Pith. sign in

REVIEW 5 major objections 4 minor 63 references

The paper claims a unified Bayesian framework that learns a low-dimensional input projection jointly with a Gaussian process (or deep GP) surrogate, propagating projection uncertainty into predictions.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 12:33 UTC pith:4DRVEJUL

load-bearing objection A useful idea—joint Bayesian projection learning for GP/DGP surrogates—but the printed core sampler is not reproducible and several equations don't parse; distinct major-revision material. the 5 major comments →

arxiv 2607.19498 v1 pith:4DRVEJUL submitted 2026-07-21 stat.ML cs.LGmath.PRstat.APstat.COstat.ME

A Bayesian Framework for Built-in Input Dimension Reduction for Gaussian Process Modeling

Eric Herrison Gyamfi , Emily L. Kang , Bledar A. Konomi , Guang Lin This is my paper
classification stat.ML cs.LGmath.PRstat.APstat.COstat.ME MSC 62F1560G1562H2562M3065C20
keywords Gaussian process regressiondimension reductionBayesian inferenceStiefel manifoldHamiltonian Monte Carlodeep Gaussian processesuncertainty quantificationcomputer experiments
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that the curse of dimensionality in Gaussian process surrogates can be addressed by learning an orthonormal projection of the inputs and the GP parameters in a single Bayesian inference, rather than in two separate stages. It argues that this joint treatment propagates projection uncertainty into predictions and yields better-calibrated uncertainty than two-stage methods. The construction extends naturally to deep Gaussian processes with two or three layers, enabling nonstationary and nonlinear response surfaces. A sympathetic reader cares because if the method works, emulators for high-dimensional simulators could be built with modest sample sizes while preserving honest uncertainty statements.

Core claim

The paper introduces a hierarchical Bayesian model in which high-dimensional inputs are projected by an orthonormal matrix W on the Stiefel manifold, a Gaussian process (or deep GP) is placed on the projected inputs, and W, kernel hyperparameters, latent layers, and noise are sampled jointly via Hamiltonian Monte Carlo with geodesic flow, elliptical slice sampling, and Gibbs updates. The central claim is that this joint inference removes the two-stage separation that ignores projection uncertainty and distorts the input-output relationship. The paper reports numerical studies showing that the built-in dimension reduction (BDR) models match or approach oracle fits that use the true projection

What carries the argument

The central object is the orthonormal projection matrix W on the Stiefel manifold V_{p,D}, assigned a Matrix Langevin prior ML(F) and sampled by Hamiltonian Monte Carlo with geodesic flow. This matrix maps high-dimensional inputs x to low-dimensional coordinates z = W^T x, and its closed-form kernel gradient with respect to W makes joint inference tractable. The same projection is embedded in deep Gaussian processes by composing it with additional GP latent layers, allowing nonlinear warping after the linear reduction.

Load-bearing premise

The load-bearing premise is that the response depends on the inputs only through a D-dimensional linear subspace; if the true active structure is nonlinear, or if the sample is too small to identify the subspace, the learned projection cannot reduce the inputs without distorting the response.

What would settle it

Construct a response whose low-dimensional structure is nonlinear, e.g., y = g(w_1^T x, w_2^T x) where g is a nonlinear function of squared or product terms, using a known orthonormal W. Fit BDR with the true D and compare RMSPE and 95% coverage to a GP fit on the full inputs. If BDR does not improve prediction and does not concentrate the posterior of W on the true column space, the linear-subspace assumption is the culprit.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If joint learning works, two-stage dimension-reduction approaches are shown to be misspecified, and users should prefer a fully Bayesian joint inference.
  • Projection uncertainty propagates through to predictions, so credible intervals stay near nominal coverage in small-data, high-dimensional regimes.
  • DGP variants with BDR can capture nonstationary and strongly nonlinear responses that plain stationary GPs, with or without BDR, cannot.
  • The higher computational cost is a real cost: per-iteration complexity is roughly O(T(n^3 + n^2 p D) + D n^3), limiting current applicability to moderate datasets.

