Pith. sign in

REVIEW 2 major objections 6 minor 43 references

When is multivariate kriging worthwhile? A design-geometry analysis of heterotopic multi-output Gaussian processes

T0 review · 2 major / 6 minor · reviewed 2026-07-10 · grok-4.5

Pith's one-line read For separable multi-output Gaussian processes, joint kriging beats separate models only when the output-specific designs are interleaved enough to open residual borrowing and make cross-dependence learnable.

desk verdict Solid geometric theory that actually explains when multi-output kriging helps under heterotopy; scoped cleanly to separable models, with usable pre-fit diagnostics. read the letter →

arxiv 2607.06832 v1 pith:3JGJFK4Z submitted 2026-07-07 stat.ME stat.CO

classification stat.MEstat.CO MSC 62M3062M2062K05
keywords multivariatestatisticskrigingmulti-outputGaussianprocessdesignofexperimentsheterotopicsamplingborrowingpotentialinteractionmassnetbenefit
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

When several related outputs are observed at different input locations (heterotopic sampling), analysts have long disagreed on whether a joint multivariate kriging model is worth the extra fitting cost. Careful simulations on shared designs often find no benefit, while multi-fidelity and geostatistics practice assume auxiliary outputs help. This paper shows that, inside the widely used separable multi-output Gaussian-process model, the answer is controlled by the geometry of those designs. It supplies model-free diagnostics—directed coverage, directed proximity and borrowing-potential indices—that can be computed before any model is fitted, proves an exact identity for the oracle prediction gain from joint modelling, and shows that the Fisher information for cross-output dependence scales with a kernel-weighted interaction mass between the designs. Zero exact overlap is therefore not a single regime: interleaved designs open both the borrowing channel and the estimation channel, while separated designs close both. The results are packaged into a first-order net-benefit screen and illustrated on synthetic designs, an M/M/1 queue and a multi-pollutant monitoring network.

What carries the argument

The residual auxiliary channel (Theorem 4.1) together with the kernel-weighted cross-design interaction mass W_pq (Theorem 5.1). The former isolates how much of an auxiliary output’s signal survives after conditioning on the target data; the latter measures how strongly the same geometry identifies the dependence parameters that borrowing requires.

What would settle it

Construct two zero-overlap designs of equal size—one interleaved at the kernel length-scale, one geometrically separated—fit both the joint and independent models under a correctly specified separable multi-output GP, and check whether the interleaved design alone produces positive oracle gain, positive interaction mass and positive net benefit while the separated design produces essentially none.

Watch

Extended reading notes

Core claim

For separable multi-output Gaussian processes the decision to model jointly rather than output-by-output is governed by output-specific design geometry. Oracle prediction gain equals a residual auxiliary quadratic form, cross-dependence information scales with kernel-weighted interaction mass, and two zero-overlap designs can be statistically opposite according to whether they interleave or separate.

Load-bearing premise

The clean geometric reconciliation and the main prediction bounds assume the covariance is separable (a single kernel shared across outputs, scaled by a coregionalisation matrix) and radial; if that structure fails, geometry alone need not decide when joint modelling pays.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. The paper studies when joint multivariate kriging (separable multi-output GPs) improves on independent univariate kriging under heterotopic designs. It argues that the answer is controlled by output-specific design geometry rather than exact co-location alone. The authors introduce model-free diagnostics (directed coverage, directed/normalised proximity, borrowing potential indices), prove an exact residual-channel identity for oracle prediction gain (Theorem 4.1) with local geometric bounds under radial kernels, show that Fisher information for cross-output dependence scales with a kernel-weighted interaction mass W_pq (Theorems 5.1–5.3), and combine oracle gain with a first-order estimation penalty into a net-benefit screen. Controlled synthetic experiments, an M/M/1 illustration, and an EPA AQS multi-pollutant case study support the geometry-based guidance and reconcile prior mixed empirical findings with classical autokrigeability.

