OTOC-Based Quantum Feature Engineering; A Framework and Empirical Study

Conventional quantum machine learning feature maps primarily employ static encoding strategies that fail to capture the dynamical evolution of quantum systems. We propose leveraging Out-of-Time-Order Correlators (OTOCs)—a diagnostic tool for quantum information scrambling—as a dynamical feature extraction mechanism for classification tasks. Our framework extracts physically meaningful features including scrambling rates (Lyapunov-like exponents), saturation values, and temporal variance from OTOC trajectories. We benchmark this approach against classical kernels and standard quantum baselines across three classification problems: distinguishing chaotic from integrable regimes, detecting quantum phase transitions, and classifying a standard tabular dataset. While OTOC-derived features encode interpretable physical dynamics, our results reveal accuracy and scalability challenges compared to optimized classical baselines, suggesting that system size and encoding strategy require further optimization for practical advantage.