Two bandwidths hide inside one word
The word bandwidth conflates two quantities that differ by many orders of magnitude. There is the raw data rate off the electrodes — broadband voltage from a million channels is hundreds of gigabits per second — and there is the information bandwidth of the BCI, the useful bits per second the user can actually convey. Today the second number, even for the best intracortical speech neuroprostheses, is a few tens of bits per second at most. The gap between them is not waste to be engineered away; it is the subject of this entire guide.
The channel-capacity ceiling
Model the interface as a communication channel from the user's intention U to the decoder's estimate \hat{U}. Shannon's theory then hands us an upper bound: no decoder, present or future, can extract more than the mutual information I(U;\hat{U}) that the neural signal actually carries about the intention. For an additive-Gaussian channel that bound takes the familiar log-SNR form, and summing independent channels gives an optimistic ceiling.
The Gaussian channel bound. The right-hand sum assumes independent channels and so is optimistic; the next section shows why shared noise makes the true capacity far smaller. Crucially, the bound depends on SNR only logarithmically — doubling signal quality adds bits, but slowly.
The practical, task-level version of the same idea is the information transfer rate popularised by Wolpaw. For a system that makes an N-way selection every T seconds at accuracy P, the achievable rate is fixed by combinatorics and error alone — no user effort beats it. Real communication BCIs live in the range of a few bits per second, and this is a large part of why.
Wolpaw's information transfer rate. Accuracy enters through the entropy term: an error-prone high-N speller can transmit less than a slow, reliable binary one. This is why raising raw accuracy, not adding classes, is usually the honest path to a faster interface.
Why information saturates, not scales
The optimistic sum above assumes each channel adds fresh information. Real cortex does not oblige. Neurons share signal and noise: their activity lies on a low-dimensional neural manifold whose dimensionality is far smaller than the neuron count, and their trial-to-trial fluctuations are correlated. This shared variability means the N-th channel largely repeats what the first N-1 already said.
Information saturation with channel count. When noise is correlated across the population, mutual information converges to a finite ceiling I_\infty rather than growing without bound. Past a scale N_0, each doubling of channels buys a diminishing sliver of new bits.
The noise floor and the observer effect
Even a perfect, infinite-channel recorder faces two irreducible walls. The first is the brain's own irreducible variability: identical intentions produce different neural responses trial to trial, so some fraction of the signal is, from the decoder's point of view, genuine noise no averaging can remove. The second is physical — Johnson-Nyquist thermal noise in the electrode and amplifier sets a hard SNR floor that shrinks the log term in the capacity bound.
There is also a subtler, deeply BCI-specific limit: the act of interfacing perturbs what it measures. Recording draws current and displaces tissue; stimulating rewrites the very state you are trying to read. These measurement and stimulation perturbation limits mean a bidirectional interface can never be a passive tap — closing the loop changes the signal, a constraint with no analogue in engineering a radio.
Sorting the fundamental from the merely hard
Discipline for the rest of this track: label each ceiling honestly. Fundamental limits follow from information theory or physics and will not yield to better engineering — the log-SNR capacity bound, saturation under shared noise, thermodynamic heat dissipation. Engineering limits are today's hard problems that a discontinuity could move — channel count, chronic longevity, wireless data rate. Speculation is everything whose feasibility is not yet even well-posed. Most public confusion comes from filing an engineering limit, or a speculation, under the wrong heading.