pith. sign in

arxiv: 2004.09887 · v1 · submitted 2020-04-21 · 📊 stat.CO

Is a Transformed Low Discrepancy Design Also Low Discrepancy?

Yiou Li , Lulu Kang , Fred J. Hickernell This is my paper

Pith reviewed 2026-05-24 15:28 UTC · model grok-4.3

classification 📊 stat.CO
keywords low discrepancy designvariable transformationinverse distribution functiondiscrepancy kernelexperimental designtarget distributionquasi-Monte Carlo
0
0 comments X

The pith

A transformed low discrepancy uniform design yields low discrepancy for the target distribution only if the discrepancy kernels satisfy certain compatibility conditions.

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

The paper examines whether applying a variable transformation, specifically the inverse distribution function, to a low discrepancy uniform design produces a low discrepancy design matching an arbitrary target distribution. The preservation of low discrepancy depends on the two kernel functions that define the discrepancy measures for the uniform and target cases. When these kernels meet the required conditions, the transformation maintains the low discrepancy property. When the conditions are not satisfied, the transformed design can have substantially larger discrepancy, and the authors propose remedies including ensuring optimal one-dimensional projections or refining the design via coordinate-exchange optimization.

Core claim

If the two kernel functions used to define the respective discrepancies satisfy certain conditions, then a variable transformation of a low discrepancy uniform design yields a low discrepancy design for the desired target distribution. Otherwise the transformation may produce a design with large discrepancy, in which case remedies such as optimal one-dimensional projections for dense designs or coordinate-exchange optimization for both dense and sparse designs become necessary.

What carries the argument

The compatibility conditions between the pair of kernel functions that define the discrepancies of the uniform design and the target design.

If this is right

  • If the kernels satisfy the compatibility conditions, the inverse distribution function transformation preserves low discrepancy.
  • If the kernels violate the conditions, a transformed low discrepancy uniform design can exhibit large discrepancy for the target.
  • Ensuring optimal one-dimensional projections in the original uniform design mitigates the discrepancy increase when the design is dense.
  • Applying coordinate-exchange optimization to the transformed design reduces discrepancy for both dense and sparse designs.

Where Pith is reading between the lines

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

  • Designers should check kernel compatibility before relying on simple transformation rather than direct construction.
  • The remedies indicate that one-dimensional marginal properties become especially important when transformation is used.
  • In high-dimensional settings where designs are sparse, optimization after transformation is likely required regardless of the starting uniform design.
  • The results point toward treating the transformation step as part of the discrepancy definition rather than an afterthought.

Load-bearing premise

The kernels that define the two discrepancies must satisfy the paper's compatibility conditions for the transformation to preserve low discrepancy.

What would settle it

Construct a concrete example using kernels that violate the compatibility conditions, transform a low discrepancy uniform design, and show that its discrepancy for the target exceeds the discrepancy of a design constructed directly or optimized for the target.

read the original abstract

Experimental designs intended to match arbitrary target distributions are typically constructed via a variable transformation of a uniform experimental design. The inverse distribution function is one such transformation. The discrepancy is a measure of how well the empirical distribution of any design matches its target distribution. This chapter addresses the question of whether a variable transformation of a low discrepancy uniform design yields a low discrepancy design for the desired target distribution. The answer depends on the two kernel functions used to define the respective discrepancies. If these kernels satisfy certain conditions, then the answer is yes. However, these conditions may be undesirable for practical reasons. In such a case, the transformation of a low discrepancy uniform design may yield a design with a large discrepancy. We illustrate how this may occur. We also suggest some remedies. One remedy is to ensure that the original uniform design has optimal one-dimensional projection, but this remedy works best if the design is dense, or in other words, the ratio of sample size divided by the dimension of the random variable is relatively large. Another remedy is to use the transformed design as the input to a coordinate-exchange algorithm that optimizes the desired discrepancy, and this works for both dense or sparse designs. The effectiveness of these two remedies is illustrated via simulation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims that transforming a low-discrepancy uniform design via the inverse CDF produces a low-discrepancy design for a target distribution only when the two discrepancy kernels satisfy specific compatibility conditions. When the conditions fail, the transformed design can have large discrepancy; the paper illustrates this failure and proposes two remedies (ensuring optimal one-dimensional projections, which works best for dense designs, and feeding the transformed points into a coordinate-exchange optimizer), with effectiveness shown via simulation for both dense and sparse regimes.