Where Pith is reading between the lines

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

  • Editorial inference: If the true active subspace is linear, the same machinery could be adapted to non-Gaussian observation models by swapping the output layer, making the approach applicable to classification or count data.
  • Editorial inference: A practical diagnostic that the paper does not provide is monitoring posterior contraction of the projection matrix W W^T; weak contraction would signal that D is misspecified or that the linear-subspace assumption is strained.
  • Editorial inference: The fixed-D treatment leaves the subspace dimension itself uncertain; a reversible-jump or model-averaging extension would be needed to make D an inferential quantity.
  • Editorial inference: A direct test of the load-bearing linear assumption would be to apply BDR to data generated from a nonlinear active manifold and observe where predictive accuracy and coverage degrade relative to the linear case.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The paper proposes a hierarchical Bayesian model for Gaussian process (GP) and deep Gaussian process (DGP) surrogates with built-in input dimension reduction. A matrix Langevin prior is placed on the projection matrix W on the Stiefel manifold, and posterior inference is performed via a hybrid Gibbs-MH-HMC-ESS sampler, with geodesic Hamiltonian Monte Carlo for W. The method is evaluated on three synthetic examples, a stochastic elliptic PDE, and the ONERA-M6 wing design problem, comparing against no-dimension-reduction and oracle versions. The central claim is that joint learning of the projection and the surrogate improves predictive accuracy and uncertainty calibration relative to two-stage or full-space methods.

Significance. If the methodology is correct and reproducible, the framework is a useful contribution: it unifies dimension reduction and GP regression in a single Bayesian formulation, and the extension to DGPs is natural. The experimental scope is broad and the comparison against oracle models is a good idea for assessing subspace recovery. However, the printed algorithm contains several mathematical inconsistencies in load-bearing components, most importantly the HMC gradient for W. Until these are corrected and the code is made available, the numerical results cannot be taken as evidence for the method exactly as described.

major comments (5)
  1. [§2.5.5 / Eq. (2.9) / Eq. (2.3)] The gradient formula (2.9) is dimensionally inconsistent. C_y is n×n, X_n−X'_n is n×p, and W is p×D; neither (C_y/θ)(X_n−X'_n)(X_n−X'_n)^T W nor the expression in Eq. (2.3) is a well-defined p×D matrix. Since Algorithm 2.1 uses ∇_W log π in the leapfrog updates, the core sampler cannot be executed from the text. In the same paragraph, the prior gradient is stated as VΛM^T, but from π(W)∝exp(tr(F^T W)) with F=MΛV^T, the derivative is F=MΛV^T, not VΛM^T. Please provide a corrected derivation, with explicit matrix dimensions, or point to the exact corrected equation in the supplement.
  2. [§2.5.2] The stated full conditional M|W,Λ,V ∼ ML(WΛV + F_M) does not follow from the stated prior. From π(W|M,Λ,V) ∝ exp(tr(VΛM^T W)) = exp(tr((W V Λ)^T M)), the concentration matrix for M should be W V Λ, not WΛV. As printed, the Gibbs update for M is inconsistent with the prior in Eq. (2.6).
  3. [§3, CRPS definition] The printed CRPS formula is the negative of the standard Gaussian CRPS. For z=(Y−μ)/σ, the standard expression is σ[z(2Φ(z)−1)+2φ(z)−1/√π]. The paper writes σ[1/√π−2φ(z)−z(2Φ(z)−1)], which is the same expression with the opposite sign. Since all reported CRPS values are positive, either the implementation uses a different formula or the printed definition is a typo. Please correct this and state the exact scoring rule used to produce the tables.
  4. [§3, reproducibility] The text says 'A full reproducibility repository is provided at GitHub' but no URL is given. Given that the gradient equations above appear to be mathematically inconsistent, the repository is the only way to determine whether the numerical experiments used corrected expressions and whether the reported results are produced by Algorithm 2.1 as written. Please provide the repository URL and a commit hash, and explicitly state whether the code implements the printed formulas or a corrected version.
  5. [§3.1.1 / Table 1] The text states that 'the tight agreement between the oracle and BDR-based fits indicates that the proposed method recovers the correct 1D manifold with high fidelity.' In Table 1, however, DGP 3-layer(1) BDR has RMSPE 0.2133 at n=280 and 0.1601 at n=480, while DGP 3-layer(1) Oracle has RMSPE 0.0592 and 0.0432, a gap of roughly 3×. If the claim is meant only for GP(1)-BDR and DGP 2-layer(1)-BDR, the text should be narrowed; otherwise the reported numbers do not support the blanket statement.
minor comments (4)
  1. [§2.2] Equation (2.3) is used for both the kernel gradient and the predictive distribution. Renumber the predictive distribution equation to avoid ambiguity.
  2. [§2.5.4 / §3] The remark that the slice-type sampling 'does not correspond to standard Gaussian-prior ESS' appears twice (Section 2.5.4 and Section 3). Consolidate to avoid repetition.
  3. [Tables] The table captions say 'Table entries report, for each metric, the median value across posterior samples,' but TC (minutes) is a runtime measure, not a posterior quantity. Clarify whether TC is the median runtime or the total runtime.
  4. [References] There are duplicated and malformed entries in the bibliography, e.g., [13] and [14] are identical, and several entries have overlapping or broken URLs. Please clean up the reference list.