Significance. If the results hold as stated, the paper supplies a useful and overdue geometric account of when multi-output kriging is worth fitting in simulation metamodelling, multi-fidelity work, and monitoring networks. The reconciliation of Kleijnen–Mehdad-type isotopic findings with the multi-fidelity/geostatistics premise is concrete and scoped correctly to separable models. Strengths include exact residual-channel and information identities with proofs, pre-fit model-free diagnostics, an explicit zero-overlap dichotomy (interleaved vs separated), a component-wise LMC extension, and experiments that isolate geometry rather than only reporting aggregate error. The operational screening procedure is practical and computationally light relative to a dense joint GP fit.

major comments (2)
  1. §6, Eq. (6.2)–(6.3) and Experiment 4 (Fig. 8.2, Table E.3): the operational net-benefit claim rests on a first-order plug-in comparison of oracle gain to excess estimation cost. In Experiment 4 the plug-in margin turns negative in the moderate/poor regimes while realised estimation cost remains slightly below oracle gain, so the plug-in rule is conservative. This is acknowledged, but the main-text guidance in §7 Step 5 still presents the plug-in check as the decision tool. Please state more explicitly in §6–§7 (and the conclusions) when the plug-in screen should be treated as a conservative triage rather than a calibrated accept/reject rule, and what a practitioner should do when the margin is near zero.
  2. §5.1–5.2 and Theorem 5.2: the clean efficient-information interpretation of W_pq is exact at independence (η=0). Away from independence the paper notes that I_ην need not vanish, yet the screening language in §7 Step 4 treats W_pq as the primary estimability diagnostic for general dependence. Please clarify in the main text how far the independence benchmark is intended to carry for nonzero Λ_pq (e.g., local detectability vs full joint estimation), so that the estimability claim is not over-read beyond the proved regime.
minor comments (6)
  1. §4.4, Theorem 4.2: the bound is described as qualitative rather than sharp; a short remark in the main text pointing to the tighter two-output residual-channel bounds (Theorem C.1 / Corollary C.1) would help readers who want quantitative envelopes.
  2. §8.1.2: Experiments 2–3 are largely deferred to Appendix E. A one-sentence pointer in the main text to the residual-channel plot (Fig. E.2) and the B_j vs gain ranking (Table E.2) would make the screening narrative self-contained.
  3. Table 8.2 / Experiment 5: the isotopic ΔRMSE and ΔMLPD are small but positive under noise; a brief cross-reference to Remark 4.1 in the table caption would prevent readers from reading this as a contradiction of autokrigeability.
  4. §8.3 and Appendix E.2: Colorado shows that large B_j need not order realised gains in small networks. Consider elevating one sentence of that caveat into the main-text case-study discussion, not only the appendix.
  5. Notation: Π is called the design proximity matrix of normalised directed proximities; a single display of the definition of ˜π_{p→q} next to Π would reduce back-referencing in §3.
  6. Typos/style: “The present paper is to show” (p. 2) → “This paper shows”; check consistent hyphenation of “multi-output” / “multioutput” and “zero-overlap” throughout.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: oracle gain, interaction-mass bounds, and zero-overlap dichotomy are derived from the separable multi-output GP model and design geometry, not fitted or self-defined to match the conclusion.

full rationale

The load-bearing chain is self-contained. Theorem 4.1 is the standard residual-variance identity from Gaussian conditioning of the joint (f_j(x⋆), y^(j), y^(-j)) under the separable model (4.1)–(4.4); Δ_j is defined as V_ind − V_joint and equals the residual quadratic form by construction of conditional Gaussians, not by fitting. Theorem 4.2 then bounds that residual channel by radial kernel monotonicity and local gaps δ_j, δ_−j—geometry enters as an assumption, not as a fitted target. Section 5 applies the classical centred-Gaussian Fisher formula to the cross-block derivative ˙Σ_u determined by K^(0)_pq, so I_uu scales with the kernel-weighted interaction mass W_pq (Theorem 5.1); the efficient-information result at independence (Theorem 5.2) and the zero-overlap dichotomy (Theorem 5.3) follow from the same spectral and coverage bounds without circular reference to the prediction claims. Diagnostics Π, B_j, W_pq are pre-fit geometric summaries (Section 3, 5.1). The net-benefit criterion (Section 6) is a first-order delta-method risk decomposition, explicitly labeled a screen rather than a tautological prediction. Self-citations [25, 26] only introduce the multi-output GP working model; autokrigeability and co-kriging equivalence are attributed to external classical sources [10, 15]. Experiments use known synthetic truth or public AQS data and do not fit parameters to force the geometric ranking. No step reduces a claimed prediction to its own fitted input or to an unverified self-citation chain.