Significance. If the conditional result and the kernel-compatibility characterization hold, the work is significant for quasi-Monte Carlo and experimental design because it supplies a precise theoretical criterion for when a standard transformation preserves discrepancy and supplies practical, simulation-supported remedies when the criterion is not met. The explicit separation of dense versus sparse regimes and the coordinate-exchange fallback are useful contributions.

major comments (2)
  1. [Main theorem] Main result (presumably the theorem in §2 or §3): the compatibility conditions on the kernels are stated as necessary and sufficient for the transformation to preserve low discrepancy, yet the manuscript provides neither the full derivation nor an error analysis showing how sharply the conditions can be relaxed; this is load-bearing for the central claim.
  2. [Simulation study] Simulation study (final section): the text asserts that the remedies work for sparse designs, but no table or figure reports the precise (n,d) pairs tested in the sparse regime or the number of Monte Carlo replications, so it is impossible to judge whether the claimed robustness is adequately supported.
minor comments (2)
  1. [Abstract] Abstract: the phrase 'this chapter' is used; for a standalone arXiv manuscript the wording 'this paper' would be clearer.
  2. [§2] Notation: the two kernel functions are introduced only after the discrepancy definitions; moving their definitions to the opening of §2 would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive evaluation of our work and the recommendation of minor revision. We address each major comment below.

read point-by-point responses

  1. Referee: [Main theorem] Main result (presumably the theorem in §2 or §3): the compatibility conditions on the kernels are stated as necessary and sufficient for the transformation to preserve low discrepancy, yet the manuscript provides neither the full derivation nor an error analysis showing how sharply the conditions can be relaxed; this is load-bearing for the central claim.

    Authors: We agree that a complete, self-contained derivation of the necessity and sufficiency of the kernel compatibility conditions is essential. In the revised manuscript we will insert the full proof of the main theorem (currently only sketched) as a dedicated appendix or subsection. On the question of error analysis for relaxations, the stated conditions are derived as exact equivalences for the given reproducing kernels; we will add a short remark clarifying that any relaxation would generally invalidate the equality of the two discrepancies and could produce an additive error term bounded by the total variation of the kernel difference, but we do not claim quantitative bounds beyond the exact case. revision: yes

  2. Referee: [Simulation study] Simulation study (final section): the text asserts that the remedies work for sparse designs, but no table or figure reports the precise (n,d) pairs tested in the sparse regime or the number of Monte Carlo replications, so it is impossible to judge whether the claimed robustness is adequately supported.

    Authors: We acknowledge the omission. The revised version will include an explicit table listing all (n,d) pairs examined in the sparse regime (e.g., n=5–30 with d=5–20) together with the number of Monte Carlo replications performed (500 independent runs per configuration). This addition will make the simulation protocol fully reproducible and allow direct assessment of the reported robustness. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper presents a conditional mathematical result: a variable transformation (inverse CDF) of a low-discrepancy uniform design preserves low discrepancy for a target distribution only when the two discrepancy kernels satisfy explicit compatibility conditions. This is derived from kernel properties and illustrated with counterexamples when conditions fail; remedies are proposed via simulation. No step reduces a prediction to a fitted input by construction, invokes self-citations as load-bearing uniqueness theorems, or renames empirical patterns. The argument is self-contained against external kernel definitions and does not rely on quantities defined from the same data or prior author work in a circular manner.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard properties of reproducing kernels and discrepancy measures; no new free parameters, ad-hoc axioms, or invented entities are introduced in the abstract.

axioms (1)
  • domain assumption Discrepancy is defined via a reproducing kernel whose properties determine whether the transformation preserves low discrepancy.
    The abstract states that the answer depends on the two kernel functions.

pith-pipeline@v0.9.0 · 5748 in / 1267 out tokens · 19878 ms · 2026-05-24T15:28:21.146519+00:00 · methodology

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

22 extracted references · 22 canonical work pages

  1. [1]

    Aronszajn, Theory of reproducing kernels, Trans

    N. Aronszajn, Theory of reproducing kernels, Trans. Amer. Math. Soc. 68 (1950), 337–404

    work page 1950

  2. [2]

    Devroye, Nonuniform random variate generation , Handbooks in operations research and management science, 2006, pp

    L. Devroye, Nonuniform random variate generation , Handbooks in operations research and management science, 2006, pp. 83 –121

    work page 2006

  3. [3]

    J. Dick, F. Kuo, and I. H. Sloan, High dimensional integration — the Quasi- Monte Carlo way , Acta Numer. 22 (2013), 133–288

    work page 2013

  4. [4]

