arxiv: 2605.03693 · v1 · submitted 2026-05-05 · 📊 stat.CO · stat.ME

Recognition: unknown

A new framework for non-stationary spatio-temporal data fusion of multi-fidelity models

Claire Miller, Fabio Sigrist, Paolo Maranzano, Pietro Colombo, Ruth O'Donnell, Xiaochen Yang

Authors on Pith no claims yet

Pith reviewed 2026-05-09 15:35 UTC · model grok-4.3

classification 📊 stat.CO stat.ME

keywords multi-fidelity Gaussian processesVecchia approximationspatio-temporal data fusionnon-stationary processescovariance decompositionenvironmental data reconstructionscalable likelihood inference

0 comments

The pith

A decomposed multi-fidelity covariance lets Vecchia approximations compute joint likelihoods for non-stationary spatio-temporal fusion without forming the full matrix.

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

The paper introduces a scalable framework for fusing abundant noisy low-fidelity data such as satellite products with sparse accurate high-fidelity observations in space and time. Its core contribution is a covariance decomposition that separates the latent low-fidelity process from the discrepancy process, allowing the Vecchia approximation to operate directly on each without ever assembling the complete joint covariance. A Woodbury identity step then reconstructs the necessary quantities for stable likelihood evaluation. The method adds a generalized least squares step with fidelity-specific offsets to keep systematic biases out of the cross-fidelity dependence structure. This combination supports fully likelihood-based inference for both stationary and non-stationary cases and is demonstrated on synthetic tests plus a large wind-speed reconstruction task.

Core claim

The central claim is that a decomposed multi-fidelity covariance formulation permits the Vecchia approximation to be applied separately to the latent low-fidelity and discrepancy processes. When combined with Woodbury-based reconstruction, the formulation produces a numerically stable and computationally efficient joint marginal likelihood without ever constructing the full multi-fidelity covariance matrix. A generalized least squares mean-removal procedure with fidelity-specific offsets prevents biases from being absorbed into the dependence parameters.

What carries the argument

The decomposed multi-fidelity covariance formulation, which isolates the low-fidelity latent process and the discrepancy process so that Vecchia approximation and Woodbury reconstruction can be performed independently on each.

If this is right

The joint marginal likelihood can be evaluated stably for large non-stationary spatio-temporal datasets.
Predictive accuracy, correlation, and local variability improve over single-fidelity spatio-temporal Gaussian processes when low- and high-fidelity sources are fused.
The method produces results that closely match exact multi-fidelity inference on controlled small problems.
High-resolution environmental reconstructions become feasible from realistic volumes of mixed-fidelity observations.

Where Pith is reading between the lines

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

The same decomposition idea could be paired with other sparse approximations beyond Vecchia for multi-fidelity models.
Fidelity-specific GLS offsets may extend to bias correction in other hierarchical spatio-temporal models.
The approach could support sequential updating when new high-fidelity observations arrive over time.

Load-bearing premise

The proposed decomposition of the multi-fidelity covariance must preserve the exact joint distribution properties so that the Vecchia approximation remains valid and the GLS bias removal works correctly in non-stationary settings.

What would settle it

On small synthetic datasets where the exact multi-fidelity Gaussian process likelihood can be computed directly, compare the approximated marginal likelihood values, parameter estimates, and predictive distributions; large discrepancies would show that the decomposition fails to preserve the required joint properties.

Figures

Figures reproduced from arXiv: 2605.03693 by Claire Miller, Fabio Sigrist, Paolo Maranzano, Pietro Colombo, Ruth O'Donnell, Xiaochen Yang.

