pith. sign in

arxiv: 2604.26535 · v2 · pith:3VVYQYWKnew · submitted 2026-04-29 · 📊 stat.ME · cs.NA· math.NA

ARMA approximation of a Non-separable Spatio-Temporal Model with Fractional Smoothnesses in Space and Time

S. Knutsen Furset , Geir-Arne Fuglstad , Espen R. Jakobsen This is my paper

Pith reviewed 2026-05-07 12:40 UTC · model grok-4.3

classification 📊 stat.ME cs.NAmath.NA
keywords spatio-temporal covariancefractional smoothnessrational approximationVARMA processMatérn modelnon-separable modeldiscretizationtemperature data
0
0 comments X

The pith

Rational approximations in time turn a non-separable fractional spatio-temporal model into a convergent vector ARMA process.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a discretization for the diffusion-based extension of the spatial Matérn model to space and time, where fractional smoothness is allowed in both dimensions. It replaces the temporal fractional operator with a rational approximation that produces a vector autoregressive moving average process. The authors prove that the covariance of this approximation converges pointwise to the true covariance and supply explicit rates in terms of spatial resolution, temporal step size, and rational-approximation accuracy. Numerical experiments confirm that low-order VARMA processes already give small pointwise errors, and a simulation study shows that correctly specifying temporal smoothness improves forecasting. The approach is applied to three months of daily mean temperatures over mainland France.

Core claim

A rational approximation to the temporal fractional operator converts the continuous non-separable space-time fractional SPDE into a vector ARMA process whose covariance converges pointwise to the covariance of the original model, with explicit convergence rates controlled by the spatial mesh size, the temporal discretization step, and the accuracy of the rational approximation.

What carries the argument

Rational approximation of the temporal fractional differential operator, which yields a discrete-time vector ARMA process while preserving the non-separable structure of the covariance.

If this is right

  • The method removes prior restrictions on allowable temporal smoothness, enabling parameter estimation via standard VARMA techniques.
  • Forecast accuracy improves when the model is specified with the correct fractional temporal smoothness.
  • Error can be controlled predictably by choosing spatial mesh, time step, and rational order together.
  • The same discretization applies directly to real data sets such as regional temperature records.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same rational-approximation idea could be paired with other spatial discretizations to treat higher-dimensional or irregularly observed data.
  • The resulting VARMA representation opens the possibility of borrowing fast filtering and smoothing algorithms from multivariate time-series literature.
  • Performance under irregular observation times or missing data remains to be quantified.

Load-bearing premise

The rational approximation of the temporal fractional operator can be made arbitrarily accurate while keeping the non-separable covariance valid and without introducing uncontrolled bias.

What would settle it

Direct numerical comparison showing that the pointwise covariance difference exceeds the derived rate bound once the rational approximation order or grid resolution is increased.

Figures

Figures reproduced from arXiv: 2604.26535 by Espen R. Jakobsen, Geir-Arne Fuglstad, S. Knutsen Furset.

