Uniformly distributed sequences generated by a greedy minimization of the \(L_2\) discrepancy (Q2084271): Difference between revisions

The paper is devoted to some new approach to generate low-discrepancy sequences in \([0;1)^d\). Choose some initial point \(x_1\) and inductively define the sequence \((x_1,\ldots,x_n,\ldots)\) as follows: suppose that \((x_1,\ldots,x_n)\) is already defined. Then \(x_{n+1}\) is defined from the condition: \(L_2\)-discrepancy of \((x_1,\ldots,x_n,x_{n+1})\) is minimal over the sequences \((x_1,\ldots,x_n,y)\). Instead of the classical \(L_2\)-discrepancy, some of it variants or generalizations can be used. The estimations of \(L_2\)-discrepancy of obtained sequences are found. These estimations can be written as \(L_{2,N}(X)\leq c_d\sqrt{N}\) and look quite far from optimal but numerical experiments (based on Warnock's formula for \(L_2\)-discrepancy) show that \(L_2\)-discrepancy of constructed sequences is very small. For example, for \(d=1\) numerical calculations shows that for most \(N\) star- and \(L_2\)-discrepancies of obtained sequence is smaller that in the case of van der Corput sequence.