With computers, it is possible to evaluate some numerical functions such as f = exp, sin, arccos, etc. The purpose of this thesis is to study how these functions can be implemented. Depending on the target architecture (software or hardware, small or high accuracy required), different problems must be addressed, but the final goal is always to eventually obtain an implementation as efficient as possible. We first study with an example the problems that arise in the case when the precision is arbitrary. When, on the contrary, the precision is known in advance, the function f is often replaced by an approximation polynomial p. Such a polynomial can then be very efficiently evaluated. In practice, the coefficients of p must be representable on...