Pith. sign in

REVIEW 3 major objections 4 minor 54 references

Optimized and regularly repeated lattice-based Latin hypercube designs for large-scale computer experiments

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Restricting Latin hypercube search to good lattice point sets, with new O(n) and O(log n) formulas for space-filling criteria, gives designs that beat numerically optimized LHDs for n ≥ 100 and one-inversion fitting of local Gaussian…

desk verdict Useful theory and new constructions for lattice-based LHDs; the empirical 'outperform' claims are stronger than the evidence because the OLHD baseline is unconverged. read the letter →

arxiv 2506.04582 v1 pith:KN5P7ZJ2 submitted 2025-06-05 stat.ME

classification stat.ME MSC 62K05
keywords Latinhypercubedesigngoodlatticepointsetspace-fillingwrap-arounddiscrepancyseparationdistanceGaussianprocessemulationlarge-scalecomputerexperimentsphysics-informedneuralnetworks
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

Computer experiments with tens of thousands of runs break both the numerical search for good designs and the matrix inversions needed to fit Gaussian process emulators. This paper argues that both bottlenecks can be removed by working inside the restricted class of good lattice point sets: the lattice structure yields shortcut formulas that evaluate standard space-filling criteria in $O(n)$ or $O(\log n)$ operations instead of $O(n^2)$, making it cheap to search many generators. The resulting optimized lattice-based Latin hypercube designs match or beat numerically optimized designs once the sample size $n$ is at least 100. A second construction, the regularly repeated lattice-based LHD, tiles the input space with translated copies of one local LHD, so a moving-window emulator fits every local Gaussian process with a single matrix inversion. This would make large-scale design and emulation substantially cheaper without sacrificing accuracy.

What carries the argument

The central object is the good-lattice-point (GLP) Latin hypercube, defined by $$L(n,v,\delta)=\{z+iv/n+\delta/n+\mathbf{1}_d/(2n): z\in\mathbb{Z}^d,\ i\in\mathbb{Z}\}+[0,1]^d,$$ where every entry of $v$ is coprime to $n$ and $w(\cdot)$ is the wrap-around distance on the torus; it is an LHD because reducing $iv/n$ modulo 1 permutes the $n$ levels in each coordinate. The load-bearing identities are Theorem 3.1's reduction of the wrap-around criteria to sums over the $n-1$ lattice displacements, for example $$c_{\mathrm{WS}}(L)=\min_{i=1}^{n-1}\Bigl(\sum_{k=1}^d w(iv_k/n)^2\Bigr)^{-1/2},$$ and Algorithm 4's use of Gaussian lattice-basis reduction to get the two-dimensional separation distance in $O(\log n)$ steps. The companion object for emulation is the regularly repeated design $$R(n,m,v,\delta)=\{zm/n+iv/n+\delta/n+\mathbf{1}_d/(2n)\}+[0,1]^d,$$ whose Theorem 3.5 and Theorem 3.6 structure makes every local $m$-point subdesign a translate of the same LHD. That translation invariance is what converts many local Gaussian process fits into one shared matrix inverse.

What would settle it

Take a deliberately non-stationary test function, for instance one with a sharp ridge in one corner and smooth behavior elsewhere, and compare an RLHD emulator that reuses a single shared correlation inverse against local Gaussian processes whose correlation parameters and inverses are re-estimated in each neighborhood; if the shared-inverse predictions are materially less accurate, the stationarity premise behind the one-inversion trick has failed.

