pith. sign in

arxiv: 2409.02331 · v4 · submitted 2024-09-03 · 📊 stat.ME

A parameterization of anisotropic Gaussian fields with penalized complexity priors

Liam Llamazares-Elias , Jonas Latz , Finn Lindgren This is my paper

Pith reviewed 2026-05-23 20:41 UTC · model grok-4.3

classification 📊 stat.ME
keywords anisotropic Gaussian fieldspenalized complexity priorsSPDEcorrelation lengthdiffusion matrixparameterizationBayesian spatial modelingGaussian random fields
0
0 comments X

The pith

A smooth invertible parameterization enables penalized complexity priors for anisotropic Gaussian fields that push correlation range to infinity and anisotropy to zero.

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

The paper develops a parameterization of the correlation length and diffusion matrix for anisotropic Gaussian fields represented as solutions to SPDEs. It then constructs penalized complexity priors on these parameters under the assumption they are constant in space. The resulting priors are weakly informative and penalize complexity specifically by favoring larger correlation ranges and reduced anisotropy. This matters because likelihoods alone give limited information about covariance structures even in in-fill asymptotics, requiring priors to ensure meaningful posterior covariance. A sympathetic reader would see this as a way to control model complexity in Bayesian spatial inference.

Core claim

The central claim is that a smooth, invertible parameterization of the correlation length and diffusion matrix of an anisotropic Gaussian field allows construction of penalized complexity priors when the parameters are constant in space. These priors are weakly informative and penalize complexity by pushing the correlation range toward infinity and the anisotropy to zero.

What carries the argument

The smooth invertible parameterization of correlation length and diffusion matrix, used to construct penalized complexity priors.

If this is right

  • Bayesian inference for anisotropic Gaussian fields gains a principled weakly informative prior on covariance structure.
  • The parameterization supports flexible yet controlled modeling of anisotropy in SPDE-based fields.
  • Posteriors are steered away from overly complex short-range or strongly anisotropic fields.
  • This construction applies directly when parameters do not vary spatially.

Where Pith is reading between the lines

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

  • The parameterization could be tested for numerical stability in high-dimensional settings with the same prior construction.
  • Extensions might explore how the penalization behaves when combined with other likelihood approximations.
  • Similar prior ideas could apply to related random field models beyond the SPDE representation.

Load-bearing premise

The correlation length and diffusion matrix parameters are constant in space.

What would settle it

A simulation or dataset where the posterior under this prior fails to push correlation range toward infinity and anisotropy toward zero when the likelihood provides no information on those parameters.

Figures

Figures reproduced from arXiv: 2409.02331 by Finn Lindgren, Jonas Latz, Liam Llamazares-Elias.

