Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Information Capacity of EEG: Theoretical and Computational Limits of Recoverable Neural Information

Ishir Rao

TL;DR

The results outline the intrinsic ceiling on how much structure about brain state or thought content can be inferred from EEG, indicating that measurement physics - not algorithmic complexity - is the dominant limitation.

Abstract

Electroencephalography (EEG) is widely used to study human brain dynamics, yet its quantitative information capacity remains unclear. Here, we combine information theory and synthetic forward modeling to estimate the mutual information between latent cortical sources and EEG recordings. Using Gaussian-channel theory and empirical simulations, we find that scalp EEG conveys only tens of bits per sample about low-dimensional neural activity. Information saturates with approximately 64-128 electrodes and scales logarithmically with signal-to-noise ratio (SNR). Linear decoders capture nearly all variance that is linearly recoverable, but the mutual information they recover remains far below the analytic channel capacity, indicating that measurement physics - not algorithmic complexity - is the dominant limitation. These results outline the intrinsic ceiling on how much structure about brain state or thought content can be inferred from EEG.

Information Capacity of EEG: Theoretical and Computational Limits of Recoverable Neural Information

TL;DR

The results outline the intrinsic ceiling on how much structure about brain state or thought content can be inferred from EEG, indicating that measurement physics - not algorithmic complexity - is the dominant limitation.

Abstract

Electroencephalography (EEG) is widely used to study human brain dynamics, yet its quantitative information capacity remains unclear. Here, we combine information theory and synthetic forward modeling to estimate the mutual information between latent cortical sources and EEG recordings. Using Gaussian-channel theory and empirical simulations, we find that scalp EEG conveys only tens of bits per sample about low-dimensional neural activity. Information saturates with approximately 64-128 electrodes and scales logarithmically with signal-to-noise ratio (SNR). Linear decoders capture nearly all variance that is linearly recoverable, but the mutual information they recover remains far below the analytic channel capacity, indicating that measurement physics - not algorithmic complexity - is the dominant limitation. These results outline the intrinsic ceiling on how much structure about brain state or thought content can be inferred from EEG.
Paper Structure (8 sections, 2 equations, 5 figures)

This paper contains 8 sections, 2 equations, 5 figures.

Table of Contents

  1. 1. Introduction
  2. 2. Model and Methods
  3. 3. Results
  4. 4. Discussion
  5. 5. Conclusion

Figures (5)

  • Figure 1: Schematic of the synthetic forward model. Cortical sources (orange) generate latent activity projected to scalp electrodes (blue) through a linear blur matrix $A$ with additive correlated noise $\varepsilon$.
  • Figure 2: Analytic (solid) and empirical (dashed) mutual information versus electrode count for different SNRs. Information growth saturates around 64--128 electrodes.
  • Figure 3: Mutual information vs. signal-to-noise ratio (SNR) for selected electrode counts. SNR dominates over electrode number in determining recoverable information.
  • Figure 4: Decoder performance ($R^2$) as a function of analytic mutual information. Both ridge (circles) and MLP (crosses) decoders achieve near-maximal linear reconstruction accuracy at high SNR.
  • Figure 5: Mutual information between true and predicted latents vs. analytic bound. Colors denote electrode count. Recovered MI reaches only a fraction of the theoretical limit.