**Figure 1.** Figure 1: Notice how generally the correlation conditioning is more precise than the view at source ↗

**Figure 1.** Figure 1: The figure illustrates predictions from a representative single replication of the view at source ↗

**Figure 2.** Figure 2: Illustration of the South Lombardy dataset. The blue dots depicts the position of view at source ↗

**Figure 3.** Figure 3: Spatial and temporal performance analysis of the MFGP framework across the view at source ↗

**Figure 4.** Figure 4: The figure depicts the comparison between view at source ↗

read the original abstract

We propose a new scalable framework for spatio-temporal data fusion with multi-fidelity Gaussian processes (MFGPs) that enables fully likelihood-based inference for both stationary and non-stationary fidelity integration. The framework is designed for environmental applications, where abundant but noisy low-fidelity data (e.g., satellite or reanalysis products) must be fused with sparse yet accurate high-fidelity in-situ observations to obtain high-resolution reconstructions. Our key methodological contribution is a decomposed multi-fidelity covariance formulation that allows the Vecchia approximation to be applied directly to the latent low-fidelity and discrepancy processes. Combined with a Woodbury-based reconstruction, this yields a numerically stable and computationally efficient evaluation of the joint marginal likelihood without ever forming the full multi-fidelity covariance matrix. In addition, we introduce a generalized least squares (GLS) mean-removal strategy with fidelity-specific offsets, preventing systematic biases from being absorbed into cross-fidelity dependence. We validate the proposed approach through extensive experiments on synthetic data and a large-scale real-world application to wind speed reconstruction in the Lombardy region of Italy. The results show that the proposed Vecchia-based MFGP closely matches exact multi-fidelity inference in controlled settings, while substantially outperforming standard single-fidelity spatio-temporal Gaussian processes in terms of predictive accuracy, correlation, and representation of local variability in realistic large-data scenarios.

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.

Desk Editor's Note private letter to a colleague

The decomposed covariance for Vecchia on separate low-fidelity and discrepancy latents, plus fidelity-specific GLS, is the actual new piece and looks workable for scalable non-stationary fusion.

read the letter

The main thing to know is that the authors decompose the multi-fidelity covariance so the Vecchia approximation can be applied directly to the latent low-fidelity process and the discrepancy process, then recover the joint marginal likelihood via Woodbury without ever building the full matrix. They pair this with a GLS mean-removal step that uses separate offsets per fidelity level to keep systematic biases from leaking into the cross-fidelity terms. Both pieces target the practical problem of fusing lots of noisy low-fidelity data (satellite, reanalysis) with sparse high-fidelity observations in non-stationary spatio-temporal settings like wind speed mapping. On synthetic data they report that the approximation recovers exact multi-fidelity inference, and on the Lombardy wind example it beats single-fidelity GPs on predictive accuracy, correlation, and local variability representation. That combination of scalability and real-world demonstration is the useful part. The soft spot is whether the decomposition is an exact algebraic identity that preserves the joint distribution properties required for the Vecchia factors plus Woodbury correction to equal the true likelihood (up to the usual truncation). Non-stationary kernels make the cross-covariance blocks location-dependent, so any hidden assumption that the conditional distributions stay Markovian after splitting could matter. The paper claims the match to exact inference on synthetics, which is reassuring, but an explicit step-by-step derivation showing the identity holds before sparsity is introduced would remove the doubt. The GLS commuting with the sparse precision is a related minor point that deserves a short check in the supplement. This is for people who already work with Vecchia or multi-fidelity GPs and need to handle large environmental datasets. A reader looking for a concrete, implementable extension of existing tools will get value from the formulation and the Lombardy results. It deserves a serious referee because the motivation is clear, the computational claim is testable, and the experiments are on the right scale even if the algebraic details need a bit more unpacking in revision.

Referee Report

2 major / 2 minor

Summary. The paper proposes a scalable framework for non-stationary spatio-temporal data fusion of multi-fidelity Gaussian processes (MFGPs). Its key contribution is a decomposed multi-fidelity covariance formulation that allows the Vecchia approximation to be applied directly to the latent low-fidelity process and the discrepancy process, combined with a Woodbury-based reconstruction to evaluate the joint marginal likelihood efficiently without forming the full covariance matrix. It further introduces a GLS mean-removal strategy with fidelity-specific offsets to prevent bias absorption into cross-fidelity dependence. The approach is validated on synthetic data (claimed to match exact inference) and a real-world wind speed reconstruction application in Lombardy, Italy, where it outperforms single-fidelity spatio-temporal GPs.