Figure 1
Figure 1. Figure 1: shows visually how the anisotropy increases with ∥v∥ and is directed towards ve. It can also be seen how the parameterization is injective (no two ellipses are the same) and smooth (the ellipses vary smoothly with v). -1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1 x1 x2 [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (b), we show simulations of the field where again we leave κ = 1 constant and rotate v by 90◦ in each plot. The figure shows that the field diffuses the most in the direction of ve and the least in the direction of ve⊥. The realizations of u are obtained using a FEM to solve the SPDE as detailed in Section 5. −4 0 4 −4 0 4 −4 0 4 −4 0 4 x1 x 2 0.25 0.50 0.75 1.00 K (a) Covariance K(x, 0) as v varies −4 0 4… view at source ↗
Figure 3
Figure 3. Figure 3: We fix κ = σu = 1 and plot the covariance K(·,(2, 2)) and realizations of solution u to (3) for ve(x) = (−x2, x1) on the left and v(x) = (x1, x2) on the right of each subfigure. In summary, we have parameterized the anisotropic field u using parameters (κ, v1, v2, σu), where u solves (3) and H := Hv is given by (10). The parameterization is identifiable and smooth, and the parameters have an intrinsic geom… view at source ↗
Figure 4
Figure 4. Figure 4: Marginal PC prior density of κ and v obtained in Theorem 2 for λθ = λv = 0.1. The hyperparameter λθ determines the flexibility of the model (how much we penalize large values of d(κ, v)), whereas λv controls the degree of anisotropy (how much we penalize large values of ∥v∥). Their values can be set to agree with desired quantiles using the following two results. Theorem 3. The prior for r := ∥v∥ satisfies… view at source ↗
Figure 5
Figure 5. Figure 5: Empirical cumulative distribution function (eCDF) of the absolute distances [PITH_FULL_IMAGE:figures/full_fig_p022_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Empirical cumulative distribution function (eCDF) of the length of symmet [PITH_FULL_IMAGE:figures/full_fig_p023_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Empirical CDF of the complexity d(κ, b vb) defined in (17). In the plots, arranged from left to right , θ true is simulated from the four distributions—πPC, πEG, πU, and πβ. Then, y = Au+ε is observed with true parameter value θ true. Finally, for each prior (red, green, teal, purple), Eπθ|y [d(κ, v)] is computed using smoothed importance sampling. 6 An application to rainfall data 6.1 Framework In this se… view at source ↗
Figure 8
Figure 8. Figure 8: The observed precipitation y, the elevation h, and the mesh of the domain for the precipitation data set. We will only study the stationary anisotropic setting where u|θ is a solution to (3) with spatially constant parameters. By incorporating β := (β0, β1) into u, our linear model (25) fits into the framework of Section 5, where now Aβ := (1m,h, A), uβ := (β,u) take the place of A,u in (21). We will consi… view at source ↗
Figure 9
Figure 9. Figure 9: Marginal densities on ρ and r = ∥v∥ of PC and EG priors The decay of the marginal PC prior on κ and of the EG prior are both exponential. The 26 [PITH_FULL_IMAGE:figures/full_fig_p026_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Predictive mean, latent field and covariance function of the anisotropic field [PITH_FULL_IMAGE:figures/full_fig_p028_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Scores for variable number of observations [PITH_FULL_IMAGE:figures/full_fig_p030_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Renormalized marginal prior and posterior densities of log( [PITH_FULL_IMAGE:figures/full_fig_p053_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Marginal prior and posterior densities of [PITH_FULL_IMAGE:figures/full_fig_p054_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Empirical cumulative distribution function (eCDF) of the absolute distances [PITH_FULL_IMAGE:figures/full_fig_p059_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Empirical cumulative distribution function (eCDF) of the length of symmetric [PITH_FULL_IMAGE:figures/full_fig_p060_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Frequency of true parameter being in the 0 [PITH_FULL_IMAGE:figures/full_fig_p061_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: eCDF of probabilities ai = Pest[θi ≤ θ true i |y] where θ true and y come from the true model. In [PITH_FULL_IMAGE:figures/full_fig_p062_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Plot of the empirical difference against [PITH_FULL_IMAGE:figures/full_fig_p064_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Kullback-Leibler divergence between the posterior and the Gaussian approxi [PITH_FULL_IMAGE:figures/full_fig_p065_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: CDF of the posterior complexity d(κ, v) for PC and not exponential-Gaussian priors 65 [PITH_FULL_IMAGE:figures/full_fig_p065_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Difference of scores between the isotropic and anisotropic PC models [PITH_FULL_IMAGE:figures/full_fig_p069_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Difference of scores between the isotropic PC and anisotropic EG models [PITH_FULL_IMAGE:figures/full_fig_p069_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Precipitation data simulated using model ( [PITH_FULL_IMAGE:figures/full_fig_p070_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Comparison of the scores of the model for each prior for anisotropic and isotropic [PITH_FULL_IMAGE:figures/full_fig_p071_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: From top to bottom, anisotropic and isotropic data is observed. From left to [PITH_FULL_IMAGE:figures/full_fig_p072_25.png] view at source ↗
read the original abstract

Gaussian random fields (GFs) are fundamental tools in spatial modeling and can be represented flexibly and efficiently as solutions to stochastic partial differential equations (SPDEs). The SPDEs depend on specific parameters, which enforce various field behaviors and can be estimated using Bayesian inference. However, even under in-fill asymptotics, the likelihood only provides limited insights into the covariance structure. In response, it is essential to leverage priors to achieve appropriate, meaningful covariance structures in the posterior. This study introduces a smooth, invertible parameterization of the correlation length and diffusion matrix of an anisotropic GF and constructs penalized complexity (PC) priors for the model when the parameters are constant in space. The formulated prior is weakly informative, effectively penalizing complexity by pushing the correlation range toward infinity and the anisotropy to zero.

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

0 major / 1 minor

Summary. The manuscript introduces a smooth, invertible parameterization of the correlation length and diffusion matrix for anisotropic Gaussian random fields (GFs) represented via SPDEs. It constructs penalized complexity (PC) priors for these parameters under the restriction that they are spatially constant. The resulting prior is weakly informative and penalizes toward infinite correlation range and zero anisotropy.

Significance. If the parameterization is smooth and invertible and the PC priors are correctly constructed from a valid base model, the work supplies a practical tool for Bayesian spatial modeling with anisotropic fields. The explicit use of the standard PC-prior distance construction (pushing to the base model of infinite range and isotropy) is a strength when the likelihood supplies limited covariance information under in-fill asymptotics.

minor comments (1)
  1. The abstract states that the parameterization is 'smooth, invertible' and that the prior 'effectively penalizes complexity,' but without the explicit mapping or distance function in the provided description it is not possible to verify that positive-definiteness of the diffusion matrix is preserved for all parameter values.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their summary of the manuscript and for recognizing the practical value of the smooth invertible parameterization together with the PC-prior construction that penalizes toward infinite range and isotropy. We note that the referee report lists no specific major comments.