Assumptions & free parameters 3 free parameters · 5 assumptions · 4 invented entities

Central claims rest on standard multi-output GP/co-kriging structure plus radial-kernel monotonicity for geometric bounds, not on fitted physical constants. Experimental free parameters (lengthscales, noise, Λ) set synthetic truth and working models but are not free knobs that force the geometric theorems. Invented objects are diagnostic indices and the interaction mass, which are definitions with external computability rather than unobserved physical mediators.

free parameters (3)
  • Kernel lengthscale ℓ (and plausible lengthscale range in screening)
    Bounds and W_pq depend on the correlation scale; Step 1 of the screen requires a plausible lengthscale. Theory is qualitative across scales, but numerical interaction mass and coverage certificates are scale-dependent.
  • Observation noise variances τ_j²
    Autokrigeability and residual-channel strength depend on noise; isotopic gain is confined to a noise-filtering channel. Experiments fix specific τ² values.
  • Cross-output dependence matrix Λ (or correlation ρ)
    Oracle gain is zero if off-diagonals vanish; experiments fix Λ entries (e.g. 0.70, 0.85). Theory treats Λ as known for oracle results and estimated for net benefit.
assumptions (5)
  • domain assumption Separable multi-output GP: Cov(f_p(x), f_q(x')) = Λ_pq k(x,x') with positive-definite Λ and scalar kernel k (Eq. 4.1–4.2).
    Working model for all main prediction and estimability theorems; nonseparable alternatives are explicitly out of scope for the clean geometry story.
  • domain assumption Radial non-increasing kernel k(x,x')=ψ(‖x−x′‖) for geometric bounds and coverage/proximity certificates (Eq. 4.6).
    Used in Theorems 4.2, 5.3 and Propositions 4.1–5.2 so that distance controls kernel support.
  • standard math Gaussian observations with independent noise; posterior mean is BLUP / co-kriging predictor (Proposition A.1).
    Standard GP conditioning; underpins prediction-gain identity Theorem 4.1.
  • standard math First-order delta-method approximation of estimation penalties via Fisher information (Theorem 6.1; Van der Vaart-style asymptotics).
    Net-benefit criterion is asymptotic; paper labels it a first-order screen, not a finite-sample guarantee.
  • standard math At independence η=0, cross-dependence parameters are information-orthogonal to nuisance covariance parameters (Theorem 5.2).
    Gives exact efficient-information interpretation of whitened interaction mass for local detectability of Λ_pq.
invented entities (4)
  • Directed coverage DC_{p→q}(r), directed/normalised proximity π, and design proximity matrix Π independent evidence
    purpose: Model-free summaries of heterotopic geometry beyond exact overlap.
    New evaluative diagnostics; computable from designs alone; not physical entities.
  • Local/global borrowing potential index b_j(x★), B_j independent evidence
    purpose: Screen outputs that have local gaps filled by auxiliary designs.
    Definitional index; experiments show it tracks average oracle gain but can misrank small networks.
  • Kernel-weighted cross-design interaction mass W_pq independent evidence
    purpose: Bridge from geometry to Fisher information for off-diagonal Λ_pq.
    Defined as ‖K_pq^{(0)}‖_F²; controls estimability bounds in Theorem 5.1.
  • First-order net benefit criterion (oracle gain vs excess estimation cost) independent evidence
    purpose: Operational decision rule for when joint modeling is worthwhile after estimation cost.
    Combines Theorems 4.1 and 6.1; plug-in version is a post-fit screen.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When is multivariate kriging worthwhile? A design-geometry analysis of heterotopic multi-output Gaussian processes." pith.science (2026). https://pith.science/paper/3JGJFK4Z

@misc{pith2026260706832,
  author       = {Pith},
  title        = {Pith review of: When is multivariate kriging worthwhile? A design-geometry analysis of heterotopic multi-output Gaussian processes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3JGJFK4Z}},
  note         = {Machine review of arXiv:2607.06832}
}
read the original abstract