Circularity Check

0 steps flagged

No significant circularity: the posterior and predictive distributions form a self-contained Bayesian derivation; the true W is used only as an oracle benchmark, and there is no self-citation chain carrying the central argument.

full rationale

The derivation chain is self-contained and no equation reduces to its own input. The model defines a proper hierarchical posterior (Eq. 2.7) from a likelihood, a matrix Langevin prior on W (Eq. 2.6), and priors on hyperparameters; the predictive distributions (Eqs. 2.10-2.15) are ordinary GP conditionings averaged over posterior draws of (W, hyperparameters). No fitted parameter is relabeled as a prediction, and the target response never enters the model definition beyond the likelihood. In the numerical studies the oracle models use the true W only as a held-out benchmark; the true W is not fed into the BDR fits, so 'close agreement with oracle' is an honest comparison whose failure would be visible. Cited prior work ([51], [54], [22], [27]) is by disjoint author sets, so the 'first Bayesian joint framework' claim is a novelty assertion rather than a self-citation-loaded deduction; the 'Following [51]' design restrictions are adopted assumptions from independent work, not claimed derived predictions. Items worth flagging as limitations but outside the circularity definition: D is selected by comparing held-out test metrics across D=1,2,3 (model selection, not a fitted-input-called-prediction); the paper itself states D is treated as fixed so 'full uncertainty propagation' excludes D; and Eq. (2.9)/Algorithm 2.1 contain a dimensionally inconsistent gradient while the promised GitHub URL is absent, which are correctness/reproducibility risks rather than circular reductions. No exhibited Eq.-to-Eq. reduction or fitted-parameter-as-prediction exists.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

No new physical or conceptual entities are introduced; the projection matrix and latent GP layers are standard. The model pulls its main ingredients from existing literature: matrix Langevin priors, Stiefel HMC, and deep GPs. The free parameters are mostly prior and sampler hyperparameters plus the model choice D, none of which are inferred from data.

free parameters (4)
  • Reduced dimension D = D ∈ {1,2,3}
    Treated as a model choice in Section 2.1; candidate values are compared, not inferred. Performance conclusions depend on selecting the correct D.
  • Matrix Langevin concentration prior hyperparameters b1, b2 = b1=5/2, b2=10/3
    Chosen by hand in Section 3 and used as the Gamma prior for the concentration parameters λ_k.
  • Lengthscale and nugget prior scales b[θr], b[θq], b[θy], b[g] = 3.9/6, 3.9/3, 3.9, 3.9
    Hand-set hyperparameters in Eq. (3.1) that encode the ordering bθy>bθq>bθr and affect mixing and smoothness.
  • HMC integration parameters ε and T = ε=0.09, T=15
    Fixed after experimentation in Section 3; the sampler's acceptance rate and convergence depend on these values.
axioms (5)
  • domain assumption The response satisfies y ⊥ x | W^T x for some D-dimensional orthonormal projection W (linear sufficient dimension reduction).
    Section 2.1; if the active structure is nonlinear, a linear projection cannot capture it and the central claim fails.
  • domain assumption A GP with squared-exponential kernel and scalar lengthscale adequately models η on the reduced space (with DGP layers handling nonstationarity).
    Sections 2.2 and 2.4; stationarity and isotropy are assumed, and the DGP extension is needed only when that fails.
  • standard math The matrix Langevin Gibbs sampler of Hoff (2009) produces correct posterior draws for M and V.
    Section 2.5.2 relies on [27] for simulation from matrix von Mises-Fisher distributions.
  • ad hoc to paper Geodesic HMC on the Stiefel manifold with the printed gradient is a valid sampler for W.
    Section 2.5.5/Algorithm 2.1; validity depends on gradient formulas (2.3)/(2.9), whose printed dimensions are inconsistent.
  • ad hoc to paper The prior rate ordering bθy > bθq > bθr is a valid structural assumption for the smoothness hierarchy.
    Section 2.4.2; an arbitrary prior choice that stabilizes sampling but is not derived from the problem.