Figure 1
Figure 1. Figure 1: Sup-norm log10-error in covariance function of the rational approximation using M = 162 spatial basis function. Different colours correspond to different orders of approximation. The x-axis represents the temporal smoothness νt. The dashed vertical bars represent integer values of γ. (A) is rt = 1.0, βs = 0.25. (B) is rt = 1.0, βs = 0.50. (C) is rt = 3.0, βs = 0.25. (D) is rt = 3.0, βs = 0.50. All cases ha… view at source ↗
Figure 2
Figure 2. Figure 2: Violin-plots of the estimated parameters over the 30 replicates, for both the ’Full’ model and the ’Simple’ model. Top row only includes ’Full’ model since these parameters are not estimated in the ’Simple’ model. The true value of the parameter is marked in red. forecasts, we compute RMSE = vuut 1 R X R r=1 Z D  u n,r Msim (s) − u n,r M\inf (s) 2 ds = vuut 1 R X R r=1 M Xsim k=1  c n,r k − cdn,r k 2 ,… view at source ↗
Figure 3
Figure 3. Figure 3: Box-plots of RMSE and CRPS of both filtered and forecasted predictions on new datasets with 30 replicates, for both the ’Full’ model and the ’Simple’ model. βs. Further, there is a high spread in the estimates for the smoothness parameters νs and νt, and we observe overestimation of temporal smoothness and underestimation of spatial smoothness. The estimate of νs is better for higher measurement error vari… view at source ↗
Figure 4
Figure 4. Figure 4: Top: Daily mean temperature measurements in France for January 1, February 1 and March 1, 2023. Bottom: The three covariates we include in our model. Elevation (left), proximity to ocean (middle), and latitude (right). see the Data availability statement. The website provides temperature measurements from 10023 stations all over Europe, in the time period January 1, 1756, to December 31, 2025. We have filt… view at source ↗
Figure 5
Figure 5. Figure 5: Daily mean temperatures in France from January 1 to March 31 at two arbitrarily selected stations. Blue line displays one step forecasts with corresponding 95% blue shaded prediction region. Red line displays filtered predictions with corresponding 95% red shaded prediction region. The temperature measurements are displayed as black points. The approximate location of the station is displayed in the bottom… view at source ↗
Figure 6
Figure 6. Figure 6: Interpolated daily mean temperature in France on February 1, 2023. Filtered temperature means with linear model (left), prediction error (middle), and filtered temperature means without linear model (right). M = 82 (top), M2 = 122 (middle), and M = 162 (bottom). Axes are in kilometres view at source ↗
Figure 7
Figure 7. Figure 7: The five folds used for spatial cross validation. Station in the training set are marked in black; stations in the test set are marked in red. 18 view at source ↗
Figure 8
Figure 8. Figure 8: Cross-validation scores for √ M = 4, 5, 6, 7, 8, 9, 10, 11, 12, 16. ”o”-lines are for the ’Full’ model. ”x”-lines are scores for the ’Simple’ model. The same M was used both for parameter inference and prediction. In the simulation study, we are able to reliably identify the degree of non-separability in the model that generated the observations. In terms of prediction, the largest improvement over a separ… view at source ↗
read the original abstract

The Mat\'ern covariance model is ubiquitous in spatial modelling, but there is no default choice for spatio-temporal modelling. In this paper, we consider the recently proposed ``diffusion-based'' extension of the spatial Mat\'ern covariance model to a spatio-temporal non-separable covariance model that allows fractional smoothnesses in space and in time. The model is described in terms of a space-time fractional stochastic partial differential equation, but currently proposed computational approaches have strong restrictions on the possible smoothnesses in time. We propose a discretization method based on rational approximations in time to handle arbitrary smoothnesses, which leads to a vector autoregressive moving average process (VARMA). We prove that the covariance function of the approximation converges pointwise, determine explicit convergence rates as a function of spatial and temporal resolutions and the accuracy of the rational approximation, and conduct numerical verification to demonstrate small pointwise error for low orders of the VARMA process. Through a simulation study, we demonstrate that the parameters can be estimated back and that correctly specifying the temporal smoothness is especially important for forecasting. The approach is illustrated for three months of daily mean temperatures in mainland France.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper presents a method to approximate a non-separable spatio-temporal covariance model with fractional smoothnesses using rational approximations to the temporal fractional operator in the underlying SPDE, resulting in a VARMA process. It proves pointwise convergence of the covariance with explicit rates based on resolutions and approximation accuracy, verifies this numerically for low orders, shows via simulation that estimating the temporal smoothness correctly improves forecasting, and applies the approach to daily temperature data in France.

Significance. If the convergence result holds, this provides an important extension for computational inference in spatio-temporal statistics, allowing arbitrary temporal fractional smoothness without the restrictions of previous methods. The strengths include the proof of pointwise convergence and explicit rates, the numerical verification of small errors for low-order VARMA, and the simulation study demonstrating the practical importance of correct temporal smoothness specification for forecasting. This could enable more flexible modeling in geostatistics and environmental applications.

