The inequality (6) indicates the requirements for the accurate regression forests: low correlation between residuals of different trees in the forest and low error for the every single tree. Moreover, randomization aimed at low correlation can be applied to decrease the average error of the trees by the factor q.