Pith. sign in

REVIEW 5 major objections 6 minor 47 references

GP-Recipe: Gaussian Process approximation to linear operations in numerical methods

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Gaussian-process regression can replace polynomial stencils in numerical methods, reaching (2r+1)th-order accuracy for interpolation and reduced-but-high order for derivatives and integrals.

desk verdict A useful kernel-based operator library with clean convergence tests, but the printed derivative kernels have the wrong sign and the new DAS kernel is incompletely specified — needs correction before it works as a recipe. read the letter →

arxiv 2506.03471 v2 pith:64JHGUTB submitted 2025-06-04 physics.comp-ph cs.NAmath.NA

Christopher DeGrendele , Dongwook Lee This is my paper
classification physics.comp-phcs.NAmath.NA MSC 65D0565D2565D3065M06 PACS 02.60.Cb
keywords Gaussianprocessregressionkernel-basedapproximationhigh-ordermethodsfinitedifferencevolumesquaredexponentialkernelDASeffectiveorderofconvergence
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

This paper argues that Gaussian process (GP) regression—a kernel-based, nonparametric Bayesian method—can serve as a general replacement for the polynomial stencil formulas used in finite difference and finite volume methods. Using the squared exponential kernel on a stencil of radius $r_{gp}$, the GP posterior mean gives ($2r_{gp}+1$)th-order interpolation and one-order-lower convergence for first derivatives, with analogous rates for integrals and data-type conversions between pointwise and cell-averaged values. The authors introduce an effective-order-of-convergence (EOC) metric based on Taylor expansion to choose multidimensional stencils without exhaustive grid studies, and they propose a new stationary 'DAS' kernel that interpolates across discontinuities without oscillations. If correct, kernel weights computed once and reused can deliver accuracy comparable to or better than conventional finite-difference schemes, directly plugging into existing PDE solvers.

What carries the argument

The load-bearing object is the prediction vector $z_*^T = k_*^T K^{-1}$: for stationary kernels on a fixed grid it is data-independent and can be precomputed, so each GP operation reduces to a dot product. The paper extends this to linear operations by replacing $K$ and $k_*$ with kernels transformed by operators $L_1$ and $L_2$ (e.g., differentiation or cell-volume integration), exploiting the fact that a GP stays a GP under linear maps. The effective order of convergence EOC — a weighted mean of Taylor-term orders — completes the toolkit by predicting which 2D stencil gives the desired accuracy before running convergence tests. The new DAS kernel is the proposed mechanism for oscillation-free interpolation across discontinuities.

What would settle it

Assemble the DAS kernel matrix from Eq. (40) on a uniform one-dimensional grid of 64 points and compute its eigenvalues. Any negative eigenvalue means $K$ is not a valid covariance and the posterior mean is not the GP prediction claimed; the oscillation-free interpolation shown in Figs. 5 and 6 should also be re-run once the formula is unambiguously defined.

Watch

Extended reading notes

Core claim

The central claim is that GP's invariance under linear operations turns one kernel into a family of high-order numerical operators: applying a desired linear operator to the kernel in the prediction vector $z_*^T = k_*^T K^{-1}$ yields weights for pointwise interpolation, derivatives, volume averages, and conversions among these data types on a fixed stencil. On uniform grids with the squared exponential kernel, grid-convergence tests show the baseline interpolation converges at order $2r_{gp}+1$, first derivatives at about $2r_{gp}$, second derivatives at about $2r_{gp}-1$, and volume-average conversions at the interpolation order; the same rates hold in 2D on a 'blocky diamond' stencil selected by a new Taylor-based effective-order metric. For discontinuous data, the new stationary DAS kernel has a narrow dependence band and low condition numbers, and in shock-interpolation tests it produces no visible grid-scale oscillations where the SE and neural-network kernels do.

Load-bearing premise

The new DAS kernel in Eq. (40) must be a genuine symmetric positive definite covariance kernel; the paper uses it in the GP posterior equations without proving this property, and the printed formula is malformed, so the oscillation-free results depend on an unverified premise.

