Information thermodynamics for feedback processes with the appearance of overshooting

We investigate feedback processes with measurement-induced protocols for particular tasks that drive systems in specified directions in state spaces. We focus on mutual information as a measure of correlation between system and memory, which has been known to play a crucial role for the second law of information thermodynamics. The performance of task is enhanced in the early stage of driving, along with the decrease of correlation and mutual information due to the passage from initial measurement. However, we find that the performance is suppressed if the time of driving exceeds a threshold, which we call feedback overshooting. We find that a type of correlation, anticorrelation, between system and memory is built up as a result of overshooting and gives rise to regaining mutual information. We examine the effect of overshooting in detail from two examples. We study the Szilard engine for the task of work extraction. We also study a recurrent feedback with finite time interval for the task to reduce the mean square distance of a colloid below the value by thermal fluctuation. We find that recurrent feedback is stable only for a moderate range of time intervals and the intensity of feedback protocol. We discuss the problem of divergence of mutual information for error-free measurement.