major comments (2)
  1. [§4 (Convergence results)] §4 (Convergence results), main theorem on pointwise convergence: the derivation establishes rates depending on spatial/temporal mesh sizes and rational approximation accuracy, but does not explicitly demonstrate that the frequency-dependent error of the rational approximant to the temporal operator remains controlled in the full non-separable space-time covariance kernel (i.e., no additional O(1) bias terms arise from coupling to the spatial fractional Laplacian across arbitrary lags). This is load-bearing for the central claim that the VARMA approximation is valid for general use including forecasting.
  2. [Numerical verification section] Numerical verification section and associated tables: the reported small pointwise errors for low-order approximations are shown for selected cases, but the experiments do not include a systematic sweep over a wider range of space-time lags or smoothness parameters near the boundary of validity; this leaves open whether the error bound holds uniformly as required by the non-separable structure.
minor comments (2)
  1. [Abstract] Abstract: the description of the rational approximation step would benefit from naming the specific technique (e.g., Padé approximants or pole-residue method) to aid immediate reproducibility.
  2. [Simulation study] Simulation study: additional detail on how the VARMA likelihood is computed and optimized (e.g., state-space representation or exact likelihood) would clarify the estimation procedure.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive overall assessment and for the constructive comments on the convergence analysis and numerical verification. We address each major comment below and have revised the manuscript to strengthen the presentation of the results.

read point-by-point responses

  1. Referee: [§4 (Convergence results)] §4 (Convergence results), main theorem on pointwise convergence: the derivation establishes rates depending on spatial/temporal mesh sizes and rational approximation accuracy, but does not explicitly demonstrate that the frequency-dependent error of the rational approximant to the temporal operator remains controlled in the full non-separable space-time covariance kernel (i.e., no additional O(1) bias terms arise from coupling to the spatial fractional Laplacian across arbitrary lags). This is load-bearing for the central claim that the VARMA approximation is valid for general use including forecasting.

    Authors: We appreciate the referee drawing attention to this point. The proof of the main theorem in Section 4 is carried out in the joint space-time Fourier domain, where the error introduced by the rational approximation to the temporal operator is bounded uniformly in frequency (by standard results on rational approximation of fractional powers). This error is then multiplied by the symbol of the spatial fractional Laplacian, whose contribution is controlled by the assumed smoothness parameters and does not produce an additional O(1) term that would persist across arbitrary lags. The resulting pointwise rates therefore already incorporate the non-separable coupling. To make the argument fully explicit, we have inserted an intermediate bound and a short clarifying paragraph immediately after the statement of the theorem. revision: yes

  2. Referee: [Numerical verification section] Numerical verification section and associated tables: the reported small pointwise errors for low-order approximations are shown for selected cases, but the experiments do not include a systematic sweep over a wider range of space-time lags or smoothness parameters near the boundary of validity; this leaves open whether the error bound holds uniformly as required by the non-separable structure.

    Authors: We agree that the original numerical section, while supportive, was limited in scope. In the revised version we have added a new set of experiments that systematically vary both the space-time lag (including larger temporal separations) and the temporal smoothness parameter over a grid that approaches the lower and upper boundaries of the admissible range. The additional tables and figures confirm that the observed pointwise errors remain small and consistent with the theoretical rates, thereby providing stronger numerical evidence for uniform control under the non-separable structure. revision: yes

Circularity Check

0 steps flagged

No significant circularity; approximation and convergence proof are independent of inputs.

full rationale

The paper extends a cited diffusion-based SPDE model for non-separable spatio-temporal Matérn covariance with fractional smoothness parameters, but the central derivation introduces a new rational approximation in time to produce a VARMA discretization. The abstract states that pointwise convergence of the covariance is proved with explicit rates depending on mesh resolutions and rational order, supported by separate numerical verification. No self-definitional steps, fitted parameters renamed as predictions, or load-bearing self-citations that reduce the claimed convergence or VARMA equivalence to prior inputs by construction are present. The rational approximation step and its error analysis stand as independent contributions.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the validity of the underlying fractional SPDE as a covariance generator and on the existence of rational approximations that converge to the temporal fractional operator while preserving non-separability.

free parameters (1)
  • Order of rational approximation
    Chosen to control temporal discretization accuracy; higher orders improve convergence but increase computational cost.
axioms (1)
  • domain assumption The diffusion-based extension of the spatial Matérn model to a non-separable space-time fractional SPDE defines a valid positive-definite covariance.
    Invoked as the starting point for the discretization; assumed from prior literature on the model.

pith-pipeline@v0.9.0 · 5520 in / 1307 out tokens · 52714 ms · 2026-05-07T12:40:15.361792+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.