Pith. sign in

REVIEW 5 major objections 5 minor 56 references

Neural Low-Discrepancy Sequences

T0 review · 5 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A neural network that maps indices to points can generate low-discrepancy sequences that beat classical constructions in dimension 4.

desk verdict A solid empirical step toward learned low-discrepancy sequences, but the 'outperforms all previous LDS constructions' claim is wider than the evidence. read the letter →

arxiv 2510.03745 v2 pith:VUDTD5BX submitted 2025-10-04 cs.LG cs.NAmath.NA

Michael Etienne Van Huffel , Nathan Kirk , Makram Chahine , Daniela Rus , T. Konstantin Rusch This is my paper
classification cs.LGcs.NAmath.NA MSC 65D3065C0568T07
keywords low-discrepancysequencesquasi-MonteCarlodiscrepancyminimizationneuralsequencegenerationprefixSobolsinusoidalpositionalencodingscientificmachinelearning
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 claims that low-discrepancy sequences — ordered point sets in the unit cube whose every initial segment is evenly spread — can be produced by a neural network rather than by number theory. The authors introduce NeuroLDS, a deterministic index-to-point network pre-trained to imitate a Sobol sequence and then fine-tuned to minimize a discrepancy loss summed over all prefixes. In dimension 4, the resulting finite sequences achieve lower L2 star, symmetric, and centered discrepancy than Sobol, Halton, and scrambled Sobol at every tested length. The authors further show the sequences improve quasi-Monte Carlo integration of an 8D borehole function, RRT robot motion planning through narrow passages, and neural surrogates for a Black-Scholes PDE. If true, this opens a machine-learning route to constructing sequences, not just fixed point sets.

What carries the argument

The load-bearing object is a feedforward network with sinusoidal positional encoding of the index. Each index i is mapped to a vector of sines and cosines at K frequency scales, echoing the digit expansions of Halton and Sobol; an L-layer MLP with ReLU and a final sigmoid turns this encoding into a point in [0,1]^d. Pretraining on Sobol stabilizes the map, and fine-tuning on a prefix-averaged discrepancy loss — with interchangeable kernels for star, symmetric, centered, extreme, periodic, or average-squared discrepancy — drives the learned sequence toward uniformity at every prefix.

What would settle it

Train the same architecture from random initialization with a much larger optimization budget and more expressive network, and check whether prefix discrepancy reaches Sobol-level or NeuroLDS-level values; if it never does, the claim that Sobol pretraining is essential is confirmed. Conversely, pretraining on Halton instead of Sobol and measuring whether the fine-tuned sequence still beats Sobol would show whether the gain comes from the neural refinement or from the specific Sobol starting point.

Watch

Extended reading notes

Core claim

The central claim is that a neural network can learn an index-to-point map fθ: {1,...,N} → [0,1]^d whose finite output sequence has lower discrepancy than classical low-discrepancy sequences, with the property holding for every prefix. The construction is two-stage: first regress onto a classical Sobol sequence (with a burn-in), then fine-tune by minimizing L_disc(θ) = Σ_{P=2}^N w_P · D_2^•({X_i}_{i=1}^P)^2, a weighted sum of closed-form L2 discrepancies over all prefixes. The paper reports that NeuroLDS outperforms Sobol, Halton, and scrambled Sobol across all tested prefix lengths in d=4 for the star, symmetric, and centered discrepancies, and demonstrates downstream gains in QMC integrati

Load-bearing premise

The method's success hinges on the premise that pretraining on a classical Sobol sequence supplies an essential inductive bias; if no classical construction is available to seed the network, direct discrepancy minimization collapses to degenerate corner-clustered points.

Editorial extensions

If this is right

  • Every prefix of a NeuroLDS sequence has lower L2 discrepancy than the corresponding Sobol, Halton, or scrambled Sobol prefix in the tested 4D cases, so users no longer need to choose N as a power of two to get good uniformity.
  • Weighted discrepancy losses let the user emphasize important coordinates, which the paper demonstrates on an 8D borehole integral where NeuroLDS is best or second-best for nearly all N.
  • Because the sequence is generated on the fly from an index, it can be extended adaptively in a planner like RRT; NeuroLDS gives higher success rates than Sobol, Halton, or uniform sampling on a narrow-passage 4D kinematic chain.
  • Training a neural PDE surrogate on NeuroLDS points lowers mean-squared price-prediction error compared with uniform random points or Sobol points for a 2D Black-Scholes basket option.
  • The framework is not tied to a fixed N: the same trained model produces any prefix, addressing a limitation of prior neural point-set generators.

