The reservoir learning power across quantum many-body localization transition

Harnessing the quantum computation power of present noisy-intermediate-size-quantum devices has received tremendous interest in last few years. Here we study learning a one-dimensional long-range randomly-coupled spin chain, within framework reservoir computing. In time sequence tasks, find system many-body localized (MBL) phase holds long-term memory, which can be attributed to emergent local integrals motion. On other hand, MBL does not provide sufficient nonlinearity highly-nonlinear sequences, show parity check task. This is reversed ergodic phase, provides but compromises memory capacity. complex task Mackey—Glass prediction that requires both capacity and nonlinearity, optimal performance near MBL-to-ergodic transition. leads guiding principle engineering at edge ergodicity reaching for generic tasks. Our theoretical finding tested with near-term NISQ devices.