Editorial extensions

If this is right

  • GP weights on a ($2r_{gp}+1$)-point stencil give ($2r_{gp}+1$)th-order point-to-point interpolation, with one order lost per derivative and the interpolation order retained for point-to-volume-average conversions.
  • Because the prediction vector is precomputed and reused, the per-evaluation cost is a dot product of length $O(2r_{gp}+1)$, comparable to a finite-difference stencil.
  • The Taylor-based EOC metric lets a user select a stencil configuration in advance, without running multiple grid-convergence studies.
  • The DAS kernel reproduces discontinuous functions without grid-scale oscillations, unlike the squared exponential and neural-network kernels, while remaining second-order on smooth data.
  • The listed linear operators can be adopted directly in finite difference, finite volume, and reconstruction routines in place of polynomial stencils.

Reading between the lines

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

  • If the DAS kernel's positive definiteness is verified, a natural next step is a hybrid scheme that uses SE kernels in smooth regions and DAS near detected shocks; the paper mentions this as future work but does not test it.
  • The EOC metric could be applied to untested operators such as divergence, curl, or surface integrals, where stencil selection is currently guesswork.
  • The accuracy of the SE-kernel results appears sensitive to the length-scale hyperparameter $\ell$; a systematic data-driven choice of $\ell$ might change the observed convergence rates at fixed stencil size.
  • The paper itself notes in Appendix A that the SE-kernel posterior variance depends only on distance and is uninformative near shocks, so the formalism as presented does not yet give uncertainty-aware shock indicators.
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

5 major / 6 minor

Summary. The paper presents a Gaussian process regression framework, termed GP-Recipe, for approximating linear operations commonly used in numerical methods: pointwise interpolation, derivatives, volume averaging, and conversions between pointwise and volume-averaged data. For smooth problems the authors use the squared exponential (SE) kernel and derive, via the linearity of GPs, the corresponding prediction vectors z*^T = t*^T C^{-1} for each operation. They also introduce a Taylor-expansion-based 'effective order of convergence' (EOC) to select 2D stencil shapes, and they propose a new 'Discontinuous ArcSin' (DAS) kernel for oscillation-free interpolation across discontinuities. Numerical convergence tables in 1D and 2D report third-, fifth-, and seventh-order accuracy matching (2 r_gp + 1) for interpolation and the expected reduced orders for derivatives, with a second-order rate for the DAS kernel. The paper concludes that GP kernel weights can replace polynomial stencils in common numerical operations.

Significance. If the formulations are correct and reproducible, the paper would provide a useful, ready-to-use catalogue of kernel-based weights for finite difference, finite volume, and interpolation tasks, with a genuinely multidimensional extension and a kernel designed for shock capturing. The manuscript has clear strengths: it ships explicit formulas for all constrained kernels (Eqs. 12-38), it reports many grid-convergence tables with raw errors and EOC rates, and the DAS kernel's low condition numbers and narrow bandwidth are demonstrated in Figure 7. The EOC-based stencil-selection idea is a potentially valuable predictive tool. However, the present version contains internal inconsistencies that prevent the reader from reproducing the reported results, and some claims are not supported by the evidence shown. The contribution is significant if the issues below are resolved, but in its current form the central 'recipe' is not self-consistent.

