piling up ziggurats

This semester. a group of Dauphine graduate students worked under my direction on simulation problems and resorted to using the Ziggurat method developed by George Marsaglia and Wai Wan Tsang, at about the time Devroye was completing his simulation bible. The algorithm covers the half-Normal density by 2², 2⁴, 2⁸, &tc., stripes, all but one rectangles and all with the same surface v. Generating uniformly from the tail strip means generating either uniformly from the rectangle part, x<r, or exactly from the Normal tail x>r, using a drifted exponential accept-reject. The choice between both does not require the surface of the rectangle but a single simulation y=vU/f(r). Furthermore, for the other rectangles, checking first that the first coordinate of the simulated point is less than the left boundary of the rectangle above avoids computing the density. This method is incredibly powerful, once the boundaries have been determined. With 2³² stripes, its efficiency is 99.3% acceptance rate. Compared with a fast algorithm by Ahrens & Dieter (1989), it is three times faster…