Significance. If the decomposition is an exact algebraic identity that preserves joint distribution properties under non-stationary kernels, the framework would enable practical, likelihood-based multi-fidelity inference at scales relevant to environmental applications with abundant low-fidelity and sparse high-fidelity data. It extends established Vecchia and GP methods with a novel decomposition for computational efficiency and numerical stability, and the real-data experiment provides a concrete demonstration of utility in spatio-temporal reconstruction.

major comments (2)

Decomposed multi-fidelity covariance formulation (key methodological contribution): the claim that Vecchia can be applied directly to the separate latent low-fidelity and discrepancy processes with Woodbury reconstruction requires that the decomposition be an exact algebraic identity preserving the joint covariance structure. In non-stationary spatio-temporal settings the cross-covariance blocks are position-dependent; it is unclear whether these blocks factor such that the conditional distributions remain compatible with Vecchia sparsity on the latents before reconstruction. An explicit derivation or numerical verification that the product of the two Vecchia factors plus Woodbury correction recovers the true joint density (up to truncation) is needed, as this is load-bearing for the central scalability claim.
GLS mean-removal strategy with fidelity-specific offsets: the approach assumes mean removal commutes with the decomposed precision after Vecchia sparsity is introduced. This commutation is not automatic for non-stationary kernels whose cross-covariance blocks depend on location; any violation would affect both bias removal and the validity of the reconstructed marginal likelihood. A concrete check (e.g., via the precision-matrix form or a small-scale counter-example) should be provided in the relevant section.

minor comments (2)

Abstract: the statements that the Vecchia-based MFGP 'closely matches exact multi-fidelity inference' and 'substantially outperforming' single-fidelity GPs would be strengthened by inclusion of at least one quantitative metric (e.g., RMSE, log-likelihood difference, or correlation value) with error bars or confidence intervals from the synthetic and real-data experiments.
Notation: the distinction between the latent low-fidelity process, discrepancy process, and observed fields could be made more explicit when first introducing the decomposed covariance to aid readability for readers unfamiliar with multi-fidelity GP decompositions.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their insightful comments and for recognizing the potential of our framework for scalable multi-fidelity spatio-temporal data fusion. We address each of the major comments below, providing clarifications and committing to revisions that strengthen the manuscript's rigor without altering the core contributions.

read point-by-point responses

Referee: Decomposed multi-fidelity covariance formulation (key methodological contribution): the claim that Vecchia can be applied directly to the separate latent low-fidelity and discrepancy processes with Woodbury reconstruction requires that the decomposition be an exact algebraic identity preserving the joint covariance structure. In non-stationary spatio-temporal settings the cross-covariance blocks are position-dependent; it is unclear whether these blocks factor such that the conditional distributions remain compatible with Vecchia sparsity on the latents before reconstruction. An explicit derivation or numerical verification that the product of the two Vecchia factors plus Woodbury correction recovers the true joint density (up to truncation) is needed, as this is load-bearing for the central scalability claim.

Authors: We appreciate the referee pointing out the need for explicit verification of the decomposition's properties. The multi-fidelity covariance decomposition is an exact algebraic identity derived from the additive structure of the low-fidelity and discrepancy processes, which holds regardless of stationarity. The Vecchia approximation is applied independently to each latent process, preserving their individual sparsity patterns, while the Woodbury matrix identity reconstructs the joint marginal likelihood from the low-rank updates induced by the cross-fidelity terms. This ensures the approximated joint density matches the true one up to the truncation error of the Vecchia approximations on the latents. We will add a detailed derivation in Section 3.2 of the revised manuscript, including the step-by-step algebraic expansion showing compatibility of the conditional distributions. Additionally, we will include numerical verification by comparing the approximated log-likelihood values to exact computations on small non-stationary synthetic datasets. revision: yes
Referee: GLS mean-removal strategy with fidelity-specific offsets: the approach assumes mean removal commutes with the decomposed precision after Vecchia sparsity is introduced. This commutation is not automatic for non-stationary kernels whose cross-covariance blocks depend on location; any violation would affect both bias removal and the validity of the reconstructed marginal likelihood. A concrete check (e.g., via the precision-matrix form or a small-scale counter-example) should be provided in the relevant section.

