Pith. sign in

REVIEW 4 major objections 4 minor 54 references

Multi-output Orthogonal Gaussian Processes for Noisy Simulator Outputs

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read MOOGP conditions each latent GP kernel to be orthogonal to the regression basis, making trend coefficients identifiable and likelihood evaluation cheap.

desk verdict A clean, competent extension of orthogonal GP to multi-output emulation; the core math checks out, but the data-fitted output basis is plugged in as fixed, so the uncertainty claims need a caveat. read the letter →

arxiv 2608.00287 v1 pith:KUVQGT64 submitted 2026-07-31 stat.ME stat.CO

Evan C. Barnett , Moses Y.-H. Chan This is my paper
classification stat.MEstat.CO
keywords multi-outputGaussianprocessesorthogonalprocessidentifiabilitysurrogatemodelingstochasticemulatorcomputerexperimentsheavy-ioncollisionslatentfactormodel
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

The paper introduces MOOGP, a surrogate model for noisy vector-valued simulators that forces the GP residual to be orthogonal to the chosen regression trend over the entire input domain. This is done by conditioning each latent covariance kernel on the event that its integral against the regression basis is zero, which removes the main source of confounding between mean and residual. As a by-product, the construction yields a factorized likelihood that only needs q inversions of n-by-n matrices instead of one inversion of an np-by-np matrix. The authors demonstrate that the orthogonalized model recovers low-order trends even under clustered designs that make standard multi-output GPs reverse the sign of the slope, and they show competitive predictive performance on a 98-output heavy-ion collision simulator.

What carries the argument

The key object is the conditioned kernel c_k^*(x,x') = c_k(x,x') - h_k(x)^T H_k^{-1} h_k(x'), which is the covariance of the latent GP conditioned on its integral against the regression basis being zero. This kernel is valid whenever H_k is positive definite, and it forces the latent process to have zero projection on the regression space with probability one. The companion factor is the output loading decomposition Ψ = Σ_ε^{1/2} Φ with Φ^T Φ = D, which diagonalizes Ψ^T Σ_ε^{-1} Ψ and turns the likelihood's nq×nq interaction term into a block-diagonal matrix, leaving only q n×n inverses.

What would settle it

Take a non-separable kernel (e.g., a non-product stationary kernel) and the same regression basis g(x)=(1,x); compute h and H by high-accuracy quadrature, form c^*, and check empirically whether ∫ g(x) z^*(x) dx is still zero across many samples. If it is not, the paper's Theorem 1 relies on assumptions that exclude this case; if it is, the method's practical scope can extend beyond the closed-form family, but the computational scaling claim would still fail.

Watch

Extended reading notes

Core claim

MOOGP defines a multi-output orthogonal Gaussian process by the condition ∫_X g(x) w(x)^T dx = 0 almost surely, where g is the regression basis and w is the residual process, and proves (Theorem 1) that this condition is enforced by replacing each latent kernel c_k with the conditioned kernel c_k^*(x,x') = c_k(x,x') - h_k(x)^T H_k^{-1} h_k(x'). Here h_k(x) = ∫ c_k(x,ξ) g(ξ) dξ and H_k = ∫∫ g(ξ) c_k(ξ,ξ') g(ξ')^T dξ dξ'. Under boundedness and positive-definiteness of H_k, each latent process and hence the mixed process w = Ψ z^* is orthogonal to span(g) with probability one. The paper also shows that for separable squared-exponential kernels and regression bases of products of distinct coordi

Load-bearing premise

The load-bearing requirement is that the integrals h_k(x) = ∫ c_k(x,ξ) g(ξ) dξ and H_k = ∫∫ g c_k g^T can be computed exactly and that H_k is invertible; with the paper's separable squared-exponential kernel and product-of-distinct-coordinates basis, those integrals are closed-form, but outside that family the construction is not guaranteed to be feasible.

