Global minimizers of periodic L2-discrepancy in dimensions at least three

Determine the appropriate generalization of permutation sets for finding approximate global minimizers of the periodic L2-discrepancy in dimensions d ≥ 3, including whether Latin hypercubes provide better candidates.

Background

The paper proves that weak Latin hypercubes generalize the exact relation between extreme and periodic L2-discrepancy known for permutation sets, but also shows that weak Latin hypercubes have discrepancy substantially worse than globally optimal point sets when d ≥ 3. The authors therefore ask which more appropriate structured class could contain approximate global minimizers of periodic L2-discrepancy.

They specifically suggest Latin hypercubes generated by coordinate permutations as possible candidates. Such point sets can achieve asymptotically optimal discrepancy rates in certain constructions, although the paper notes that higher-dimensional numerical evidence is unavailable and that no known rational lattices attain Roth’s bound for d ≥ 3.

References

What is the right generalization for finding (approximate) global minimizers of the periodic $L_2$-discrepancy in dimensions $d \geq 3$? Perhaps Latin hypercubes, which gives sets of the form $$ \left{\frac{1}{M}(\sigma_1(m), \dots, \sigma_d(m)) : m=0, 1, \dots, M-1\right} $$ for permutations $\sigma_1, \dots, \sigma_d$ of ${0, 1, \dots, M-1}$ are better candidates for this task.

On the $L_2$-discrepancy of Latin hypercubes  (2502.20828 - Nagel, 28 Feb 2025) in Section 6, Conclusion and outlook, item (i)