Experimental Realization of an Information Machine with Tunable Temporal Correlations

We experimentally realize a Maxwell's demon that converts information gained by measurements to work. Our setup is composed of a colloidal particle in a channel filled with a flowing fluid. A barrier made by light prevents the particle from being carried away by the flow. The colloidal particle then performs biased Brownian motion in the vicinity of the barrier. The particle's position is measured periodically. When the particle is found to be far enough from the barrier, feedback is applied by moving the barrier upstream while maintaining a given minimal distance from the particle. At steady state, the net effect of this measurement and feedback loop is to steer the particle upstream while applying very little direct work on it. This clean example of a Maxwell's demon is also naturally operated in a parameter regime where correlations between outcomes of consecutive measurements are important. Interestingly, we find a tradeoff between output power and efficiency. The efficiency is maximal at quasistatic operating conditions, whereas both the power output and rate of information gain are maximal for very frequent measurements.