major comments (5)
  1. [Section 6.1.3, Eqs. (24), (26), (28), (29)] The printed odd-derivative kernels have the wrong sign relative to the stated definition in Eq. (23) and to direct differentiation of the SE kernel in Eq. (12). For the first derivative, direct differentiation gives (d/dx)K_SE(x,x') = -(x-x')/ell^2 K_SE(x,x'), so at x = x* the kernel should be t_i = -(x* - x_i)/ell^2 exp(-(x*-x_i)^2/(2ell^2)). Eq. (24) instead prints t_i = + (x* - x_i)/ell^2 exp(...), which is the opposite sign. The same sign reversal appears in the third-derivative kernel Eq. (26) and in the 2D first-derivative kernels Eqs. (28)-(29). Because the prediction vector is linear in t*, a reader implementing the formulas exactly as written would obtain derivatives that are the negatives of the true GP posterior derivatives, and the clean convergence shown in Tables 5 and 7 could not be reproduced. This is a load-bearing inconsistency for a paper whose title promises a 'recipe'; the signs must be corrected or an explicit convention stated.
  2. [Section 4.2, Eq. (18)] The definition of the effective order of convergence is not consistent with its use in Table 1. Eq. (18) defines c_h as the Taylor coefficients of the target function f itself, c_h = (1/alpha!) partial^{|alpha|} f(x*)/partial x^alpha with |alpha| = h. As written, EOC depends only on f at the expansion point and is independent of the GP weights, the stencil, and the linear operation, so it cannot produce the varied EOC values reported for different stencils and different derivative operations in Table 1. For the EOC to be a property of the approximation scheme, c_h must be the coefficients of the error expansion that involve the kernel weights, e.g., c_h = sum_i w_i (x_i - x*)^h / h!, or an analogous moment of the prediction vector. The current definition is ill-posed and requires correction, since the stencil-selection analysis in Section 4.3 relies on it.
  3. [Section 7, Eq. (40)] The DAS kernel is not fully defined and its mathematical validity is not established. The printed formula in Eq. (40) has unbalanced parentheses: the denominator expression 'qQd i=1 (1 + 2(1 + (xi - x'i)^2)' is missing a closing parenthesis and the structure of the product is ambiguous. More importantly, the paper uses KDAS in the GP posterior equations (Eqs. (2) and (19)) without proving that KDAS is a symmetric positive definite covariance kernel, which is a prerequisite for K to be invertible and for the posterior mean to be meaningful. The claims of oscillation-free interpolation in Figs. 5 and 6 and the convergence results in Table 15 all depend on KDAS being a valid kernel. The authors should provide a correct expression, state the range of the argument of sin^{-1}, and either prove positive definiteness or cite a theorem from which it follows.
  4. [Table 7, Section 6.1.3] The convergence claim in Section 6.1.3 is contradicted by the reported data for the third derivative with r_gp = 1. Table 7 shows that the L1, L2, and L-infinity errors grow monotonically under refinement (e.g., L1 error increases from 2.0660e+03 at N=16 to 3.3232e+03 at N=256), and the corresponding EOCs are negative. The text states that 'these test results satisfy the anticipated and effective convergence rates in Table 1 very well as the order of the derivative varies,' which is not true for this case. This outlier either indicates a bug in the implementation or reveals a genuine limitation of the recipe for high-order derivatives on small stencils; it must be acknowledged, explained, and the affected claims qualified. The same table also shows EOC degradation at fine resolutions for r_gp = 2 (L1 EOC drops from 1.84 to 1.93 to 1.93, below the anticipated 3) and r_gp = 3 (EOC asymptotes near 3.9 instead of 5), so the blanket statement in Section 6.1.3 is not supported.
  5. [Abstract and Conclusion] The abstract and conclusion claim that 'GP approximated solutions feature improved solution accuracy compared to the conventional finite-difference counterparts,' but the manuscript contains no direct comparison against finite-difference stencils of the same order. All convergence tables compare the GP prediction to the exact solution only. The claim of improvement over finite differences is therefore unsupported by the reported evidence. Either add explicit numerical comparisons against standard finite-difference weights on the same grids and stencils, or remove/qualify the claim to refer only to the demonstrated convergence rates of the GP method itself.
minor comments (6)
  1. [Section 7, paragraph after Eq. (40)] The sentence 'First, we remark that the SE and DAS kernels are astationary kernel' contains a typo and an agreement error; it should read 'are stationary kernels.'
  2. [Eq. (40) and Section 7] The kernel name is spelled 'Discontinous ArcSin' in the text; the intended spelling is likely 'Discontinuous ArcSin'.
  3. [Table 13 caption] The caption reads 'Grid convergence results for the 2D volume-average-to-point conversion in Section 6.1.9,' but the section describes volume-average-to-derivative conversion for the cross derivative; the caption should be corrected to match the content.
  4. [Eq. (16)] The Taylor expansion in Eq. (16) uses both h and x_i - x* for the same quantity and the remainder is written as O(h^{alpha+1}), which is not a well-formed order symbol when alpha is the summation index; the notation should be cleaned up.
  5. [Eq. (26)] The third-derivative kernel mixes x and x* in the factor (x - x'): after evaluation at x = x* the factor should be (x* - x'), and the printed formula is inconsistent with Eq. (24) in this respect.
  6. [Reference [45]] Reference [45] is listed without a journal or conference venue; if the work is a preprint, this should be stated.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the GP weights, convergence rates, and DAS kernel behavior are computed from fixed kernels and measured against analytic test functions, not recycled from fitted parameters or self-citations.

