JOVANA
Explore Library Glossary Getting Started Three Levels Fields How it works Mission
Join the mission
All guides

Diminishing Returns: The Information Theory of Many Channels

Why decodable information saturates as channels grow — and what shared variability, low-dimensional manifolds, and the noise floor have to do with it.

Bandwidth is not channel count

The information bandwidth of a BCI is what the user can actually communicate or control per second, measured in bits — not the number of electrodes. The classic scalar is the information transfer rate (ITR), which combines the number of choices, the accuracy, and the speed of selection.

B = \log_2 N + P\log_2 P + (1-P)\log_2\!\frac{1-P}{N-1}

Wolpaw's bits per selection for N equally likely targets at accuracy P. Multiplied by selections per minute it gives ITR. Note that it says nothing about channels — you can add electrodes indefinitely and move B not at all if accuracy and speed do not improve.

This reframes the whole scaling story. Channels are an input; bandwidth is the output; and the throughput a user achieves is bounded by how much task-relevant information the added channels actually carry. The rest of this guide is about why that stops growing.

Why channels saturate

Neurons are not independent sensors. They share correlated variability, and population activity for a given behaviour lives on a low-dimensional neural manifold — the dimensionality of motor-cortical activity during reaching, for instance, is on the order of ten, not hundreds. Once you have sampled those dimensions, more channels mostly re-measure them.

I = \tfrac{1}{2}\log_2 \det\!\left(\mathbf{I} + \Sigma_s\,\Sigma_n^{-1}\right)

Gaussian population information: signal covariance \Sigma_s against noise covariance \Sigma_n. When both are low-rank and the noise is shared along the informative directions, adding channels leaves the determinant almost unchanged — the mathematical face of saturation.

Revisit the scaling curve with the theory in hand: raise the channel count and watch information rise steeply, then flatten. The plateau height is set by the true dimensionality and the shared noise, not by how many electrodes you add.

I(C) \approx I_{\infty}\left(1 - e^{-C/C_0}\right)

A schematic saturating law: information approaches a ceiling I_{\infty} with a characteristic scale C_0. It is a caricature, not a derivation — but it captures the qualitative truth that beyond a few C_0, extra channels buy vanishing information.

This is information saturation with channel count, and it is not a failure of hardware — it is a property of the brain's own low-dimensional coding. It sets the plateau you saw in Guide 1 and explains why raw channel count is a poor headline number.

Redundancy: waste or robustness?

If extra channels are redundant, are they wasted? Not entirely. Spatial oversampling buys robustness: when individual electrodes fail, drift, or get walled off, a redundant array degrades gracefully instead of going dark. For a chronic implant meant to last years, redundancy that adds little information can still add crucial reliability.

The noise floor

There is a deeper limit beneath the saturation curve. Even with perfect, infinite channels, decoding is bounded by the irreducible variability of the neural computation itself — trial-to-trial fluctuation that is not measurement noise but the brain genuinely doing something slightly different each time. No electrode count reaches through that floor.

The practical consequence is a reordering of priorities. If the information plateau is close and the noise floor is real, the frontier payoff often lies not in more channels but in better models — decoders that extract more of the information already present. That is why the neural-foundation-model and latent-dynamics tracks are, in a sense, the other half of the bandwidth story.