Simulation experiments, multi-fidelity computer models and monitoring networks often produce several related outputs observed at different input locations, a sampling pattern known as heterotopic. Whether a joint multivariate kriging metamodel then predicts better than separate univariate metamodels has remained unresolved: careful simulation comparisons on common designs report little or no benefit from multivariate kriging, yet the multi-fidelity and geostatistical literatures are built on the premise that auxiliary outputs help. We show that, for separable multi-output Gaussian processes, the answer is governed by the geometry of the output-specific designs. We introduce model-free diagnostics that can be computed before fitting, namely directed coverage, directed proximity and borrowing potential indices. We derive an exact identity for the oracle prediction gain of joint modelling and bound this gain using local geometry under radial functions. We further prove that the estimability of cross-output dependence is controlled by a kernel-weighted cross-design interaction mass, and extend this result component by component to the linear model of coregionalisation. One consequence is that interleaved and separated designs are not statistically equivalent, even when both have zero overlap. We combine these results into a first-order net benefit criterion for deciding when joint modelling is worthwhile. Controlled synthetic experiments, an M/M/1 queueing illustration and a case study of a multi-pollutant monitoring network turn this criterion into practical guidance.

Figures

Figures reproduced from arXiv: 2607.06832 by the authors.

Figure 8.1
Figure 8.1. Zero-overlap designs for Experiment 1. Left: interleaved. Middle: moderate. Right: [PITH_FULL_IMAGE:figures/full_fig_p019_8_1.png] view at source ↗
Figure 8.2
Figure 8.2. Dependence recovery and net benefit trade-off in a three-output design for [PITH_FULL_IMAGE:figures/full_fig_p020_8_2.png] view at source ↗
Figure 8.3
Figure 8.3. Texas AQS case study. Left: monitoring geometry for PM [PITH_FULL_IMAGE:figures/full_fig_p022_8_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 43 canonical work pages

  1. [1]

    Jack P. C. Kleijnen. Kriging metamodeling in simulation: A review.European Journal of Operational Research, 192(3):707–716, 2009

  2. [2]

    Nelson, and Jeremy Staum

    Bruce Ankenman, Barry L. Nelson, and Jeremy Staum. Stochastic kriging for simulation metamodeling.Operations Research, 58(2):371–382, 2010

  3. [3]

    Jack P. C. Kleijnen. Regression and Kriging metamodels with their experimental designs in simulation: A review.European Journal of Operational Research, 256(1):1–16, 2017

  4. [4]

    Erickson, Bruce E

    Collin B. Erickson, Bruce E. Ankenman, and Susan M. Sanchez. Comparison of Gaussian process modeling software.European Journal of Operational Research, 266(1):179–192, 2018

  5. [5]

    Jack P. C. Kleijnen and Ehsan Mehdad. Multivariate versus univariate Kriging metamodels for multi-response simulation models.European Journal of Operational Research, 236(2): 573–582, 2014

  6. [6]

    Fricker, Jeremy E

    Thomas E. Fricker, Jeremy E. Oakley, and Nathan M. Urban. Multivariate Gaussian process emulators with nonseparable covariance structures.Technometrics, 55(1):47–56, 2013

  7. [7]

    Linear coregionalization model: tools for estimation and choice of cross-variogram matrix.Mathematical Geology, 24(3):269–286, 1992

    Michel Goulard and Marc Voltz. Linear coregionalization model: tools for estimation and choice of cross-variogram matrix.Mathematical Geology, 24(3):269–286, 1992

  8. [8]

    Multi-task Gaussian process prediction.Advances in Neural Information Processing Systems, 20, 2007

    Edwin V Bonilla, Kian Chai, and Christopher Williams. Multi-task Gaussian process prediction.Advances in Neural Information Processing Systems, 20, 2007

Show all 43 references
  1. [9]

    Kernels for vector-valued functions: A review.Foundations and Trends in Machine Learning, 4(3):195–266, 2012

    Mauricio A Alvarez, Lorenzo Rosasco, and Neil D Lawrence. Kernels for vector-valued functions: A review.Foundations and Trends in Machine Learning, 4(3):195–266, 2012

  2. [10]

    Springer Science & Business Media, 2013

    Hans Wackernagel.Multivariate Geostatistics: An Introduction with Applications. Springer Science & Business Media, 2013

  3. [11]

    Genton and William Kleiber

    Marc G. Genton and William Kleiber. Cross-covariance functions for multivariate geo- statistics.Statistical Science, 30(2):147–163, 2015. doi: 10.1214/14-STS487

  4. [12]

