Pith. sign in

REVIEW 1 major objections 4 minor 31 references

This paper establishes that the local spacing of scrambled quasi-Monte Carlo point sets is set by the randomization type: full Owen scrambling forces pairs polynomially closer than optimal, while matrix and linear scrambling keep the mesh r

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

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

T0 review · deepseek-v4-flash

2026-08-03 14:21 UTC pith:P6PQORYZ

load-bearing objection Useful sharp results and a careful Owen analysis, but the abstract's unconditional O_P(log N) mesh-ratio claim for matrix/linear scrambling is false: a simple (1,m,2)-net counterexample forces ρ=Ω(N^{1/2}). the 1 major comments →

arxiv 2607.29063 v1 pith:P6PQORYZ submitted 2026-07-31 math.NA cs.NA

Separation properties of scrambled digital nets and related random point sets

Kosuke Suzuki This is my paper
classification math.NA cs.NA MSC 11K3611K4552C17
keywords quasi-Monte Carlodigital netsOwen scramblingmatrix scramblinglinear scramblingseparation radiusmesh ratiocoincidence bound
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper's central claim is that the geometric effect of randomizing a low-discrepancy point set is governed by whether the randomization introduces local independence or preserves shared algebraic randomness. For full Owen scrambling of any fixed-t net, pairs of points can be polynomially closer than the optimal N^{-1/d} scale, so the mesh ratio grows at least like N^{1/(2d)} and a single scrambling of a (t,d)-sequence is almost surely not quasi-uniform. For matrix and linear scrambling of binary digital nets with fixed t, the same geometry is much better: under the uniform coincidence bound the mesh ratio is O_P(log N), and in the balanced-prefix affine-tail model it is Θ_P(log N). The paper also pins down sharp probabilistic orders for Monte Carlo, jittered, and Latin hypercube sampling, whose mesh ratios all diverge polynomially.

Core claim

On the paper's own terms, the discovery is a sharp contrast between two scrambling mechanisms. A single full Owen scrambling (independent uniform permutations applied at every node of the digit tree) of any (t,d)-sequence is almost surely non-quasi-uniform: its minimum distance is at most O_P(N^{-3/(2d)}) and its mesh ratio at least Ω_P(N^{1/(2d)}), with matching upper bounds up to logarithmic factors under uniform coincidence and common-prefix conditions. By contrast, for binary digital fixed-t nets in dimension d≥2 satisfying the uniform coincidence bound, matrix and linear scrambling (random nonsingular lower-triangular matrices, with fixed or random digital shift) give R_∞(P_scr)=Ω_P(N^{

What carries the argument

The argument runs on two competing mechanisms. Under full Owen scrambling, Lemma 4.2 gives subtree independence: once the first k scrambled digits are fixed, the tails of points in distinct prefixes are independent uniforms; applying this at level k≈m/d produces close pairs at scale N^{-3/(2d)}. Under matrix/linear scrambling, that independence is replaced by a shared linear structure: a digitwise difference is the image of the input difference under a random lower-triangular matrix, and Lemmas 5.2–5.3 count how many such differences can be small. The load-bearing global condition is the elementary-interval coincidence bound (Definition 4.8, Eq. (21)): M_m(k)≤ A N(N-1) b^{-|k|} with A indepe

Load-bearing premise

The paper assumes, rather than proves for every fixed-t family, that the uniform elementary-interval coincidence bound M_m(k) ≤ A N(N-1) b^{-|k|} holds with A independent of m; if a family violates it, close pairs can be far more common and the O_P(log N) mesh-ratio conclusion may fail.

What would settle it

Compute the coincidence count M_m(k) for an explicit fixed-t binary digital net family in dimension d≥2 (for example, Sobol' points in dimension 3) and check whether M_m(k) exceeds A 2^m(2^m-1)2^{-|k|} for every constant A as m grows, at levels k≈m/d. If such a family exists, the matrix/linear-scrambling mesh-ratio bound O_P(log N) would be false for it; conversely, verifying the bound suggests the logarithmic order is universal for fixed-t digital nets.