full rationale

I walked the derivation chain from the GP posterior mean (Eq. 2) through the constrained kernels (Eqs. 19-38) and the DAS kernel (Eq. 40). The prediction vector z* = k*^T K^{-1} is determined entirely by the kernel and grid, independent of the data f, and no parameter appearing in the reported accuracy tables is fitted to those same error tables. The length scale ℓ is set a priori (e.g., ℓ = 0.05 in most tests), not optimized against the measured errors, so the convergence orders in Tables 3-15 are emergent results rather than fitted predictions. The EOC heuristic in Eq. (18) is computed from Taylor coefficients of the kernel weights and used as a stencil-selection diagnostic; it is not the measured grid-convergence rate, and the paper independently verifies the chosen stencil with grid-convergence studies, so no EOC result is being passed off as an independent prediction that was actually used as an input. The invariance of GPs under linear operators, which underlies the kernel construction, is proved in Appendix B, and the (2rgp+1)-order interpolation claim is cited to prior work [23,25,37] but is also demonstrated directly in Tables 3 and 4, so the self-citations are context and support, not load-bearing circularity. The DAS kernel's oscillation-free behavior is demonstrated on test profiles without fitting DAS hyperparameters to those profiles. The malformed Eq. (40) and the sign inconsistency in the printed derivative kernels (Eqs. 24, 26, 28, 29 relative to direct differentiation of Eq. 12) are serious reproducibility and correctness concerns, but they are not circularity: they do not reduce the paper's conclusions to its own inputs. I therefore find no circular step.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claims rest on standard GP theory plus two paper-specific assumptions: the DAS kernel is a valid covariance kernel, and the EOC metric as printed includes hidden weight moments. The length scale is a free parameter chosen by hand. No new physical entities are introduced; the DAS kernel is a mathematical function, not a postulated entity.

free parameters (3)
  • SE kernel length scale ℓ = 0.05 for most smooth tests; π/10 for Laplacian; ℓ=12Δ in Fig. 6
    Chosen by hand as a characteristic length scale; not optimized against the reported errors, but it controls the kernel matrix conditioning and the effective stencil weights.
  • NN kernel hyperparameters σ and σ0 = σ=σ0=1 (and σ=100 in Fig. 5)
    Used only for comparison in the discontinuous interpolation study; not part of the main method.
  • Taylor truncation order k = 12
    Chosen for the EOC computation in Section 4.2; the results may depend on this cutoff.
assumptions (5)
  • standard math Standard GP posterior formulas (Eq. 2) are valid
    Assumed from Rasmussen and Williams; the GP prior and conditioning are standard.
  • standard math SE kernel (Eq. 12) is symmetric positive definite and C∞
    Well-known property of the squared exponential kernel.
  • ad hoc to paper DAS kernel (Eq. 40) is a valid symmetric positive definite covariance kernel
    Used without proof in Section 7; validity is assumed and only condition numbers are shown.
  • ad hoc to paper The EOC in Eq. (18) actually includes kernel-weight moments in c_h
    As printed, Eq. (18) omits the moments and cannot produce stencil-dependent values in Table 1.
  • domain assumption The smooth test functions are analytic and the stencils are on uniform Cartesian grids
    All convergence studies use smooth functions on uniform grids; extension to general grids is not established.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GP-Recipe: Gaussian Process approximation to linear operations in numerical methods." pith.science (2026). https://pith.science/paper/64JHGUTB

@misc{pith2026250603471,
  author       = {Pith},
  title        = {Pith review of: GP-Recipe: Gaussian Process approximation to linear operations in numerical methods},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/64JHGUTB}},
  note         = {Machine review of arXiv:2506.03471}
}
read the original abstract

We introduce new Gaussian Process (GP) high-order approximations to linear operations that are frequently used in various numerical methods. Our method employs the kernel-based GP regression modeling, a non-parametric Bayesian approach to regression that operates on the probability distribution over all admissible functions that fit observed data. We begin in the first part with discrete data approximations to various linear operators applied to smooth data using the most popular squared exponential kernel function. In the second part, we discuss data interpolation across discontinuities with sharp gradients, for which we introduce a new GP kernel that fits discontinuous data without oscillations. The current study extends our previous GP work on polynomial-free shock-capturing methods in finite difference and finite volume methods to a suite of linear operator approximations on smooth data. The formulations introduced in this paper can be readily adopted in daily practices in numerical methods, including numerical approximations of finite differences, quadrature rules, interpolations, and reconstructions, which are most frequently used in numerical modeling in modern science and engineering applications. In the test problems, we demonstrate that the GP approximated solutions feature improved solution accuracy compared to the conventional finite-difference counterparts.

