Do the samples have to be uniform or how can it be adapted to be efficient but also sample enough in areas of large gradients? Does it become a nested problem of sampling gradients then?