REVIEW 3 major objections 3 minor 1 cited by
On the optimization of discrepancy measures
T0 review · 3 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash
Pith's one-line read The average squared discrepancy—averaging the L2 star discrepancy over all unit-cube vertices—is computable in O(d n^2) time, equivalent to a symmetric weighted L2 criterion, and uniquely avoids the pathologies that undermine the usual L2 s
desk verdict A plausible and likely useful new QMC criterion; the equivalence and numerical claims are checkable and the authors are serious, but the abstract alone can't verify the constant factor or the robustness of the experiments. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the average squared discrepancy, defined as the mean of $2^d$ $L_2$ star discrepancies computed from each vertex of the unit cube as anchor. Its equivalence to a weighted symmetric $L_2$ criterion is the mechanism that simultaneously gives a low-degree polynomial evaluation formula, removes the anchor-dependence that causes pathologies, and permits transferring optimization algorithms designed for symmetric discrepancies.
What would settle it
Take the same point set in dimensions $d=1,2,4,8$ and compute the ratio of its average squared discrepancy to the weighted symmetric $L_2$ criterion. If the ratio is not essentially constant across dimensions and across point sets of very different quality (e.g., a grid versus a random cloud), the core equivalence is disconfirmed. A second empirical check: minimize the average squared discrepancy for $n$ points and evaluate the resulting $L_2$ star discrepancy; if it is not competitive with known low-discrepancy constructions on the same sample size, the transfer claim fails.
Extended reading notes
Core claim
The paper's central claim is that symmetrizing the $L_2$ star discrepancy by averaging its vertex-anchored versions yields a discrepancy measure that is both easy to minimize and free of the known counterexamples that rule out the standard $L_2$ star discrepancy as an optimization criterion. The average squared discrepancy is shown to be equivalent to a weighted symmetric $L_2$ criterion up to a constant, which explains its smooth behavior and $O(d n^2)$ cost. The paper reports that it is the only measure in its comparison that avoids the pathological behavior, and that optimizing it transfers to good performance for the $L_2$ star discrepancy, but not conversely.
Load-bearing premise
The constant factor linking the average squared discrepancy to the weighted symmetric $L_2$ criterion is independent of the point configuration and the dimension; if it changes with geometry or $d$, the equivalence and the computational-practicality story collapse.
Editorial extensions
If this is right
- For quasi-Monte Carlo, one can now generate high-quality point sets by directly minimizing a smooth, $O(d n^2)$-cost objective, rather than relying on $L_\infty$ or empirically tuned heuristics.
- Because the average squared discrepancy is equivalent to a symmetric weighted criterion, existing software and theory for symmetrized discrepancies apply to it unchanged.
- The asymmetric failure of transfer—good for $L_2$ star when optimizing the new measure, but not vice versa—suggests that averaging over anchors removes bias that the usual star discrepancy cannot detect.
- The paper clarifies which discrepancy notions are safe to use as objectives: among those tested, only the symmetrized one avoids pathologies.
- For fixed $n$ and $d$, evaluation cost $O(d n^2)$ makes the criterion feasible for moderate- to large-scale point-set optimization.
Reading between the lines
- The same vertex-averaging trick could be applied to other anchored discrepancies (e.g., $L_p$ star discrepancies) to define pathology-free objectives; the paper only compares $L_2$ variants.
- In high dimensions, enumerating all $2^d$ vertices is impossible, but the proven equivalence to a weighted symmetric criterion suggests that the average squared discrepancy can be evaluated in closed form without enumeration, which is presumably the basis of the $O(d n^2)$ claim.
- One might expect the average squared discrepancy to have connections to the concept of roto-symmetric discrepancy or to deep lattice-rooted designs, which could be explored using the paper's equivalence.
- A testable prediction: for any fixed point set, the average squared discrepancy is always bounded between constant multiples of the symmetric criterion; if that fails, the paper's main equivalence is wrong.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a new criterion, the average squared discrepancy, obtained by averaging the L2 star discrepancy over the 2^d anchor vertices of the unit cube. It claims that this criterion can be evaluated in O(d n^2) time, that it is equivalent to Hickernell's weighted symmetric L2 criterion up to a constant factor, and that among a wide range of traditional discrepancy measures it is the only one avoiding the pathologies identified by J. Matoušek. A numerical study reportedly shows that optimizing the average squared discrepancy yields strong performance for the L2 star discrepancy, while the reverse transfer does not hold.
