Computing the global optimum of a multivariate polynomial over the reals

Let f be a polynomial in Q[X₁, ..., X_n] of degree D. We provide an efficient algorithm in practice to compute the global supremum sup_{x∈ Rⁿ} f(x) of f (or its infimum inf_{{x∈ Rⁿ}}f(x)). The complexity of our method is bounded by D^O(n)}. In a probabilistic model, a more precise result yields a complexity bounded by O(n⁷D⁴ⁿ) arithmetic operations in Q. Our implementation is more efficient by several orders of magnitude than previous ones based on quantifier elimination. Sometimes, it can tackle problems that numerical techniques do not reach. Our algorithm is based on the computation of generalized critical values of the mapping x-> f(x), i.e. the set of points {c∈ C mid exists (x_ll)_ll∈ N}⊂ Cⁿ ;f(x_ll)-> c, ;||x_ll||||d_{x_ll} f||-> 0 { when }ll-> ∞}. We prove that the global optimum of f lies in its set of generalized critical values and provide an efficient way of deciding which value is the global optimum.