Authors: We agree that explicit verification of the commutation property is warranted, particularly in non-stationary settings. In our framework, the GLS mean removal with fidelity-specific offsets is performed on the observed data prior to applying the Vecchia approximation to the covariance structure. This ordering ensures that the mean estimation does not interfere with the sparsity pattern of the precision matrices for the latent processes. The fidelity-specific offsets are estimated jointly via GLS on the decomposed covariance, preventing bias absorption. To address the concern, we will provide a concrete check in the revised manuscript, including a small-scale numerical example and an analysis of the precision matrix form demonstrating that the mean removal commutes appropriately under the non-stationary cross-covariance structure. If any adjustments are needed based on this check, they will be incorporated. revision: yes

Circularity Check

0 steps flagged

No circularity: novel decomposition is presented as algebraic identity independent of fitted outputs

full rationale

The paper's central contribution is a decomposed multi-fidelity covariance that permits direct Vecchia application to separate latent low-fidelity and discrepancy processes, followed by Woodbury reconstruction of the joint likelihood. This is introduced as a new formulation rather than derived from or equivalent to any fitted parameter or self-citation chain. No equation reduces a claimed prediction to a quantity defined solely by the inputs (e.g., no 'predict the ratio that was already fitted'). GLS mean removal with fidelity offsets is likewise a proposed strategy, not a renaming of known results. External Vecchia and GP literature is cited for the approximation technique itself, but the load-bearing novelty (decomposition preserving joint properties for non-stationary spatio-temporal kernels) is self-contained and validated against exact inference on synthetic data. No self-definitional, fitted-input, or uniqueness-imported steps appear in the derivation.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The framework rests on standard Gaussian process modeling assumptions and the validity of the new covariance decomposition for Vecchia and GLS steps; hyperparameters of the kernels and offsets are fitted during likelihood maximization.

free parameters (2)

fidelity-specific offsets
Introduced in the GLS mean-removal strategy and fitted to data to prevent bias absorption.
kernel hyperparameters
Standard GP length scales, variances, and noise terms fitted via marginal likelihood.

axioms (2)

domain assumption The multi-fidelity processes follow a joint Gaussian distribution that admits the proposed decomposition.
Invoked to justify applying Vecchia separately to latent low-fidelity and discrepancy processes.
domain assumption Vecchia approximation remains accurate and stable under the decomposed covariance structure.
Central to the efficiency claim but not proven in the abstract.

pith-pipeline@v0.9.0 · 5552 in / 1379 out tokens · 35219 ms · 2026-05-09T15:35:41.626706+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

47 extracted references · 2 canonical work pages

[1]

Interpreting Mixed Membership Models: Implications of Erosheva's Representation Theorem , Year =

April Galyardt , Booktitle =. Interpreting Mixed Membership Models: Implications of Erosheva's Representation Theorem , Year =
[2]

Mixed Membership Distributions with Applications to Modeling Multiple Strategy Usage , Year =

April Galyardt , Date-Added =. Mixed Membership Distributions with Applications to Modeling Multiple Strategy Usage , Year =
[3]

Modeling Student Metacognitive Strategies in a Intelligent Tutoring System , Year =

April Galyardt and Ilya Goldin , Booktitle =. Modeling Student Metacognitive Strategies in a Intelligent Tutoring System , Year =
[4]

Campbell and Shauna Austin , Date-Added =

Jamie I.D. Campbell and Shauna Austin , Date-Added =. Effects of response time deadlines on adults' strategy choices for simple addition , Volume =. Memory & Cognition , Keywords =
[5]

Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , volume=

Nonlinear information fusion algorithms for data-efficient multi-fidelity modelling , author=. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , volume=. 2017 , publisher=

2017
[6]

arXiv preprint arXiv:1604.07484 , year=

Deep multi-fidelity Gaussian processes , author=. arXiv preprint arXiv:1604.07484 , year=

work page arXiv
[7]

2024 , url =

Pietrostat193 , title =. 2024 , url =

2024
[8]

Characterizing novice-expert differences in macrocognition: an exploratory study of cognitive work in the emergency department , Volume =