pith-pipeline@v1.3.0-alltime-deepseek · 44816 in / 14778 out tokens · 150470 ms · 2026-08-01T12:33:57.685940+00:00 · methodology

0 comments
read the original abstract

Gaussian process (GP) modeling is widely used in computational science and engineering. However, fitting a GP to high-dimensional inputs remains challenging due to the curse of dimensionality. While various methods have been proposed to reduce input dimensionality, they typically follow a two-stage approach, performing dimension reduction and GP fitting separately. We introduce a Bayesian framework that seamlessly integrates dimensionality reduction with GP modeling and inference. Our approach, built on a hierarchical Bayesian model with priors on the Stiefel manifold, enforces orthonormality on the projection matrix and enables posterior inference via Hamiltonian Monte Carlo with geodesic flow. Additionally, we extend this framework by incorporating Deep Gaussian Processes (DGP) with built-in dimension reduction, providing a more flexible and powerful tool for complex datasets. Through extensive numerical studies, we demonstrate that while the proposed Bayesian method incurs higher computational costs, it improves predictive performance and uncertainty quantification, providing a principled and robust alternative to existing methods.

Figures

Figures reproduced from arXiv: 2607.19498 by Bledar A. Konomi, Emily L. Kang, Eric Herrison Gyamfi, Guang Lin.

Figure 1
Figure 1. Figure 1: Schematic illustration of built-in dimension reduction: high-dimensional inputs [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Model structure of a standard Gaussian process with built-in dimension reduction. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Two-layer DGP with built-in dimension reduction. The projection [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Three-layer DGP with built-in dimension reduction. The projection [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: RMSPE for 1D input subspace at ntrain = 280 and 480 based on the response surface of a polynomial function with known structure. 3.1.2. 2D Input Subspace. For the two-dimensional input subspace, the true 10 × 2 projection matrix W is W =   0.00840 −0.18426 0.34300 −0.05347 0.08108 0.06556 −0.41219 0.65424 0.48483 0.03966 0.06720 −0.41480 0.48210 0.07550 0.21010 0.53750 0.07810 −0.20020 −0.29… view at source ↗
Figure 6
Figure 6. Figure 6: MLPPD for 1D input subspace at ntrain = 280 and 480 based on the response surface of a polynomial function with known structure. The remaining components of η(·) are a0 = −0.06976, a = (0.4376, 0.9870)⊤, A =  −0.9257 −0.3840 −0.4174 −0.6766 , σ2 ϵ = 0.01. The results for the two-dimensional case closely parallel the one-dimensional setting, with GP(2)–BDR achieving the strongest overall performance acros… view at source ↗
Figure 7
Figure 7. Figure 7: RMSPE for 2D input subspace at ntrain = 280 and 480 based on the response surface of a polynomial function with known structure. is obtained by projecting high-dimensional vectors x ∈ [0, 1]10 using a known 10 × 1 matrix W (as defined in Subsection 3.1.1). The response function is (3.2) η(z) =    1.35 cos(12πz), z < 0.333, 1.35, 0.333 ≤ z ≤ 0.666, 1.35 cos(6πz), z > 0.666, which introduces oscillatory… view at source ↗
Figure 8
Figure 8. Figure 8: MLPPD for 2D input subspace at ntrain = 280 and 480 based on the response surface of a polynomial function with known structure. n = 500. Each dataset is split into training (80%) and test (20%) sets, and Gaussian noise with mean 0 and standard deviation 0.1 is added to produce noisy outputs y. One-, two-, and three-layer DGP models are then fitted using the observed data, with the analysis focusing on the… view at source ↗
Figure 9
Figure 9. Figure 9: RMSPE for 1D input subspace at ntrain = 240 and 400 based on the response surface of a generated piecewise function. Increasing the training size to ntrain = 400 leads to systematic improvements across all BDR models, particularly in NSME and CRPS. The two-layer DGP(2)–BDR model becomes increasingly competitive as its latent mappings stabilize with additional data. These gains are more pronounced than in t… view at source ↗
Figure 10
Figure 10. Figure 10: MLPPD for 1D input subspace at ntrain = 240 and 400 based on the response surface of a generated piecewise function. mance limit, and the close agreement between the BDR fits and these benchmarks confirms that the framework recovers the governing 2D latent manifold even under strong nonlinearity. The CP and ALCI results in [PITH_FULL_IMAGE:figures/full_fig_p032_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: RMSPE for 2D input subspace at ntrain = 240 and 400 based on the response surface of a generated exponential function. with (3.10) a(0)u ′ (0) = Z 1 0 s + 1 a(s; ϵ) ds Z 1 0 1 a(s; ϵ) ds . For this study, we set u(0) = 0, a0 = 0, and σ = 1, and examine both correlation-length scenarios (β = 1 and β = 0.01). We use a training set of 270 noiseless observations and a test set of 30 points. The surrogate mode… view at source ↗
Figure 12
Figure 12. Figure 12: MLPPD for 2D input subspace at ntrain = 240 and 400 based on the response surface of a generated exponential function. The results, shown in [PITH_FULL_IMAGE:figures/full_fig_p035_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Elliptic PDE: RMSPE for BDR and W/o across different input subspaces and [PITH_FULL_IMAGE:figures/full_fig_p037_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Elliptic PDE: MLPPD for BDR and W/o across different input subspaces and [PITH_FULL_IMAGE:figures/full_fig_p038_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: ONERA M6 (drag and lift): RMSPE for BDR and W/o across different input [PITH_FULL_IMAGE:figures/full_fig_p040_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: ONERA M6 (drag and lift): MLPPD for BDR and W/o across different input [PITH_FULL_IMAGE:figures/full_fig_p041_16.png] view at source ↗

