Information-theoretic limits of BCI

Framing a BCI as a noisy communication channel lets us bound its performance with Shannon's theory: the achievable command rate cannot exceed the mutual information I(X;Y) between the user's intended state X and the decoded output Y, and the channel capacity C, the maximum of that mutual information over input distributions, sets a hard ceiling that no decoder, however clever, can beat. This reframes decoder engineering as approaching a capacity rather than chasing arbitrary accuracy, and makes information transfer rate, in bits per unit time, the common currency in which read and write channels can be compared.

The subtlety is that the true capacity of the neural channel is unknown and probably not stationary: the effective noise includes trial-to-trial neural variability, drift, and the user's own uncertainty, and the input alphabet, meaning what the brain can reliably and voluntarily produce, is itself the bottleneck rather than the electrode. Estimated transfer rates of today's systems, from a few bits per minute for classic P300 spellers to far higher for intracortical speech decoders, are lower bounds on capacity achieved by specific decoders, not measurements of the channel itself.