In this letter, we compute the variation of the mutual information, resulting from a small variation in its argument. Although the result can be applied in many problems, we consider only one example: the result is used for deriving a new method for blind source separation in linear mixtures. The experimental results emphasize the performance of the resulting algorithm.