Reading between the lines

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

  • NeuroLDS is best read as a refinement of Sobol rather than a replacement: since direct discrepancy minimization collapses and pretraining on Sobol is described as essential, the learned sequences inherit Sobol's structural strengths and weaknesses, including its coordinate-correlation issues in higher dimensions.
  • A natural stress test is to pretrain on Halton or a scrambled Sobol target; if discrepancy gains persist, the framework is more general than the paper's Sobol-specific story suggests.
  • The same prefix-loss scheme could be applied to non-uniform targets via Stein or other kernel discrepancies, which the authors mention as future flexibility; that would turn the method into an extensible sampler for arbitrary distributions.
  • Because the sinusoidal encoding is exactly a Fourier feature embedding, the method resembles learned positional encodings in sequence models, suggesting that low-discrepancy structure could be learned end-to-end inside such models rather than bolted on as a fixed sampling scheme.
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 / 5 minor

Summary. The paper introduces NeuroLDS, a neural index-to-point map for generating finite low-discrepancy sequences. The method uses a sinusoidal positional encoding of the index, passes it through an MLP, pre-trains the network to match a Sobol' sequence, and then fine-tunes it to minimize L2-based prefix discrepancy losses. The authors report that NeuroLDS achieves lower L2 discrepancy than Sobol', Halton, and scrambled Sobol' in d=4, and demonstrate improvements on a Borehole QMC integration task, an RRT motion-planning benchmark, and a Black-Scholes PDE surrogate-learning task. The abstract claims that NeuroLDS 'outperforms all previous LDS constructions by a significant margin with respect to discrepancy measures.'

Significance. If established, a machine-learned, extensible low-discrepancy sequence generator that improves on classical constructions would be a useful engineering contribution, especially given the fixed-N limitation of MPMC. The two-stage pretrain-then-fine-tune idea is sensible, and the paper includes useful ablations (e.g., the collapse of direct discrepancy minimization in §3.3). The code is made publicly available. However, the central claim is not supported by the current evidence: the discrepancy comparisons omit L2-optimized sequence constructions that the paper itself cites, the hyperparameters are tuned on the reported metric, experiments are limited to d=4 and N≤10000, and several application claims are overstated relative to the tables. The paper is publishable only after the claims are substantively qualified and the missing comparisons are added.

