Overview

• The study, published in Science on September 25, 2025, reports learning a noisy optical channel in about 15 minutes versus an estimated 20 million years for a comparable classical method.
• Two squeezed light beams were entangled so one probed the channel and the other served as a reference, and a joint homodyne measurement extracted more information per shot than classical readouts.
• The setup operated at telecom wavelengths with standard optical components and tolerated ordinary losses, indicating the advantage arises from the measurement strategy rather than ideal hardware.
• The experiment validates a 2024 theoretical proposal predicting an entanglement-enabled learning advantage for bosonic random displacement channels.
• DTU's bigQ center led the work with collaborators from the University of Chicago, Perimeter Institute, University of Waterloo, Caltech, MIT, and KAIST, and the team emphasizes this is a proof-of-principle rather than a targeted real-world deployment.