. . . "Here is one approach: using Chernoff bounds, you can show that if you take polynomially many random samples and average to estimate F, the additive error is at most M/poly(n) except with negilgible probability, where M is an upper bound on f." .