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

63 extracted references · 10 canonical work pages

  1. [1]

    Chinese Journal of Aeronautics , year=

    Application of a PCA-DBN-based surrogate model to robust aerodynamic design optimization , author=. Chinese Journal of Aeronautics , year=

  2. [2]

    2016 IEEE Symposium Series on Computational Intelligence (SSCI) , year=

    The use of Kernel PCA in evolutionary optimization for computationally demanding engineering applications , author=. 2016 IEEE Symposium Series on Computational Intelligence (SSCI) , year=

    2016

  3. [3]

    Computers & Structures , year=

    Kernel principal component analysis-based Gaussian process regression modelling for high-dimensional reliability analysis , author=. Computers & Structures , year=

  4. [4]

    Mathematical Problems in Engineering , year=

    An Improved Approach for Estimating the Hyperparameters of the Kriging Model for High-Dimensional Problems through the Partial Least Squares Method , author=. Mathematical Problems in Engineering , year=

  5. [6]

    and Dow, Eric and Wang, Qiqi , title =

    Constantine, Paul G. and Dow, Eric and Wang, Qiqi , title =. SIAM Journal on Scientific Computing , volume =. 2014 , doi =. https://doi.org/10.1137/130916138 , abstract =

    doi:10.1137/130916138 2014

  6. [7]

    AIAA AVIATION 2022 Forum , year=

    A Multi-Fidelity Approximation of the Active Subspace Method for Surrogate Models with High-Dimensional Inputs , author=. AIAA AVIATION 2022 Forum , year=

    2022

  7. [8]

    , title =

    Constantine, Paul G. , title =. 2015 , isbn =

    2015

  8. [9]

    2011 , eprint=

    Gradient-based kernel dimension reduction for supervised learning , author=. 2011 , eprint=

    2011

  9. [10]

    Dimension Reduction for Gaussian Process Emulation: An Application to the Influence of Bathymetry on Tsunami Heights , Url =

    Liu, Xiaoyu and Guillas, Serge , Doi =. Dimension Reduction for Gaussian Process Emulation: An Application to the Influence of Bathymetry on Tsunami Heights , Url =. 2017 , Bdsk-Url-1 =. https://doi.org/10.1137/16M1090648 , Journal =

    work page doi:10.1137/16m1090648 2017

  10. [11]

    International Journal for Uncertainty Quantification , DOI=

    A Fully Bayesian Gradient-Free Supervised Dimension Reduction Method using Gaussian Processes , author=. International Journal for Uncertainty Quantification , DOI=. 2022 , volumn=. 2008.03534 , archivePrefix=

    Pith/arXiv arXiv 2022

  11. [12]

    Gaussian processes with built-in dimensionality reduction: Applications to high-dimensional uncertainty propagation , journal =

    Rohit Tripathy and Ilias Bilionis and Marcial Gonzalez , keywords =. Gaussian processes with built-in dimensionality reduction: Applications to high-dimensional uncertainty propagation , journal =. 2016 , issn =. doi:https://doi.org/10.1016/j.jcp.2016.05.039 , url =

    work page doi:10.1016/j.jcp.2016.05.039 2016

  12. [13]

