Experimental study of a noisy information engine with memory

The investigation of information engines, which use observations of the fluctuating state of a system to extract work from a heat bath, has refined our understanding of the second law of thermodynamics. Inspired by the classic thought experiment of Maxwell, our engine consists of an optically trapped bead in water, with favorable thermal fluctuations rectified by shifting the trap and converted to directed motion and even controlled trajectories. But what if the information acquired is noisy? The performance of these engines deteriorates when the feedback is based on inaccurate data. Here, we report experiments on an information engine that functions well, even when the noise in each observation is roughly equal to the size of thermal fluctuations. The key innovation is to use the memory of past observations to improve the estimate of particle position. By basing feedback on improved estimates of the system state, we can extract power even when the naive use of observations leads to a net loss of power, as too many wrong feedback decisions are made. Our engine overcomes these difficulties and shows that effective information engines can be made without requiring highly accurate observations.