Schubert, Christiane C and Denmark, T Kent and Crandall, Beth and Grome, Anna and Pappas, James , Date-Added =. Characterizing novice-expert differences in macrocognition: an exploratory study of cognitive work in the emergency department , Volume =. Annals of emergency medicine , Keywords =. 2013 , Bdsk-File-1 =

2013
[9]

Siegler , Date-Added =

Robert S. Siegler , Date-Added =. The perils of averaging data over strategies: An example from children's addition , Volume =. Journal of Experimental Psychology: General , Keywords =. 1987 , Bdsk-File-1 =

1987
[10]

and Feltovich, Paul J

Chi, Michelene T.H. and Feltovich, Paul J. and Glaser, Robert , Date-Added =. Categorization and representation of physics problems by experts and novices , Volume =. Cognitive science , Keywords =. 1981 , Bdsk-File-1 =

1981
[11]

Cognitive Psychology and Educational Assessment , Year =

Robert Mislevy , Booktitle =. Cognitive Psychology and Educational Assessment , Year =
[12]

Mislevy and John T

Robert J. Mislevy and John T. Behrens and Kristen E. Design and Discovery in Educational Assessment: Evidence-Centered Design, Psychometrics, and Educational Data Mining , Volume =. Journal of Educational Data Mining , Number =. 2012 , Bdsk-File-1 =

2012
[13]

Intelligent Tutoring Systems for Continuous, Embedded Assessment , Year =

Kurt VanLehn , Booktitle =. Intelligent Tutoring Systems for Continuous, Embedded Assessment , Year =
[14]

Fienberg and John Lafferty , Date-Added =

Elena Erosheva and Stephen E. Fienberg and John Lafferty , Date-Added =. Mixed-Membership Models of Scientific Publications , Volume =. PNAS , Number =. 2004 , Bdsk-Url-1 =

2004
[15]

Topic Models , Year =

David Blei and John Lafferty , Booktitle =. Topic Models , Year =
[16]

and Mclaren, B

Aleven, V. and Mclaren, B. and Roll, I. and Koedinger, K. , Date-Added =. Toward meta-cognitive tutoring: A model of help seeking with a Cognitive Tutor , Volume =. International Journal of Artificial Intelligence in Education , Pages =
[17]

and Koedinger, Kenneth R

Goldin, Ilya M. and Koedinger, Kenneth R. and Aleven, Vincent A. W. M. M. , Booktitle =. Hints: You Can't Have Just One , Year =
[18]

and Koedinger, Kenneth R

Goldin, Ilya M. and Koedinger, Kenneth R. and Aleven, Vincent A. W. M. M. , Booktitle =. Learner Differences in Hint Processing , Year =
[19]

and Carlson, Ryan , Booktitle =

Goldin, Ilya M. and Carlson, Ryan , Booktitle =. Learner Differences and Hint Content , Year =
[20]

Sequential Activity Profiling: Latent Dirichlet Allocation of Markov Chains , Volume =

Mark Girolami and Ata Kaban , Date-Added =. Sequential Activity Profiling: Latent Dirichlet Allocation of Markov Chains , Volume =. Data Mining and Knowledge Discovery , Keywords =. 2005 , Bdsk-File-1 =

2005
[21]

Journal of Computational and Graphical Statistics , volume=

Vecchia-approximated deep Gaussian processes for computer experiments , author=. Journal of Computational and Graphical Statistics , volume=. 2023 , publisher=

2023
[22]

Earth and Space Science , volume=

A multifidelity framework and uncertainty quantification for sea surface temperature in the massachusetts and cape cod bays , author=. Earth and Space Science , volume=. 2020 , publisher=

2020
[23]

Journal of the Royal Statistical Society Series C: Applied Statistics , pages=

Warped multifidelity Gaussian processes for data fusion of skewed environmental data , author=. Journal of the Royal Statistical Society Series C: Applied Statistics , pages=. 2025 , publisher=

2025
[24]

Environmental and Ecological Statistics , volume=

Spatiotemporal modelling of PM 2.5 concentrations in Lombardy (Italy): a comparative study , author=. Environmental and Ecological Statistics , volume=. 2024 , publisher=

2024
[25]

Dati e Indicatori - ARPA Lombardia , year =
[26]

