Artificial increasing returns to scale and the problem of sampling from lognormals