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

If this is right

  • Randomization of a QMC construction does not automatically preserve quasi-uniformity; full Owen scrambling typically destroys it, while matrix/linear scrambling keeps mesh ratio to O_P(log N) under the coincidence bound.
  • A single full Owen scrambling of any (t,d)-sequence is almost surely non-quasi-uniform, with mesh ratio diverging like N^{1/(2d)} up to logarithmic factors.
  • Monte Carlo, jittered, and Latin hypercube sampling all have mesh ratios diverging as positive powers of N; jittered sampling satisfies an explicit Weibull limit law for its minimum distance.
  • For binary digital fixed-t nets in d≥2 satisfying the uniform coincidence bound, matrix and linear scrambling guarantee minimum distance Ω_P(N^{-1/d}(log N)^{-1}), which is at most a logarithmic factor off the optimal scale.
  • The balanced-prefix affine-tail model and the 1D binary digital (0,m,1)-net analysis show the Θ_P(log N) order is genuinely attained, so the O_P(log N) bound is sharp rather than an artifact of the method.

Where Pith is reading between the lines

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

  • If the uniform coincidence bound holds for the standard fixed-t digital constructions (e.g., Sobol' or Niederreiter nets in dimension ≥3), then matrix/linear scrambling would give a practical recipe for randomized space-filling designs that stay nearly quasi-uniform; verifying the bound for those explicit families is a direct testable step.
  • The sharp contrast suggests a design heuristic: when a randomized QMC point set is needed for both integration and scattered-data approximation, prefer randomizations that commute with the digital linear structure over fully independent digit permutations.
  • The one-dimensional equivalence between Owen-scrambled (0,1)-sequences and jittered samples suggests the Weibull limit law for jittered minimum distance should also describe the minimum distance of balanced-prefix affine-tail models at the 1D level; that would unify the two benchmark scales.
  • An open testable extension is whether the exact Θ_P(log N) order of the balanced-prefix affine-tail model carries over to higher-dimensional matrix/linear scrambling beyond the paper's one-sided bounds; the model's rigidity (few distinct difference vectors among adjacent dyadic cells) is the suspected reason.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

1 major / 4 minor

Summary. The paper studies the separation radius and mesh ratio of randomized quasi-Monte Carlo point sets, proving sharp probabilistic orders for Monte Carlo, jittered, and Latin hypercube sampling, and contrasting them with the behavior of full Owen scrambling versus matrix/linear scrambling of digital nets. For full Owen scrambling, the paper proves a universal upper bound on the minimum distance and, under uniform coincidence and common-prefix conditions, matching lower bounds. For matrix and linear scrambling, it proves a logarithmic upper bound on the mesh ratio for binary digital fixed-t nets satisfying a uniform elementary-interval coincidence bound, and a sharp Θ(log N) result in a balanced-prefix affine-tail model. The main technical contributions are a subtree-independence lemma for Owen scrambling, a pair-counting framework using common-prefix lengths, and a linear-algebraic tail model for matrix scrambling.

Significance. If the results are correct, the paper makes a substantive contribution to the quasi-uniformity literature: it shows that the local geometry of scrambled QMC point sets depends delicately on the scrambling mechanism, with full Owen scrambling causing polynomial deterioration of the mesh ratio while shared algebraic randomness in matrix/linear scrambling can preserve logarithmic mesh ratios. The proofs are detailed and largely self-contained, with explicit constants and careful probabilistic arguments; the affine-tail model is a useful new tool. However, the abstract advertises an unconditional O_P(log N) statement for matrix/linear scrambling that is false without the uniform coincidence bound, and the paper should be revised to make the hypotheses and the necessity of the coincidence bound clear.

