The Bioelectrical Information Theory: Investigating the theoretical compression limit of bioelectrical signals under artificial intelligence

Technical Summary: The Bioelectrical Information Theory

Problem Statement
Bioelectrical signals are increasingly acquired at scales that exceed the bandwidth, storage, and energy constraints of practical sensing platforms, particularly in high-density neural probes and cortical microelectrode arrays. Current compression strategies are predominantly framed as waveform preservation problems, limited by the entropy of the raw signal itself. This approach treats compression as an engineering convenience for data movement rather than a fundamental prerequisite for deploying physiological intelligence under constrained power and latency budgets. The paper argues that framing compression solely around signal fidelity fails to account for the effective information content of bioelectrical data, which is determined by physiological structure, model capacity, and downstream task requirements.

Methodology: A Three-Level Hierarchical Framework
The authors propose an information-theoretic framework that reframes bioelectrical compression as a hierarchical process of information reduction, consisting of three progressive levels:

1. Signal-Level Purification:
   The raw multi-channel signal ($Z_{raw}$) is modeled as a noisy superposition of latent physiological sources ($Z_{sig}$) and noise ($N$), expressed as $Z_{raw} = AZ_{sig} + N$. The objective at this stage is not to preserve the raw waveform but to isolate the information the observation carries about the latent source. The effective information content is quantified using Gaussian vector-channel mutual information, $I(Z_{raw}; Z_{sig})$, which measures the informative part of the signal relative to the total entropy of the measurement.

2. Physiological-Level Abstraction:
   Once purified, the latent source is mapped into a compact, structured, and quantized representation ($Z_{phys}$) using parametric encoders (e.g., vector-quantized autoencoders). This stage shifts the goal from waveform reconstruction to preserving salient physiological dynamics. Entropy reduction is achieved through two mechanisms:

   • Dimensionality Reduction: Lowering the effective dimension $d$ of the representation.
   • Resolution Coarsening: Increasing the quantization step size $\delta$.
     The resulting entropy is approximated as $H(Q_{\delta}(Z)) \approx h(Z) + d \log_2(1/\delta)$, where the reduction depends on the encoder's ability to discard signal-level redundancy while retaining physiological structure.

3. Semantic-Level Conditioning:
   The final layer maps the structured physiological representation into a task-relevant semantic code ($Z_{sem}$). This level operates on the premise that not all variability in $Z_{phys}$ is necessary for a specific downstream task $T$ with target $Y_T$.

   • Task Relevance: It discards "semantic nuisance" components, retaining only the mutual information $I(Z_{phys}; Y_T)$.
   • Causal Modeling: It exploits causal and contextual dependencies within the structured representation. Unlike conventional marginal coding (e.g., Huffman), which relies on shared priors, this level utilizes deep learning models (specifically large language models or world models, denoted $M_\theta$) to model long-range dependencies. This allows the marginal entropy to be reduced to conditional entropy: $H(Z_{sem}) = \sum H_{M_\theta}(n_i | n_{<i})$.

Key Contributions

• Theoretical Reframing: The paper establishes a theoretical framework where the compression limit of bioelectrical signals is not a fixed property of the waveform but a model- and task-conditioned quantity.
• Hierarchical Decomposition: It formalizes bioelectrical compression as a three-stage hierarchy (Signal $\to$ Physiological $\to$ Semantic), providing a conceptual decomposition of total entropy reduction (Equation 7).
• Role of Generative Models: It posits that large language models and world models can significantly lower effective entropy by predicting structured physiological components based on causal-semantic context, thereby replacing the transmission of marginal data with the transmission of residual information required for task-level interpretation.

Results and Claims
The paper does not present specific experimental datasets or empirical performance metrics (e.g., specific compression ratios or reconstruction error rates). Instead, the "results" are theoretical and conceptual:

• The derivation of the mutual information bound for signal-level purification.
• The formulation of entropy reduction equations for physiological abstraction and semantic conditioning.
• The conceptual demonstration that extreme entropy reduction is possible if the system prioritizes task-relevant information and causal predictability over raw waveform fidelity.

Significance
The authors claim that this perspective fundamentally shifts the paradigm of bioelectrical compression from "signal preservation" to "semantic coding." By integrating expressive models with neural interfaces, the system need not transmit the full signal but only the residual information required to update a shared structure toward the correct task-level interpretation. This approach suggests that as models become more integrated with bioelectronic systems, compression can achieve substantially lower entropy while preserving the high-level information necessary for neural communication prostheses, health monitoring, and closed-loop control. The framework provides a theoretical basis for moving beyond the limitations of raw signal entropy in the design of future physiological intelligence systems.