On the uniform distribution modulo 1 of multidimensional LS-sequences
This result warns practitioners using LS-sequences for Quasi-Monte Carlo integration that naive multidimensional extension may fail, but the finding is incremental as it identifies specific number-theoretic conditions for failure.
The paper proves that combining two one-dimensional low-discrepancy LS-sequences coordinatewise can yield a two-dimensional sequence that is not even dense in [0,1]^2, contradicting the expectation that such combinations produce low-discrepancy sequences.
Ingrid Carbone introduced the notion of so-called LS-sequences of points, which are obtained by a generalization of Kakutani's interval splitting procedure. Under an appropriate choice of the parameters $L$ and $S$, such sequences have low discrepancy, which means that they are natural candidates for Quasi-Monte Carlo integration. It is tempting to assume that LS-sequences can be combined coordinatewise to obtain a multidimensional low-discrepancy sequence. However, in the present paper we prove that this is not always the case: if the parameters $L_1,S_1$ and $L_2,S_2$ of two one-dimensional low-discrepancy LS-sequences satisfy certain number-theoretic conditions, then their two-dimensional combination is not even dense in $[0,1]^2$.