    Gramacy and David Higdon , title =

    Annie Sauer and Robert B. Gramacy and David Higdon , title =. Technometrics , volume =. 2023 , publisher =

    2023

  13. [14]

    Hoff , title =

    Peter D. Hoff , title =. Journal of Computational and Graphical Statistics , volume =. 2009 , publisher =. doi:10.1198/jcgs.2009.07177 , URL =

    Pith/arXiv arXiv 2009

  14. [15]

    and Ghanem, R

    Tsilifis, P. and Ghanem, R. G. , title =. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , volume =. 2018 , doi =. https://royalsocietypublishing.org/doi/pdf/10.1098/rspa.2018.0285 , abstract =

    arXiv 2018

  15. [16]

    Bayesian learning of orthogonal embeddings for multi-fidelity Gaussian Processes , journal =

    Panagiotis Tsilifis and Piyush Pandita and Sayan Ghosh and Valeria Andreoli and Thomas Vandeputte and Liping Wang , keywords =. Bayesian learning of orthogonal embeddings for multi-fidelity Gaussian Processes , journal =. 2021 , issn =. doi:https://doi.org/10.1016/j.cma.2021.114147 , url =

    arXiv 2021

  16. [17]

    Concentrated matrix Langevin distributions , journal =

    Yasuko Chikuse , keywords =. Concentrated matrix Langevin distributions , journal =. 2003 , issn =. doi:https://doi.org/10.1016/S0047-259X(02)00065-9 , url =

    work page doi:10.1016/s0047-259x(02)00065-9 2003

  17. [18]

    , booktitle =

    Damianou, Andreas and Lawrence, Neil D. , booktitle =. Deep. 2013 , editor =

    2013

  18. [19]

    Inference in Deep Gaussian Processes using Stochastic Gradient Hamiltonian Monte Carlo , url =

    Havasi, Marton and Hern\'. Inference in Deep Gaussian Processes using Stochastic Gradient Hamiltonian Monte Carlo , url =. Advances in Neural Information Processing Systems , editor =

  19. [20]

    Proceedings of The 33rd International Conference on Machine Learning , pages =

    Deep Gaussian Processes for Regression using Approximate Expectation Propagation , author =. Proceedings of The 33rd International Conference on Machine Learning , pages =. 2016 , editor =

    2016

  20. [21]

    and Girolami, Mark A

    Dunlop, Matthew M. and Girolami, Mark A. and Stuart, Andrew M. and Teckentrup, Aretha L. , title =. J. Mach. Learn. Res. , month = jan, pages =. 2018 , issue_date =

    2018

  21. [22]

    SIAM/ASA Journal on Uncertainty Quantification , volume =

    Fadikar, Arindam and Higdon, Dave and Chen, Jiangzhuo and Lewis, Bryan and Venkatramanan, Srinivasan and Marathe, Madhav , title =. SIAM/ASA Journal on Uncertainty Quantification , volume =. 2018 , doi =. https://doi.org/10.1137/17M1161233 , abstract =

    work page doi:10.1137/17m1161233 2018

  22. [23]

    and Winskill, Peter and Ghani, Azra and Bhatt, Samir and Flaxman, Seth , title =

    Charles, Giovanni and Wolock, Timothy M. and Winskill, Peter and Ghani, Azra and Bhatt, Samir and Flaxman, Seth , title =. Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence...

    work page doi:10.1609/aaai.v37i12.26658 2023

  23. [24]

    PLoS Computational Biology , year=

    Bayesian History Matching of Complex Infectious Disease Models Using Emulation: A Tutorial and a Case Study on HIV in Uganda , author=. PLoS Computational Biology , year=

  24. [25]

    Technometrics , year=

    Building Accurate Emulators for Stochastic Simulations via Quantile Kriging , author=. Technometrics , year=

  25. [26]

    Journal of the American Statistical Association , year=

    Statistical Agent-Based Models for Discrete Spatio-Temporal Systems , author=. Journal of the American Statistical Association , year=

  26. [27]

    SIAM/ASA J

    Efficient History Matching of a High Dimensional Individual-Based HIV Transmission Model , author=. SIAM/ASA J. Uncertain. Quantification , year=

  27. [28]

    Journal of Advances in Modeling Earth Systems , volume =

    Lu, Dan and Ricciuto, Daniel and Stoyanov, Miroslav and Gu, Lianhong , title =. Journal of Advances in Modeling Earth Systems , volume =. doi:https://doi.org/10.1002/2017MS001134 , url =. https://agupubs.onlinelibrary.wiley.com/doi/pdf/10.1002/2017MS001134 , abstract =

    work page doi:10.1002/2017ms001134

  28. [29]