major comments (1)
  1. [Abstract; §5.2, Definition 4.8, Theorem 5.5] The abstract states that for matrix and linear scrambling of binary digital nets with fixed t in dimension d≥2 the mesh ratio is O_P(log N), without qualification. This is false. For m≥2, take C1=I_m and C2 with C2_{r,m-r}=1 (1≤r≤m-1) and C2_{m,m}=1. This is a (1,m,2)-net: for e1+e2=m-1 the combined first e1 and e2 rows of C1 and C2 have disjoint supports and rank m-1. Let h=e_m. Then C1h=C2h=e_m, so for every scrambling realization and every shift the pair (a,a⊕h) satisfies ∥x_{a⊕h}-x_a∥∞≤2^{-m+1}, giving R∞≤2^{-m+1} deterministically. Since h∞≥2^{-m/2-1}, ρ∞=Ω(N^{1/2}), contradicting O_P(log N). This family also violates (21): M_m(m-1,m-1)=N=2^m, while the bound would require a constant A independent of m. Thus the uniform coincidence bound is essential. Theorem 5.5 is correct as a conditional statement; the abstract must be amended and the paper should include an explicit example show
minor comments (4)
  1. [§3.2 proof of Theorem 3.3] The variable t used for the Weibull threshold conflicts with the net quality parameter t used throughout the paper. Rename the threshold (e.g., s or τ) to avoid confusion.
  2. [Abstract] The qualification 'with coincidence bound' appears in Table 1 and in Theorem 5.5 but not in the abstract. This inconsistency should be fixed; the abstract should state the uniform coincidence-bound hypothesis for the matrix/linear O_P(log N) result.
  3. [§5.2] After Theorem 5.5, a remark noting which families satisfy the uniform coincidence bound (e.g., (0,m,d)-nets and (0,e,d)-sequences by Lemma 4.10) and that not all fixed-t digital nets do would help readers assess the theorem's applicability.
  4. [§2.2] The O_P notation is defined for sequences indexed by m→∞, but several results are stated as N→∞. A brief clarification that m=log_b N and that all statements are equivalent would improve readability.

Circularity Check

0 steps flagged

No significant circularity: all main bounds are derived from stated deterministic hypotheses (net property, coincidence bound, common-prefix cutoff) and from explicit probabilistic calculations; no prediction is an input by construction.

full rationale

Walked the derivation chain. Section 3 derives the Monte Carlo, jittered and Latin hypercube sampling orders from Poisson approximation, elementary pair-count estimates, and negative association; these are self-contained and parameter-free. Section 4's Owen-scrambling bounds use two explicitly stated input conditions, Definition 4.8 (Eq. (21)) and Definition 4.9, and then prove the minimum-distance tail by conditioning on scrambled prefixes and applying Lemma 4.7. Lemma 4.10 verifies those conditions for (0,m,d)-nets and (0,e,d)-sequences, and Lemma 4.15 derives the common-prefix cutoff from the coincidence bound for digital nets. The conclusion R_infty = O_P(N^{-3/(2d)}) etc. is a probabilistic statement about scrambled copies and is not a restatement of M_m(k). Section 5 proves the matrix/linear-scrambling logarithmic mesh-ratio bound under the uniform coincidence bound; the balanced-prefix affine-tail model is introduced as a model, Theorem 5.11 establishes Theta_P(2^{-k}/k) within that model, and Corollary 5.12 proves that one-dimensional matrix/linear scrambling actually realizes the model. The self-citations [3,4,8,9] are contextual/comparative; the only internally cited technical tool is Lemma 2.2 (from [3]), an elementary monotonicity transfer that is auxiliary and not load-bearing for the central claims. The abstract's unconditional 'O_P(log N)' wording for matrix/linear scrambling omits the coincidence-bound hypothesis of Theorem 5.5; this is a correctness/qualification issue, and the skeptic's counterexample (if valid) attacks the truth of that unqualified abstract claim, not circularity. No fitted parameter, renamed known result, or definition-in-terms-of-target occurs anywhere in the derivation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

