On the best approximation for bootstrapped empirical processes