Figures

Figures reproduced from arXiv: 2506.03471 by the authors.

Figure 1
Figure 1. 1D GP stencil. The red “*” marks indicate the [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. 2D GP stencils of rgp = 2. The central cell xm = (xim , y jm ) is the cell where GP predictions are computed at the surrounding interface locations, (xim±1/2, y jm ) or (xim , y jm±1/2), or corner locations, (xim±1/2, y jm±1/2). To decide on an optimal 2D stencil, we examine the so-called effective order of convergence (or EOC) rates for various GP prediction operations on each configuration (more details to follow … view at source ↗
Figure 3
Figure 3. GP’s prediction of ∇ 2 sin(x) sin(y) on [0, 2π] × [0, 2π] using a 128 × 128 grid resolution with rgp = 2 and ℓ = π/10. The GP result shows a great visual agreement with the exact Laplacian calculation. The convergence study for this case is shown in [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Heat map comparison of kernels. (Top left) SE kernel with ℓ = 1.0. (Top right) SE kernel with ℓ = 0.35. (Bottom left) NN kernel with σ = σ0 = 1. (Bottom right) DAS kernel. i.e., 10 or more cells away from each other, the correlation of them rapidly drops with the incre…
Figure 5
Figure 5. Figure 5: We test the problem of interpolating across a single discontinuity using our GP point-to-point interpolation and compare di [PITH_FULL_IMAGE:figures/full_fig_p027_5.png]
Figure 6
Figure 6. Figure 6: Comparison of point-to-point interpolation on three di [PITH_FULL_IMAGE:figures/full_fig_p028_6.png]
Figure 7
Figure 7. Figure 7: Condition number of K plotted as a function of grid spacing ∆x. (Left) Comparison of the three kernels for various stencil sizes. The hyperparameters are held at ℓ = 1.8 for SE and σ = σ0 = 1 for NN, while rgp is varied for all cases. (Right) Comparison of the three ke…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 42 canonical work pages

  1. [1]

    Gaussian processes for regression

    Christopher Williams and Carl Rasmussen. Gaussian processes for regression. Advances in neural information processing systems, 8, 1995

    work page 1995

  2. [2]

    Rasmussen and C.K.I

    C.E. Rasmussen and C.K.I. Williams. Gaussian Processes for Machine Learning. Adaptive Computation And Machine Learning. MIT Press, 2005

    work page 2005

  3. [3]

    Gaussian processes in machine learning

    Carl Edward Rasmussen. Gaussian processes in machine learning. In Summer school on machine learning, pages 63–71. Springer, 2003

    2003

  4. [4]

    Kernel methods in machine learning

    Thomas Hofmann, Bernhard Sch ¨olkopf, and Alexander J Smola. Kernel methods in machine learning. 2008

    work page 2008

  5. [5]

    A tutorial on gaussian process regression: Modelling, exploring, and exploiting functions

    Eric Schulz, Maarten Speekenbrink, and Andreas Krause. A tutorial on gaussian process regression: Modelling, exploring, and exploiting functions. Journal of mathematical psychology, 85:1–16, 2018

    work page 2018

  6. [6]

    When gaussian process meets big data: A review of scalable gps.IEEE transactions on neural networks and learning systems, 31(11):4405–4423, 2020

    Haitao Liu, Yew-Soon Ong, Xiaobo Shen, and Jianfei Cai. When gaussian process meets big data: A review of scalable gps.IEEE transactions on neural networks and learning systems, 31(11):4405–4423, 2020

    2020

  7. [7]

    Fast gaussian process regression for big data

    Sourish Das, Sasanka Roy, and Rajiv Sambasivan. Fast gaussian process regression for big data. Big data research, 14:12–26, 2018

    work page 2018

  8. [8]