major comments (5)
  1. [§1.2/§3.1 and Abstract] The central claim that NeuroLDS 'outperforms all previous LDS constructions' is unsupported by the comparison set. The discrepancy experiments in §3.1 include only Sobol', Halton, and scrambled Sobol'. The paper cites Kritzinger (2022), a greedy extensible sequence specifically designed to minimize the L2 discrepancy, but never evaluates it. Since the training loss in Eq. (3) is exactly the L2 discrepancy reported, and hyperparameters are selected via Optuna on the same metric, lower values than generic untuned baselines are expected. Add a comparison against L2-optimized sequences (e.g., Kritzinger's method) or restrict the claim to 'classical Sobol'/Halton baselines.'
  2. [§2.2/§3.3] The method's success depends on Sobol' pretraining. §3.3 reports that direct discrepancy minimization 'collapsed to a degenerate solution' and that pretraining is 'essential.' Thus NeuroLDS is a learned refinement of a classical construction rather than an independent construction. No experiments with alternative pretraining references (e.g., Halton, scrambled Sobol', or a stable random initialization) are reported, so it is unclear whether the approach can bootstrap without a classical sequence or whether it inherits Sobol's limitations. This weakens the novelty claim and the generality of the conclusion; please add such ablations or substantially temper the claims.
  3. [§3.2.1/Table 1] The Borehole claim that NeuroLDS 'consistently delivers superior accuracy' is too strong. In Table 1, Sobol' is better at N=60 (0.1840 vs 0.1864); NM-Greedy is far better at N=100 (0.0299 vs 0.5765) and Sobol' is also better there; at N=380 Sobol' and Halton are both better than NeuroLDS. The experiment is a single deterministic run with no variance estimate, so 'consistently' is not established. Report multiple runs or scrambling replicates with error bars, and use a summary statistic rather than per-N best counts.
  4. [§3.2.2/Table 2] The RRT results report success rates over 160 repetitions but give no confidence intervals or significance tests. In Table 2, NeuroLDS is not the best at passage width 0.52 (79.87 vs Halton 82.54) and 0.56 (83.60 vs Halton 83.66), so the statement 'consistently delivers the highest success' is inaccurate. Provide confidence intervals and a multiple-comparison analysis across the seven widths.
  5. [§3.2.3/Table 3] The Black-Scholes experiment tunes hyperparameters separately for each sampling method via a light random search on the reported MSE, and reports averages over 20 runs without standard errors. The differences between NeuroLDS variants (3.34–4.01) and Sobol' (4.04), all ×10^-4, are small and may be within noise. Report per-method variability (e.g., 95% CIs) and use a fixed network budget or nested model selection to avoid in-sample tuning on the evaluation metric.
minor comments (5)
  1. [Figure 5 vs §3.1] The main text says Figure 5 was trained with D_sym^2, but the Appendix caption says D_star^2. Please correct this inconsistency.
  2. [Table 5] The column 'Final LR ratio' is unclear. Define what it is and how it is used in training.
  3. [Figure 2] The panels would benefit from explicit log-scale axis labels and a legend that is consistent across panels; currently the line styles are hard to distinguish in a small printed figure.
  4. [§3.1] Please state whether the discrepancy values in Table 6 are computed on the same burn-in-adjusted prefixes as the baselines, and give the precise definition of the 'sym' kernel used, since the kernel in Table 4 can take negative values.
  5. [§3.2.2] The text says 10 precomputed sequences of length 10^5 are used, but Table 2 reports success rates from 160 repetitions per sequence; clarify how the repetitions are distributed across the 10 sequences.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; discrepancy comparisons are in-sample but external applications provide independent support.

full rationale

The derivation chain is not circular. NeuroLDS is a construction method: it takes classical Sobol' as initialization and fine-tunes an MLP to minimize the prefix-discrepancy loss in Eq. (3). The discrepancy numbers in Table 6 are therefore the training objective, not an independent prediction; this weakens the 'outperforms all previous constructions' claim as evidence, but it is not a definitional reduction because the optimization could fail and the comparison to baselines is empirical. The paper's stronger independent support comes from out-of-objective applications (Borehole integration, RRT planning, Black-Scholes PDE), where the trained sequences are evaluated on tasks not used in training. The admitted reliance on Sobol'/Halton pretraining (§3.3 and §4) is a limitation on the method's independence, not a circular step. Self-citations to Clément et al. (2025) and Rusch et al. (2024) are not load-bearing: the discrepancy formulas are provided in Appendix A, and MPMC is background. The unsupported 'all previous LDS constructions' claim, due to omitting Kritzinger's greedy L2-optimized sequence, is a completeness/correctness issue, not circularity.

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

The central empirical gains come from choosing a neural parameterization and training it on a classical discrepancy objective. The method does not derive a new theoretical construction; it imports standard math and a Sobol initialization, plus several tunable hyperparameters.

free parameters (5)
  • Optuna hyperparameters (hidden sizes, layers, K, learning rates) = hidden sizes 512/768, layers 5/7, K 32/64, learning rates in Table 5
    Selected per loss on the target discrepancy metric; they are needed to obtain the reported results and are not derived from theory.
  • Burn-in period = 128
    Applied uniformly to pretraining and baselines; an arbitrary convention that affects prefix comparisons at small N.
  • Borehole sensitivity weights gamma = (1.0000, 0.0010, 0.0010, 0.0633, 0.0010, 0.0634, 0.0610, 0.0158)
    Derived from Sobol sensitivity analysis of the specific integrand; used to optimize NeuroLDS and NM-Greedy but not classical baselines, making that application comparison application-tuned.
  • Prefix loss weights w_P = w_P = 1/(N-2)
    Uniform weighting chosen; the paper shows an alternative length-proportional weighting changes early vs late prefix performance, so this is a design choice affecting all reported discrepancies.
  • Training sequence length N = 10^4 for discrepancy, 10^5 for RRT
    The generator is trained for a fixed horizon; behavior beyond N is not tested.
assumptions (5)
  • standard math L2 discrepancy bounds quadrature error via the RKHS Cauchy-Schwarz inequality (Eq. 2).
    Standard QMC theory; used to justify discrepancy as a surrogate for integration error.
  • domain assumption Minimizing the sum of prefix L2 discrepancies produces sequences that are good for integration and planning.
    The method's objective; no theorem guarantees O((log N)^d/N) prefix rates.
  • ad hoc to paper Sobol pretraining is necessary; direct discrepancy minimization collapses (Section 3.3).
    The authors' ablation shows the method depends on a classical LDS for initialization.
  • ad hoc to paper Sinusoidal positional encoding with K bands and i/N normalization captures the index structure needed for LD sequences.
    Chosen by analogy to Fourier features; no uniqueness or optimality argument.
  • domain assumption Weighted product kernels with gamma from sensitivity analysis are appropriate for the Borehole problem.
    Used to tailor NeuroLDS and NM-Greedy; not used for baseline classical sequences.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural Low-Discrepancy Sequences." pith.science (2026). https://pith.science/paper/VUDTD5BX