The central technical contributions rest on two uniformity conditions on the input net family (coincidence bound and common-prefix cutoff). These are honest, stated assumptions rather than fitted parameters. The balanced-prefix affine-tail model is a new analytic construction but introduces no new entity or fitted constant; it is a probability model whose only randomness comes from independent uniform matrix entries.

axioms (5)
  • domain assumption Uniform elementary-interval coincidence bound: M_m(k) <= A N(N-1) b^(-|k|) with A independent of m (Definition 4.8, Eq. (21)).
    Assumed for the net family in Theorem 5.5 and Corollary 4.12. It is not proved to follow from fixed t alone, and the abstract's matrix/linear O_P(log N) claim depends on it.
  • domain assumption Uniform common-prefix cutoff: no elementary interval with |k| >= m+kappa contains more than one point (Definition 4.9).
    Needed for the full-Owen lower-tail union bound in Theorem 4.11. For digital nets it follows from the coincidence bound via Lemma 4.15; for non-digital nets it is an additional condition.
  • standard math (t,m,d)-net covering bound of Niederreiter: h_infty(P) <= b^((d-1+t)/d) N^(-1/d) (Eq. (2)).
    Used to turn separation-radius bounds into mesh-ratio bounds after scrambling preserves the net property.
  • standard math External probability theorems: Chen-Stein/Poisson approximation (Arratia-Goldstein-Gordon [1]), negative association (Joag-Dev-Proschan [13]), covering radius of random points (Reznikov-Saff [23]).
    Used in Sections 3 and 4 for jittered and Latin hypercube sampling; cited precisely and not re-proved.
  • domain assumption Scrambling maps (full Owen, matrix, linear) preserve the net property and act by independent digit permutations or random linear maps (Definitions in Section 2.4).
    These are the standard scrambling models from [17,18,5]; they are the object of study rather than derived facts.

pith-pipeline@v1.3.0-daily-deepseek · 3468 in / 4081 out tokens · 224368 ms · 2026-08-03T14:21:47.509101+00:00 · methodology

0 comments
read the original abstract

We study how standard randomization procedures affect the local geometry of quasi-Monte Carlo point sets, as measured by their minimum distance and mesh ratio. Although probabilistic selection within structured lattice families can produce quasi-uniform point sets, randomizing an existing low-discrepancy construction need not preserve quasi-uniformity. We first determine sharp probabilistic orders for Monte Carlo, jittered, and Latin hypercube sampling, whose mesh ratios diverge as positive powers of $N$. The orders are $\Theta_{\mathbb{P}}(N^{1/d}(\log N)^{1/d})$ for Monte Carlo sampling, $\Theta_{\mathbb{P}}(N^{1/(d+1)})$ for jittered sampling, and, for $d\ge 2$, $\Theta_{\mathbb{P}}(N^{1/d}(\log N)^{1/d})$ for Latin hypercube sampling. We also obtain a Weibull limit law for the minimum distance of jittered samples. For full Owen scrambling, every family of fixed-$t$ nets has minimum distance $O_{\mathbb{P}}(N^{-3/(2d)})$, and its mesh ratio is therefore $\Omega_{\mathbb{P}}(N^{1/(2d)})$. Under uniform coincidence and common-prefix conditions, these bounds are sharp up to logarithmic factors. Moreover, a single full Owen scrambling of any $(t,d)$-sequence is almost surely non-quasi-uniform. By contrast, for matrix and linear scrambling of binary digital nets with fixed $t$ in dimension $d\ge 2$, the mesh ratio is $O_{\mathbb{P}}(\log N)$, whereas it is $\Theta_{\mathbb{P}}(\log N)$ in the separate balanced-prefix affine-tail model. The model also yields the exact probabilistic order for one-dimensional binary digital $(0,m,1)$-nets under matrix or linear scrambling. These results demonstrate that the geometric effect of randomization is governed by whether it introduces local independence or shared algebraic randomness.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 1 linked inside Pith

  1. [1]