Significance. If all claims hold, the paper offers a practically valuable optimization target: a smooth, efficiently computable criterion that is theoretically connected to existing discrepancy measures and empirically transfers to the more standard L2 star discrepancy. The O(d n^2) complexity bound and the claimed constant-factor equivalence are crisp, testable statements that would be useful to the community. The manuscript is, however, available to the referee only in abstract form, so the central derivations and numerical evidence cannot be independently verified here.
major comments (3)
- [Abstract] The claim that the average squared discrepancy is equivalent to Hickernell's weighted symmetric L2 criterion 'by a constant factor' is load-bearing. The full proof must specify whether the constant depends on the dimension d, the number of points n, or the point configuration. If the constant varies with d or with the shape of the point set, then the O(d n^2) computation does not by itself imply a useful equivalence, and the numerical transfer to L2 star discrepancy would require a quantitative relation rather than a bare constant-factor statement. The abstract alone cannot establish this.
- [Abstract] The statement that 'only the average squared discrepancy avoids the problems raised by Matoušek' is a strong uniqueness claim. The manuscript must define precisely which class of discrepancy measures is being compared, what 'avoids' means in this context (e.g., convexity, differentiability, boundedness, or a specific counterexample to Matoušek's pathology), and over what domain the comparison is made. Without this specification, the claim is not falsifiable and cannot be assessed from the abstract.
- [Abstract / Numerical study] The reported transfer result—optimizing the average squared discrepancy leads to strong L2 star discrepancy performance, whereas the converse does not—needs a precise statement of the experimental protocol. The abstract does not specify the dimensions, the range of n, the number of optimization runs, the point-set families tested, or the performance metric used. These details are essential to interpret the claim, and their absence in the abstract is acceptable, but the full text must provide them; otherwise the numerical conclusion cannot be checked.
minor comments (3)
- [Abstract] Please define n and d explicitly at first use: n is presumably the number of points and d the dimension, but this should be stated.
- [Abstract] The phrase 'averages over 2^d versions of the L2 star discrepancy anchored in the different vertices' is somewhat concise. Clarify whether the average is over the full anchored boxes or over a normalizing factor, and whether the vertices are all corners of [0,1]^d.
- [Abstract] A citation to Hickernell's weighted symmetric L2 criterion would help the reader; the abstract mentions it without a reference. Also, the diacritic in 'Matoušek' is rendered as a LaTeX escape in the abstract; the final version should use proper Unicode or LaTeX source.
Circularity Check
No significant circularity evident from the abstract-level derivation.
full rationale
The abstract reports a new criterion defined as an average of anchored L2 star discrepancies, states an O(d n^2) computational cost, and claims a mathematical equivalence to Hickernell's weighted symmetric L2 criterion. These are independent mathematical claims, not fitted to the evaluation metric. The numerical study optimizes the average squared discrepancy and evaluates against L2 star discrepancy; since these are different functionals, the transfer is non-circular by design. No self-citation load-bearing step, no fitted parameter renamed as a prediction, and no uniqueness claim imported from the authors appear in the abstract. Because only the abstract was available for review, no specific equation-level reduction could be exhibited; under the rule requiring quoted evidence of circularity, the appropriate finding is no significant circularity.
Assumptions & free parameters
assumptions (2)
- standard math The standard definition and properties of discrepancy measures, particularly the L2 star discrepancy, are assumed.
- domain assumption Matoušek's characterization of pathologies in the L2 star discrepancy is accepted as the baseline for what a better criterion should avoid.
Cite this review
Pith. "Pith review of On the optimization of discrepancy measures." pith.science (2026). https://pith.science/paper/MTSPKPLG
@misc{pith2026250804926,
author = {Pith},
title = {Pith review of: On the optimization of discrepancy measures},
year = {2026},
howpublished = {\url{https://pith.science/paper/MTSPKPLG}},
note = {Machine review of arXiv:2508.04926}
}
abstract
Points in the unit cube with low discrepancy can be constructed using algebra or, more recently, by direct computational optimization of a criterion. The usual $L_\infty$ star discrepancy is a poor criterion for this because it is computationally expensive and lacks differentiability. Its usual replacement, the $L_2$ star discrepancy, is smooth but exhibits other pathologies shown by J. Matou\v{s}ek. In an attempt to address these problems, we introduce the \textit{average squared discrepancy} which averages over $2^d$ versions of the $L_2$ star discrepancy anchored in the different vertices of $[0,1]^d$. Not only can this criterion be computed in $O(dn^2)$ time, like the $L_2$ star discrepancy, but also we show that it is equivalent to a weighted symmetric $L_2$ criterion of Hickernell's by a constant factor. We compare this criterion with a wide range of traditional discrepancy measures, and show that only the average squared discrepancy avoids the problems raised by Matou\v{s}ek. Furthermore, we present a comprehensive numerical study showing in particular that optimizing for the average squared discrepancy leads to strong performance for the $L_2$ star discrepancy, whereas the converse does not hold.
Forward citations
Cited by 1 Pith paper
-
Neural Low-Discrepancy Sequences
A trained index-to-point network generates low-discrepancy sequences whose prefixes beat Sobol, Halton, and scrambled Sobol discrepancy in four dimensions.
Reviewed August 5, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.