Fast timescale reparameterization of stable neural manifolds

Abstract The prevailing opinion in systems neuroscience is that motor cortex generates consistent behavior through a stable neural manifold—a low-dimensional mapping from neuron population activity to kinematics. However, standard population analyses group many repeated trials to reduce estimation noise, obscuring the time-varying activity of the network. Here, a block-by-block analysis of task trials in two species revealed that motor cortex rotates its principal subspaces as the contribution of individual neurons fluctuates, maintaining stable dynamics that are coupled to the neural state. While static decoders performed inconsistently due to non-stationary neuron activity, we confirmed that decoding performance could be rescued through subspace alignment, finding that roughly half of this geometric correction was driven by the internal rank-swapping of dominant state eigenvectors. Rather than relying on a static mapping to execute reaching behavior, motor cortex maintains a library of redundant state eigenvectors that generalize over different reach directions. Model simulations demonstrated that strong excitatory coupling combined with spike-rate adaptation is sufficient to produce a latent library whose elements can swap rank during continuous reparameterization of the network state. Together, these results refine our understanding of motor control: stable neural manifolds drift and jump within bounded limits, continuously outrunning neuron fatigue during repetitive activation.