    Real-time seismic damage simulation for urban building portfolio based on basic building information and machine learning , journal =

    Shuang Li and Changqing Li and Yuwen Huang and Changhai Zhai , keywords =. Real-time seismic damage simulation for urban building portfolio based on basic building information and machine learning , journal =. 2024 , issn =. doi:https://doi.org/10.1016/j.ijdrr.2024.104687 , url =

    arXiv 2024

  29. [30]

    and Rahimian, M

    Ahmed, Abdulrahman A. and Rahimian, M. Amin and Roberts, Mark S. , title =. Proceedings of the Winter Simulation Conference , pages =. 2024 , isbn =

    2024

  30. [31]

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

    Bayesian calibration of computer models , author=. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , year=

  31. [32]

    Geoscientific Model Development , year=

    Correcting a bias in a climate model with an augmented emulator , author=. Geoscientific Model Development , year=

  32. [33]

    2024 , issue_date =

    Donnelly, James and Daneshkhah, Alireza and Abolfathi, Soroush , title =. 2024 , issue_date =. doi:10.1016/j.engappai.2023.107536 , journal =

    arXiv 2024

  33. [34]

    2022 , MONTH = Jan, DOI =

    Marmin, S. 2022 , MONTH = Jan, DOI =

    2022

  34. [35]

    Muhit and Annalisa Occhipinti and Nashwan Dawood , keywords =

    Delbaz Samadian and Imrose B. Muhit and Annalisa Occhipinti and Nashwan Dawood , keywords =. Surrogate models for seismic and pushover response prediction of steel special moment resisting frames , journal =. 2024 , issn =. doi:https://doi.org/10.1016/j.engstruct.2024.118307 , url =

    arXiv 2024

  35. [36]

    Hybrid surrogate model combining physics and data for seismic drift estimation of shear‐wall structures , volume=

    Fei, Yifan and Liao, Wenjie and Zhao, Pengju and Lu, Xinzheng and Guan, Hong , year=. Hybrid surrogate model combining physics and data for seismic drift estimation of shear‐wall structures , volume=. Earthquake Engineering. doi:10.1002/eqe.4151 , number=

    work page doi:10.1002/eqe.4151

  36. [37]

    Gramacy, R. B. , Isbn =. Surrogates: Gaussian Process Modeling, Design, and Optimization for the Applied Sciences , Url =. 2020 , Bdsk-Url-1 =

    2020

  37. [38]

    Mehta and Andrew Walker and Earl Lawrence and Richard Linares and David Higdon and Josef Koller , keywords =

    Piyush M. Mehta and Andrew Walker and Earl Lawrence and Richard Linares and David Higdon and Josef Koller , keywords =. Modeling satellite drag coefficients with response surfaces , journal =. 2014 , issn =. doi:https://doi.org/10.1016/j.asr.2014.06.033 , url =

    doi:10.1016/j.asr.2014.06.033 2014

  38. [39]

    Springer Series in Statistics , year=

    The Design and Analysis of Computer Experiments , author=. Springer Series in Statistics , year=

  39. [40]

    Adaptive computation and machine learning , year=

    Gaussian Processes for Machine Learning , author=. Adaptive computation and machine learning , year=

  40. [41]

    Radaideh and Tomasz Kozlowski , keywords =

    Majdi I. Radaideh and Tomasz Kozlowski , keywords =. Surrogate modeling of advanced computer simulations using deep Gaussian processes , journal =. 2020 , issn =. doi:https://doi.org/10.1016/j.ress.2019.106731 , url =

    arXiv 2020

  41. [42]

    Proceedings of the 31st International Conference on Neural Information Processing Systems , pages =

    Salimbeni, Hugh and Deisenroth, Marc Peter , title =. Proceedings of the 31st International Conference on Neural Information Processing Systems , pages =. 2017 , isbn =

    2017

  42. [43]

    Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics , pages =

    Elliptical slice sampling , author =. Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics , pages =. 2010 , editor =

    2010

  43. [44]

    , title =

    Smith, Ralph C. , title =. 2024 , doi =

    2024

  44. [45]

    and Casella, George , title =

    Robert, Christian P. and Casella, George , title =. 2005 , isbn =

    2005

  45. [46]