Watch

Extended reading notes

Core claim

The paper's central claim is that the lattice constraint is not a restriction to be tolerated but the mechanism that makes large-scale design and estimation feasible. For any good-lattice-point Latin hypercube $L(n,v,\delta)$ with generator $v$, Theorem 3.1 expresses the wrap-around separation, wrap-around $\phi_p$, wrap-around projection, and wrap-around discrepancy as functions of the $n-1$ displacements $iv/n$ only, reducing evaluation from $O(n^2d)$ to $O(nd)$; Theorem 3.2 and Algorithm 4 further reduce the bivariate separation and fill-distance criteria to $O(\log n)$ via shortest-vector computation. The criteria are shift-invariant, and the paper reports that they are positively correlated on lattice designs, so Algorithm 3 can optimize the generator alone and the resulting LLHDs have lower criterion values than optimized unrestricted LHDs for $n\ge 100$ in the reported comparisons. The paper's second claim is that the design $R(n,m,v,\delta)=\{zm/n+iv/n+\delta/n+\mathbf{1}_d/(2n)\}$ contains an $m$-point LHD in every hypercube of width $m/n$ (Theorem 3.5), and that the local designs are exact translations whenever the hypercube origins differ by a lattice vector (Theorem 3.6). Consequently Algorithm 5 fits all local Gaussian process models with one inversion of a single $m\times m$ correlation matrix, and the reported emulation comparisons show lower prediction error than partitioned, nearest-neighbor, local approximate, or sparse-grid alternatives.

Load-bearing premise

The single-inversion speedup for emulation rests on the assumption that the correlation function describing the simulation output is stationary, meaning the statistical relationship between nearby runs is the same everywhere in the input space; if the output behaves very differently in different regions, the translated local designs no longer share a common correlation matrix and Algorithm 5's cost saving evaporates.

Editorial extensions

If this is right

  • For sample sizes of 100 and above, restricting the search to good lattice point sets gives designs with lower wrap-around discrepancy, projective separation, and $\phi_p$ values than the same number of iterations of unrestricted LHD optimization, and the gap widens as $n$ grows.
  • Because the wrap-around criteria are positively correlated on lattice designs, optimizing a single cheap criterion such as the bivariate separation distance also improves the other space-filling properties, so practitioners can use the cheapest criterion without sacrificing overall quality.
  • The optimized lattice designs reduce mean absolute integration error in uncertainty quantification and lower the training error of physics-informed neural networks compared with random LHDs, optimized LHDs, and power-generator lattice designs on the tested PDEs.
  • For large-scale emulation, RLHDs make the cost of fitting local Gaussian process models at many prediction sites essentially the cost of one $m$-point model, while achieving lower prediction error than partitioned, nearest-neighbor, local approximate, and sparse-grid emulators in the reported test functions.
  • Every lattice-based LHD with composite $n$ is automatically a sliced LHD whose slices have identical space-filling properties, so sliced designs for multi-fidelity or sequential experiments come for free from the same construction.

Reading between the lines

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

  • The paper's final remarks only gesture at other lattice families, such as densest-packing, interleaved, and sphere-packing lattices; if the translation argument carries over to those families, the RLHD machinery should extend well beyond GLP generators, and a concrete check is whether the positive correlation among criteria persists there.
  • The reported correlations suggest a cheap-proxy strategy: optimize WS2 and inherit good WD and WP behavior. The paper's own tables show this proxy weakens at $n=1000$, $d=10$, so a useful extension is to map exactly where the proxy stops being reliable as dimension grows.
  • The integration-accuracy gains plausibly extend beyond the two PINN examples to any training objective that approximates an integral by a weighted sum over design points, since low wrap-around discrepancy is exactly what such sums need; this is an inference, not a claim the paper tests.
  • Because design evaluation no longer dominates the computation for very large $n$, the practical bottleneck shifts to materializing the $O(nd)$ design points and to nearest-lattice-point queries during prediction, so memory layout and spatial indexing become the next scalability question.
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

3 major / 4 minor

Summary. The manuscript proposes two families of lattice-based Latin hypercube designs: optimized LLHDs and regularly repeated RLHDs. For the former, the authors derive shortcut formulas (Theorems 3.1–3.3) that reduce the per-evaluation cost of several wrap-around space-filling criteria from O(n^2 d) to O(nd) or O(log(n)d^2), and they embed these formulas in a neighborhood search (Algorithm 3). For the latter, they show that RLHDs contain many translated m-point local LHDs, so that a moving-window Gaussian process predictor can share one correlation-matrix inverse (Theorems 3.5–3.7, Algorithm 5). The numerical sections compare LLHDs with OLHDs, PLHDs, and other designs for space-filling criteria, uncertainty quantification, PINN training, and large-scale emulation. Proofs are provided in a supplementary file, and R code is supplied.