Editorial extensions

If this is right

  • With orthogonality enforced, regression coefficients become identifiable: the fitted trend does not depend on whether the training design happens to be space-filling, so scientists can interpret low-order effects in the surrogate without worrying that the residual absorbed them.
  • Under clustered designs, MOOGP recovers a known linear trend to within about 0.1, while a non-orthogonal multi-output GP reverses the sign of the slope; orthogonality fixes the confounding even when points are concentrated near one edge.
  • Because the likelihood factorizes into q n×n components, MOOGP can handle many outputs p through a low-rank latent basis (q ≤ p) at a cost that grows with q rather than with p cubed.
  • In the heavy-ion collision application, the orthogonalized model gives lower standardized RMSE and narrower prediction intervals than its non-orthogonal counterpart on ten of eleven observable families, and its stable coefficient maps differ most where orthogonality matters.
  • The predictive benefit of MOOGP over non-orthogonal MOGP is small or even slightly negative in some benchmarks; the paper's ablation attributes the main predictive gain to output sharing, whereas orthogonality mainly improves trend interpretation, not point prediction.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the exact-integral requirement could be relaxed with numerical quadrature or approximate conditioning, the identifiability property might extend to other stationary kernels; the closed-form scaling would then be lost, but the qualitative guarantee of separating trend from residual could persist.
  • Because the orthogonality is defined against the Lebesgue measure on the input domain, MOOGP implicitly assumes the target trend is the projection of the response onto the regression space under that measure; under non-uniform sampling that measure may not reflect the design distribution, which is exactly why ordinary least squares fails under clustering—an assumption worth testing when designing
  • The same conditioning trick could be applied to the discrepancy term in Bayesian calibration, potentially making model parameters and code discrepancy identifiable in multi-output systems—an extension the paper names but does not implement.
  • A testable extension is to let the output loading matrix Ψ be random and to check whether the orthogonality condition still holds after marginalization; the current theorem treats Ψ as fixed, so a fully Bayesian treatment would require re-proving orthogonality of the mixed process.
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

4 major / 4 minor

Summary. The paper introduces MOOGP, a multi-output Gaussian process emulator for noisy vector-valued simulator outputs. The key idea is to extend scalar orthogonal GP (Plumlee and Joseph, 2018) to the multi-output setting by requiring the residual process Ψz*(x) to be orthogonal to the regression basis in L²(X) with probability one. Orthogonality is enforced by conditioning each latent GP on the event that its integral against the regression basis is zero, yielding the kernel c*_k = c_k − h_kᵀ H_k⁻¹ h_k (Eq. 11, Theorem 1). The manuscript also develops a Woodbury-based likelihood reduction (Section 3.4.2) that evaluates the marginal likelihood with q inversions of n×n matrices rather than one np×np inversion. Numerical experiments include a trend-recovery illustration, a 2×2 ablation separating orthogonality from output sharing, a multi-output Borehole benchmark, and a heavy-ion collision (VAH) cross-validation. The central mathematical derivation is clean and appears correct, but the implementation uses a data-dependent SVD-based output basis that is treated as fixed, and several claims in the abstract and conclusion are stronger than the reported evidence.

Significance. If read as a method whose inference is conditional on the estimated output basis, the paper makes a sound and useful contribution: it provides a principled multi-output extension of orthogonal GP, with a transparent likelihood derivation (Propositions 1–2), a correct Woodbury-based complexity reduction, and an open-source implementation with reproducible experiments. The controlled trend-recovery experiment clearly illustrates the population-level confounding in non-orthogonal multi-output GP and shows that orthogonality substantially improves trend identification. The 2×2 ablation is well designed to separate orthogonality from output sharing. However, the advertised 'closed-form uncertainty quantification' is not fully supported for the implemented procedure because the output basis is estimated from the training response and its uncertainty is not propagated; observed coverages (VAH 0.922, Borehole ~0.90) are below nominal. The abstract's sign-reversal claim is not supported by Table 1, and the conclusion's 'sharper predictions' claim is contradicted by Table 2 and Figure 2. These issues are fixable with careful reframing and additional diagnostics, but they are load-bearing for

