Digital nets in dimension two with the optimal order of \(L_p\) discrepancy (Q2199494)

The paper deals with digital \((0,m,2)\)-nets over \(\mathbb{Z}_2\). Such nets are point sets with \(2^m\), \(m\in\mathbb{N}\), points in the unit square with very regular distribution properties. The points of a \((0,m,2)\)-net are generated by means of linear algebra, and are characterized by two generating matrices \(C_1\) and \(C_2\) of size \(m\times m\) over \(\mathbb{Z}_2\). Digital nets are frequently studied in the theory of uniform distribution modulo one, where one considers to which extent the points are evenly distributed. The level of uniform distribution can be assessed by studying the so-called discrepancy function, which compares the number of points of a given point set in axes-parallel boxes to the volumes of these boxes. Taking \(L_p\)-norms of the discrepancy function then yields the notions of \(L_p\)-discrepancy for \(p\in[1,\infty]\). Given a point set of \(N\) points in the unit square, it is an intriguing question what the order of magnitude with respect to \(N\) of the \(L_p\)-discrepancy of the point set is. The authors consider a particular choice of the generating matrices \(C_1\) and \(C_2\) of a digital \((0,m,2)\)-net over \(\mathbb{Z}_2\), namely \[ C_1= \begin{pmatrix} 0 & 0 & \ldots & 0 & 1\\ 0 & 0 & \ldots & 1 & 0\\ \ldots & \ldots & \ldots & \ldots & \ldots\\ 0 & 1 & \ldots & 0 & 0\\ 1 & 0 & \ldots & 0 & 0\\ \end{pmatrix},\ \ \ C_2=\begin{pmatrix} 1 & a_1 & a_1 & \ldots & a_1 & a_1\\ 0 & 1 & a_2 & \ldots & a_2 & a_2\\ \ldots & \ldots & \ldots & \ldots & \ldots & \ldots\\ 0 & 0 & 0 &\ldots & 1 & a_{m-1}\\ 0 & 0 & 0 &\ldots & 0 & 1\\ \end{pmatrix}, \] i.e., \(C_2\) is an upper-triangular matrix with a special structure. Previously, it was known that one can obtain point sets with optimal order of \(L_2\)-discrepancy by applying a symmetrization method to digital \((0,m,2)\)-nets. The present paper shows further progress on this question by proving that by choosing the matrices \(C_1\) and \(C_2\) as above, and making a suitable assumption on the \(a_i\) in the matrix \(C_2\), one can also obtain digital nets with optimal order of \(L_p\)-discrepancy for \(p\in [1,\infty)\), but without the need to symmetrize the net anymore. Indeed, the main result of the paper implies that the \(L_p\)-discrepancy of a digital \((0,m,2)\)-net generated by \(C_1\) and \(C_2\) of the above form, and with \(N=2^m\) points, is of order \[ \mathcal{O} \left(\frac{\max (\sqrt{\log N}, h_m (\boldsymbol{a}))}{N}\right), \] where \(h_m (\boldsymbol{a})\) is defined as \(h_m (\boldsymbol{a})=\sum_{i=1}^m (1-a_i)\), i.e., \(h_m (\boldsymbol{a})\) counts how many of the \(a_i\) in \(C_2\) are zero. In particular, if \(h_m (\boldsymbol{a})\) is of order \(\mathcal{O} (\sqrt{\log N})\), the order of magnitude of the \(L_p\)-discrepancy of the corresponding digital net is \(\mathcal{O} (\sqrt{\log N} /N)\), which is optimal due to classical results in discrepancy theory.