Table of Contents
Fetching ...

Uniform distribution via lattices: from point sets to sequences

Damir Ferizović

TL;DR

This work develops a lattice-based framework to generate low-discrepancy sequences in $[0,1)^d$ by constructing the $S^{\Box}_{b,d}$ and $S^{\boxplus}_{b,d}$ sequences and proving a general transfer theorem: discrepancy bounds for empirical measures of perturbed lattice images $P(T_v(\omega))$ bound the discrepancy of initial segments $Z_N$ of the mapped sequence. It unifies and extends classical one-dimensional results (van der Corput and Niederreiter) to higher dimensions and to sphere domains using projections, yielding explicit bounds in terms of digit sums and $L^p$ norms. The paper also provides refined quantitative bounds on the size of index sets with favorable discrepancy and extends the construction to self-similar sets via Hutchinson's framework. These results offer a versatile, computationally efficient approach to produce highly uniform point sequences on $[0,1)^d$ and $\mathbb{S}^2$ with practical implications for numerical integration and quasi-Monte Carlo methods.

Abstract

In this work we construct many sequences $S=S^\Box_{b,d}$, or $S=S^\boxplus_{b,d}$ in the $d$--dimensional unit hypercube, which for $d=1$ are (generalized) van der Corput sequences or Niederreiter's $(0,1)$-sequences in base $b$ respectively. Further, we introduce the notion of $f$-sublinearity and use it to define discrepancy functions which subsume the notion of $L^p$-discrepancy, Wasserstein $p$-distance, and many more methods to compare empirical measures to an underlying base measure. We will relate bounds for a given discrepancy functions $\mathscr{D}$ of the multiset of projected lattice sets $P(b^{-m}\mathbb{Z}^d$), to bounds of $\mathscr{D}(Z_N)$, i.e. the initial segments of the sequence $Z=P(S)$ for any $N\in\mathbb{N}$. We show that this relation holds in any dimension $d$, for any map $P$ defined on a hypercube, and any discrepancy function as introduced in this work for which bounds on $P(b^{-m}\mathbb{Z}^d+v$) can be obtained. We apply this theorem in $d=1$ to obtain bounds for the $L^p$--discrepancy of van der Corput and Niederreiter (0,1) sequences in terms of digit sums for all $0<p\leq \infty$. In $d=2$ an application of our construction yields many sequences on the two-sphere, such that the initial segments $Z_N$ have low $L^\infty$--discrepancy.

Uniform distribution via lattices: from point sets to sequences

TL;DR

This work develops a lattice-based framework to generate low-discrepancy sequences in by constructing the and sequences and proving a general transfer theorem: discrepancy bounds for empirical measures of perturbed lattice images bound the discrepancy of initial segments of the mapped sequence. It unifies and extends classical one-dimensional results (van der Corput and Niederreiter) to higher dimensions and to sphere domains using projections, yielding explicit bounds in terms of digit sums and norms. The paper also provides refined quantitative bounds on the size of index sets with favorable discrepancy and extends the construction to self-similar sets via Hutchinson's framework. These results offer a versatile, computationally efficient approach to produce highly uniform point sequences on and with practical implications for numerical integration and quasi-Monte Carlo methods.

Abstract

In this work we construct many sequences , or in the --dimensional unit hypercube, which for are (generalized) van der Corput sequences or Niederreiter's -sequences in base respectively. Further, we introduce the notion of -sublinearity and use it to define discrepancy functions which subsume the notion of -discrepancy, Wasserstein -distance, and many more methods to compare empirical measures to an underlying base measure. We will relate bounds for a given discrepancy functions of the multiset of projected lattice sets ), to bounds of , i.e. the initial segments of the sequence for any . We show that this relation holds in any dimension , for any map defined on a hypercube, and any discrepancy function as introduced in this work for which bounds on ) can be obtained. We apply this theorem in to obtain bounds for the --discrepancy of van der Corput and Niederreiter (0,1) sequences in terms of digit sums for all . In an application of our construction yields many sequences on the two-sphere, such that the initial segments have low --discrepancy.
Paper Structure (14 sections, 10 theorems, 105 equations, 7 figures)

This paper contains 14 sections, 10 theorems, 105 equations, 7 figures.

Key Result

Theorem 1

Let $P:[0,1)^d\rightarrow\mathbb{M}$ be a map. Fix an admissible function-pair $(\mathscr{D},f)$, integer $b>1$ and $u\in\mathbb{N}\cup\{\infty\}$. Let $(L_m,\ \mathscr{Z}^u)$ have one of two meanings for $0\leq m\leq u$ (where $\mathscr{Z}^\infty$ is the set of sequences): Define the sets $W^{k,b}_m= [0,b^{-m})^d\cap (\mathbb{Z}^d\cdot b^{-k-1})$ and let $W^{\infty,b}_m=\bigcup_{k=0}^\infty W^{k

Figures (7)

  • Figure 1: The first 11 elements of a possible choice for $S^{\Box}_{2,2}$ in a). In b) we see the underlying construction via shifting $S_{2^2}$ clearly, and its order is seen in a).
  • Figure 2: Here we see how Lemma \ref{['lem_latticeTracing']} works for $N=39=3^3+3^2+3$ and $S=S^\Box_{3,1}$ with $\pi_1=\pi_2=\pi_4$ the identity permutation of elements $(1,2)$, and $\pi_3\neq \pi_1$.
  • Figure 3: Let $S_N$ be like in Figure \ref{['fig:Lem8Visualization']} and $\mathscr{D}$ be 1-subadditive. Here we see how Lemma \ref{['lem_latticeTracing']} and 1-subadditivity work hand in hand to bound $\mathscr{D}(E_{S_N})$.
  • Figure 4: We see the first 11 elements of a possible choice for $S^{\boxplus}_{2,2}$, and how lattice filling and the $S^\Box$-guidance conditions are satisfied, where $S^\Box_{2,2}$ is as in Figure \ref{['fig:Sbox']}, which is also the base for the color scheme. The point $q_{11}$ could be placed in one of the boxes to the right of $q_3$ or $q_7$; the point $q_{12}$ must be placed in the box between $q_9$ and $q_8$; $q_{13}$ must be placed between $q_5$ and $q_4$, etc.
  • Figure 5: This visual aid intends to help the reader verify that $M^\mathds{1}(b^m)\leq 1$ in the proof of Theorem B. The curve represents $L^{-1}(\partial C)$ of a spherical cap $C$. The dots represent $\omega_{b^m}$ and $T_v(\omega_{b^m})$ respectively.
  • ...and 2 more figures

Theorems & Definitions (40)

  • Definition 1
  • Definition 2
  • Definition 3
  • Remark
  • Theorem 1: Main result
  • Remark
  • Definition 4
  • Remark
  • Corollary 2
  • proof
  • ...and 30 more