Dynamic compression of whole-brain neural trajectories during human motor learning

Abstract

Motor learning involves the dynamic reconfiguration of brain activity across widely distributed networks. Yet, the moment-to-moment evolution of the whole-brain functional states underpinning this process remains unknown. Here, applying manifold-based trajectory analyses to human fMRI data, we uncover a fundamental signature of motor learning: Neural state transitions are sharply constrained during initial learning—manifesting as a sharp compression of trajectory geometry—and relax these constraints as performance stabilizes. This effect, which closely tracked behavioral error, was recapitulated during relearning a day later and was further validated in an independent motor learning dataset. Regional analyses indicated that these global changes were driven by a shift in the dominant source of regional activity modulation from sensorimotor to cognitive control networks. Together, our results suggest a fundamental principle of learning, where whole-brain functional dynamics are compressed in response to errors, providing a framework for understanding how large-scale neural activity guides behavioral adaptation.