Significance. The theoretical contribution is real and useful: if the formulas are correct, they make it feasible to search much larger classes of lattice LHDs than before, and the RLHD translation property is an elegant way to reduce local GP fitting costs under stationarity. The supplementary proofs and the accompanying R package are strengths. The empirical part, however, does not yet establish the broad claim that the designs 'consistently outperform others for n>=100', because the OLHD baseline is not converged and because almost no replication uncertainty is reported. These issues are fixable with additional experiments and more careful wording; they do not invalidate the theorems.

major comments (3)
  1. [Section 4, Figures 2-4] The main empirical support for the abstract's claim that LLHDs consistently outperform others for n>=100 is a comparison with OLHD after exactly 10^5 iterations of Algorithm 2. Figure 2 shows that the OLHD algorithm has not converged within 10^5 iterations in any tested case. Since every LLHD is an LHD (Lemma 2.1), a globally optimized OLHD can never be worse than the best LLHD on the criterion being optimized; the observed LLHD advantage therefore reflects the incompleteness of the OLHD optimization, not an intrinsic superiority of the lattice restriction. The same limited-budget OLHD-WD design is carried into the uncertainty-quantification comparison in Figure 4. Please either run the OLHD baseline to a much larger budget or to an explicit convergence criterion using a state-of-the-art LHD optimizer, or explicitly rephrase the conclusions as valid only for a fixed small optimization budget. This is the load-bearing point for the headline empirical claim.
  2. [Section 4, Figures 4-5 and Table 3] The applied comparisons report only point estimates, with no replication count, standard error, or statistical test. In Figure 5 the PINN training is stochastic, and in Table 3 the test points are randomly generated; differences such as 0.478 versus 0.510 for Emichalewicz, or several close curves in Figure 4, may be within noise. Please report means and standard errors (or boxplots) over at least 10 independent replications and state the number of replications in the captions. Without this, the statements that LLHD-WD 'clearly indicates' superiority and that RLHD 'achieves the smallest RMSE' are not supported.
  3. [Sections 3.1 and 4] Algorithm 3 is a heuristic neighborhood search, and the recommendation to use Q = max(ceil(T/(5p(n)d)), 1) rests entirely on the statement in Section 3.1 that '5p(n)d iterations seem to be enough,' with no formal convergence guarantee and no diagnostic table. Yet Section 4 repeatedly refers to 'the optimal LLHD' and uses that term to conclude that PLHDs are inferior. Unless the supplementary material provides evidence that the algorithm certifies local (or global) optimality for the tested cases, please replace 'optimal' by 'optimized' throughout and support the 5p(n)d heuristic with a small convergence study.
minor comments (4)
  1. [Theorem 3.3 and Algorithm 4] Theorem 3.3 refers to 'the a and b are computed from the end of Algorithm 4,' but Algorithm 4 as written returns only ||a||; please return a and b or state the quantities explicitly in the theorem.
  2. [Algorithm 5] Line 2 of Algorithm 5 contains the typo 'corvariance'; it should be 'covariance'.
  3. [Throughout] The text contains many OCR or encoding artifacts, such as '±' used for assignment, '2' for the minus sign, 'f' for less-than-or-equal-to, and garbled mathematical expressions (e.g., Algorithm 1). The final typeset version should be carefully proofread.
  4. [Section 3.3] The stationarity assumption behind the shared-inverse trick in Algorithm 5 is stated in Section 3.3, but the abstract and Section 4 could briefly note that the RLHD speedup applies to stationary correlation functions; a short discussion of behavior when the output is nonstationary would help readers judge the method's scope.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: criteria are external benchmarks, theorems are direct derivations, and the RLHD single-inversion claim explicitly assumes stationarity.