Circularity Check

0 steps flagged

No significant circularity; construction is self-contained

full rationale

The paper introduces a smooth invertible reparameterization of correlation length and diffusion matrix for anisotropic Gaussian fields (under the explicit restriction to spatially constant parameters) and then applies the standard penalized complexity prior construction to the resulting parameters. No equation reduces a claimed prediction or uniqueness result to a fitted input by construction, no load-bearing step depends on a self-citation chain whose base result is unverified, and the PC-prior distance penalty toward infinite range and zero anisotropy follows directly from the given base model once the one-to-one mapping is supplied. The derivation therefore consists of an explicit modeling choice rather than an internal reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The contribution rests on the standard SPDE representation of Gaussian fields and the modeling choice that parameters are spatially constant.

axioms (2)
  • domain assumption Gaussian random fields can be represented flexibly as solutions to stochastic partial differential equations
    Stated in the abstract as the foundational representation
  • domain assumption Parameters are constant in space
    Explicitly required for the PC prior construction in the abstract

pith-pipeline@v0.9.0 · 5659 in / 1186 out tokens · 20474 ms · 2026-05-23T20:41:26.718422+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages

  1. [1]

    J., Taylor, J

    Adler, R. J., Taylor, J. E., et al. (2007). Random fields and geometry , volume 80. Springer

    work page 2007

  2. [2]

    P., and Gelfand, A

    Banerjee, S., Carlin, B. P., and Gelfand, A. E. (2003). Hierarchical modeling and analysis for spatial data . Chapman and Hall/CRC

    work page 2003

  3. [3]

    Bhatt, S., Weiss, D., Cameron, E., Bisanzio, D., Mappin, B., Dalrymple, U., Battle, K., Moyes, C., Henry, A., Eckhoff, P., et al. (2015). The effect of malaria control on Plasmodium falciparum in Africa between 2000 and 2015. Nature , 526(7572):207--211

    work page 2015

  4. [4]

    B., and Xiong, Z

    Bolin, D., Simas, A. B., and Xiong, Z. (2023). Wasserstein complexity penalization priors: a new class of penalizing complexity priors. arXiv preprint arXiv:2312.04481

    work page arXiv 2023

  5. [5]

    and Wikle, C

    Cressie, N. and Wikle, C. K. (2015). Statistics for spatio-temporal data . John Wiley & Sons

    work page 2015

  6. [6]

    Dawid, A. P. and Sebastiani, P. (1999). Coherent dispersion criteria for optimal experimental design. Annals of Statistics , pages 65--81

    work page 1999

  7. [7]

    M., Girolami, M

    Dunlop, M. M., Girolami, M. A., Stuart, A. M., and Teckentrup, A. L. (2018). How deep are deep Gaussian processes? Journal of Machine Learning Research , 19(54):1--46

    work page 2018

  8. [8]

    M., Iglesias, M

    Dunlop, M. M., Iglesias, M. A., and Stuart, A. M. (2017). Hierarchical bayesian level set inversion. Statistics and Computing , 27:1555--1584

    work page 2017

  9. [9]

    Evans, L. C. (2010). Partial differential equations , volume 19. American Mathematical Soc

    work page 2010

  10. [10]

    Fuglstad, G.-A., Lindgren, F., Simpson, D., and Rue, H. (2015). Exploring a new class of non-stationary spatial Gaussian random fields with varying local anisotropy. Statistica Sinica , pages 115--133

    work page 2015

  11. [11]

    Fuglstad, G.-A., Simpson, D., Lindgren, F., and Rue, H. (2019). Constructing priors that penalize the complexity of Gaussian random fields. Journal of the American Statistical Association , 114(525):445--452

    work page 2019

  12. [12]

    Gelbrich, M. (1990). On a formula for the l2 wasserstein metric between measures on euclidean and hilbert spaces. Mathematische Nachrichten , 147(1):185--203

    work page 1990

  13. [13]

    Gel'fand, I. M. and Vilenkin, N. Y. (2014). Generalized functions: Applications of harmonic analysis , volume 4. Academic press

    work page 2014

  14. [14]

    and Raftery, A

    Gneiting, T. and Raftery, A. E. (2007). Strictly proper scoring rules, prediction, and estimation. Journal of the American statistical Association , 102(477):359--378

    work page 2007

  15. [15]