major comments (4)
  1. [§3.4.2, Eq. (14)] The implementation sets Ψ = Σ_ε^{1/2} Φ with Φ and D obtained from the thin SVD of the training response Y, then treats Φ and D as fixed plug-in quantities. Consequently Proposition 1 and the predictive equations (7)–(8) are not the likelihood and predictive distribution of the generative GP described by Definition 1; they condition on a data-dependent estimate of the output basis. Basis-selection uncertainty is not propagated, so the reported intervals are conditional intervals and can undercover. Table 3 (MOOGP coverage 0.922; also Borehole coverages ~0.90 in Fig. 2) is consistent with this concern. The paper discloses the omission, but the advertised 'closed-form uncertainty quantification' is not supported for the implemented procedure unless this is addressed or the inferential target is explicitly redefined.
  2. [Abstract; §4.1, Table 1] The abstract states that the non-orthogonalized counterpart fails 'to the extent it reverses the sign of the trend.' Table 1 does not support this. For all three outputs, the mean MOGP slope has the same sign as the truth (17.37 vs 12; −4.61 vs −10; 9.28 vs 6). The demonstrated failure is a large magnitude bias, not sign reversal. Please correct the abstract and any related text.
  3. [§6; Table 2; Fig. 2] The conclusion states that under uneven coverage MOOGP 'produces sharper predictions.' This is not supported by the reported experiments. In the Forrester ablation (Table 2), MOOGP has slightly higher sRRMSE than MOGP under both designs (0.22 vs 0.21 uniform; 0.46 vs 0.44 clustered). In the Borehole benchmark (Fig. 2), neither MOOGP nor MOGP is consistently better. The only clear predictive gain over MOGP appears in the VAH application (Table 3). Please either temper the conclusion or restrict it to the VAH setting.
  4. [§4.1] The target coefficients in Table 1 are the population L² projection of the noise-free signal for the raw outputs, but Section 3.1 says outputs are standardized before fitting. If MOOGP/MOGP standardize internally and the table reports back-transformed coefficients, this should be stated explicitly; if the outputs are not standardized in this experiment, the text contradicts Section 3.1. As written, the comparison between least squares on raw y and GP fits on standardized y is unclear.
minor comments (4)
  1. [§3.4.1; Appendix B] The closed-form implementation is restricted to separable squared-exponential kernels and regression bases whose factors use distinct input coordinates. This restriction is disclosed, but it should be stated more prominently: for general kernels or basis functions, h_k and H_k are intractable and the method loses its closed-form structure.
  2. [§6] The sentence 'the specific difference, most notably in a_high and (ζ/s)_max, resulted from enforcing orthogonoality demonstrates interpretative quality of the proposed method' is grammatically awkward and contains a typo ('orthogonoality'). Please revise.
  3. [§3.2; Prop. 2 proof] Minor notation issues: extra commas appear in the statement of the error distribution, and the citation to 'H¨ardle and Hl´avka (2015, ch. 5)' is not harmonized with the reference list entry.
  4. [§4.2, Table 2] The coverage column in Table 2 would benefit from Monte Carlo standard errors. With 200 paired replications the standard error is small, but reporting it would make the coverage differences more interpretable.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; disclosed SVD plug-in basis is a self-referential estimation caveat, not a reduction of a prediction to a fitted value.

full rationale

The derivation chain is self-contained. Theorem 1 enforces Definition 1 by conditioning each latent kernel as c*_k = c_k - h_k^T H_k^{-1} h_k (Eq. 11); Appendix A computes Cov(∫ g z*) = H - H H^{-1} H = 0, the standard conditional-covariance identity, so the orthogonality result is a direct consequence of the construction rather than a restatement of the conclusion. Proposition 1 and the predictive equations are ordinary multivariate-normal conditionals conditional on fixed Ψ. The paper does not reduce a prediction to a fitted value. The one self-referential element is Section 3.4.2: Φ and D are obtained from the thin SVD of the training response Y and then 'treated as fixed plug-in quantities during optimization, and their estimation uncertainty is not propagated.' This makes the implemented likelihood a profile/composite likelihood and can explain undercoverage (Borehole ~0.90, VAH 0.922), but the paper explicitly discloses this limitation, and it is a statistical plug-in issue, not a circular derivation. External Borehole and VAH benchmarks against LCGP, OILMM, PUQ, plus the 2×2 ablation, provide independent support. No load-bearing self-citation chain is used: scalar OGP is cited to Plumlee and Joseph (2018), not to the authors' own prior work.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The model introduces no new physical entities. The free parameters are standard GP hyperparameters, noise variances, a rank-selection choice, and an empirical SVD basis. The axioms are standard GP assumptions, analytic-tractability restrictions specific to this implementation, and a low-rank output subspace assumption.