full rationale

The main derivations are self-contained. Theorems 3.1-3.3 evaluate wrap-around criteria (WS, WA, WP, WD, WS2/WF2) for lattice-based LHDs by substituting the lattice parametrization L(n,v,delta) from (1) into the externally defined criterion formulas; the supplementary proofs perform direct algebraic reductions, using Lemma 2.1 (Fang et al. 2018) and standard lattice-basis facts (Bremner 2011), with no free parameters fitted to the conclusions. Theorem 3.2's equivalence cWS=cRS for two-dimensional LLHDs is proven by an explicit lattice-geometry argument, not imported. Theorems 3.4-3.7 establish slicedness, translation equivalence, and expected size by set-theoretic identities from the definitions of R(n,m,v,delta) and L(m,v,cdot); Theorem 3.6 is a direct "if and only if" statement about lattice translations. Algorithm 5's claim that all local GPs share one correlation-matrix inversion is exactly the translation property of Theorem 3.6 applied under the paper's explicit stationarity assumption ("Since our method makes no assumption about the correlation kernel other than stationarity"); stationarity is an assumption, not a conclusion derived from the design. The numerical sections compare designs on externally fixed test functions (Prpeak, G, Borehole, Oscil, Burgers, wave, Ackley, Shekel, Emichalewicz) and external benchmarks (OLHD, PLHD, laGP, SG); these are empirical evaluations rather than predictions whose constants were set by the method. The self-citations to He (2017a,b, 2019a,b, 2021) appear in final-remarks suggestions for future lattice constructions and are not load-bearing for any theorem or numerical claim. Concerns that the OLHD baseline may be under-optimized after 10^5 iterations, or that non-stationary outputs would break the single-inversion trick, are legitimate empirical or assumption limitations but not circularity: the paper does not define its criteria in terms of its conclusions, fit parameters to target results, or import an unverified self-cited theorem as the basis of the derivation.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The paper's derivations rest on standard number-theoretic facts and lattice reduction; the domain assumptions are the choice of wrap-around criteria, stationarity, and the effect hierarchy principle. No hidden fitted parameters enter the theorems; the only hand-set quantity is the iteration multiplier 5 in the convergence recommendation.

free parameters (1)
  • iteration multiplier 5 = 5
    The recommendation that 5p(n)d iterations of Algorithm 3 suffice for convergence is a hand-set heuristic based on limited numerical experiments (Section 3.1). The efficiency claim depends on this heuristic holding for the chosen n and d.
assumptions (5)
  • domain assumption Wrap-around distance based criteria (WD, WS, WA, WP) are appropriate for constructing statistically-uniform designs and are nearly optimal under the standard CD, RS, AS, PS criteria.
    Adopted from Vorechovsky and Elias (2020); the paper restricts its optimization to these criteria based on this prior result, which is a domain belief rather than derived here.
  • domain assumption The correlation function of the simulation output is stationary, so all translated local designs have identical correlation matrices.
    Explicitly stated in Section 3.3: 'our method makes no assumption about the correlation kernel other than stationarity.' Algorithm 5 relies on a single matrix inversion.
  • domain assumption Lower-order interactions are more important than higher-order interactions (effect hierarchy principle), justifying WS2 and WF2 criteria.
    Cites Shi and Xu (2024); used to motivate bivariate projection criteria in Section 3.1.
  • standard math Lattice basis reduction (Gaussian algorithm) correctly computes the shortest vector in a 2D lattice in O(log n) steps.
    Used in Algorithm 4 and Theorem 3.3; standard result from Bremner (2011).
  • ad hoc to paper The heuristic that 5p(n)d iterations suffice for convergence of Algorithm 3.
    Empirical rule stated in Section 3.1 and used to set Q; no proof is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimized and regularly repeated lattice-based Latin hypercube designs for large-scale computer experiments." pith.science (2026). https://pith.science/paper/KN5P7ZJ2

@misc{pith2026250604582,
  author       = {Pith},
  title        = {Pith review of: Optimized and regularly repeated lattice-based Latin hypercube designs for large-scale computer experiments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KN5P7ZJ2}},
  note         = {Machine review of arXiv:2506.04582}
}
read the original abstract

Computer simulations serve as powerful tools for scientists and engineers to gain insights into complex systems. Less costly than physical experiments, computer experiments sometimes involve large number of trials. Conventional design optimization and model fitting methods for computer experiments are inefficient for large-scale problems. In this paper, we propose new methods to optimize good lattice point sets, using less computation to construct designs with enhanced space-filling properties such as high separation distance, low discrepancy, and high separation distance on projections. These designs show promising performance in uncertainty quantification as well as physics-informed neural networks. We also propose a new type of space-filling design called regularly repeated lattice-based Latin hypercube designs, which contain lots of local space-filling Latin hypercube designs as subdesigns. Such designs facilitate rapid fitting of multiple local Gaussian process models in a moving window type of modeling approach and thus are useful for large-scale emulation problems.