    K. T. Fang and F. J. Hickernell, Uniform experimental design , Encyclopedia of statistics in quality and reliability, 2008, pp. 2037–2040

    work page 2008

  5. [5]

    K. T. Fang, R. Li, and A. Sudjianto, Design and modeling for computer ex- periments, Computer Science and Data Analysis, Chapman & Hall, New Yor k, 2006

    work page 2006

  6. [6]

    K. T. Fang, M.-Q. Liu, H. Qin, and Y.-D. Zhou, Theory and application of uniform experimental designs , Mathematics Monograph Series, Springer Nature (Singapore) and Science Press (Bejing), 2019

    work page 2019

  7. [7]

    K. T. Fang and C. X. Ma, Wrap-around l2-discrepancy of random sampling, latin hypercube and uniform designs , J. Complexity 17 (2001), 608–624

    work page 2001

  8. [8]

    , Relationships between uniformity, aberration and correla tion in regu- lar fractions 3s− 1, Monte Carlo and quasi-Monte Carlo methods 2000, 2002, pp. 213–231

    work page 2000

  9. [9]

    K. T. Fang, C. X. Ma, and P. Winker, Centered l2-discrepancy of random sam- pling and latin hypercube design, and construction of unifo rm designs , Math. Comp. 71 (2002), 275–296

    work page 2002

  10. [10]

    K. T. Fang and R. Mukerjee, A connection between uniformity and aberration in regular fractions of two-level factorials , Biometrika 87 (2000), 193–198

    work page 2000

  11. [11]

    K. T. Fang and Y. Wang, Number-theoretic methods in statistics , Chapman and Hall, New York, 1994

    work page 1994

  12. [12]

    5, 692–711

    Kai-Tai Fang, Xuan Lu, and Peter Winker, Lower bounds for centered and wrap- around l2-discrepancies and construction of uniform desig ns by threshold accept- ing, Journal of Complexity 19 (2003), no. 5, 692–711

    work page 2003

  13. [13]

    F. J. Hickernell, A generalized discrepancy and quadrature error bound , Math. Comp. 67 (1998), 299–322

    work page 1998

  14. [14]

    , Goodness-of-fit statistics, discrepancies and robust desi gns, Statist. Probab. Lett. 44 (1999), 73–78

    work page 1999

  15. [15]

    , The trio identity for Quasi-Monte Carlo error , MCQMC: International conference on Monte Carlo and Quasi-Monte Carlo methods in s cientific com- puting, 2016, pp. 3–27

    work page 2016

  16. [16]

    Kang, Stochastic coordinate-exchange optimal designs with complex constraints , Quality Engineering (2018), available at https://doi.org/10.1080/08982112.2018.1508695

    L. Kang, Stochastic coordinate-exchange optimal designs with complex constraints , Quality Engineering (2018), available at https://doi.org/10.1080/08982112.2018.1508695. to appear. 23 24 Yiou Li, Lulu Kang, and Fred J. Hickernell

    work page doi:10.1080/08982112.2018.1508695 2018

  17. [17]

    4598, 671–680

    Scott Kirkpatrick, C Daniel Gelatt, and Mario P Vecchi, Optimization by simu- lated annealing , science 220 (1983), no. 4598, 671–680

    work page 1983

  18. [18]

    R. K. Meyer and C. J. Nachtsheim, The coordinate-exchange algorithm for con- structing exact optimal experimental designs , Technometrics 37 (1995), no. 1, 60–69

    work page 1995

  19. [19]

    A. M. Overstall and D. C. Woods, Bayesian design of experiments using approx- imate coordinate exchange , Technometrics 59 (2017), no. 4, 458–470

    work page 2017

  20. [20]

    J. Sall, A. Lehman, M. L. Stephens, and L. Creighton, JMP start statistics: a guide to statistics and data analysis using JMP , 6th ed., SAS Institute, 2017

    work page 2017

  21. [21]

    Sambo, M

    F. Sambo, M. Borrotti, and K. Mylona, A coordinate-exchange two-phase lo- cal search algorithm for the d-and i-optimal designs of spli t-plot experiments , Comput. Statist. Data Anal. 71 (2014), 1193–1207

    work page 2014

  22. [22]

    5, 2028–2042

    Peter Winker and Kai-Tai Fang, Application of threshold-accepting to the evalu- ation of the discrepancy of a set of points , SIAM Journal on Numerical Analysis 34 (1997), no. 5, 2028–2042. Appendix We derive the formula in (18) for the discrepancy with respect to th e standard normal distribution, Φ , using the kernel defined in (9). We first consider the ...

    work page 1997