When dealing with noisy and/or partial inputs to describe a shape, the standard approach in geometry processing is to reconstruct a clean implicit surface first, and to do geometry on it. But why this one in particular? What about all the other surfaces that would have been coherent with the input?