Figures

Figures reproduced from arXiv: 2506.04582 by the authors.

Figure 2
Figure 2. ). Consequently, it requires a total of [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 1
Figure 1. An RLHD in two dimensions, showing the designs points (circle [PITH_FULL_IMAGE:figures/full_fig_p017_1.png] view at source ↗
Figure 2
Figure 2. The WD (top), WP (middle), and WA (bottom) values versus [PITH_FULL_IMAGE:figures/full_fig_p018_2.png] view at source ↗
Figures from the paper (3 more)
Figure 3
Figure 3. Figure 3: The WD (top), WP (middle), and WA (bottom) values versus [PITH_FULL_IMAGE:figures/full_fig_p020_3.png]
Figure 4
Figure 4. Figure 4: Mean absolute integration errors versus n for the 3-dimensional Prpeak function (topleft), the 5-dimensional G-function (topright), the 8-dimensional Borehole function (bottomleft), and the 12-dimensional Oscil function (bottomright). 23 [PITH_FULL_IMAGE:figures/full_…
Figure 5
Figure 5. Figure 5: Mean integrated relative error versus n for the Burgers’ equation (left) and the wave equation (right). This approach requires O(m32 d ) = O(nm2 ) operations, which is smaller than O(n 3 ) but larger than O(m3 ) that is required for RLHD. In the nearest neighbor approa…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 54 canonical work pages

  1. [1]

    Ba, S. (2015). SLHD: Maximin-distance (sliced) L atin hypercube designs . R package version 2.1-1

  2. [2]

    A., and Myers, W

    Ba, S., Brenneman, W. A., and Myers, W. R. (2015). Optimal sliced L atin hypercube designs. Technometrics , 57(4):479--487

  3. [3]

    and Joseph, V

    Ba, S. and Joseph, V. R. (2015). MaxPro: Maximum projection designs . R package version 3.1-2

  4. [4]

    Bremner, M. R. (2011). Lattice Basis Reduction: An Introduction to the LLL Algorithm and Its Applications . Boca Raton: CRC Press

  5. [5]

    Chen, H., Ding, L., and Tuo, R. (2022). Kernel packet: An exact and scalable algorithm for G aussian process regression with M at\'ern correlations. Journal of Machine Learning Research , 23(127):1--32

  6. [6]

    and Johannesson, G

    Cressie, N. and Johannesson, G. (2008). Efficient emulators of computer experiments using compactly supported correlation functions, with an application to cosmology. Journal of the Royal Statistical Society, Series B , 70:209–226

  7. [7]

    Eliáš, J., Vořechovský, M., and Sadílek, V. (2020). Periodic version of the minimax distance criterion for M onte C arlo integration. Advances in Engineering Software , 149:102900

  8. [8]

    F., Liu, M

    Fang, K. F., Liu, M. Q., Qin, H., and Zhou, Y. D. (2018). Theory and Application of Uniform Experimental Designs . Singapore: Springer

Show all 54 references
  1. [9]

    F., Maringer, D., Tang, Y., and Winker, P

    Fang, K. F., Maringer, D., Tang, Y., and Winker, P. (2006). Lower bounds and stochastic optimization algorithms for uniform designs with three or four levels. Mathematics of Computation , 75(254):859--878

  2. [10]

    Franco, J., Dupuy, D., Roustant, O., Damblin, G., and Iooss., B. (2014). DiceDesign: Designs of Computer Experiments . R package version 1.6

  3. [11]

    and Apley, D

    Gramacy, R. and Apley, D. (2015). Local G aussian process approximation for large computer experiments. Journal of Computational and Graphical Statistics , 24:561--578

