Robust Adaptive Quantum Sensing under Decoherence Distribution Shift: Distributionally Robust Training of Learned Experiment-Design Agents

Abstract

Autonomous, learned experiment-design agents, neural / reinforcement-learning policies that choose each measurement to maximize information, increasingly drive quantum sensing and self-driving labs. We show that such agents, trained on an assumed decoherence model, are brittle to decoherence distribution shift: a neural design policy trained at a nominal dephasing rate learns long interrogations that collapse when the true rate differs. Our central contribution is a distributionally robust training objective that hedges the decoherence law over a Wasserstein / Γ-minimax ambiguity set; the resulting agent degrades gracefully, cutting estimation error by ≈17–20% under a 3× mismatch and halving the worst-case design risk, for only a modest in-distribution premium. We prove the robust objective is tractable, a nested pair of one-dimensional convex problems with strong duality for a continuous decoherence law, and equip the agent with a fast, calibrated skew-normal belief whose correctness is certified by an adaptive, multiparameter skewed quantum Bernstein–von Mises theorem, including an exact skewness self-annihilation identity. The result is a trustworthy, distribution-shift-robust agent for autonomous quantum experiment design.

Keywords