M. C. Kennedy and A. O'Hagan , title =. Biometrika , volume =. 2000 , month =

2000
[27]

Artificial intelligence and statistics , pages=

Deep gaussian processes , author=. Artificial intelligence and statistics , pages=. 2013 , organization=

2013
[28]

Statistical Science , volume=

A general framework for Vecchia approximations of Gaussian processes , author=. Statistical Science , volume=. 2021 , publisher=

2021
[29]

ERA5: Fifth Generation of ECMWF Atmospheric Reanalyses of the Global Climate , year =
[30]

Earth , volume=

Air quality in Lombardy, Italy: an overview of the environmental monitoring system of ARPA Lombardia , author=. Earth , volume=. 2022 , publisher=

2022
[31]

Journal of Aircraft , volume=

Efficient multipoint aerodynamic design optimization via cokriging , author=. Journal of Aircraft , volume=
[32]

SIAM/ASA Journal on Uncertainty Quantification , volume=

Multifidelity Monte Carlo estimation with adaptive low-fidelity models , author=. SIAM/ASA Journal on Uncertainty Quantification , volume=. 2019 , publisher=

2019
[33]

Structural and Multidisciplinary Optimization , volume=

A hierarchical kriging approach for multi-fidelity optimization of automotive crashworthiness problems , author=. Structural and Multidisciplinary Optimization , volume=. 2022 , publisher=

2022
[34]

Journal of Agricultural, Biological and Environmental Statistics , volume=

Multi-scale Vecchia approximations of Gaussian processes , author=. Journal of Agricultural, Biological and Environmental Statistics , volume=. 2022 , publisher=

2022
[35]

Journal of the American Statistical Association , pages=

Iterative methods for vecchia-laplace approximations for latent gaussian process models , author=. Journal of the American Statistical Association , pages=. 2024 , publisher=

2024
[36]

Spatial Statistics , volume=

Hierarchical Bayesian nearest neighbor co-kriging Gaussian process models; an application to intersatellite calibration , author=. Spatial Statistics , volume=. 2021 , publisher=

2021
[37]

arXiv preprint arXiv:2501.11448 , year=

An accuracy-runtime trade-off comparison of scalable Gaussian process approximations for spatial data , author=. arXiv preprint arXiv:2501.11448 , year=

work page arXiv
[38]

Environmetrics , volume=

Recursive nearest neighbor co-kriging models for big multi-fidelity spatial data sets , author=. Environmetrics , volume=. 2024 , publisher=

2024
[39]

International Journal for Uncertainty Quantification , volume=

Recursive co-kriging model for design of computer experiments with multiple levels of fidelity , author=. International Journal for Uncertainty Quantification , volume=. 2014 , publisher=

2014
[40]

Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

Estimation and model identification for continuous spatial processes , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 1988 , publisher=

1988
[41]

Journal of the American Statistical Association , volume=

Hierarchical nearest-neighbor Gaussian process models for large geostatistical datasets , author=. Journal of the American Statistical Association , volume=. 2016 , publisher=

2016
[42]

Advances in Neural Information Processing Systems , volume=

Sparse gaussian process hyperparameters: Optimize or integrate? , author=. Advances in Neural Information Processing Systems , volume=
[43]

Spatial Statistics , volume=

Spatially clustered regression , author=. Spatial Statistics , volume=. 2021 , publisher=

2021
[44]

Environmental and Ecological Statistics , volume=

ARPALData: an R package for retrieving and analyzing air quality and weather data from ARPA Lombardia (Italy) , author=. Environmental and Ecological Statistics , volume=. 2024 , publisher=

2024
[45]

Advances in Computational Science and Engineering , volume =

Review of multi-fidelity models , author =. Advances in Computational Science and Engineering , volume =. 2023 , doi =

2023
[46]

Aerospace Science and Technology , volume=

Overview of Gaussian process based multi-fidelity techniques with variable relationship between fidelities, application to aerospace systems , author=. Aerospace Science and Technology , volume=. 2020 , publisher=

2020
[47]

Earth , volume =

Maranzano, Paolo , title =. Earth , volume =. 2022 , number =

2022