  4. [12]

    and Sun, F

    Gramacy, R. and Sun, F. (2023). laGP: Local Approximate G aussian Process Regression . R package version 1.5-0

  5. [13]

    Gramacy, R. B. (2020). Surrogates: G aussian Process Modeling, Design, and Optimization for the Applied Sciences . New York: Chapman and Hall/CRC, 1st edition

  6. [14]

    Gramacy, R. B. and Lee, H. K. H. (2008). Bayesian treed G aussian process models with an application to computer modeling. Journal of the American Statistical Association , 103(483):1119--1130

  7. [15]

    Haaland, B., Wang, W., and Maheshwari, V. (2018). A framework for controlling sources of inaccuracy in G aussian process emulation of deterministic computer experiments. SIAM/ASA Journal on Uncertainty Quantification , 6(2):497--521

  8. [16]

    He, X. (2017a). Interleaved lattice-based minimax distance designs. Biometrika , 104(3):713--725

  9. [17]

    He, X. (2017b). Rotated sphere packing designs. Journal of the American Statistical Association , 112(520):1612--1622

  10. [18]

    He, X. (2019a). Interleaved lattice-based maximin distance designs. Biometrika , 106(2):453--464

  11. [19]

    He, X. (2019b). Sliced rotated sphere packing designs. Technometrics , 61(1):66--76

  12. [20]

    He, X. (2021). Lattice-based designs possessing quasi-optimal separation distance on all projections. Biometrika , 108(2):443–454

  13. [21]

    and Tang, B

    He, Y. and Tang, B. (2013). Strong orthogonal arrays and associated L atin hypercubes for computer experiments. Biometrika , 100:254--260

  14. [22]

    Hickernell, F. J. (1998). Lattice Rules: How Well Do They Measure Up? , page 109–166. Springer, 1st edition

  15. [23]

    Hua, L. K. and Wang, Y. (1981). Applications of Number Theory to Numerical Analysis . Berlin: Springer

  16. [24]

    C., Chen, W., and Sudjianto, A

    Jin, R. C., Chen, W., and Sudjianto, A. (2005). An efficient algorithm for constructing optimal design of computer experiments. Journal of Statistical Planning and Inference , 134(1):268--287

  17. [25]

    E., Moore, L

    Johnson, M. E., Moore, L. M., and Ylvisaker, D. (1990). Minimax and maximin distance designs. Journal of Statistical Planning and Inference , 26:131--148

  18. [26]

    G., Bingham, D., Habib, S., Heitmann, K., and Frieman, J

    Kaufman, C. G., Bingham, D., Habib, S., Heitmann, K., and Frieman, J. A. (2011). Efficient emulators of computer experiments using compactly supported correlation functions, with an application to cosmology. Annals of Applied Statistics , 5(4):2470--2492

  19. [27]

    K., and Holmes, C

    Kim, H.-M., Mallick, B. K., and Holmes, C. C. (2005). Analyzing nonstationary spatial data using piecewise G aussian processes. Journal of the American Statistical Association , 100(470):653--668

  20. [28]

    Korobov, N. M. (1959). The approximate computation of multiple integrals. Doklady Akademii Nauk SSSR , 124:1207--1210

  21. [29]

    and Yaguchi, T

    Matsubara, T. and Yaguchi, T. (2023). Good lattice accelerates physics-informed neural networks. In 1st Workshop on the Synergy of Scientific and Machine Learning Modeling @ ICML2023

  22. [30]

    D., Beckman, R

    McKay, M. D., Beckman, R. J., and Conover, W. J. (1979). A comparison of three methods for selecting values of input variables in the analysis of output from a computer code. Technometrics , 21:239--45

  23. [31]

    Morris, M. D. and Mitchell, T. J. (1995). Exploratory designs for computational experiments. Journal of Statistical Planning and Inference , 43(3):381--402

  24. [32]

    Mullen, K. M. (2014). Continuous global optimization in R . Journal of Statistical Software , 060(1):1--45

  25. [33]