    Dynamic Programming and Lagrange Multipliers , urldate =

    Richard Bellman , journal =. Dynamic Programming and Lagrange Multipliers , urldate =

  46. [47]

    The Curse of Highly Variable Functions for Local Kernel Machines , url =

    Bengio, Yoshua and Delalleau, Olivier and Roux, Nicolas , booktitle =. The Curse of Highly Variable Functions for Local Kernel Machines , url =

  47. [48]

    2008 , publisher=

    Global Sensitivity Analysis: The Primer , author=. 2008 , publisher=

    2008

  48. [49]

    2012 , publisher=

    Bayesian Learning for Neural Networks , author=. 2012 , publisher=

    2012

  49. [50]

    2003 , publisher=

    Stochastic Finite Elements: A Spectral Approach , author=. 2003 , publisher=

    2003

  50. [51]

    Gramacy , title =

    Annie Sauer and Andrew Cooper and Robert B. Gramacy , title =. Journal of Computational and Graphical Statistics , volume =. 2023 , publisher =. doi:10.1080/10618600.2022.2129662 , URL =

    arXiv 2023

  51. [53]

    Lukaczyk and Paul Constantine and Francisco Palacios and Juan J

    Trent W. Lukaczyk and Paul Constantine and Francisco Palacios and Juan J. Alonso , title =. 10th AIAA Multidisciplinary Design Optimization Conference , year =. doi:10.2514/6.2014-1171 , note =

    work page doi:10.2514/6.2014-1171 2014

  52. [54]

    Kenway, Gaetan K. W. and Martins, Joaquim R. R. A. , title =. Journal of Aircraft , volume =. 2014 , doi =. https://doi.org/10.2514/1.C032150 , abstract =

    work page doi:10.2514/1.c032150 2014

  53. [55]

    Scandinavian Journal of Statistics , volume =

    Byrne, Simon and Girolami, Mark , title =. Scandinavian Journal of Statistics , volume =. doi:https://doi.org/10.1111/sjos.12036 , url =. https://onlinelibrary.wiley.com/doi/pdf/10.1111/sjos.12036 , abstract =

    doi:10.1111/sjos.12036

  54. [56]

    and Kent, Carson and Bui-Thanh, Tan , title =

    Constantine, Paul G. and Kent, Carson and Bui-Thanh, Tan , title =. SIAM Journal on Scientific Computing , volume =. 2016 , doi =

    2016

  55. [57]

    Welch and Toby J

    Jerome Sacks and William J. Welch and Toby J. Mitchell and Henry P. Wynn , title =. Statistical Science , number =. 1989 , doi =

    1989

  56. [58]

    Electronic Journal of Statistics , number =

    Wenjia Wang , title =. Electronic Journal of Statistics , number =. 2021 , doi =

    2021

  57. [59]

    A Survey on High-dimensional Gaussian Process Modeling with Application to Bayesian Optimization , year =

    Binois, Micka\". A Survey on High-dimensional Gaussian Process Modeling with Application to Bayesian Optimization , year =. ACM Trans. Evol. Learn. Optim. , month = aug, articleno =. doi:10.1145/3545611 , abstract =

    doi:10.1145/3545611

  58. [60]

    IISE Transactions , volume =

    Gecheng Chen and Rui Tuo , title =. IISE Transactions , volume =. 2023 , publisher =. doi:10.1080/24725854.2022.2121882 , URL =

    arXiv 2023

  59. [61]

    and Lee, Herbert K

    Gramacy, Robert B. and Lee, Herbert K. , title =. Statistics and Computing , month = may, pages =. 2012 , issue_date =. doi:10.1007/s11222-010-9224-x , abstract =

    work page doi:10.1007/s11222-010-9224-x 2012

  60. [62]

    Rubin , title =

    Andrew Gelman and Donald B. Rubin , title =. Statistical Science , number =. 1992 , doi =

    1992

  61. [63]

    Welch , journal =

    Philip Heidelberger and Peter D. Welch , journal =. Simulation Run Length Control in the Presence of an Initial Transient , urldate =

  62. [64]

    CODA: convergence diagnosis and output analysis for MCMC , journal =

    Plummer, Martyn and Best, Nicky and Cowles, Kate and Vines, Karen , url =. CODA: convergence diagnosis and output analysis for MCMC , journal =. 2006 , number =

    2006

  63. [65]

    Geyer , journal =

    Charles J. Geyer , journal =. Practical Markov Chain Monte Carlo , urldate =