free parameters (4)
  • Latent kernel length scales and variances (ℓ_kj, σ_k²)
    Estimated by maximum likelihood; they control the smoothness and amplitude of each latent residual process and enter the orthogonalized kernel c*_k through h_k and H_k.
  • Observation noise variances (σ_ℓ²)
    Estimated by maximum likelihood; part of the diagonal noise covariance Σ_ε.
  • Latent rank q = q=5 in VAH; 99% cumulative SVD threshold in Borehole
    Model-selection choice determining the number of latent processes; affects the output covariance rank and computation time.
  • Empirical output basis Φ and diagonal D
    Set from the thin SVD of the standardized training response Y and treated as fixed plug-in quantities; estimation uncertainty is not propagated (Section 3.4.2).
assumptions (5)
  • domain assumption Data follow a Gaussian process with independent, identically distributed Gaussian observation errors (Equation 5).
    This is the standard GP emulator assumption; it underlies the likelihood and predictive equations.
  • domain assumption The input domain X is bounded, the kernel c_k is bounded and continuous, and H_k is positive definite for each latent component (Lemma 1).
    Needed for the stochastic integrals in Equation (10) to exist and for the conditional covariance c*_k to be a valid positive semidefinite kernel.
  • standard math The conditioning on A_k = 0 defines a Gaussian process via conditional covariance; standard modification results (Doob 1953) apply.
    Used in the proof of Theorem 1 (Appendix A) to justify measurability and Fubini.
  • ad hoc to paper The implementation uses separable squared-exponential kernels and regression basis functions that are products of distinct input coordinates (Section 3.4.1).
    This is the specific kernel/basis family for which the integrals h_k and H_k have closed forms. The general construction would require numerical integration.
  • domain assumption The output dependence is approximated by a low-rank linear subspace captured by the SVD of the standardized training response.
    The plug-in basis Φ and D (Section 3.4.2) assume the response lies approximately in a q-dimensional linear subspace; this is an empirical modeling choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-output Orthogonal Gaussian Processes for Noisy Simulator Outputs." pith.science (2026). https://pith.science/paper/KUVQGT64

@misc{pith2026260800287,
  author       = {Pith},
  title        = {Pith review of: Multi-output Orthogonal Gaussian Processes for Noisy Simulator Outputs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KUVQGT64}},
  note         = {Machine review of arXiv:2608.00287}
}
read the original abstract

Computer simulations can model physical processes but are often too expensive to produce enough runs for calibration, sensitivity analysis, prediction, and uncertainty quantification. As a result, statistical surrogates are frequently used as cheaper alternatives that can be trained on a small number of simulation runs. Gaussian processes (GP) are well-suited as surrogates, as they provide flexible, nonlinear regression and closed-form uncertainty quantification. However, standard GPs are insufficient replacements for more complex simulators, such as stochastic or multi-output simulators. In addition, typical GPs also struggle to separate fitted regression coefficients from residual process variation. We introduce multi-output orthogonal Gaussian process (MOOGP) to tackle the problems mentioned above. This paper has three main contributions: (i) a definition of orthogonality for multi-output Gaussian processes, (ii) a construction that enforces orthogonality by conditioning each covariance function, and (iii) a structured likelihood formulation that significantly improves computational scaling. In a trend-recovery illustrative example, MOOGP recovers the true trend while the non-orthogonalized counterpart fails to do so, to the extent it reverses the sign of the trend. Numerical experiments and an application in heavy-ion collision simulations further demonstrate the interpretability and predictive advantage of MOOGP.