    Z., and Ding, Y

    Park, C., Huang, J. Z., and Ding, Y. (2011). Domain decomposition approach for fast G aussian process regression of large spatial data sets. Journal of Machine Learning Research , 12(47):1697--1728

  26. [34]

    Plumlee, M. (2014). Fast prediction of deterministic functions using sparse grid experimental designs. Journal of Computational and Graphical Statistics , 109:1581--1591

  27. [35]

    Qian, P. Z. G. (2012). Sliced L atin hypercube designs. Journal of the American Statistical Association , 107(497):393--399

  28. [36]

    Raissi, M., Perdikaris, P., and Karniadakis, G. E. (2019). Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational physics , 378:686--707

  29. [37]

    J., Mitchell, T

    Sacks, J., Welch, W. J., Mitchell, T. J., and Wynn, H. (1989). Design and analysis of computer experiments. Statistical Science , 4:409--423

  30. [38]

    J., Williams, B

    Santner, T. J., Williams, B. J., and Notz, W. I. (2018). The Design and Analysis of Computer Experiments . New York: Springer, 2nd edition

  31. [39]

    and Xu, H

    Shi, C. and Xu, H. (2024). A Projection Space-Filling Criterion and Related Optimality Results . Journal of the American Statistical Association , 119(548):2658--2669

  32. [40]

    Sloan, I. H. and Joe, S. (1994). Lattice methods for multiple integration . Oxford: Oxford University Press

  33. [41]

    Sun, F., Wang, Y., and Xu, H. (2019). Uniform projection designs. Annals of Statistics , 47(1):641--661

  34. [42]

    and Bingham, D

    Surjanovic, S. and Bingham, D. (2013). Virtual library of simulation experiments: Test functions and datasets

  35. [43]

    Tang, B. (1993). Orthogonal array-based Latin hypercubes. Journal of the American Statistical Association , 88(424):1392--1397

  36. [44]

    and Wang, W

    Tuo, R. and Wang, W. (2020). K riging prediction with isotropic M atern correlations: Robustness and experimental designs. Journal of Machine Learning Research , 21(187):1--38

  37. [45]

    and Xu, H

    Vazquez, A. and Xu, H. (2024). An integer programming algorithm for constructing maximin distance designs from good lattice point sets. Statistica Sinica , 34:1347--1366

  38. [46]

    and Eliáš, J

    Vořechovský, M. and Eliáš, J. (2020). Modification of the maximin and _p (phi) criteria to achieve statistically uniform distribution of sampling points. Technometrics , 62(3):371--386

  39. [47]

    Wang, L., Xiao, Q., and Xu, H. (2018). Optimal maximin L_1 -distance L atin hypercube designs based on good lattice point designs. Annals of Statistics , 46(6B):3741--3766

  40. [48]

    and Haaland, B

    Wang, W. and Haaland, B. (2019). Controlling sources of inaccuracy in stochastic K riging. Technometrics , 61(3):309--321

  41. [49]

    Wang, Y., Sun, F., and Xu, H. (2022). On design orthogonality, maximin distance, and projection uniformity for computer experiments. Journal of the American Statistical Association , 117(537):375--385

  42. [50]

    Wu, C., Zhu, M., Tan, Q., Kartha, Y., and Lu, L. (2023). A comprehensive study of non-adaptive and residual-based adaptive sampling for physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering , 403:115671

  43. [51]

    and Xu, H

    Xiao, Q. and Xu, H. (2017). Construction of maximin distance L atin squares and related L atin hypercube designs . Biometrika , 104(2):455--464

  44. [52]

    and Xu, H

    Xiao, Q. and Xu, H. (2018). Construction of maximin distance designs via level permutation and expansion . Statistica Sinica , 28:1395--1414

  45. [53]

    Zhou, W., Yang, J., and Liu, M. (2020). Optimal maximin L_2 -distance L atin hypercube designs. Journal of Statistical Planning and Inference , 207:113--122

  46. [54]

    and Xu, H

    Zhou, Y. and Xu, H. (2015). Space-filling properties of good lattice point sets . Biometrika , 102(4):959--966

Pith tools

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