@misc{pith2026251003745,
  author       = {Pith},
  title        = {Pith review of: Neural Low-Discrepancy Sequences},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VUDTD5BX}},
  note         = {Machine review of arXiv:2510.03745}
}
read the original abstract

Low-discrepancy points are designed to efficiently fill the space in a uniform manner. This uniformity is highly advantageous in many problems in science and engineering, including in numerical integration, computer vision, machine perception, computer graphics, machine learning, and simulation. Whereas most previous low-discrepancy constructions rely on abstract algebra and number theory, Message-Passing Monte Carlo (MPMC) was recently introduced to exploit machine learning methods for generating point sets with lower discrepancy than previously possible. However, MPMC is limited to generating point sets and cannot be extended to low-discrepancy sequences (LDS), i.e., sequences of points in which every prefix has low discrepancy, a property essential for many applications. To address this limitation, we introduce Neural Low-Discrepancy Sequences (NeuroLDS), the first machine learning-based framework for generating finite LDS. Drawing inspiration from classical LDS, we train a neural network to map indices to points such that the resulting sequences exhibit minimal discrepancy across all prefixes. To this end, we deploy a two-stage learning process: supervised approximation of classical constructions followed by unsupervised fine-tuning to minimize prefix discrepancies. We demonstrate that NeuroLDS outperforms all previous LDS constructions by a significant margin with respect to discrepancy measures. Moreover, we demonstrate the effectiveness of NeuroLDS across diverse applications, including numerical integration, robot motion planning, and scientific machine learning. These results highlight the promise and broad significance of Neural Low-Discrepancy Sequences. Our code can be found at https://github.com/camail-official/neuro-lds.

Figures

Figures reproduced from arXiv: 2510.03745 by the authors.

Figure 1
Figure 1. Schematic of the proposed NEUROLDS model. First, each index i ∈ {1, . . . , N} is mapped to a sinusoidal feature vector ψi ∈ R 1+2K. Second, the encoded features are passed through an L-layer multilayer perceptron (MLP), which outputs Xi ∈ [0, 1]d . Finally, the col￾lection {X1, . . . , XN } forms a (learned) low-discrepancy sequence within the unit hypercube. choices include symmetric, centered, and periodic varian… view at source ↗
Figure 2
Figure 2. Comparison plots of different discrepancy metrics for 4d point sequences up to a length [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Sobol’ indices for the Borehole function. [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Visualization of kinematic chain task for a 4d chain in a semi-circular tunnel. [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: First N ∈ {64, 128, 256} points of Sobol’ (blue) and NEUROLDS (orange) in 2d. NEU￾ROLDS was trained with the Dstar 2 discrepancy and Optuna-tuned hyperparameters. Sobol’ exhibits visible structure and clustering, while NEUROLDS distributes points more irregularly yet e…
Figure 6
Figure 6. Figure 6: NeuroLDS discrepancy under D sym 2 in 4d as a function of prefix length P. The three panels show K ∈ {8, 16, 32} (one per panel). Models differ only in K; all remaining hyperparameters and the training setup are identical. The y-axis is log-scaled; lower is better. To …
Figure 7
Figure 7. Figure 7: Comparison of fine-tuning with uniform weights ( [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 2 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Akiba, S

