Suppose f(x) is a strictly increasing function on some interval (x0, xl], and let g(x) be the inverse function. (For example, f and g might be “exp” and “ln”, or “tan” and “arctan”.) If x is a floating point number such that x0 5 x 2 x1, let f(x) = round(f(x)), and if y is a floating point number such that f(xo) 5 y <> 0 for all x in [x0, xl]), then repeated application of h will be stable in the sense that WV+))) = W(x)), for all x such that both sides of this equation are defined. In other words, there will be no “drift” if the subroutines are properly implemented.