    Kennedy and Anthony O’Hagan

    Marc C. Kennedy and Anthony O’Hagan. Predicting the output from a complex computer code when fast approximations are available.Biometrika, 87(1):1–13, 2000

  5. [13]

    Alexander I. J. Forrester, András Sóbester, and Andy J. Keane. Multi-fidelity optimization via surrogate modelling.Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 463(2088):3251–3269, 2007

  6. [14]

    Recursive co-kriging model for design of com- puter experiments with multiple levels of fidelity.International Journal for Uncertainty Quantification, 4(5):365–386, 2014

    Loic Le Gratiet and Josselin Garnier. Recursive co-kriging model for design of com- puter experiments with multiple levels of fidelity.International Journal for Uncertainty Quantification, 4(5):365–386, 2014. 48

  7. [15]

    John Wiley & Sons, Hoboken, NJ, 2nd edition, 2012

    Jean-Paul Chilès and Pierre Delfiner.Geostatistics: Modeling Spatial Uncertainty. John Wiley & Sons, Hoboken, NJ, 2nd edition, 2012

  8. [16]

    Peter Z. G. Qian. Nested Latin hypercube designs.Biometrika, 96(4):957–970, 2009

  9. [17]

    A method for the updating of stochastic kriging metamodels.European Journal of Operational Research, 247(3):859–866, 2015

    Bogumił Kamiński. A method for the updating of stochastic kriging metamodels.European Journal of Operational Research, 247(3):859–866, 2015

  10. [18]

    Efficient space-filling and non- collapsing sequential design strategies for simulation-based modeling.European Journal of Operational Research, 214(3):683–696, 2011

    Karel Crombecq, Eric Laermans, and Tom Dhaene. Efficient space-filling and non- collapsing sequential design strategies for simulation-based modeling.European Journal of Operational Research, 214(3):683–696, 2011

  11. [19]

    Wim C. M. van Beers and Jack P. C. Kleijnen. Customized sequential designs for random simulation experiments: Kriging metamodeling and bootstrapping.European Journal of Operational Research, 186(3):1099–1113, 2008

  12. [20]

    Xi Chen and Qiang Zhou. Sequential design strategies for mean response surface meta- modeling via stochastic kriging with adaptive exploration and exploitation.European Journal of Operational Research, 262(2):575–585, 2017

  13. [21]

    Comparison of Kriging- based algorithms for simulation optimization with heterogeneous noise.European Journal of Operational Research, 261(1):279–301, 2017

    Hamed Jalali, Inneke Van Nieuwenhuyse, and Victor Picheny. Comparison of Kriging- based algorithms for simulation optimization with heterogeneous noise.European Journal of Operational Research, 261(1):279–301, 2017

  14. [22]

    Computationally efficient convolved multiple output Gaussian processes.The Journal of Machine Learning Research, 12:1459–1500, 2011

    Mauricio A Alvarez and Neil D Lawrence. Computationally efficient convolved multiple output Gaussian processes.The Journal of Machine Learning Research, 12:1459–1500, 2011

  15. [23]

    Generic inference in latent Gaussian process models.Journal of Machine Learning Research, 20:117:1–117:63, 2019

    Edwin V Bonilla, Karl Krauth, and Amir Dezfouli. Generic inference in latent Gaussian process models.Journal of Machine Learning Research, 20:117:1–117:63, 2019

  16. [24]

    Remarks on multi-output Gaussian process regression.Knowledge-Based Systems, 144:102–121, 2018

    Haitao Liu, Jianfei Cai, and Yew-Soon Ong. Remarks on multi-output Gaussian process regression.Knowledge-Based Systems, 144:102–121, 2018

  17. [25]

    Multivariate Gaussian and Student-t process regression for multi-output prediction.Neural Computing and Applications, 32(8): 3005–3028, 2020

    Zexun Chen, Bo Wang, and Alexander N Gorban. Multivariate Gaussian and Student-t process regression for multi-output prediction.Neural Computing and Applications, 32(8): 3005–3028, 2020

  18. [26]

    Multivariate Gaussian processes: definitions, examples and applications.Metron, 81(2):181–191, 2023

    Zexun Chen, Jun Fan, and Kuo Wang. Multivariate Gaussian processes: definitions, examples and applications.Metron, 81(2):181–191, 2023

  19. [27]

