Dimensionality of neural activity

Although a population of N neurons in principle spans an N-dimensional firing-rate space, the trial-averaged activity during structured tasks typically explores only a handful of dimensions: a low-dimensional subspace captures most of the variance. Dimensionality is usually quantified from the eigenvalue spectrum of the covariance matrix, for example by the participation ratio (the squared sum of eigenvalues divided by the sum of their squares) or by the number of components needed to reach a variance threshold. This empirical low-dimensionality is the observation that motivates neural manifolds, latent-state models and dynamical analyses.

Measured dimensionality is not an intrinsic constant of the tissue; it depends on how rich the task is, how many neurons and trials were recorded, and how the data were smoothed. Simple, stereotyped tasks yield low estimates almost by construction, and undersampling can make activity look lower-dimensional than it is. Theory relates the observed dimensionality to the number and smoothness of task conditions rather than to any hard limit of the circuit, so a low number should be read as 'low relative to this task', not as evidence that the cortex only ever uses a few dimensions.