    P., Gneiting, T., Berrocal, V

    Grimit, E. P., Gneiting, T., Berrocal, V. J., and Johnson, N. A. (2006). The continuous ranked probability score for circular variables and its application to mesoscale forecast ensemble verification. Quarterly Journal of the Royal Meteorological Society: A journal of the atmospheric sciences, applied meteorology and physical oceanography , 132(621C):2925--2942

    work page 2006

  16. [16]

    Ingebrigtsen, R., Lindgren, F., Steinsland, I., and Martino, S. (2015). Estimation of a non-stationary model for annual precipitation in southern Norway using replicates of the spatial field. Spatial Statistics , 14:338--364

    work page 2015

  17. [17]

    Jaynes, E. T. (2003). Probability theory: The logic of science . Cambridge university press

    work page 2003

  18. [18]

    Lindgren, F., Bakka, H., Bolin, D., Krainski, E., and Rue, H. (2020). A diffusion-based spatio-temporal extension of Gaussian Mat \'e rn fields. arXiv preprint arXiv:2006.04917

    work page arXiv 2020

  19. [19]

    Lindgren, F., Bolin, D., and Rue, H. (2022). The SPDE approach for Gaussian and non- Gaussian fields: 10 years and still running. Spatial Statistics , page 100599

    work page 2022

  20. [20]

    Lindgren, F., Rue, H., and Lindstr \"o m, J. (2011). An explicit link between Gaussian fields and Gaussian Markov random fields: the stochastic partial differential equation approach. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , 73

    work page 2011

  21. [21]

    Lindgren, G. (2012). Stationary stochastic processes: Theory and applications . CRC Press

    work page 2012

  22. [22]

    H., Kim, S., B \"u rkner, P., Huurre, N., Faltejskov \'a , K., Gelman, A., and Vehtari, A

    Modr \'a k, M., Moon, A. H., Kim, S., B \"u rkner, P., Huurre, N., Faltejskov \'a , K., Gelman, A., and Vehtari, A. (2023). Simulation-based calibration checking for bayesian computation: The choice of test quantities shapes sensitivity. Bayesian Analysis , 1(1):1--28

    work page 2023

  23. [23]

    Rasmussen, C. E. (2003). Gaussian processes in machine learning. In Summer school on machine learning , pages 63--71. Springer

    work page 2003

  24. [24]

    Roques, L., Allard, D., and Soubeyrand, S. (2022). Spatial statistics and stochastic partial differential equations: A mechanistic viewpoint. Spatial Statistics , 50:100591

    work page 2022

  25. [25]

    and Held, L

    Rue, H. and Held, L. (2005). Gaussian Markov random fields: theory and applications . Chapman and Hall/CRC

    work page 2005

  26. [26]

    Sampson, P. D. and Guttorp, P. (1992). Nonparametric estimation of nonstationary spatial covariance structure. Journal of the American Statistical Association , 87(417):108--119

    work page 1992

  27. [27]

    Simpson, D., Lindgren, F., and Rue, H. (2012). Think continuous: Markov ian Gaussian models in spatial statistics. Spatial Statistics , 1:16--29

    work page 2012

  28. [28]

    P., Rue, H., Martins, T

    Simpson, D. P., Rue, H., Martins, T. G., Riebler, A., and S rbye, S. H. (2014). Penalising model component complexity: A principled, practical approach to constructing priors. Statistical Science , 32:1--28

    work page 2014

  29. [29]

    Stein, M. L. (2004). Equivalence of gaussian measures for some nonstationary random fields. Journal of Statistical Planning and Inference , 123(1):1--11

    work page 2004

  30. [30]

    Talts, S., Betancourt, M., Simpson, D., Vehtari, A., and Gelman, A. (2018). Validating bayesian inference algorithms with simulation-based calibration. arXiv preprint arXiv:1804.06788

    work page arXiv 2018

  31. [31]

    Uribe, F., Papaioannou, I., Latz, J., Betz, W., Ullmann, E., and Straub, D. (2021). Bayesian inference with subset simulation in varying dimensions applied to the karhunen–loève expansion. International Journal for Numerical Methods in Engineering , 122(18):5100--5127

    work page 2021

  32. [32]

    Vehtari, A., Simpson, D., Gelman, A., Yao, Y., and Gabry, J. (2015). Pareto smoothed importance sampling. arXiv preprint arXiv:1507.02646

    work page arXiv 2015

  33. [33]

    and Zuo, R

    Wang, J. and Zuo, R. (2021). Spatial modelling of hydrothermal mineralization-related geochemical patterns using INLA + SPDE and local singularity analysis. Computers & Geosciences , 154:104822

    work page 2021

  34. [34]

    Zhang, H. (2004). Inconsistent estimation and asymptotically equal interpolations in model-based geostatistics. Journal of the American Statistical Association , 99(465):250--261

    work page 2004