    Gaussian process behaviour in wide deep neural networks

    Alexander G de G Matthews, Mark Rowland, Jiri Hron, Richard E Turner, and Zoubin Ghahramani. Gaussian process behaviour in wide deep neural networks. arXiv preprint arXiv:1804.11271, 2018

    arXiv 2018

Show all 47 references
  1. [9]

    Deep convolutional networks as shallow gaussian processes.arXiv preprint arXiv:1808.05587, 2018

    Adri `a Garriga-Alonso, Carl Edward Rasmussen, and Laurence Aitchison. Deep convolutional networks as shallow gaussian processes.arXiv preprint arXiv:1808.05587, 2018

    2018 arXiv

  2. [10]

    Gaussian processes for time-series modelling

    Stephen Roberts, Michael Osborne, Mark Ebden, Steven Reece, Neale Gibson, and Suzanne Aigrain. Gaussian processes for time-series modelling. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences , 371(1984):20110550, 2013

    1984

  3. [11]

    Fast and scalable gaussian process modeling with applications to astronomical time series

    Daniel Foreman-Mackey, Eric Agol, Sivaram Ambikasaran, and Ruth Angus. Fast and scalable gaussian process modeling with applications to astronomical time series. The Astronomical Journal, 154(6):220, 2017

    2017

  4. [12]

    Time series forecasting with gaussian processes needs priors

    Giorgio Corani, Alessio Benavoli, and Marco Za ffalon. Time series forecasting with gaussian processes needs priors. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 103–117. Springer, 2021

    2021

  5. [13]

    Gaussian process for nonstationary time series prediction

    Sofiane Brahim-Belhouari and Amine Bermak. Gaussian process for nonstationary time series prediction. Computational Statistics & Data Analysis, 47(4):705–712, 2004

    2004

  6. [14]

    Gaussian processes for data-e fficient learning in robotics and control

    Marc Peter Deisenroth, Dieter Fox, and Carl Edward Rasmussen. Gaussian processes for data-e fficient learning in robotics and control. IEEE transactions on pattern analysis and machine intelligence, 37(2):408–423, 2013. 24

    2013

  7. [15]

    Gaussian processes in reinforcement learning

    Malte Kuss and Carl Rasmussen. Gaussian processes in reinforcement learning. Advances in neural information processing systems , 16, 2003

    2003

  8. [16]

    A gaussian process upsampling model for improvements in optical character recognition

    Steven I Reeves, Dongwook Lee, Anurag Singh, and Kunal Verma. A gaussian process upsampling model for improvements in optical character recognition. In Advances in Visual Computing: 15th International Symposium, ISVC 2020, San Diego, CA, USA, October 5–7, 2020, Proceedings, Pa...

    2020

  9. [17]

    Individualized gaussian process-based prediction and detection of local and global gray matter abnormalities in elderly subjects

    Gabriel Ziegler, Gerard R Ridgway, Robert Dahnke, Christian Gaser, Alzheimer’s Disease Neuroimaging Initiative, et al. Individualized gaussian process-based prediction and detection of local and global gray matter abnormalities in elderly subjects. NeuroImage, 97:333–348, 2014

    2014

  10. [18]

    Gaussian process regression stochastic volatility model for financial time series.IEEE Journal of Selected Topics in Signal Processing, 10(6):1015–1028, 2016

    Jianan Han, Xiao-Ping Zhang, and Fang Wang. Gaussian process regression stochastic volatility model for financial time series.IEEE Journal of Selected Topics in Signal Processing, 10(6):1015–1028, 2016

    2016

  11. [19]

    Price forecasts of ten steel products using gaussian process regressions

    Xiaojie Xu and Yun Zhang. Price forecasts of ten steel products using gaussian process regressions. Engineering Applications of Artificial Intelligence, 126:106870, 2023

    2023

  12. [20]

    Prediction with gaussian processes: From linear regression to linear prediction and beyond

    Christopher KI Williams. Prediction with gaussian processes: From linear regression to linear prediction and beyond. In Learning in graphical models, pages 599–621. Springer, 1998

    1998

  13. [21]

    New high-order methods using gaussian processes for computational fluid dynamics simulations

