ABSTRACT
Recorded activity of cortical neurons during behavior is often analyzed with dimensionality reduction methods. The resulting oscillatory trajectories have been interpreted as signatures of latent oscillatory dynamical systems. Here we show that oscillatory trajectories arise as a consequence of the horseshoe effect after applying dimensionality reduction methods on signals that approximately exhibit continuous variation in time, regardless of whether latent oscillatory dynamical systems are present or not. We show that task-relevant information is instead contained in deviations between oscillatory trajectories, which we account for by introducing a new model that goes beyond the dichotomy between representational and dynamical systems models of neural activity. Our findings help interpret the results of dimensionality reduction analyses.
Competing Interest Statement
The authors have declared no competing interest.
