If b is small compared to a, generate U[-a,a] and reject values in (-b,b) -- that is, generate a new value when that happens. This is the general solution to the problem of generating a random variate from an oddball distribution: you generate from a distribution that dominates yours and reject values you don't want. When the PDF of your desired distribution lies between zero and the envelope distribution you need another random variate to act as a surrogate: if U[0,envelope-value(x)] < [my-funky-dist-value(x)] we keep the first variate x, else we generate a new pair.

For instance, I would like to generate uniform_real numbers in the intervall [-a;-b]U[b;a].

The previous answer is an example of a less general but often more efficient approach: cut up another distribution and paste it back together to make up yours. These approaches can be combines: paste up something close and use surrogate screening to fine-tune. See any good book on simulation. -swn

Do you have any reference ?

The text I used in school is Sheldon M. Ross, _Simulation_ 2nd ed, Academic Press 1997 ISBN 0-12-598410-3, section 4.4 (pp 53-56) "The Accepance-Rejection Technique". -swn