Figures

Figures reproduced from arXiv: 2608.00287 by the authors.

Figure 1
Figure 1. Trend recovery for the data-generating function in Equation (17) under the [PITH_FULL_IMAGE:figures/full_fig_p016_1.png] view at source ↗
Figure 2
Figure 2. Borehole test performance across five paired replications for each ( [PITH_FULL_IMAGE:figures/full_fig_p021_2.png] view at source ↗
Figure 3
Figure 3. VAH five-fold cross-validation by observable family. Each mark pools all held-out [PITH_FULL_IMAGE:figures/full_fig_p024_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Standardized low-order trend coefficients for the VAH application. Rows are the [PITH_FULL_IMAGE:figures/full_fig_p025_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 1 canonical work pages

  1. [1]

    2008 , publisher=

    Matrix Algebra From a Statistician's Perspective , author=. 2008 , publisher=

    2008

  2. [2]

    Statistica Sinica , volume=

    Orthogonal Gaussian process models , author=. Statistica Sinica , volume=. 2018 , doi=

    2018

  3. [3]

    Technometrics , volume=

    Bayesian Design and Analysis of Computer Experiments: Use of Derivatives in Surface Prediction , author=. Technometrics , volume=. 1993 , doi=

    1993

  4. [4]

    arXiv preprint arXiv:2205.12361 , year=

    Bayesian Modeling of Nearly Mutually Orthogonal Processes , author=. arXiv preprint arXiv:2205.12361 , year=

    arXiv

  5. [5]

    Statistica Sinica , pages=

    Orthogonal-maximin Latin hypercube designs , author=. Statistica Sinica , pages=. 2008 , publisher=

    2008

  6. [6]

    Multi-Fidelity Analysis , booktitle =

    Forrester, Sóbester, Keane , publisher =. Multi-Fidelity Analysis , booktitle =. doi:https://doi.org/10.1002/9780470770801.ch8 , url =. https://onlinelibrary.wiley.com/doi/pdf/10.1002/9780470770801.ch8 , year =

    work page doi:10.1002/9780470770801.ch8

  7. [7]

    arXiv preprint arXiv:1908.08868 , year=

    BdryGP: a new Gaussian process model for incorporating boundary information , author=. arXiv preprint arXiv:1908.08868 , year=

    arXiv 1908

  8. [8]

    Statistica Sinica , number =

    Gaussian process modeling with boundary information , author=. Statistica Sinica , number =

Show all 54 references
  1. [9]

    Journal of the American Statistical Association , volume=

    Rational kriging , author=. Journal of the American Statistical Association , volume=. 2025 , publisher=

    2025

  2. [10]

    2015 , publisher=

    Multivariate statistics: exercises and solutions , author=. 2015 , publisher=

    2015

  3. [11]

    Statistical Science , volume=

    Design and Analysis of Computer Experiments , author=. Statistical Science , volume=. 1989 , doi=

    1989

  4. [12]

    2003 , publisher=

    The Design and Analysis of Computer Experiments , author=. 2003 , publisher=

    2003

  5. [13]

    2026 , publisher=

    Experimental Design for Data Science and Engineering , author=. 2026 , publisher=

    2026

  6. [14]

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

    Bayesian Calibration of Computer Models , author=. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , volume=. 2001 , doi=

    2001

  7. [15]

    2006 , publisher=

    Gaussian Processes for Machine Learning , author=. 2006 , publisher=

    2006

  8. [16]

    2020 , publisher=

    Surrogates: Gaussian Process Modeling, Design, and Optimization for the Applied Sciences , author=. 2020 , publisher=

    2020

  9. [17]

    Operations Research , volume=

    Stochastic Kriging for Simulation Metamodeling , author=. Operations Research , volume=. 2010 , doi=

    2010

  10. [18]

    Statistics and Computing , volume=

    Cases for the Nugget in Modeling Computer Experiments , author=. Statistics and Computing , volume=. 2012 , doi=

