I think they wanted to express that learning to predict the correct output ("error minimization") puts a limit on the achievable performance. While ranking (not

I think they wanted to express that learning to predict the correct output ("error minimization") puts a limit on the achievable performance. While ranking (not just RL, really) allows to improve beyond the current best-known answer.