Shared variability and noise correlations

Repeated presentations of the same movement or stimulus never produce identical population responses; the leftover trial-to-trial fluctuations are partly private to each neuron and partly shared across many. The shared part — measured as noise correlations, the correlations of residual firing after subtracting the condition mean — is typically low-dimensional and is modelled by factor-analysis-style decompositions that split total covariance into a few common latent factors plus per-neuron private noise. This shared variability is precisely what latent-state and manifold analyses treat as signal, so the line between 'noise' and 'latent dynamics' is a modelling choice rather than a fact of nature.

How much shared variability hurts or helps coding depends on its geometry relative to the signal: fluctuations aligned with the directions that distinguish conditions are the damaging kind that cannot be averaged away by adding neurons, whereas fluctuations orthogonal to the signal are comparatively harmless. In BCI this matters twice over — shared variability sets the noise floor the decoder must fight, and, because a user can only volitionally move activity along the dimensions their circuitry actually co-varies, the structure of shared variability constrains which decoder mappings a person can readily learn to control.