    2012

  11. [19]

    Statistical Science , volume=

    Analyzing Stochastic Computer Models: A Review with Opportunities , author=. Statistical Science , volume=. 2022 , doi=

    2022

  12. [20]

    SIAM/ASA Journal on Uncertainty Quantification , volume=

    Calibrating a Stochastic, Agent-Based Model Using Quantile-Based Emulation , author=. SIAM/ASA Journal on Uncertainty Quantification , volume=. 2018 , doi=

    2018

  13. [21]

    Statistics and Computing , volume=

    Global Sensitivity Analysis of Stochastic Computer Models with Joint Metamodels , author=. Statistics and Computing , volume=. 2012 , doi=

    2012

  14. [22]

    Mark , journal=

    Herbei, Radu and Berliner, L. Mark , journal=. Estimating Ocean Circulation: An. 2014 , doi=

    2014

  15. [23]

    Journal of Computational and Graphical Statistics , volume=

    Practical Heteroscedastic Gaussian Process Modeling for Large Simulation Experiments , author=. Journal of Computational and Graphical Statistics , volume=. 2018 , doi=

    2018

  16. [24]

    Journal of Statistical Planning and Inference , volume=

    Bayesian Emulation of Complex Multi-Output and Dynamic Computer Models , author=. Journal of Statistical Planning and Inference , volume=. 2010 , doi=

    2010

  17. [25]

    Journal of the American Statistical Association , volume=

    Computer Model Calibration Using High-Dimensional Output , author=. Journal of the American Statistical Association , volume=. 2008 , doi=

    2008

  18. [26]

    and Furnstahl, Richard J

    Phillips, Daniel R. and Furnstahl, Richard J. and Heinz, Ulrich and Maiti, Tapabrata and Nazarewicz, Witold and Nunes, Filomena M. and Plumlee, Matthew and Pratola, Matthew T. and Pratt, Scott and Viens, Frederi G. and Wild, Stefan M. , journal=. Get on the. 2021 , doi=

    2021

  19. [27]

    Physical Review C , volume=

    Bayesian Calibration of Viscous Anisotropic Hydrodynamic Simulations of Heavy-Ion Collisions , author=. Physical Review C , volume=. 2023 , doi=

    2023

  20. [28]

    The Annals of Applied Statistics , volume=

    Parallel Partial Gaussian Process Emulation for Computer Models with Massive Output , author=. The Annals of Applied Statistics , volume=. 2016 , doi=

    2016

  21. [29]

    Journal of Open Source Software , volume=

    O'Gara, David and Binois, Micka. Journal of Open Source Software , volume=. 2025 , doi=

    2025

  22. [30]

    2023 , month=

    High-Dimensional Gaussian Process Methods for Uncertainty Quantification , author=. 2023 , month=

    2023

  23. [31]

    Technometrics , volume=

    Constructing a Simulation Surrogate with Partially Observed Output , author=. Technometrics , volume=. 2024 , doi=

    2024

  24. [32]

    Journal of Computational and Graphical Statistics , volume=

    Efficient Emulators for Multivariate Deterministic Functions , author=. Journal of Computational and Graphical Statistics , volume=. 2008 , doi=

    2008

  25. [33]

    Proceedings of the Tenth International Workshop on Artificial Intelligence and Statistics , pages=

    Semiparametric Latent Factor Models , author=. Proceedings of the Tenth International Workshop on Artificial Intelligence and Statistics , pages=. 2005 , editor=

    2005

  26. [34]

    Advances in Neural Information Processing Systems 20 , pages=

    Multi-Task Gaussian Process Prediction , author=. Advances in Neural Information Processing Systems 20 , pages=. 2008 , editor=

    2008

  27. [35]

    Journal of Machine Learning Research , volume=

    Computationally Efficient Convolved Multiple Output Gaussian Processes , author=. Journal of Machine Learning Research , volume=. 2011 , url=

    2011

  28. [36]

    Foundations and Trends in Machine Learning , volume=

    Kernels for Vector-Valued Functions: A Review , author=. Foundations and Trends in Machine Learning , volume=. 2012 , doi=

