Sharp and smooth state transitions during resting state brain dynamics

Resting-state fMRI is routinely described in terms of a small set of recurring whole-brain states. Far less is known about what happens between them. Two pictures compete. In one, the brain toggles: it switches between qualitatively distinct regimes, and a transition is a discrete event. In the other, apparent states are convenient labels applied to a continuous trajectory, and transitions are mostly boundaries imposed by the analyst rather than events in the brain. We adjudicated between them with a multi-level Switching Linear Dynamical System fit to Human Connectome Project data ( $N=500$ ; 254 cortical and subcortical regions), in which every participant has their own equations of motion and switching parameters, tied to group-level parameters and estimated jointly with them. The model estimates the dynamics on either side of a boundary, so transitions can be characterized rather than merely counted. Transitions were structured and strongly heterogeneous. Four of the eight transitions examined reversed the estimated flow between two samples acquired 720 ms apart, two of them deeply, while others merely reoriented and were "one-sided", with activity flat before the switch and changing only afterwards. Sharp and smooth transitions shared destination states and engaged overlapping networks, so the heterogeneity cannot be attributed simply to hemodynamic filtering. What predicted sharpness was the direction of travel, such that the brain drifts out of a dominant, long-dwell state regime and is sharply switched into it. To link these system-level dynamics to anatomy we introduce transition importance, which asks how much the model's evidence for a particular switch depends on a region's signal. The evidence accumulated over seconds before a switch, raising its odds by 1.8 to 3.2 times, and discharged immediately after. In several transitions the regions that steer a state shared no region with those that end it, so a region's activity magnitude does not determine its contribution to system-level dynamics. Our findings suggest that spontaneous brain activity is organized as much by how the brain moves between configurations as by which ones it occupies, making transitions themselves an important target of study.