Power-law scaling of high-dimensional noise predicts continued information growth in mouse V1 populations

Researchers from Kyoto, Harvard, and UCLA reanalyzed existing two-photon calcium imaging datasets while mice viewed orientation or contrast differences. They decomposed trial-to-trial variability into orthogonal activity patterns, measured the strength of each pattern, and quantified its alignment with the stimulus-encoding axis. Random subsampling across population sizes revealed two consistent power laws: one governing the distribution of noise variances and another describing the cosine similarity between noise and signal directions. These exponents remained stable after size correction. The measured exponents imply that information grows as a power function of neuron count rather than saturating. Stronger noise modes align more with the signal, yet weaker modes that still carry stimulus information extend across many additional dimensions, offsetting the penalty. This pattern held in every animal and stimulus set examined, overturning the expectation that noise correlations must impose an asymptotic ceiling. The result challenges three decades of theoretical work that treated noise correlations as a hard limit on population coding. It also supplies a quantitative framework for predicting performance in noisy artificial systems whose components exhibit analogous shared fluctuations. Next experiments must record from substantially larger populations or causally manipulate the dominant noise modes to test whether the predicted scaling persists when the recorded fraction of cortex approaches completeness.