    Arratia, L

    R. Arratia, L. Goldstein, and L. Gordon,Two moments suffice for Poisson approximations: the Chen–Stein method, Annals of Probability17(1989), no. 1, 9–25

    1989

  2. [2]

    S. R. Blackburn, C. Homberger, and P. Winkler,The minimum Manhattan distance and minimum jump of permutations, Journal of Combinatorial Theory, Series A161(2019), 364–386

    2019

  3. [3]

    J. Dick, T. Goda, G. Larcher, F. Pillichshammer, and K. Suzuki,On the quasi-uniformity properties of quasi-Monte Carlo point sets and sequences – Part I: Lattices and Kronecker sequences, Mathematics of Computation (2026), to appear

    2026

  4. [4]

    J. Dick, T. Goda, and K. Suzuki,On the quasi-uniformity properties of quasi-Monte Carlo point sets and sequences – Part II: Digital nets and sequences, Mathematics of Computation (2026), to appear

    2026

  5. [5]

    Dick and F

    J. Dick and F. Pillichshammer,Digital nets and sequences: Discrepancy theory and quasi-Monte Carlo integration, Cambridge University Press, Cambridge, 2010

    2010

  6. [6]

    Doerr,A sharp discrepancy bound for jittered sampling, Mathematics of Computation91(2022), no

    B. Doerr,A sharp discrepancy bound for jittered sampling, Mathematics of Computation91(2022), no. 336, 1871–1892

    2022

  7. [7]

    D. P. Dubhashi and D. Ranjan,Balls and bins: a study in negative dependence, Random Structures & Algorithms13(1998), no. 2, 99–124

    1998

  8. [8]

    Goda,The Sobol’ sequence is not quasi-uniform in dimension 2, Proceedings of the American Mathe- matical Society152(2024), no

    T. Goda,The Sobol’ sequence is not quasi-uniform in dimension 2, Proceedings of the American Mathe- matical Society152(2024), no. 8, 3209–3213

    2024

  9. [9]

    T. Goda, R. Hofer, and K. Suzuki,Disproving the quasi-uniformity of the Halton sequences and of some Halton-type sequences, Journal of Complexity95(2026), 102047

    2026

  10. [10]

    T. Goda, Y. Liu, and R. Tempone,Quasi-Monte Carlo with a Hankel random digital net, arXiv:2604.24105, 2026

    Pith/arXiv arXiv 2026

  11. [11]

    Gr¨ unschloß, J

    L. Gr¨ unschloß, J. Hanika, R. Schwede, and A. Keller, (t, m, s)-nets and maximized minimum distance, Monte Carlo and Quasi-Monte Carlo Methods 2006, Springer, Berlin, 2008, pp. 397–412

    2006

  12. [12]

    Gr¨ unschloß and A

    L. Gr¨ unschloß and A. Keller, (t, m, s)-nets and maximized minimum distance, Part II, Monte Carlo and Quasi-Monte Carlo Methods 2008, Springer, Berlin, 2009, pp. 395–409

    2008

  13. [13]

    Joag-Dev and F

    K. Joag-Dev and F. Proschan,Negative association of random variables with applications, Annals of Sta- tistics11(1983), no. 1, 286–295

    1983

  14. [14]

    Matouˇ sek,On theL 2-discrepancy for anchored boxes, Journal of Complexity14(1998), no

    J. Matouˇ sek,On theL 2-discrepancy for anchored boxes, Journal of Complexity14(1998), no. 4, 527–556

    1998

  15. [15]

    M. D. McKay, R. J. Beckman, and W. J. Conover,A comparison of three methods for selecting values of input variables in the analysis of output from a computer code, Technometrics21(1979), no. 2, 239–245. SEPARATION PROPERTIES OF SCRAMBLED POINT SETS 33

    1979

  16. [16]

    Niederreiter,Random number generation and quasi-Monte Carlo methods, CBMS-NSF Regional Con- ference Series in Applied Mathematics, vol

    H. Niederreiter,Random number generation and quasi-Monte Carlo methods, CBMS-NSF Regional Con- ference Series in Applied Mathematics, vol. 63, Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 1992