    2012

  29. [37]

    Proceedings of the 37th International Conference on Machine Learning , pages=

    Scalable Exact Inference in Multi-Output Gaussian Processes , author=. Proceedings of the 37th International Conference on Machine Learning , pages=. 2020 , editor=

    2020

  30. [38]

    Technometrics , volume=

    Multivariate Gaussian Process Emulators with Nonseparable Covariance Structures , author=. Technometrics , volume=. 2013 , doi=

    2013

  31. [39]

    Computational Statistics & Data Analysis , volume=

    Calibration of Computer Models with Multivariate Output , author=. Computational Statistics & Data Analysis , volume=. 2012 , doi=

    2012

  32. [40]

    Technometrics , volume=

    Sequential Bayesian Experimental Design for Calibration of Expensive Simulation Models , author=. Technometrics , volume=. 2024 , doi=

    2024

  33. [41]

    Knowledge-Based Systems , volume=

    Remarks on Multi-Output Gaussian Process Regression , author=. Knowledge-Based Systems , volume=. 2018 , doi=

    2018

  34. [42]

    Journal of the American Statistical Association , volume=

    Strictly Proper Scoring Rules, Prediction, and Estimation , author=. Journal of the American Statistical Association , volume=. 2007 , doi=

    2007

  35. [43]

    Journal of the American Statistical Association , volume=

    Additive multi-index Gaussian process modeling, with application to multi-physics surrogate modeling of the quark-gluon plasma , author=. Journal of the American Statistical Association , volume=. 2026 , publisher=

    2026

  36. [44]

    arXiv preprint arXiv:2211.02465 , year=

    Multi-output Gaussian processes for inverse uncertainty quantification in neutron noise analysis , author=. arXiv preprint arXiv:2211.02465 , year=

    arXiv

  37. [45]

    The Annals of Applied Statistics , volume=

    High-fidelity hurricane surge forecasting using emulation and sequential experiments , author=. The Annals of Applied Statistics , volume=. 2021 , publisher=

    2021

  38. [46]

    Journal of Physics G: Nuclear and Particle Physics , volume=

    A Bayesian approach for parameter estimation and prediction using a computationally intensive model , author=. Journal of Physics G: Nuclear and Particle Physics , volume=. 2015 , publisher=

    2015

  39. [47]

    Technometrics , volume=

    Analysis of computer experiments with functional response , author=. Technometrics , volume=. 2015 , publisher=

    2015

  40. [48]

    Journal of the American Statistical Association , volume=

    Bayesian calibration of inexact computer models , author=. Journal of the American Statistical Association , volume=. 2017 , publisher=

    2017

  41. [49]

    SIAM/ASA Journal on Uncertainty Quantification , volume=

    Adjustments to computer models via projected kernel calibration , author=. SIAM/ASA Journal on Uncertainty Quantification , volume=. 2019 , publisher=

    2019

  42. [50]

    Journal of the American Statistical Association , volume=

    Bayesian projected calibration of computer models , author=. Journal of the American Statistical Association , volume=. 2021 , publisher=

    2021

  43. [51]

    arXiv preprint arXiv:2603.05545 , year=

    VarP-GP: cost-efficient Bayesian emulation of quark-gluon plasma modeling with variable statistical precision , author=. arXiv preprint arXiv:2603.05545 , year=

  44. [52]

    Physical Review C , volume=

    (3+ 1)-dimensional anisotropic fluid dynamics with a lattice QCD equation of state , author=. Physical Review C , volume=. 2018 , publisher=

    2018

  45. [53]

    Anisotropic fluid dynamical simulations of heavy-ion collisions , journal =

    Mike McNelis and Dennis Bazow and Ulrich Heinz , keywords =. Anisotropic fluid dynamical simulations of heavy-ion collisions , journal =. 2021 , issn =. doi:https://doi.org/10.1016/j.cpc.2021.108077 , url =

    2021

  46. [54]

    , title =

    Doob, Joseph L. , title =. 1953 , language =

    1953

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.