    Dongwook Lee, Adam Reyes, Carlo Graziani, and Petros Tzeferacos. New high-order methods using gaussian processes for computational fluid dynamics simulations. In Journal of Physics: Conference Series, volume 837, page 012018. IOP Publishing, 2017

    2017

  14. [22]

    E fficient implementation of essentially non-oscillatory shock-capturing schemes, II

    Chi-Wang Shu and Stanley Osher. E fficient implementation of essentially non-oscillatory shock-capturing schemes, II. Journal of Computa- tional Physics, 83(1):32–78, 1989

    1989

  15. [23]

    A new class of high-order methods for fluid dynamics simulations using gaussian process modeling: One-dimensional case

    Adam Reyes, Dongwook Lee, Carlo Graziani, and Petros Tzeferacos. A new class of high-order methods for fluid dynamics simulations using gaussian process modeling: One-dimensional case. Journal of Scientific Computing, 76(1):443–480, 2018

    2018

  16. [24]

    E fficient implementation of weighted ENO schemes

    Guang-Shan Jiang and Chi-Wang Shu. E fficient implementation of weighted ENO schemes. Journal of computational physics, 126(1):202– 228, 1996

    1996

  17. [25]

    A variable high-order shock-capturing finite di fference method with GP-WENO

    Adam Reyes, Dongwook Lee, Carlo Graziani, and Petros Tzeferacos. A variable high-order shock-capturing finite di fference method with GP-WENO. Journal of Computational Physics, 381:189–217, 2019

    2019

  18. [26]

    High-order genuinely multidimensional finite volume methods via kernel-based weno

    ICT May and D Lee. High-order genuinely multidimensional finite volume methods via kernel-based weno. InJournal of Physics: Conference Series, volume 2742, page 012025. IOP Publishing, 2024

    2024

  19. [27]

    Kfvm-weno: A high-order accurate kernel-based finite volume method for compressible hydrodynamics

    Ian CT May and Dongwook Lee. Kfvm-weno: A high-order accurate kernel-based finite volume method for compressible hydrodynamics. The Astrophysical Journal, 967(1):6, 2024

    2024

  20. [28]

    Balsara, Sudip Garain, and Chi Wang Shu

    Dinshaw S. Balsara, Sudip Garain, and Chi Wang Shu. An e fficient class of WENO schemes with adaptive order. Journal of Computational Physics, 326:780–804, 2016

    2016

  21. [29]

    Essentially non-oscillatory and weighted essentially non-oscillatory schemes for hyperbolic conservation laws

    Chi-Wang Shu. Essentially non-oscillatory and weighted essentially non-oscillatory schemes for hyperbolic conservation laws. In Advanced numerical approximation of nonlinear hyperbolic equations, pages 325–432. Springer, 1998

    1998

  22. [30]

    A survey of the isentropic Euler vortex problem using high-order methods

    Seth C Spiegel, HT Huynh, and James R DeBonis. A survey of the isentropic Euler vortex problem using high-order methods. In 22nd AIAA Computational Fluid Dynamics Conference, page 2444, 2015

    2015

  23. [31]

    A high-order finite volume method for systems of conservation laws—multi-dimensional optimal order detection (MOOD)

    St ´ephane Clain, Steven Diot, and Rapha¨el Loub`ere. A high-order finite volume method for systems of conservation laws—multi-dimensional optimal order detection (MOOD). Journal of computational Physics, 230(10):4028–4050, 2011

    2011

  24. [32]

    Improved detection criteria for the multi-dimensional optimal order detection (MOOD) on unstructured meshes with very high-order polynomials

    Steven Diot, St ´ephane Clain, and Rapha¨el Loub`ere. Improved detection criteria for the multi-dimensional optimal order detection (MOOD) on unstructured meshes with very high-order polynomials. Computers & Fluids, 64:43–63, 2012

    2012

  25. [33]

    The multidimensional optimal order detection method in the three-dimensional case: very high-order finite volume method for hyperbolic systems

    Steven Diot, Rapha ¨el Loub `ere, and Stephane Clain. The multidimensional optimal order detection method in the three-dimensional case: very high-order finite volume method for hyperbolic systems. International Journal for Numerical Methods in Fluids, 73(4):362–392, 2013

    2013

  26. [34]

