that extends their applicability. We then propose their asymmetric space-time extension, which come from a substantially different perspective than Lagrangian asymmetric space-time covariances. There are fewer parameters in the new models, one controls both the spatial and temporal marginal covariances, and the standard separable model is a special case. In simulation studies and analysis of the frequently-studied Irish wind data, these new models also improve model fit and prediction performance, and they can be easier to estimate. These features indicate broad applicability for improved analysis in environmental and other space-time data.

💡 Research Summary

The paper introduces a novel class of “reflective asymmetric” covariance functions for space‑time Gaussian processes, built on recent developments in multivariate spatial statistics. The authors first extend the asymmetric cross‑covariance construction of Yarger et al. (2026), which multiplies the spatial spectral density by the factor (-i,\text{sign}(\langle x,\tilde x\rangle)). While earlier results were limited to one‑dimensional space, the new work derives closed‑form expressions for squared‑exponential and Cauchy kernels in any spatial dimension (d). The resulting cross‑covariance decomposes into a symmetric part (the real component of a positive‑definite matrix (\sigma)) and an antisymmetric part (the imaginary component), the latter being an odd function reflected about the origin—hence the term “reflective.”

To obtain space‑time covariances, the same spectral trick is applied jointly to spatial and temporal frequencies: the symmetric space‑time spectral density (f(x,\eta)) is multiplied by (\text{sign}(\langle x,\tilde x\rangle),\text{sign}(\eta)). Consequently the full covariance separates into a conventional symmetric, non‑separable component (e.g., a Gneiting‑type kernel) and a reflective asymmetric term. This construction differs fundamentally from the widely used Lagrangian approach, where asymmetry arises from advective transport (V) and forces the temporal marginal covariance to depend on the spatial parameters. In the reflective framework the temporal marginal (C(0,u)) can be specified independently, granting greater flexibility and improving the asymptotic behavior of estimators.

Parameter‑wise, the reflective models are parsimonious. Apart from the usual variance, spatial and temporal range parameters, and a non‑separability weight (b), only the complex off‑diagonal entries of (\sigma) (specifically their imaginary parts) are needed to induce asymmetry. In contrast, a full Lagrangian model requires a mean vector and a full covariance matrix for the transport vector (V), leading to (1+d+d(d+1)/2) parameters, which grows rapidly with dimension. The reduced parameter set simplifies likelihood optimization and reduces the risk of over‑fitting, especially in high‑dimensional settings.

Computationally, the authors adopt Vecchia’s approximation to obtain sparse precision matrices, enabling scalable inference for large space‑time datasets. They also outline unconditional simulation via spectral sampling, extensions to non‑stationary and anisotropic settings, and a likelihood‑ratio test for detecting reflective asymmetry.

Simulation experiments compare three families: (i) symmetric Gneiting models, (ii) Lagrangian models, and (iii) reflective asymmetric models. When data are generated from a reflective asymmetric process, the reflective models achieve higher log‑likelihoods, lower prediction RMSE, and more accurate parameter recovery than the Lagrangian alternatives. Conversely, when the true process follows a Lagrangian structure, the Lagrangian models outperform the reflective ones, confirming that the two approaches are not interchangeable and that model selection is crucial.

The methodology is applied to the classic Irish wind dataset (Haslett & Raftery, 1989). Five model variants are fitted: a symmetric Gneiting model, a Lagrangian model, and three reflective asymmetric models differing in their marginal spatial and temporal kernels. Across AIC, BIC, cross‑validated log‑likelihood, and out‑of‑sample prediction error, the reflective asymmetric models consistently rank best. Notably, the ability to model the temporal marginal independently yields more realistic seasonal and trend components, and the estimation converges faster with fewer numerical issues than the Lagrangian counterpart.

The discussion acknowledges limitations: the reflective asymmetry is introduced solely via sign functions in the spectral domain, which may not capture more complex, nonlinear asymmetries observed in practice. Extensions to richer spectral transformations, non‑stationary spectral densities, and fully Bayesian treatments are suggested as future work. Nonetheless, the paper demonstrates that reflective asymmetric covariances provide a parsimonious, flexible, and computationally tractable alternative to Lagrangian models for a wide range of environmental and spatio‑temporal applications.