    T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama. Optuna: A next-generation hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp.\ 2623--2631. ACM, 2019. doi:10.1145/3292500.3330701

    arXiv 2019

  3. [3]

    An and A

    J. An and A. Owen. Quasi-regression. J.\ Complexity, 17: 0 588--607, 2001

    2001

  4. [4]

    R. E. Caflisch and W. Morokoff. Valuation of mortgage backed securities using the quasi- M onte C arlo method. In International Association of Financial Engineers First Annual Computational Finance Conference, 1996

    1996

  5. [5]

    Chahine, T

    M. Chahine, T. K. Rusch, Z. J. Patterson, and D. Rus. Improving E fficiency of S ampling-based motion planning via M essage- P assing M onte C arlo. In 9th Annual Conference on Robot Learning, 2025

    2025

  6. [6]

    J. Chen, H. Jiang, and N. Kirk. High- D imensional Q uasi- M onte C arlo via C ombinatorial D iscrepancy, 2025. URL https://arxiv.org/abs/2508.18426

    arXiv 2025

  7. [7]

    Ye Chen, L

    W. Ye Chen, L. Mackey, J. Gorham, F.-X. Briol, and C. Oates. Stein points. In Jennifer Dy and Andreas Krause (eds.), Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pp.\ 844--853. PMLR, 10--15 Jul 2018. URL https://proceedings.mlr.press/v80/chen18f.html

    2018

  8. [8]

    Cl \'e ment, N

    F. Cl \'e ment, N. Kirk, A. B. Owen, and T. K. Rusch. On the optimization of discrepancy measures. arXiv preprint arXiv:2508.04926, 2025

    arXiv 2025

Show all 56 references
  1. [9]

    Clément, C

    F. Clément, C. Doerr, K. Klamroth, and L. Paquete. Constructing optimal star discrepancy sets. Proceedings of the American Mathematical Society, Series B, 12 0 (7): 0 78--90, 2025

    2025

  2. [10]

    Dick and F

    J. Dick and F. Pillichshammer. Digital Nets and Sequences: Discrepancy Theory and Quasi- M onte C arlo Integration . Cambridge University Press, Cambridge, 2010

    2010

  3. [11]

    J. Dick, I. H. Sloan, X. Wang, and H. Wo\' z niakowski. Good lattice rules in weighted K orobov spaces with general weights. Numer.\ Math., 103 0 (1): 0 63--97, 2006. ISSN 0029-599X

    2006

  4. [12]

    J. Dick, F. Kuo, and I. H. Sloan. High dimensional integration --- the Q uasi- M onte C arlo way. Acta Numer., 22: 0 133--288, 2013. doi:10.1017/S0962492913000044

    2013 doi

  5. [13]

    J. Dick, P. Kritzer, and F. Pillichshammer. Lattice Rules: N umerical Integration, Approximation, and Discrepancy . Springer Series in Computational Mathematics. Springer Cham, 2022. doi:https://doi.org/10.1007/978-3-031-09951-9

    2022 doi

  6. [14]

    L. Ding, S. Mak, and J. C. F. Wu. The B dry M at\'ern G P : R eliable incorporation of boundary information on irregular domains for G aussian process modeling, 2025. URL https://arxiv.org/abs/2507.09178

    2025 arXiv

  7. [15]

    Dwivedi and L

    R. Dwivedi and L. Mackey. Kernel thinning. Journal of Machine Learning Research, 25 0 (152): 0 1--77, 2024

    2024

  8. [16]

    Gnewuch, A

    M. Gnewuch, A. Mayer, and K. Ritter. On weighted H ilbert spaces and integration of functions of infinitely many variables. Journal of Complexity, 30 0 (2): 0 29--47, 2014. Dagstuhl 2012

    2014

  9. [17]

    J. H. Halton. On the efficiency of certain quasi-random sequences of points in evaluating multi-dimensional integrals. Numer.\ Math., 2: 0 84--90, 1960

    1960

  10. [18]

    J. M. Hammersley. Monte carlo methods for solving multivariable problems. Proceedings of the New York Academy of Sciences, 86 0 (5): 0 844--874, 1960

    1960

  11. [19]