    1992

  17. [17]

    A. B. Owen,Randomly permuted(t, m, s)-nets and(t, s)-sequences, Monte Carlo and Quasi-Monte Carlo Methods in Scientific Computing, Lecture Notes in Statistics, vol. 106, Springer, New York, 1995, pp. 299– 317

    1995

  18. [18]

    4, 363–378

    ,Variance with alternative scramblings of digital nets, ACM Transactions on Modeling and Com- puter Simulation13(2003), no. 4, 363–378

    2003

  19. [19]

    Pan,Automatic optimal-rate convergence of randomized nets using median-of-means, Mathematics of Computation95(2026), no

    Z. Pan,Automatic optimal-rate convergence of randomized nets using median-of-means, Mathematics of Computation95(2026), no. 359, 1415–1446

    2026

  20. [20]

    Pausinger and S

    F. Pausinger and S. Steinerberger,On the discrepancy of jittered sampling, Journal of Complexity33 (2016), 199–216

    2016

  21. [21]

    Pronzato and W

    L. Pronzato and W. G. M¨ uller,Design of computer experiments: space filling and beyond, Statistics and Computing22(2012), no. 3, 681–701

    2012

  22. [22]

    Pronzato and A

    L. Pronzato and A. Zhigljavsky,Quasi-uniform designs with optimal and near-optimal uniformity constant, Journal of Approximation Theory294(2023), 105931

    2023

  23. [23]

    Reznikov and E

    A. Reznikov and E. B. Saff,The covering radius of randomly distributed points on a manifold, International Mathematics Research Notices (2016), no. 19, 6065–6094

    2016

  24. [24]

    Sakai and T

    N. Sakai and T. Goda,Space-filling lattice designs for computer experiments, arXiv:2602.15390, 2026

    arXiv 2026

  25. [25]

    Schaback and H

    R. Schaback and H. Wendland,Kernel techniques: from machine learning to meshless methods, Acta Numerica15(2006), 543–639

    2006

  26. [26]

    Schulte and C

    M. Schulte and C. Th¨ ale,Poisson point process convergence and extreme values in stochastic geometry, Stochastic Analysis for Poisson Point Processes, Bocconi & Springer Series, vol. 7, Springer, Cham, 2016, pp. 255–294

    2016

  27. [27]

    Stein,Large sample properties of simulations using Latin hypercube sampling, Technometrics29(1987), no

    M. Stein,Large sample properties of simulations using Latin hypercube sampling, Technometrics29(1987), no. 2, 143–151

    1987

  28. [28]

    Suzuki,Exactℓ ∞-separation radius of Sobol’ sequences in dimension 2, arXiv:2508.14803, 2025

    K. Suzuki,Exactℓ ∞-separation radius of Sobol’ sequences in dimension 2, arXiv:2508.14803, 2025

    arXiv 2025

  29. [29]

    Tezuka,On the discrepancy of generalized Niederreiter sequences, Journal of Complexity29(2013), no

    S. Tezuka,On the discrepancy of generalized Niederreiter sequences, Journal of Complexity29(2013), no. 3–4, 240–247

    2013

  30. [30]

    Wendland,Scattered data approximation, Cambridge Monographs on Applied and Computational Math- ematics, vol

    H. Wendland,Scattered data approximation, Cambridge Monographs on Applied and Computational Math- ematics, vol. 17, Cambridge University Press, Cambridge, 2005

    2005

  31. [31]

    Wiart, C

    J. Wiart, C. Lemieux, and G. Y. Dong,On the dependence structure and quality of scrambled(t, m, s)-nets, Monte Carlo Methods and Applications27(2021), no. 1, 1–26. (K. Suzuki)F aculty of Science, Yamagata University, 1-4-12 Kojirakawa-machi, Yamagata, 990-8560, Japan Email address:kosuke-suzuki@sci.kj.yamagata-u.ac.jp

    2021