Convergence

Mean (star-) discrepancy for 10 input point sets optimized by Lloyd's as well as our method. Each input point set initially contained 1024 random points. Our method drastically improves discrepancy already within the first iteration and converges after about 130 iterations on average. Although Lloyd's method eventually achieves a comparable discrepancy and is computationally less expensive, it converges much slower and needs about 1350 iterations on average using the same convergence criterion. In contrast to Lloyd's method, also note that our method produces point distributions of low-discrepancy complexity class already after a single iteration.