    Nonstationary multivariate process modeling through spatially varying coregionalization.TEST, 13(2): 263–312, 2004

    Alan E Gelfand, Alexandra M Schmidt, Sudipto Banerjee, and C F Sirmans. Nonstationary multivariate process modeling through spatially varying coregionalization.TEST, 13(2): 263–312, 2004. doi: 10.1007/BF02595775

  20. [28]

    Multivariate spatial modeling for geosta- tistical data using convolved covariance functions.Mathematical Geology, 39(2):225–245,

    Anandamayee Majumdar and Alan E Gelfand. Multivariate spatial modeling for geosta- tistical data using convolved covariance functions.Mathematical Geology, 39(2):225–245,

  21. [29]

    doi: 10.1007/s11004-006-9072-6. 49

  22. [30]

    Matérn cross-covariance functions for multivariate random fields.Journal of the American Statistical Association, 105(491):1167–1177, 2010

    Tilmann Gneiting, William Kleiber, and Martin Schlather. Matérn cross-covariance functions for multivariate random fields.Journal of the American Statistical Association, 105(491):1167–1177, 2010. doi: 10.1198/jasa.2010.tm09420

  23. [31]

    Apanasovich, Marc G

    Tatiyana V. Apanasovich, Marc G. Genton, and Ying Sun. A valid matérn class of cross-covariance functions for multivariate random fields with any number of components. Journal of the American Statistical Association, 107(497):180–193, 2012. doi: 10.1080/01 621459.2011.643197

  24. [32]

    Inside-out cross-covariance for spatial multivariate data.Journal of the American Statistical Association, pages 1–22, 2026

    Michele Peruzzi. Inside-out cross-covariance for spatial multivariate data.Journal of the American Statistical Association, pages 1–22, 2026. doi: 10.1080/01621459.2026.2640644

  25. [33]

    Matrix formulation of co-kriging.Journal of the International Association for Mathematical Geology, 14(3):249–257, 1982

    Donald E Myers. Matrix formulation of co-kriging.Journal of the International Association for Mathematical Geology, 14(3):249–257, 1982

  26. [34]

    Ver Hoef and Noel Cressie

    Jay M. Ver Hoef and Noel Cressie. Multivariable spatial prediction.Mathematical Geology, 25(2):219–240, 1993. doi: 10.1007/BF00893273

  27. [35]

    John Wiley & Sons, 2015

    Noel Cressie.Statistics for spatial data. John Wiley & Sons, 2015

  28. [36]

    D. R. Cox and Nancy Reid. Parameter orthogonality and approximate conditional inference.Journal of the Royal Statistical Society. Series B (Methodological), 49(1):1–18, 1987

  29. [37]

    A. W. Van der Vaart.Asymptotic Statistics. Cambridge University Press, Cambridge, 1998

  30. [38]

    Spatial sampling design for parameter estimation of the covariance function.Journal of statistical planning and inference, 134(2):583–603, 2005

    Zhengyuan Zhu and Michael L Stein. Spatial sampling design for parameter estimation of the covariance function.Journal of statistical planning and inference, 134(2):583–603, 2005

  31. [39]

    Spatial sampling design for prediction with estimated parameters.Journal of agricultural, biological, and environmental statistics, 11(1):24–44, 2006

    Zhengyuan Zhu and Michael L Stein. Spatial sampling design for prediction with estimated parameters.Journal of agricultural, biological, and environmental statistics, 11(1):24–44, 2006

  32. [40]

    Springer, 2007

    Werner G Müller.Collecting spatial data: optimum design of experiments for random fields. Springer, 2007

  33. [41]

    Springer, New York, 1999

    Michael L Stein.Interpolation of Spatial Data: Some Theory for Kriging. Springer, New York, 1999

  34. [42]

    Cambridge University Press, Cambridge, 2004

    Holger Wendland.Scattered Data Approximation. Cambridge University Press, Cambridge, 2004

  35. [43]

    Strictly proper scoring rules, prediction, and estimation.Journal of the American Statistical Association, 102(477):359–378, 2007

    Tilmann Gneiting and Adrian E Raftery. Strictly proper scoring rules, prediction, and estimation.Journal of the American Statistical Association, 102(477):359–378, 2007. 50

Pith tools

Reviewed July 10, 2026 · model on record in the stance chip above.