    Herman and W

    J. Herman and W. Usher. S A L ib: an open-source python library for sensitivity analysis. The Journal of Open Source Software, 2017. doi:10.21105/joss.00097

    2017 doi

  12. [20]

    Herrmann and C

    L. Herrmann and C. Schwab. Multilevel quasi- M onte C arlo uncertainty quantification for advection-diffusion-reaction. In M onte C arlo and Quasi- M onte C arlo Methods: MCQMC , R ennes, France, J uly 2018 , pp.\ 31--67, 2020. doi:10.1007/978-3-030-43465-6\_2

    2018 doi

  13. [21]

    F. J. Hickernell. A generalized discrepancy and quadrature error bound. Mathematics of Computation, 67 0 (221): 0 299--322, 1998

    1998

  14. [23]

    Janson, B

    L. Janson, B. Ichter, and M. Pavone. Deterministic sampling-based motion planning: Optimality, complexity, and performance. The International Journal of Robotics Research, 37 0 (1): 0 46--61, 2018

    2018

  15. [24]

    A. Keller. Quasi- M onte C arlo image synthesis in a nutshell. In M onte C arlo and Quasi- M onte C arlo Methods 2012 , pp.\ 213--249, 2013

    2012

  16. [25]

    Kersaudy, B

    P. Kersaudy, B. Sudret, N. Varsier, O. Picon, and J. Wiart. A new surrogate modeling technique combining kriging and polynomial chaos expansions – application to uncertainty analysis in computational dosimetry. Journal of Computational Physics, 286: 0 103--117, 2015

    2015

  17. [26]

    N. Kirk. On P roinov's lower bound for the diaphony. Uniform Distribution Theory, 2 0 (15): 0 39--72, 2020

    2020

  18. [27]

    Kirk and C

    N. Kirk and C. Lemieux. An improved H alton sequence for implementation in quasi- M onte C arlo methods. In Proceedings of the Winter Simulation Conference, WSC '24, pp.\ 431–442. IEEE Press, 2025

    2025

  19. [28]

    N. Kirk, T. K. Rusch, J. Zech, and D. Rus. Low stein discrepancy via M essage- P assing M onte C arlo. arXiv preprint arXiv:2503.21103, 2025

    2025 arXiv

  20. [29]

    Korn and S

    R. Korn and S. Zeytun. Efficient basket M onte C arlo option pricing via a simple analytical approximation. Journal of Computational and Applied Mathematics, 243: 0 48--59, 2013

    2013

  21. [30]

    Kritzinger

    R. Kritzinger. Uniformly distributed sequences generated by a greedy minimization of the l_2 discrepancy. Moscow Journal of Combinatorics and Number Theory, 11 0 (3): 0 215--236, 2022

    2022

  22. [31]

    Q. Liu, J. Lee, and M. Jordan. A kernelized stein discrepancy for goodness-of-fit tests. In Maria Florina Balcan and Kilian Q. Weinberger (eds.), Proceedings of The 33rd International Conference on Machine Learning, volume 48 of Proceedings of Machine Learning Research, pp.\ 2...

    2016

  23. [32]

    Longo, S

    M. Longo, S. Mishra, T. K. Rusch, and C. Schwab. Higher-order quasi- M onte C arlo training of deep neural networks. SIAM Journal on Scientific Computing, 43 0 (6): 0 A3938--A3966, 2021

    2021

  24. [33]

    Mak and V

    S. Mak and V. R. Joseph. Support points. Ann. Statist., 46 0 (6A): 0 2562--2592, 2018

    2018

  25. [34]

    Mak and V

    S. Mak and V. R. Joseph. Robust designs for G aussian process emulation of computer experiments, 2025. URL https://arxiv.org/abs/2507.09156

    2025 arXiv

  26. [35]

    J. J. Miller, S. Mak, B. Sun, S. R. Narayanan, S. Yang, Z. Sun, K. S. Kim, and C.-B. M. Kweon. Expected diverse utility (edu): Diverse bayesian optimization of expensive computer simulators, 2025. URL https://arxiv.org/abs/2410.01196

    2025 arXiv

  27. [36]

    Mishra and T

    S. Mishra and T. K. Rusch. Enhancing accuracy of deep learning algorithms by training with low-discrepancy sequences. SIAM Journal on Numerical Analysis, 59 0 (3): 0 1811--1834, 2021

    2021

  28. [37]

