| name | efficient-coding-criticality |
| description | Theoretical framework linking efficient coding to criticality in neural populations. Shows that maximizing Fisher information under resource constraints naturally leads to soft modes, diverging correlation lengths, and power-law neural avalanches, unifying statistical and dynamical perspectives of criticality. Also explains sloppiness in neural systems. Use when studying: critical brain hypothesis, neural avalanches, efficient coding theory, Fisher information in neural populations, soft modes, critical slowing down, power-law neural dynamics, or the relation between coding efficiency and brain criticality.
|
| arxiv_id | 2605.22598 |
| published | 2026-05-21 |
| authors | He Xiao, Xinyue Zhao, Weikang Wang |
| tags | ["criticality","efficient coding","Fisher information","neural avalanches","critical brain","soft modes","sloppiness","neural dynamics"] |
Efficient Coding Under Constraint Drives Neural Systems Towards Criticality and Sloppiness
arXiv:2605.22598 (Xiao, Zhao, Wang, May 2026)
Category: q-bio.NC (Neurons and Cognition)
MSC: 92B20
Core Idea
The brain operates near a critical state — neural avalanches follow power-law distributions. But why? This paper provides a theoretical framework showing that maximizing coding efficiency under resource constraints naturally drives neural populations toward criticality.
Key Contributions
1. Efficient Coding → Criticality (Mathematical Proof)
Using a Gaussian population coding model:
- Fisher Information (FI) measures coding accuracy — how well a neural population encodes a stimulus.
- Under resource constraints (limited neural firing, metabolic cost), maximizing FI forces the Fisher information matrix to develop near-zero eigenvalues (soft modes).
- Soft modes → diverging correlation lengths → hallmark of statistical criticality.
- The optimization objective:
max L = FI - λ·R where R is a resource constraint (e.g., total firing rate). At optimality, the FI matrix has zero modes — a critical point.
2. Unification of Two Criticality Perspectives
| Statistical Criticality | Dynamical Criticality |
|---|
| Diverging correlation lengths | Critical slowing down, bifurcation |
| Spatial correlations span the system | Recovery from perturbation takes infinite time |
| Arises from FI matrix soft modes | Arises from spectral properties of the dynamical operator |
The framework unifies both: introducing spatial structure (neighboring neurons have correlated tuning) connects the static FI matrix soft modes to dynamical critical slowing down — the same resource-constrained optimization that produces spatial criticality also produces temporal criticality.
3. Explanation of Sloppiness
Sloppiness = the phenomenon where neural systems are insensitive to changes in most parameter directions (only a few "stiff" directions matter).
- The soft modes of the FI matrix define sloppy directions — parameter changes along these directions barely affect coding accuracy.
- This is a natural consequence of operating at criticality: the system sacrifices sensitivity along irrelevant dimensions to maximize coding along relevant ones.
- Provides a mechanistic link between the critical brain hypothesis and sloppy model phenomenology.
4. Numerical Verification
Power-law neural avalanches emerge from the optimization, confirming the theoretical predictions.
Mathematical Framework
Gaussian Population Code
Each neuron has a tuning curve: r_i(s) = f_i(s) + ε_i where f_i(s) is the mean response to stimulus s, and ε_i is noise (Gaussian with covariance Σ).
Fisher Information Matrix
FI_ij = E[∂log p(r|s)/∂θ_i · ∂log p(r|s)/∂θ_j]
For Gaussian noise: FI(s) = J(s)^T Σ⁻¹ J(s) where J is the Jacobian of the tuning curves w.r.t. stimulus parameters.
Resource-Constrained Optimization
max_{θ} Tr(FI) - λ·||θ||² or similar regularized objectives. The key is that the constraint prevents the system from having all eigenvalues large — some must go to zero.
Soft Modes and Criticality
When FI has zero eigenvalues, the system is at a critical point in the sense of parameter space: changes along soft-mode directions don't affect the coding accuracy (neutral directions). These soft modes correspond to the diverging length scales of statistical criticality.
Unification via Spatial Structure
Introduce spatial coupling between neurons (nearby neurons have similar tuning). The FI matrix becomes approximately:
FI ≈ N·(I + α·L) where L is the graph Laplacian of the neural network. Near criticality, α → α_c at which point (I + α·L) becomes singular — the correlation length diverges AND the dynamical time scale diverges.
Relation to Existing Theories
| Theory | Connection |
|---|
| Critical Brain Hypothesis | Provides the mechanistic why — criticality is a consequence of optimal coding under constraints |
| Self-Organized Criticality (SOC) | The optimization process naturally self-organizes to criticality without fine-tuning |
| Efficient Coding Hypothesis | Directly linked — the efficiency objective itself drives criticality |
| Sloppy Model Theory | Sloppiness is a byproduct of criticality, not a separate phenomenon |
Key Results
- Fisher information maximization under resource constraints → soft modes → criticality
- Spatial structure unifies statistical and dynamical criticality
- Sloppiness emerges naturally as a consequence of critical dynamics
- Power-law avalanches confirmed numerically
Limitations
- Gaussian approximation: Real neural noise is not Gaussian — Poisson-like (mean-variance coupling). The framework may need extension to non-Gaussian noise.
- Single population: Model considers one neural population; real brains have hierarchical, multi-region organization.
- Static optimization: The optimization is at equilibrium — doesn't capture how the brain dynamically adapts to changing constraints in real time.
Activation Keywords
- critical brain hypothesis
- neural avalanches
- efficient coding
- Fisher information neural population
- soft modes neural dynamics
- critical slowing down brain
- sloppiness neural systems
- power-law neural activity
- resource-constrained coding