    La m´ ethode MOOD Multi-dimensional Optimal Order Detection: la premi` ere approche a posteriori aux m´ ethodes volumes finis d’ordre tr` es ´ elev´ e

    Steven Diot. La m´ ethode MOOD Multi-dimensional Optimal Order Detection: la premi` ere approche a posteriori aux m´ ethodes volumes finis d’ordre tr` es ´ elev´ e. PhD thesis, Universit ´e de Toulouse, Universit´e Toulouse III-Paul Sabatier, 2012

    2012

  27. [35]

    A priori neural networks versus a posteriori mood loop: A high accurate 1d fv scheme testing bed

    Alexandre Bourriaud, Rapha ¨el Loub`ere, and Rodolphe Turpault. A priori neural networks versus a posteriori mood loop: A high accurate 1d fv scheme testing bed. Journal of Scientific Computing, 84(2):1–36, 2020

    2020

  28. [36]

    Gp-mood: a positivity-preserving high-order finite volume method for hyperbolic conservation laws

    Dongwook Lee and R ´emi Bourgeois. Gp-mood: a positivity-preserving high-order finite volume method for hyperbolic conservation laws. Proceedings of the International Astronomical Union, 16(S362):373–379, 2020

    2020

  29. [37]

    Gp-mood: A positivity-preserving high-order finite volume method for hyperbolic conservation laws

    R ´emi Bourgeois and Dongwook Lee. Gp-mood: A positivity-preserving high-order finite volume method for hyperbolic conservation laws. Journal of Computational Physics, 471:111603, 2022

    2022

  30. [38]

    An application of gaussian process modeling for high- order accurate adaptive mesh refinement prolongation

    Steve Reeves, Dongwook Lee, Adam Reyes, Carlo Graziani, and Petros Tzeferacos. An application of gaussian process modeling for high- order accurate adaptive mesh refinement prolongation. arXiv preprint arXiv:2003.08508; Accepted for publication in CAMCoS, 2021

    2003 arXiv

  31. [39]

    A high-order finite-volume method for conservation laws on locally refined grids

    Peter McCorquodale and Phillip Colella. A high-order finite-volume method for conservation laws on locally refined grids. Communications in Applied Mathematics and Computational Science, 6(1):1–25, 2011

    2011

  32. [40]

    Improved accuracy of high-order WENO finite volume methods on Cartesian grids

    Pawel Buchm ¨uller and Christiane Helzel. Improved accuracy of high-order WENO finite volume methods on Cartesian grids. Journal of Scientific Computing, 61(2):343–368, 2014

    2014

  33. [41]

    C. Bishop. Pattern recognition and machine learning (information science and statistics), 1st edn. 2006. corr. 2nd printing edn. Springer, New York, 2007

    2006

  34. [42]

    Interpolation of spatial data: some theory for kriging

    Michael L Stein. Interpolation of spatial data: some theory for kriging. Springer Science & Business Media, 2012

    2012

  35. [43]

    Kernel-based approximation methods using Matlab, volume 19

    Gregory E Fasshauer and Michael J McCourt. Kernel-based approximation methods using Matlab, volume 19. World Scientific Publishing Company, 2015

    2015

  36. [44]

    Randall J. LeVeque. Finite Difference Methods for Ordinary and Partial Differential Equations. Society for Industrial and Applied Mathe- matics, 2007

    2007

  37. [45]

    Abhishek Mishra, David Salac, and Matthew G. Knepley. On the order of accuracy for finite di fference approximations of partial differential equations using stencil composition, 2023

    2023

  38. [46]

    Le, Alex J

    Quoc V . Le, Alex J. Smola, and St ´ephane Canu. Heteroscedastic gaussian process regression. In Proceedings of the 22nd International Conference on Machine Learning, ICML ’05, page 489–496, New York, NY , USA, 2005. Association for Computing Machinery. 25

    2005

  39. [47]

    Most likely heteroscedastic gaussian process regression

    Kristian Kersting, Christian Plagemann, Patrick Pfa ff, and Wolfram Burgard. Most likely heteroscedastic gaussian process regression. In Proceedings of the 24th International Conference on Machine Learning, ICML ’07, page 393–400, New York, NY , USA, 2007. Association for Comp...

    2007

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.