    M. D. Morris, T. J. Mitchell, and D. Ylvisaker. Bayesian design and analysis of computer experiments: Use of derivatives in surface prediction. Technometrics, 35 0 (3): 0 243--255, 1993

    1993

  29. [38]

    J. A. Nelder and R. Mead. A simplex method for function minimization. The Computer Journal, 4 0 (7): 0 308--313, 1965

    1965

  30. [39]

    Niederreiter

    H. Niederreiter. Random Number Generation and Quasi-Monte Carlo Methods, volume 63 of CBMS-NSF Regional Conference Series in Applied Mathematics. SIAM, 1992

    1992

  31. [40]

    A. B. Owen. Randomly permuted (t,m,s) -nets and (t,s) -sequences. Monte Carlo and Quasi-Monte Carlo Methods in Scientific Computing, 106: 0 299--317, 1995. doi:10.1007/978-1-4612-2552-2_23

    1995 doi

  32. [41]

    A. B. Owen. Scrambled net variance for integrals of smooth functions. The Annals of Statistics, 25 0 (4): 0 1541--1562, 1997. doi:10.1214/aos/1069362388

    1997

  33. [42]

    A. B. Owen. Multidimensional variation for quasi- M onte C arlo. In International Conference on Statistics in honour of Professor Kai-Tai Fang’s 65th birthday, pp.\ 49--74. World Scientific, 2005

    2005

  34. [43]

    A. B. Owen. On dropping the first Sobol' point. In M onte C arlo and Quasi- M onte C arlo Methods: MCQMC , O xford, England, A ugust 2020 , pp.\ 71--86, 2022

    2020

  35. [44]

    A. B. Owen. Practical Quasi-Monte Carlo . At https://artowen.su.domains/mc/practicalqmc.pdf, 2023

    2023

  36. [45]

    Paulin, N

    L. Paulin, N. Bonneel, D. Coeurjolly, J.-C. Iehl, A. Keller, and V. Ostromoukhov. Matbuilder: Mastering sampling uniformity over projections. ACM Transactions on Graphics (Proceedings of SIGGRAPH), 41 0 (4), August 2022. doi:https://doi.org/10.1145/3528223.3530063

    2022

  37. [46]

    K. F. Roth. On irregularities of distribution. Mathematika, 1: 0 73--79, 1954

    1954

  38. [47]

    T. K. Rusch, N. Kirk, M.M. Bronstein, C. Lemieux, and D. Rus. Message- P assing M onte C arlo: Generating low-discrepancy point sets via graph neural networks. Proceedings of the National Academy of Sciences, 121 0 (40): 0 e2409913121, 2024

    2024

  39. [48]

    Saltelli

    A. Saltelli. Making best use of model evaluations to compute sensitivity indices. Computer Physics Communications, 2 0 (145): 0 280--297, 2002

    2002

  40. [49]

    Saltelli, P

    A. Saltelli, P. Annoni, I. Azzini, F. Campolongo, M. Ratto, and S. Tarantola. Variance based sensitivity analysis of model output. design and estimator for the total sensitivity index. Computer Physics Communications, 2 0 (181): 0 259--270, 2010

    2010

  41. [50]

    I. H. Sloan and H. Wozniakowski. When are quasi- M onte C arlo algorithms efficient for high dimensional integrals? J. Complexity, 14: 0 1--33, 1998

    1998

  42. [51]

    I. H. Sloan and H. Woźniakowski. Tractability of multivariate integration for weighted K orobov classes. Journal of Complexity, 17 0 (4): 0 697--721, 2001

    2001

  43. [52]

    I. M. Sobol'. On the distribution of points in a cube and the approximate evaluation of integrals. USSR Comput. Math. and Math. Phys., 7 0 (4): 0 86--112, 1967

    1967

  44. [53]

    Wang and I

    X. Wang and I. H. Sloan. Why are high-dimensional finance problems often of low effective dimension? SIAM J. Sci. Comput., 27 (1): 0 159--183, 2005

    2005

  45. [54]

    T. T. Warnock. Computing the discrepancy of sequences in the unit square. Mathematics of Computation, 26 0 (120): 0 815--819, 1972

    1972

  46. [55]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  47. [56]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  48. [57]

    #1 " @preamble

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

    2012 arXiv

Pith tools

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