In recent years, quantum experiments have become increasingly precise, fast, and capable of high resolution. Particular interest has been given to quantum control, which aims to prepare, manipulate, and steer quantum states toward desired outcomes. Common applications of quantum control include state preparation for quantum computing algorithms, protocols to implement nanoscale machines, and feedback to guide a system’s evolution. Feedback involves collecting information from quantum measurements, then acting on the system based on measurement outcomes. The standard measurement model in quantum mechanics is the projective measurement, which destroys quantum coherence by causing the wave function to collapse to a subspace spanned by the eigenstates of the measured operator.
This thesis explores the theory of weak measurement processes, a class of measurement protocols that extract information from a quantum system while (partially) preserving coherence. The weak measurement protocol has a tunable parameter that controls the information obtained per measurement cycle and the disturbance (decoherence) introduced into the quantum system. Using this nondestructive form of measurement, one can extract information during the system’s evolution and apply real-time feedback to drive the system’s evolution to specific target states. A general master equation is derived to describe continuous feedback using weak measurements with general filtering processing. Particular cases of low-pass and band-pass filters are studied in detail and applied to a harmonic oscillator cooling protocol. Results show that ground-state cooling of the quantum harmonic oscillator can be achieved.
Finally, this dissertation discusses an experimental and computational project that uses ma- chine learning to estimate the temperature and the number of atoms of a cold atomic cloud. The goal is to use non-destructive measurements to infer hidden properties of the atomic ensemble without disturbing the atomic trap. Results show that reasonable accuracy can be achieved using various neural network architectures, depending on the complexity of the input data. The accu- racy and responsiveness of the trained models make them suitable for real-time estimators that can be used in closed-loop feedback.